blob: 8273691b8570f9c802cce1c4387763c79c467f47 [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",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700131 "src/f32-conv-hwc/3x3s2p0p1c3x4-scalar-1x1.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700132 "src/f32-conv-hwc/3x3s2p1c3x4-scalar-1x1.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -0700133 "src/f32-conv-hwc2chw/3x3s2p1c3x4-scalar-1x1.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700134 "src/f32-dwconv/gen/up1x4-minmax-scalar-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700135 "src/f32-dwconv/gen/up1x4-minmax-scalar.c",
136 "src/f32-dwconv/gen/up1x4-scalar-acc2.c",
137 "src/f32-dwconv/gen/up1x4-scalar.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700138 "src/f32-dwconv/gen/up1x9-minmax-scalar-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700139 "src/f32-dwconv/gen/up1x9-minmax-scalar.c",
140 "src/f32-dwconv/gen/up1x9-scalar-acc2.c",
141 "src/f32-dwconv/gen/up1x9-scalar.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700142 "src/f32-dwconv/gen/up1x25-minmax-scalar-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700143 "src/f32-dwconv/gen/up1x25-minmax-scalar.c",
144 "src/f32-dwconv/gen/up1x25-scalar-acc2.c",
145 "src/f32-dwconv/gen/up1x25-scalar.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700146 "src/f32-dwconv/gen/up2x4-minmax-scalar-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700147 "src/f32-dwconv/gen/up2x4-minmax-scalar.c",
148 "src/f32-dwconv/gen/up2x4-scalar-acc2.c",
149 "src/f32-dwconv/gen/up2x4-scalar.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700150 "src/f32-dwconv/gen/up2x9-minmax-scalar-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700151 "src/f32-dwconv/gen/up2x9-minmax-scalar.c",
152 "src/f32-dwconv/gen/up2x9-scalar-acc2.c",
153 "src/f32-dwconv/gen/up2x9-scalar.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700154 "src/f32-dwconv/gen/up2x25-minmax-scalar-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700155 "src/f32-dwconv/gen/up2x25-minmax-scalar.c",
156 "src/f32-dwconv/gen/up2x25-scalar-acc2.c",
157 "src/f32-dwconv/gen/up2x25-scalar.c",
Frank Barchard35db7d02020-10-26 13:37:34 -0700158 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-scalar-1x1-acc2.c",
159 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-scalar-1x1-acc3.c",
160 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-scalar-1x1-acc4.c",
Marat Dukhan91249d22020-10-24 12:02:51 -0700161 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-scalar-1x1.c",
Frank Barchard35db7d02020-10-26 13:37:34 -0700162 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-scalar-2x1-acc2.c",
Marat Dukhan91249d22020-10-24 12:02:51 -0700163 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-scalar-2x1.c",
164 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-scalar-3x1.c",
165 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-scalar-4x1.c",
166 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-scalar-5x1.c",
167 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-scalar-6x1.c",
Marat Dukhancf5b3c32020-10-25 19:21:10 -0700168 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-scalar-1x1-acc2.c",
169 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-scalar-1x1-acc3.c",
170 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-scalar-1x1-acc4.c",
Frank Barchard35db7d02020-10-26 13:37:34 -0700171 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-scalar-1x1.c",
Marat Dukhancf5b3c32020-10-25 19:21:10 -0700172 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-scalar-2x1-acc2.c",
Frank Barchard35db7d02020-10-26 13:37:34 -0700173 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-scalar-2x1.c",
174 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-scalar-3x1.c",
175 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-scalar-4x1.c",
Marat Dukhanc4efb002020-10-25 23:14:47 -0700176 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-scalar-1x1-acc2.c",
177 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-scalar-1x1-acc3.c",
178 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-scalar-1x1-acc4.c",
179 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-scalar-1x1-acc5.c",
Frank Barchard35db7d02020-10-26 13:37:34 -0700180 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-scalar-1x1.c",
Marat Dukhanc4efb002020-10-25 23:14:47 -0700181 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-scalar-2x1-acc2.c",
182 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-scalar-2x1-acc3.c",
Frank Barchard35db7d02020-10-26 13:37:34 -0700183 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-scalar-2x1.c",
Marat Dukhanc4efb002020-10-25 23:14:47 -0700184 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-scalar-3x1-acc2.c",
Frank Barchard35db7d02020-10-26 13:37:34 -0700185 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-scalar-3x1.c",
Marat Dukhan29c0c332020-10-28 22:11:00 -0700186 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-scalar-1x1-acc2.c",
187 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-scalar-1x1-acc3.c",
188 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-scalar-1x1-acc4.c",
189 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-scalar-1x1-acc5.c",
190 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-scalar-1x1.c",
191 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-scalar-2x1-acc2.c",
192 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-scalar-2x1-acc3.c",
193 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-scalar-2x1.c",
194 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-scalar-3x1-acc2.c",
195 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-scalar-3x1.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -0700196 "src/f32-gavgpool-cw/scalar-x1.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700197 "src/f32-gavgpool/7p7x-minmax-scalar-c1.c",
198 "src/f32-gavgpool/7x-minmax-scalar-c1.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700199 "src/f32-gemm/gen-inc/1x4inc-minmax-scalar.c",
200 "src/f32-gemm/gen-inc/2x4inc-minmax-scalar.c",
201 "src/f32-gemm/gen-inc/4x4inc-minmax-scalar.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700202 "src/f32-gemm/gen/1x4-minmax-scalar.c",
203 "src/f32-gemm/gen/1x4-relu-scalar.c",
204 "src/f32-gemm/gen/1x4-scalar.c",
205 "src/f32-gemm/gen/2x4-minmax-scalar.c",
206 "src/f32-gemm/gen/2x4-relu-scalar.c",
207 "src/f32-gemm/gen/2x4-scalar.c",
208 "src/f32-gemm/gen/4x2-minmax-scalar.c",
209 "src/f32-gemm/gen/4x2-relu-scalar.c",
210 "src/f32-gemm/gen/4x2-scalar.c",
211 "src/f32-gemm/gen/4x4-minmax-scalar.c",
212 "src/f32-gemm/gen/4x4-relu-scalar.c",
213 "src/f32-gemm/gen/4x4-scalar.c",
XNNPACK Team21432672020-10-19 19:58:48 -0700214 "src/f32-ibilinear-chw/gen/scalar-p1.c",
215 "src/f32-ibilinear-chw/gen/scalar-p2.c",
216 "src/f32-ibilinear-chw/gen/scalar-p4.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700217 "src/f32-ibilinear/gen/scalar-c1.c",
218 "src/f32-ibilinear/gen/scalar-c2.c",
219 "src/f32-ibilinear/gen/scalar-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700220 "src/f32-igemm/gen/1x4-minmax-scalar.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700221 "src/f32-igemm/gen/1x4-relu-scalar.c",
222 "src/f32-igemm/gen/1x4-scalar.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700223 "src/f32-igemm/gen/2x4-minmax-scalar.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700224 "src/f32-igemm/gen/2x4-relu-scalar.c",
225 "src/f32-igemm/gen/2x4-scalar.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700226 "src/f32-igemm/gen/4x2-minmax-scalar.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700227 "src/f32-igemm/gen/4x2-relu-scalar.c",
228 "src/f32-igemm/gen/4x2-scalar.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700229 "src/f32-igemm/gen/4x4-minmax-scalar.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700230 "src/f32-igemm/gen/4x4-relu-scalar.c",
231 "src/f32-igemm/gen/4x4-scalar.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700232 "src/f32-maxpool/9p8x-minmax-scalar-c1.c",
233 "src/f32-pavgpool/9p8x-minmax-scalar-c1.c",
234 "src/f32-pavgpool/9x-minmax-scalar-c1.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700235 "src/f32-ppmm/gen/2x4-minmax-scalar.c",
236 "src/f32-ppmm/gen/3x3-minmax-scalar.c",
237 "src/f32-ppmm/gen/4x2-minmax-scalar.c",
238 "src/f32-ppmm/gen/4x4-minmax-scalar.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800239 "src/f32-prelu/gen/scalar-2x1.c",
240 "src/f32-prelu/gen/scalar-2x4.c",
Marat Dukhanf46f6752020-01-21 11:03:49 -0800241 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x1.c",
Marat Dukhanf46f6752020-01-21 11:03:49 -0800242 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x2-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700243 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x2.c",
Marat Dukhanf46f6752020-01-21 11:03:49 -0800244 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x4-acc2.c",
245 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x4-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700246 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x4.c",
Marat Dukhanf46f6752020-01-21 11:03:49 -0800247 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x1.c",
Marat Dukhanf46f6752020-01-21 11:03:49 -0800248 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x2-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700249 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x2.c",
Marat Dukhanf46f6752020-01-21 11:03:49 -0800250 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x4-acc2.c",
251 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x4-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700252 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x4.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700253 "src/f32-rmax/scalar.c",
Marat Dukhan355ab432020-04-09 19:01:52 -0700254 "src/f32-spmm/gen/1x1-minmax-scalar-pipelined.c",
255 "src/f32-spmm/gen/1x1-minmax-scalar.c",
256 "src/f32-spmm/gen/2x1-minmax-scalar-pipelined.c",
257 "src/f32-spmm/gen/2x1-minmax-scalar.c",
258 "src/f32-spmm/gen/4x1-minmax-scalar-pipelined.c",
259 "src/f32-spmm/gen/4x1-minmax-scalar.c",
260 "src/f32-spmm/gen/8x1-minmax-scalar-pipelined.c",
261 "src/f32-spmm/gen/8x1-minmax-scalar.c",
262 "src/f32-spmm/gen/8x2-minmax-scalar.c",
263 "src/f32-spmm/gen/8x4-minmax-scalar.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700264 "src/f32-vbinary/gen/vadd-minmax-scalar-x1.c",
265 "src/f32-vbinary/gen/vadd-minmax-scalar-x2.c",
266 "src/f32-vbinary/gen/vadd-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700267 "src/f32-vbinary/gen/vadd-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700268 "src/f32-vbinary/gen/vadd-relu-scalar-x1.c",
269 "src/f32-vbinary/gen/vadd-relu-scalar-x2.c",
270 "src/f32-vbinary/gen/vadd-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700271 "src/f32-vbinary/gen/vadd-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700272 "src/f32-vbinary/gen/vadd-scalar-x1.c",
273 "src/f32-vbinary/gen/vadd-scalar-x2.c",
274 "src/f32-vbinary/gen/vadd-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700275 "src/f32-vbinary/gen/vadd-scalar-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700276 "src/f32-vbinary/gen/vaddc-minmax-scalar-x1.c",
277 "src/f32-vbinary/gen/vaddc-minmax-scalar-x2.c",
278 "src/f32-vbinary/gen/vaddc-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700279 "src/f32-vbinary/gen/vaddc-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700280 "src/f32-vbinary/gen/vaddc-relu-scalar-x1.c",
281 "src/f32-vbinary/gen/vaddc-relu-scalar-x2.c",
282 "src/f32-vbinary/gen/vaddc-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700283 "src/f32-vbinary/gen/vaddc-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700284 "src/f32-vbinary/gen/vaddc-scalar-x1.c",
285 "src/f32-vbinary/gen/vaddc-scalar-x2.c",
286 "src/f32-vbinary/gen/vaddc-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700287 "src/f32-vbinary/gen/vaddc-scalar-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700288 "src/f32-vbinary/gen/vdiv-minmax-scalar-x1.c",
289 "src/f32-vbinary/gen/vdiv-minmax-scalar-x2.c",
290 "src/f32-vbinary/gen/vdiv-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700291 "src/f32-vbinary/gen/vdiv-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700292 "src/f32-vbinary/gen/vdiv-relu-scalar-x1.c",
293 "src/f32-vbinary/gen/vdiv-relu-scalar-x2.c",
294 "src/f32-vbinary/gen/vdiv-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700295 "src/f32-vbinary/gen/vdiv-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700296 "src/f32-vbinary/gen/vdiv-scalar-x1.c",
297 "src/f32-vbinary/gen/vdiv-scalar-x2.c",
298 "src/f32-vbinary/gen/vdiv-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700299 "src/f32-vbinary/gen/vdiv-scalar-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700300 "src/f32-vbinary/gen/vdivc-minmax-scalar-x1.c",
301 "src/f32-vbinary/gen/vdivc-minmax-scalar-x2.c",
302 "src/f32-vbinary/gen/vdivc-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700303 "src/f32-vbinary/gen/vdivc-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700304 "src/f32-vbinary/gen/vdivc-relu-scalar-x1.c",
305 "src/f32-vbinary/gen/vdivc-relu-scalar-x2.c",
306 "src/f32-vbinary/gen/vdivc-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700307 "src/f32-vbinary/gen/vdivc-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700308 "src/f32-vbinary/gen/vdivc-scalar-x1.c",
309 "src/f32-vbinary/gen/vdivc-scalar-x2.c",
310 "src/f32-vbinary/gen/vdivc-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700311 "src/f32-vbinary/gen/vdivc-scalar-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800312 "src/f32-vbinary/gen/vmax-scalar-x1.c",
313 "src/f32-vbinary/gen/vmax-scalar-x2.c",
314 "src/f32-vbinary/gen/vmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700315 "src/f32-vbinary/gen/vmax-scalar-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800316 "src/f32-vbinary/gen/vmaxc-scalar-x1.c",
317 "src/f32-vbinary/gen/vmaxc-scalar-x2.c",
318 "src/f32-vbinary/gen/vmaxc-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700319 "src/f32-vbinary/gen/vmaxc-scalar-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800320 "src/f32-vbinary/gen/vmin-scalar-x1.c",
321 "src/f32-vbinary/gen/vmin-scalar-x2.c",
322 "src/f32-vbinary/gen/vmin-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700323 "src/f32-vbinary/gen/vmin-scalar-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800324 "src/f32-vbinary/gen/vminc-scalar-x1.c",
325 "src/f32-vbinary/gen/vminc-scalar-x2.c",
326 "src/f32-vbinary/gen/vminc-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700327 "src/f32-vbinary/gen/vminc-scalar-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700328 "src/f32-vbinary/gen/vmul-minmax-scalar-x1.c",
329 "src/f32-vbinary/gen/vmul-minmax-scalar-x2.c",
330 "src/f32-vbinary/gen/vmul-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700331 "src/f32-vbinary/gen/vmul-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700332 "src/f32-vbinary/gen/vmul-relu-scalar-x1.c",
333 "src/f32-vbinary/gen/vmul-relu-scalar-x2.c",
334 "src/f32-vbinary/gen/vmul-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700335 "src/f32-vbinary/gen/vmul-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700336 "src/f32-vbinary/gen/vmul-scalar-x1.c",
337 "src/f32-vbinary/gen/vmul-scalar-x2.c",
338 "src/f32-vbinary/gen/vmul-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700339 "src/f32-vbinary/gen/vmul-scalar-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700340 "src/f32-vbinary/gen/vmulc-minmax-scalar-x1.c",
341 "src/f32-vbinary/gen/vmulc-minmax-scalar-x2.c",
342 "src/f32-vbinary/gen/vmulc-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700343 "src/f32-vbinary/gen/vmulc-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700344 "src/f32-vbinary/gen/vmulc-relu-scalar-x1.c",
345 "src/f32-vbinary/gen/vmulc-relu-scalar-x2.c",
346 "src/f32-vbinary/gen/vmulc-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700347 "src/f32-vbinary/gen/vmulc-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700348 "src/f32-vbinary/gen/vmulc-scalar-x1.c",
349 "src/f32-vbinary/gen/vmulc-scalar-x2.c",
350 "src/f32-vbinary/gen/vmulc-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700351 "src/f32-vbinary/gen/vmulc-scalar-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700352 "src/f32-vbinary/gen/vrdivc-minmax-scalar-x1.c",
353 "src/f32-vbinary/gen/vrdivc-minmax-scalar-x2.c",
354 "src/f32-vbinary/gen/vrdivc-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700355 "src/f32-vbinary/gen/vrdivc-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700356 "src/f32-vbinary/gen/vrdivc-relu-scalar-x1.c",
357 "src/f32-vbinary/gen/vrdivc-relu-scalar-x2.c",
358 "src/f32-vbinary/gen/vrdivc-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700359 "src/f32-vbinary/gen/vrdivc-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700360 "src/f32-vbinary/gen/vrdivc-scalar-x1.c",
361 "src/f32-vbinary/gen/vrdivc-scalar-x2.c",
362 "src/f32-vbinary/gen/vrdivc-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700363 "src/f32-vbinary/gen/vrdivc-scalar-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700364 "src/f32-vbinary/gen/vrsubc-minmax-scalar-x1.c",
365 "src/f32-vbinary/gen/vrsubc-minmax-scalar-x2.c",
366 "src/f32-vbinary/gen/vrsubc-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700367 "src/f32-vbinary/gen/vrsubc-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700368 "src/f32-vbinary/gen/vrsubc-relu-scalar-x1.c",
369 "src/f32-vbinary/gen/vrsubc-relu-scalar-x2.c",
370 "src/f32-vbinary/gen/vrsubc-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700371 "src/f32-vbinary/gen/vrsubc-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700372 "src/f32-vbinary/gen/vrsubc-scalar-x1.c",
373 "src/f32-vbinary/gen/vrsubc-scalar-x2.c",
374 "src/f32-vbinary/gen/vrsubc-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700375 "src/f32-vbinary/gen/vrsubc-scalar-x8.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -0700376 "src/f32-vbinary/gen/vsqrdiff-scalar-x1.c",
377 "src/f32-vbinary/gen/vsqrdiff-scalar-x2.c",
378 "src/f32-vbinary/gen/vsqrdiff-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700379 "src/f32-vbinary/gen/vsqrdiff-scalar-x8.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -0700380 "src/f32-vbinary/gen/vsqrdiffc-scalar-x1.c",
381 "src/f32-vbinary/gen/vsqrdiffc-scalar-x2.c",
382 "src/f32-vbinary/gen/vsqrdiffc-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700383 "src/f32-vbinary/gen/vsqrdiffc-scalar-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700384 "src/f32-vbinary/gen/vsub-minmax-scalar-x1.c",
385 "src/f32-vbinary/gen/vsub-minmax-scalar-x2.c",
386 "src/f32-vbinary/gen/vsub-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700387 "src/f32-vbinary/gen/vsub-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700388 "src/f32-vbinary/gen/vsub-relu-scalar-x1.c",
389 "src/f32-vbinary/gen/vsub-relu-scalar-x2.c",
390 "src/f32-vbinary/gen/vsub-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700391 "src/f32-vbinary/gen/vsub-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700392 "src/f32-vbinary/gen/vsub-scalar-x1.c",
393 "src/f32-vbinary/gen/vsub-scalar-x2.c",
394 "src/f32-vbinary/gen/vsub-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700395 "src/f32-vbinary/gen/vsub-scalar-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700396 "src/f32-vbinary/gen/vsubc-minmax-scalar-x1.c",
397 "src/f32-vbinary/gen/vsubc-minmax-scalar-x2.c",
398 "src/f32-vbinary/gen/vsubc-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700399 "src/f32-vbinary/gen/vsubc-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700400 "src/f32-vbinary/gen/vsubc-relu-scalar-x1.c",
401 "src/f32-vbinary/gen/vsubc-relu-scalar-x2.c",
402 "src/f32-vbinary/gen/vsubc-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700403 "src/f32-vbinary/gen/vsubc-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700404 "src/f32-vbinary/gen/vsubc-scalar-x1.c",
405 "src/f32-vbinary/gen/vsubc-scalar-x2.c",
406 "src/f32-vbinary/gen/vsubc-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700407 "src/f32-vbinary/gen/vsubc-scalar-x8.c",
Marat Dukhanb1eec082021-05-05 23:24:55 -0700408 "src/f32-vclamp/gen/vclamp-scalar-x1.c",
409 "src/f32-vclamp/gen/vclamp-scalar-x2.c",
410 "src/f32-vclamp/gen/vclamp-scalar-x4.c",
Marat Dukhaned6baaf2020-12-01 15:07:08 -0800411 "src/f32-velu/gen/velu-scalar-rr2-lut16-p3-x1.c",
412 "src/f32-velu/gen/velu-scalar-rr2-lut16-p3-x2.c",
413 "src/f32-velu/gen/velu-scalar-rr2-lut16-p3-x3.c",
414 "src/f32-velu/gen/velu-scalar-rr2-lut16-p3-x4.c",
415 "src/f32-velu/gen/velu-scalar-rr2-lut16-p3-x5.c",
416 "src/f32-velu/gen/velu-scalar-rr2-lut16-p3-x6.c",
417 "src/f32-velu/gen/velu-scalar-rr2-p6-x1.c",
418 "src/f32-velu/gen/velu-scalar-rr2-p6-x2.c",
419 "src/f32-velu/gen/velu-scalar-rr2-p6-x3.c",
420 "src/f32-velu/gen/velu-scalar-rr2-p6-x4.c",
421 "src/f32-velu/gen/velu-scalar-rr2-p6-x5.c",
422 "src/f32-velu/gen/velu-scalar-rr2-p6-x6.c",
Marat Dukhan6674d692021-05-05 22:27:00 -0700423 "src/f32-vhswish/gen/vhswish-scalar-x1.c",
424 "src/f32-vhswish/gen/vhswish-scalar-x2.c",
425 "src/f32-vhswish/gen/vhswish-scalar-x4.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700426 "src/f32-vlrelu/gen/vlrelu-scalar-x1.c",
427 "src/f32-vlrelu/gen/vlrelu-scalar-x2.c",
428 "src/f32-vlrelu/gen/vlrelu-scalar-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700429 "src/f32-vmulcaddc/gen/c1-minmax-scalar-2x.c",
430 "src/f32-vmulcaddc/gen/c2-minmax-scalar-2x.c",
431 "src/f32-vmulcaddc/gen/c4-minmax-scalar-2x.c",
Marat Dukhanb1eec082021-05-05 23:24:55 -0700432 "src/f32-vrelu/gen/vrelu-scalar-x1.c",
433 "src/f32-vrelu/gen/vrelu-scalar-x2.c",
434 "src/f32-vrelu/gen/vrelu-scalar-x4.c",
435 "src/f32-vrelu/gen/vrelu-scalar-x8.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -0700436 "src/f32-vrnd/gen/vrndd-scalar-libm-x1.c",
437 "src/f32-vrnd/gen/vrndd-scalar-libm-x2.c",
438 "src/f32-vrnd/gen/vrndd-scalar-libm-x4.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700439 "src/f32-vrnd/gen/vrndne-scalar-libm-x1.c",
440 "src/f32-vrnd/gen/vrndne-scalar-libm-x2.c",
441 "src/f32-vrnd/gen/vrndne-scalar-libm-x4.c",
442 "src/f32-vrnd/gen/vrndu-scalar-libm-x1.c",
443 "src/f32-vrnd/gen/vrndu-scalar-libm-x2.c",
444 "src/f32-vrnd/gen/vrndu-scalar-libm-x4.c",
445 "src/f32-vrnd/gen/vrndz-scalar-libm-x1.c",
446 "src/f32-vrnd/gen/vrndz-scalar-libm-x2.c",
447 "src/f32-vrnd/gen/vrndz-scalar-libm-x4.c",
Marat Dukhanb1eec082021-05-05 23:24:55 -0700448 "src/f32-vsigmoid/gen/vsigmoid-scalar-lut64-p2-div-x1.c",
449 "src/f32-vsigmoid/gen/vsigmoid-scalar-lut64-p2-div-x2.c",
450 "src/f32-vsigmoid/gen/vsigmoid-scalar-lut64-p2-div-x4.c",
451 "src/f32-vsigmoid/gen/vsigmoid-scalar-lut2048-p1-div-x1.c",
452 "src/f32-vsigmoid/gen/vsigmoid-scalar-lut2048-p1-div-x2.c",
453 "src/f32-vsigmoid/gen/vsigmoid-scalar-lut2048-p1-div-x4.c",
454 "src/f32-vsigmoid/gen/vsigmoid-scalar-p5-div-x1.c",
455 "src/f32-vsigmoid/gen/vsigmoid-scalar-p5-div-x2.c",
456 "src/f32-vsigmoid/gen/vsigmoid-scalar-p5-div-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 Dukhan44542882021-05-10 09:21:48 -0700494 "src/qs8-dwconv/gen/up1x9-minmax-scalar.c",
Marat Dukhan4ed14882021-05-12 17:50:40 -0700495 "src/qs8-dwconv/gen/up1x25-minmax-scalar.c",
Marat Dukhan44542882021-05-10 09:21:48 -0700496 "src/qs8-dwconv/gen/up2x9-minmax-scalar.c",
Marat Dukhan4ed14882021-05-12 17:50:40 -0700497 "src/qs8-dwconv/gen/up2x25-minmax-scalar.c",
Marat Dukhan44542882021-05-10 09:21:48 -0700498 "src/qs8-dwconv/gen/up4x9-minmax-scalar.c",
Marat Dukhan4ed14882021-05-12 17:50:40 -0700499 "src/qs8-dwconv/gen/up4x25-minmax-scalar.c",
Marat Dukhan047b6202021-05-11 20:32:25 -0700500 "src/qs8-gavgpool/gen/7p7x-minmax-scalar-c1.c",
501 "src/qs8-gavgpool/gen/7p7x-minmax-scalar-c2.c",
502 "src/qs8-gavgpool/gen/7p7x-minmax-scalar-c4.c",
503 "src/qs8-gavgpool/gen/7x-minmax-scalar-c1.c",
504 "src/qs8-gavgpool/gen/7x-minmax-scalar-c2.c",
505 "src/qs8-gavgpool/gen/7x-minmax-scalar-c4.c",
Marat Dukhana1a4e782021-05-07 17:49:47 -0700506 "src/qs8-gemm/gen/1x2-minmax-scalar.c",
507 "src/qs8-gemm/gen/2x2-minmax-scalar.c",
508 "src/qs8-gemm/gen/3x2-minmax-scalar.c",
509 "src/qs8-gemm/gen/4x2-minmax-scalar.c",
510 "src/qs8-gemm/gen/1x4-minmax-scalar.c",
511 "src/qs8-gemm/gen/2x4-minmax-scalar.c",
512 "src/qs8-gemm/gen/3x4-minmax-scalar.c",
513 "src/qs8-gemm/gen/4x4-minmax-scalar.c",
514 "src/qs8-igemm/gen/1x2-minmax-scalar.c",
515 "src/qs8-igemm/gen/2x2-minmax-scalar.c",
516 "src/qs8-igemm/gen/3x2-minmax-scalar.c",
517 "src/qs8-igemm/gen/4x2-minmax-scalar.c",
518 "src/qs8-igemm/gen/1x4-minmax-scalar.c",
519 "src/qs8-igemm/gen/2x4-minmax-scalar.c",
520 "src/qs8-igemm/gen/3x4-minmax-scalar.c",
521 "src/qs8-igemm/gen/4x4-minmax-scalar.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -0700522 "src/qs8-requantization/fp32-scalar-lrintf.c",
523 "src/qs8-requantization/fp32-scalar-magic.c",
524 "src/qs8-requantization/precise-scalar-signed64.c",
525 "src/qs8-requantization/precise-scalar-unsigned32.c",
526 "src/qs8-requantization/precise-scalar-unsigned64.c",
527 "src/qs8-requantization/q31-scalar.c",
Marat Dukhand481c282021-05-11 23:48:31 -0700528 "src/qs8-vadd/gen/minmax-scalar-x1.c",
529 "src/qs8-vadd/gen/minmax-scalar-x2.c",
530 "src/qs8-vadd/gen/minmax-scalar-x4.c",
531 "src/qs8-vaddc/gen/minmax-scalar-x1.c",
532 "src/qs8-vaddc/gen/minmax-scalar-x2.c",
533 "src/qs8-vaddc/gen/minmax-scalar-x4.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -0700534 "src/qu8-avgpool/9p8x-minmax-scalar-c1.c",
535 "src/qu8-avgpool/9x-minmax-scalar-c1.c",
536 "src/qu8-dwconv/up1x9-minmax-scalar.c",
537 "src/qu8-gavgpool/7p7x-minmax-scalar-c1.c",
538 "src/qu8-gavgpool/7x-minmax-scalar-c1.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700539 "src/qu8-gemm/2x2-minmax-scalar.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -0700540 "src/qu8-igemm/2x2-minmax-scalar.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -0700541 "src/qu8-requantization/fp32-scalar-lrintf.c",
542 "src/qu8-requantization/fp32-scalar-magic.c",
543 "src/qu8-requantization/precise-scalar-signed64.c",
544 "src/qu8-requantization/precise-scalar-unsigned32.c",
545 "src/qu8-requantization/precise-scalar-unsigned64.c",
546 "src/qu8-requantization/q31-scalar.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -0700547 "src/qu8-vadd/minmax-scalar.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700548 "src/u8-lut32norm/scalar.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700549 "src/u8-maxpool/9p8x-minmax-scalar-c1.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700550 "src/u8-rmax/scalar.c",
Marat Dukhan6674d692021-05-05 22:27:00 -0700551 "src/u8-vclamp/scalar-x4.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700552 "src/x8-lut/scalar.c",
553 "src/x8-zip/x2-scalar.c",
554 "src/x8-zip/x3-scalar.c",
555 "src/x8-zip/x4-scalar.c",
556 "src/x8-zip/xm-scalar.c",
Marat Dukhanad71b9a2020-11-20 00:01:51 -0800557 "src/x32-depthtospace2d-chw2hwc/scalar.c",
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -0700558 "src/x32-fill/scalar-float.c",
559 "src/x32-fill/scalar-int.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700560 "src/x32-packx/x2-scalar.c",
561 "src/x32-packx/x3-scalar.c",
562 "src/x32-packx/x4-scalar.c",
Marat Dukhan63523d42020-05-22 17:07:33 -0700563 "src/x32-pad/scalar-float.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700564 "src/x32-pad/scalar-int.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700565 "src/x32-unpool/scalar.c",
566 "src/x32-zip/x2-scalar.c",
567 "src/x32-zip/x3-scalar.c",
568 "src/x32-zip/x4-scalar.c",
569 "src/x32-zip/xm-scalar.c",
Marat Dukhan048931b2020-11-24 20:53:54 -0800570 "src/xx-copy/memcpy.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700571]
572
Marat Dukhan436ebe62019-12-04 15:10:12 -0800573WASM_UKERNELS = [
Marat Dukhan99936602020-04-11 16:47:01 -0700574 "src/f32-avgpool/9p8x-minmax-wasm-c1.c",
575 "src/f32-avgpool/9x-minmax-wasm-c1.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700576 "src/f32-dwconv/gen/up1x4-minmax-wasm-acc2.c",
577 "src/f32-dwconv/gen/up1x4-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700578 "src/f32-dwconv/gen/up1x4-wasm-acc2.c",
579 "src/f32-dwconv/gen/up1x4-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700580 "src/f32-dwconv/gen/up1x9-minmax-wasm-acc2.c",
581 "src/f32-dwconv/gen/up1x9-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700582 "src/f32-dwconv/gen/up1x9-wasm-acc2.c",
583 "src/f32-dwconv/gen/up1x9-wasm.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700584 "src/f32-dwconv/gen/up1x25-minmax-wasm-acc2.c",
585 "src/f32-dwconv/gen/up1x25-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700586 "src/f32-dwconv/gen/up1x25-wasm-acc2.c",
587 "src/f32-dwconv/gen/up1x25-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700588 "src/f32-dwconv/gen/up2x4-minmax-wasm-acc2.c",
589 "src/f32-dwconv/gen/up2x4-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700590 "src/f32-dwconv/gen/up2x4-wasm-acc2.c",
591 "src/f32-dwconv/gen/up2x4-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700592 "src/f32-dwconv/gen/up2x9-minmax-wasm-acc2.c",
593 "src/f32-dwconv/gen/up2x9-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700594 "src/f32-dwconv/gen/up2x9-wasm-acc2.c",
595 "src/f32-dwconv/gen/up2x9-wasm.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700596 "src/f32-dwconv/gen/up2x25-minmax-wasm-acc2.c",
597 "src/f32-dwconv/gen/up2x25-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700598 "src/f32-dwconv/gen/up2x25-wasm-acc2.c",
599 "src/f32-dwconv/gen/up2x25-wasm.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700600 "src/f32-gavgpool/7p7x-minmax-wasm-c1.c",
601 "src/f32-gavgpool/7x-minmax-wasm-c1.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700602 "src/f32-gemm/gen-inc/1x4inc-minmax-wasm.c",
603 "src/f32-gemm/gen-inc/2x4inc-minmax-wasm.c",
604 "src/f32-gemm/gen-inc/4x4inc-minmax-wasm.c",
605 "src/f32-gemm/gen/1x4-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700606 "src/f32-gemm/gen/1x4-relu-wasm.c",
607 "src/f32-gemm/gen/1x4-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700608 "src/f32-gemm/gen/2x4-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700609 "src/f32-gemm/gen/2x4-relu-wasm.c",
610 "src/f32-gemm/gen/2x4-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700611 "src/f32-gemm/gen/4x2-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700612 "src/f32-gemm/gen/4x2-relu-wasm.c",
613 "src/f32-gemm/gen/4x2-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700614 "src/f32-gemm/gen/4x4-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700615 "src/f32-gemm/gen/4x4-relu-wasm.c",
616 "src/f32-gemm/gen/4x4-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700617 "src/f32-igemm/gen/1x4-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700618 "src/f32-igemm/gen/1x4-relu-wasm.c",
619 "src/f32-igemm/gen/1x4-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700620 "src/f32-igemm/gen/2x4-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700621 "src/f32-igemm/gen/2x4-relu-wasm.c",
622 "src/f32-igemm/gen/2x4-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700623 "src/f32-igemm/gen/4x2-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700624 "src/f32-igemm/gen/4x2-relu-wasm.c",
625 "src/f32-igemm/gen/4x2-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700626 "src/f32-igemm/gen/4x4-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700627 "src/f32-igemm/gen/4x4-relu-wasm.c",
628 "src/f32-igemm/gen/4x4-wasm.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700629 "src/f32-maxpool/9p8x-minmax-wasm-c1.c",
630 "src/f32-pavgpool/9p8x-minmax-wasm-c1.c",
631 "src/f32-pavgpool/9x-minmax-wasm-c1.c",
Marat Dukhan7c1f8082020-06-25 13:26:20 -0700632 "src/f32-prelu/gen/wasm-2x1.c",
633 "src/f32-prelu/gen/wasm-2x4.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700634 "src/f32-vbinary/gen/vadd-minmax-wasm-x1.c",
635 "src/f32-vbinary/gen/vadd-minmax-wasm-x2.c",
636 "src/f32-vbinary/gen/vadd-minmax-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700637 "src/f32-vbinary/gen/vadd-minmax-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700638 "src/f32-vbinary/gen/vadd-relu-wasm-x1.c",
639 "src/f32-vbinary/gen/vadd-relu-wasm-x2.c",
640 "src/f32-vbinary/gen/vadd-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700641 "src/f32-vbinary/gen/vadd-relu-wasm-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700642 "src/f32-vbinary/gen/vaddc-minmax-wasm-x1.c",
643 "src/f32-vbinary/gen/vaddc-minmax-wasm-x2.c",
644 "src/f32-vbinary/gen/vaddc-minmax-wasm-x4.c",
645 "src/f32-vbinary/gen/vaddc-minmax-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700646 "src/f32-vbinary/gen/vaddc-relu-wasm-x1.c",
647 "src/f32-vbinary/gen/vaddc-relu-wasm-x2.c",
648 "src/f32-vbinary/gen/vaddc-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700649 "src/f32-vbinary/gen/vaddc-relu-wasm-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700650 "src/f32-vbinary/gen/vdiv-minmax-wasm-x1.c",
651 "src/f32-vbinary/gen/vdiv-minmax-wasm-x2.c",
652 "src/f32-vbinary/gen/vdiv-minmax-wasm-x4.c",
653 "src/f32-vbinary/gen/vdiv-minmax-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700654 "src/f32-vbinary/gen/vdiv-relu-wasm-x1.c",
655 "src/f32-vbinary/gen/vdiv-relu-wasm-x2.c",
656 "src/f32-vbinary/gen/vdiv-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700657 "src/f32-vbinary/gen/vdiv-relu-wasm-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700658 "src/f32-vbinary/gen/vdivc-minmax-wasm-x1.c",
659 "src/f32-vbinary/gen/vdivc-minmax-wasm-x2.c",
660 "src/f32-vbinary/gen/vdivc-minmax-wasm-x4.c",
661 "src/f32-vbinary/gen/vdivc-minmax-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700662 "src/f32-vbinary/gen/vdivc-relu-wasm-x1.c",
663 "src/f32-vbinary/gen/vdivc-relu-wasm-x2.c",
664 "src/f32-vbinary/gen/vdivc-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700665 "src/f32-vbinary/gen/vdivc-relu-wasm-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800666 "src/f32-vbinary/gen/vmax-wasm-x1.c",
667 "src/f32-vbinary/gen/vmax-wasm-x2.c",
668 "src/f32-vbinary/gen/vmax-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700669 "src/f32-vbinary/gen/vmax-wasm-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800670 "src/f32-vbinary/gen/vmaxc-wasm-x1.c",
671 "src/f32-vbinary/gen/vmaxc-wasm-x2.c",
672 "src/f32-vbinary/gen/vmaxc-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700673 "src/f32-vbinary/gen/vmaxc-wasm-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800674 "src/f32-vbinary/gen/vmin-wasm-x1.c",
675 "src/f32-vbinary/gen/vmin-wasm-x2.c",
676 "src/f32-vbinary/gen/vmin-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700677 "src/f32-vbinary/gen/vmin-wasm-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800678 "src/f32-vbinary/gen/vminc-wasm-x1.c",
679 "src/f32-vbinary/gen/vminc-wasm-x2.c",
680 "src/f32-vbinary/gen/vminc-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700681 "src/f32-vbinary/gen/vminc-wasm-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700682 "src/f32-vbinary/gen/vmul-minmax-wasm-x1.c",
683 "src/f32-vbinary/gen/vmul-minmax-wasm-x2.c",
684 "src/f32-vbinary/gen/vmul-minmax-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700685 "src/f32-vbinary/gen/vmul-minmax-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700686 "src/f32-vbinary/gen/vmul-relu-wasm-x1.c",
687 "src/f32-vbinary/gen/vmul-relu-wasm-x2.c",
688 "src/f32-vbinary/gen/vmul-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700689 "src/f32-vbinary/gen/vmul-relu-wasm-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700690 "src/f32-vbinary/gen/vmulc-minmax-wasm-x1.c",
691 "src/f32-vbinary/gen/vmulc-minmax-wasm-x2.c",
692 "src/f32-vbinary/gen/vmulc-minmax-wasm-x4.c",
693 "src/f32-vbinary/gen/vmulc-minmax-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700694 "src/f32-vbinary/gen/vmulc-relu-wasm-x1.c",
695 "src/f32-vbinary/gen/vmulc-relu-wasm-x2.c",
696 "src/f32-vbinary/gen/vmulc-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700697 "src/f32-vbinary/gen/vmulc-relu-wasm-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700698 "src/f32-vbinary/gen/vrdivc-minmax-wasm-x1.c",
699 "src/f32-vbinary/gen/vrdivc-minmax-wasm-x2.c",
700 "src/f32-vbinary/gen/vrdivc-minmax-wasm-x4.c",
701 "src/f32-vbinary/gen/vrdivc-minmax-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700702 "src/f32-vbinary/gen/vrdivc-relu-wasm-x1.c",
703 "src/f32-vbinary/gen/vrdivc-relu-wasm-x2.c",
704 "src/f32-vbinary/gen/vrdivc-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700705 "src/f32-vbinary/gen/vrdivc-relu-wasm-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700706 "src/f32-vbinary/gen/vrsubc-minmax-wasm-x1.c",
707 "src/f32-vbinary/gen/vrsubc-minmax-wasm-x2.c",
708 "src/f32-vbinary/gen/vrsubc-minmax-wasm-x4.c",
709 "src/f32-vbinary/gen/vrsubc-minmax-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700710 "src/f32-vbinary/gen/vrsubc-relu-wasm-x1.c",
711 "src/f32-vbinary/gen/vrsubc-relu-wasm-x2.c",
712 "src/f32-vbinary/gen/vrsubc-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700713 "src/f32-vbinary/gen/vrsubc-relu-wasm-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700714 "src/f32-vbinary/gen/vsub-minmax-wasm-x1.c",
715 "src/f32-vbinary/gen/vsub-minmax-wasm-x2.c",
716 "src/f32-vbinary/gen/vsub-minmax-wasm-x4.c",
717 "src/f32-vbinary/gen/vsub-minmax-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700718 "src/f32-vbinary/gen/vsub-relu-wasm-x1.c",
719 "src/f32-vbinary/gen/vsub-relu-wasm-x2.c",
720 "src/f32-vbinary/gen/vsub-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700721 "src/f32-vbinary/gen/vsub-relu-wasm-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700722 "src/f32-vbinary/gen/vsubc-minmax-wasm-x1.c",
723 "src/f32-vbinary/gen/vsubc-minmax-wasm-x2.c",
724 "src/f32-vbinary/gen/vsubc-minmax-wasm-x4.c",
725 "src/f32-vbinary/gen/vsubc-minmax-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700726 "src/f32-vbinary/gen/vsubc-relu-wasm-x1.c",
727 "src/f32-vbinary/gen/vsubc-relu-wasm-x2.c",
728 "src/f32-vbinary/gen/vsubc-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700729 "src/f32-vbinary/gen/vsubc-relu-wasm-x8.c",
Marat Dukhanb1eec082021-05-05 23:24:55 -0700730 "src/f32-vclamp/gen/vclamp-wasm-x1.c",
731 "src/f32-vclamp/gen/vclamp-wasm-x2.c",
732 "src/f32-vclamp/gen/vclamp-wasm-x4.c",
Marat Dukhaned6baaf2020-12-01 15:07:08 -0800733 "src/f32-velu/gen/velu-wasm-rr2-lut16-p3-x1.c",
734 "src/f32-velu/gen/velu-wasm-rr2-lut16-p3-x2.c",
735 "src/f32-velu/gen/velu-wasm-rr2-lut16-p3-x3.c",
736 "src/f32-velu/gen/velu-wasm-rr2-lut16-p3-x4.c",
737 "src/f32-velu/gen/velu-wasm-rr2-lut16-p3-x5.c",
738 "src/f32-velu/gen/velu-wasm-rr2-lut16-p3-x6.c",
739 "src/f32-velu/gen/velu-wasm-rr2-p6-x1.c",
740 "src/f32-velu/gen/velu-wasm-rr2-p6-x2.c",
741 "src/f32-velu/gen/velu-wasm-rr2-p6-x3.c",
742 "src/f32-velu/gen/velu-wasm-rr2-p6-x4.c",
743 "src/f32-velu/gen/velu-wasm-rr2-p6-x5.c",
744 "src/f32-velu/gen/velu-wasm-rr2-p6-x6.c",
Marat Dukhan6674d692021-05-05 22:27:00 -0700745 "src/f32-vhswish/gen/vhswish-wasm-x1.c",
746 "src/f32-vhswish/gen/vhswish-wasm-x2.c",
747 "src/f32-vhswish/gen/vhswish-wasm-x4.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -0700748 "src/f32-vlrelu/gen/vlrelu-wasm-x1.c",
749 "src/f32-vlrelu/gen/vlrelu-wasm-x2.c",
750 "src/f32-vlrelu/gen/vlrelu-wasm-x4.c",
Marat Dukhanb1eec082021-05-05 23:24:55 -0700751 "src/f32-vrelu/gen/vrelu-wasm-x1.c",
752 "src/f32-vrelu/gen/vrelu-wasm-x2.c",
753 "src/f32-vrelu/gen/vrelu-wasm-x4.c",
754 "src/f32-vrelu/gen/vrelu-wasm-x8.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700755 "src/f32-vmulcaddc/gen/c1-minmax-wasm-2x.c",
756 "src/f32-vmulcaddc/gen/c2-minmax-wasm-2x.c",
757 "src/f32-vmulcaddc/gen/c4-minmax-wasm-2x.c",
Marat Dukhan436ebe62019-12-04 15:10:12 -0800758]
759
Marat Dukhan290055c2020-06-09 12:24:29 -0700760WASMSIMD_UKERNELS = [
Marat Dukhan40f05522020-07-16 22:33:12 -0700761 "src/f32-argmaxpool/4x-wasmsimd-c4.c",
762 "src/f32-argmaxpool/9p8x-wasmsimd-c4.c",
763 "src/f32-argmaxpool/9x-wasmsimd-c4.c",
Marat Dukhan3b7432d2020-07-16 17:46:32 -0700764 "src/f32-avgpool/9p8x-minmax-wasmsimd-arm-c4.c",
765 "src/f32-avgpool/9p8x-minmax-wasmsimd-x86-c4.c",
766 "src/f32-avgpool/9x-minmax-wasmsimd-arm-c4.c",
767 "src/f32-avgpool/9x-minmax-wasmsimd-x86-c4.c",
Frank Barchard22136062020-11-24 18:44:46 -0800768 "src/f32-conv-hwc2chw/3x3s2p1c3x4-wasmsimd-2x2.c",
Frank Barchard0725b8d2020-12-07 11:07:35 -0800769 "src/f32-dwconv/gen/up4x4-minmax-wasmsimd-arm-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700770 "src/f32-dwconv/gen/up4x4-minmax-wasmsimd-arm.c",
Frank Barchard0725b8d2020-12-07 11:07:35 -0800771 "src/f32-dwconv/gen/up4x4-minmax-wasmsimd-x86-acc2.c",
Marat Dukhanac014d72020-06-16 08:36:47 -0700772 "src/f32-dwconv/gen/up4x4-minmax-wasmsimd-x86.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700773 "src/f32-dwconv/gen/up4x4-wasmsimd.c",
Frank Barchard0725b8d2020-12-07 11:07:35 -0800774 "src/f32-dwconv/gen/up4x9-minmax-wasmsimd-arm-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700775 "src/f32-dwconv/gen/up4x9-minmax-wasmsimd-arm.c",
Frank Barchard0725b8d2020-12-07 11:07:35 -0800776 "src/f32-dwconv/gen/up4x9-minmax-wasmsimd-x86-acc2.c",
Marat Dukhanac014d72020-06-16 08:36:47 -0700777 "src/f32-dwconv/gen/up4x9-minmax-wasmsimd-x86.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700778 "src/f32-dwconv/gen/up4x9-wasmsimd.c",
Frank Barchard0725b8d2020-12-07 11:07:35 -0800779 "src/f32-dwconv/gen/up4x25-minmax-wasmsimd-arm-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700780 "src/f32-dwconv/gen/up4x25-minmax-wasmsimd-arm.c",
Frank Barchard0725b8d2020-12-07 11:07:35 -0800781 "src/f32-dwconv/gen/up4x25-minmax-wasmsimd-x86-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700782 "src/f32-dwconv/gen/up4x25-minmax-wasmsimd-x86.c",
783 "src/f32-dwconv/gen/up4x25-wasmsimd.c",
Frank Barchard0725b8d2020-12-07 11:07:35 -0800784 "src/f32-dwconv/gen/up8x4-minmax-wasmsimd-arm-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700785 "src/f32-dwconv/gen/up8x4-minmax-wasmsimd-arm.c",
Frank Barchard0725b8d2020-12-07 11:07:35 -0800786 "src/f32-dwconv/gen/up8x4-minmax-wasmsimd-x86-acc2.c",
Marat Dukhanac014d72020-06-16 08:36:47 -0700787 "src/f32-dwconv/gen/up8x4-minmax-wasmsimd-x86.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700788 "src/f32-dwconv/gen/up8x4-wasmsimd.c",
Frank Barchard0725b8d2020-12-07 11:07:35 -0800789 "src/f32-dwconv/gen/up8x9-minmax-wasmsimd-arm-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700790 "src/f32-dwconv/gen/up8x9-minmax-wasmsimd-arm.c",
Frank Barchard0725b8d2020-12-07 11:07:35 -0800791 "src/f32-dwconv/gen/up8x9-minmax-wasmsimd-x86-acc2.c",
Marat Dukhanac014d72020-06-16 08:36:47 -0700792 "src/f32-dwconv/gen/up8x9-minmax-wasmsimd-x86.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700793 "src/f32-dwconv/gen/up8x9-wasmsimd.c",
Frank Barchard0725b8d2020-12-07 11:07:35 -0800794 "src/f32-dwconv/gen/up8x25-minmax-wasmsimd-arm-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700795 "src/f32-dwconv/gen/up8x25-minmax-wasmsimd-arm.c",
Frank Barchard0725b8d2020-12-07 11:07:35 -0800796 "src/f32-dwconv/gen/up8x25-minmax-wasmsimd-x86-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700797 "src/f32-dwconv/gen/up8x25-minmax-wasmsimd-x86.c",
798 "src/f32-dwconv/gen/up8x25-wasmsimd.c",
Frank Barchard412e2f42020-12-11 11:40:50 -0800799 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-loadsplat-1x4-acc2.c",
800 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-loadsplat-1x4-acc3.c",
801 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-loadsplat-1x4-acc4.c",
802 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-loadsplat-1x4.c",
803 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-loadsplat-2x4-acc2.c",
804 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-loadsplat-2x4.c",
805 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-loadsplat-3x4.c",
806 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-loadsplat-4x4.c",
807 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-loadsplat-5x4.c",
808 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-loadsplat-6x4.c",
Frank Barchard02bb4292020-12-15 18:25:32 -0800809 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-splat-1x4-acc2.c",
810 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-splat-1x4-acc3.c",
811 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-splat-1x4-acc4.c",
812 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-splat-1x4.c",
813 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-splat-2x4-acc2.c",
814 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-splat-2x4.c",
815 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-splat-3x4.c",
816 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-splat-4x4.c",
817 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-splat-5x4.c",
818 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-splat-6x4.c",
Frank Barchard412e2f42020-12-11 11:40:50 -0800819 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-loadsplat-1x4-acc2.c",
820 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-loadsplat-1x4-acc3.c",
821 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-loadsplat-1x4-acc4.c",
822 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-loadsplat-1x4.c",
823 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-loadsplat-2x4-acc2.c",
824 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-loadsplat-2x4.c",
825 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-loadsplat-3x4.c",
826 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-loadsplat-4x4.c",
827 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-loadsplat-5x4.c",
828 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-loadsplat-6x4.c",
Frank Barchard02bb4292020-12-15 18:25:32 -0800829 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-splat-1x4-acc2.c",
830 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-splat-1x4-acc3.c",
831 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-splat-1x4-acc4.c",
832 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-splat-1x4.c",
833 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-splat-2x4-acc2.c",
834 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-splat-2x4.c",
835 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-splat-3x4.c",
836 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-splat-4x4.c",
837 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-splat-5x4.c",
838 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-splat-6x4.c",
Frank Barchardc5704bf2020-12-21 23:09:00 -0800839 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-arm-loadsplat-1x4-acc2.c",
840 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-arm-loadsplat-1x4-acc3.c",
841 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-arm-loadsplat-1x4-acc4.c",
842 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-arm-loadsplat-1x4.c",
843 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-arm-loadsplat-2x4-acc2.c",
844 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-arm-loadsplat-2x4.c",
845 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-arm-loadsplat-3x4.c",
846 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-arm-loadsplat-4x4.c",
Frank Barchard412e2f42020-12-11 11:40:50 -0800847 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-arm-splat-1x4-acc2.c",
848 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-arm-splat-1x4-acc3.c",
849 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-arm-splat-1x4-acc4.c",
850 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-arm-splat-1x4.c",
851 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-arm-splat-2x4-acc2.c",
852 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-arm-splat-2x4.c",
853 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-arm-splat-3x4.c",
854 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-arm-splat-4x4.c",
Frank Barchardcadd4222021-01-20 16:27:25 -0800855 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-x86-loadsplat-1x4-acc2.c",
856 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-x86-loadsplat-1x4-acc3.c",
857 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-x86-loadsplat-1x4-acc4.c",
858 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-x86-loadsplat-1x4.c",
859 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-x86-loadsplat-2x4-acc2.c",
860 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-x86-loadsplat-2x4.c",
861 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-x86-loadsplat-3x4.c",
862 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-x86-loadsplat-4x4.c",
Frank Barchard412e2f42020-12-11 11:40:50 -0800863 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-x86-splat-1x4-acc2.c",
864 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-x86-splat-1x4-acc3.c",
865 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-x86-splat-1x4-acc4.c",
866 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-x86-splat-1x4.c",
867 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-x86-splat-2x4-acc2.c",
868 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-x86-splat-2x4.c",
869 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-x86-splat-3x4.c",
870 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-x86-splat-4x4.c",
Frank Barchardb20dcd62020-12-15 16:46:14 -0800871 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-loadsplat-1x4-acc2.c",
872 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-loadsplat-1x4-acc3.c",
873 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-loadsplat-1x4-acc4.c",
874 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-loadsplat-1x4-acc5.c",
875 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-loadsplat-1x4.c",
876 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-loadsplat-2x4-acc2.c",
877 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-loadsplat-2x4-acc3.c",
878 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-loadsplat-2x4.c",
879 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-loadsplat-3x4-acc2.c",
880 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-loadsplat-3x4.c",
881 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-loadsplat-4x4-acc2.c",
882 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-loadsplat-4x4.c",
883 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-loadsplat-5x4.c",
Frank Barchard412e2f42020-12-11 11:40:50 -0800884 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-splat-1x4-acc2.c",
885 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-splat-1x4-acc3.c",
886 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-splat-1x4-acc4.c",
887 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-splat-1x4-acc5.c",
888 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-splat-1x4.c",
889 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-splat-2x4-acc2.c",
890 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-splat-2x4-acc3.c",
891 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-splat-2x4.c",
892 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-splat-3x4-acc2.c",
893 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-splat-3x4.c",
894 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-splat-4x4-acc2.c",
895 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-splat-4x4.c",
896 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-splat-5x4.c",
Frank Barchardb20dcd62020-12-15 16:46:14 -0800897 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-loadsplat-1x4-acc2.c",
898 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-loadsplat-1x4-acc3.c",
899 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-loadsplat-1x4-acc4.c",
900 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-loadsplat-1x4-acc5.c",
901 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-loadsplat-1x4.c",
902 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-loadsplat-2x4-acc2.c",
903 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-loadsplat-2x4-acc3.c",
904 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-loadsplat-2x4.c",
905 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-loadsplat-3x4-acc2.c",
906 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-loadsplat-3x4.c",
907 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-loadsplat-4x4-acc2.c",
908 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-loadsplat-4x4.c",
909 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-loadsplat-5x4.c",
Frank Barchard412e2f42020-12-11 11:40:50 -0800910 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-splat-1x4-acc2.c",
911 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-splat-1x4-acc3.c",
912 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-splat-1x4-acc4.c",
913 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-splat-1x4-acc5.c",
914 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-splat-1x4.c",
915 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-splat-2x4-acc2.c",
916 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-splat-2x4-acc3.c",
917 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-splat-2x4.c",
918 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-splat-3x4-acc2.c",
919 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-splat-3x4.c",
920 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-splat-4x4-acc2.c",
921 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-splat-4x4.c",
922 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-splat-5x4.c",
Frank Barchardc6889b32020-12-21 11:27:22 -0800923 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-arm-loadsplat-1x4-acc2.c",
924 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-arm-loadsplat-1x4-acc3.c",
925 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-arm-loadsplat-1x4-acc4.c",
926 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-arm-loadsplat-1x4-acc5.c",
927 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-arm-loadsplat-1x4.c",
928 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-arm-loadsplat-2x4-acc2.c",
929 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-arm-loadsplat-2x4-acc3.c",
930 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-arm-loadsplat-2x4.c",
931 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-arm-loadsplat-3x4-acc2.c",
932 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-arm-loadsplat-3x4.c",
Frank Barchard412e2f42020-12-11 11:40:50 -0800933 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-arm-splat-1x4-acc2.c",
934 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-arm-splat-1x4-acc3.c",
935 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-arm-splat-1x4-acc4.c",
936 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-arm-splat-1x4-acc5.c",
937 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-arm-splat-1x4.c",
938 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-arm-splat-2x4-acc2.c",
939 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-arm-splat-2x4-acc3.c",
940 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-arm-splat-2x4.c",
941 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-arm-splat-3x4-acc2.c",
942 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-arm-splat-3x4.c",
Frank Barchardc6889b32020-12-21 11:27:22 -0800943 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-x86-loadsplat-1x4-acc2.c",
944 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-x86-loadsplat-1x4-acc3.c",
945 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-x86-loadsplat-1x4-acc4.c",
946 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-x86-loadsplat-1x4-acc5.c",
947 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-x86-loadsplat-1x4.c",
948 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-x86-loadsplat-2x4-acc2.c",
949 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-x86-loadsplat-2x4-acc3.c",
950 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-x86-loadsplat-2x4.c",
951 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-x86-loadsplat-3x4-acc2.c",
952 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-x86-loadsplat-3x4.c",
Frank Barchard412e2f42020-12-11 11:40:50 -0800953 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-x86-splat-1x4-acc2.c",
954 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-x86-splat-1x4-acc3.c",
955 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-x86-splat-1x4-acc4.c",
956 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-x86-splat-1x4-acc5.c",
957 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-x86-splat-1x4.c",
958 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-x86-splat-2x4-acc2.c",
959 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-x86-splat-2x4-acc3.c",
960 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-x86-splat-2x4.c",
961 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-x86-splat-3x4-acc2.c",
962 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-x86-splat-3x4.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700963 "src/f32-gavgpool-cw/wasmsimd-arm-x4.c",
964 "src/f32-gavgpool-cw/wasmsimd-x86-x4.c",
Marat Dukhanc6016802020-07-16 18:51:28 -0700965 "src/f32-gavgpool/7p7x-minmax-wasmsimd-arm-c4.c",
966 "src/f32-gavgpool/7p7x-minmax-wasmsimd-x86-c4.c",
967 "src/f32-gavgpool/7x-minmax-wasmsimd-arm-c4.c",
968 "src/f32-gavgpool/7x-minmax-wasmsimd-x86-c4.c",
Frank Barchard0725b8d2020-12-07 11:07:35 -0800969 "src/f32-gemm/gen-inc/1x8inc-minmax-wasmsimd-arm-loadsplat.c",
970 "src/f32-gemm/gen-inc/1x8inc-minmax-wasmsimd-arm-splat.c",
971 "src/f32-gemm/gen-inc/1x8inc-minmax-wasmsimd-x86-loadsplat.c",
972 "src/f32-gemm/gen-inc/1x8inc-minmax-wasmsimd-x86-splat.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700973 "src/f32-gemm/gen-inc/1x8s4inc-minmax-wasmsimd-arm.c",
974 "src/f32-gemm/gen-inc/1x8s4inc-minmax-wasmsimd-x86.c",
Frank Barchard0725b8d2020-12-07 11:07:35 -0800975 "src/f32-gemm/gen-inc/3x8inc-minmax-wasmsimd-arm-loadsplat.c",
976 "src/f32-gemm/gen-inc/3x8inc-minmax-wasmsimd-arm-splat.c",
977 "src/f32-gemm/gen-inc/3x8inc-minmax-wasmsimd-x86-loadsplat.c",
978 "src/f32-gemm/gen-inc/3x8inc-minmax-wasmsimd-x86-splat.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700979 "src/f32-gemm/gen-inc/3x8s4inc-minmax-wasmsimd-arm.c",
980 "src/f32-gemm/gen-inc/3x8s4inc-minmax-wasmsimd-x86.c",
Frank Barchard0725b8d2020-12-07 11:07:35 -0800981 "src/f32-gemm/gen-inc/4x8inc-minmax-wasmsimd-arm-loadsplat.c",
982 "src/f32-gemm/gen-inc/4x8inc-minmax-wasmsimd-arm-splat.c",
983 "src/f32-gemm/gen-inc/4x8inc-minmax-wasmsimd-x86-loadsplat.c",
984 "src/f32-gemm/gen-inc/4x8inc-minmax-wasmsimd-x86-splat.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700985 "src/f32-gemm/gen-inc/4x8s4inc-minmax-wasmsimd-arm.c",
986 "src/f32-gemm/gen-inc/4x8s4inc-minmax-wasmsimd-x86.c",
Frank Barchard0725b8d2020-12-07 11:07:35 -0800987 "src/f32-gemm/gen-inc/5x8inc-minmax-wasmsimd-arm-loadsplat.c",
988 "src/f32-gemm/gen-inc/5x8inc-minmax-wasmsimd-arm-splat.c",
989 "src/f32-gemm/gen-inc/5x8inc-minmax-wasmsimd-x86-loadsplat.c",
990 "src/f32-gemm/gen-inc/5x8inc-minmax-wasmsimd-x86-splat.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700991 "src/f32-gemm/gen-inc/5x8s4inc-minmax-wasmsimd-arm.c",
992 "src/f32-gemm/gen-inc/5x8s4inc-minmax-wasmsimd-x86.c",
Frank Barchard0725b8d2020-12-07 11:07:35 -0800993 "src/f32-gemm/gen-inc/6x8inc-minmax-wasmsimd-arm-loadsplat.c",
994 "src/f32-gemm/gen-inc/6x8inc-minmax-wasmsimd-arm-splat.c",
995 "src/f32-gemm/gen-inc/6x8inc-minmax-wasmsimd-x86-loadsplat.c",
996 "src/f32-gemm/gen-inc/6x8inc-minmax-wasmsimd-x86-splat.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700997 "src/f32-gemm/gen-inc/6x8s4inc-minmax-wasmsimd-arm.c",
998 "src/f32-gemm/gen-inc/6x8s4inc-minmax-wasmsimd-x86.c",
Frank Barchard0725b8d2020-12-07 11:07:35 -0800999 "src/f32-gemm/gen/1x8-minmax-wasmsimd-arm-loadsplat.c",
1000 "src/f32-gemm/gen/1x8-minmax-wasmsimd-arm-splat.c",
1001 "src/f32-gemm/gen/1x8-minmax-wasmsimd-x86-loadsplat.c",
1002 "src/f32-gemm/gen/1x8-minmax-wasmsimd-x86-splat.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001003 "src/f32-gemm/gen/1x8-relu-wasmsimd-splat.c",
1004 "src/f32-gemm/gen/1x8-wasmsimd-splat.c",
1005 "src/f32-gemm/gen/1x8s4-minmax-wasmsimd-arm.c",
1006 "src/f32-gemm/gen/1x8s4-minmax-wasmsimd-x86.c",
Frank Barchard0725b8d2020-12-07 11:07:35 -08001007 "src/f32-gemm/gen/3x8-minmax-wasmsimd-arm-loadsplat.c",
1008 "src/f32-gemm/gen/3x8-minmax-wasmsimd-arm-splat.c",
1009 "src/f32-gemm/gen/3x8-minmax-wasmsimd-x86-loadsplat.c",
1010 "src/f32-gemm/gen/3x8-minmax-wasmsimd-x86-splat.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001011 "src/f32-gemm/gen/3x8s4-minmax-wasmsimd-arm.c",
1012 "src/f32-gemm/gen/3x8s4-minmax-wasmsimd-x86.c",
1013 "src/f32-gemm/gen/4x2c4-minmax-wasmsimd-arm.c",
1014 "src/f32-gemm/gen/4x2c4-minmax-wasmsimd-x86.c",
1015 "src/f32-gemm/gen/4x2c4-relu-wasmsimd.c",
1016 "src/f32-gemm/gen/4x2c4-wasmsimd.c",
Frank Barchard0725b8d2020-12-07 11:07:35 -08001017 "src/f32-gemm/gen/4x8-minmax-wasmsimd-arm-loadsplat.c",
1018 "src/f32-gemm/gen/4x8-minmax-wasmsimd-arm-splat.c",
1019 "src/f32-gemm/gen/4x8-minmax-wasmsimd-x86-loadsplat.c",
1020 "src/f32-gemm/gen/4x8-minmax-wasmsimd-x86-splat.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001021 "src/f32-gemm/gen/4x8-relu-wasmsimd-splat.c",
1022 "src/f32-gemm/gen/4x8-wasmsimd-splat.c",
1023 "src/f32-gemm/gen/4x8s4-minmax-wasmsimd-arm.c",
1024 "src/f32-gemm/gen/4x8s4-minmax-wasmsimd-x86.c",
Frank Barchard0725b8d2020-12-07 11:07:35 -08001025 "src/f32-gemm/gen/5x8-minmax-wasmsimd-arm-loadsplat.c",
1026 "src/f32-gemm/gen/5x8-minmax-wasmsimd-arm-splat.c",
1027 "src/f32-gemm/gen/5x8-minmax-wasmsimd-x86-loadsplat.c",
1028 "src/f32-gemm/gen/5x8-minmax-wasmsimd-x86-splat.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001029 "src/f32-gemm/gen/5x8-relu-wasmsimd-splat.c",
1030 "src/f32-gemm/gen/5x8-wasmsimd-splat.c",
1031 "src/f32-gemm/gen/5x8s4-minmax-wasmsimd-arm.c",
1032 "src/f32-gemm/gen/5x8s4-minmax-wasmsimd-x86.c",
Frank Barchard0725b8d2020-12-07 11:07:35 -08001033 "src/f32-gemm/gen/6x8-minmax-wasmsimd-arm-loadsplat.c",
1034 "src/f32-gemm/gen/6x8-minmax-wasmsimd-arm-splat.c",
1035 "src/f32-gemm/gen/6x8-minmax-wasmsimd-x86-loadsplat.c",
1036 "src/f32-gemm/gen/6x8-minmax-wasmsimd-x86-splat.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001037 "src/f32-gemm/gen/6x8s4-minmax-wasmsimd-arm.c",
1038 "src/f32-gemm/gen/6x8s4-minmax-wasmsimd-x86.c",
XNNPACK Team965272b2020-10-23 21:10:15 -07001039 "src/f32-ibilinear-chw/gen/wasmsimd-p4.c",
1040 "src/f32-ibilinear-chw/gen/wasmsimd-p8.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001041 "src/f32-ibilinear/gen/wasmsimd-c4.c",
1042 "src/f32-ibilinear/gen/wasmsimd-c8.c",
Frank Barchard0725b8d2020-12-07 11:07:35 -08001043 "src/f32-igemm/gen/1x8-minmax-wasmsimd-arm-loadsplat.c",
1044 "src/f32-igemm/gen/1x8-minmax-wasmsimd-arm-splat.c",
1045 "src/f32-igemm/gen/1x8-minmax-wasmsimd-x86-loadsplat.c",
1046 "src/f32-igemm/gen/1x8-minmax-wasmsimd-x86-splat.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001047 "src/f32-igemm/gen/1x8-relu-wasmsimd-splat.c",
1048 "src/f32-igemm/gen/1x8-wasmsimd-splat.c",
1049 "src/f32-igemm/gen/1x8s4-minmax-wasmsimd-arm.c",
1050 "src/f32-igemm/gen/1x8s4-minmax-wasmsimd-x86.c",
Frank Barchard0725b8d2020-12-07 11:07:35 -08001051 "src/f32-igemm/gen/3x8-minmax-wasmsimd-arm-loadsplat.c",
1052 "src/f32-igemm/gen/3x8-minmax-wasmsimd-arm-splat.c",
1053 "src/f32-igemm/gen/3x8-minmax-wasmsimd-x86-loadsplat.c",
1054 "src/f32-igemm/gen/3x8-minmax-wasmsimd-x86-splat.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001055 "src/f32-igemm/gen/3x8s4-minmax-wasmsimd-arm.c",
1056 "src/f32-igemm/gen/3x8s4-minmax-wasmsimd-x86.c",
1057 "src/f32-igemm/gen/4x2c4-minmax-wasmsimd-arm.c",
1058 "src/f32-igemm/gen/4x2c4-minmax-wasmsimd-x86.c",
1059 "src/f32-igemm/gen/4x2c4-relu-wasmsimd.c",
1060 "src/f32-igemm/gen/4x2c4-wasmsimd.c",
Frank Barchard0725b8d2020-12-07 11:07:35 -08001061 "src/f32-igemm/gen/4x8-minmax-wasmsimd-arm-loadsplat.c",
1062 "src/f32-igemm/gen/4x8-minmax-wasmsimd-arm-splat.c",
1063 "src/f32-igemm/gen/4x8-minmax-wasmsimd-x86-loadsplat.c",
1064 "src/f32-igemm/gen/4x8-minmax-wasmsimd-x86-splat.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001065 "src/f32-igemm/gen/4x8-relu-wasmsimd-splat.c",
1066 "src/f32-igemm/gen/4x8-wasmsimd-splat.c",
1067 "src/f32-igemm/gen/4x8s4-minmax-wasmsimd-arm.c",
1068 "src/f32-igemm/gen/4x8s4-minmax-wasmsimd-x86.c",
Frank Barchard0725b8d2020-12-07 11:07:35 -08001069 "src/f32-igemm/gen/5x8-minmax-wasmsimd-arm-loadsplat.c",
1070 "src/f32-igemm/gen/5x8-minmax-wasmsimd-arm-splat.c",
1071 "src/f32-igemm/gen/5x8-minmax-wasmsimd-x86-loadsplat.c",
1072 "src/f32-igemm/gen/5x8-minmax-wasmsimd-x86-splat.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001073 "src/f32-igemm/gen/5x8-relu-wasmsimd-splat.c",
1074 "src/f32-igemm/gen/5x8-wasmsimd-splat.c",
1075 "src/f32-igemm/gen/5x8s4-minmax-wasmsimd-arm.c",
1076 "src/f32-igemm/gen/5x8s4-minmax-wasmsimd-x86.c",
Frank Barchard0725b8d2020-12-07 11:07:35 -08001077 "src/f32-igemm/gen/6x8-minmax-wasmsimd-arm-loadsplat.c",
1078 "src/f32-igemm/gen/6x8-minmax-wasmsimd-arm-splat.c",
1079 "src/f32-igemm/gen/6x8-minmax-wasmsimd-x86-loadsplat.c",
1080 "src/f32-igemm/gen/6x8-minmax-wasmsimd-x86-splat.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001081 "src/f32-igemm/gen/6x8s4-minmax-wasmsimd-arm.c",
1082 "src/f32-igemm/gen/6x8s4-minmax-wasmsimd-x86.c",
Marat Dukhanf6e24802020-07-08 22:20:40 -07001083 "src/f32-maxpool/9p8x-minmax-wasmsimd-arm-c4.c",
1084 "src/f32-maxpool/9p8x-minmax-wasmsimd-x86-c4.c",
Marat Dukhan1483c532020-07-16 18:08:19 -07001085 "src/f32-pavgpool/9p8x-minmax-wasmsimd-arm-c4.c",
1086 "src/f32-pavgpool/9p8x-minmax-wasmsimd-x86-c4.c",
1087 "src/f32-pavgpool/9x-minmax-wasmsimd-arm-c4.c",
1088 "src/f32-pavgpool/9x-minmax-wasmsimd-x86-c4.c",
Frank Barchard0725b8d2020-12-07 11:07:35 -08001089 "src/f32-ppmm/gen/4x8-minmax-wasmsimd-arm-splat.c",
1090 "src/f32-ppmm/gen/4x8-minmax-wasmsimd-x86-splat.c",
Frank Barcharda5316982020-07-23 13:19:28 -07001091 "src/f32-prelu/gen/wasmsimd-bitselect-1x4.c",
1092 "src/f32-prelu/gen/wasmsimd-bitselect-1x8.c",
1093 "src/f32-prelu/gen/wasmsimd-bitselect-1x16.c",
Marat Dukhan195f8eb2020-06-25 12:50:57 -07001094 "src/f32-prelu/gen/wasmsimd-bitselect-2x4.c",
1095 "src/f32-prelu/gen/wasmsimd-bitselect-2x8.c",
Frank Barcharda5316982020-07-23 13:19:28 -07001096 "src/f32-prelu/gen/wasmsimd-bitselect-2x16.c",
1097 "src/f32-prelu/gen/wasmsimd-bitselect-4x4.c",
1098 "src/f32-prelu/gen/wasmsimd-bitselect-4x8.c",
1099 "src/f32-prelu/gen/wasmsimd-bitselect-4x16.c",
1100 "src/f32-prelu/gen/wasmsimd-minmax-1x4.c",
1101 "src/f32-prelu/gen/wasmsimd-minmax-1x8.c",
1102 "src/f32-prelu/gen/wasmsimd-minmax-1x16.c",
Marat Dukhan195f8eb2020-06-25 12:50:57 -07001103 "src/f32-prelu/gen/wasmsimd-minmax-2x4.c",
1104 "src/f32-prelu/gen/wasmsimd-minmax-2x8.c",
Frank Barcharda5316982020-07-23 13:19:28 -07001105 "src/f32-prelu/gen/wasmsimd-minmax-2x16.c",
1106 "src/f32-prelu/gen/wasmsimd-minmax-4x4.c",
1107 "src/f32-prelu/gen/wasmsimd-minmax-4x8.c",
1108 "src/f32-prelu/gen/wasmsimd-minmax-4x16.c",
Marat Dukhan52238f02020-07-16 15:30:28 -07001109 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x4.c",
Marat Dukhan52238f02020-07-16 15:30:28 -07001110 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x8-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001111 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x8.c",
Marat Dukhan52238f02020-07-16 15:30:28 -07001112 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x12-acc2.c",
1113 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x12-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001114 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x12.c",
Marat Dukhan52238f02020-07-16 15:30:28 -07001115 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x16-acc2.c",
1116 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x16-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001117 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x16.c",
Marat Dukhan52238f02020-07-16 15:30:28 -07001118 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x20-acc2.c",
1119 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x20-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001120 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x20.c",
Marat Dukhan8c417962020-07-08 12:27:50 -07001121 "src/f32-rmax/wasmsimd-arm.c",
1122 "src/f32-rmax/wasmsimd-x86.c",
Frank Barchard8ef44cd2020-11-03 12:30:23 -08001123 "src/f32-spmm/gen/4x1-minmax-wasmsimd-arm-pipelined-x2.c",
1124 "src/f32-spmm/gen/4x1-minmax-wasmsimd-arm-pipelined.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07001125 "src/f32-spmm/gen/4x1-minmax-wasmsimd-arm-x2.c",
1126 "src/f32-spmm/gen/4x1-minmax-wasmsimd-arm-x4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001127 "src/f32-spmm/gen/4x1-minmax-wasmsimd-arm.c",
Frank Barchard8ef44cd2020-11-03 12:30:23 -08001128 "src/f32-spmm/gen/4x1-minmax-wasmsimd-x86-pipelined-x2.c",
1129 "src/f32-spmm/gen/4x1-minmax-wasmsimd-x86-pipelined.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07001130 "src/f32-spmm/gen/4x1-minmax-wasmsimd-x86-x2.c",
1131 "src/f32-spmm/gen/4x1-minmax-wasmsimd-x86-x4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001132 "src/f32-spmm/gen/4x1-minmax-wasmsimd-x86.c",
Frank Barchard8ef44cd2020-11-03 12:30:23 -08001133 "src/f32-spmm/gen/8x1-minmax-wasmsimd-arm-pipelined-x2.c",
1134 "src/f32-spmm/gen/8x1-minmax-wasmsimd-arm-pipelined.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07001135 "src/f32-spmm/gen/8x1-minmax-wasmsimd-arm-x2.c",
1136 "src/f32-spmm/gen/8x1-minmax-wasmsimd-arm-x4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001137 "src/f32-spmm/gen/8x1-minmax-wasmsimd-arm.c",
Frank Barchard8ef44cd2020-11-03 12:30:23 -08001138 "src/f32-spmm/gen/8x1-minmax-wasmsimd-x86-pipelined-x2.c",
1139 "src/f32-spmm/gen/8x1-minmax-wasmsimd-x86-pipelined.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07001140 "src/f32-spmm/gen/8x1-minmax-wasmsimd-x86-x2.c",
1141 "src/f32-spmm/gen/8x1-minmax-wasmsimd-x86-x4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001142 "src/f32-spmm/gen/8x1-minmax-wasmsimd-x86.c",
Frank Barchard8ef44cd2020-11-03 12:30:23 -08001143 "src/f32-spmm/gen/16x1-minmax-wasmsimd-arm-pipelined-x2.c",
1144 "src/f32-spmm/gen/16x1-minmax-wasmsimd-arm-pipelined.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07001145 "src/f32-spmm/gen/16x1-minmax-wasmsimd-arm-x2.c",
1146 "src/f32-spmm/gen/16x1-minmax-wasmsimd-arm-x4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001147 "src/f32-spmm/gen/16x1-minmax-wasmsimd-arm.c",
Frank Barchard8ef44cd2020-11-03 12:30:23 -08001148 "src/f32-spmm/gen/16x1-minmax-wasmsimd-x86-pipelined-x2.c",
1149 "src/f32-spmm/gen/16x1-minmax-wasmsimd-x86-pipelined.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07001150 "src/f32-spmm/gen/16x1-minmax-wasmsimd-x86-x2.c",
1151 "src/f32-spmm/gen/16x1-minmax-wasmsimd-x86-x4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001152 "src/f32-spmm/gen/16x1-minmax-wasmsimd-x86.c",
Frank Barchard8ef44cd2020-11-03 12:30:23 -08001153 "src/f32-spmm/gen/32x1-minmax-wasmsimd-arm-pipelined-x2.c",
1154 "src/f32-spmm/gen/32x1-minmax-wasmsimd-arm-pipelined.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07001155 "src/f32-spmm/gen/32x1-minmax-wasmsimd-arm-x2.c",
1156 "src/f32-spmm/gen/32x1-minmax-wasmsimd-arm-x4.c",
Frank Barchard846c0c62020-10-26 15:01:39 -07001157 "src/f32-spmm/gen/32x1-minmax-wasmsimd-arm.c",
Frank Barchard8ef44cd2020-11-03 12:30:23 -08001158 "src/f32-spmm/gen/32x1-minmax-wasmsimd-x86-pipelined-x2.c",
1159 "src/f32-spmm/gen/32x1-minmax-wasmsimd-x86-pipelined.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07001160 "src/f32-spmm/gen/32x1-minmax-wasmsimd-x86-x2.c",
1161 "src/f32-spmm/gen/32x1-minmax-wasmsimd-x86-x4.c",
Frank Barchard846c0c62020-10-26 15:01:39 -07001162 "src/f32-spmm/gen/32x1-minmax-wasmsimd-x86.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -07001163 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-arm-x4.c",
1164 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001165 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-arm-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -07001166 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-x86-x4.c",
1167 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001168 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -07001169 "src/f32-vbinary/gen/vadd-relu-wasmsimd-x4.c",
1170 "src/f32-vbinary/gen/vadd-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001171 "src/f32-vbinary/gen/vadd-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07001172 "src/f32-vbinary/gen/vadd-wasmsimd-x4.c",
1173 "src/f32-vbinary/gen/vadd-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001174 "src/f32-vbinary/gen/vadd-wasmsimd-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001175 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-arm-x4.c",
1176 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001177 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-arm-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001178 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-x86-x4.c",
1179 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001180 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -07001181 "src/f32-vbinary/gen/vaddc-relu-wasmsimd-x4.c",
1182 "src/f32-vbinary/gen/vaddc-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001183 "src/f32-vbinary/gen/vaddc-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07001184 "src/f32-vbinary/gen/vaddc-wasmsimd-x4.c",
1185 "src/f32-vbinary/gen/vaddc-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001186 "src/f32-vbinary/gen/vaddc-wasmsimd-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -07001187 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-arm-x4.c",
1188 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001189 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-arm-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -07001190 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-x86-x4.c",
1191 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001192 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -07001193 "src/f32-vbinary/gen/vdiv-relu-wasmsimd-x4.c",
1194 "src/f32-vbinary/gen/vdiv-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001195 "src/f32-vbinary/gen/vdiv-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07001196 "src/f32-vbinary/gen/vdiv-wasmsimd-x4.c",
1197 "src/f32-vbinary/gen/vdiv-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001198 "src/f32-vbinary/gen/vdiv-wasmsimd-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001199 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-arm-x4.c",
1200 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001201 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-arm-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001202 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-x86-x4.c",
1203 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001204 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -07001205 "src/f32-vbinary/gen/vdivc-relu-wasmsimd-x4.c",
1206 "src/f32-vbinary/gen/vdivc-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001207 "src/f32-vbinary/gen/vdivc-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07001208 "src/f32-vbinary/gen/vdivc-wasmsimd-x4.c",
1209 "src/f32-vbinary/gen/vdivc-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001210 "src/f32-vbinary/gen/vdivc-wasmsimd-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -07001211 "src/f32-vbinary/gen/vmax-wasmsimd-arm-x4.c",
1212 "src/f32-vbinary/gen/vmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001213 "src/f32-vbinary/gen/vmax-wasmsimd-arm-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -07001214 "src/f32-vbinary/gen/vmax-wasmsimd-x86-x4.c",
1215 "src/f32-vbinary/gen/vmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001216 "src/f32-vbinary/gen/vmax-wasmsimd-x86-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001217 "src/f32-vbinary/gen/vmaxc-wasmsimd-arm-x4.c",
1218 "src/f32-vbinary/gen/vmaxc-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001219 "src/f32-vbinary/gen/vmaxc-wasmsimd-arm-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001220 "src/f32-vbinary/gen/vmaxc-wasmsimd-x86-x4.c",
1221 "src/f32-vbinary/gen/vmaxc-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001222 "src/f32-vbinary/gen/vmaxc-wasmsimd-x86-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -07001223 "src/f32-vbinary/gen/vmin-wasmsimd-arm-x4.c",
1224 "src/f32-vbinary/gen/vmin-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001225 "src/f32-vbinary/gen/vmin-wasmsimd-arm-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -07001226 "src/f32-vbinary/gen/vmin-wasmsimd-x86-x4.c",
1227 "src/f32-vbinary/gen/vmin-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001228 "src/f32-vbinary/gen/vmin-wasmsimd-x86-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001229 "src/f32-vbinary/gen/vminc-wasmsimd-arm-x4.c",
1230 "src/f32-vbinary/gen/vminc-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001231 "src/f32-vbinary/gen/vminc-wasmsimd-arm-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001232 "src/f32-vbinary/gen/vminc-wasmsimd-x86-x4.c",
1233 "src/f32-vbinary/gen/vminc-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001234 "src/f32-vbinary/gen/vminc-wasmsimd-x86-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -07001235 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-arm-x4.c",
1236 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001237 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-arm-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -07001238 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-x86-x4.c",
1239 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001240 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -07001241 "src/f32-vbinary/gen/vmul-relu-wasmsimd-x4.c",
1242 "src/f32-vbinary/gen/vmul-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001243 "src/f32-vbinary/gen/vmul-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07001244 "src/f32-vbinary/gen/vmul-wasmsimd-x4.c",
1245 "src/f32-vbinary/gen/vmul-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001246 "src/f32-vbinary/gen/vmul-wasmsimd-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001247 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-arm-x4.c",
1248 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001249 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-arm-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001250 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-x86-x4.c",
1251 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001252 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -07001253 "src/f32-vbinary/gen/vmulc-relu-wasmsimd-x4.c",
1254 "src/f32-vbinary/gen/vmulc-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001255 "src/f32-vbinary/gen/vmulc-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07001256 "src/f32-vbinary/gen/vmulc-wasmsimd-x4.c",
1257 "src/f32-vbinary/gen/vmulc-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001258 "src/f32-vbinary/gen/vmulc-wasmsimd-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001259 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-arm-x4.c",
1260 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001261 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-arm-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001262 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-x86-x4.c",
1263 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001264 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -07001265 "src/f32-vbinary/gen/vrdivc-relu-wasmsimd-x4.c",
1266 "src/f32-vbinary/gen/vrdivc-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001267 "src/f32-vbinary/gen/vrdivc-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07001268 "src/f32-vbinary/gen/vrdivc-wasmsimd-x4.c",
1269 "src/f32-vbinary/gen/vrdivc-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001270 "src/f32-vbinary/gen/vrdivc-wasmsimd-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001271 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-arm-x4.c",
1272 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001273 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-arm-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001274 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-x86-x4.c",
1275 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001276 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -07001277 "src/f32-vbinary/gen/vrsubc-relu-wasmsimd-x4.c",
1278 "src/f32-vbinary/gen/vrsubc-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001279 "src/f32-vbinary/gen/vrsubc-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07001280 "src/f32-vbinary/gen/vrsubc-wasmsimd-x4.c",
1281 "src/f32-vbinary/gen/vrsubc-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001282 "src/f32-vbinary/gen/vrsubc-wasmsimd-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -07001283 "src/f32-vbinary/gen/vsqrdiff-wasmsimd-x4.c",
1284 "src/f32-vbinary/gen/vsqrdiff-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001285 "src/f32-vbinary/gen/vsqrdiff-wasmsimd-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001286 "src/f32-vbinary/gen/vsqrdiffc-wasmsimd-x4.c",
1287 "src/f32-vbinary/gen/vsqrdiffc-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001288 "src/f32-vbinary/gen/vsqrdiffc-wasmsimd-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -07001289 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-arm-x4.c",
1290 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001291 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-arm-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -07001292 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-x86-x4.c",
1293 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001294 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -07001295 "src/f32-vbinary/gen/vsub-relu-wasmsimd-x4.c",
1296 "src/f32-vbinary/gen/vsub-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001297 "src/f32-vbinary/gen/vsub-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07001298 "src/f32-vbinary/gen/vsub-wasmsimd-x4.c",
1299 "src/f32-vbinary/gen/vsub-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001300 "src/f32-vbinary/gen/vsub-wasmsimd-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001301 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-arm-x4.c",
1302 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001303 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-arm-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001304 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-x86-x4.c",
1305 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001306 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -07001307 "src/f32-vbinary/gen/vsubc-relu-wasmsimd-x4.c",
1308 "src/f32-vbinary/gen/vsubc-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001309 "src/f32-vbinary/gen/vsubc-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07001310 "src/f32-vbinary/gen/vsubc-wasmsimd-x4.c",
1311 "src/f32-vbinary/gen/vsubc-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001312 "src/f32-vbinary/gen/vsubc-wasmsimd-x16.c",
Marat Dukhanb1eec082021-05-05 23:24:55 -07001313 "src/f32-vclamp/gen/vclamp-wasmsimd-arm-x4.c",
1314 "src/f32-vclamp/gen/vclamp-wasmsimd-arm-x8.c",
1315 "src/f32-vclamp/gen/vclamp-wasmsimd-x86-x4.c",
1316 "src/f32-vclamp/gen/vclamp-wasmsimd-x86-x8.c",
Marat Dukhaned6baaf2020-12-01 15:07:08 -08001317 "src/f32-velu/gen/velu-wasmsimd-arm-rr2-lut16-p3-x4.c",
1318 "src/f32-velu/gen/velu-wasmsimd-arm-rr2-lut16-p3-x8.c",
1319 "src/f32-velu/gen/velu-wasmsimd-arm-rr2-lut16-p3-x12.c",
1320 "src/f32-velu/gen/velu-wasmsimd-arm-rr2-lut16-p3-x16.c",
1321 "src/f32-velu/gen/velu-wasmsimd-arm-rr2-lut16-p3-x20.c",
1322 "src/f32-velu/gen/velu-wasmsimd-arm-rr2-lut16-p3-x24.c",
Marat Dukhaned6baaf2020-12-01 15:07:08 -08001323 "src/f32-velu/gen/velu-wasmsimd-arm-rr2-p6-x4.c",
1324 "src/f32-velu/gen/velu-wasmsimd-arm-rr2-p6-x8.c",
1325 "src/f32-velu/gen/velu-wasmsimd-arm-rr2-p6-x12.c",
1326 "src/f32-velu/gen/velu-wasmsimd-arm-rr2-p6-x16.c",
1327 "src/f32-velu/gen/velu-wasmsimd-arm-rr2-p6-x20.c",
1328 "src/f32-velu/gen/velu-wasmsimd-arm-rr2-p6-x24.c",
Frank Barcharde7223ee2020-12-04 19:04:01 -08001329 "src/f32-velu/gen/velu-wasmsimd-x86-rr2-lut16-p3-x4.c",
1330 "src/f32-velu/gen/velu-wasmsimd-x86-rr2-lut16-p3-x8.c",
1331 "src/f32-velu/gen/velu-wasmsimd-x86-rr2-lut16-p3-x12.c",
1332 "src/f32-velu/gen/velu-wasmsimd-x86-rr2-lut16-p3-x16.c",
1333 "src/f32-velu/gen/velu-wasmsimd-x86-rr2-lut16-p3-x20.c",
1334 "src/f32-velu/gen/velu-wasmsimd-x86-rr2-lut16-p3-x24.c",
Marat Dukhaned6baaf2020-12-01 15:07:08 -08001335 "src/f32-velu/gen/velu-wasmsimd-x86-rr2-p6-x4.c",
1336 "src/f32-velu/gen/velu-wasmsimd-x86-rr2-p6-x8.c",
1337 "src/f32-velu/gen/velu-wasmsimd-x86-rr2-p6-x12.c",
1338 "src/f32-velu/gen/velu-wasmsimd-x86-rr2-p6-x16.c",
1339 "src/f32-velu/gen/velu-wasmsimd-x86-rr2-p6-x20.c",
1340 "src/f32-velu/gen/velu-wasmsimd-x86-rr2-p6-x24.c",
Marat Dukhan6674d692021-05-05 22:27:00 -07001341 "src/f32-vhswish/gen/vhswish-wasmsimd-x4.c",
1342 "src/f32-vhswish/gen/vhswish-wasmsimd-x8.c",
1343 "src/f32-vhswish/gen/vhswish-wasmsimd-x16.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07001344 "src/f32-vlrelu/gen/vlrelu-wasmsimd-bitselect-x4.c",
1345 "src/f32-vlrelu/gen/vlrelu-wasmsimd-bitselect-x8.c",
1346 "src/f32-vlrelu/gen/vlrelu-wasmsimd-minmax-x4.c",
1347 "src/f32-vlrelu/gen/vlrelu-wasmsimd-minmax-x8.c",
Marat Dukhand816f622020-07-15 10:14:39 -07001348 "src/f32-vmulcaddc/gen/c4-minmax-wasmsimd-arm-2x.c",
Marat Dukhand816f622020-07-15 10:14:39 -07001349 "src/f32-vmulcaddc/gen/c4-minmax-wasmsimd-x86-2x.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001350 "src/f32-vmulcaddc/gen/c8-minmax-wasmsimd-arm-2x.c",
Marat Dukhand816f622020-07-15 10:14:39 -07001351 "src/f32-vmulcaddc/gen/c8-minmax-wasmsimd-x86-2x.c",
Marat Dukhanb1eec082021-05-05 23:24:55 -07001352 "src/f32-vrelu/gen/vrelu-wasmsimd-x4.c",
1353 "src/f32-vrelu/gen/vrelu-wasmsimd-x8.c",
1354 "src/f32-vrelu/gen/vrelu-wasmsimd-x16.c",
Marat Dukhanb82b2cd2020-07-16 02:23:42 -07001355 "src/f32-vrnd/gen/vrndd-wasmsimd-addsub-x4.c",
1356 "src/f32-vrnd/gen/vrndd-wasmsimd-addsub-x8.c",
1357 "src/f32-vrnd/gen/vrndd-wasmsimd-cvt-x4.c",
1358 "src/f32-vrnd/gen/vrndd-wasmsimd-cvt-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001359 "src/f32-vrnd/gen/vrndne-wasmsimd-addsub-x4.c",
1360 "src/f32-vrnd/gen/vrndne-wasmsimd-addsub-x8.c",
1361 "src/f32-vrnd/gen/vrndu-wasmsimd-addsub-x4.c",
1362 "src/f32-vrnd/gen/vrndu-wasmsimd-addsub-x8.c",
1363 "src/f32-vrnd/gen/vrndu-wasmsimd-cvt-x4.c",
1364 "src/f32-vrnd/gen/vrndu-wasmsimd-cvt-x8.c",
1365 "src/f32-vrnd/gen/vrndz-wasmsimd-addsub-x4.c",
1366 "src/f32-vrnd/gen/vrndz-wasmsimd-addsub-x8.c",
1367 "src/f32-vrnd/gen/vrndz-wasmsimd-cvt-x4.c",
1368 "src/f32-vrnd/gen/vrndz-wasmsimd-cvt-x8.c",
Marat Dukhanb1eec082021-05-05 23:24:55 -07001369 "src/f32-vsigmoid/gen/vsigmoid-wasmsimd-lut64-p2-div-x4.c",
1370 "src/f32-vsigmoid/gen/vsigmoid-wasmsimd-lut64-p2-div-x8.c",
1371 "src/f32-vsigmoid/gen/vsigmoid-wasmsimd-lut64-p2-div-x12.c",
1372 "src/f32-vsigmoid/gen/vsigmoid-wasmsimd-lut64-p2-div-x16.c",
1373 "src/f32-vsigmoid/gen/vsigmoid-wasmsimd-lut64-p2-div-x20.c",
1374 "src/f32-vsigmoid/gen/vsigmoid-wasmsimd-lut64-p2-div-x24.c",
1375 "src/f32-vsigmoid/gen/vsigmoid-wasmsimd-p5-div-x4.c",
1376 "src/f32-vsigmoid/gen/vsigmoid-wasmsimd-p5-div-x8.c",
1377 "src/f32-vsigmoid/gen/vsigmoid-wasmsimd-p5-div-x12.c",
1378 "src/f32-vsigmoid/gen/vsigmoid-wasmsimd-p5-div-x16.c",
1379 "src/f32-vsigmoid/gen/vsigmoid-wasmsimd-p5-div-x20.c",
1380 "src/f32-vsigmoid/gen/vsigmoid-wasmsimd-p5-div-x24.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07001381 "src/f32-vsqrt/gen/wasmsimd-sqrt-x4.c",
1382 "src/f32-vsqrt/gen/wasmsimd-sqrt-x8.c",
Marat Dukhan37c83512020-06-29 13:25:53 -07001383 "src/f32-vunary/gen/vabs-wasmsimd-x4.c",
1384 "src/f32-vunary/gen/vabs-wasmsimd-x8.c",
1385 "src/f32-vunary/gen/vneg-wasmsimd-x4.c",
1386 "src/f32-vunary/gen/vneg-wasmsimd-x8.c",
1387 "src/f32-vunary/gen/vsqr-wasmsimd-x4.c",
1388 "src/f32-vunary/gen/vsqr-wasmsimd-x8.c",
Marat Dukhande390d42020-11-29 19:32:18 -08001389 "src/math/expm1minus-wasmsimd-rr2-lut16-p3-andnot.c",
1390 "src/math/expm1minus-wasmsimd-rr2-lut16-p3-max.c",
1391 "src/math/expm1minus-wasmsimd-rr2-p6-andnot.c",
1392 "src/math/expm1minus-wasmsimd-rr2-p6-max.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001393 "src/math/roundd-wasmsimd-addsub.c",
1394 "src/math/roundd-wasmsimd-cvt.c",
1395 "src/math/roundne-wasmsimd-addsub.c",
1396 "src/math/roundu-wasmsimd-addsub.c",
1397 "src/math/roundu-wasmsimd-cvt.c",
1398 "src/math/roundz-wasmsimd-addsub.c",
1399 "src/math/roundz-wasmsimd-cvt.c",
1400 "src/math/sigmoid-wasmsimd-rr2-lut64-p2-div.c",
1401 "src/math/sigmoid-wasmsimd-rr2-p5-div.c",
Marat Dukhancc8f34c2020-08-05 16:36:38 -07001402 "src/qs8-dwconv/gen/up8x9-minmax-wasmsimd-mul16.c",
Marat Dukhan4ed14882021-05-12 17:50:40 -07001403 "src/qs8-dwconv/gen/up8x25-minmax-wasmsimd-mul16.c",
Marat Dukhancc8f34c2020-08-05 16:36:38 -07001404 "src/qs8-dwconv/gen/up16x9-minmax-wasmsimd-mul16.c",
Marat Dukhan4ed14882021-05-12 17:50:40 -07001405 "src/qs8-dwconv/gen/up16x25-minmax-wasmsimd-mul16.c",
Marat Dukhancc8f34c2020-08-05 16:36:38 -07001406 "src/qs8-dwconv/gen/up24x9-minmax-wasmsimd-mul16.c",
Marat Dukhan4ed14882021-05-12 17:50:40 -07001407 "src/qs8-dwconv/gen/up24x25-minmax-wasmsimd-mul16.c",
Marat Dukhanb5e3d172020-08-06 13:29:53 -07001408 "src/qs8-gavgpool/gen/7p7x-minmax-wasmsimd-c8-acc2.c",
1409 "src/qs8-gavgpool/gen/7p7x-minmax-wasmsimd-c16-acc2.c",
1410 "src/qs8-gavgpool/gen/7p7x-minmax-wasmsimd-c24-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001411 "src/qs8-gavgpool/gen/7x-minmax-wasmsimd-c8-acc2.c",
1412 "src/qs8-gavgpool/gen/7x-minmax-wasmsimd-c16-acc2.c",
1413 "src/qs8-gavgpool/gen/7x-minmax-wasmsimd-c24-acc2.c",
Marat Dukhan27203da2020-08-05 15:19:03 -07001414 "src/qs8-gemm/gen/1x4c8-minmax-wasmsimd-ld64.c",
Marat Dukhan27203da2020-08-05 15:19:03 -07001415 "src/qs8-gemm/gen/1x4c8-minmax-wasmsimd-ld128.c",
Marat Dukhan27203da2020-08-05 15:19:03 -07001416 "src/qs8-gemm/gen/1x4c8-xw-minmax-wasmsimd.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001417 "src/qs8-gemm/gen/2x4c8-minmax-wasmsimd-ld64.c",
1418 "src/qs8-gemm/gen/2x4c8-minmax-wasmsimd-ld128.c",
Marat Dukhan27203da2020-08-05 15:19:03 -07001419 "src/qs8-gemm/gen/2x4c8-xw-minmax-wasmsimd.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001420 "src/qs8-gemm/gen/3x4c8-minmax-wasmsimd-ld64.c",
1421 "src/qs8-gemm/gen/3x4c8-minmax-wasmsimd-ld128.c",
Marat Dukhan27203da2020-08-05 15:19:03 -07001422 "src/qs8-gemm/gen/3x4c8-xw-minmax-wasmsimd.c",
1423 "src/qs8-igemm/gen/1x4c8-minmax-wasmsimd-ld64.c",
Marat Dukhan27203da2020-08-05 15:19:03 -07001424 "src/qs8-igemm/gen/1x4c8-minmax-wasmsimd-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001425 "src/qs8-igemm/gen/2x4c8-minmax-wasmsimd-ld64.c",
Marat Dukhan27203da2020-08-05 15:19:03 -07001426 "src/qs8-igemm/gen/2x4c8-minmax-wasmsimd-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001427 "src/qs8-igemm/gen/3x4c8-minmax-wasmsimd-ld64.c",
Marat Dukhan27203da2020-08-05 15:19:03 -07001428 "src/qs8-igemm/gen/3x4c8-minmax-wasmsimd-ld128.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -07001429 "src/qs8-requantization/fp32-wasmsimd.c",
Marat Dukhan138560c2020-08-03 18:57:34 -07001430 "src/qs8-requantization/q31-wasmsimd.c",
Marat Dukhan5df27f82020-09-02 23:59:21 -07001431 "src/qs8-vadd/gen/minmax-wasmsimd-x8.c",
1432 "src/qs8-vadd/gen/minmax-wasmsimd-x16.c",
1433 "src/qs8-vadd/gen/minmax-wasmsimd-x24.c",
1434 "src/qs8-vadd/gen/minmax-wasmsimd-x32.c",
1435 "src/qs8-vaddc/gen/minmax-wasmsimd-x8.c",
1436 "src/qs8-vaddc/gen/minmax-wasmsimd-x16.c",
1437 "src/qs8-vaddc/gen/minmax-wasmsimd-x24.c",
1438 "src/qs8-vaddc/gen/minmax-wasmsimd-x32.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07001439 "src/qu8-requantization/fp32-wasmsimd.c",
Marat Dukhan22de5e72020-08-03 21:53:18 -07001440 "src/qu8-requantization/q31-wasmsimd.c",
Marat Dukhan8ee37012020-07-16 13:17:13 -07001441 "src/x32-fill/wasmsimd.c",
Marat Dukhan66d99e92020-07-16 12:56:21 -07001442 "src/x32-packx/x4-wasmsimd.c",
Marat Dukhan9306ae02020-07-16 15:51:13 -07001443 "src/x32-pad/wasmsimd.c",
Marat Dukhan9d4bfa22020-07-16 19:07:04 -07001444 "src/x32-unpool/wasmsimd.c",
Marat Dukhane3b78762020-07-16 20:02:58 -07001445 "src/x32-zip/x2-wasmsimd.c",
1446 "src/x32-zip/x3-wasmsimd.c",
1447 "src/x32-zip/x4-wasmsimd.c",
1448 "src/x32-zip/xm-wasmsimd.c",
Marat Dukhan290055c2020-06-09 12:24:29 -07001449]
1450
Marat Dukhan08c4a432019-10-03 09:29:21 -07001451# ISA-specific micro-kernels
1452NEON_UKERNELS = [
Marat Dukhanef25c6d2020-07-24 00:59:40 -07001453 "src/f32-argmaxpool/4x-neon-c4.c",
1454 "src/f32-argmaxpool/9p8x-neon-c4.c",
1455 "src/f32-argmaxpool/9x-neon-c4.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001456 "src/f32-avgpool/9p8x-minmax-neon-c4.c",
1457 "src/f32-avgpool/9x-minmax-neon-c4.c",
Marat Dukhan56b10cd2020-05-18 09:35:49 -07001458 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neon-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001459 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neon-2x2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001460 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neon-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001461 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neon-2x2.c",
Marat Dukhan56b10cd2020-05-18 09:35:49 -07001462 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neon-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001463 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neon-2x2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001464 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neon-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001465 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neon-2x2.c",
Marat Dukhanc7634882020-12-07 15:11:12 -08001466 "src/f32-conv-hwc2chw/3x3s2p1c3x4-neon-2x2.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -07001467 "src/f32-dwconv/gen/up4x4-minmax-neon-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001468 "src/f32-dwconv/gen/up4x4-minmax-neon.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001469 "src/f32-dwconv/gen/up4x9-minmax-neon-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001470 "src/f32-dwconv/gen/up4x9-minmax-neon.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -07001471 "src/f32-dwconv/gen/up4x25-minmax-neon-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001472 "src/f32-dwconv/gen/up4x25-minmax-neon.c",
1473 "src/f32-dwconv/gen/up8x4-minmax-neon-acc2.c",
1474 "src/f32-dwconv/gen/up8x4-minmax-neon.c",
1475 "src/f32-dwconv/gen/up8x9-minmax-neon-acc2.c",
1476 "src/f32-dwconv/gen/up8x9-minmax-neon.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -07001477 "src/f32-dwconv/gen/up8x25-minmax-neon-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001478 "src/f32-dwconv/gen/up8x25-minmax-neon.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001479 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-1x4-acc2.c",
1480 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-1x4-acc3.c",
1481 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-1x4-acc4.c",
Marat Dukhanc581e482020-10-24 01:28:11 -07001482 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-1x4.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001483 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-2x4-acc2.c",
Marat Dukhanc581e482020-10-24 01:28:11 -07001484 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-2x4.c",
1485 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-3x4.c",
1486 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-4x4.c",
1487 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-5x4.c",
1488 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-6x4.c",
Marat Dukhan82f0c322020-10-25 19:17:35 -07001489 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neon-1x4-acc2.c",
1490 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neon-1x4-acc3.c",
1491 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neon-1x4-acc4.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001492 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neon-1x4.c",
Marat Dukhan82f0c322020-10-25 19:17:35 -07001493 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neon-2x4-acc2.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001494 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neon-2x4.c",
1495 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neon-3x4.c",
1496 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neon-4x4.c",
Marat Dukhan149f0ea2020-10-26 12:50:33 -07001497 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-1x4-acc2.c",
1498 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-1x4-acc3.c",
1499 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-1x4-acc4.c",
1500 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-1x4-acc5.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001501 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-1x4.c",
Marat Dukhan149f0ea2020-10-26 12:50:33 -07001502 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-2x4-acc2.c",
1503 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-2x4-acc3.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001504 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-2x4.c",
Marat Dukhan149f0ea2020-10-26 12:50:33 -07001505 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-3x4-acc2.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001506 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-3x4.c",
Marat Dukhan149f0ea2020-10-26 12:50:33 -07001507 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-4x4-acc2.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001508 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-4x4.c",
1509 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-5x4.c",
Marat Dukhan30d4b252020-10-29 16:33:22 -07001510 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neon-1x4-acc2.c",
1511 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neon-1x4-acc3.c",
1512 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neon-1x4-acc4.c",
1513 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neon-1x4-acc5.c",
1514 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neon-1x4.c",
1515 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neon-2x4-acc2.c",
1516 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neon-2x4-acc3.c",
1517 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neon-2x4.c",
Marat Dukhan30d4b252020-10-29 16:33:22 -07001518 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neon-3x4-acc2.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07001519 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neon-3x4.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -07001520 "src/f32-gavgpool-cw/neon-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001521 "src/f32-gavgpool/7p7x-minmax-neon-c4.c",
1522 "src/f32-gavgpool/7x-minmax-neon-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001523 "src/f32-gemm/gen-inc/1x8inc-minmax-neon-dup-ld64.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001524 "src/f32-gemm/gen-inc/1x8inc-minmax-neon-lane-ld64.c",
1525 "src/f32-gemm/gen-inc/1x8s4inc-minmax-neon.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001526 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-dup-ld64.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001527 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-dup-ld128.c",
1528 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-lane-ld64.c",
1529 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-lane-ld128.c",
1530 "src/f32-gemm/gen-inc/4x8s4inc-minmax-neon.c",
1531 "src/f32-gemm/gen-inc/5x8inc-minmax-neon-lane-ld64.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001532 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-dup-ld64.c",
1533 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-dup-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001534 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-lane-ld64.c",
1535 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-lane-ld128.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001536 "src/f32-gemm/gen-inc/6x8s4inc-minmax-neon.c",
1537 "src/f32-gemm/gen-inc/8x8s4inc-minmax-neon.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001538 "src/f32-gemm/gen/1x8-minmax-neon-dup-ld64.c",
1539 "src/f32-gemm/gen/1x8-minmax-neon-lane-ld64.c",
1540 "src/f32-gemm/gen/1x8s4-minmax-neon.c",
1541 "src/f32-gemm/gen/4x2-minmax-neon-lane-ld64.c",
1542 "src/f32-gemm/gen/4x8-minmax-neon-dup-ld64.c",
1543 "src/f32-gemm/gen/4x8-minmax-neon-dup-ld128.c",
1544 "src/f32-gemm/gen/4x8-minmax-neon-lane-ld64.c",
1545 "src/f32-gemm/gen/4x8-minmax-neon-lane-ld128.c",
1546 "src/f32-gemm/gen/4x8s4-minmax-neon.c",
1547 "src/f32-gemm/gen/5x8-minmax-neon-lane-ld64.c",
1548 "src/f32-gemm/gen/6x8-minmax-neon-dup-ld64.c",
1549 "src/f32-gemm/gen/6x8-minmax-neon-dup-ld128.c",
1550 "src/f32-gemm/gen/6x8-minmax-neon-lane-ld64.c",
1551 "src/f32-gemm/gen/6x8-minmax-neon-lane-ld128.c",
1552 "src/f32-gemm/gen/6x8s4-minmax-neon.c",
1553 "src/f32-gemm/gen/8x8s4-minmax-neon.c",
Artsiom Ablavatski2202c812021-01-22 14:16:43 -08001554 "src/f32-ibilinear-chw/gen/neon-p4.c",
1555 "src/f32-ibilinear-chw/gen/neon-p8.c",
Frank Barchard8247e212021-02-03 18:12:33 -08001556 "src/f32-ibilinear/gen/neon-c4.c",
1557 "src/f32-ibilinear/gen/neon-c8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001558 "src/f32-igemm/gen/1x8-minmax-neon-dup-ld64.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001559 "src/f32-igemm/gen/1x8-minmax-neon-lane-ld64.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001560 "src/f32-igemm/gen/1x8s4-minmax-neon.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001561 "src/f32-igemm/gen/4x2-minmax-neon-lane-ld64.c",
1562 "src/f32-igemm/gen/4x4-minmax-neon-lane-ld64.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001563 "src/f32-igemm/gen/4x8-minmax-neon-dup-ld64.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001564 "src/f32-igemm/gen/4x8-minmax-neon-dup-ld128.c",
1565 "src/f32-igemm/gen/4x8-minmax-neon-lane-ld64.c",
1566 "src/f32-igemm/gen/4x8-minmax-neon-lane-ld128.c",
1567 "src/f32-igemm/gen/4x8s4-minmax-neon.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001568 "src/f32-igemm/gen/6x8-minmax-neon-dup-ld64.c",
1569 "src/f32-igemm/gen/6x8-minmax-neon-dup-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001570 "src/f32-igemm/gen/6x8-minmax-neon-lane-ld64.c",
1571 "src/f32-igemm/gen/6x8-minmax-neon-lane-ld128.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001572 "src/f32-igemm/gen/6x8s4-minmax-neon.c",
1573 "src/f32-igemm/gen/8x8s4-minmax-neon.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001574 "src/f32-maxpool/9p8x-minmax-neon-c4.c",
1575 "src/f32-pavgpool/9p8x-minmax-neon-c4.c",
1576 "src/f32-pavgpool/9x-minmax-neon-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001577 "src/f32-ppmm/gen/4x8-minmax-neon.c",
1578 "src/f32-ppmm/gen/8x8-minmax-neon.c",
Frank Barcharda5316982020-07-23 13:19:28 -07001579 "src/f32-prelu/gen/neon-1x4.c",
1580 "src/f32-prelu/gen/neon-1x8.c",
1581 "src/f32-prelu/gen/neon-1x16.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001582 "src/f32-prelu/gen/neon-2x4.c",
1583 "src/f32-prelu/gen/neon-2x8.c",
Frank Barcharda5316982020-07-23 13:19:28 -07001584 "src/f32-prelu/gen/neon-2x16.c",
1585 "src/f32-prelu/gen/neon-4x4.c",
1586 "src/f32-prelu/gen/neon-4x8.c",
1587 "src/f32-prelu/gen/neon-4x16.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001588 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x4.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001589 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x8-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001590 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x8.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001591 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12-acc2.c",
1592 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001593 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001594 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16-acc2.c",
1595 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001596 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001597 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20-acc2.c",
1598 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001599 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20.c",
1600 "src/f32-raddstoreexpminusmax/gen/neon-p5-x4.c",
1601 "src/f32-raddstoreexpminusmax/gen/neon-p5-x8-acc2.c",
1602 "src/f32-raddstoreexpminusmax/gen/neon-p5-x8.c",
1603 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12-acc2.c",
1604 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12-acc3.c",
1605 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12.c",
1606 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16-acc2.c",
1607 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16-acc4.c",
1608 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16.c",
1609 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20-acc2.c",
1610 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20-acc5.c",
1611 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001612 "src/f32-rmax/neon.c",
Marat Dukhan5b86c432020-12-06 19:15:03 -08001613 "src/f32-spmm/gen/4x1-minmax-neon-pipelined.c",
1614 "src/f32-spmm/gen/4x1-minmax-neon-x2.c",
1615 "src/f32-spmm/gen/4x1-minmax-neon.c",
1616 "src/f32-spmm/gen/8x1-minmax-neon-pipelined.c",
1617 "src/f32-spmm/gen/8x1-minmax-neon-x2.c",
1618 "src/f32-spmm/gen/8x1-minmax-neon.c",
1619 "src/f32-spmm/gen/12x1-minmax-neon.c",
1620 "src/f32-spmm/gen/16x1-minmax-neon-pipelined.c",
1621 "src/f32-spmm/gen/16x1-minmax-neon-x2.c",
1622 "src/f32-spmm/gen/16x1-minmax-neon.c",
1623 "src/f32-spmm/gen/32x1-minmax-neon-pipelined.c",
1624 "src/f32-spmm/gen/32x1-minmax-neon-x2.c",
1625 "src/f32-spmm/gen/32x1-minmax-neon.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001626 "src/f32-vbinary/gen/vadd-minmax-neon-x4.c",
1627 "src/f32-vbinary/gen/vadd-minmax-neon-x8.c",
1628 "src/f32-vbinary/gen/vaddc-minmax-neon-x4.c",
1629 "src/f32-vbinary/gen/vaddc-minmax-neon-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -08001630 "src/f32-vbinary/gen/vmax-neon-x4.c",
1631 "src/f32-vbinary/gen/vmax-neon-x8.c",
1632 "src/f32-vbinary/gen/vmaxc-neon-x4.c",
1633 "src/f32-vbinary/gen/vmaxc-neon-x8.c",
1634 "src/f32-vbinary/gen/vmin-neon-x4.c",
1635 "src/f32-vbinary/gen/vmin-neon-x8.c",
1636 "src/f32-vbinary/gen/vminc-neon-x4.c",
1637 "src/f32-vbinary/gen/vminc-neon-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001638 "src/f32-vbinary/gen/vmul-minmax-neon-x4.c",
1639 "src/f32-vbinary/gen/vmul-minmax-neon-x8.c",
1640 "src/f32-vbinary/gen/vmulc-minmax-neon-x4.c",
1641 "src/f32-vbinary/gen/vmulc-minmax-neon-x8.c",
1642 "src/f32-vbinary/gen/vrsubc-minmax-neon-x4.c",
1643 "src/f32-vbinary/gen/vrsubc-minmax-neon-x8.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07001644 "src/f32-vbinary/gen/vsqrdiff-neon-x4.c",
1645 "src/f32-vbinary/gen/vsqrdiff-neon-x8.c",
1646 "src/f32-vbinary/gen/vsqrdiffc-neon-x4.c",
1647 "src/f32-vbinary/gen/vsqrdiffc-neon-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001648 "src/f32-vbinary/gen/vsub-minmax-neon-x4.c",
1649 "src/f32-vbinary/gen/vsub-minmax-neon-x8.c",
1650 "src/f32-vbinary/gen/vsubc-minmax-neon-x4.c",
1651 "src/f32-vbinary/gen/vsubc-minmax-neon-x8.c",
Marat Dukhanb1eec082021-05-05 23:24:55 -07001652 "src/f32-vclamp/gen/vclamp-neon-x4.c",
1653 "src/f32-vclamp/gen/vclamp-neon-x8.c",
Marat Dukhaned6baaf2020-12-01 15:07:08 -08001654 "src/f32-velu/gen/velu-neon-rr2-lut16-p3-x4.c",
1655 "src/f32-velu/gen/velu-neon-rr2-lut16-p3-x8.c",
1656 "src/f32-velu/gen/velu-neon-rr2-lut16-p3-x12.c",
1657 "src/f32-velu/gen/velu-neon-rr2-lut16-p3-x16.c",
1658 "src/f32-velu/gen/velu-neon-rr2-lut16-p3-x20.c",
1659 "src/f32-velu/gen/velu-neon-rr2-lut16-p3-x24.c",
1660 "src/f32-velu/gen/velu-neon-rr2-p6-x4.c",
1661 "src/f32-velu/gen/velu-neon-rr2-p6-x8.c",
1662 "src/f32-velu/gen/velu-neon-rr2-p6-x12.c",
1663 "src/f32-velu/gen/velu-neon-rr2-p6-x16.c",
1664 "src/f32-velu/gen/velu-neon-rr2-p6-x20.c",
1665 "src/f32-velu/gen/velu-neon-rr2-p6-x24.c",
Marat Dukhan6674d692021-05-05 22:27:00 -07001666 "src/f32-vhswish/gen/vhswish-neon-x4.c",
1667 "src/f32-vhswish/gen/vhswish-neon-x8.c",
1668 "src/f32-vhswish/gen/vhswish-neon-x16.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07001669 "src/f32-vlrelu/gen/vlrelu-neon-x4.c",
1670 "src/f32-vlrelu/gen/vlrelu-neon-x8.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001671 "src/f32-vmulcaddc/gen/c4-minmax-neon-2x.c",
1672 "src/f32-vmulcaddc/gen/c8-minmax-neon-2x.c",
Marat Dukhanb1eec082021-05-05 23:24:55 -07001673 "src/f32-vrelu/gen/vrelu-neon-x4.c",
1674 "src/f32-vrelu/gen/vrelu-neon-x8.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07001675 "src/f32-vrnd/gen/vrndd-neon-x4.c",
1676 "src/f32-vrnd/gen/vrndd-neon-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001677 "src/f32-vrnd/gen/vrndne-neon-x4.c",
1678 "src/f32-vrnd/gen/vrndne-neon-x8.c",
1679 "src/f32-vrnd/gen/vrndu-neon-x4.c",
1680 "src/f32-vrnd/gen/vrndu-neon-x8.c",
1681 "src/f32-vrnd/gen/vrndz-neon-x4.c",
1682 "src/f32-vrnd/gen/vrndz-neon-x8.c",
Marat Dukhanb1eec082021-05-05 23:24:55 -07001683 "src/f32-vsigmoid/gen/vsigmoid-neon-rr2-lut64-p2-nr2recps-x4.c",
1684 "src/f32-vsigmoid/gen/vsigmoid-neon-rr2-lut64-p2-nr2recps-x8.c",
1685 "src/f32-vsigmoid/gen/vsigmoid-neon-rr2-lut64-p2-nr2recps-x12.c",
1686 "src/f32-vsigmoid/gen/vsigmoid-neon-rr2-lut64-p2-nr2recps-x16.c",
1687 "src/f32-vsigmoid/gen/vsigmoid-neon-rr2-lut64-p2-nr2recps-x20.c",
1688 "src/f32-vsigmoid/gen/vsigmoid-neon-rr2-lut64-p2-nr2recps-x24.c",
1689 "src/f32-vsigmoid/gen/vsigmoid-neon-rr2-lut2048-p1-nr2recps-x4.c",
1690 "src/f32-vsigmoid/gen/vsigmoid-neon-rr2-lut2048-p1-nr2recps-x8.c",
1691 "src/f32-vsigmoid/gen/vsigmoid-neon-rr2-lut2048-p1-nr2recps-x12.c",
1692 "src/f32-vsigmoid/gen/vsigmoid-neon-rr2-lut2048-p1-nr2recps-x16.c",
1693 "src/f32-vsigmoid/gen/vsigmoid-neon-rr2-lut2048-p1-nr2recps-x20.c",
1694 "src/f32-vsigmoid/gen/vsigmoid-neon-rr2-lut2048-p1-nr2recps-x24.c",
1695 "src/f32-vsigmoid/gen/vsigmoid-neon-rr2-p5-nr2recps-x4.c",
1696 "src/f32-vsigmoid/gen/vsigmoid-neon-rr2-p5-nr2recps-x8.c",
1697 "src/f32-vsigmoid/gen/vsigmoid-neon-rr2-p5-nr2recps-x12.c",
1698 "src/f32-vsigmoid/gen/vsigmoid-neon-rr2-p5-nr2recps-x16.c",
1699 "src/f32-vsigmoid/gen/vsigmoid-neon-rr2-p5-nr2recps-x20.c",
1700 "src/f32-vsigmoid/gen/vsigmoid-neon-rr2-p5-nr2recps-x24.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07001701 "src/f32-vunary/gen/vabs-neon-x4.c",
1702 "src/f32-vunary/gen/vabs-neon-x8.c",
1703 "src/f32-vunary/gen/vneg-neon-x4.c",
1704 "src/f32-vunary/gen/vneg-neon-x8.c",
1705 "src/f32-vunary/gen/vsqr-neon-x4.c",
1706 "src/f32-vunary/gen/vsqr-neon-x8.c",
Marat Dukhande390d42020-11-29 19:32:18 -08001707 "src/math/expm1minus-neon-rr2-lut16-p3.c",
1708 "src/math/expm1minus-neon-rr2-p6.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001709 "src/math/roundd-neon-addsub.c",
1710 "src/math/roundd-neon-cvt.c",
1711 "src/math/roundne-neon-addsub.c",
1712 "src/math/roundu-neon-addsub.c",
1713 "src/math/roundu-neon-cvt.c",
1714 "src/math/roundz-neon-addsub.c",
1715 "src/math/roundz-neon-cvt.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001716 "src/math/sigmoid-neon-rr2-lut64-p2-nr2recps.c",
1717 "src/math/sigmoid-neon-rr2-lut2048-p1-nr2recps.c",
1718 "src/math/sigmoid-neon-rr2-p5-nr2recps.c",
1719 "src/math/sqrt-neon-nr1rsqrts.c",
1720 "src/math/sqrt-neon-nr2rsqrts.c",
1721 "src/math/sqrt-neon-nr3rsqrts.c",
Marat Dukhan023bcf92020-08-10 12:40:50 -07001722 "src/qs8-dwconv/gen/up8x9-minmax-neon-mul16.c",
Marat Dukhan4ed14882021-05-12 17:50:40 -07001723 "src/qs8-dwconv/gen/up8x25-minmax-neon-mul16.c",
Marat Dukhan023bcf92020-08-10 12:40:50 -07001724 "src/qs8-dwconv/gen/up16x9-minmax-neon-mul16.c",
Marat Dukhan4ed14882021-05-12 17:50:40 -07001725 "src/qs8-dwconv/gen/up16x25-minmax-neon-mul16.c",
Marat Dukhan023bcf92020-08-10 12:40:50 -07001726 "src/qs8-dwconv/gen/up24x9-minmax-neon-mul16.c",
Marat Dukhan4ed14882021-05-12 17:50:40 -07001727 "src/qs8-dwconv/gen/up24x25-minmax-neon-mul16.c",
Marat Dukhan023bcf92020-08-10 12:40:50 -07001728 "src/qs8-dwconv/gen/up32x9-minmax-neon-mul16.c",
Marat Dukhan4ed14882021-05-12 17:50:40 -07001729 "src/qs8-dwconv/gen/up32x25-minmax-neon-mul16.c",
Marat Dukhan281262d2020-08-10 13:23:21 -07001730 "src/qs8-gavgpool/gen/7p7x-minmax-neon-c8-acc2.c",
1731 "src/qs8-gavgpool/gen/7p7x-minmax-neon-c16-acc2.c",
1732 "src/qs8-gavgpool/gen/7p7x-minmax-neon-c24-acc2.c",
1733 "src/qs8-gavgpool/gen/7p7x-minmax-neon-c32-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001734 "src/qs8-gavgpool/gen/7x-minmax-neon-c8-acc2.c",
1735 "src/qs8-gavgpool/gen/7x-minmax-neon-c16-acc2.c",
1736 "src/qs8-gavgpool/gen/7x-minmax-neon-c24-acc2.c",
1737 "src/qs8-gavgpool/gen/7x-minmax-neon-c32-acc2.c",
Frank Barchard1f51d382021-04-20 14:07:14 -07001738 "src/qs8-gemm/gen/1x8-minmax-neon-mlal-lane-prfm.c",
Frank Barchardec0bf142021-01-15 10:14:05 -08001739 "src/qs8-gemm/gen/1x8-minmax-neon-mlal-lane.c",
Frank Barchardec0bf142021-01-15 10:14:05 -08001740 "src/qs8-gemm/gen/1x8-minmax-neon-mull-addw-dup.c",
Frank Barchard8247e212021-02-03 18:12:33 -08001741 "src/qs8-gemm/gen/1x8c2-minmax-neon-mlal-padal-dup.c",
Frank Barchard2302ffd2021-01-22 14:08:02 -08001742 "src/qs8-gemm/gen/1x8c2-minmax-neon-mull-padal-dup.c",
Frank Barchardda78da12021-03-02 14:28:00 -08001743 "src/qs8-gemm/gen/1x8c8-minmax-neon-mlal-padal.c",
Frank Barchard8247e212021-02-03 18:12:33 -08001744 "src/qs8-gemm/gen/1x8c8-minmax-neon-mull-padal.c",
Frank Barchard671d1b02021-03-10 15:42:08 -08001745 "src/qs8-gemm/gen/1x8c16-minmax-neon-mlal-padal.c",
Frank Barchard1f51d382021-04-20 14:07:14 -07001746 "src/qs8-gemm/gen/1x16-minmax-neon-mlal-lane-prfm.c",
Frank Barchardcadd4222021-01-20 16:27:25 -08001747 "src/qs8-gemm/gen/1x16-minmax-neon-mlal-lane.c",
1748 "src/qs8-gemm/gen/1x16-minmax-neon-mull-addw-dup.c",
Frank Barchard8247e212021-02-03 18:12:33 -08001749 "src/qs8-gemm/gen/1x16c2-minmax-neon-mlal-padal-dup.c",
Frank Barchard2302ffd2021-01-22 14:08:02 -08001750 "src/qs8-gemm/gen/1x16c2-minmax-neon-mull-padal-dup.c",
Frank Barchardda78da12021-03-02 14:28:00 -08001751 "src/qs8-gemm/gen/1x16c8-minmax-neon-mlal-padal.c",
Frank Barchard8247e212021-02-03 18:12:33 -08001752 "src/qs8-gemm/gen/1x16c8-minmax-neon-mull-padal.c",
Frank Barchard671d1b02021-03-10 15:42:08 -08001753 "src/qs8-gemm/gen/1x16c16-minmax-neon-mlal-padal.c",
Frank Barchard1f51d382021-04-20 14:07:14 -07001754 "src/qs8-gemm/gen/2x8-minmax-neon-mlal-lane-prfm.c",
Frank Barchardcadd4222021-01-20 16:27:25 -08001755 "src/qs8-gemm/gen/2x8-minmax-neon-mlal-lane.c",
Frank Barchardec0bf142021-01-15 10:14:05 -08001756 "src/qs8-gemm/gen/2x8-minmax-neon-mull-addw-dup.c",
Frank Barchard8247e212021-02-03 18:12:33 -08001757 "src/qs8-gemm/gen/2x8c2-minmax-neon-mlal-padal-dup.c",
Frank Barchard2302ffd2021-01-22 14:08:02 -08001758 "src/qs8-gemm/gen/2x8c2-minmax-neon-mull-padal-dup.c",
Frank Barchardda78da12021-03-02 14:28:00 -08001759 "src/qs8-gemm/gen/2x8c8-minmax-neon-mlal-padal.c",
Frank Barchard8247e212021-02-03 18:12:33 -08001760 "src/qs8-gemm/gen/2x8c8-minmax-neon-mull-padal.c",
Frank Barchard671d1b02021-03-10 15:42:08 -08001761 "src/qs8-gemm/gen/2x8c16-minmax-neon-mlal-padal.c",
Frank Barchard1f51d382021-04-20 14:07:14 -07001762 "src/qs8-gemm/gen/2x16-minmax-neon-mlal-lane-prfm.c",
Frank Barchardcadd4222021-01-20 16:27:25 -08001763 "src/qs8-gemm/gen/2x16-minmax-neon-mlal-lane.c",
1764 "src/qs8-gemm/gen/2x16-minmax-neon-mull-addw-dup.c",
Frank Barchard8247e212021-02-03 18:12:33 -08001765 "src/qs8-gemm/gen/2x16c2-minmax-neon-mlal-padal-dup.c",
Frank Barchard2302ffd2021-01-22 14:08:02 -08001766 "src/qs8-gemm/gen/2x16c2-minmax-neon-mull-padal-dup.c",
Frank Barchardda78da12021-03-02 14:28:00 -08001767 "src/qs8-gemm/gen/2x16c8-minmax-neon-mlal-padal.c",
Frank Barchard8247e212021-02-03 18:12:33 -08001768 "src/qs8-gemm/gen/2x16c8-minmax-neon-mull-padal.c",
Frank Barchard671d1b02021-03-10 15:42:08 -08001769 "src/qs8-gemm/gen/2x16c16-minmax-neon-mlal-padal.c",
Frank Barchard1f51d382021-04-20 14:07:14 -07001770 "src/qs8-gemm/gen/3x8-minmax-neon-mlal-lane-prfm.c",
Frank Barchard2302ffd2021-01-22 14:08:02 -08001771 "src/qs8-gemm/gen/3x8-minmax-neon-mlal-lane.c",
1772 "src/qs8-gemm/gen/3x8-minmax-neon-mull-addw-dup.c",
Frank Barchard8247e212021-02-03 18:12:33 -08001773 "src/qs8-gemm/gen/3x8c2-minmax-neon-mlal-padal-dup.c",
Frank Barchard2302ffd2021-01-22 14:08:02 -08001774 "src/qs8-gemm/gen/3x8c2-minmax-neon-mull-padal-dup.c",
Frank Barchardda78da12021-03-02 14:28:00 -08001775 "src/qs8-gemm/gen/3x8c8-minmax-neon-mlal-padal.c",
Frank Barchard8247e212021-02-03 18:12:33 -08001776 "src/qs8-gemm/gen/3x8c8-minmax-neon-mull-padal.c",
Frank Barchard671d1b02021-03-10 15:42:08 -08001777 "src/qs8-gemm/gen/3x8c16-minmax-neon-mlal-padal.c",
Frank Barchard1f51d382021-04-20 14:07:14 -07001778 "src/qs8-gemm/gen/3x16-minmax-neon-mlal-lane-prfm.c",
Frank Barchard8247e212021-02-03 18:12:33 -08001779 "src/qs8-gemm/gen/3x16-minmax-neon-mlal-lane.c",
1780 "src/qs8-gemm/gen/3x16-minmax-neon-mull-addw-dup.c",
1781 "src/qs8-gemm/gen/3x16c2-minmax-neon-mlal-padal-dup.c",
1782 "src/qs8-gemm/gen/3x16c2-minmax-neon-mull-padal-dup.c",
Frank Barchardda78da12021-03-02 14:28:00 -08001783 "src/qs8-gemm/gen/3x16c8-minmax-neon-mlal-padal.c",
Frank Barchard8247e212021-02-03 18:12:33 -08001784 "src/qs8-gemm/gen/3x16c8-minmax-neon-mull-padal.c",
Frank Barchard671d1b02021-03-10 15:42:08 -08001785 "src/qs8-gemm/gen/3x16c16-minmax-neon-mlal-padal.c",
Frank Barchard1f51d382021-04-20 14:07:14 -07001786 "src/qs8-gemm/gen/4x8-minmax-neon-mlal-lane-prfm.c",
Frank Barchardcadd4222021-01-20 16:27:25 -08001787 "src/qs8-gemm/gen/4x8-minmax-neon-mlal-lane.c",
Frank Barchardec0bf142021-01-15 10:14:05 -08001788 "src/qs8-gemm/gen/4x8-minmax-neon-mull-addw-dup.c",
Frank Barchard8247e212021-02-03 18:12:33 -08001789 "src/qs8-gemm/gen/4x8c2-minmax-neon-mlal-padal-dup.c",
Frank Barchard2302ffd2021-01-22 14:08:02 -08001790 "src/qs8-gemm/gen/4x8c2-minmax-neon-mull-padal-dup.c",
Frank Barchardda78da12021-03-02 14:28:00 -08001791 "src/qs8-gemm/gen/4x8c8-minmax-neon-mlal-padal.c",
Frank Barchard8247e212021-02-03 18:12:33 -08001792 "src/qs8-gemm/gen/4x8c8-minmax-neon-mull-padal.c",
Frank Barchard671d1b02021-03-10 15:42:08 -08001793 "src/qs8-gemm/gen/4x8c16-minmax-neon-mlal-padal.c",
Frank Barchard1f51d382021-04-20 14:07:14 -07001794 "src/qs8-gemm/gen/4x16-minmax-neon-mlal-lane-prfm.c",
Frank Barchardcadd4222021-01-20 16:27:25 -08001795 "src/qs8-gemm/gen/4x16-minmax-neon-mlal-lane.c",
1796 "src/qs8-gemm/gen/4x16-minmax-neon-mull-addw-dup.c",
Frank Barchard8247e212021-02-03 18:12:33 -08001797 "src/qs8-gemm/gen/4x16c2-minmax-neon-mlal-padal-dup.c",
Frank Barchard2302ffd2021-01-22 14:08:02 -08001798 "src/qs8-gemm/gen/4x16c2-minmax-neon-mull-padal-dup.c",
Frank Barchardda78da12021-03-02 14:28:00 -08001799 "src/qs8-gemm/gen/4x16c8-minmax-neon-mlal-padal.c",
Frank Barchard58990122021-01-29 16:09:56 -08001800 "src/qs8-gemm/gen/4x16c8-minmax-neon-mull-padal.c",
Frank Barchard71c4d1a2021-02-19 10:07:23 -08001801 "src/qs8-gemm/gen/4x16c16-minmax-neon-mlal-padal.c",
Frank Barcharddddb38f2021-05-03 16:51:48 -07001802 "src/qs8-gemm/gen/6x8-minmax-neon-mlal-lane-prfm.c",
1803 "src/qs8-gemm/gen/6x8-minmax-neon-mlal-lane.c",
1804 "src/qs8-gemm/gen/6x16-minmax-neon-mlal-lane-prfm.c",
1805 "src/qs8-gemm/gen/6x16-minmax-neon-mlal-lane.c",
Frank Barchard1f51d382021-04-20 14:07:14 -07001806 "src/qs8-igemm/gen/1x8-minmax-neon-mlal-lane-prfm.c",
Frank Barchardec0bf142021-01-15 10:14:05 -08001807 "src/qs8-igemm/gen/1x8-minmax-neon-mlal-lane.c",
Frank Barchard671d1b02021-03-10 15:42:08 -08001808 "src/qs8-igemm/gen/1x8-minmax-neon-mull-addw-dup.c",
1809 "src/qs8-igemm/gen/1x8c2-minmax-neon-mlal-padal-dup.c",
1810 "src/qs8-igemm/gen/1x8c2-minmax-neon-mull-padal-dup.c",
Frank Barchardda78da12021-03-02 14:28:00 -08001811 "src/qs8-igemm/gen/1x8c8-minmax-neon-mlal-padal.c",
1812 "src/qs8-igemm/gen/1x8c8-minmax-neon-mull-padal.c",
Frank Barchard671d1b02021-03-10 15:42:08 -08001813 "src/qs8-igemm/gen/1x8c16-minmax-neon-mlal-padal.c",
Frank Barchard1f51d382021-04-20 14:07:14 -07001814 "src/qs8-igemm/gen/1x16-minmax-neon-mlal-lane-prfm.c",
Frank Barchard671d1b02021-03-10 15:42:08 -08001815 "src/qs8-igemm/gen/1x16-minmax-neon-mlal-lane.c",
1816 "src/qs8-igemm/gen/1x16-minmax-neon-mull-addw-dup.c",
1817 "src/qs8-igemm/gen/1x16c2-minmax-neon-mlal-padal-dup.c",
1818 "src/qs8-igemm/gen/1x16c2-minmax-neon-mull-padal-dup.c",
1819 "src/qs8-igemm/gen/1x16c8-minmax-neon-mlal-padal.c",
1820 "src/qs8-igemm/gen/1x16c8-minmax-neon-mull-padal.c",
1821 "src/qs8-igemm/gen/1x16c16-minmax-neon-mlal-padal.c",
Frank Barchard1f51d382021-04-20 14:07:14 -07001822 "src/qs8-igemm/gen/2x8-minmax-neon-mlal-lane-prfm.c",
Frank Barchard671d1b02021-03-10 15:42:08 -08001823 "src/qs8-igemm/gen/2x8-minmax-neon-mlal-lane.c",
1824 "src/qs8-igemm/gen/2x8-minmax-neon-mull-addw-dup.c",
1825 "src/qs8-igemm/gen/2x8c2-minmax-neon-mlal-padal-dup.c",
1826 "src/qs8-igemm/gen/2x8c2-minmax-neon-mull-padal-dup.c",
Frank Barchardda78da12021-03-02 14:28:00 -08001827 "src/qs8-igemm/gen/2x8c8-minmax-neon-mlal-padal.c",
1828 "src/qs8-igemm/gen/2x8c8-minmax-neon-mull-padal.c",
Frank Barchard671d1b02021-03-10 15:42:08 -08001829 "src/qs8-igemm/gen/2x8c16-minmax-neon-mlal-padal.c",
Frank Barchard1f51d382021-04-20 14:07:14 -07001830 "src/qs8-igemm/gen/2x16-minmax-neon-mlal-lane-prfm.c",
Frank Barchard671d1b02021-03-10 15:42:08 -08001831 "src/qs8-igemm/gen/2x16-minmax-neon-mlal-lane.c",
1832 "src/qs8-igemm/gen/2x16-minmax-neon-mull-addw-dup.c",
1833 "src/qs8-igemm/gen/2x16c2-minmax-neon-mlal-padal-dup.c",
1834 "src/qs8-igemm/gen/2x16c2-minmax-neon-mull-padal-dup.c",
1835 "src/qs8-igemm/gen/2x16c8-minmax-neon-mlal-padal.c",
1836 "src/qs8-igemm/gen/2x16c8-minmax-neon-mull-padal.c",
1837 "src/qs8-igemm/gen/2x16c16-minmax-neon-mlal-padal.c",
Frank Barchard1f51d382021-04-20 14:07:14 -07001838 "src/qs8-igemm/gen/3x8-minmax-neon-mlal-lane-prfm.c",
Frank Barchard671d1b02021-03-10 15:42:08 -08001839 "src/qs8-igemm/gen/3x8-minmax-neon-mlal-lane.c",
1840 "src/qs8-igemm/gen/3x8-minmax-neon-mull-addw-dup.c",
1841 "src/qs8-igemm/gen/3x8c2-minmax-neon-mlal-padal-dup.c",
1842 "src/qs8-igemm/gen/3x8c2-minmax-neon-mull-padal-dup.c",
Frank Barchardda78da12021-03-02 14:28:00 -08001843 "src/qs8-igemm/gen/3x8c8-minmax-neon-mlal-padal.c",
1844 "src/qs8-igemm/gen/3x8c8-minmax-neon-mull-padal.c",
Frank Barchard671d1b02021-03-10 15:42:08 -08001845 "src/qs8-igemm/gen/3x8c16-minmax-neon-mlal-padal.c",
Frank Barchard1f51d382021-04-20 14:07:14 -07001846 "src/qs8-igemm/gen/3x16-minmax-neon-mlal-lane-prfm.c",
Frank Barchard671d1b02021-03-10 15:42:08 -08001847 "src/qs8-igemm/gen/3x16-minmax-neon-mlal-lane.c",
1848 "src/qs8-igemm/gen/3x16-minmax-neon-mull-addw-dup.c",
1849 "src/qs8-igemm/gen/3x16c2-minmax-neon-mlal-padal-dup.c",
1850 "src/qs8-igemm/gen/3x16c2-minmax-neon-mull-padal-dup.c",
1851 "src/qs8-igemm/gen/3x16c8-minmax-neon-mlal-padal.c",
1852 "src/qs8-igemm/gen/3x16c8-minmax-neon-mull-padal.c",
1853 "src/qs8-igemm/gen/3x16c16-minmax-neon-mlal-padal.c",
Frank Barchard1f51d382021-04-20 14:07:14 -07001854 "src/qs8-igemm/gen/4x8-minmax-neon-mlal-lane-prfm.c",
Frank Barchard671d1b02021-03-10 15:42:08 -08001855 "src/qs8-igemm/gen/4x8-minmax-neon-mlal-lane.c",
1856 "src/qs8-igemm/gen/4x8-minmax-neon-mull-addw-dup.c",
1857 "src/qs8-igemm/gen/4x8c2-minmax-neon-mlal-padal-dup.c",
1858 "src/qs8-igemm/gen/4x8c2-minmax-neon-mull-padal-dup.c",
Frank Barchardda78da12021-03-02 14:28:00 -08001859 "src/qs8-igemm/gen/4x8c8-minmax-neon-mlal-padal.c",
1860 "src/qs8-igemm/gen/4x8c8-minmax-neon-mull-padal.c",
Frank Barchardda78da12021-03-02 14:28:00 -08001861 "src/qs8-igemm/gen/4x8c16-minmax-neon-mlal-padal.c",
Frank Barchard1f51d382021-04-20 14:07:14 -07001862 "src/qs8-igemm/gen/4x16-minmax-neon-mlal-lane-prfm.c",
Frank Barchard671d1b02021-03-10 15:42:08 -08001863 "src/qs8-igemm/gen/4x16-minmax-neon-mlal-lane.c",
Frank Barchardda78da12021-03-02 14:28:00 -08001864 "src/qs8-igemm/gen/4x16-minmax-neon-mull-addw-dup.c",
Frank Barchard671d1b02021-03-10 15:42:08 -08001865 "src/qs8-igemm/gen/4x16c2-minmax-neon-mlal-padal-dup.c",
1866 "src/qs8-igemm/gen/4x16c2-minmax-neon-mull-padal-dup.c",
1867 "src/qs8-igemm/gen/4x16c8-minmax-neon-mlal-padal.c",
1868 "src/qs8-igemm/gen/4x16c8-minmax-neon-mull-padal.c",
1869 "src/qs8-igemm/gen/4x16c16-minmax-neon-mlal-padal.c",
Frank Barcharddddb38f2021-05-03 16:51:48 -07001870 "src/qs8-igemm/gen/6x8-minmax-neon-mlal-lane-prfm.c",
1871 "src/qs8-igemm/gen/6x8-minmax-neon-mlal-lane.c",
1872 "src/qs8-igemm/gen/6x16-minmax-neon-mlal-lane-prfm.c",
1873 "src/qs8-igemm/gen/6x16-minmax-neon-mlal-lane.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -07001874 "src/qs8-requantization/fp32-neon.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001875 "src/qs8-requantization/precise-neon.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -07001876 "src/qs8-requantization/q31-neon.c",
Marat Dukhanba7b2792020-09-02 14:26:45 -07001877 "src/qs8-vadd/gen/minmax-neon-ld64-x8.c",
1878 "src/qs8-vadd/gen/minmax-neon-ld64-x16.c",
1879 "src/qs8-vadd/gen/minmax-neon-ld64-x24.c",
1880 "src/qs8-vadd/gen/minmax-neon-ld64-x32.c",
1881 "src/qs8-vaddc/gen/minmax-neon-ld64-x8.c",
1882 "src/qs8-vaddc/gen/minmax-neon-ld64-x16.c",
1883 "src/qs8-vaddc/gen/minmax-neon-ld64-x24.c",
1884 "src/qs8-vaddc/gen/minmax-neon-ld64-x32.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -07001885 "src/qu8-avgpool/9p8x-minmax-neon-c8.c",
1886 "src/qu8-avgpool/9x-minmax-neon-c8.c",
1887 "src/qu8-dwconv/up8x9-minmax-neon.c",
1888 "src/qu8-gavgpool/7p7x-minmax-neon-c8.c",
1889 "src/qu8-gavgpool/7x-minmax-neon-c8.c",
1890 "src/qu8-gemm/4x8-minmax-neon.c",
1891 "src/qu8-gemm/8x8-minmax-neon.c",
1892 "src/qu8-igemm/4x8-minmax-neon.c",
1893 "src/qu8-igemm/8x8-minmax-neon.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07001894 "src/qu8-requantization/fp32-neon.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001895 "src/qu8-requantization/precise-neon.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07001896 "src/qu8-requantization/q31-neon.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -07001897 "src/qu8-vadd/minmax-neon.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001898 "src/u8-maxpool/9p8x-minmax-neon-c16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001899 "src/u8-rmax/neon.c",
Marat Dukhan6674d692021-05-05 22:27:00 -07001900 "src/u8-vclamp/neon-x64.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001901 "src/x8-zip/x2-neon.c",
1902 "src/x8-zip/x3-neon.c",
1903 "src/x8-zip/x4-neon.c",
1904 "src/x8-zip/xm-neon.c",
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -07001905 "src/x32-fill/neon.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001906 "src/x32-packx/x4-neon-st4.c",
Marat Dukhan63523d42020-05-22 17:07:33 -07001907 "src/x32-pad/neon.c",
Marat Dukhan57dccd82020-04-14 00:53:10 -07001908 "src/x32-unpool/neon.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001909 "src/x32-zip/x2-neon.c",
1910 "src/x32-zip/x3-neon.c",
1911 "src/x32-zip/x4-neon.c",
1912 "src/x32-zip/xm-neon.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001913]
1914
1915NEONFMA_UKERNELS = [
Frank Barchard04336c12020-10-22 16:48:55 -07001916 "src/f32-dwconv/gen/up4x4-minmax-neonfma-acc2.c",
1917 "src/f32-dwconv/gen/up4x4-minmax-neonfma.c",
1918 "src/f32-dwconv/gen/up4x9-minmax-neonfma-acc2.c",
1919 "src/f32-dwconv/gen/up4x9-minmax-neonfma.c",
1920 "src/f32-dwconv/gen/up4x25-minmax-neonfma-acc2.c",
1921 "src/f32-dwconv/gen/up4x25-minmax-neonfma.c",
1922 "src/f32-dwconv/gen/up8x4-minmax-neonfma-acc2.c",
1923 "src/f32-dwconv/gen/up8x4-minmax-neonfma.c",
1924 "src/f32-dwconv/gen/up8x9-minmax-neonfma-acc2.c",
1925 "src/f32-dwconv/gen/up8x9-minmax-neonfma.c",
1926 "src/f32-dwconv/gen/up8x25-minmax-neonfma-acc2.c",
1927 "src/f32-dwconv/gen/up8x25-minmax-neonfma.c",
1928 "src/f32-gemm/gen-inc/1x8inc-minmax-neonfma-dup-ld64.c",
1929 "src/f32-gemm/gen-inc/1x8s4inc-minmax-neonfma.c",
1930 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-dup-ld64.c",
1931 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-dup-ld128.c",
1932 "src/f32-gemm/gen-inc/4x8s4inc-minmax-neonfma.c",
1933 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-dup-ld64.c",
1934 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-dup-ld128.c",
1935 "src/f32-gemm/gen-inc/6x8s4inc-minmax-neonfma.c",
1936 "src/f32-gemm/gen-inc/8x8s4inc-minmax-neonfma.c",
1937 "src/f32-gemm/gen/1x8-minmax-neonfma-dup-ld64.c",
1938 "src/f32-gemm/gen/1x8s4-minmax-neonfma.c",
1939 "src/f32-gemm/gen/4x8-minmax-neonfma-dup-ld64.c",
1940 "src/f32-gemm/gen/4x8-minmax-neonfma-dup-ld128.c",
1941 "src/f32-gemm/gen/4x8s4-minmax-neonfma.c",
1942 "src/f32-gemm/gen/6x8-minmax-neonfma-dup-ld64.c",
1943 "src/f32-gemm/gen/6x8-minmax-neonfma-dup-ld128.c",
1944 "src/f32-gemm/gen/6x8s4-minmax-neonfma.c",
1945 "src/f32-gemm/gen/8x8s4-minmax-neonfma.c",
Artsiom Ablavatski2202c812021-01-22 14:16:43 -08001946 "src/f32-ibilinear-chw/gen/neonfma-p4.c",
1947 "src/f32-ibilinear-chw/gen/neonfma-p8.c",
Frank Barchard8247e212021-02-03 18:12:33 -08001948 "src/f32-ibilinear/gen/neonfma-c4.c",
1949 "src/f32-ibilinear/gen/neonfma-c8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001950 "src/f32-igemm/gen/1x8-minmax-neonfma-dup-ld64.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001951 "src/f32-igemm/gen/1x8s4-minmax-neonfma.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001952 "src/f32-igemm/gen/4x8-minmax-neonfma-dup-ld64.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001953 "src/f32-igemm/gen/4x8-minmax-neonfma-dup-ld128.c",
1954 "src/f32-igemm/gen/4x8s4-minmax-neonfma.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001955 "src/f32-igemm/gen/6x8-minmax-neonfma-dup-ld64.c",
1956 "src/f32-igemm/gen/6x8-minmax-neonfma-dup-ld128.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001957 "src/f32-igemm/gen/6x8s4-minmax-neonfma.c",
1958 "src/f32-igemm/gen/8x8s4-minmax-neonfma.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001959 "src/f32-ppmm/gen/4x8-minmax-neonfma.c",
1960 "src/f32-ppmm/gen/8x8-minmax-neonfma.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001961 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x4.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001962 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x8-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001963 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x8.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001964 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12-acc2.c",
1965 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001966 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001967 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16-acc2.c",
1968 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001969 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001970 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20-acc2.c",
1971 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001972 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20.c",
1973 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x4.c",
1974 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x8-acc2.c",
1975 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x8.c",
1976 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12-acc2.c",
1977 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12-acc3.c",
1978 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12.c",
1979 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16-acc2.c",
1980 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16-acc4.c",
1981 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16.c",
1982 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20-acc2.c",
1983 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20-acc5.c",
1984 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20.c",
Marat Dukhan2fa7a0c2020-12-06 19:09:02 -08001985 "src/f32-spmm/gen/4x1-minmax-neonfma-pipelined.c",
1986 "src/f32-spmm/gen/4x1-minmax-neonfma-x2.c",
1987 "src/f32-spmm/gen/4x1-minmax-neonfma.c",
1988 "src/f32-spmm/gen/8x1-minmax-neonfma-pipelined.c",
1989 "src/f32-spmm/gen/8x1-minmax-neonfma-x2.c",
1990 "src/f32-spmm/gen/8x1-minmax-neonfma.c",
1991 "src/f32-spmm/gen/12x1-minmax-neonfma.c",
1992 "src/f32-spmm/gen/16x1-minmax-neonfma-pipelined.c",
1993 "src/f32-spmm/gen/16x1-minmax-neonfma-x2.c",
1994 "src/f32-spmm/gen/16x1-minmax-neonfma.c",
1995 "src/f32-spmm/gen/32x1-minmax-neonfma-pipelined.c",
1996 "src/f32-spmm/gen/32x1-minmax-neonfma-x2.c",
1997 "src/f32-spmm/gen/32x1-minmax-neonfma.c",
Marat Dukhaned6baaf2020-12-01 15:07:08 -08001998 "src/f32-velu/gen/velu-neonfma-rr1-lut16-p3-x4.c",
1999 "src/f32-velu/gen/velu-neonfma-rr1-lut16-p3-x8.c",
2000 "src/f32-velu/gen/velu-neonfma-rr1-lut16-p3-x12.c",
2001 "src/f32-velu/gen/velu-neonfma-rr1-lut16-p3-x16.c",
2002 "src/f32-velu/gen/velu-neonfma-rr1-lut16-p3-x20.c",
2003 "src/f32-velu/gen/velu-neonfma-rr1-lut16-p3-x24.c",
2004 "src/f32-velu/gen/velu-neonfma-rr1-p6-x4.c",
2005 "src/f32-velu/gen/velu-neonfma-rr1-p6-x8.c",
2006 "src/f32-velu/gen/velu-neonfma-rr1-p6-x12.c",
2007 "src/f32-velu/gen/velu-neonfma-rr1-p6-x16.c",
2008 "src/f32-velu/gen/velu-neonfma-rr1-p6-x20.c",
2009 "src/f32-velu/gen/velu-neonfma-rr1-p6-x24.c",
Marat Dukhan99936602020-04-11 16:47:01 -07002010 "src/f32-vmulcaddc/gen/c4-minmax-neonfma-2x.c",
2011 "src/f32-vmulcaddc/gen/c8-minmax-neonfma-2x.c",
Marat Dukhanb1eec082021-05-05 23:24:55 -07002012 "src/f32-vsigmoid/gen/vsigmoid-neonfma-rr1-lut64-p2-nr1recps1fma-x4.c",
2013 "src/f32-vsigmoid/gen/vsigmoid-neonfma-rr1-lut64-p2-nr1recps1fma-x8.c",
2014 "src/f32-vsigmoid/gen/vsigmoid-neonfma-rr1-lut64-p2-nr1recps1fma-x12.c",
2015 "src/f32-vsigmoid/gen/vsigmoid-neonfma-rr1-lut64-p2-nr1recps1fma-x16.c",
2016 "src/f32-vsigmoid/gen/vsigmoid-neonfma-rr1-lut64-p2-nr1recps1fma-x20.c",
2017 "src/f32-vsigmoid/gen/vsigmoid-neonfma-rr1-lut64-p2-nr1recps1fma-x24.c",
2018 "src/f32-vsigmoid/gen/vsigmoid-neonfma-rr1-lut64-p2-nr2fma-x4.c",
2019 "src/f32-vsigmoid/gen/vsigmoid-neonfma-rr1-lut64-p2-nr2fma-x8.c",
2020 "src/f32-vsigmoid/gen/vsigmoid-neonfma-rr1-lut64-p2-nr2fma-x12.c",
2021 "src/f32-vsigmoid/gen/vsigmoid-neonfma-rr1-lut64-p2-nr2fma-x16.c",
2022 "src/f32-vsigmoid/gen/vsigmoid-neonfma-rr1-lut64-p2-nr2fma-x20.c",
2023 "src/f32-vsigmoid/gen/vsigmoid-neonfma-rr1-lut64-p2-nr2fma-x24.c",
2024 "src/f32-vsigmoid/gen/vsigmoid-neonfma-rr1-lut64-p2-nr2recps-x4.c",
2025 "src/f32-vsigmoid/gen/vsigmoid-neonfma-rr1-lut64-p2-nr2recps-x8.c",
2026 "src/f32-vsigmoid/gen/vsigmoid-neonfma-rr1-lut64-p2-nr2recps-x12.c",
2027 "src/f32-vsigmoid/gen/vsigmoid-neonfma-rr1-lut64-p2-nr2recps-x16.c",
2028 "src/f32-vsigmoid/gen/vsigmoid-neonfma-rr1-lut64-p2-nr2recps-x20.c",
2029 "src/f32-vsigmoid/gen/vsigmoid-neonfma-rr1-lut64-p2-nr2recps-x24.c",
2030 "src/f32-vsigmoid/gen/vsigmoid-neonfma-rr1-lut2048-p1-nr1recps1fma-x4.c",
2031 "src/f32-vsigmoid/gen/vsigmoid-neonfma-rr1-lut2048-p1-nr1recps1fma-x8.c",
2032 "src/f32-vsigmoid/gen/vsigmoid-neonfma-rr1-lut2048-p1-nr1recps1fma-x12.c",
2033 "src/f32-vsigmoid/gen/vsigmoid-neonfma-rr1-lut2048-p1-nr1recps1fma-x16.c",
2034 "src/f32-vsigmoid/gen/vsigmoid-neonfma-rr1-lut2048-p1-nr1recps1fma-x20.c",
2035 "src/f32-vsigmoid/gen/vsigmoid-neonfma-rr1-lut2048-p1-nr1recps1fma-x24.c",
2036 "src/f32-vsigmoid/gen/vsigmoid-neonfma-rr1-lut2048-p1-nr2fma-x4.c",
2037 "src/f32-vsigmoid/gen/vsigmoid-neonfma-rr1-lut2048-p1-nr2fma-x8.c",
2038 "src/f32-vsigmoid/gen/vsigmoid-neonfma-rr1-lut2048-p1-nr2fma-x12.c",
2039 "src/f32-vsigmoid/gen/vsigmoid-neonfma-rr1-lut2048-p1-nr2fma-x16.c",
2040 "src/f32-vsigmoid/gen/vsigmoid-neonfma-rr1-lut2048-p1-nr2fma-x20.c",
2041 "src/f32-vsigmoid/gen/vsigmoid-neonfma-rr1-lut2048-p1-nr2fma-x24.c",
2042 "src/f32-vsigmoid/gen/vsigmoid-neonfma-rr1-lut2048-p1-nr2recps-x4.c",
2043 "src/f32-vsigmoid/gen/vsigmoid-neonfma-rr1-lut2048-p1-nr2recps-x8.c",
2044 "src/f32-vsigmoid/gen/vsigmoid-neonfma-rr1-lut2048-p1-nr2recps-x12.c",
2045 "src/f32-vsigmoid/gen/vsigmoid-neonfma-rr1-lut2048-p1-nr2recps-x16.c",
2046 "src/f32-vsigmoid/gen/vsigmoid-neonfma-rr1-lut2048-p1-nr2recps-x20.c",
2047 "src/f32-vsigmoid/gen/vsigmoid-neonfma-rr1-lut2048-p1-nr2recps-x24.c",
2048 "src/f32-vsigmoid/gen/vsigmoid-neonfma-rr1-p5-nr1recps1fma-x4.c",
2049 "src/f32-vsigmoid/gen/vsigmoid-neonfma-rr1-p5-nr1recps1fma-x8.c",
2050 "src/f32-vsigmoid/gen/vsigmoid-neonfma-rr1-p5-nr1recps1fma-x12.c",
2051 "src/f32-vsigmoid/gen/vsigmoid-neonfma-rr1-p5-nr1recps1fma-x16.c",
2052 "src/f32-vsigmoid/gen/vsigmoid-neonfma-rr1-p5-nr1recps1fma-x20.c",
2053 "src/f32-vsigmoid/gen/vsigmoid-neonfma-rr1-p5-nr1recps1fma-x24.c",
2054 "src/f32-vsigmoid/gen/vsigmoid-neonfma-rr1-p5-nr2fma-x4.c",
2055 "src/f32-vsigmoid/gen/vsigmoid-neonfma-rr1-p5-nr2fma-x8.c",
2056 "src/f32-vsigmoid/gen/vsigmoid-neonfma-rr1-p5-nr2fma-x12.c",
2057 "src/f32-vsigmoid/gen/vsigmoid-neonfma-rr1-p5-nr2fma-x16.c",
2058 "src/f32-vsigmoid/gen/vsigmoid-neonfma-rr1-p5-nr2fma-x20.c",
2059 "src/f32-vsigmoid/gen/vsigmoid-neonfma-rr1-p5-nr2fma-x24.c",
2060 "src/f32-vsigmoid/gen/vsigmoid-neonfma-rr1-p5-nr2recps-x4.c",
2061 "src/f32-vsigmoid/gen/vsigmoid-neonfma-rr1-p5-nr2recps-x8.c",
2062 "src/f32-vsigmoid/gen/vsigmoid-neonfma-rr1-p5-nr2recps-x12.c",
2063 "src/f32-vsigmoid/gen/vsigmoid-neonfma-rr1-p5-nr2recps-x16.c",
2064 "src/f32-vsigmoid/gen/vsigmoid-neonfma-rr1-p5-nr2recps-x20.c",
2065 "src/f32-vsigmoid/gen/vsigmoid-neonfma-rr1-p5-nr2recps-x24.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07002066 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x4.c",
2067 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x8.c",
2068 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x12.c",
2069 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x16.c",
2070 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x20.c",
2071 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x24.c",
2072 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x28.c",
2073 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x32.c",
2074 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x36.c",
2075 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x40.c",
2076 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x4.c",
2077 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x8.c",
2078 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x12.c",
2079 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x16.c",
2080 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x20.c",
2081 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x24.c",
2082 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x28.c",
2083 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x32.c",
2084 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x36.c",
2085 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x40.c",
Marat Dukhanb7633f22020-11-20 16:34:56 -08002086 "src/math/exp-neonfma-rr2-lut64-p2.c",
2087 "src/math/exp-neonfma-rr2-p5.c",
Frank Barcharde7223ee2020-12-04 19:04:01 -08002088 "src/math/expm1minus-neonfma-rr1-lut16-p3.c",
2089 "src/math/expm1minus-neonfma-rr1-p6.c",
Marat Dukhan9dd119a2020-11-20 18:20:04 -08002090 "src/math/expminus-neonfma-rr2-lut64-p2.c",
2091 "src/math/expminus-neonfma-rr2-lut2048-p1.c",
2092 "src/math/expminus-neonfma-rr2-p5.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08002093 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr1recps1fma.c",
2094 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr2fma.c",
2095 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr2recps.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002096 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr1recps1fma.c",
2097 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr2fma.c",
2098 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr2recps.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08002099 "src/math/sigmoid-neonfma-rr1-p5-nr1recps1fma.c",
2100 "src/math/sigmoid-neonfma-rr1-p5-nr2fma.c",
2101 "src/math/sigmoid-neonfma-rr1-p5-nr2recps.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08002102 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr1recps1fma.c",
2103 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr2fma.c",
2104 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr2recps.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002105 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr1recps1fma.c",
2106 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr2fma.c",
2107 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr2recps.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08002108 "src/math/sigmoid-neonfma-rr2-p5-nr1recps1fma.c",
2109 "src/math/sigmoid-neonfma-rr2-p5-nr2fma.c",
2110 "src/math/sigmoid-neonfma-rr2-p5-nr2recps.c",
Marat Dukhan84000762020-06-29 18:38:43 -07002111 "src/math/sqrt-neonfma-nr1fma.c",
Marat Dukhan84000762020-06-29 18:38:43 -07002112 "src/math/sqrt-neonfma-nr1rsqrts1fma1adj.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002113 "src/math/sqrt-neonfma-nr2fma.c",
2114 "src/math/sqrt-neonfma-nr2fma1adj.c",
2115 "src/math/sqrt-neonfma-nr3fma.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002116]
2117
2118AARCH64_NEONFMA_UKERNELS = [
Marat Dukhan56b10cd2020-05-18 09:35:49 -07002119 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neonfma-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07002120 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neonfma-2x2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002121 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neonfma-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07002122 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neonfma-2x2.c",
Marat Dukhan56b10cd2020-05-18 09:35:49 -07002123 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neonfma-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07002124 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neonfma-2x2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002125 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neonfma-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07002126 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neonfma-2x2.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -07002127 "src/f32-conv-hwc2chw/3x3s2p1c3x4-neonfma-2x2.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07002128 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-1x4-acc2.c",
2129 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-1x4-acc3.c",
2130 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-1x4-acc4.c",
Marat Dukhan1268a242020-10-24 00:36:32 -07002131 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-1x4.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07002132 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-2x4-acc2.c",
Marat Dukhan1268a242020-10-24 00:36:32 -07002133 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-2x4.c",
2134 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-3x4.c",
2135 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-4x4.c",
2136 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-5x4.c",
2137 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-6x4.c",
Marat Dukhan82f0c322020-10-25 19:17:35 -07002138 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neonfma-1x4-acc2.c",
2139 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neonfma-1x4-acc3.c",
2140 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neonfma-1x4-acc4.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07002141 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neonfma-1x4.c",
Marat Dukhan82f0c322020-10-25 19:17:35 -07002142 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neonfma-2x4-acc2.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07002143 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neonfma-2x4.c",
2144 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neonfma-3x4.c",
2145 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neonfma-4x4.c",
Marat Dukhan149f0ea2020-10-26 12:50:33 -07002146 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-1x4-acc2.c",
2147 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-1x4-acc3.c",
2148 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-1x4-acc4.c",
2149 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-1x4-acc5.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07002150 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-1x4.c",
Marat Dukhan149f0ea2020-10-26 12:50:33 -07002151 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-2x4-acc2.c",
2152 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-2x4-acc3.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07002153 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-2x4.c",
Marat Dukhan149f0ea2020-10-26 12:50:33 -07002154 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-3x4-acc2.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07002155 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-3x4.c",
Marat Dukhan149f0ea2020-10-26 12:50:33 -07002156 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-4x4-acc2.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07002157 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-4x4.c",
2158 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-5x4.c",
Marat Dukhan30d4b252020-10-29 16:33:22 -07002159 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neonfma-1x4-acc2.c",
2160 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neonfma-1x4-acc3.c",
2161 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neonfma-1x4-acc4.c",
2162 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neonfma-1x4-acc5.c",
2163 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neonfma-1x4.c",
2164 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neonfma-2x4-acc2.c",
2165 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neonfma-2x4-acc3.c",
2166 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neonfma-2x4.c",
Marat Dukhan30d4b252020-10-29 16:33:22 -07002167 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neonfma-3x4-acc2.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07002168 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neonfma-3x4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002169 "src/f32-gemm/gen-inc/1x8inc-minmax-neonfma-lane-ld64.c",
2170 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-lane-ld64.c",
2171 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-lane-ld128.c",
2172 "src/f32-gemm/gen-inc/5x8inc-minmax-neonfma-lane-ld64.c",
2173 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-lane-ld64.c",
2174 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-lane-ld128.c",
2175 "src/f32-gemm/gen/1x8-minmax-neonfma-lane-ld64.c",
2176 "src/f32-gemm/gen/4x2-minmax-neonfma-lane-ld64.c",
2177 "src/f32-gemm/gen/4x8-minmax-neonfma-lane-ld64.c",
2178 "src/f32-gemm/gen/4x8-minmax-neonfma-lane-ld128.c",
2179 "src/f32-gemm/gen/5x8-minmax-neonfma-lane-ld64.c",
2180 "src/f32-gemm/gen/6x8-minmax-neonfma-lane-ld64.c",
2181 "src/f32-gemm/gen/6x8-minmax-neonfma-lane-ld128.c",
2182 "src/f32-igemm/gen/1x8-minmax-neonfma-lane-ld64.c",
2183 "src/f32-igemm/gen/4x2-minmax-neonfma-lane-ld64.c",
2184 "src/f32-igemm/gen/4x4-minmax-neonfma-lane-ld64.c",
2185 "src/f32-igemm/gen/4x8-minmax-neonfma-lane-ld64.c",
2186 "src/f32-igemm/gen/4x8-minmax-neonfma-lane-ld128.c",
2187 "src/f32-igemm/gen/6x8-minmax-neonfma-lane-ld64.c",
2188 "src/f32-igemm/gen/6x8-minmax-neonfma-lane-ld128.c",
Marat Dukhan355ab432020-04-09 19:01:52 -07002189 "src/f32-spmm/gen/4x2-minmax-neonfma.c",
2190 "src/f32-spmm/gen/4x4-minmax-neonfma.c",
Marat Dukhan355ab432020-04-09 19:01:52 -07002191 "src/f32-spmm/gen/8x2-minmax-neonfma.c",
2192 "src/f32-spmm/gen/8x4-minmax-neonfma.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002193 "src/f32-spmm/gen/12x2-minmax-neonfma.c",
2194 "src/f32-spmm/gen/12x4-minmax-neonfma.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002195 "src/f32-spmm/gen/16x2-minmax-neonfma.c",
2196 "src/f32-spmm/gen/16x4-minmax-neonfma.c",
Frank Barchard846c0c62020-10-26 15:01:39 -07002197 "src/f32-spmm/gen/32x2-minmax-neonfma.c",
2198 "src/f32-spmm/gen/32x4-minmax-neonfma.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002199 "src/f32-vbinary/gen/vdiv-minmax-neon-x4.c",
2200 "src/f32-vbinary/gen/vdiv-minmax-neon-x8.c",
2201 "src/f32-vbinary/gen/vdivc-minmax-neon-x4.c",
2202 "src/f32-vbinary/gen/vdivc-minmax-neon-x8.c",
2203 "src/f32-vbinary/gen/vrdivc-minmax-neon-x4.c",
2204 "src/f32-vbinary/gen/vrdivc-minmax-neon-x8.c",
Marat Dukhanb1eec082021-05-05 23:24:55 -07002205 "src/f32-vsigmoid/gen/vsigmoid-neonfma-rr1-lut64-p2-div-x4.c",
2206 "src/f32-vsigmoid/gen/vsigmoid-neonfma-rr1-lut64-p2-div-x8.c",
2207 "src/f32-vsigmoid/gen/vsigmoid-neonfma-rr1-lut64-p2-div-x12.c",
2208 "src/f32-vsigmoid/gen/vsigmoid-neonfma-rr1-lut64-p2-div-x16.c",
2209 "src/f32-vsigmoid/gen/vsigmoid-neonfma-rr1-lut64-p2-div-x20.c",
2210 "src/f32-vsigmoid/gen/vsigmoid-neonfma-rr1-lut64-p2-div-x24.c",
2211 "src/f32-vsigmoid/gen/vsigmoid-neonfma-rr1-lut2048-p1-div-x4.c",
2212 "src/f32-vsigmoid/gen/vsigmoid-neonfma-rr1-lut2048-p1-div-x8.c",
2213 "src/f32-vsigmoid/gen/vsigmoid-neonfma-rr1-lut2048-p1-div-x12.c",
2214 "src/f32-vsigmoid/gen/vsigmoid-neonfma-rr1-lut2048-p1-div-x16.c",
2215 "src/f32-vsigmoid/gen/vsigmoid-neonfma-rr1-lut2048-p1-div-x20.c",
2216 "src/f32-vsigmoid/gen/vsigmoid-neonfma-rr1-lut2048-p1-div-x24.c",
2217 "src/f32-vsigmoid/gen/vsigmoid-neonfma-rr1-p5-div-x4.c",
2218 "src/f32-vsigmoid/gen/vsigmoid-neonfma-rr1-p5-div-x8.c",
2219 "src/f32-vsigmoid/gen/vsigmoid-neonfma-rr1-p5-div-x12.c",
2220 "src/f32-vsigmoid/gen/vsigmoid-neonfma-rr1-p5-div-x16.c",
2221 "src/f32-vsigmoid/gen/vsigmoid-neonfma-rr1-p5-div-x20.c",
2222 "src/f32-vsigmoid/gen/vsigmoid-neonfma-rr1-p5-div-x24.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07002223 "src/f32-vsqrt/gen/neon-sqrt-x4.c",
2224 "src/f32-vsqrt/gen/neon-sqrt-x8.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08002225 "src/math/sigmoid-neonfma-rr1-lut64-p2-div.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002226 "src/math/sigmoid-neonfma-rr1-lut2048-p1-div.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08002227 "src/math/sigmoid-neonfma-rr1-p5-div.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08002228 "src/math/sigmoid-neonfma-rr2-lut64-p2-div.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002229 "src/math/sigmoid-neonfma-rr2-lut2048-p1-div.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08002230 "src/math/sigmoid-neonfma-rr2-p5-div.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002231]
2232
Marat Dukhan8853b822020-05-07 12:19:01 -07002233NEONV8_UKERNELS = [
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07002234 "src/f32-vrnd/gen/vrndd-neonv8-x4.c",
2235 "src/f32-vrnd/gen/vrndd-neonv8-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002236 "src/f32-vrnd/gen/vrndne-neonv8-x4.c",
2237 "src/f32-vrnd/gen/vrndne-neonv8-x8.c",
2238 "src/f32-vrnd/gen/vrndu-neonv8-x4.c",
2239 "src/f32-vrnd/gen/vrndu-neonv8-x8.c",
2240 "src/f32-vrnd/gen/vrndz-neonv8-x4.c",
2241 "src/f32-vrnd/gen/vrndz-neonv8-x8.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07002242 "src/math/roundd-neonv8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002243 "src/math/roundne-neonv8.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07002244 "src/math/roundu-neonv8.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07002245 "src/math/roundz-neonv8.c",
Marat Dukhan8853b822020-05-07 12:19:01 -07002246]
2247
Marat Dukhan08c4a432019-10-03 09:29:21 -07002248AARCH64_NEONFP16ARITH_UKERNELS = [
Frank Barchard5a599a62020-06-04 20:12:44 -07002249 "src/f16-dwconv/gen/up8x4-minmax-neonfp16arith-acc2.c",
2250 "src/f16-dwconv/gen/up8x4-minmax-neonfp16arith.c",
2251 "src/f16-dwconv/gen/up8x9-minmax-neonfp16arith-acc2.c",
2252 "src/f16-dwconv/gen/up8x9-minmax-neonfp16arith.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002253 "src/f16-dwconv/gen/up8x25-minmax-neonfp16arith-acc2.c",
2254 "src/f16-dwconv/gen/up8x25-minmax-neonfp16arith.c",
2255 "src/f16-dwconv/gen/up16x4-minmax-neonfp16arith-acc2.c",
2256 "src/f16-dwconv/gen/up16x4-minmax-neonfp16arith.c",
2257 "src/f16-dwconv/gen/up16x9-minmax-neonfp16arith-acc2.c",
2258 "src/f16-dwconv/gen/up16x9-minmax-neonfp16arith.c",
2259 "src/f16-dwconv/gen/up16x25-minmax-neonfp16arith-acc2.c",
2260 "src/f16-dwconv/gen/up16x25-minmax-neonfp16arith.c",
Frank Barchard0bb49a72020-06-04 11:35:11 -07002261 "src/f16-gavgpool/7p7x-minmax-neonfp16arith-c8.c",
2262 "src/f16-gavgpool/7x-minmax-neonfp16arith-c8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002263 "src/f16-gemm/gen-inc/1x8inc-minmax-neonfp16arith-ld64.c",
2264 "src/f16-gemm/gen-inc/1x16inc-minmax-neonfp16arith-ld64.c",
2265 "src/f16-gemm/gen-inc/4x8inc-minmax-neonfp16arith-ld64.c",
2266 "src/f16-gemm/gen-inc/4x16inc-minmax-neonfp16arith-ld64.c",
2267 "src/f16-gemm/gen-inc/6x8inc-minmax-neonfp16arith-ld64.c",
2268 "src/f16-gemm/gen-inc/6x16inc-minmax-neonfp16arith-ld64.c",
2269 "src/f16-gemm/gen-inc/8x8inc-minmax-neonfp16arith-ld64.c",
2270 "src/f16-gemm/gen-inc/8x16inc-minmax-neonfp16arith-ld64.c",
2271 "src/f16-gemm/gen/1x8-minmax-neonfp16arith-ld64.c",
2272 "src/f16-gemm/gen/1x16-minmax-neonfp16arith-ld64.c",
2273 "src/f16-gemm/gen/4x8-minmax-neonfp16arith-ld64.c",
2274 "src/f16-gemm/gen/4x16-minmax-neonfp16arith-ld64.c",
2275 "src/f16-gemm/gen/6x8-minmax-neonfp16arith-ld64.c",
2276 "src/f16-gemm/gen/6x16-minmax-neonfp16arith-ld64.c",
2277 "src/f16-gemm/gen/8x8-minmax-neonfp16arith-ld64.c",
2278 "src/f16-gemm/gen/8x16-minmax-neonfp16arith-ld64.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002279 "src/f16-igemm/gen/1x8-minmax-neonfp16arith-ld64.c",
2280 "src/f16-igemm/gen/1x16-minmax-neonfp16arith-ld64.c",
2281 "src/f16-igemm/gen/4x8-minmax-neonfp16arith-ld64.c",
2282 "src/f16-igemm/gen/4x16-minmax-neonfp16arith-ld64.c",
2283 "src/f16-igemm/gen/6x8-minmax-neonfp16arith-ld64.c",
2284 "src/f16-igemm/gen/6x16-minmax-neonfp16arith-ld64.c",
2285 "src/f16-igemm/gen/8x8-minmax-neonfp16arith-ld64.c",
2286 "src/f16-igemm/gen/8x16-minmax-neonfp16arith-ld64.c",
Frank Barchardb1966592020-05-12 13:47:06 -07002287 "src/f16-prelu/gen/neonfp16arith-2x8.c",
Frank Barcharda5316982020-07-23 13:19:28 -07002288 "src/f16-prelu/gen/neonfp16arith-2x16.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07002289 "src/f16-spmm/gen/8x1-minmax-neonfp16arith-x2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002290 "src/f16-spmm/gen/8x1-minmax-neonfp16arith.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07002291 "src/f16-spmm/gen/16x1-minmax-neonfp16arith-x2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002292 "src/f16-spmm/gen/16x1-minmax-neonfp16arith.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07002293 "src/f16-spmm/gen/24x1-minmax-neonfp16arith-x2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002294 "src/f16-spmm/gen/24x1-minmax-neonfp16arith.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07002295 "src/f16-spmm/gen/32x1-minmax-neonfp16arith-x2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002296 "src/f16-spmm/gen/32x1-minmax-neonfp16arith.c",
2297 "src/f16-vbinary/gen/vadd-minmax-neonfp16arith-x8.c",
2298 "src/f16-vbinary/gen/vadd-minmax-neonfp16arith-x16.c",
2299 "src/f16-vbinary/gen/vaddc-minmax-neonfp16arith-x8.c",
2300 "src/f16-vbinary/gen/vaddc-minmax-neonfp16arith-x16.c",
2301 "src/f16-vbinary/gen/vdiv-minmax-neonfp16arith-x8.c",
2302 "src/f16-vbinary/gen/vdiv-minmax-neonfp16arith-x16.c",
2303 "src/f16-vbinary/gen/vdivc-minmax-neonfp16arith-x8.c",
2304 "src/f16-vbinary/gen/vdivc-minmax-neonfp16arith-x16.c",
2305 "src/f16-vbinary/gen/vmax-neonfp16arith-x8.c",
2306 "src/f16-vbinary/gen/vmax-neonfp16arith-x16.c",
2307 "src/f16-vbinary/gen/vmaxc-neonfp16arith-x8.c",
2308 "src/f16-vbinary/gen/vmaxc-neonfp16arith-x16.c",
2309 "src/f16-vbinary/gen/vmin-neonfp16arith-x8.c",
2310 "src/f16-vbinary/gen/vmin-neonfp16arith-x16.c",
2311 "src/f16-vbinary/gen/vminc-neonfp16arith-x8.c",
2312 "src/f16-vbinary/gen/vminc-neonfp16arith-x16.c",
2313 "src/f16-vbinary/gen/vmul-minmax-neonfp16arith-x8.c",
2314 "src/f16-vbinary/gen/vmul-minmax-neonfp16arith-x16.c",
2315 "src/f16-vbinary/gen/vmulc-minmax-neonfp16arith-x8.c",
2316 "src/f16-vbinary/gen/vmulc-minmax-neonfp16arith-x16.c",
2317 "src/f16-vbinary/gen/vrdivc-minmax-neonfp16arith-x8.c",
2318 "src/f16-vbinary/gen/vrdivc-minmax-neonfp16arith-x16.c",
2319 "src/f16-vbinary/gen/vrsubc-minmax-neonfp16arith-x8.c",
2320 "src/f16-vbinary/gen/vrsubc-minmax-neonfp16arith-x16.c",
2321 "src/f16-vbinary/gen/vsub-minmax-neonfp16arith-x8.c",
2322 "src/f16-vbinary/gen/vsub-minmax-neonfp16arith-x16.c",
2323 "src/f16-vbinary/gen/vsubc-minmax-neonfp16arith-x8.c",
2324 "src/f16-vbinary/gen/vsubc-minmax-neonfp16arith-x16.c",
Marat Dukhanb1eec082021-05-05 23:24:55 -07002325 "src/f16-vclamp/gen/vclamp-neonfp16arith-x8.c",
2326 "src/f16-vclamp/gen/vclamp-neonfp16arith-x16.c",
Marat Dukhan6674d692021-05-05 22:27:00 -07002327 "src/f16-vhswish/gen/vhswish-neonfp16arith-x8.c",
2328 "src/f16-vhswish/gen/vhswish-neonfp16arith-x16.c",
Marat Dukhanb1eec082021-05-05 23:24:55 -07002329 "src/f16-vrelu/gen/vrelu-neonfp16arith-x8.c",
2330 "src/f16-vrelu/gen/vrelu-neonfp16arith-x16.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002331 "src/f16-vmulcaddc/gen/c8-minmax-neonfp16arith-2x.c",
2332 "src/f16-vmulcaddc/gen/c16-minmax-neonfp16arith-2x.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002333]
2334
Benoit Jacoba9644732020-08-13 12:48:55 -07002335NEONDOT_UKERNELS = [
Marat Dukhand4c83032020-09-09 23:20:12 -07002336 "src/qs8-gemm/gen/1x8c4-minmax-neondot.c",
Marat Dukhand4c83032020-09-09 23:20:12 -07002337 "src/qs8-gemm/gen/1x16c4-minmax-neondot.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002338 "src/qs8-gemm/gen/4x8c4-minmax-neondot.c",
Benoit Jacoba9644732020-08-13 12:48:55 -07002339 "src/qs8-gemm/gen/4x16c4-minmax-neondot.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002340 "src/qs8-gemm/gen/6x8c4-minmax-neondot.c",
Frank Barchard2fa17452020-09-22 09:51:46 -07002341 "src/qs8-gemm/gen/6x16c4-minmax-neondot.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002342 "src/qs8-gemm/gen/8x8c4-minmax-neondot.c",
Frank Barcharda48848f2020-09-24 00:59:24 -07002343 "src/qs8-gemm/gen/8x16c4-minmax-neondot.c",
Marat Dukhand4c83032020-09-09 23:20:12 -07002344 "src/qs8-igemm/gen/1x8c4-minmax-neondot.c",
Marat Dukhand4c83032020-09-09 23:20:12 -07002345 "src/qs8-igemm/gen/1x16c4-minmax-neondot.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002346 "src/qs8-igemm/gen/4x8c4-minmax-neondot.c",
Benoit Jacoba05487f2020-08-14 13:37:21 -07002347 "src/qs8-igemm/gen/4x16c4-minmax-neondot.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002348 "src/qs8-igemm/gen/6x8c4-minmax-neondot.c",
Frank Barchard2fa17452020-09-22 09:51:46 -07002349 "src/qs8-igemm/gen/6x16c4-minmax-neondot.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002350 "src/qs8-igemm/gen/8x8c4-minmax-neondot.c",
Frank Barcharda48848f2020-09-24 00:59:24 -07002351 "src/qs8-igemm/gen/8x16c4-minmax-neondot.c",
Benoit Jacoba9644732020-08-13 12:48:55 -07002352]
2353
Marat Dukhan08c4a432019-10-03 09:29:21 -07002354SSE_UKERNELS = [
Marat Dukhan99936602020-04-11 16:47:01 -07002355 "src/f32-avgpool/9p8x-minmax-sse-c4.c",
2356 "src/f32-avgpool/9x-minmax-sse-c4.c",
Erich Elsenb1233402020-06-08 15:53:15 -07002357 "src/f32-conv-hwc2chw/3x3s2p1c3x4-sse-1x1.c",
2358 "src/f32-conv-hwc2chw/3x3s2p1c3x4-sse-2x2.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002359 "src/f32-dwconv/gen/up4x4-minmax-sse-acc2.c",
2360 "src/f32-dwconv/gen/up4x4-minmax-sse.c",
2361 "src/f32-dwconv/gen/up4x9-minmax-sse-acc2.c",
2362 "src/f32-dwconv/gen/up4x9-minmax-sse.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002363 "src/f32-dwconv/gen/up4x25-minmax-sse-acc2.c",
2364 "src/f32-dwconv/gen/up4x25-minmax-sse.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002365 "src/f32-dwconv/gen/up8x4-minmax-sse-acc2.c",
2366 "src/f32-dwconv/gen/up8x4-minmax-sse.c",
2367 "src/f32-dwconv/gen/up8x9-minmax-sse-acc2.c",
2368 "src/f32-dwconv/gen/up8x9-minmax-sse.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002369 "src/f32-dwconv/gen/up8x25-minmax-sse-acc2.c",
2370 "src/f32-dwconv/gen/up8x25-minmax-sse.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07002371 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-1x4-acc2.c",
2372 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-1x4-acc3.c",
2373 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-1x4-acc4.c",
Marat Dukhan470078a2020-10-23 22:36:52 -07002374 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-1x4.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07002375 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-2x4-acc2.c",
Marat Dukhan470078a2020-10-23 22:36:52 -07002376 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-2x4.c",
2377 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-3x4.c",
2378 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-4x4.c",
2379 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-5x4.c",
2380 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-6x4.c",
Marat Dukhan0ff97182020-10-25 19:14:03 -07002381 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-sse-1x4-acc2.c",
2382 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-sse-1x4-acc3.c",
2383 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-sse-1x4-acc4.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07002384 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-sse-1x4.c",
Marat Dukhan0ff97182020-10-25 19:14:03 -07002385 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-sse-2x4-acc2.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07002386 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-sse-2x4.c",
2387 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-sse-3x4.c",
2388 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-sse-4x4.c",
Marat Dukhand0503892020-10-30 08:22:04 -07002389 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-1x4-acc2.c",
2390 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-1x4-acc3.c",
2391 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-1x4-acc4.c",
2392 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-1x4-acc5.c",
2393 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-1x4.c",
2394 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-2x4-acc2.c",
2395 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-2x4-acc3.c",
2396 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-2x4.c",
2397 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-3x4-acc2.c",
2398 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-3x4.c",
2399 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-4x4-acc2.c",
2400 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-4x4.c",
2401 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-5x4.c",
Marat Dukhanccca2142020-10-30 17:32:45 -07002402 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-sse-1x4-acc2.c",
2403 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-sse-1x4-acc3.c",
2404 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-sse-1x4-acc4.c",
2405 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-sse-1x4-acc5.c",
2406 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-sse-1x4.c",
2407 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-sse-2x4-acc2.c",
2408 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-sse-2x4-acc3.c",
2409 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-sse-2x4.c",
Marat Dukhanccca2142020-10-30 17:32:45 -07002410 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-sse-3x4-acc2.c",
Frank Barchard8ef44cd2020-11-03 12:30:23 -08002411 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-sse-3x4.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -07002412 "src/f32-gavgpool-cw/sse-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -07002413 "src/f32-gavgpool/7p7x-minmax-sse-c4.c",
2414 "src/f32-gavgpool/7x-minmax-sse-c4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002415 "src/f32-gemm/gen-inc/1x8inc-minmax-sse-dup.c",
2416 "src/f32-gemm/gen-inc/1x8inc-minmax-sse-load1.c",
2417 "src/f32-gemm/gen-inc/1x8s4inc-minmax-sse.c",
Marat Dukhan802fcae2020-12-11 14:37:25 -08002418 "src/f32-gemm/gen-inc/3x8inc-minmax-sse-dup.c",
2419 "src/f32-gemm/gen-inc/3x8inc-minmax-sse-load1.c",
2420 "src/f32-gemm/gen-inc/3x8s4inc-minmax-sse.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002421 "src/f32-gemm/gen-inc/4x8inc-minmax-sse-dup.c",
2422 "src/f32-gemm/gen-inc/4x8inc-minmax-sse-load1.c",
2423 "src/f32-gemm/gen-inc/4x8s4inc-minmax-sse.c",
Marat Dukhan802fcae2020-12-11 14:37:25 -08002424 "src/f32-gemm/gen-inc/5x8inc-minmax-sse-dup.c",
2425 "src/f32-gemm/gen-inc/5x8inc-minmax-sse-load1.c",
2426 "src/f32-gemm/gen-inc/5x8s4inc-minmax-sse.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002427 "src/f32-gemm/gen/1x8-minmax-sse-dup.c",
2428 "src/f32-gemm/gen/1x8-minmax-sse-load1.c",
2429 "src/f32-gemm/gen/1x8s4-minmax-sse.c",
Marat Dukhan802fcae2020-12-11 14:37:25 -08002430 "src/f32-gemm/gen/3x8-minmax-sse-dup.c",
2431 "src/f32-gemm/gen/3x8-minmax-sse-load1.c",
2432 "src/f32-gemm/gen/3x8s4-minmax-sse.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002433 "src/f32-gemm/gen/4x2c4-minmax-sse.c",
2434 "src/f32-gemm/gen/4x8-minmax-sse-dup.c",
2435 "src/f32-gemm/gen/4x8-minmax-sse-load1.c",
2436 "src/f32-gemm/gen/4x8s4-minmax-sse.c",
Marat Dukhan802fcae2020-12-11 14:37:25 -08002437 "src/f32-gemm/gen/5x8-minmax-sse-dup.c",
2438 "src/f32-gemm/gen/5x8-minmax-sse-load1.c",
2439 "src/f32-gemm/gen/5x8s4-minmax-sse.c",
Artsiom Ablavatskib3ffd582021-03-31 13:00:08 -07002440 "src/f32-ibilinear-chw/gen/sse-p4.c",
2441 "src/f32-ibilinear-chw/gen/sse-p8.c",
Frank Barchard4a352042021-04-13 15:52:08 -07002442 "src/f32-ibilinear/gen/sse-c4.c",
2443 "src/f32-ibilinear/gen/sse-c8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002444 "src/f32-igemm/gen/1x8-minmax-sse-dup.c",
2445 "src/f32-igemm/gen/1x8-minmax-sse-load1.c",
2446 "src/f32-igemm/gen/1x8s4-minmax-sse.c",
Marat Dukhan802fcae2020-12-11 14:37:25 -08002447 "src/f32-igemm/gen/3x8-minmax-sse-dup.c",
2448 "src/f32-igemm/gen/3x8-minmax-sse-load1.c",
2449 "src/f32-igemm/gen/3x8s4-minmax-sse.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002450 "src/f32-igemm/gen/4x2c4-minmax-sse.c",
2451 "src/f32-igemm/gen/4x8-minmax-sse-dup.c",
2452 "src/f32-igemm/gen/4x8-minmax-sse-load1.c",
2453 "src/f32-igemm/gen/4x8s4-minmax-sse.c",
Marat Dukhan802fcae2020-12-11 14:37:25 -08002454 "src/f32-igemm/gen/5x8-minmax-sse-dup.c",
2455 "src/f32-igemm/gen/5x8-minmax-sse-load1.c",
2456 "src/f32-igemm/gen/5x8s4-minmax-sse.c",
Marat Dukhan99936602020-04-11 16:47:01 -07002457 "src/f32-maxpool/9p8x-minmax-sse-c4.c",
2458 "src/f32-pavgpool/9p8x-minmax-sse-c4.c",
2459 "src/f32-pavgpool/9x-minmax-sse-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002460 "src/f32-ppmm/gen/4x8-minmax-sse.c",
Marat Dukhan39b5e942020-06-24 15:03:48 -07002461 "src/f32-prelu/gen/sse-2x4.c",
2462 "src/f32-prelu/gen/sse-2x8.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002463 "src/f32-rmax/sse.c",
Marat Dukhan355ab432020-04-09 19:01:52 -07002464 "src/f32-spmm/gen/4x1-minmax-sse.c",
2465 "src/f32-spmm/gen/8x1-minmax-sse.c",
Erich Elsen6e80fdc2020-06-09 15:35:37 -07002466 "src/f32-spmm/gen/16x1-minmax-sse.c",
Frank Barchard846c0c62020-10-26 15:01:39 -07002467 "src/f32-spmm/gen/32x1-minmax-sse.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002468 "src/f32-vbinary/gen/vadd-minmax-sse-x4.c",
2469 "src/f32-vbinary/gen/vadd-minmax-sse-x8.c",
2470 "src/f32-vbinary/gen/vaddc-minmax-sse-x4.c",
2471 "src/f32-vbinary/gen/vaddc-minmax-sse-x8.c",
2472 "src/f32-vbinary/gen/vdiv-minmax-sse-x4.c",
2473 "src/f32-vbinary/gen/vdiv-minmax-sse-x8.c",
2474 "src/f32-vbinary/gen/vdivc-minmax-sse-x4.c",
2475 "src/f32-vbinary/gen/vdivc-minmax-sse-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -08002476 "src/f32-vbinary/gen/vmax-sse-x4.c",
2477 "src/f32-vbinary/gen/vmax-sse-x8.c",
2478 "src/f32-vbinary/gen/vmaxc-sse-x4.c",
2479 "src/f32-vbinary/gen/vmaxc-sse-x8.c",
2480 "src/f32-vbinary/gen/vmin-sse-x4.c",
2481 "src/f32-vbinary/gen/vmin-sse-x8.c",
2482 "src/f32-vbinary/gen/vminc-sse-x4.c",
2483 "src/f32-vbinary/gen/vminc-sse-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002484 "src/f32-vbinary/gen/vmul-minmax-sse-x4.c",
2485 "src/f32-vbinary/gen/vmul-minmax-sse-x8.c",
2486 "src/f32-vbinary/gen/vmulc-minmax-sse-x4.c",
2487 "src/f32-vbinary/gen/vmulc-minmax-sse-x8.c",
2488 "src/f32-vbinary/gen/vrdivc-minmax-sse-x4.c",
2489 "src/f32-vbinary/gen/vrdivc-minmax-sse-x8.c",
2490 "src/f32-vbinary/gen/vrsubc-minmax-sse-x4.c",
2491 "src/f32-vbinary/gen/vrsubc-minmax-sse-x8.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07002492 "src/f32-vbinary/gen/vsqrdiff-sse-x4.c",
2493 "src/f32-vbinary/gen/vsqrdiff-sse-x8.c",
2494 "src/f32-vbinary/gen/vsqrdiffc-sse-x4.c",
2495 "src/f32-vbinary/gen/vsqrdiffc-sse-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002496 "src/f32-vbinary/gen/vsub-minmax-sse-x4.c",
2497 "src/f32-vbinary/gen/vsub-minmax-sse-x8.c",
2498 "src/f32-vbinary/gen/vsubc-minmax-sse-x4.c",
2499 "src/f32-vbinary/gen/vsubc-minmax-sse-x8.c",
Marat Dukhanb1eec082021-05-05 23:24:55 -07002500 "src/f32-vclamp/gen/vclamp-sse-x4.c",
2501 "src/f32-vclamp/gen/vclamp-sse-x8.c",
Marat Dukhan6674d692021-05-05 22:27:00 -07002502 "src/f32-vhswish/gen/vhswish-sse-x4.c",
2503 "src/f32-vhswish/gen/vhswish-sse-x8.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07002504 "src/f32-vlrelu/gen/vlrelu-sse-x4.c",
2505 "src/f32-vlrelu/gen/vlrelu-sse-x8.c",
Marat Dukhan99936602020-04-11 16:47:01 -07002506 "src/f32-vmulcaddc/gen/c4-minmax-sse-2x.c",
2507 "src/f32-vmulcaddc/gen/c8-minmax-sse-2x.c",
Marat Dukhanb1eec082021-05-05 23:24:55 -07002508 "src/f32-vrelu/gen/vrelu-sse-x4.c",
2509 "src/f32-vrelu/gen/vrelu-sse-x8.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07002510 "src/f32-vsqrt/gen/sse-sqrt-x4.c",
2511 "src/f32-vsqrt/gen/sse-sqrt-x8.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07002512 "src/f32-vunary/gen/vabs-sse-x4.c",
2513 "src/f32-vunary/gen/vabs-sse-x8.c",
2514 "src/f32-vunary/gen/vneg-sse-x4.c",
2515 "src/f32-vunary/gen/vneg-sse-x8.c",
2516 "src/f32-vunary/gen/vsqr-sse-x4.c",
2517 "src/f32-vunary/gen/vsqr-sse-x8.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07002518 "src/math/roundd-sse-addsub.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002519 "src/math/roundne-sse-addsub.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07002520 "src/math/roundu-sse-addsub.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07002521 "src/math/roundz-sse-addsub.c",
Marat Dukhan84000762020-06-29 18:38:43 -07002522 "src/math/sqrt-sse-hh1mac.c",
2523 "src/math/sqrt-sse-nr1mac.c",
2524 "src/math/sqrt-sse-nr2mac.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002525 "src/x32-fill/sse.c",
2526 "src/x32-packx/x4-sse.c",
2527 "src/x32-pad/sse.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002528]
2529
2530SSE2_UKERNELS = [
Marat Dukhan329da642019-11-19 21:44:39 -08002531 "src/f32-argmaxpool/4x-sse2-c4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002532 "src/f32-argmaxpool/9p8x-sse2-c4.c",
Marat Dukhan329da642019-11-19 21:44:39 -08002533 "src/f32-argmaxpool/9x-sse2-c4.c",
Marat Dukhan802fcae2020-12-11 14:37:25 -08002534 "src/f32-gemm/gen-inc/1x8inc-minmax-sse2-dup.c",
2535 "src/f32-gemm/gen-inc/3x8inc-minmax-sse2-dup.c",
2536 "src/f32-gemm/gen-inc/4x8inc-minmax-sse2-dup.c",
2537 "src/f32-gemm/gen-inc/5x8inc-minmax-sse2-dup.c",
2538 "src/f32-gemm/gen/1x8-minmax-sse2-dup.c",
2539 "src/f32-gemm/gen/3x8-minmax-sse2-dup.c",
2540 "src/f32-gemm/gen/4x8-minmax-sse2-dup.c",
2541 "src/f32-gemm/gen/5x8-minmax-sse2-dup.c",
2542 "src/f32-igemm/gen/1x8-minmax-sse2-dup.c",
2543 "src/f32-igemm/gen/3x8-minmax-sse2-dup.c",
2544 "src/f32-igemm/gen/4x8-minmax-sse2-dup.c",
2545 "src/f32-igemm/gen/5x8-minmax-sse2-dup.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08002546 "src/f32-prelu/gen/sse2-2x4.c",
2547 "src/f32-prelu/gen/sse2-2x8.c",
Marat Dukhanb39689d2020-01-24 13:32:20 -08002548 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x4.c",
Marat Dukhanb39689d2020-01-24 13:32:20 -08002549 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x8-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002550 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x8.c",
Marat Dukhanb39689d2020-01-24 13:32:20 -08002551 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12-acc2.c",
2552 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002553 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12.c",
Marat Dukhanb39689d2020-01-24 13:32:20 -08002554 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16-acc2.c",
2555 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002556 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16.c",
Marat Dukhanb39689d2020-01-24 13:32:20 -08002557 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20-acc2.c",
2558 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002559 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20.c",
Marat Dukhaned6baaf2020-12-01 15:07:08 -08002560 "src/f32-velu/gen/velu-sse2-rr2-lut16-p3-x4.c",
2561 "src/f32-velu/gen/velu-sse2-rr2-lut16-p3-x8.c",
2562 "src/f32-velu/gen/velu-sse2-rr2-lut16-p3-x12.c",
2563 "src/f32-velu/gen/velu-sse2-rr2-lut16-p3-x16.c",
2564 "src/f32-velu/gen/velu-sse2-rr2-lut16-p3-x20.c",
2565 "src/f32-velu/gen/velu-sse2-rr2-lut16-p3-x24.c",
2566 "src/f32-velu/gen/velu-sse2-rr2-p6-x4.c",
2567 "src/f32-velu/gen/velu-sse2-rr2-p6-x8.c",
2568 "src/f32-velu/gen/velu-sse2-rr2-p6-x12.c",
2569 "src/f32-velu/gen/velu-sse2-rr2-p6-x16.c",
2570 "src/f32-velu/gen/velu-sse2-rr2-p6-x20.c",
2571 "src/f32-velu/gen/velu-sse2-rr2-p6-x24.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07002572 "src/f32-vlrelu/gen/vlrelu-sse2-x4.c",
2573 "src/f32-vlrelu/gen/vlrelu-sse2-x8.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07002574 "src/f32-vrnd/gen/vrndd-sse2-x4.c",
2575 "src/f32-vrnd/gen/vrndd-sse2-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002576 "src/f32-vrnd/gen/vrndne-sse2-x4.c",
2577 "src/f32-vrnd/gen/vrndne-sse2-x8.c",
2578 "src/f32-vrnd/gen/vrndu-sse2-x4.c",
2579 "src/f32-vrnd/gen/vrndu-sse2-x8.c",
2580 "src/f32-vrnd/gen/vrndz-sse2-x4.c",
2581 "src/f32-vrnd/gen/vrndz-sse2-x8.c",
Marat Dukhanb1eec082021-05-05 23:24:55 -07002582 "src/f32-vsigmoid/gen/vsigmoid-sse2-lut64-p2-div-x4.c",
2583 "src/f32-vsigmoid/gen/vsigmoid-sse2-lut64-p2-div-x8.c",
2584 "src/f32-vsigmoid/gen/vsigmoid-sse2-lut64-p2-div-x12.c",
2585 "src/f32-vsigmoid/gen/vsigmoid-sse2-lut64-p2-div-x16.c",
2586 "src/f32-vsigmoid/gen/vsigmoid-sse2-lut64-p2-div-x20.c",
2587 "src/f32-vsigmoid/gen/vsigmoid-sse2-lut64-p2-div-x24.c",
2588 "src/f32-vsigmoid/gen/vsigmoid-sse2-p5-div-x4.c",
2589 "src/f32-vsigmoid/gen/vsigmoid-sse2-p5-div-x8.c",
2590 "src/f32-vsigmoid/gen/vsigmoid-sse2-p5-div-x12.c",
2591 "src/f32-vsigmoid/gen/vsigmoid-sse2-p5-div-x16.c",
2592 "src/f32-vsigmoid/gen/vsigmoid-sse2-p5-div-x20.c",
2593 "src/f32-vsigmoid/gen/vsigmoid-sse2-p5-div-x24.c",
Marat Dukhanb7633f22020-11-20 16:34:56 -08002594 "src/math/exp-sse2-rr2-lut64-p2.c",
2595 "src/math/exp-sse2-rr2-p5.c",
Marat Dukhande390d42020-11-29 19:32:18 -08002596 "src/math/expm1minus-sse2-rr2-lut16-p3.c",
Marat Dukhana438aca2020-11-20 15:45:01 -08002597 "src/math/expm1minus-sse2-rr2-p6.c",
Frank Barchard3b800452020-11-22 12:12:35 -08002598 "src/math/expminus-sse2-rr2-p5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002599 "src/math/roundd-sse2-cvt.c",
2600 "src/math/roundne-sse2-cvt.c",
2601 "src/math/roundu-sse2-cvt.c",
2602 "src/math/roundz-sse2-cvt.c",
2603 "src/math/sigmoid-sse2-rr2-lut64-p2-div.c",
2604 "src/math/sigmoid-sse2-rr2-lut64-p2-nr1.c",
2605 "src/math/sigmoid-sse2-rr2-lut64-p2-nr2.c",
2606 "src/math/sigmoid-sse2-rr2-p5-div.c",
2607 "src/math/sigmoid-sse2-rr2-p5-nr1.c",
2608 "src/math/sigmoid-sse2-rr2-p5-nr2.c",
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07002609 "src/qs8-dwconv/gen/up8x9-minmax-sse2-mul16.c",
Marat Dukhan4ed14882021-05-12 17:50:40 -07002610 "src/qs8-dwconv/gen/up8x25-minmax-sse2-mul16.c",
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07002611 "src/qs8-dwconv/gen/up16x9-minmax-sse2-mul16.c",
Marat Dukhan4ed14882021-05-12 17:50:40 -07002612 "src/qs8-dwconv/gen/up16x25-minmax-sse2-mul16.c",
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07002613 "src/qs8-dwconv/gen/up24x9-minmax-sse2-mul16.c",
Marat Dukhan4ed14882021-05-12 17:50:40 -07002614 "src/qs8-dwconv/gen/up24x25-minmax-sse2-mul16.c",
Marat Dukhan159688f2020-08-06 10:34:29 -07002615 "src/qs8-gavgpool/gen/7p7x-minmax-sse2-c8-acc2.c",
2616 "src/qs8-gavgpool/gen/7p7x-minmax-sse2-c16-acc2.c",
2617 "src/qs8-gavgpool/gen/7p7x-minmax-sse2-c24-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002618 "src/qs8-gavgpool/gen/7x-minmax-sse2-c8-acc2.c",
2619 "src/qs8-gavgpool/gen/7x-minmax-sse2-c16-acc2.c",
2620 "src/qs8-gavgpool/gen/7x-minmax-sse2-c24-acc2.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002621 "src/qs8-gemm/gen/1x4c2-minmax-sse2-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002622 "src/qs8-gemm/gen/1x4c2-minmax-sse2-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002623 "src/qs8-gemm/gen/1x4c2-xw-minmax-sse2.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002624 "src/qs8-gemm/gen/1x4c8-minmax-sse2-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002625 "src/qs8-gemm/gen/1x4c8-minmax-sse2-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002626 "src/qs8-gemm/gen/1x4c8-xw-minmax-sse2.c",
Marat Dukhan2e427872021-04-09 23:40:07 -07002627 "src/qs8-gemm/gen/2x4c2-minmax-sse2-ld64.c",
2628 "src/qs8-gemm/gen/2x4c2-minmax-sse2-ld128.c",
2629 "src/qs8-gemm/gen/2x4c2-xw-minmax-sse2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002630 "src/qs8-gemm/gen/2x4c8-minmax-sse2-ld64.c",
2631 "src/qs8-gemm/gen/2x4c8-minmax-sse2-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002632 "src/qs8-gemm/gen/2x4c8-xw-minmax-sse2.c",
Marat Dukhan2e427872021-04-09 23:40:07 -07002633 "src/qs8-gemm/gen/3x4c2-minmax-sse2-ld64.c",
2634 "src/qs8-gemm/gen/3x4c2-minmax-sse2-ld128.c",
2635 "src/qs8-gemm/gen/3x4c2-xw-minmax-sse2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002636 "src/qs8-gemm/gen/3x4c8-minmax-sse2-ld64.c",
2637 "src/qs8-gemm/gen/3x4c8-minmax-sse2-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002638 "src/qs8-gemm/gen/3x4c8-xw-minmax-sse2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002639 "src/qs8-gemm/gen/4x4c2-minmax-sse2-ld64.c",
2640 "src/qs8-gemm/gen/4x4c2-minmax-sse2-ld128.c",
2641 "src/qs8-gemm/gen/4x4c2-xw-minmax-sse2.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002642 "src/qs8-igemm/gen/1x4c2-minmax-sse2-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002643 "src/qs8-igemm/gen/1x4c2-minmax-sse2-ld128.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002644 "src/qs8-igemm/gen/1x4c8-minmax-sse2-ld64.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002645 "src/qs8-igemm/gen/1x4c8-minmax-sse2-ld128.c",
Marat Dukhan2e427872021-04-09 23:40:07 -07002646 "src/qs8-igemm/gen/2x4c2-minmax-sse2-ld64.c",
2647 "src/qs8-igemm/gen/2x4c2-minmax-sse2-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002648 "src/qs8-igemm/gen/2x4c8-minmax-sse2-ld64.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002649 "src/qs8-igemm/gen/2x4c8-minmax-sse2-ld128.c",
Marat Dukhan2e427872021-04-09 23:40:07 -07002650 "src/qs8-igemm/gen/3x4c2-minmax-sse2-ld64.c",
2651 "src/qs8-igemm/gen/3x4c2-minmax-sse2-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002652 "src/qs8-igemm/gen/3x4c8-minmax-sse2-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002653 "src/qs8-igemm/gen/3x4c8-minmax-sse2-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002654 "src/qs8-igemm/gen/4x4c2-minmax-sse2-ld64.c",
2655 "src/qs8-igemm/gen/4x4c2-minmax-sse2-ld128.c",
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07002656 "src/qs8-requantization/fp32-sse2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002657 "src/qs8-requantization/precise-sse2.c",
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07002658 "src/qs8-requantization/q31-sse2.c",
Marat Dukhand9f3ad42020-08-10 12:30:58 -07002659 "src/qs8-vadd/gen/minmax-sse2-mul16-ld64-x8.c",
2660 "src/qs8-vadd/gen/minmax-sse2-mul16-ld64-x16.c",
2661 "src/qs8-vadd/gen/minmax-sse2-mul16-ld64-x24.c",
2662 "src/qs8-vadd/gen/minmax-sse2-mul16-ld64-x32.c",
Marat Dukhan0270d9f2020-08-11 00:56:46 -07002663 "src/qs8-vaddc/gen/minmax-sse2-mul16-ld64-x8.c",
2664 "src/qs8-vaddc/gen/minmax-sse2-mul16-ld64-x16.c",
2665 "src/qs8-vaddc/gen/minmax-sse2-mul16-ld64-x24.c",
2666 "src/qs8-vaddc/gen/minmax-sse2-mul16-ld64-x32.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -07002667 "src/qu8-avgpool/9p8x-minmax-sse2-c8.c",
2668 "src/qu8-avgpool/9x-minmax-sse2-c8.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -07002669 "src/qu8-dwconv/up8x9-minmax-sse2.c",
2670 "src/qu8-gavgpool/7p7x-minmax-sse2-c8.c",
2671 "src/qu8-gavgpool/7x-minmax-sse2-c8.c",
2672 "src/qu8-gemm/2x4c8-minmax-sse2.c",
2673 "src/qu8-gemm/4x4c2-minmax-sse2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002674 "src/qu8-igemm/4x4c2-minmax-sse2.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07002675 "src/qu8-requantization/fp32-sse2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002676 "src/qu8-requantization/precise-sse2.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07002677 "src/qu8-requantization/q31-sse2.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -07002678 "src/qu8-vadd/minmax-sse2.c",
Marat Dukhan99936602020-04-11 16:47:01 -07002679 "src/u8-maxpool/9p8x-minmax-sse2-c16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002680 "src/u8-rmax/sse2.c",
Marat Dukhan6674d692021-05-05 22:27:00 -07002681 "src/u8-vclamp/sse2-x64.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002682 "src/x8-zip/x2-sse2.c",
2683 "src/x8-zip/x3-sse2.c",
2684 "src/x8-zip/x4-sse2.c",
2685 "src/x8-zip/xm-sse2.c",
Marat Dukhan57dccd82020-04-14 00:53:10 -07002686 "src/x32-unpool/sse2.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002687 "src/x32-zip/x2-sse2.c",
2688 "src/x32-zip/x3-sse2.c",
2689 "src/x32-zip/x4-sse2.c",
2690 "src/x32-zip/xm-sse2.c",
Marat Dukhanfe7acb62020-03-09 19:30:05 -07002691]
2692
2693SSSE3_UKERNELS = [
Frank Barchard35db7d02020-10-26 13:37:34 -07002694 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-1x4-acc2.c",
2695 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-1x4-acc3.c",
2696 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-1x4-acc4.c",
Marat Dukhan98f2eeb2020-10-23 23:13:41 -07002697 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-1x4.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07002698 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-2x4-acc2.c",
Marat Dukhan98f2eeb2020-10-23 23:13:41 -07002699 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-2x4.c",
2700 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-3x4.c",
2701 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-4x4.c",
2702 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-5x4.c",
2703 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-6x4.c",
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07002704 "src/qs8-dwconv/gen/up8x9-minmax-ssse3-mul16.c",
Marat Dukhan4ed14882021-05-12 17:50:40 -07002705 "src/qs8-dwconv/gen/up8x25-minmax-ssse3-mul16.c",
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07002706 "src/qs8-dwconv/gen/up16x9-minmax-ssse3-mul16.c",
Marat Dukhan4ed14882021-05-12 17:50:40 -07002707 "src/qs8-dwconv/gen/up16x25-minmax-ssse3-mul16.c",
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07002708 "src/qs8-dwconv/gen/up24x9-minmax-ssse3-mul16.c",
Marat Dukhan4ed14882021-05-12 17:50:40 -07002709 "src/qs8-dwconv/gen/up24x25-minmax-ssse3-mul16.c",
Marat Dukhan159688f2020-08-06 10:34:29 -07002710 "src/qs8-gavgpool/gen/7p7x-minmax-ssse3-c8-acc2.c",
2711 "src/qs8-gavgpool/gen/7p7x-minmax-ssse3-c16-acc2.c",
2712 "src/qs8-gavgpool/gen/7p7x-minmax-ssse3-c24-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002713 "src/qs8-gavgpool/gen/7x-minmax-ssse3-c8-acc2.c",
2714 "src/qs8-gavgpool/gen/7x-minmax-ssse3-c16-acc2.c",
2715 "src/qs8-gavgpool/gen/7x-minmax-ssse3-c24-acc2.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002716 "src/qs8-gemm/gen/1x4c2-minmax-ssse3-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002717 "src/qs8-gemm/gen/1x4c2-minmax-ssse3-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002718 "src/qs8-gemm/gen/1x4c2-xw-minmax-ssse3.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002719 "src/qs8-gemm/gen/1x4c8-minmax-ssse3-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002720 "src/qs8-gemm/gen/1x4c8-minmax-ssse3-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002721 "src/qs8-gemm/gen/1x4c8-xw-minmax-ssse3.c",
Marat Dukhan2e427872021-04-09 23:40:07 -07002722 "src/qs8-gemm/gen/2x4c2-minmax-ssse3-ld64.c",
2723 "src/qs8-gemm/gen/2x4c2-minmax-ssse3-ld128.c",
2724 "src/qs8-gemm/gen/2x4c2-xw-minmax-ssse3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002725 "src/qs8-gemm/gen/2x4c8-minmax-ssse3-ld64.c",
2726 "src/qs8-gemm/gen/2x4c8-minmax-ssse3-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002727 "src/qs8-gemm/gen/2x4c8-xw-minmax-ssse3.c",
Marat Dukhan2e427872021-04-09 23:40:07 -07002728 "src/qs8-gemm/gen/3x4c2-minmax-ssse3-ld64.c",
2729 "src/qs8-gemm/gen/3x4c2-minmax-ssse3-ld128.c",
2730 "src/qs8-gemm/gen/3x4c2-xw-minmax-ssse3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002731 "src/qs8-gemm/gen/3x4c8-minmax-ssse3-ld64.c",
2732 "src/qs8-gemm/gen/3x4c8-minmax-ssse3-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002733 "src/qs8-gemm/gen/3x4c8-xw-minmax-ssse3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002734 "src/qs8-gemm/gen/4x4c2-minmax-ssse3-ld64.c",
2735 "src/qs8-gemm/gen/4x4c2-minmax-ssse3-ld128.c",
2736 "src/qs8-gemm/gen/4x4c2-xw-minmax-ssse3.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002737 "src/qs8-igemm/gen/1x4c2-minmax-ssse3-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002738 "src/qs8-igemm/gen/1x4c2-minmax-ssse3-ld128.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002739 "src/qs8-igemm/gen/1x4c8-minmax-ssse3-ld64.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002740 "src/qs8-igemm/gen/1x4c8-minmax-ssse3-ld128.c",
Marat Dukhan2e427872021-04-09 23:40:07 -07002741 "src/qs8-igemm/gen/2x4c2-minmax-ssse3-ld64.c",
2742 "src/qs8-igemm/gen/2x4c2-minmax-ssse3-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002743 "src/qs8-igemm/gen/2x4c8-minmax-ssse3-ld64.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002744 "src/qs8-igemm/gen/2x4c8-minmax-ssse3-ld128.c",
Marat Dukhan2e427872021-04-09 23:40:07 -07002745 "src/qs8-igemm/gen/3x4c2-minmax-ssse3-ld64.c",
2746 "src/qs8-igemm/gen/3x4c2-minmax-ssse3-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002747 "src/qs8-igemm/gen/3x4c8-minmax-ssse3-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002748 "src/qs8-igemm/gen/3x4c8-minmax-ssse3-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002749 "src/qs8-igemm/gen/4x4c2-minmax-ssse3-ld64.c",
2750 "src/qs8-igemm/gen/4x4c2-minmax-ssse3-ld128.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -07002751 "src/qs8-requantization/precise-ssse3.c",
2752 "src/qs8-requantization/q31-ssse3.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07002753 "src/qu8-requantization/precise-ssse3.c",
2754 "src/qu8-requantization/q31-ssse3.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002755]
2756
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08002757SSE41_UKERNELS = [
Marat Dukhan40a672f2019-11-25 03:08:22 -08002758 "src/f32-prelu/gen/sse41-2x4.c",
2759 "src/f32-prelu/gen/sse41-2x8.c",
Marat Dukhaned6baaf2020-12-01 15:07:08 -08002760 "src/f32-velu/gen/velu-sse41-rr2-lut16-p3-x4.c",
2761 "src/f32-velu/gen/velu-sse41-rr2-lut16-p3-x8.c",
2762 "src/f32-velu/gen/velu-sse41-rr2-lut16-p3-x12.c",
2763 "src/f32-velu/gen/velu-sse41-rr2-lut16-p3-x16.c",
2764 "src/f32-velu/gen/velu-sse41-rr2-lut16-p3-x20.c",
2765 "src/f32-velu/gen/velu-sse41-rr2-lut16-p3-x24.c",
2766 "src/f32-velu/gen/velu-sse41-rr2-p6-x4.c",
2767 "src/f32-velu/gen/velu-sse41-rr2-p6-x8.c",
2768 "src/f32-velu/gen/velu-sse41-rr2-p6-x12.c",
2769 "src/f32-velu/gen/velu-sse41-rr2-p6-x16.c",
2770 "src/f32-velu/gen/velu-sse41-rr2-p6-x20.c",
2771 "src/f32-velu/gen/velu-sse41-rr2-p6-x24.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07002772 "src/f32-vlrelu/gen/vlrelu-sse41-x4.c",
2773 "src/f32-vlrelu/gen/vlrelu-sse41-x8.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07002774 "src/f32-vrnd/gen/vrndd-sse41-x4.c",
2775 "src/f32-vrnd/gen/vrndd-sse41-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002776 "src/f32-vrnd/gen/vrndne-sse41-x4.c",
2777 "src/f32-vrnd/gen/vrndne-sse41-x8.c",
2778 "src/f32-vrnd/gen/vrndu-sse41-x4.c",
2779 "src/f32-vrnd/gen/vrndu-sse41-x8.c",
2780 "src/f32-vrnd/gen/vrndz-sse41-x4.c",
2781 "src/f32-vrnd/gen/vrndz-sse41-x8.c",
Marat Dukhanb1eec082021-05-05 23:24:55 -07002782 "src/f32-vsigmoid/gen/vsigmoid-sse41-lut64-p2-div-x4.c",
2783 "src/f32-vsigmoid/gen/vsigmoid-sse41-lut64-p2-div-x8.c",
2784 "src/f32-vsigmoid/gen/vsigmoid-sse41-lut64-p2-div-x12.c",
2785 "src/f32-vsigmoid/gen/vsigmoid-sse41-lut64-p2-div-x16.c",
2786 "src/f32-vsigmoid/gen/vsigmoid-sse41-lut64-p2-div-x20.c",
2787 "src/f32-vsigmoid/gen/vsigmoid-sse41-lut64-p2-div-x24.c",
2788 "src/f32-vsigmoid/gen/vsigmoid-sse41-p5-div-x4.c",
2789 "src/f32-vsigmoid/gen/vsigmoid-sse41-p5-div-x8.c",
2790 "src/f32-vsigmoid/gen/vsigmoid-sse41-p5-div-x12.c",
2791 "src/f32-vsigmoid/gen/vsigmoid-sse41-p5-div-x16.c",
2792 "src/f32-vsigmoid/gen/vsigmoid-sse41-p5-div-x20.c",
2793 "src/f32-vsigmoid/gen/vsigmoid-sse41-p5-div-x24.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002794 "src/math/roundd-sse41.c",
2795 "src/math/roundne-sse41.c",
2796 "src/math/roundu-sse41.c",
2797 "src/math/roundz-sse41.c",
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07002798 "src/qs8-dwconv/gen/up8x9-minmax-sse41-mul16.c",
Marat Dukhan07feec82021-04-02 22:41:15 -07002799 "src/qs8-dwconv/gen/up8x9-minmax-sse41-mul32.c",
Marat Dukhan4ed14882021-05-12 17:50:40 -07002800 "src/qs8-dwconv/gen/up8x25-minmax-sse41-mul16.c",
2801 "src/qs8-dwconv/gen/up8x25-minmax-sse41-mul32.c",
Frank Barchard4a352042021-04-13 15:52:08 -07002802 "src/qs8-dwconv/gen/up16x9-minmax-sse41-mul16.c",
Marat Dukhan07feec82021-04-02 22:41:15 -07002803 "src/qs8-dwconv/gen/up16x9-minmax-sse41-mul32.c",
Marat Dukhan4ed14882021-05-12 17:50:40 -07002804 "src/qs8-dwconv/gen/up16x25-minmax-sse41-mul16.c",
2805 "src/qs8-dwconv/gen/up16x25-minmax-sse41-mul32.c",
Frank Barchard4a352042021-04-13 15:52:08 -07002806 "src/qs8-dwconv/gen/up24x9-minmax-sse41-mul16.c",
Marat Dukhan07feec82021-04-02 22:41:15 -07002807 "src/qs8-dwconv/gen/up24x9-minmax-sse41-mul32.c",
Marat Dukhan4ed14882021-05-12 17:50:40 -07002808 "src/qs8-dwconv/gen/up24x25-minmax-sse41-mul16.c",
2809 "src/qs8-dwconv/gen/up24x25-minmax-sse41-mul32.c",
Marat Dukhan159688f2020-08-06 10:34:29 -07002810 "src/qs8-gavgpool/gen/7p7x-minmax-sse41-c8-acc2.c",
2811 "src/qs8-gavgpool/gen/7p7x-minmax-sse41-c16-acc2.c",
2812 "src/qs8-gavgpool/gen/7p7x-minmax-sse41-c24-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002813 "src/qs8-gavgpool/gen/7x-minmax-sse41-c8-acc2.c",
2814 "src/qs8-gavgpool/gen/7x-minmax-sse41-c16-acc2.c",
2815 "src/qs8-gavgpool/gen/7x-minmax-sse41-c24-acc2.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002816 "src/qs8-gemm/gen/1x4c2-minmax-sse41-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002817 "src/qs8-gemm/gen/1x4c2-minmax-sse41-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002818 "src/qs8-gemm/gen/1x4c2-xw-minmax-sse41.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002819 "src/qs8-gemm/gen/1x4c8-minmax-sse41-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002820 "src/qs8-gemm/gen/1x4c8-minmax-sse41-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002821 "src/qs8-gemm/gen/1x4c8-xw-minmax-sse41.c",
Marat Dukhan2e427872021-04-09 23:40:07 -07002822 "src/qs8-gemm/gen/2x4c2-minmax-sse41-ld64.c",
2823 "src/qs8-gemm/gen/2x4c2-minmax-sse41-ld128.c",
2824 "src/qs8-gemm/gen/2x4c2-xw-minmax-sse41.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002825 "src/qs8-gemm/gen/2x4c8-minmax-sse41-ld64.c",
2826 "src/qs8-gemm/gen/2x4c8-minmax-sse41-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002827 "src/qs8-gemm/gen/2x4c8-xw-minmax-sse41.c",
Marat Dukhan2e427872021-04-09 23:40:07 -07002828 "src/qs8-gemm/gen/3x4c2-minmax-sse41-ld64.c",
2829 "src/qs8-gemm/gen/3x4c2-minmax-sse41-ld128.c",
2830 "src/qs8-gemm/gen/3x4c2-xw-minmax-sse41.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002831 "src/qs8-gemm/gen/3x4c8-minmax-sse41-ld64.c",
2832 "src/qs8-gemm/gen/3x4c8-minmax-sse41-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002833 "src/qs8-gemm/gen/3x4c8-xw-minmax-sse41.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002834 "src/qs8-gemm/gen/4x4c2-minmax-sse41-ld64.c",
2835 "src/qs8-gemm/gen/4x4c2-minmax-sse41-ld128.c",
2836 "src/qs8-gemm/gen/4x4c2-xw-minmax-sse41.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002837 "src/qs8-igemm/gen/1x4c2-minmax-sse41-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002838 "src/qs8-igemm/gen/1x4c2-minmax-sse41-ld128.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002839 "src/qs8-igemm/gen/1x4c8-minmax-sse41-ld64.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002840 "src/qs8-igemm/gen/1x4c8-minmax-sse41-ld128.c",
Marat Dukhan2e427872021-04-09 23:40:07 -07002841 "src/qs8-igemm/gen/2x4c2-minmax-sse41-ld64.c",
2842 "src/qs8-igemm/gen/2x4c2-minmax-sse41-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002843 "src/qs8-igemm/gen/2x4c8-minmax-sse41-ld64.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002844 "src/qs8-igemm/gen/2x4c8-minmax-sse41-ld128.c",
Marat Dukhan2e427872021-04-09 23:40:07 -07002845 "src/qs8-igemm/gen/3x4c2-minmax-sse41-ld64.c",
2846 "src/qs8-igemm/gen/3x4c2-minmax-sse41-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002847 "src/qs8-igemm/gen/3x4c8-minmax-sse41-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002848 "src/qs8-igemm/gen/3x4c8-minmax-sse41-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002849 "src/qs8-igemm/gen/4x4c2-minmax-sse41-ld64.c",
2850 "src/qs8-igemm/gen/4x4c2-minmax-sse41-ld128.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -07002851 "src/qs8-requantization/fp32-sse4.c",
2852 "src/qs8-requantization/precise-sse4.c",
2853 "src/qs8-requantization/q31-sse4.c",
Marat Dukhand9f3ad42020-08-10 12:30:58 -07002854 "src/qs8-vadd/gen/minmax-sse41-mul16-ld64-x8.c",
2855 "src/qs8-vadd/gen/minmax-sse41-mul16-ld64-x16.c",
2856 "src/qs8-vadd/gen/minmax-sse41-mul16-ld64-x24.c",
2857 "src/qs8-vadd/gen/minmax-sse41-mul16-ld64-x32.c",
Marat Dukhanbb9225e2020-09-06 22:40:56 -07002858 "src/qs8-vadd/gen/minmax-sse41-mul32-ld32-x8.c",
2859 "src/qs8-vadd/gen/minmax-sse41-mul32-ld32-x16.c",
2860 "src/qs8-vadd/gen/minmax-sse41-mul32-ld32-x24.c",
2861 "src/qs8-vadd/gen/minmax-sse41-mul32-ld32-x32.c",
Marat Dukhan0270d9f2020-08-11 00:56:46 -07002862 "src/qs8-vaddc/gen/minmax-sse41-mul16-ld64-x8.c",
2863 "src/qs8-vaddc/gen/minmax-sse41-mul16-ld64-x16.c",
2864 "src/qs8-vaddc/gen/minmax-sse41-mul16-ld64-x24.c",
2865 "src/qs8-vaddc/gen/minmax-sse41-mul16-ld64-x32.c",
Marat Dukhanbb9225e2020-09-06 22:40:56 -07002866 "src/qs8-vaddc/gen/minmax-sse41-mul32-ld32-x8.c",
2867 "src/qs8-vaddc/gen/minmax-sse41-mul32-ld32-x16.c",
2868 "src/qs8-vaddc/gen/minmax-sse41-mul32-ld32-x24.c",
2869 "src/qs8-vaddc/gen/minmax-sse41-mul32-ld32-x32.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07002870 "src/qu8-requantization/precise-sse4.c",
2871 "src/qu8-requantization/q31-sse4.c",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08002872]
2873
Marat Dukhan08c4a432019-10-03 09:29:21 -07002874AVX_UKERNELS = [
Marat Dukhan1c587112020-04-08 20:04:28 -07002875 "src/f32-dwconv/gen/up8x4-minmax-avx-acc2.c",
2876 "src/f32-dwconv/gen/up8x4-minmax-avx.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002877 "src/f32-dwconv/gen/up8x9-minmax-avx-acc2.c",
2878 "src/f32-dwconv/gen/up8x9-minmax-avx.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002879 "src/f32-dwconv/gen/up8x25-minmax-avx-acc2.c",
2880 "src/f32-dwconv/gen/up8x25-minmax-avx.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002881 "src/f32-dwconv/gen/up16x4-minmax-avx-acc2.c",
2882 "src/f32-dwconv/gen/up16x4-minmax-avx.c",
2883 "src/f32-dwconv/gen/up16x9-minmax-avx-acc2.c",
2884 "src/f32-dwconv/gen/up16x9-minmax-avx.c",
2885 "src/f32-dwconv/gen/up16x25-minmax-avx-acc2.c",
2886 "src/f32-dwconv/gen/up16x25-minmax-avx.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002887 "src/f32-gemm/gen-inc/1x8inc-minmax-avx-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002888 "src/f32-gemm/gen-inc/1x16inc-minmax-avx-broadcast.c",
2889 "src/f32-gemm/gen-inc/3x16inc-minmax-avx-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002890 "src/f32-gemm/gen-inc/4x8inc-minmax-avx-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002891 "src/f32-gemm/gen-inc/4x16inc-minmax-avx-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002892 "src/f32-gemm/gen-inc/5x8inc-minmax-avx-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002893 "src/f32-gemm/gen-inc/5x16inc-minmax-avx-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002894 "src/f32-gemm/gen-inc/6x8inc-minmax-avx-broadcast.c",
2895 "src/f32-gemm/gen-inc/7x8inc-minmax-avx-broadcast.c",
2896 "src/f32-gemm/gen/1x8-minmax-avx-broadcast.c",
2897 "src/f32-gemm/gen/1x16-minmax-avx-broadcast.c",
2898 "src/f32-gemm/gen/3x16-minmax-avx-broadcast.c",
2899 "src/f32-gemm/gen/4x8-minmax-avx-broadcast.c",
2900 "src/f32-gemm/gen/4x16-minmax-avx-broadcast.c",
2901 "src/f32-gemm/gen/5x8-minmax-avx-broadcast.c",
2902 "src/f32-gemm/gen/5x16-minmax-avx-broadcast.c",
2903 "src/f32-gemm/gen/6x8-minmax-avx-broadcast.c",
2904 "src/f32-gemm/gen/7x8-minmax-avx-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002905 "src/f32-igemm/gen/1x8-minmax-avx-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002906 "src/f32-igemm/gen/1x16-minmax-avx-broadcast.c",
2907 "src/f32-igemm/gen/3x16-minmax-avx-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002908 "src/f32-igemm/gen/4x8-minmax-avx-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002909 "src/f32-igemm/gen/4x16-minmax-avx-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002910 "src/f32-igemm/gen/5x8-minmax-avx-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002911 "src/f32-igemm/gen/5x16-minmax-avx-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002912 "src/f32-igemm/gen/6x8-minmax-avx-broadcast.c",
2913 "src/f32-igemm/gen/7x8-minmax-avx-broadcast.c",
Marat Dukhan90eca0a2020-03-11 00:52:23 -07002914 "src/f32-prelu/gen/avx-2x8.c",
2915 "src/f32-prelu/gen/avx-2x16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002916 "src/f32-rmax/avx.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002917 "src/f32-vbinary/gen/vadd-minmax-avx-x8.c",
2918 "src/f32-vbinary/gen/vadd-minmax-avx-x16.c",
2919 "src/f32-vbinary/gen/vaddc-minmax-avx-x8.c",
2920 "src/f32-vbinary/gen/vaddc-minmax-avx-x16.c",
2921 "src/f32-vbinary/gen/vdiv-minmax-avx-x8.c",
2922 "src/f32-vbinary/gen/vdiv-minmax-avx-x16.c",
2923 "src/f32-vbinary/gen/vdivc-minmax-avx-x8.c",
2924 "src/f32-vbinary/gen/vdivc-minmax-avx-x16.c",
Marat Dukhan9a88efe2019-12-10 15:54:24 -08002925 "src/f32-vbinary/gen/vmax-avx-x8.c",
2926 "src/f32-vbinary/gen/vmax-avx-x16.c",
2927 "src/f32-vbinary/gen/vmaxc-avx-x8.c",
2928 "src/f32-vbinary/gen/vmaxc-avx-x16.c",
2929 "src/f32-vbinary/gen/vmin-avx-x8.c",
2930 "src/f32-vbinary/gen/vmin-avx-x16.c",
2931 "src/f32-vbinary/gen/vminc-avx-x8.c",
2932 "src/f32-vbinary/gen/vminc-avx-x16.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002933 "src/f32-vbinary/gen/vmul-minmax-avx-x8.c",
2934 "src/f32-vbinary/gen/vmul-minmax-avx-x16.c",
2935 "src/f32-vbinary/gen/vmulc-minmax-avx-x8.c",
2936 "src/f32-vbinary/gen/vmulc-minmax-avx-x16.c",
2937 "src/f32-vbinary/gen/vrdivc-minmax-avx-x8.c",
2938 "src/f32-vbinary/gen/vrdivc-minmax-avx-x16.c",
2939 "src/f32-vbinary/gen/vrsubc-minmax-avx-x8.c",
2940 "src/f32-vbinary/gen/vrsubc-minmax-avx-x16.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07002941 "src/f32-vbinary/gen/vsqrdiff-avx-x8.c",
2942 "src/f32-vbinary/gen/vsqrdiff-avx-x16.c",
2943 "src/f32-vbinary/gen/vsqrdiffc-avx-x8.c",
2944 "src/f32-vbinary/gen/vsqrdiffc-avx-x16.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002945 "src/f32-vbinary/gen/vsub-minmax-avx-x8.c",
2946 "src/f32-vbinary/gen/vsub-minmax-avx-x16.c",
2947 "src/f32-vbinary/gen/vsubc-minmax-avx-x8.c",
2948 "src/f32-vbinary/gen/vsubc-minmax-avx-x16.c",
Marat Dukhanb1eec082021-05-05 23:24:55 -07002949 "src/f32-vclamp/gen/vclamp-avx-x8.c",
2950 "src/f32-vclamp/gen/vclamp-avx-x16.c",
Marat Dukhaned6baaf2020-12-01 15:07:08 -08002951 "src/f32-velu/gen/velu-avx-rr2-lut4-p4-perm-x8.c",
2952 "src/f32-velu/gen/velu-avx-rr2-lut4-p4-perm-x16.c",
2953 "src/f32-velu/gen/velu-avx-rr2-lut4-p4-perm-x24.c",
2954 "src/f32-velu/gen/velu-avx-rr2-lut4-p4-perm-x32.c",
2955 "src/f32-velu/gen/velu-avx-rr2-lut4-p4-perm-x40.c",
2956 "src/f32-velu/gen/velu-avx-rr2-lut4-p4-perm-x48.c",
2957 "src/f32-velu/gen/velu-avx-rr2-lut16-p3-x8.c",
2958 "src/f32-velu/gen/velu-avx-rr2-lut16-p3-x16.c",
2959 "src/f32-velu/gen/velu-avx-rr2-lut16-p3-x24.c",
2960 "src/f32-velu/gen/velu-avx-rr2-lut16-p3-x32.c",
2961 "src/f32-velu/gen/velu-avx-rr2-lut16-p3-x40.c",
2962 "src/f32-velu/gen/velu-avx-rr2-lut16-p3-x48.c",
2963 "src/f32-velu/gen/velu-avx-rr2-p6-x8.c",
2964 "src/f32-velu/gen/velu-avx-rr2-p6-x16.c",
2965 "src/f32-velu/gen/velu-avx-rr2-p6-x24.c",
2966 "src/f32-velu/gen/velu-avx-rr2-p6-x32.c",
2967 "src/f32-velu/gen/velu-avx-rr2-p6-x40.c",
2968 "src/f32-velu/gen/velu-avx-rr2-p6-x48.c",
Marat Dukhan6674d692021-05-05 22:27:00 -07002969 "src/f32-vhswish/gen/vhswish-avx-x8.c",
2970 "src/f32-vhswish/gen/vhswish-avx-x16.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07002971 "src/f32-vlrelu/gen/vlrelu-avx-x8.c",
2972 "src/f32-vlrelu/gen/vlrelu-avx-x16.c",
Marat Dukhanb1eec082021-05-05 23:24:55 -07002973 "src/f32-vrelu/gen/vrelu-avx-x8.c",
2974 "src/f32-vrelu/gen/vrelu-avx-x16.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07002975 "src/f32-vrnd/gen/vrndd-avx-x8.c",
2976 "src/f32-vrnd/gen/vrndd-avx-x16.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002977 "src/f32-vrnd/gen/vrndne-avx-x8.c",
2978 "src/f32-vrnd/gen/vrndne-avx-x16.c",
2979 "src/f32-vrnd/gen/vrndu-avx-x8.c",
2980 "src/f32-vrnd/gen/vrndu-avx-x16.c",
2981 "src/f32-vrnd/gen/vrndz-avx-x8.c",
2982 "src/f32-vrnd/gen/vrndz-avx-x16.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07002983 "src/f32-vscale/avx-x32.c",
Marat Dukhanb1eec082021-05-05 23:24:55 -07002984 "src/f32-vsigmoid/gen/vsigmoid-avx-rr2-p5-div-x8.c",
2985 "src/f32-vsigmoid/gen/vsigmoid-avx-rr2-p5-div-x16.c",
2986 "src/f32-vsigmoid/gen/vsigmoid-avx-rr2-p5-div-x24.c",
2987 "src/f32-vsigmoid/gen/vsigmoid-avx-rr2-p5-div-x32.c",
2988 "src/f32-vsigmoid/gen/vsigmoid-avx-rr2-p5-div-x40.c",
2989 "src/f32-vsigmoid/gen/vsigmoid-avx-rr2-p5-div-x48.c",
2990 "src/f32-vsigmoid/gen/vsigmoid-avx-rr2-p5-div-x56.c",
2991 "src/f32-vsigmoid/gen/vsigmoid-avx-rr2-p5-div-x64.c",
2992 "src/f32-vsigmoid/gen/vsigmoid-avx-rr2-p5-div-x72.c",
2993 "src/f32-vsigmoid/gen/vsigmoid-avx-rr2-p5-div-x80.c",
2994 "src/f32-vsigmoid/gen/vsigmoid-avx-rr2-p5-nr2-x8.c",
2995 "src/f32-vsigmoid/gen/vsigmoid-avx-rr2-p5-nr2-x16.c",
2996 "src/f32-vsigmoid/gen/vsigmoid-avx-rr2-p5-nr2-x24.c",
2997 "src/f32-vsigmoid/gen/vsigmoid-avx-rr2-p5-nr2-x32.c",
2998 "src/f32-vsigmoid/gen/vsigmoid-avx-rr2-p5-nr2-x40.c",
2999 "src/f32-vsigmoid/gen/vsigmoid-avx-rr2-p5-nr2-x48.c",
3000 "src/f32-vsigmoid/gen/vsigmoid-avx-rr2-p5-nr2-x56.c",
3001 "src/f32-vsigmoid/gen/vsigmoid-avx-rr2-p5-nr2-x64.c",
3002 "src/f32-vsigmoid/gen/vsigmoid-avx-rr2-p5-nr2-x72.c",
3003 "src/f32-vsigmoid/gen/vsigmoid-avx-rr2-p5-nr2-x80.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07003004 "src/f32-vsqrt/gen/avx-sqrt-x8.c",
3005 "src/f32-vsqrt/gen/avx-sqrt-x16.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07003006 "src/f32-vunary/gen/vabs-avx-x8.c",
3007 "src/f32-vunary/gen/vabs-avx-x16.c",
3008 "src/f32-vunary/gen/vneg-avx-x8.c",
3009 "src/f32-vunary/gen/vneg-avx-x16.c",
3010 "src/f32-vunary/gen/vsqr-avx-x8.c",
3011 "src/f32-vunary/gen/vsqr-avx-x16.c",
Frank Barchard4a352042021-04-13 15:52:08 -07003012 "src/math/exp-avx-rr2-p5.c",
3013 "src/math/expm1minus-avx-rr2-lut4-p4-perm.c",
3014 "src/math/expm1minus-avx-rr2-lut16-p3.c",
3015 "src/math/expm1minus-avx-rr2-p6.c",
3016 "src/math/sigmoid-avx-rr2-lut64-p2-div.c",
3017 "src/math/sigmoid-avx-rr2-p5-div.c",
3018 "src/math/sigmoid-avx-rr2-p5-nr1.c",
3019 "src/math/sigmoid-avx-rr2-p5-nr2.c",
Marat Dukhanfa0ab852021-04-02 17:30:49 -07003020 "src/qs8-dwconv/gen/up8x9-minmax-avx-mul16.c",
Marat Dukhan07feec82021-04-02 22:41:15 -07003021 "src/qs8-dwconv/gen/up8x9-minmax-avx-mul32.c",
Marat Dukhan4ed14882021-05-12 17:50:40 -07003022 "src/qs8-dwconv/gen/up8x25-minmax-avx-mul16.c",
3023 "src/qs8-dwconv/gen/up8x25-minmax-avx-mul32.c",
Frank Barchard4a352042021-04-13 15:52:08 -07003024 "src/qs8-dwconv/gen/up16x9-minmax-avx-mul16.c",
Marat Dukhan07feec82021-04-02 22:41:15 -07003025 "src/qs8-dwconv/gen/up16x9-minmax-avx-mul32.c",
Marat Dukhan4ed14882021-05-12 17:50:40 -07003026 "src/qs8-dwconv/gen/up16x25-minmax-avx-mul16.c",
3027 "src/qs8-dwconv/gen/up16x25-minmax-avx-mul32.c",
Frank Barchard4a352042021-04-13 15:52:08 -07003028 "src/qs8-dwconv/gen/up24x9-minmax-avx-mul16.c",
Marat Dukhan07feec82021-04-02 22:41:15 -07003029 "src/qs8-dwconv/gen/up24x9-minmax-avx-mul32.c",
Marat Dukhan4ed14882021-05-12 17:50:40 -07003030 "src/qs8-dwconv/gen/up24x25-minmax-avx-mul16.c",
3031 "src/qs8-dwconv/gen/up24x25-minmax-avx-mul32.c",
Marat Dukhana3c16332021-04-02 15:03:27 -07003032 "src/qs8-gemm/gen/1x4c2-minmax-avx-ld64.c",
3033 "src/qs8-gemm/gen/1x4c2-minmax-avx-ld128.c",
3034 "src/qs8-gemm/gen/1x4c2-xw-minmax-avx.c",
3035 "src/qs8-gemm/gen/1x4c8-minmax-avx-ld64.c",
3036 "src/qs8-gemm/gen/1x4c8-minmax-avx-ld128.c",
3037 "src/qs8-gemm/gen/1x4c8-xw-minmax-avx.c",
Marat Dukhan2e427872021-04-09 23:40:07 -07003038 "src/qs8-gemm/gen/2x4c2-minmax-avx-ld64.c",
3039 "src/qs8-gemm/gen/2x4c2-minmax-avx-ld128.c",
3040 "src/qs8-gemm/gen/2x4c2-xw-minmax-avx.c",
Marat Dukhana3c16332021-04-02 15:03:27 -07003041 "src/qs8-gemm/gen/2x4c8-minmax-avx-ld64.c",
3042 "src/qs8-gemm/gen/2x4c8-minmax-avx-ld128.c",
3043 "src/qs8-gemm/gen/2x4c8-xw-minmax-avx.c",
Marat Dukhan2e427872021-04-09 23:40:07 -07003044 "src/qs8-gemm/gen/3x4c2-minmax-avx-ld64.c",
3045 "src/qs8-gemm/gen/3x4c2-minmax-avx-ld128.c",
3046 "src/qs8-gemm/gen/3x4c2-xw-minmax-avx.c",
Marat Dukhana3c16332021-04-02 15:03:27 -07003047 "src/qs8-gemm/gen/3x4c8-minmax-avx-ld64.c",
3048 "src/qs8-gemm/gen/3x4c8-minmax-avx-ld128.c",
3049 "src/qs8-gemm/gen/3x4c8-xw-minmax-avx.c",
3050 "src/qs8-gemm/gen/4x4c2-minmax-avx-ld64.c",
3051 "src/qs8-gemm/gen/4x4c2-minmax-avx-ld128.c",
3052 "src/qs8-gemm/gen/4x4c2-xw-minmax-avx.c",
3053 "src/qs8-igemm/gen/1x4c2-minmax-avx-ld64.c",
3054 "src/qs8-igemm/gen/1x4c2-minmax-avx-ld128.c",
3055 "src/qs8-igemm/gen/1x4c8-minmax-avx-ld64.c",
3056 "src/qs8-igemm/gen/1x4c8-minmax-avx-ld128.c",
Marat Dukhan2e427872021-04-09 23:40:07 -07003057 "src/qs8-igemm/gen/2x4c2-minmax-avx-ld64.c",
3058 "src/qs8-igemm/gen/2x4c2-minmax-avx-ld128.c",
Marat Dukhana3c16332021-04-02 15:03:27 -07003059 "src/qs8-igemm/gen/2x4c8-minmax-avx-ld64.c",
3060 "src/qs8-igemm/gen/2x4c8-minmax-avx-ld128.c",
Marat Dukhan2e427872021-04-09 23:40:07 -07003061 "src/qs8-igemm/gen/3x4c2-minmax-avx-ld64.c",
3062 "src/qs8-igemm/gen/3x4c2-minmax-avx-ld128.c",
Marat Dukhana3c16332021-04-02 15:03:27 -07003063 "src/qs8-igemm/gen/3x4c8-minmax-avx-ld64.c",
3064 "src/qs8-igemm/gen/3x4c8-minmax-avx-ld128.c",
3065 "src/qs8-igemm/gen/4x4c2-minmax-avx-ld64.c",
3066 "src/qs8-igemm/gen/4x4c2-minmax-avx-ld128.c",
Marat Dukhane9c4b962021-04-02 16:56:55 -07003067 "src/qs8-vadd/gen/minmax-avx-mul16-ld64-x8.c",
3068 "src/qs8-vadd/gen/minmax-avx-mul16-ld64-x16.c",
3069 "src/qs8-vadd/gen/minmax-avx-mul16-ld64-x24.c",
3070 "src/qs8-vadd/gen/minmax-avx-mul16-ld64-x32.c",
3071 "src/qs8-vadd/gen/minmax-avx-mul32-ld32-x8.c",
3072 "src/qs8-vadd/gen/minmax-avx-mul32-ld32-x16.c",
3073 "src/qs8-vadd/gen/minmax-avx-mul32-ld32-x24.c",
3074 "src/qs8-vadd/gen/minmax-avx-mul32-ld32-x32.c",
3075 "src/qs8-vaddc/gen/minmax-avx-mul16-ld64-x8.c",
3076 "src/qs8-vaddc/gen/minmax-avx-mul16-ld64-x16.c",
3077 "src/qs8-vaddc/gen/minmax-avx-mul16-ld64-x24.c",
3078 "src/qs8-vaddc/gen/minmax-avx-mul16-ld64-x32.c",
3079 "src/qs8-vaddc/gen/minmax-avx-mul32-ld32-x8.c",
3080 "src/qs8-vaddc/gen/minmax-avx-mul32-ld32-x16.c",
3081 "src/qs8-vaddc/gen/minmax-avx-mul32-ld32-x24.c",
3082 "src/qs8-vaddc/gen/minmax-avx-mul32-ld32-x32.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003083]
3084
Marat Dukhan1566fee2020-08-02 21:55:41 -07003085XOP_UKERNELS = [
Marat Dukhan3fd4e272021-04-10 11:16:42 -07003086 "src/qs8-dwconv/gen/up8x9-minmax-xop-mul32.c",
Marat Dukhan4ed14882021-05-12 17:50:40 -07003087 "src/qs8-dwconv/gen/up8x25-minmax-xop-mul32.c",
Marat Dukhan3fd4e272021-04-10 11:16:42 -07003088 "src/qs8-dwconv/gen/up16x9-minmax-xop-mul32.c",
Marat Dukhan4ed14882021-05-12 17:50:40 -07003089 "src/qs8-dwconv/gen/up16x25-minmax-xop-mul32.c",
Marat Dukhan3fd4e272021-04-10 11:16:42 -07003090 "src/qs8-dwconv/gen/up24x9-minmax-xop-mul32.c",
Marat Dukhan4ed14882021-05-12 17:50:40 -07003091 "src/qs8-dwconv/gen/up24x25-minmax-xop-mul32.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07003092 "src/qs8-gemm/gen/1x4c2-minmax-xop-ld64.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07003093 "src/qs8-gemm/gen/1x4c2-minmax-xop-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07003094 "src/qs8-gemm/gen/1x4c2-xw-minmax-xop.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07003095 "src/qs8-gemm/gen/1x4c8-minmax-xop-ld64.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07003096 "src/qs8-gemm/gen/1x4c8-minmax-xop-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07003097 "src/qs8-gemm/gen/1x4c8-xw-minmax-xop.c",
Marat Dukhan2e427872021-04-09 23:40:07 -07003098 "src/qs8-gemm/gen/2x4c2-minmax-xop-ld64.c",
3099 "src/qs8-gemm/gen/2x4c2-minmax-xop-ld128.c",
3100 "src/qs8-gemm/gen/2x4c2-xw-minmax-xop.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003101 "src/qs8-gemm/gen/2x4c8-minmax-xop-ld64.c",
3102 "src/qs8-gemm/gen/2x4c8-minmax-xop-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07003103 "src/qs8-gemm/gen/2x4c8-xw-minmax-xop.c",
Marat Dukhan2e427872021-04-09 23:40:07 -07003104 "src/qs8-gemm/gen/3x4c2-minmax-xop-ld64.c",
3105 "src/qs8-gemm/gen/3x4c2-minmax-xop-ld128.c",
3106 "src/qs8-gemm/gen/3x4c2-xw-minmax-xop.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003107 "src/qs8-gemm/gen/3x4c8-minmax-xop-ld64.c",
3108 "src/qs8-gemm/gen/3x4c8-minmax-xop-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07003109 "src/qs8-gemm/gen/3x4c8-xw-minmax-xop.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003110 "src/qs8-gemm/gen/4x4c2-minmax-xop-ld64.c",
3111 "src/qs8-gemm/gen/4x4c2-minmax-xop-ld128.c",
3112 "src/qs8-gemm/gen/4x4c2-xw-minmax-xop.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07003113 "src/qs8-igemm/gen/1x4c2-minmax-xop-ld64.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07003114 "src/qs8-igemm/gen/1x4c2-minmax-xop-ld128.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07003115 "src/qs8-igemm/gen/1x4c8-minmax-xop-ld64.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07003116 "src/qs8-igemm/gen/1x4c8-minmax-xop-ld128.c",
Marat Dukhan2e427872021-04-09 23:40:07 -07003117 "src/qs8-igemm/gen/2x4c2-minmax-xop-ld64.c",
3118 "src/qs8-igemm/gen/2x4c2-minmax-xop-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003119 "src/qs8-igemm/gen/2x4c8-minmax-xop-ld64.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07003120 "src/qs8-igemm/gen/2x4c8-minmax-xop-ld128.c",
Marat Dukhan2e427872021-04-09 23:40:07 -07003121 "src/qs8-igemm/gen/3x4c2-minmax-xop-ld64.c",
3122 "src/qs8-igemm/gen/3x4c2-minmax-xop-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003123 "src/qs8-igemm/gen/3x4c8-minmax-xop-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07003124 "src/qs8-igemm/gen/3x4c8-minmax-xop-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003125 "src/qs8-igemm/gen/4x4c2-minmax-xop-ld64.c",
3126 "src/qs8-igemm/gen/4x4c2-minmax-xop-ld128.c",
Marat Dukhanbb9225e2020-09-06 22:40:56 -07003127 "src/qs8-vadd/gen/minmax-xop-mul32-ld32-x8.c",
3128 "src/qs8-vadd/gen/minmax-xop-mul32-ld32-x16.c",
3129 "src/qs8-vadd/gen/minmax-xop-mul32-ld32-x24.c",
3130 "src/qs8-vadd/gen/minmax-xop-mul32-ld32-x32.c",
3131 "src/qs8-vaddc/gen/minmax-xop-mul32-ld32-x8.c",
3132 "src/qs8-vaddc/gen/minmax-xop-mul32-ld32-x16.c",
3133 "src/qs8-vaddc/gen/minmax-xop-mul32-ld32-x24.c",
3134 "src/qs8-vaddc/gen/minmax-xop-mul32-ld32-x32.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07003135]
3136
Marat Dukhanfda12b82019-11-21 12:27:59 -08003137FMA3_UKERNELS = [
Marat Dukhan1c587112020-04-08 20:04:28 -07003138 "src/f32-dwconv/gen/up8x4-minmax-fma3-acc2.c",
3139 "src/f32-dwconv/gen/up8x4-minmax-fma3.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07003140 "src/f32-dwconv/gen/up8x9-minmax-fma3-acc2.c",
3141 "src/f32-dwconv/gen/up8x9-minmax-fma3.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07003142 "src/f32-dwconv/gen/up8x25-minmax-fma3-acc2.c",
3143 "src/f32-dwconv/gen/up8x25-minmax-fma3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003144 "src/f32-dwconv/gen/up16x4-minmax-fma3-acc2.c",
3145 "src/f32-dwconv/gen/up16x4-minmax-fma3.c",
3146 "src/f32-dwconv/gen/up16x9-minmax-fma3-acc2.c",
3147 "src/f32-dwconv/gen/up16x9-minmax-fma3.c",
3148 "src/f32-dwconv/gen/up16x25-minmax-fma3-acc2.c",
3149 "src/f32-dwconv/gen/up16x25-minmax-fma3.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07003150 "src/f32-gemm/gen-inc/1x8inc-minmax-fma3-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003151 "src/f32-gemm/gen-inc/1x16inc-minmax-fma3-broadcast.c",
3152 "src/f32-gemm/gen-inc/1x16s4inc-minmax-fma3-broadcast.c",
3153 "src/f32-gemm/gen-inc/3x16inc-minmax-fma3-broadcast.c",
3154 "src/f32-gemm/gen-inc/3x16s4inc-minmax-fma3-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07003155 "src/f32-gemm/gen-inc/4x8inc-minmax-fma3-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003156 "src/f32-gemm/gen-inc/4x16inc-minmax-fma3-broadcast.c",
3157 "src/f32-gemm/gen-inc/4x16s4inc-minmax-fma3-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07003158 "src/f32-gemm/gen-inc/5x8inc-minmax-fma3-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003159 "src/f32-gemm/gen-inc/5x16inc-minmax-fma3-broadcast.c",
3160 "src/f32-gemm/gen-inc/5x16s4inc-minmax-fma3-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07003161 "src/f32-gemm/gen-inc/6x8inc-minmax-fma3-broadcast.c",
3162 "src/f32-gemm/gen-inc/7x8inc-minmax-fma3-broadcast.c",
3163 "src/f32-gemm/gen-inc/8x8inc-minmax-fma3-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003164 "src/f32-gemm/gen/1x8-minmax-fma3-broadcast.c",
3165 "src/f32-gemm/gen/1x16-minmax-fma3-broadcast.c",
3166 "src/f32-gemm/gen/1x16s4-minmax-fma3-broadcast.c",
3167 "src/f32-gemm/gen/3x16-minmax-fma3-broadcast.c",
3168 "src/f32-gemm/gen/3x16s4-minmax-fma3-broadcast.c",
3169 "src/f32-gemm/gen/4x8-minmax-fma3-broadcast.c",
3170 "src/f32-gemm/gen/4x16-minmax-fma3-broadcast.c",
3171 "src/f32-gemm/gen/4x16s4-minmax-fma3-broadcast.c",
3172 "src/f32-gemm/gen/5x8-minmax-fma3-broadcast.c",
3173 "src/f32-gemm/gen/5x16-minmax-fma3-broadcast.c",
3174 "src/f32-gemm/gen/5x16s4-minmax-fma3-broadcast.c",
3175 "src/f32-gemm/gen/6x8-minmax-fma3-broadcast.c",
3176 "src/f32-gemm/gen/7x8-minmax-fma3-broadcast.c",
3177 "src/f32-gemm/gen/8x8-minmax-fma3-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07003178 "src/f32-igemm/gen/1x8-minmax-fma3-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003179 "src/f32-igemm/gen/1x16-minmax-fma3-broadcast.c",
3180 "src/f32-igemm/gen/1x16s4-minmax-fma3-broadcast.c",
3181 "src/f32-igemm/gen/3x16-minmax-fma3-broadcast.c",
3182 "src/f32-igemm/gen/3x16s4-minmax-fma3-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07003183 "src/f32-igemm/gen/4x8-minmax-fma3-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003184 "src/f32-igemm/gen/4x16-minmax-fma3-broadcast.c",
3185 "src/f32-igemm/gen/4x16s4-minmax-fma3-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07003186 "src/f32-igemm/gen/5x8-minmax-fma3-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003187 "src/f32-igemm/gen/5x16-minmax-fma3-broadcast.c",
3188 "src/f32-igemm/gen/5x16s4-minmax-fma3-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07003189 "src/f32-igemm/gen/6x8-minmax-fma3-broadcast.c",
3190 "src/f32-igemm/gen/7x8-minmax-fma3-broadcast.c",
3191 "src/f32-igemm/gen/8x8-minmax-fma3-broadcast.c",
Marat Dukhan6674d692021-05-05 22:27:00 -07003192 "src/f32-vhswish/gen/vhswish-fma3-x8.c",
3193 "src/f32-vhswish/gen/vhswish-fma3-x16.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07003194 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x8.c",
3195 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x16.c",
3196 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x24.c",
3197 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x32.c",
3198 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x40.c",
3199 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x48.c",
3200 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x56.c",
3201 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x64.c",
Marat Dukhan84000762020-06-29 18:38:43 -07003202 "src/math/sqrt-fma3-nr1fma.c",
Marat Dukhan84000762020-06-29 18:38:43 -07003203 "src/math/sqrt-fma3-nr1fma1adj.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003204 "src/math/sqrt-fma3-nr2fma.c",
Marat Dukhanfda12b82019-11-21 12:27:59 -08003205]
3206
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003207AVX2_UKERNELS = [
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003208 "src/f32-raddexpminusmax/gen/avx2-p5-x64-acc2.c",
3209 "src/f32-raddexpminusmax/gen/avx2-p5-x64-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003210 "src/f32-raddexpminusmax/gen/avx2-p5-x64.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003211 "src/f32-raddexpminusmax/gen/avx2-p5-x72-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003212 "src/f32-raddexpminusmax/gen/avx2-p5-x72.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003213 "src/f32-raddexpminusmax/gen/avx2-p5-x80-acc2.c",
3214 "src/f32-raddexpminusmax/gen/avx2-p5-x80-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003215 "src/f32-raddexpminusmax/gen/avx2-p5-x80.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003216 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc2.c",
3217 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc3.c",
3218 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc6.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003219 "src/f32-raddexpminusmax/gen/avx2-p5-x96.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003220 "src/f32-raddextexp/gen/avx2-p5-x64-acc2.c",
3221 "src/f32-raddextexp/gen/avx2-p5-x64-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003222 "src/f32-raddextexp/gen/avx2-p5-x64.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003223 "src/f32-raddextexp/gen/avx2-p5-x72-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003224 "src/f32-raddextexp/gen/avx2-p5-x72.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003225 "src/f32-raddextexp/gen/avx2-p5-x80-acc2.c",
3226 "src/f32-raddextexp/gen/avx2-p5-x80-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003227 "src/f32-raddextexp/gen/avx2-p5-x80.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003228 "src/f32-raddextexp/gen/avx2-p5-x96-acc2.c",
3229 "src/f32-raddextexp/gen/avx2-p5-x96-acc3.c",
3230 "src/f32-raddextexp/gen/avx2-p5-x96-acc6.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003231 "src/f32-raddextexp/gen/avx2-p5-x96.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003232 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64-acc2.c",
3233 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003234 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003235 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x72-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003236 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x72.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003237 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80-acc2.c",
3238 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003239 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003240 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc2.c",
3241 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc3.c",
3242 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc6.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003243 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96.c",
Marat Dukhaned6baaf2020-12-01 15:07:08 -08003244 "src/f32-velu/gen/velu-avx2-rr1-lut4-p4-perm-x8.c",
3245 "src/f32-velu/gen/velu-avx2-rr1-lut4-p4-perm-x16.c",
3246 "src/f32-velu/gen/velu-avx2-rr1-lut4-p4-perm-x24.c",
3247 "src/f32-velu/gen/velu-avx2-rr1-lut4-p4-perm-x32.c",
3248 "src/f32-velu/gen/velu-avx2-rr1-lut4-p4-perm-x40.c",
3249 "src/f32-velu/gen/velu-avx2-rr1-lut4-p4-perm-x48.c",
3250 "src/f32-velu/gen/velu-avx2-rr1-lut4-p4-perm-x56.c",
3251 "src/f32-velu/gen/velu-avx2-rr1-lut4-p4-perm-x64.c",
3252 "src/f32-velu/gen/velu-avx2-rr1-lut4-p4-perm-x72.c",
3253 "src/f32-velu/gen/velu-avx2-rr1-lut4-p4-perm-x80.c",
3254 "src/f32-velu/gen/velu-avx2-rr1-lut8-p4-perm-x8.c",
3255 "src/f32-velu/gen/velu-avx2-rr1-lut8-p4-perm-x16.c",
3256 "src/f32-velu/gen/velu-avx2-rr1-lut8-p4-perm-x24.c",
3257 "src/f32-velu/gen/velu-avx2-rr1-lut8-p4-perm-x32.c",
3258 "src/f32-velu/gen/velu-avx2-rr1-lut8-p4-perm-x40.c",
3259 "src/f32-velu/gen/velu-avx2-rr1-lut8-p4-perm-x48.c",
3260 "src/f32-velu/gen/velu-avx2-rr1-lut8-p4-perm-x56.c",
3261 "src/f32-velu/gen/velu-avx2-rr1-lut8-p4-perm-x64.c",
3262 "src/f32-velu/gen/velu-avx2-rr1-lut8-p4-perm-x72.c",
3263 "src/f32-velu/gen/velu-avx2-rr1-lut8-p4-perm-x80.c",
3264 "src/f32-velu/gen/velu-avx2-rr1-lut16-p3-gather-x8.c",
3265 "src/f32-velu/gen/velu-avx2-rr1-lut16-p3-gather-x16.c",
3266 "src/f32-velu/gen/velu-avx2-rr1-lut16-p3-gather-x24.c",
3267 "src/f32-velu/gen/velu-avx2-rr1-lut16-p3-gather-x32.c",
3268 "src/f32-velu/gen/velu-avx2-rr1-lut16-p3-gather-x40.c",
3269 "src/f32-velu/gen/velu-avx2-rr1-lut16-p3-gather-x48.c",
3270 "src/f32-velu/gen/velu-avx2-rr1-lut16-p3-gather-x56.c",
3271 "src/f32-velu/gen/velu-avx2-rr1-lut16-p3-gather-x64.c",
3272 "src/f32-velu/gen/velu-avx2-rr1-lut16-p3-gather-x72.c",
3273 "src/f32-velu/gen/velu-avx2-rr1-lut16-p3-gather-x80.c",
3274 "src/f32-velu/gen/velu-avx2-rr1-p6-x8.c",
3275 "src/f32-velu/gen/velu-avx2-rr1-p6-x16.c",
3276 "src/f32-velu/gen/velu-avx2-rr1-p6-x24.c",
3277 "src/f32-velu/gen/velu-avx2-rr1-p6-x32.c",
3278 "src/f32-velu/gen/velu-avx2-rr1-p6-x40.c",
3279 "src/f32-velu/gen/velu-avx2-rr1-p6-x48.c",
3280 "src/f32-velu/gen/velu-avx2-rr1-p6-x56.c",
3281 "src/f32-velu/gen/velu-avx2-rr1-p6-x64.c",
3282 "src/f32-velu/gen/velu-avx2-rr1-p6-x72.c",
3283 "src/f32-velu/gen/velu-avx2-rr1-p6-x80.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003284 "src/f32-vscaleexpminusmax/gen/avx2-p5-x8.c",
3285 "src/f32-vscaleexpminusmax/gen/avx2-p5-x16.c",
3286 "src/f32-vscaleexpminusmax/gen/avx2-p5-x24.c",
3287 "src/f32-vscaleexpminusmax/gen/avx2-p5-x32.c",
3288 "src/f32-vscaleexpminusmax/gen/avx2-p5-x40.c",
3289 "src/f32-vscaleexpminusmax/gen/avx2-p5-x48.c",
3290 "src/f32-vscaleexpminusmax/gen/avx2-p5-x56.c",
3291 "src/f32-vscaleexpminusmax/gen/avx2-p5-x64.c",
3292 "src/f32-vscaleexpminusmax/gen/avx2-p5-x72.c",
3293 "src/f32-vscaleexpminusmax/gen/avx2-p5-x80.c",
3294 "src/f32-vscaleexpminusmax/gen/avx2-p5-x88.c",
3295 "src/f32-vscaleexpminusmax/gen/avx2-p5-x96.c",
3296 "src/f32-vscaleextexp/gen/avx2-p5-x8.c",
3297 "src/f32-vscaleextexp/gen/avx2-p5-x16.c",
3298 "src/f32-vscaleextexp/gen/avx2-p5-x24.c",
3299 "src/f32-vscaleextexp/gen/avx2-p5-x32.c",
3300 "src/f32-vscaleextexp/gen/avx2-p5-x40.c",
3301 "src/f32-vscaleextexp/gen/avx2-p5-x48.c",
3302 "src/f32-vscaleextexp/gen/avx2-p5-x56.c",
3303 "src/f32-vscaleextexp/gen/avx2-p5-x64.c",
3304 "src/f32-vscaleextexp/gen/avx2-p5-x72.c",
3305 "src/f32-vscaleextexp/gen/avx2-p5-x80.c",
3306 "src/f32-vscaleextexp/gen/avx2-p5-x88.c",
3307 "src/f32-vscaleextexp/gen/avx2-p5-x96.c",
Marat Dukhanb1eec082021-05-05 23:24:55 -07003308 "src/f32-vsigmoid/gen/vsigmoid-avx2-rr1-p5-div-x8.c",
3309 "src/f32-vsigmoid/gen/vsigmoid-avx2-rr1-p5-div-x16.c",
3310 "src/f32-vsigmoid/gen/vsigmoid-avx2-rr1-p5-div-x24.c",
3311 "src/f32-vsigmoid/gen/vsigmoid-avx2-rr1-p5-div-x32.c",
3312 "src/f32-vsigmoid/gen/vsigmoid-avx2-rr1-p5-div-x40.c",
3313 "src/f32-vsigmoid/gen/vsigmoid-avx2-rr1-p5-div-x48.c",
3314 "src/f32-vsigmoid/gen/vsigmoid-avx2-rr1-p5-div-x56.c",
3315 "src/f32-vsigmoid/gen/vsigmoid-avx2-rr1-p5-div-x64.c",
3316 "src/f32-vsigmoid/gen/vsigmoid-avx2-rr1-p5-div-x72.c",
3317 "src/f32-vsigmoid/gen/vsigmoid-avx2-rr1-p5-div-x80.c",
3318 "src/f32-vsigmoid/gen/vsigmoid-avx2-rr1-p5-nr1fma-x8.c",
3319 "src/f32-vsigmoid/gen/vsigmoid-avx2-rr1-p5-nr1fma-x16.c",
3320 "src/f32-vsigmoid/gen/vsigmoid-avx2-rr1-p5-nr1fma-x24.c",
3321 "src/f32-vsigmoid/gen/vsigmoid-avx2-rr1-p5-nr1fma-x32.c",
3322 "src/f32-vsigmoid/gen/vsigmoid-avx2-rr1-p5-nr1fma-x40.c",
3323 "src/f32-vsigmoid/gen/vsigmoid-avx2-rr1-p5-nr1fma-x48.c",
3324 "src/f32-vsigmoid/gen/vsigmoid-avx2-rr1-p5-nr1fma-x56.c",
3325 "src/f32-vsigmoid/gen/vsigmoid-avx2-rr1-p5-nr1fma-x64.c",
3326 "src/f32-vsigmoid/gen/vsigmoid-avx2-rr1-p5-nr1fma-x72.c",
3327 "src/f32-vsigmoid/gen/vsigmoid-avx2-rr1-p5-nr1fma-x80.c",
3328 "src/f32-vsigmoid/gen/vsigmoid-avx2-rr1-p5-nr2fma-x8.c",
3329 "src/f32-vsigmoid/gen/vsigmoid-avx2-rr1-p5-nr2fma-x16.c",
3330 "src/f32-vsigmoid/gen/vsigmoid-avx2-rr1-p5-nr2fma-x24.c",
3331 "src/f32-vsigmoid/gen/vsigmoid-avx2-rr1-p5-nr2fma-x32.c",
3332 "src/f32-vsigmoid/gen/vsigmoid-avx2-rr1-p5-nr2fma-x40.c",
3333 "src/f32-vsigmoid/gen/vsigmoid-avx2-rr1-p5-nr2fma-x48.c",
3334 "src/f32-vsigmoid/gen/vsigmoid-avx2-rr1-p5-nr2fma-x56.c",
3335 "src/f32-vsigmoid/gen/vsigmoid-avx2-rr1-p5-nr2fma-x64.c",
3336 "src/f32-vsigmoid/gen/vsigmoid-avx2-rr1-p5-nr2fma-x72.c",
3337 "src/f32-vsigmoid/gen/vsigmoid-avx2-rr1-p5-nr2fma-x80.c",
Marat Dukhanb7633f22020-11-20 16:34:56 -08003338 "src/math/exp-avx2-rr2-lut8-p3-perm.c",
3339 "src/math/exp-avx2-rr2-lut8-p4-perm.c",
3340 "src/math/exp-avx2-rr2-p5.c",
Marat Dukhande390d42020-11-29 19:32:18 -08003341 "src/math/expm1minus-avx2-rr1-lut4-p4-perm.c",
3342 "src/math/expm1minus-avx2-rr1-lut8-p4-perm.c",
3343 "src/math/expm1minus-avx2-rr1-lut16-p3-gather.c",
3344 "src/math/expm1minus-avx2-rr1-p6.c",
Frank Barcharde7223ee2020-12-04 19:04:01 -08003345 "src/math/expminus-avx2-rr2-p5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003346 "src/math/extexp-avx2-p5.c",
3347 "src/math/sigmoid-avx2-rr1-lut64-p2-gather-div.c",
3348 "src/math/sigmoid-avx2-rr1-lut64-p2-gather-nr1fma.c",
3349 "src/math/sigmoid-avx2-rr1-lut64-p2-gather-nr2fma.c",
3350 "src/math/sigmoid-avx2-rr1-lut64-p2-gather-nr2fma1adj.c",
3351 "src/math/sigmoid-avx2-rr1-p5-div.c",
3352 "src/math/sigmoid-avx2-rr1-p5-nr1fma.c",
3353 "src/math/sigmoid-avx2-rr1-p5-nr2fma.c",
3354 "src/math/sigmoid-avx2-rr2-lut64-p2-gather-div.c",
3355 "src/math/sigmoid-avx2-rr2-lut64-p2-gather-nr1fma.c",
3356 "src/math/sigmoid-avx2-rr2-lut64-p2-gather-nr2fma.c",
3357 "src/math/sigmoid-avx2-rr2-lut64-p2-gather-nr2fma1adj.c",
3358 "src/math/sigmoid-avx2-rr2-p5-div.c",
3359 "src/math/sigmoid-avx2-rr2-p5-nr1fma.c",
3360 "src/math/sigmoid-avx2-rr2-p5-nr2fma.c",
Marat Dukhand65a1522020-08-04 19:28:18 -07003361 "src/qs8-dwconv/gen/up8x9-minmax-avx2-mul32.c",
Marat Dukhan4ed14882021-05-12 17:50:40 -07003362 "src/qs8-dwconv/gen/up8x25-minmax-avx2-mul32.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003363 "src/qs8-dwconv/gen/up16x9-minmax-avx2-mul16.c",
Marat Dukhand65a1522020-08-04 19:28:18 -07003364 "src/qs8-dwconv/gen/up16x9-minmax-avx2-mul32.c",
Marat Dukhan4ed14882021-05-12 17:50:40 -07003365 "src/qs8-dwconv/gen/up16x25-minmax-avx2-mul16.c",
3366 "src/qs8-dwconv/gen/up16x25-minmax-avx2-mul32.c",
Marat Dukhand65a1522020-08-04 19:28:18 -07003367 "src/qs8-dwconv/gen/up24x9-minmax-avx2-mul32.c",
Marat Dukhan4ed14882021-05-12 17:50:40 -07003368 "src/qs8-dwconv/gen/up24x25-minmax-avx2-mul32.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003369 "src/qs8-dwconv/gen/up32x9-minmax-avx2-mul16.c",
Marat Dukhand65a1522020-08-04 19:28:18 -07003370 "src/qs8-dwconv/gen/up32x9-minmax-avx2-mul32.c",
Marat Dukhan4ed14882021-05-12 17:50:40 -07003371 "src/qs8-dwconv/gen/up32x25-minmax-avx2-mul16.c",
3372 "src/qs8-dwconv/gen/up32x25-minmax-avx2-mul32.c",
Marat Dukhan12809522020-08-02 22:23:51 -07003373 "src/qs8-gemm/gen/1x8c8-minmax-avx2.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07003374 "src/qs8-gemm/gen/1x8c8-xw-minmax-avx2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003375 "src/qs8-gemm/gen/2x8c8-minmax-avx2.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07003376 "src/qs8-gemm/gen/2x8c8-xw-minmax-avx2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003377 "src/qs8-gemm/gen/3x8c8-minmax-avx2.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07003378 "src/qs8-gemm/gen/3x8c8-xw-minmax-avx2.c",
Marat Dukhan12809522020-08-02 22:23:51 -07003379 "src/qs8-igemm/gen/1x8c8-minmax-avx2.c",
3380 "src/qs8-igemm/gen/2x8c8-minmax-avx2.c",
3381 "src/qs8-igemm/gen/3x8c8-minmax-avx2.c",
Marat Dukhane6dc0b62020-09-08 23:57:14 -07003382 "src/qs8-vadd/gen/minmax-avx2-mul32-ld64-x8.c",
3383 "src/qs8-vadd/gen/minmax-avx2-mul32-ld64-x16.c",
3384 "src/qs8-vadd/gen/minmax-avx2-mul32-ld64-x24.c",
3385 "src/qs8-vadd/gen/minmax-avx2-mul32-ld64-x32.c",
3386 "src/qs8-vaddc/gen/minmax-avx2-mul32-ld64-x8.c",
3387 "src/qs8-vaddc/gen/minmax-avx2-mul32-ld64-x16.c",
3388 "src/qs8-vaddc/gen/minmax-avx2-mul32-ld64-x24.c",
3389 "src/qs8-vaddc/gen/minmax-avx2-mul32-ld64-x32.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003390]
3391
Marat Dukhan08c4a432019-10-03 09:29:21 -07003392AVX512F_UKERNELS = [
Marat Dukhan1c587112020-04-08 20:04:28 -07003393 "src/f32-dwconv/gen/up16x4-minmax-avx512f-acc2.c",
3394 "src/f32-dwconv/gen/up16x4-minmax-avx512f.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07003395 "src/f32-dwconv/gen/up16x9-minmax-avx512f-acc2.c",
3396 "src/f32-dwconv/gen/up16x9-minmax-avx512f.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07003397 "src/f32-dwconv/gen/up16x25-minmax-avx512f-acc2.c",
3398 "src/f32-dwconv/gen/up16x25-minmax-avx512f.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003399 "src/f32-dwconv/gen/up32x4-minmax-avx512f-acc2.c",
3400 "src/f32-dwconv/gen/up32x4-minmax-avx512f.c",
3401 "src/f32-dwconv/gen/up32x9-minmax-avx512f-acc2.c",
3402 "src/f32-dwconv/gen/up32x9-minmax-avx512f.c",
3403 "src/f32-dwconv/gen/up32x25-minmax-avx512f-acc2.c",
3404 "src/f32-dwconv/gen/up32x25-minmax-avx512f.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07003405 "src/f32-gemm/gen-inc/1x16inc-minmax-avx512f-broadcast.c",
3406 "src/f32-gemm/gen-inc/4x16inc-minmax-avx512f-broadcast.c",
3407 "src/f32-gemm/gen-inc/5x16inc-minmax-avx512f-broadcast.c",
3408 "src/f32-gemm/gen-inc/6x16inc-minmax-avx512f-broadcast.c",
3409 "src/f32-gemm/gen-inc/7x16inc-minmax-avx512f-broadcast.c",
3410 "src/f32-gemm/gen-inc/8x16inc-minmax-avx512f-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003411 "src/f32-gemm/gen/1x16-minmax-avx512f-broadcast.c",
3412 "src/f32-gemm/gen/4x16-minmax-avx512f-broadcast.c",
3413 "src/f32-gemm/gen/5x16-minmax-avx512f-broadcast.c",
3414 "src/f32-gemm/gen/6x16-minmax-avx512f-broadcast.c",
3415 "src/f32-gemm/gen/7x16-minmax-avx512f-broadcast.c",
3416 "src/f32-gemm/gen/8x16-minmax-avx512f-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07003417 "src/f32-igemm/gen/1x16-minmax-avx512f-broadcast.c",
3418 "src/f32-igemm/gen/4x16-minmax-avx512f-broadcast.c",
3419 "src/f32-igemm/gen/5x16-minmax-avx512f-broadcast.c",
3420 "src/f32-igemm/gen/6x16-minmax-avx512f-broadcast.c",
3421 "src/f32-igemm/gen/7x16-minmax-avx512f-broadcast.c",
3422 "src/f32-igemm/gen/8x16-minmax-avx512f-broadcast.c",
Marat Dukhan90eca0a2020-03-11 00:52:23 -07003423 "src/f32-prelu/gen/avx512f-2x16.c",
3424 "src/f32-prelu/gen/avx512f-2x32.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003425 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128-acc2.c",
3426 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003427 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003428 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x144-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003429 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x144.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003430 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160-acc2.c",
3431 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003432 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003433 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc2.c",
3434 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc3.c",
3435 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc6.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003436 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003437 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128-acc2.c",
3438 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003439 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003440 "src/f32-raddextexp/gen/avx512f-p5-scalef-x144-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003441 "src/f32-raddextexp/gen/avx512f-p5-scalef-x144.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003442 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160-acc2.c",
3443 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003444 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003445 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc2.c",
3446 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc3.c",
3447 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc6.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003448 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003449 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128-acc2.c",
3450 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003451 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003452 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x144-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003453 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x144.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003454 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160-acc2.c",
3455 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003456 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003457 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc2.c",
3458 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc3.c",
3459 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc6.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003460 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003461 "src/f32-rmax/avx512f.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07003462 "src/f32-vbinary/gen/vadd-minmax-avx512f-x16.c",
3463 "src/f32-vbinary/gen/vadd-minmax-avx512f-x32.c",
3464 "src/f32-vbinary/gen/vaddc-minmax-avx512f-x16.c",
3465 "src/f32-vbinary/gen/vaddc-minmax-avx512f-x32.c",
3466 "src/f32-vbinary/gen/vdiv-minmax-avx512f-x16.c",
3467 "src/f32-vbinary/gen/vdiv-minmax-avx512f-x32.c",
3468 "src/f32-vbinary/gen/vdivc-minmax-avx512f-x16.c",
3469 "src/f32-vbinary/gen/vdivc-minmax-avx512f-x32.c",
Marat Dukhan9a88efe2019-12-10 15:54:24 -08003470 "src/f32-vbinary/gen/vmax-avx512f-x16.c",
3471 "src/f32-vbinary/gen/vmax-avx512f-x32.c",
3472 "src/f32-vbinary/gen/vmaxc-avx512f-x16.c",
3473 "src/f32-vbinary/gen/vmaxc-avx512f-x32.c",
3474 "src/f32-vbinary/gen/vmin-avx512f-x16.c",
3475 "src/f32-vbinary/gen/vmin-avx512f-x32.c",
3476 "src/f32-vbinary/gen/vminc-avx512f-x16.c",
3477 "src/f32-vbinary/gen/vminc-avx512f-x32.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07003478 "src/f32-vbinary/gen/vmul-minmax-avx512f-x16.c",
3479 "src/f32-vbinary/gen/vmul-minmax-avx512f-x32.c",
3480 "src/f32-vbinary/gen/vmulc-minmax-avx512f-x16.c",
3481 "src/f32-vbinary/gen/vmulc-minmax-avx512f-x32.c",
3482 "src/f32-vbinary/gen/vrdivc-minmax-avx512f-x16.c",
3483 "src/f32-vbinary/gen/vrdivc-minmax-avx512f-x32.c",
3484 "src/f32-vbinary/gen/vrsubc-minmax-avx512f-x16.c",
3485 "src/f32-vbinary/gen/vrsubc-minmax-avx512f-x32.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07003486 "src/f32-vbinary/gen/vsqrdiff-avx512f-x16.c",
3487 "src/f32-vbinary/gen/vsqrdiff-avx512f-x32.c",
3488 "src/f32-vbinary/gen/vsqrdiffc-avx512f-x16.c",
3489 "src/f32-vbinary/gen/vsqrdiffc-avx512f-x32.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07003490 "src/f32-vbinary/gen/vsub-minmax-avx512f-x16.c",
3491 "src/f32-vbinary/gen/vsub-minmax-avx512f-x32.c",
3492 "src/f32-vbinary/gen/vsubc-minmax-avx512f-x16.c",
3493 "src/f32-vbinary/gen/vsubc-minmax-avx512f-x32.c",
Marat Dukhanb1eec082021-05-05 23:24:55 -07003494 "src/f32-vclamp/gen/vclamp-avx512f-x16.c",
3495 "src/f32-vclamp/gen/vclamp-avx512f-x32.c",
Marat Dukhaned6baaf2020-12-01 15:07:08 -08003496 "src/f32-velu/gen/velu-avx512f-rr1-lut16-p3-perm-x16.c",
3497 "src/f32-velu/gen/velu-avx512f-rr1-lut16-p3-perm-x32.c",
3498 "src/f32-velu/gen/velu-avx512f-rr1-lut16-p3-perm-x48.c",
3499 "src/f32-velu/gen/velu-avx512f-rr1-lut16-p3-perm-x64.c",
3500 "src/f32-velu/gen/velu-avx512f-rr1-lut16-p3-perm-x80.c",
3501 "src/f32-velu/gen/velu-avx512f-rr1-lut16-p3-perm-x96.c",
3502 "src/f32-velu/gen/velu-avx512f-rr1-lut16-p3-perm-x112.c",
3503 "src/f32-velu/gen/velu-avx512f-rr1-lut16-p3-perm-x128.c",
3504 "src/f32-velu/gen/velu-avx512f-rr1-p6-x16.c",
3505 "src/f32-velu/gen/velu-avx512f-rr1-p6-x32.c",
3506 "src/f32-velu/gen/velu-avx512f-rr1-p6-x48.c",
3507 "src/f32-velu/gen/velu-avx512f-rr1-p6-x64.c",
3508 "src/f32-velu/gen/velu-avx512f-rr1-p6-x80.c",
3509 "src/f32-velu/gen/velu-avx512f-rr1-p6-x96.c",
3510 "src/f32-velu/gen/velu-avx512f-rr1-p6-x112.c",
3511 "src/f32-velu/gen/velu-avx512f-rr1-p6-x128.c",
Marat Dukhan6674d692021-05-05 22:27:00 -07003512 "src/f32-vhswish/gen/vhswish-avx512f-x16.c",
3513 "src/f32-vhswish/gen/vhswish-avx512f-x32.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07003514 "src/f32-vlrelu/gen/vlrelu-avx512f-x16.c",
3515 "src/f32-vlrelu/gen/vlrelu-avx512f-x32.c",
Marat Dukhanb1eec082021-05-05 23:24:55 -07003516 "src/f32-vrelu/gen/vrelu-avx512f-x16.c",
3517 "src/f32-vrelu/gen/vrelu-avx512f-x32.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003518 "src/f32-vrnd/gen/vrndd-avx512f-x16.c",
3519 "src/f32-vrnd/gen/vrndd-avx512f-x32.c",
3520 "src/f32-vrnd/gen/vrndne-avx512f-x16.c",
3521 "src/f32-vrnd/gen/vrndne-avx512f-x32.c",
3522 "src/f32-vrnd/gen/vrndu-avx512f-x16.c",
3523 "src/f32-vrnd/gen/vrndu-avx512f-x32.c",
3524 "src/f32-vrnd/gen/vrndz-avx512f-x16.c",
3525 "src/f32-vrnd/gen/vrndz-avx512f-x32.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07003526 "src/f32-vscale/avx512f-x64.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003527 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x16.c",
3528 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x32.c",
3529 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x48.c",
3530 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x64.c",
3531 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x80.c",
3532 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x96.c",
3533 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x112.c",
3534 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x128.c",
3535 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x144.c",
3536 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x160.c",
3537 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x176.c",
3538 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x192.c",
3539 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x16.c",
3540 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x32.c",
3541 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x48.c",
3542 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x64.c",
3543 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x80.c",
3544 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x96.c",
3545 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x112.c",
3546 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x128.c",
3547 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x144.c",
3548 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x160.c",
3549 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x176.c",
3550 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x192.c",
Marat Dukhanb1eec082021-05-05 23:24:55 -07003551 "src/f32-vsigmoid/gen/vsigmoid-avx512f-rr1-lut16-p3-perm-scalef-div-x16.c",
3552 "src/f32-vsigmoid/gen/vsigmoid-avx512f-rr1-lut16-p3-perm-scalef-div-x32.c",
3553 "src/f32-vsigmoid/gen/vsigmoid-avx512f-rr1-lut16-p3-perm-scalef-div-x48.c",
3554 "src/f32-vsigmoid/gen/vsigmoid-avx512f-rr1-lut16-p3-perm-scalef-div-x64.c",
3555 "src/f32-vsigmoid/gen/vsigmoid-avx512f-rr1-lut16-p3-perm-scalef-div-x80.c",
3556 "src/f32-vsigmoid/gen/vsigmoid-avx512f-rr1-lut16-p3-perm-scalef-div-x96.c",
3557 "src/f32-vsigmoid/gen/vsigmoid-avx512f-rr1-lut16-p3-perm-scalef-div-x112.c",
3558 "src/f32-vsigmoid/gen/vsigmoid-avx512f-rr1-lut16-p3-perm-scalef-div-x128.c",
3559 "src/f32-vsigmoid/gen/vsigmoid-avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x16.c",
3560 "src/f32-vsigmoid/gen/vsigmoid-avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x32.c",
3561 "src/f32-vsigmoid/gen/vsigmoid-avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x48.c",
3562 "src/f32-vsigmoid/gen/vsigmoid-avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x64.c",
3563 "src/f32-vsigmoid/gen/vsigmoid-avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x80.c",
3564 "src/f32-vsigmoid/gen/vsigmoid-avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x96.c",
3565 "src/f32-vsigmoid/gen/vsigmoid-avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x112.c",
3566 "src/f32-vsigmoid/gen/vsigmoid-avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x128.c",
3567 "src/f32-vsigmoid/gen/vsigmoid-avx512f-rr1-p5-scalef-div-x16.c",
3568 "src/f32-vsigmoid/gen/vsigmoid-avx512f-rr1-p5-scalef-div-x32.c",
3569 "src/f32-vsigmoid/gen/vsigmoid-avx512f-rr1-p5-scalef-div-x48.c",
3570 "src/f32-vsigmoid/gen/vsigmoid-avx512f-rr1-p5-scalef-div-x64.c",
3571 "src/f32-vsigmoid/gen/vsigmoid-avx512f-rr1-p5-scalef-div-x80.c",
3572 "src/f32-vsigmoid/gen/vsigmoid-avx512f-rr1-p5-scalef-div-x96.c",
3573 "src/f32-vsigmoid/gen/vsigmoid-avx512f-rr1-p5-scalef-div-x112.c",
3574 "src/f32-vsigmoid/gen/vsigmoid-avx512f-rr1-p5-scalef-div-x128.c",
3575 "src/f32-vsigmoid/gen/vsigmoid-avx512f-rr1-p5-scalef-nr1fma-x16.c",
3576 "src/f32-vsigmoid/gen/vsigmoid-avx512f-rr1-p5-scalef-nr1fma-x32.c",
3577 "src/f32-vsigmoid/gen/vsigmoid-avx512f-rr1-p5-scalef-nr1fma-x48.c",
3578 "src/f32-vsigmoid/gen/vsigmoid-avx512f-rr1-p5-scalef-nr1fma-x64.c",
3579 "src/f32-vsigmoid/gen/vsigmoid-avx512f-rr1-p5-scalef-nr1fma-x80.c",
3580 "src/f32-vsigmoid/gen/vsigmoid-avx512f-rr1-p5-scalef-nr1fma-x96.c",
3581 "src/f32-vsigmoid/gen/vsigmoid-avx512f-rr1-p5-scalef-nr1fma-x112.c",
3582 "src/f32-vsigmoid/gen/vsigmoid-avx512f-rr1-p5-scalef-nr1fma-x128.c",
3583 "src/f32-vsigmoid/gen/vsigmoid-avx512f-rr2-lut32-p2-perm2-scalef-div-x16.c",
3584 "src/f32-vsigmoid/gen/vsigmoid-avx512f-rr2-lut32-p2-perm2-scalef-div-x32.c",
3585 "src/f32-vsigmoid/gen/vsigmoid-avx512f-rr2-lut32-p2-perm2-scalef-div-x48.c",
3586 "src/f32-vsigmoid/gen/vsigmoid-avx512f-rr2-lut32-p2-perm2-scalef-div-x64.c",
3587 "src/f32-vsigmoid/gen/vsigmoid-avx512f-rr2-lut32-p2-perm2-scalef-div-x80.c",
3588 "src/f32-vsigmoid/gen/vsigmoid-avx512f-rr2-lut32-p2-perm2-scalef-div-x96.c",
3589 "src/f32-vsigmoid/gen/vsigmoid-avx512f-rr2-lut32-p2-perm2-scalef-div-x112.c",
3590 "src/f32-vsigmoid/gen/vsigmoid-avx512f-rr2-lut32-p2-perm2-scalef-div-x128.c",
3591 "src/f32-vsigmoid/gen/vsigmoid-avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x16.c",
3592 "src/f32-vsigmoid/gen/vsigmoid-avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x32.c",
3593 "src/f32-vsigmoid/gen/vsigmoid-avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x48.c",
3594 "src/f32-vsigmoid/gen/vsigmoid-avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x64.c",
3595 "src/f32-vsigmoid/gen/vsigmoid-avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x80.c",
3596 "src/f32-vsigmoid/gen/vsigmoid-avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x96.c",
3597 "src/f32-vsigmoid/gen/vsigmoid-avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x112.c",
3598 "src/f32-vsigmoid/gen/vsigmoid-avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x128.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07003599 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x16.c",
3600 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x32.c",
3601 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x48.c",
3602 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x64.c",
3603 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x80.c",
3604 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x96.c",
3605 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x112.c",
3606 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x128.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07003607 "src/f32-vunary/gen/vabs-avx512f-x16.c",
3608 "src/f32-vunary/gen/vabs-avx512f-x32.c",
3609 "src/f32-vunary/gen/vneg-avx512f-x16.c",
3610 "src/f32-vunary/gen/vneg-avx512f-x32.c",
3611 "src/f32-vunary/gen/vsqr-avx512f-x16.c",
3612 "src/f32-vunary/gen/vsqr-avx512f-x32.c",
Marat Dukhanb7633f22020-11-20 16:34:56 -08003613 "src/math/exp-avx512f-rr2-lut16-p3-perm-scalef.c",
3614 "src/math/exp-avx512f-rr2-lut16-p3-perm.c",
3615 "src/math/exp-avx512f-rr2-lut32-p2-perm2-scalef.c",
3616 "src/math/exp-avx512f-rr2-lut32-p2-perm2.c",
3617 "src/math/exp-avx512f-rr2-p5-scalef.c",
3618 "src/math/exp-avx512f-rr2-p5.c",
Marat Dukhande390d42020-11-29 19:32:18 -08003619 "src/math/expm1minus-avx512f-rr1-lut16-p3-perm.c",
3620 "src/math/expm1minus-avx512f-rr1-p6.c",
Marat Dukhan98ba4412019-10-23 02:14:28 -07003621 "src/math/extexp-avx512f-p5.c",
Marat Dukhan6a34c5f2020-09-22 21:44:15 -07003622 "src/math/sigmoid-avx512f-rr1-lut16-p3-perm-scalef-div.c",
Marat Dukhan6a34c5f2020-09-22 21:44:15 -07003623 "src/math/sigmoid-avx512f-rr1-lut16-p3-perm-scalef-nr1fma.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003624 "src/math/sigmoid-avx512f-rr1-lut16-p3-perm-scalef-nr1fma1adj.c",
Marat Dukhan6a34c5f2020-09-22 21:44:15 -07003625 "src/math/sigmoid-avx512f-rr1-lut32-p2-perm2-scalef-div.c",
Marat Dukhan6a34c5f2020-09-22 21:44:15 -07003626 "src/math/sigmoid-avx512f-rr1-lut32-p2-perm2-scalef-nr1fma.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003627 "src/math/sigmoid-avx512f-rr1-lut32-p2-perm2-scalef-nr1fma1adj.c",
Marat Dukhan36173d22020-10-15 17:14:26 -07003628 "src/math/sigmoid-avx512f-rr1-lut64-p2-gather-scalef-div.c",
Marat Dukhan36173d22020-10-15 17:14:26 -07003629 "src/math/sigmoid-avx512f-rr1-lut64-p2-gather-scalef-nr1fma.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003630 "src/math/sigmoid-avx512f-rr1-lut64-p2-gather-scalef-nr1fma1adj.c",
3631 "src/math/sigmoid-avx512f-rr1-p5-scalef-div.c",
3632 "src/math/sigmoid-avx512f-rr1-p5-scalef-nr1fma.c",
3633 "src/math/sigmoid-avx512f-rr1-p5-scalef-nr1fma1adj.c",
3634 "src/math/sigmoid-avx512f-rr2-lut16-p3-perm-scalef-div.c",
3635 "src/math/sigmoid-avx512f-rr2-lut16-p3-perm-scalef-nr1fma.c",
3636 "src/math/sigmoid-avx512f-rr2-lut16-p3-perm-scalef-nr1fma1adj.c",
3637 "src/math/sigmoid-avx512f-rr2-lut32-p2-perm2-scalef-div.c",
3638 "src/math/sigmoid-avx512f-rr2-lut32-p2-perm2-scalef-nr1fma.c",
3639 "src/math/sigmoid-avx512f-rr2-lut32-p2-perm2-scalef-nr1fma1adj.c",
Marat Dukhan36173d22020-10-15 17:14:26 -07003640 "src/math/sigmoid-avx512f-rr2-lut64-p2-gather-scalef-div.c",
Marat Dukhan36173d22020-10-15 17:14:26 -07003641 "src/math/sigmoid-avx512f-rr2-lut64-p2-gather-scalef-nr1fma.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003642 "src/math/sigmoid-avx512f-rr2-lut64-p2-gather-scalef-nr1fma1adj.c",
3643 "src/math/sigmoid-avx512f-rr2-p5-scalef-div.c",
3644 "src/math/sigmoid-avx512f-rr2-p5-scalef-nr1fma.c",
3645 "src/math/sigmoid-avx512f-rr2-p5-scalef-nr1fma1adj.c",
Marat Dukhan84000762020-06-29 18:38:43 -07003646 "src/math/sqrt-avx512f-nr1fma.c",
Marat Dukhan84000762020-06-29 18:38:43 -07003647 "src/math/sqrt-avx512f-nr1fma1adj.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003648 "src/math/sqrt-avx512f-nr2fma.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003649]
3650
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07003651AVX512SKX_UKERNELS = [
Marat Dukhan2ffc5e62020-09-06 22:33:38 -07003652 "src/qs8-dwconv/gen/up16x9-minmax-avx512skx-mul32.c",
Marat Dukhan4ed14882021-05-12 17:50:40 -07003653 "src/qs8-dwconv/gen/up16x25-minmax-avx512skx-mul32.c",
Marat Dukhan2ffc5e62020-09-06 22:33:38 -07003654 "src/qs8-dwconv/gen/up32x9-minmax-avx512skx-mul32.c",
Marat Dukhan4ed14882021-05-12 17:50:40 -07003655 "src/qs8-dwconv/gen/up32x25-minmax-avx512skx-mul32.c",
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07003656 "src/qs8-gemm/gen/1x16c8-minmax-avx512skx.c",
3657 "src/qs8-gemm/gen/2x16c8-minmax-avx512skx.c",
3658 "src/qs8-gemm/gen/3x16c8-minmax-avx512skx.c",
3659 "src/qs8-gemm/gen/4x16c8-minmax-avx512skx.c",
3660 "src/qs8-igemm/gen/1x16c8-minmax-avx512skx.c",
3661 "src/qs8-igemm/gen/2x16c8-minmax-avx512skx.c",
3662 "src/qs8-igemm/gen/3x16c8-minmax-avx512skx.c",
3663 "src/qs8-igemm/gen/4x16c8-minmax-avx512skx.c",
3664]
3665
Frank Barchardbcedc082020-08-17 18:00:51 -07003666WASM32_ASM_UKERNELS = [
Marat Dukhan6674d692021-05-05 22:27:00 -07003667 "src/f32-vrelu/wasm_shr_x1.S",
3668 "src/f32-vrelu/wasm_shr_x2.S",
3669 "src/f32-vrelu/wasm_shr_x4.S",
Frank Barchardbcedc082020-08-17 18:00:51 -07003670]
3671
Marat Dukhan08c4a432019-10-03 09:29:21 -07003672AARCH32_ASM_UKERNELS = [
Marat Dukhan32f93812020-05-17 20:31:21 -07003673 "src/f32-gemm/4x4-aarch32-vfp-ld64.S",
Marat Dukhan3b98f6b2020-05-17 10:09:22 -07003674 "src/f32-gemm/4x4-minmax-aarch32-vfp-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003675 "src/f32-gemm/4x8-minmax-aarch32-neon-cortex-a53.S",
3676 "src/f32-gemm/4x8-minmax-aarch32-neon-cortex-a55.S",
Frank Barchard490febe2020-07-16 18:42:17 -07003677 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-cortex-a7.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003678 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-cortex-a75.S",
Frank Barchard569561d2020-06-17 13:11:12 -07003679 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-ld64.S",
Frank Barchard490febe2020-07-16 18:42:17 -07003680 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-pld-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003681 "src/f32-igemm/4x8-minmax-aarch32-neon-cortex-a53.S",
3682 "src/f32-igemm/4x8-minmax-aarch32-neon-cortex-a55.S",
Frank Barchard490febe2020-07-16 18:42:17 -07003683 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-cortex-a7.S",
3684 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-cortex-a75.S",
3685 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-ld64.S",
3686 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-pld-cortex-a75.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003687]
3688
3689AARCH64_ASM_UKERNELS = [
Frank Barchardbddfbcd2020-04-15 12:32:41 -07003690 "src/f16-gemm/gen-inc/1x8inc-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchard04336c12020-10-22 16:48:55 -07003691 "src/f16-gemm/gen-inc/1x16inc-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchardbddfbcd2020-04-15 12:32:41 -07003692 "src/f16-gemm/gen-inc/4x8inc-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchard04336c12020-10-22 16:48:55 -07003693 "src/f16-gemm/gen-inc/4x16inc-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchardbddfbcd2020-04-15 12:32:41 -07003694 "src/f16-gemm/gen-inc/6x8inc-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchard04336c12020-10-22 16:48:55 -07003695 "src/f16-gemm/gen-inc/6x16inc-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchard3b8e5662020-04-20 12:12:53 -07003696 "src/f16-gemm/gen-inc/8x8inc-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchard04336c12020-10-22 16:48:55 -07003697 "src/f16-gemm/gen/1x8-minmax-aarch64-neonfp16arith-ld64.S",
3698 "src/f16-gemm/gen/1x16-minmax-aarch64-neonfp16arith-ld32.S",
3699 "src/f16-gemm/gen/4x8-minmax-aarch64-neonfp16arith-ld64.S",
3700 "src/f16-gemm/gen/4x16-minmax-aarch64-neonfp16arith-ld32.S",
3701 "src/f16-gemm/gen/6x8-minmax-aarch64-neonfp16arith-ld64.S",
3702 "src/f16-gemm/gen/6x16-minmax-aarch64-neonfp16arith-ld32.S",
3703 "src/f16-gemm/gen/8x8-minmax-aarch64-neonfp16arith-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003704 "src/f32-dwconv/up4x9-minmax-aarch64-neonfma-cortex-a55.S",
3705 "src/f32-dwconv/up4x9-minmax-aarch64-neonfma.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003706 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a53.S",
3707 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a57.S",
3708 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a75.S",
Frank Barchard04336c12020-10-22 16:48:55 -07003709 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-ld64.S",
3710 "src/f32-gemm/gen-inc/1x12inc-minmax-aarch64-neonfma-cortex-a53.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003711 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a53.S",
3712 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a55.S",
3713 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a57.S",
3714 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003715 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-ld64.S",
Frank Barchard04336c12020-10-22 16:48:55 -07003716 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-ld128.S",
3717 "src/f32-gemm/gen-inc/4x12inc-minmax-aarch64-neonfma-cortex-a53.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003718 "src/f32-gemm/gen-inc/5x8inc-minmax-aarch64-neonfma-cortex-a57.S",
3719 "src/f32-gemm/gen-inc/5x8inc-minmax-aarch64-neonfma-cortex-a75.S",
3720 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a53.S",
3721 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a55.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003722 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a57.S",
Frank Barchard04336c12020-10-22 16:48:55 -07003723 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a73.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003724 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003725 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-ld64.S",
Frank Barchard04336c12020-10-22 16:48:55 -07003726 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-ld128.S",
3727 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a53.S",
3728 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a57.S",
3729 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a75.S",
3730 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-ld64.S",
3731 "src/f32-gemm/gen/1x12-minmax-aarch64-neonfma-cortex-a53.S",
3732 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a53.S",
3733 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a55.S",
3734 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a57.S",
3735 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a75.S",
3736 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-ld64.S",
3737 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-ld128.S",
3738 "src/f32-gemm/gen/4x12-minmax-aarch64-neonfma-cortex-a53.S",
3739 "src/f32-gemm/gen/5x8-minmax-aarch64-neonfma-cortex-a57.S",
3740 "src/f32-gemm/gen/5x8-minmax-aarch64-neonfma-cortex-a75.S",
3741 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a53.S",
3742 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a55.S",
3743 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a57.S",
3744 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a73.S",
3745 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a75.S",
3746 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-ld64.S",
3747 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-ld128.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003748 "src/f32-igemm/1x8-minmax-aarch64-neonfma-cortex-a53.S",
Frank Barchard04336c12020-10-22 16:48:55 -07003749 "src/f32-igemm/1x12-minmax-aarch64-neonfma-cortex-a53.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003750 "src/f32-igemm/4x8-minmax-aarch64-neonfma-cortex-a53.S",
3751 "src/f32-igemm/4x8-minmax-aarch64-neonfma-cortex-a55.S",
Frank Barchard04336c12020-10-22 16:48:55 -07003752 "src/f32-igemm/4x12-minmax-aarch64-neonfma-cortex-a53.S",
3753 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a53.S",
3754 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a55.S",
3755 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a73.S",
3756 "src/f32-igemm/gen/1x8-minmax-aarch64-neonfma-cortex-a57.S",
3757 "src/f32-igemm/gen/1x8-minmax-aarch64-neonfma-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003758 "src/f32-igemm/gen/4x8-minmax-aarch64-neonfma-cortex-a57.S",
3759 "src/f32-igemm/gen/4x8-minmax-aarch64-neonfma-cortex-a75.S",
3760 "src/f32-igemm/gen/5x8-minmax-aarch64-neonfma-cortex-a57.S",
3761 "src/f32-igemm/gen/5x8-minmax-aarch64-neonfma-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003762 "src/f32-igemm/gen/6x8-minmax-aarch64-neonfma-cortex-a57.S",
3763 "src/f32-igemm/gen/6x8-minmax-aarch64-neonfma-cortex-a75.S",
Frank Barchard4a4be4e2021-03-01 15:00:28 -08003764 "src/qs8-gemm/1x16c4-aarch64-neondot-ld32.S",
Frank Barchardf1fd89e2020-10-06 23:27:11 -07003765 "src/qs8-gemm/1x16c4-aarch64-neondot-ld64.S",
Frank Barchard1dc9fef2021-03-10 23:50:19 -08003766 "src/qs8-gemm/2x8c8-aarch64-neon-mull-padal.S",
Frank Barchard5655cb72021-03-09 20:57:14 -08003767 "src/qs8-gemm/2x8c16-aarch64-neon-mlal-padal.S",
Frank Barchard671d1b02021-03-10 15:42:08 -08003768 "src/qs8-gemm/4x16c4-aarch64-neondot-cortex-a55.S",
Frank Barchardda78da12021-03-02 14:28:00 -08003769 "src/qs8-gemm/4x16c4-aarch64-neondot-ld32.S",
Frank Barchard671d1b02021-03-10 15:42:08 -08003770 "src/qs8-gemm/4x16c4-aarch64-neondot-ld64.S",
Frank Barchard938ea812021-05-07 13:41:35 -07003771 "src/qs8-gemm/gen/1x8c8-aarch64-neon-mlal-padal.S",
3772 "src/qs8-gemm/gen/1x8c8-aarch64-neon-mlal-padal-prfm.S",
Frank Barcharddddb38f2021-05-03 16:51:48 -07003773 "src/qs8-gemm/gen/1x8c8-aarch64-neon-mlal-padal-cortex-a53.S",
3774 "src/qs8-gemm/gen/1x8c8-aarch64-neon-mlal-padal-prfm-cortex-a53.S",
3775 "src/qs8-gemm/gen/2x8c8-aarch64-neon-mlal-padal-cortex-a53.S",
3776 "src/qs8-gemm/gen/2x8c8-aarch64-neon-mlal-padal-prfm-cortex-a53.S",
Frank Barchard4a352042021-04-13 15:52:08 -07003777 "src/qs8-gemm/gen/2x8c8-aarch64-neon-mlal-padal-prfm.S",
3778 "src/qs8-gemm/gen/2x8c8-aarch64-neon-mlal-padal.S",
Frank Barchard7ca54df2021-03-12 00:09:02 -08003779 "src/qs8-igemm/2x8c16-aarch64-neon-mlal-padal.S",
Frank Barchard89e12f82021-03-08 11:42:44 -08003780 "src/qs8-igemm/4x16c4-aarch64-neondot-cortex-a55.S",
Frank Barchard671d1b02021-03-10 15:42:08 -08003781 "src/qs8-igemm/4x16c4-aarch64-neondot-ld64.S",
Frank Barchard938ea812021-05-07 13:41:35 -07003782 "src/qs8-igemm/gen/1x8c8-aarch64-neon-mlal-padal.S",
3783 "src/qs8-igemm/gen/1x8c8-aarch64-neon-mlal-padal-prfm.S",
Frank Barcharddddb38f2021-05-03 16:51:48 -07003784 "src/qs8-igemm/gen/1x8c8-aarch64-neon-mlal-padal-cortex-a53.S",
3785 "src/qs8-igemm/gen/1x8c8-aarch64-neon-mlal-padal-prfm-cortex-a53.S",
3786 "src/qs8-igemm/gen/2x8c8-aarch64-neon-mlal-padal-cortex-a53.S",
3787 "src/qs8-igemm/gen/2x8c8-aarch64-neon-mlal-padal-prfm-cortex-a53.S",
Frank Barchard4a352042021-04-13 15:52:08 -07003788 "src/qs8-igemm/gen/2x8c8-aarch64-neon-mlal-padal-prfm.S",
3789 "src/qs8-igemm/gen/2x8c8-aarch64-neon-mlal-padal.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003790]
3791
Marat Dukhan1b354632020-03-23 12:50:22 -07003792INTERNAL_MICROKERNEL_HDRS = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003793 "src/xnnpack/argmaxpool.h",
3794 "src/xnnpack/avgpool.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003795 "src/xnnpack/common.h",
3796 "src/xnnpack/conv.h",
Yury Kartynnike7841862020-11-04 18:22:18 -08003797 "src/xnnpack/depthtospace.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003798 "src/xnnpack/dwconv.h",
Frank Barchard04336c12020-10-22 16:48:55 -07003799 "src/xnnpack/fill.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003800 "src/xnnpack/gavgpool.h",
3801 "src/xnnpack/gemm.h",
Marat Dukhan660fd192020-03-10 04:55:30 -07003802 "src/xnnpack/ibilinear.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003803 "src/xnnpack/igemm.h",
Marat Dukhancfb31342019-12-05 10:42:57 -08003804 "src/xnnpack/intrinsics-polyfill.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003805 "src/xnnpack/lut.h",
3806 "src/xnnpack/math.h",
3807 "src/xnnpack/maxpool.h",
3808 "src/xnnpack/packx.h",
3809 "src/xnnpack/pad.h",
3810 "src/xnnpack/params.h",
3811 "src/xnnpack/pavgpool.h",
3812 "src/xnnpack/ppmm.h",
3813 "src/xnnpack/prelu.h",
Marat Dukhan97579532019-10-18 16:40:39 -07003814 "src/xnnpack/raddexpminusmax.h",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07003815 "src/xnnpack/raddextexp.h",
Marat Dukhan97579532019-10-18 16:40:39 -07003816 "src/xnnpack/raddstoreexpminusmax.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003817 "src/xnnpack/rmax.h",
3818 "src/xnnpack/scalar-utils.h",
3819 "src/xnnpack/spmm.h",
3820 "src/xnnpack/unpool.h",
3821 "src/xnnpack/vadd.h",
Marat Dukhan1e782c42019-11-21 17:02:40 -08003822 "src/xnnpack/vbinary.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003823 "src/xnnpack/vmulcaddc.h",
Marat Dukhan05ac8e32019-10-21 15:39:33 -07003824 "src/xnnpack/vscale.h",
Marat Dukhan97579532019-10-18 16:40:39 -07003825 "src/xnnpack/vscaleexpminusmax.h",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07003826 "src/xnnpack/vscaleextexp.h",
Marat Dukhan1e782c42019-11-21 17:02:40 -08003827 "src/xnnpack/vunary.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003828 "src/xnnpack/zip.h",
Marat Dukhan1b354632020-03-23 12:50:22 -07003829]
3830
3831INTERNAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003832 "include/xnnpack.h",
3833 "src/xnnpack/allocator.h",
3834 "src/xnnpack/compute.h",
3835 "src/xnnpack/im2col.h",
3836 "src/xnnpack/indirection.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003837 "src/xnnpack/math-stubs.h",
Chao Mei6ddfc602020-05-13 22:29:36 -07003838 "src/xnnpack/memory-planner.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003839 "src/xnnpack/operator.h",
3840 "src/xnnpack/pack.h",
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07003841 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003842 "src/xnnpack/requantization-stubs.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003843 "src/xnnpack/requantization.h",
Marat Dukhan1d75a542020-02-03 12:23:01 -08003844 "src/xnnpack/subgraph.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003845]
3846
Marat Dukhan1b354632020-03-23 12:50:22 -07003847ACCURACY_EVAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003848 "src/xnnpack/math-stubs.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003849]
3850
Marat Dukhan1b354632020-03-23 12:50:22 -07003851MICROKERNEL_BENCHMARK_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003852 "include/xnnpack.h",
Frank Barchard35db7d02020-10-26 13:37:34 -07003853 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003854]
3855
Marat Dukhan1b354632020-03-23 12:50:22 -07003856MICROKERNEL_TEST_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Frank Barchard35db7d02020-10-26 13:37:34 -07003857 "include/xnnpack.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003858 "src/xnnpack/isa-checks.h",
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07003859 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003860 "src/xnnpack/requantization.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003861]
3862
3863OPERATOR_TEST_PARAMS_HDRS = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003864 "src/xnnpack/common.h",
Frank Barchard04336c12020-10-22 16:48:55 -07003865 "src/xnnpack/params.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003866]
3867
3868WEIGHTS_PACK_HDRS = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003869 "src/xnnpack/compute.h",
Frank Barchard04336c12020-10-22 16:48:55 -07003870 "src/xnnpack/operator.h",
3871 "src/xnnpack/pack.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003872]
3873
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07003874LOGGING_COPTS = select({
3875 # No logging in optimized mode
3876 ":optimized_build": ["-DXNN_LOG_LEVEL=0"],
3877 # Full logging in debug mode
3878 ":debug_build": ["-DXNN_LOG_LEVEL=5"],
3879 # Error-only logging in default (fastbuild) mode
3880 "//conditions:default": ["-DXNN_LOG_LEVEL=2"],
3881})
3882
Marat Dukhan3b59de22020-06-03 20:15:19 -07003883LOGGING_SRCS = select({
3884 # No logging in optimized mode
3885 ":optimized_build": [],
3886 "//conditions:default": [
Marat Dukhanccd3a1d2021-03-29 16:03:12 -07003887 "src/datatype-strings.c",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003888 "src/operator-strings.c",
3889 "src/subgraph-strings.c",
3890 ],
3891})
3892
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07003893LOGGING_HDRS = [
3894 "src/xnnpack/log.h",
3895]
3896
Marat Dukhan08c4a432019-10-03 09:29:21 -07003897xnnpack_cc_library(
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003898 name = "tables",
3899 srcs = TABLE_SRCS,
3900 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003901 gcc_copts = xnnpack_gcc_std_copts(),
3902 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003903)
3904
3905xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003906 name = "scalar_ukernels",
3907 srcs = SCALAR_UKERNELS,
3908 hdrs = INTERNAL_HDRS,
3909 aarch32_copts = ["-marm"],
Marat Dukhan10a38082020-04-17 03:58:35 -07003910 gcc_copts = xnnpack_gcc_std_copts(),
3911 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003912 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003913 ":tables",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003914 "@FP16",
3915 "@FXdiv",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003916 "@pthreadpool",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003917 ],
3918)
3919
3920xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003921 name = "scalar_ukernels_test_mode",
3922 srcs = SCALAR_UKERNELS,
3923 hdrs = INTERNAL_HDRS,
3924 aarch32_copts = ["-marm"],
3925 copts = [
3926 "-UNDEBUG",
3927 "-DXNN_TEST_MODE=1",
3928 ],
3929 gcc_copts = xnnpack_gcc_std_copts(),
3930 msvc_copts = xnnpack_msvc_std_copts(),
3931 deps = [
3932 ":tables",
3933 "@FP16",
3934 "@FXdiv",
3935 "@pthreadpool",
3936 ],
3937)
3938
3939xnnpack_cc_library(
Marat Dukhan436ebe62019-12-04 15:10:12 -08003940 name = "wasm_ukernels",
3941 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003942 gcc_copts = xnnpack_gcc_std_copts(),
3943 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan436ebe62019-12-04 15:10:12 -08003944 wasm_srcs = WASM_UKERNELS,
Marat Dukhan1c6e3892020-06-25 23:56:10 -07003945 wasmsimd_srcs = WASM_UKERNELS + WASMSIMD_UKERNELS,
Marat Dukhan436ebe62019-12-04 15:10:12 -08003946 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003947 ":tables",
Marat Dukhan436ebe62019-12-04 15:10:12 -08003948 "@FP16",
3949 "@FXdiv",
3950 "@pthreadpool",
3951 ],
3952)
3953
3954xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003955 name = "wasm_ukernels_test_mode",
3956 hdrs = INTERNAL_HDRS,
3957 copts = [
3958 "-UNDEBUG",
3959 "-DXNN_TEST_MODE=1",
3960 ],
3961 gcc_copts = xnnpack_gcc_std_copts(),
3962 msvc_copts = xnnpack_msvc_std_copts(),
3963 wasm_srcs = WASM_UKERNELS,
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07003964 wasmsimd_srcs = WASM_UKERNELS + WASMSIMD_UKERNELS,
Marat Dukhan33fcf782020-05-24 14:27:15 -07003965 deps = [
3966 ":tables",
3967 "@FP16",
3968 "@FXdiv",
3969 "@pthreadpool",
3970 ],
3971)
3972
3973xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003974 name = "neon_ukernels",
3975 hdrs = INTERNAL_HDRS,
3976 aarch32_copts = [
3977 "-marm",
Marat Dukhan8853b822020-05-07 12:19:01 -07003978 "-march=armv7-a",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003979 "-mfpu=neon",
3980 ],
3981 aarch32_srcs = NEON_UKERNELS,
3982 aarch64_srcs = NEON_UKERNELS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003983 gcc_copts = xnnpack_gcc_std_copts(),
3984 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08003985 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003986 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003987 "@FP16",
3988 "@pthreadpool",
3989 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003990)
3991
3992xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003993 name = "neon_ukernels_test_mode",
3994 hdrs = INTERNAL_HDRS,
3995 aarch32_copts = [
3996 "-marm",
3997 "-march=armv7-a",
3998 "-mfpu=neon",
3999 ],
4000 aarch32_srcs = NEON_UKERNELS,
4001 aarch64_srcs = NEON_UKERNELS,
4002 copts = [
4003 "-UNDEBUG",
4004 "-DXNN_TEST_MODE=1",
4005 ],
4006 gcc_copts = xnnpack_gcc_std_copts(),
4007 msvc_copts = xnnpack_msvc_std_copts(),
4008 deps = [
4009 ":tables",
4010 "@FP16",
4011 "@pthreadpool",
4012 ],
4013)
4014
4015xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004016 name = "neonfma_ukernels",
4017 hdrs = INTERNAL_HDRS,
4018 aarch32_copts = [
4019 "-marm",
Marat Dukhan8853b822020-05-07 12:19:01 -07004020 "-march=armv7-a",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004021 "-mfpu=neon-vfpv4",
4022 ],
4023 aarch32_srcs = NEONFMA_UKERNELS,
4024 aarch64_srcs = NEONFMA_UKERNELS + AARCH64_NEONFMA_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07004025 apple_aarch32_copts = [
4026 "-mcpu=swift",
4027 "-mtune=generic",
4028 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07004029 gcc_copts = xnnpack_gcc_std_copts(),
4030 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08004031 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08004032 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08004033 "@FP16",
4034 "@pthreadpool",
4035 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004036)
4037
4038xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07004039 name = "neonfma_ukernels_test_mode",
4040 hdrs = INTERNAL_HDRS,
4041 aarch32_copts = [
4042 "-marm",
4043 "-march=armv7-a",
4044 "-mfpu=neon-vfpv4",
4045 ],
4046 aarch32_srcs = NEONFMA_UKERNELS,
4047 aarch64_srcs = NEONFMA_UKERNELS + AARCH64_NEONFMA_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07004048 apple_aarch32_copts = [
4049 "-mcpu=swift",
4050 "-mtune=generic",
4051 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07004052 copts = [
4053 "-UNDEBUG",
4054 "-DXNN_TEST_MODE=1",
4055 ],
4056 gcc_copts = xnnpack_gcc_std_copts(),
4057 msvc_copts = xnnpack_msvc_std_copts(),
4058 deps = [
4059 ":tables",
4060 "@FP16",
4061 "@pthreadpool",
4062 ],
4063)
4064
4065xnnpack_cc_library(
Marat Dukhan8853b822020-05-07 12:19:01 -07004066 name = "neonv8_ukernels",
4067 hdrs = INTERNAL_HDRS,
4068 aarch32_copts = [
4069 "-marm",
4070 "-march=armv8-a",
4071 "-mfpu=neon-fp-armv8",
4072 ],
4073 aarch32_srcs = NEONV8_UKERNELS,
4074 aarch64_srcs = NEONV8_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07004075 apple_aarch32_copts = [
4076 "-mcpu=cyclone",
4077 "-mtune=generic",
4078 ],
Marat Dukhan8853b822020-05-07 12:19:01 -07004079 gcc_copts = xnnpack_gcc_std_copts(),
4080 msvc_copts = xnnpack_msvc_std_copts(),
4081 deps = [
4082 ":tables",
4083 "@FP16",
4084 "@pthreadpool",
4085 ],
4086)
4087
4088xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07004089 name = "neonv8_ukernels_test_mode",
4090 hdrs = INTERNAL_HDRS,
4091 aarch32_copts = [
4092 "-marm",
4093 "-march=armv8-a",
4094 "-mfpu=neon-fp-armv8",
4095 ],
4096 aarch32_srcs = NEONV8_UKERNELS,
4097 aarch64_srcs = NEONV8_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07004098 apple_aarch32_copts = [
4099 "-mcpu=cyclone",
4100 "-mtune=generic",
4101 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07004102 copts = [
4103 "-UNDEBUG",
4104 "-DXNN_TEST_MODE=1",
4105 ],
4106 gcc_copts = xnnpack_gcc_std_copts(),
4107 msvc_copts = xnnpack_msvc_std_copts(),
4108 deps = [
4109 ":tables",
4110 "@FP16",
4111 "@pthreadpool",
4112 ],
4113)
4114
4115xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004116 name = "neonfp16arith_ukernels",
4117 hdrs = INTERNAL_HDRS,
4118 aarch64_copts = ["-march=armv8.2-a+fp16"],
4119 aarch64_srcs = AARCH64_NEONFP16ARITH_UKERNELS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004120 gcc_copts = xnnpack_gcc_std_copts(),
4121 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08004122 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08004123 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08004124 "@FP16",
4125 "@pthreadpool",
4126 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004127)
4128
4129xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07004130 name = "neonfp16arith_ukernels_test_mode",
4131 hdrs = INTERNAL_HDRS,
4132 aarch64_copts = ["-march=armv8.2-a+fp16"],
4133 aarch64_srcs = AARCH64_NEONFP16ARITH_UKERNELS,
4134 copts = [
4135 "-UNDEBUG",
4136 "-DXNN_TEST_MODE=1",
4137 ],
4138 gcc_copts = xnnpack_gcc_std_copts(),
4139 msvc_copts = xnnpack_msvc_std_copts(),
4140 deps = [
4141 ":tables",
4142 "@FP16",
4143 "@pthreadpool",
4144 ],
4145)
4146
4147xnnpack_cc_library(
Benoit Jacoba9644732020-08-13 12:48:55 -07004148 name = "neondot_ukernels",
4149 hdrs = INTERNAL_HDRS,
Marat Dukhan799ac752020-08-13 16:08:03 -07004150 aarch32_copts = [
4151 "-marm",
4152 "-march=armv8.2-a+dotprod",
4153 "-mfpu=neon-fp-armv8",
4154 ],
Benoit Jacoba9644732020-08-13 12:48:55 -07004155 aarch32_srcs = NEONDOT_UKERNELS,
4156 aarch64_copts = ["-march=armv8.2-a+dotprod"],
4157 aarch64_srcs = NEONDOT_UKERNELS,
4158 gcc_copts = xnnpack_gcc_std_copts(),
4159 msvc_copts = xnnpack_msvc_std_copts(),
4160 deps = [
4161 ":tables",
4162 "@FP16",
4163 "@pthreadpool",
4164 ],
4165)
4166
4167xnnpack_cc_library(
4168 name = "neondot_ukernels_test_mode",
4169 hdrs = INTERNAL_HDRS,
Marat Dukhan799ac752020-08-13 16:08:03 -07004170 aarch32_copts = [
4171 "-marm",
4172 "-march=armv8.2-a+dotprod",
4173 "-mfpu=neon-fp-armv8",
4174 ],
Benoit Jacoba9644732020-08-13 12:48:55 -07004175 aarch32_srcs = NEONDOT_UKERNELS,
4176 aarch64_copts = ["-march=armv8.2-a+dotprod"],
4177 aarch64_srcs = NEONDOT_UKERNELS,
4178 copts = [
4179 "-UNDEBUG",
4180 "-DXNN_TEST_MODE=1",
4181 ],
4182 gcc_copts = xnnpack_gcc_std_copts(),
4183 msvc_copts = xnnpack_msvc_std_copts(),
4184 deps = [
4185 ":tables",
4186 "@FP16",
4187 "@pthreadpool",
4188 ],
4189)
4190
4191xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004192 name = "sse2_ukernels",
4193 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004194 gcc_copts = xnnpack_gcc_std_copts(),
4195 gcc_x86_copts = ["-msse2"],
4196 msvc_copts = xnnpack_msvc_std_copts(),
4197 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004198 x86_srcs = SSE_UKERNELS + SSE2_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08004199 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08004200 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08004201 "@FP16",
4202 "@pthreadpool",
4203 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004204)
4205
4206xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07004207 name = "sse2_ukernels_test_mode",
4208 hdrs = INTERNAL_HDRS,
4209 copts = [
4210 "-UNDEBUG",
4211 "-DXNN_TEST_MODE=1",
4212 ],
4213 gcc_copts = xnnpack_gcc_std_copts(),
4214 gcc_x86_copts = ["-msse2"],
4215 msvc_copts = xnnpack_msvc_std_copts(),
4216 msvc_x86_32_copts = ["/arch:SSE2"],
4217 x86_srcs = SSE_UKERNELS + SSE2_UKERNELS,
4218 deps = [
4219 ":tables",
4220 "@FP16",
4221 "@pthreadpool",
4222 ],
4223)
4224
4225xnnpack_cc_library(
Marat Dukhanfe7acb62020-03-09 19:30:05 -07004226 name = "ssse3_ukernels",
4227 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004228 gcc_copts = xnnpack_gcc_std_copts(),
4229 gcc_x86_copts = ["-mssse3"],
4230 msvc_copts = xnnpack_msvc_std_copts(),
4231 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhanfe7acb62020-03-09 19:30:05 -07004232 x86_srcs = SSSE3_UKERNELS,
4233 deps = [
4234 ":tables",
4235 "@FP16",
4236 "@pthreadpool",
4237 ],
4238)
4239
4240xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07004241 name = "ssse3_ukernels_test_mode",
4242 hdrs = INTERNAL_HDRS,
4243 copts = [
4244 "-UNDEBUG",
4245 "-DXNN_TEST_MODE=1",
4246 ],
4247 gcc_copts = xnnpack_gcc_std_copts(),
4248 gcc_x86_copts = ["-mssse3"],
4249 msvc_copts = xnnpack_msvc_std_copts(),
4250 msvc_x86_32_copts = ["/arch:SSE2"],
4251 x86_srcs = SSSE3_UKERNELS,
4252 deps = [
4253 ":tables",
4254 "@FP16",
4255 "@pthreadpool",
4256 ],
4257)
4258
4259xnnpack_cc_library(
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08004260 name = "sse41_ukernels",
4261 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004262 gcc_copts = xnnpack_gcc_std_copts(),
4263 gcc_x86_copts = ["-msse4.1"],
4264 msvc_copts = xnnpack_msvc_std_copts(),
4265 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08004266 x86_srcs = SSE41_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08004267 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08004268 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08004269 "@FP16",
4270 "@pthreadpool",
4271 ],
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08004272)
4273
4274xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07004275 name = "sse41_ukernels_test_mode",
4276 hdrs = INTERNAL_HDRS,
4277 copts = [
4278 "-UNDEBUG",
4279 "-DXNN_TEST_MODE=1",
4280 ],
4281 gcc_copts = xnnpack_gcc_std_copts(),
4282 gcc_x86_copts = ["-msse4.1"],
4283 msvc_copts = xnnpack_msvc_std_copts(),
4284 msvc_x86_32_copts = ["/arch:SSE2"],
4285 x86_srcs = SSE41_UKERNELS,
4286 deps = [
4287 ":tables",
4288 "@FP16",
4289 "@pthreadpool",
4290 ],
4291)
4292
4293xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004294 name = "avx_ukernels",
4295 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004296 gcc_copts = xnnpack_gcc_std_copts(),
4297 gcc_x86_copts = ["-mavx"],
4298 msvc_copts = xnnpack_msvc_std_copts(),
4299 msvc_x86_32_copts = ["/arch:AVX"],
4300 msvc_x86_64_copts = ["/arch:AVX"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004301 x86_srcs = AVX_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08004302 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08004303 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08004304 "@FP16",
4305 "@pthreadpool",
4306 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004307)
4308
4309xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07004310 name = "avx_ukernels_test_mode",
4311 hdrs = INTERNAL_HDRS,
4312 copts = [
4313 "-UNDEBUG",
4314 "-DXNN_TEST_MODE=1",
4315 ],
4316 gcc_copts = xnnpack_gcc_std_copts(),
4317 gcc_x86_copts = ["-mavx"],
4318 msvc_copts = xnnpack_msvc_std_copts(),
4319 msvc_x86_32_copts = ["/arch:AVX"],
4320 msvc_x86_64_copts = ["/arch:AVX"],
4321 x86_srcs = AVX_UKERNELS,
4322 deps = [
4323 ":tables",
4324 "@FP16",
4325 "@pthreadpool",
4326 ],
4327)
4328
4329xnnpack_cc_library(
Marat Dukhan1566fee2020-08-02 21:55:41 -07004330 name = "xop_ukernels",
4331 hdrs = INTERNAL_HDRS,
4332 gcc_copts = xnnpack_gcc_std_copts(),
4333 gcc_x86_copts = ["-mxop"],
4334 msvc_copts = xnnpack_msvc_std_copts(),
4335 msvc_x86_32_copts = ["/arch:AVX"],
4336 msvc_x86_64_copts = ["/arch:AVX"],
4337 x86_srcs = XOP_UKERNELS,
4338 deps = [
4339 ":tables",
4340 "@FP16",
4341 "@pthreadpool",
4342 ],
4343)
4344
4345xnnpack_cc_library(
4346 name = "xop_ukernels_test_mode",
4347 hdrs = INTERNAL_HDRS,
4348 copts = [
4349 "-UNDEBUG",
4350 "-DXNN_TEST_MODE=1",
4351 ],
4352 gcc_copts = xnnpack_gcc_std_copts(),
4353 gcc_x86_copts = ["-mxop"],
4354 msvc_copts = xnnpack_msvc_std_copts(),
4355 msvc_x86_32_copts = ["/arch:AVX"],
4356 msvc_x86_64_copts = ["/arch:AVX"],
4357 x86_srcs = XOP_UKERNELS,
4358 deps = [
4359 ":tables",
4360 "@FP16",
4361 "@pthreadpool",
4362 ],
4363)
4364
4365xnnpack_cc_library(
Marat Dukhanfda12b82019-11-21 12:27:59 -08004366 name = "fma3_ukernels",
4367 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004368 gcc_copts = xnnpack_gcc_std_copts(),
4369 gcc_x86_copts = ["-mfma"],
4370 msvc_copts = xnnpack_msvc_std_copts(),
4371 msvc_x86_32_copts = ["/arch:AVX"],
4372 msvc_x86_64_copts = ["/arch:AVX"],
Marat Dukhanfda12b82019-11-21 12:27:59 -08004373 x86_srcs = FMA3_UKERNELS,
4374 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08004375 ":tables",
Marat Dukhanfda12b82019-11-21 12:27:59 -08004376 "@FP16",
4377 "@pthreadpool",
4378 ],
4379)
4380
4381xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07004382 name = "fma3_ukernels_test_mode",
4383 hdrs = INTERNAL_HDRS,
4384 copts = [
4385 "-UNDEBUG",
4386 "-DXNN_TEST_MODE=1",
4387 ],
4388 gcc_copts = xnnpack_gcc_std_copts(),
4389 gcc_x86_copts = ["-mfma"],
4390 msvc_copts = xnnpack_msvc_std_copts(),
4391 msvc_x86_32_copts = ["/arch:AVX"],
4392 msvc_x86_64_copts = ["/arch:AVX"],
4393 x86_srcs = FMA3_UKERNELS,
4394 deps = [
4395 ":tables",
4396 "@FP16",
4397 "@pthreadpool",
4398 ],
4399)
4400
4401xnnpack_cc_library(
Marat Dukhan6adff4e2019-10-14 18:32:07 -07004402 name = "avx2_ukernels",
4403 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004404 gcc_copts = xnnpack_gcc_std_copts(),
4405 gcc_x86_copts = [
Marat Dukhan6adff4e2019-10-14 18:32:07 -07004406 "-mfma",
4407 "-mavx2",
4408 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07004409 msvc_copts = xnnpack_msvc_std_copts(),
4410 msvc_x86_32_copts = ["/arch:AVX2"],
4411 msvc_x86_64_copts = ["/arch:AVX2"],
Marat Dukhan6adff4e2019-10-14 18:32:07 -07004412 x86_srcs = AVX2_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08004413 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08004414 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08004415 "@FP16",
4416 "@pthreadpool",
4417 ],
Marat Dukhan6adff4e2019-10-14 18:32:07 -07004418)
4419
4420xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07004421 name = "avx2_ukernels_test_mode",
4422 hdrs = INTERNAL_HDRS,
4423 copts = [
4424 "-UNDEBUG",
4425 "-DXNN_TEST_MODE=1",
4426 ],
4427 gcc_copts = xnnpack_gcc_std_copts(),
4428 gcc_x86_copts = [
4429 "-mfma",
4430 "-mavx2",
4431 ],
4432 msvc_copts = xnnpack_msvc_std_copts(),
4433 msvc_x86_32_copts = ["/arch:AVX2"],
4434 msvc_x86_64_copts = ["/arch:AVX2"],
4435 x86_srcs = AVX2_UKERNELS,
4436 deps = [
4437 ":tables",
4438 "@FP16",
4439 "@pthreadpool",
4440 ],
4441)
4442
4443xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004444 name = "avx512f_ukernels",
4445 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004446 gcc_copts = xnnpack_gcc_std_copts(),
4447 gcc_x86_copts = ["-mavx512f"],
4448 mingw_copts = ["-fno-asynchronous-unwind-tables"],
4449 msvc_copts = xnnpack_msvc_std_copts(),
4450 msvc_x86_32_copts = ["/arch:AVX512"],
4451 msvc_x86_64_copts = ["/arch:AVX512"],
4452 msys_copts = ["-fno-asynchronous-unwind-tables"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004453 x86_srcs = AVX512F_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08004454 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08004455 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08004456 "@FP16",
4457 "@pthreadpool",
4458 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004459)
4460
4461xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07004462 name = "avx512f_ukernels_test_mode",
4463 hdrs = INTERNAL_HDRS,
4464 copts = [
4465 "-UNDEBUG",
4466 "-DXNN_TEST_MODE=1",
4467 ],
4468 gcc_copts = xnnpack_gcc_std_copts(),
4469 gcc_x86_copts = ["-mavx512f"],
4470 mingw_copts = ["-fno-asynchronous-unwind-tables"],
4471 msvc_copts = xnnpack_msvc_std_copts(),
4472 msvc_x86_32_copts = ["/arch:AVX512"],
4473 msvc_x86_64_copts = ["/arch:AVX512"],
4474 msys_copts = ["-fno-asynchronous-unwind-tables"],
4475 x86_srcs = AVX512F_UKERNELS,
4476 deps = [
4477 ":tables",
4478 "@FP16",
4479 "@pthreadpool",
4480 ],
4481)
4482
4483xnnpack_cc_library(
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07004484 name = "avx512skx_ukernels",
4485 hdrs = INTERNAL_HDRS,
4486 gcc_copts = xnnpack_gcc_std_copts(),
4487 gcc_x86_copts = [
4488 "-mavx512f",
4489 "-mavx512cd",
4490 "-mavx512bw",
4491 "-mavx512dq",
4492 "-mavx512vl",
4493 ],
4494 mingw_copts = ["-fno-asynchronous-unwind-tables"],
4495 msvc_copts = xnnpack_msvc_std_copts(),
4496 msvc_x86_32_copts = ["/arch:AVX512"],
4497 msvc_x86_64_copts = ["/arch:AVX512"],
4498 msys_copts = ["-fno-asynchronous-unwind-tables"],
4499 x86_srcs = AVX512SKX_UKERNELS,
4500 deps = [
4501 ":tables",
4502 "@FP16",
4503 "@pthreadpool",
4504 ],
4505)
4506
4507xnnpack_cc_library(
4508 name = "avx512skx_ukernels_test_mode",
4509 hdrs = INTERNAL_HDRS,
4510 copts = [
4511 "-UNDEBUG",
4512 "-DXNN_TEST_MODE=1",
4513 ],
4514 gcc_copts = xnnpack_gcc_std_copts(),
4515 gcc_x86_copts = [
4516 "-mavx512f",
4517 "-mavx512cd",
4518 "-mavx512bw",
4519 "-mavx512dq",
4520 "-mavx512vl",
4521 ],
4522 mingw_copts = ["-fno-asynchronous-unwind-tables"],
4523 msvc_copts = xnnpack_msvc_std_copts(),
4524 msvc_x86_32_copts = ["/arch:AVX512"],
4525 msvc_x86_64_copts = ["/arch:AVX512"],
4526 msys_copts = ["-fno-asynchronous-unwind-tables"],
4527 x86_srcs = AVX512SKX_UKERNELS,
4528 deps = [
4529 ":tables",
4530 "@FP16",
4531 "@pthreadpool",
4532 ],
4533)
4534
4535xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004536 name = "asm_ukernels",
4537 hdrs = ["src/xnnpack/assembly.h"],
4538 aarch32_srcs = AARCH32_ASM_UKERNELS,
Frank Barchard31bb45b2020-10-06 00:26:33 -07004539 aarch64_copts = ["-march=armv8.2-a+fp16+dotprod"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004540 aarch64_srcs = AARCH64_ASM_UKERNELS,
Frank Barchardbcedc082020-08-17 18:00:51 -07004541 wasm_srcs = WASM32_ASM_UKERNELS,
4542 wasmsimd_srcs = WASM32_ASM_UKERNELS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07004543)
4544
Marat Dukhan3b59de22020-06-03 20:15:19 -07004545xnnpack_cc_library(
4546 name = "logging_utils",
4547 srcs = LOGGING_SRCS,
4548 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
4549 copts = LOGGING_COPTS + [
4550 "-Isrc",
4551 "-Iinclude",
4552 ] + select({
4553 ":debug_build": [],
4554 "//conditions:default": xnnpack_min_size_copts(),
4555 }),
4556 gcc_copts = xnnpack_gcc_std_copts(),
4557 msvc_copts = xnnpack_msvc_std_copts(),
4558 visibility = xnnpack_visibility(),
4559 deps = [
4560 "@FP16",
4561 "@clog",
4562 "@pthreadpool",
4563 ],
4564)
4565
Marat Dukhan08c4a432019-10-03 09:29:21 -07004566xnnpack_aggregate_library(
4567 name = "ukernels",
Marat Dukhan6e8c0ce2021-04-13 14:35:08 -07004568 aarch32_ios_deps = [
4569 ":neon_ukernels",
4570 ":neonfma_ukernels",
4571 ":neonv8_ukernels",
4572 ":asm_ukernels",
4573 ],
4574 aarch32_nonios_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07004575 ":neon_ukernels",
4576 ":neonfma_ukernels",
Marat Dukhan8853b822020-05-07 12:19:01 -07004577 ":neonv8_ukernels",
Benoit Jacoba9644732020-08-13 12:48:55 -07004578 ":neondot_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004579 ":asm_ukernels",
4580 ],
4581 aarch64_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07004582 ":neon_ukernels",
4583 ":neonfma_ukernels",
Marat Dukhan8853b822020-05-07 12:19:01 -07004584 ":neonv8_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004585 ":neonfp16arith_ukernels",
Benoit Jacoba9644732020-08-13 12:48:55 -07004586 ":neondot_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004587 ":asm_ukernels",
4588 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07004589 generic_deps = [
4590 ":scalar_ukernels",
4591 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07004592 wasm_deps = [
4593 ":wasm_ukernels",
Frank Barchardbcedc082020-08-17 18:00:51 -07004594 ":asm_ukernels",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004595 ],
4596 wasmsimd_deps = [
4597 ":wasm_ukernels",
Frank Barchardbcedc082020-08-17 18:00:51 -07004598 ":asm_ukernels",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004599 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004600 x86_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07004601 ":sse2_ukernels",
Marat Dukhanfe7acb62020-03-09 19:30:05 -07004602 ":ssse3_ukernels",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08004603 ":sse41_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004604 ":avx_ukernels",
Marat Dukhan1566fee2020-08-02 21:55:41 -07004605 ":xop_ukernels",
Marat Dukhanfda12b82019-11-21 12:27:59 -08004606 ":fma3_ukernels",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07004607 ":avx2_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004608 ":avx512f_ukernels",
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07004609 ":avx512skx_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004610 ],
4611)
4612
Marat Dukhan33fcf782020-05-24 14:27:15 -07004613xnnpack_aggregate_library(
4614 name = "ukernels_test_mode",
Marat Dukhan6e8c0ce2021-04-13 14:35:08 -07004615 aarch32_ios_deps = [
4616 ":neon_ukernels_test_mode",
4617 ":neonfma_ukernels_test_mode",
4618 ":neonv8_ukernels_test_mode",
4619 ":asm_ukernels",
4620 ],
4621 aarch32_nonios_deps = [
Marat Dukhan33fcf782020-05-24 14:27:15 -07004622 ":neon_ukernels_test_mode",
4623 ":neonfma_ukernels_test_mode",
4624 ":neonv8_ukernels_test_mode",
Benoit Jacoba9644732020-08-13 12:48:55 -07004625 ":neondot_ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004626 ":asm_ukernels",
4627 ],
4628 aarch64_deps = [
4629 ":neon_ukernels_test_mode",
4630 ":neonfma_ukernels_test_mode",
4631 ":neonv8_ukernels_test_mode",
4632 ":neonfp16arith_ukernels_test_mode",
Benoit Jacoba9644732020-08-13 12:48:55 -07004633 ":neondot_ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004634 ":asm_ukernels",
4635 ],
4636 generic_deps = [
4637 ":scalar_ukernels_test_mode",
4638 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07004639 wasm_deps = [
4640 ":wasm_ukernels_test_mode",
Frank Barchardbcedc082020-08-17 18:00:51 -07004641 ":asm_ukernels",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004642 ],
4643 wasmsimd_deps = [
4644 ":wasm_ukernels_test_mode",
Frank Barchardbcedc082020-08-17 18:00:51 -07004645 ":asm_ukernels",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004646 ],
4647 x86_deps = [
4648 ":sse2_ukernels_test_mode",
4649 ":ssse3_ukernels_test_mode",
4650 ":sse41_ukernels_test_mode",
4651 ":avx_ukernels_test_mode",
Marat Dukhan1566fee2020-08-02 21:55:41 -07004652 ":xop_ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004653 ":fma3_ukernels_test_mode",
4654 ":avx2_ukernels_test_mode",
4655 ":avx512f_ukernels_test_mode",
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07004656 ":avx512skx_ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004657 ],
4658)
4659
Marat Dukhan08c4a432019-10-03 09:29:21 -07004660xnnpack_cc_library(
4661 name = "im2col",
4662 srcs = ["src/im2col.c"],
4663 hdrs = [
4664 "src/xnnpack/common.h",
4665 "src/xnnpack/im2col.h",
4666 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07004667 gcc_copts = xnnpack_gcc_std_copts(),
4668 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004669)
4670
4671xnnpack_cc_library(
4672 name = "indirection",
4673 srcs = ["src/indirection.c"],
4674 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004675 gcc_copts = xnnpack_gcc_std_copts(),
4676 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004677 deps = [
4678 "@FP16",
4679 "@FXdiv",
4680 "@pthreadpool",
4681 ],
4682)
4683
4684xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07004685 name = "indirection_test_mode",
4686 srcs = ["src/indirection.c"],
4687 hdrs = INTERNAL_HDRS,
4688 copts = [
4689 "-UNDEBUG",
4690 "-DXNN_TEST_MODE=1",
4691 ],
4692 gcc_copts = xnnpack_gcc_std_copts(),
4693 msvc_copts = xnnpack_msvc_std_copts(),
4694 deps = [
4695 "@FP16",
4696 "@FXdiv",
4697 "@pthreadpool",
4698 ],
4699)
4700
4701xnnpack_cc_library(
Marat Dukhanab582382020-07-06 13:32:08 -07004702 name = "packing",
4703 srcs = ["src/packing.c"],
4704 hdrs = INTERNAL_HDRS,
4705 gcc_copts = xnnpack_gcc_std_copts(),
4706 msvc_copts = xnnpack_msvc_std_copts(),
4707 deps = [
4708 "@FP16",
4709 "@FXdiv",
4710 "@pthreadpool",
4711 ],
4712)
4713
4714xnnpack_cc_library(
4715 name = "packing_test_mode",
4716 srcs = ["src/packing.c"],
4717 hdrs = INTERNAL_HDRS,
4718 copts = [
4719 "-UNDEBUG",
4720 "-DXNN_TEST_MODE=1",
4721 ],
4722 gcc_copts = xnnpack_gcc_std_copts(),
4723 msvc_copts = xnnpack_msvc_std_copts(),
4724 deps = [
4725 "@FP16",
4726 "@FXdiv",
4727 "@pthreadpool",
4728 ],
4729)
4730
4731xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004732 name = "operator_run",
4733 srcs = ["src/operator-run.c"],
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07004734 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004735 copts = LOGGING_COPTS + select({
Marat Dukhan05702cf2020-03-26 15:41:33 -07004736 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4737 "//conditions:default": [],
4738 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07004739 gcc_copts = xnnpack_gcc_std_copts(),
4740 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004741 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07004742 ":logging_utils",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004743 "@FP16",
4744 "@FXdiv",
4745 "@clog",
4746 "@pthreadpool",
4747 ],
4748)
4749
Chao Mei6ddfc602020-05-13 22:29:36 -07004750xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07004751 name = "operator_run_test_mode",
4752 srcs = ["src/operator-run.c"],
4753 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
4754 copts = LOGGING_COPTS + [
4755 "-UNDEBUG",
4756 "-DXNN_TEST_MODE=1",
4757 ] + select({
4758 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4759 "//conditions:default": [],
4760 }),
4761 gcc_copts = xnnpack_gcc_std_copts(),
4762 msvc_copts = xnnpack_msvc_std_copts(),
4763 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07004764 ":logging_utils",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004765 "@FP16",
4766 "@FXdiv",
4767 "@clog",
4768 "@pthreadpool",
4769 ],
4770)
4771
4772xnnpack_cc_library(
Chao Mei6ddfc602020-05-13 22:29:36 -07004773 name = "memory_planner",
4774 srcs = ["src/memory-planner.c"],
4775 hdrs = INTERNAL_HDRS,
4776 defines = select({
4777 ":xnn_enable_memopt_explicit_true": ["XNN_ENABLE_MEMOPT=1"],
4778 ":xnn_enable_memopt_explicit_false": ["XNN_ENABLE_MEMOPT=0"],
4779 "//conditions:default": ["XNN_ENABLE_MEMOPT=1"],
4780 }),
4781 gcc_copts = xnnpack_gcc_std_copts(),
4782 msvc_copts = xnnpack_msvc_std_copts(),
4783 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07004784 ":logging_utils",
Chao Mei6ddfc602020-05-13 22:29:36 -07004785 "@pthreadpool",
4786 ],
4787)
4788
Marat Dukhan33fcf782020-05-24 14:27:15 -07004789xnnpack_cc_library(
4790 name = "memory_planner_test_mode",
4791 srcs = ["src/memory-planner.c"],
4792 hdrs = INTERNAL_HDRS,
4793 copts = [
4794 "-UNDEBUG",
4795 "-DXNN_TEST_MODE=1",
4796 ],
4797 defines = select({
4798 ":xnn_enable_memopt_explicit_true": ["XNN_ENABLE_MEMOPT=1"],
4799 ":xnn_enable_memopt_explicit_false": ["XNN_ENABLE_MEMOPT=0"],
4800 "//conditions:default": ["XNN_ENABLE_MEMOPT=1"],
4801 }),
4802 gcc_copts = xnnpack_gcc_std_copts(),
4803 msvc_copts = xnnpack_msvc_std_copts(),
4804 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07004805 ":logging_utils",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004806 "@pthreadpool",
4807 ],
4808)
4809
Marat Dukhan08c4a432019-10-03 09:29:21 -07004810cc_library(
4811 name = "enable_assembly",
4812 defines = select({
4813 ":xnn_enable_assembly_explicit_true": ["XNN_ENABLE_ASSEMBLY=1"],
4814 ":xnn_enable_assembly_explicit_false": ["XNN_ENABLE_ASSEMBLY=0"],
Frank Barchard810171d2019-10-10 10:34:51 -07004815 "//conditions:default": ["XNN_ENABLE_ASSEMBLY=1"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004816 }),
4817)
4818
Marat Dukhan9de90e02020-06-18 16:04:12 -07004819cc_library(
4820 name = "enable_sparse",
4821 defines = select({
4822 ":xnn_enable_sparse_explicit_true": ["XNN_ENABLE_SPARSE=1"],
4823 ":xnn_enable_sparse_explicit_false": ["XNN_ENABLE_SPARSE=0"],
Marat Dukhanb36582b2020-12-08 11:16:28 -08004824 "//conditions:default": ["XNN_ENABLE_SPARSE=1"],
Marat Dukhan9de90e02020-06-18 16:04:12 -07004825 }),
4826)
4827
Marat Dukhancf056b22019-10-07 10:26:29 -07004828xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004829 name = "operators",
4830 srcs = OPERATOR_SRCS + [
Marat Dukhan496389f2021-04-07 15:47:12 -07004831 "src/allocator.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004832 "src/operator-delete.c",
Marat Dukhancf056b22019-10-07 10:26:29 -07004833 ],
4834 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004835 copts = LOGGING_COPTS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07004836 "-Isrc",
4837 "-Iinclude",
4838 ] + select({
4839 ":debug_build": [],
4840 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07004841 }) + select({
4842 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4843 "//conditions:default": [],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004844 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07004845 gcc_copts = xnnpack_gcc_std_copts(),
4846 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004847 deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07004848 ":indirection",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004849 ":logging_utils",
Marat Dukhanab582382020-07-06 13:32:08 -07004850 ":packing",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004851 "@FP16",
4852 "@FXdiv",
4853 "@clog",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004854 "@pthreadpool",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004855 ],
4856)
4857
Marat Dukhan10a38082020-04-17 03:58:35 -07004858xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07004859 name = "operators_test_mode",
4860 srcs = OPERATOR_SRCS + [
Marat Dukhan496389f2021-04-07 15:47:12 -07004861 "src/allocator.c",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004862 "src/operator-delete.c",
4863 ],
4864 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
4865 copts = LOGGING_COPTS + [
4866 "-Isrc",
4867 "-Iinclude",
4868 "-UNDEBUG",
4869 "-DXNN_TEST_MODE=1",
4870 ] + select({
4871 ":debug_build": [],
4872 "//conditions:default": xnnpack_min_size_copts(),
4873 }) + select({
4874 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4875 "//conditions:default": [],
4876 }),
4877 gcc_copts = xnnpack_gcc_std_copts(),
4878 msvc_copts = xnnpack_msvc_std_copts(),
4879 deps = [
4880 ":indirection_test_mode",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004881 ":logging_utils",
Marat Dukhanab582382020-07-06 13:32:08 -07004882 ":packing_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004883 "@FP16",
4884 "@FXdiv",
4885 "@clog",
4886 "@pthreadpool",
4887 ],
4888)
4889
4890xnnpack_cc_library(
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004891 name = "XNNPACK",
4892 srcs = [
4893 "src/init.c",
Marat Dukhanccfdbd12020-02-03 14:27:45 -08004894 "src/runtime.c",
4895 "src/subgraph.c",
4896 "src/tensor.c",
Marat Dukhanf03da0d2020-06-10 16:00:20 -07004897 ] + SUBGRAPH_SRCS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004898 hdrs = ["include/xnnpack.h"],
4899 copts = LOGGING_COPTS + [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004900 "-Isrc",
4901 "-Iinclude",
4902 ] + select({
4903 ":debug_build": [],
4904 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07004905 }) + select({
4906 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4907 "//conditions:default": [],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004908 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07004909 gcc_copts = xnnpack_gcc_std_copts(),
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004910 includes = ["include"],
Marat Dukhan10a38082020-04-17 03:58:35 -07004911 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004912 visibility = xnnpack_visibility(),
4913 deps = [
4914 ":enable_assembly",
Marat Dukhan9de90e02020-06-18 16:04:12 -07004915 ":enable_sparse",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004916 ":logging_utils",
Chao Mei6ddfc602020-05-13 22:29:36 -07004917 ":memory_planner",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004918 ":operator_run",
4919 ":operators",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004920 ":ukernels",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004921 "@clog",
Marat Dukhanab2946c2020-05-21 20:04:13 -07004922 "@FP16",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004923 "@pthreadpool",
Marat Dukhand343c222019-10-07 09:22:14 -07004924 ] + select({
4925 ":emscripten": [],
4926 "//conditions:default": ["@cpuinfo"],
4927 }),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004928)
4929
Marat Dukhan10a38082020-04-17 03:58:35 -07004930xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07004931 name = "XNNPACK_test_mode",
4932 srcs = [
4933 "src/init.c",
4934 "src/runtime.c",
4935 "src/subgraph.c",
4936 "src/tensor.c",
Marat Dukhanf03da0d2020-06-10 16:00:20 -07004937 ] + SUBGRAPH_SRCS,
Marat Dukhan33fcf782020-05-24 14:27:15 -07004938 hdrs = ["include/xnnpack.h"],
4939 copts = LOGGING_COPTS + [
4940 "-Isrc",
4941 "-Iinclude",
4942 "-UNDEBUG",
4943 "-DXNN_TEST_MODE=1",
4944 ] + select({
4945 ":debug_build": [],
4946 "//conditions:default": xnnpack_min_size_copts(),
4947 }) + select({
4948 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4949 "//conditions:default": [],
4950 }),
4951 gcc_copts = xnnpack_gcc_std_copts(),
4952 includes = ["include"],
4953 msvc_copts = xnnpack_msvc_std_copts(),
4954 visibility = xnnpack_visibility(),
4955 deps = [
4956 ":enable_assembly",
Marat Dukhan9de90e02020-06-18 16:04:12 -07004957 ":enable_sparse",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004958 ":logging_utils",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004959 ":memory_planner_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004960 ":operator_run_test_mode",
4961 ":operators_test_mode",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004962 ":ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004963 "@clog",
4964 "@FP16",
4965 "@pthreadpool",
4966 ] + select({
4967 ":emscripten": [],
4968 "//conditions:default": ["@cpuinfo"],
4969 }),
4970)
4971
Marat Dukhanf0cb70a2021-03-30 15:45:15 -07004972# Specialized XNNPACK version for TensorFlow Lite. Excludes operators currently
4973# not used by the TensorFlow Lite XNNPACK delegate to minimize code size.
Marat Dukhanae046f52020-06-15 13:16:14 -07004974xnnpack_cc_library(
Marat Dukhanf0cb70a2021-03-30 15:45:15 -07004975 name = "xnnpack_for_tflite",
4976 srcs = [
4977 "src/init.c",
4978 "src/runtime.c",
4979 "src/subgraph.c",
4980 "src/tensor.c",
4981 ] + SUBGRAPH_SRCS,
4982 hdrs = ["include/xnnpack.h"],
4983 copts = LOGGING_COPTS + [
4984 "-Isrc",
4985 "-Iinclude",
4986 ] + select({
4987 ":debug_build": [],
4988 "//conditions:default": xnnpack_min_size_copts(),
4989 }) + select({
4990 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4991 "//conditions:default": [],
4992 }),
4993 defines = [
Marat Dukhanf0cb70a2021-03-30 15:45:15 -07004994 "XNN_NO_QU8_OPERATORS",
4995 "XNN_NO_U8_OPERATORS",
4996 "XNN_NO_X8_OPERATORS",
4997 "XNN_NO_F16_OPERATORS",
4998 "XNN_NO_X16_OPERATORS",
Marat Dukhanb939cdb2021-03-30 18:51:51 -07004999 ] + select({
5000 ":xnn_enable_qs8_explicit_true": [],
5001 ":xnn_enable_qs8_explicit_false": ["XNN_NO_QS8_OPERATORS"],
5002 "//conditions:default": ["XNN_NO_QS8_OPERATORS"],
5003 }),
Marat Dukhanf0cb70a2021-03-30 15:45:15 -07005004 gcc_copts = xnnpack_gcc_std_copts(),
5005 includes = ["include"],
5006 msvc_copts = xnnpack_msvc_std_copts(),
5007 visibility = xnnpack_visibility(),
5008 deps = [
5009 ":enable_assembly",
5010 ":enable_sparse",
5011 ":logging_utils",
5012 ":memory_planner",
5013 ":operator_run",
5014 ":operators",
Marat Dukhand09ca262021-03-30 16:17:12 -07005015 ":ukernels",
Marat Dukhanf0cb70a2021-03-30 15:45:15 -07005016 "@clog",
5017 "@FP16",
5018 "@pthreadpool",
5019 ] + select({
5020 ":emscripten": [],
5021 "//conditions:default": ["@cpuinfo"],
5022 }),
5023)
5024
5025# Specialized XNNPACK version for TensorFlow.js. Excludes operators currently
5026# not used by the TensorFlow.js WebAssembly backend to minimize code size.
5027xnnpack_cc_library(
5028 name = "xnnpack_for_tfjs",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07005029 srcs = [
5030 "src/init.c",
5031 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07005032 hdrs = ["include/xnnpack.h"],
5033 copts = LOGGING_COPTS + [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07005034 "-Isrc",
5035 "-Iinclude",
5036 ] + select({
5037 ":debug_build": [],
5038 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07005039 }) + select({
5040 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
5041 "//conditions:default": [],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07005042 }),
5043 defines = [
Marat Dukhan16f1e1a2020-08-04 16:38:22 -07005044 "XNN_NO_QS8_OPERATORS",
Marat Dukhan08b7a972020-07-14 18:17:29 -07005045 "XNN_NO_QU8_OPERATORS",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07005046 "XNN_NO_U8_OPERATORS",
5047 "XNN_NO_X8_OPERATORS",
Marat Dukhanefc47b82019-11-18 09:25:38 -08005048 "XNN_NO_NCHW_OPERATORS",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07005049 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07005050 gcc_copts = xnnpack_gcc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07005051 includes = ["include"],
Marat Dukhan10a38082020-04-17 03:58:35 -07005052 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07005053 visibility = xnnpack_visibility(),
5054 deps = [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07005055 ":enable_assembly",
Marat Dukhan3b59de22020-06-03 20:15:19 -07005056 ":logging_utils",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005057 ":operator_run",
5058 ":operators",
Marat Dukhan3b59de22020-06-03 20:15:19 -07005059 ":ukernels",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07005060 "@clog",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005061 "@pthreadpool",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07005062 ] + select({
5063 ":emscripten": [],
5064 "//conditions:default": ["@cpuinfo"],
5065 }),
Marat Dukhan08c4a432019-10-03 09:29:21 -07005066)
5067
Marat Dukhancf056b22019-10-07 10:26:29 -07005068xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005069 name = "bench_utils",
5070 srcs = ["bench/utils.cc"],
5071 hdrs = ["bench/utils.h"],
Marat Dukhanbad48fe2019-11-04 10:35:22 -08005072 deps = [
5073 "@com_google_benchmark//:benchmark",
5074 "@cpuinfo",
5075 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005076)
5077
Frank Barchard7e955972019-10-11 10:34:25 -07005078######################### Benchmarks for micro-kernels #########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07005079
5080xnnpack_benchmark(
Marat Dukhan595e1702020-07-31 10:12:52 -07005081 name = "qs8_gemm_bench",
5082 srcs = [
5083 "bench/gemm.h",
5084 "bench/qs8-gemm.cc",
5085 "src/xnnpack/AlignedAllocator.h",
5086 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Frank Barchard31328cb2020-10-12 11:55:18 -07005087 copts = xnnpack_optional_ruy_copts() + xnnpack_optional_gemmlowp_copts(),
5088 deps = MICROKERNEL_BENCHMARK_DEPS + [":packing"] + xnnpack_optional_ruy_deps() + xnnpack_optional_gemmlowp_deps(),
Marat Dukhan595e1702020-07-31 10:12:52 -07005089)
5090
5091xnnpack_benchmark(
Marat Dukhan56bdd4a2020-08-03 19:47:04 -07005092 name = "qs8_requantization_bench",
5093 srcs = [
5094 "bench/qs8-requantization.cc",
Marat Dukhan56bdd4a2020-08-03 19:47:04 -07005095 "src/xnnpack/AlignedAllocator.h",
Frank Barchard04336c12020-10-22 16:48:55 -07005096 "src/xnnpack/requantization-stubs.h",
Marat Dukhan56bdd4a2020-08-03 19:47:04 -07005097 ] + MICROKERNEL_BENCHMARK_HDRS,
5098 deps = MICROKERNEL_BENCHMARK_DEPS,
5099)
5100
5101xnnpack_benchmark(
Marat Dukhan08b7a972020-07-14 18:17:29 -07005102 name = "qu8_gemm_bench",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005103 srcs = [
5104 "bench/gemm.h",
Marat Dukhan08b7a972020-07-14 18:17:29 -07005105 "bench/qu8-gemm.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005106 "src/xnnpack/AlignedAllocator.h",
5107 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07005108 copts = xnnpack_optional_ruy_copts() + xnnpack_optional_gemmlowp_copts(),
Marat Dukhanab582382020-07-06 13:32:08 -07005109 deps = MICROKERNEL_BENCHMARK_DEPS + [":packing"] + xnnpack_optional_ruy_deps() + xnnpack_optional_gemmlowp_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07005110)
5111
5112xnnpack_benchmark(
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07005113 name = "qu8_requantization_bench",
5114 srcs = [
5115 "bench/qu8-requantization.cc",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07005116 "src/xnnpack/AlignedAllocator.h",
Frank Barchard04336c12020-10-22 16:48:55 -07005117 "src/xnnpack/requantization-stubs.h",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07005118 ] + MICROKERNEL_BENCHMARK_HDRS,
5119 deps = MICROKERNEL_BENCHMARK_DEPS,
5120)
5121
5122xnnpack_benchmark(
Frank Barchard40d20fe2020-05-05 00:37:45 -07005123 name = "f16_igemm_bench",
5124 srcs = [
5125 "bench/f16-igemm.cc",
5126 "bench/conv.h",
5127 "bench/google/conv.h",
5128 "src/xnnpack/AlignedAllocator.h",
5129 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005130 deps = MICROKERNEL_BENCHMARK_DEPS + [
5131 ":indirection",
5132 ":packing",
5133 ],
Frank Barchard40d20fe2020-05-05 00:37:45 -07005134)
5135
5136xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005137 name = "f16_gemm_bench",
5138 srcs = [
5139 "bench/f16-gemm.cc",
5140 "bench/gemm.h",
5141 "src/xnnpack/AlignedAllocator.h",
5142 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005143 deps = MICROKERNEL_BENCHMARK_DEPS + [
5144 ":packing",
5145 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005146)
5147
5148xnnpack_benchmark(
Marat Dukhanbdb56f52020-02-05 21:42:49 -08005149 name = "f16_spmm_bench",
5150 srcs = [
5151 "bench/f16-spmm.cc",
Marat Dukhan1631e3e2020-12-06 19:29:31 -08005152 "bench/spmm.h",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08005153 "src/xnnpack/AlignedAllocator.h",
5154 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanbdb56f52020-02-05 21:42:49 -08005155 deps = MICROKERNEL_BENCHMARK_DEPS,
5156)
5157
5158xnnpack_benchmark(
Marat Dukhan6674d692021-05-05 22:27:00 -07005159 name = "f16_vrelu_bench",
5160 srcs = [
5161 "bench/f16-vrelu.cc",
5162 "src/xnnpack/AlignedAllocator.h",
5163 ] + MICROKERNEL_BENCHMARK_HDRS,
5164 deps = MICROKERNEL_BENCHMARK_DEPS,
5165)
5166
5167xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005168 name = "f32_igemm_bench",
5169 srcs = [
5170 "bench/f32-igemm.cc",
5171 "bench/conv.h",
5172 "src/xnnpack/AlignedAllocator.h",
5173 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005174 deps = MICROKERNEL_BENCHMARK_DEPS + [
5175 ":indirection",
5176 ":packing",
5177 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005178)
5179
5180xnnpack_benchmark(
5181 name = "f32_conv_hwc_bench",
5182 srcs = [
5183 "bench/f32-conv-hwc.cc",
5184 "bench/dconv.h",
5185 "src/xnnpack/AlignedAllocator.h",
5186 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005187 deps = MICROKERNEL_BENCHMARK_DEPS + [
5188 ":packing",
5189 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005190)
5191
5192xnnpack_benchmark(
Marat Dukhan1f29b802020-05-15 23:46:39 -07005193 name = "f32_conv_hwc2chw_bench",
Erich Elsen563df5f2019-10-23 08:02:21 -07005194 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07005195 "bench/f32-conv-hwc2chw.cc",
Erich Elsen563df5f2019-10-23 08:02:21 -07005196 "bench/dconv.h",
5197 "src/xnnpack/AlignedAllocator.h",
5198 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005199 deps = MICROKERNEL_BENCHMARK_DEPS + [
5200 ":packing",
5201 ],
Erich Elsen563df5f2019-10-23 08:02:21 -07005202)
5203
5204xnnpack_benchmark(
Frank Barchard5a599a62020-06-04 20:12:44 -07005205 name = "f16_dwconv_bench",
5206 srcs = [
5207 "bench/f16-dwconv.cc",
5208 "bench/dwconv.h",
5209 "bench/google/dwconv.h",
5210 "src/xnnpack/AlignedAllocator.h",
5211 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005212 deps = MICROKERNEL_BENCHMARK_DEPS + [
5213 ":indirection",
5214 ":packing",
5215 ],
Frank Barchard5a599a62020-06-04 20:12:44 -07005216)
5217
5218xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005219 name = "f32_dwconv_bench",
5220 srcs = [
5221 "bench/f32-dwconv.cc",
5222 "bench/dwconv.h",
5223 "src/xnnpack/AlignedAllocator.h",
5224 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005225 deps = MICROKERNEL_BENCHMARK_DEPS + [
5226 ":indirection",
5227 ":packing",
5228 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005229)
5230
5231xnnpack_benchmark(
Marat Dukhanbf715f92020-10-23 20:17:00 -07005232 name = "f32_dwconv2d_chw_bench",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005233 srcs = [
Marat Dukhanbf715f92020-10-23 20:17:00 -07005234 "bench/f32-dwconv2d-chw.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005235 "bench/dwconv.h",
5236 "src/xnnpack/AlignedAllocator.h",
5237 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005238 deps = MICROKERNEL_BENCHMARK_DEPS + [
5239 ":indirection",
5240 ":packing",
5241 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005242)
5243
5244xnnpack_benchmark(
5245 name = "f32_gemm_bench",
5246 srcs = [
5247 "bench/f32-gemm.cc",
5248 "bench/gemm.h",
5249 "src/xnnpack/AlignedAllocator.h",
5250 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07005251 copts = xnnpack_optional_ruy_copts(),
Marat Dukhanab582382020-07-06 13:32:08 -07005252 deps = MICROKERNEL_BENCHMARK_DEPS + [":packing"] + xnnpack_optional_ruy_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07005253)
5254
5255xnnpack_benchmark(
Marat Dukhan4c4eb002019-12-08 21:27:49 -08005256 name = "f32_raddexpminusmax_bench",
5257 srcs = [
5258 "bench/f32-raddexpminusmax.cc",
5259 "src/xnnpack/AlignedAllocator.h",
5260 ] + MICROKERNEL_BENCHMARK_HDRS,
5261 deps = MICROKERNEL_BENCHMARK_DEPS,
5262)
5263
5264xnnpack_benchmark(
5265 name = "f32_raddextexp_bench",
5266 srcs = [
5267 "bench/f32-raddextexp.cc",
5268 "src/xnnpack/AlignedAllocator.h",
5269 ] + MICROKERNEL_BENCHMARK_HDRS,
5270 deps = MICROKERNEL_BENCHMARK_DEPS,
5271)
5272
5273xnnpack_benchmark(
5274 name = "f32_raddstoreexpminusmax_bench",
5275 srcs = [
5276 "bench/f32-raddstoreexpminusmax.cc",
5277 "src/xnnpack/AlignedAllocator.h",
5278 ] + MICROKERNEL_BENCHMARK_HDRS,
5279 deps = MICROKERNEL_BENCHMARK_DEPS,
5280)
5281
5282xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005283 name = "f32_rmax_bench",
5284 srcs = [
5285 "bench/f32-rmax.cc",
5286 "src/xnnpack/AlignedAllocator.h",
5287 ] + MICROKERNEL_BENCHMARK_HDRS,
5288 deps = MICROKERNEL_BENCHMARK_DEPS,
5289)
5290
5291xnnpack_benchmark(
5292 name = "f32_spmm_bench",
5293 srcs = [
5294 "bench/f32-spmm.cc",
Marat Dukhan1631e3e2020-12-06 19:29:31 -08005295 "bench/spmm.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005296 "src/xnnpack/AlignedAllocator.h",
5297 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005298 deps = MICROKERNEL_BENCHMARK_DEPS,
5299)
5300
5301xnnpack_benchmark(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08005302 name = "f32_softmax_bench",
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07005303 srcs = [
Marat Dukhanfd8e6892020-01-27 15:25:25 -08005304 "bench/f32-softmax.cc",
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07005305 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07005306 copts = xnnpack_optional_dnnl_copts(),
Marat Dukhan8d3c6932020-03-06 20:27:27 -08005307 deps = MICROKERNEL_BENCHMARK_DEPS + xnnpack_optional_dnnl_deps(),
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07005308)
5309
5310xnnpack_benchmark(
Marat Dukhaned6baaf2020-12-01 15:07:08 -08005311 name = "f32_velu_bench",
5312 srcs = [
5313 "bench/f32-velu.cc",
5314 "src/xnnpack/AlignedAllocator.h",
5315 ] + MICROKERNEL_BENCHMARK_HDRS,
5316 deps = MICROKERNEL_BENCHMARK_DEPS,
5317)
5318
5319xnnpack_benchmark(
Marat Dukhan6674d692021-05-05 22:27:00 -07005320 name = "f32_vhswish_bench",
5321 srcs = [
5322 "bench/f32-vhswish.cc",
5323 "src/xnnpack/AlignedAllocator.h",
5324 ] + MICROKERNEL_BENCHMARK_HDRS,
5325 deps = MICROKERNEL_BENCHMARK_DEPS,
5326)
5327
5328xnnpack_benchmark(
5329 name = "f32_vrelu_bench",
5330 srcs = [
5331 "bench/f32-vrelu.cc",
5332 "src/xnnpack/AlignedAllocator.h",
5333 ] + MICROKERNEL_BENCHMARK_HDRS,
5334 deps = MICROKERNEL_BENCHMARK_DEPS,
5335)
5336
5337xnnpack_benchmark(
Marat Dukhan4c4eb002019-12-08 21:27:49 -08005338 name = "f32_vscaleexpminusmax_bench",
5339 srcs = [
5340 "bench/f32-vscaleexpminusmax.cc",
5341 "src/xnnpack/AlignedAllocator.h",
5342 ] + MICROKERNEL_BENCHMARK_HDRS,
5343 deps = MICROKERNEL_BENCHMARK_DEPS,
5344)
5345
5346xnnpack_benchmark(
5347 name = "f32_vscaleextexp_bench",
5348 srcs = [
5349 "bench/f32-vscaleextexp.cc",
5350 "src/xnnpack/AlignedAllocator.h",
5351 ] + MICROKERNEL_BENCHMARK_HDRS,
5352 deps = MICROKERNEL_BENCHMARK_DEPS,
5353)
5354
5355xnnpack_benchmark(
Marat Dukhan6674d692021-05-05 22:27:00 -07005356 name = "f32_vsigmoid_bench",
5357 srcs = [
5358 "bench/f32-vsigmoid.cc",
5359 "src/xnnpack/AlignedAllocator.h",
5360 ] + MICROKERNEL_BENCHMARK_HDRS,
5361 deps = MICROKERNEL_BENCHMARK_DEPS,
5362)
5363
5364xnnpack_benchmark(
Marat Dukhanf4db2f32020-06-30 10:55:30 -07005365 name = "f32_vsqrt_bench",
5366 srcs = [
5367 "bench/f32-vsqrt.cc",
5368 "src/xnnpack/AlignedAllocator.h",
5369 ] + MICROKERNEL_BENCHMARK_HDRS,
5370 deps = MICROKERNEL_BENCHMARK_DEPS,
5371)
5372
5373xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005374 name = "f32_im2col_gemm_bench",
5375 srcs = [
5376 "bench/f32-im2col-gemm.cc",
5377 "bench/conv.h",
5378 "src/xnnpack/AlignedAllocator.h",
5379 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005380 deps = MICROKERNEL_BENCHMARK_DEPS + [
5381 ":im2col",
5382 ":packing",
5383 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005384)
5385
Marat Dukhanfe7acb62020-03-09 19:30:05 -07005386xnnpack_benchmark(
Marat Dukhanffbf96a2020-05-14 02:59:08 -07005387 name = "rounding_bench",
5388 srcs = [
5389 "bench/rounding.cc",
Marat Dukhanffbf96a2020-05-14 02:59:08 -07005390 "src/xnnpack/AlignedAllocator.h",
Frank Barchard04336c12020-10-22 16:48:55 -07005391 "src/xnnpack/math-stubs.h",
Marat Dukhanffbf96a2020-05-14 02:59:08 -07005392 ] + MICROKERNEL_BENCHMARK_HDRS,
5393 deps = MICROKERNEL_BENCHMARK_DEPS,
5394)
5395
Marat Dukhan08c4a432019-10-03 09:29:21 -07005396########################### Benchmarks for operators ###########################
5397
5398xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005399 name = "average_pooling_bench",
5400 srcs = ["bench/average-pooling.cc"],
Marat Dukhan7a16d8b2020-03-11 04:22:44 -07005401 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07005402 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07005403 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07005404)
5405
5406xnnpack_benchmark(
Marat Dukhan87727142020-06-24 15:24:10 -07005407 name = "bankers_rounding_bench",
5408 srcs = ["bench/bankers-rounding.cc"],
5409 copts = xnnpack_optional_tflite_copts(),
5410 tags = ["nowin32"],
5411 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
5412)
5413
5414xnnpack_benchmark(
5415 name = "ceiling_bench",
5416 srcs = ["bench/ceiling.cc"],
5417 copts = xnnpack_optional_tflite_copts(),
5418 tags = ["nowin32"],
5419 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
5420)
5421
5422xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005423 name = "channel_shuffle_bench",
5424 srcs = ["bench/channel-shuffle.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07005425 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005426)
5427
5428xnnpack_benchmark(
5429 name = "convolution_bench",
5430 srcs = ["bench/convolution.cc"],
5431 copts = xnnpack_optional_tflite_copts() + xnnpack_optional_armcl_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07005432 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07005433 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps() + xnnpack_optional_armcl_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07005434)
5435
5436xnnpack_benchmark(
5437 name = "deconvolution_bench",
5438 srcs = ["bench/deconvolution.cc"],
5439 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07005440 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07005441 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07005442)
5443
5444xnnpack_benchmark(
Marat Dukhanb6bd4bc2020-12-01 17:01:40 -08005445 name = "elu_bench",
5446 srcs = ["bench/elu.cc"],
5447 copts = xnnpack_optional_tflite_copts(),
5448 tags = ["nowin32"],
5449 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
5450)
5451
5452xnnpack_benchmark(
Marat Dukhan87727142020-06-24 15:24:10 -07005453 name = "floor_bench",
5454 srcs = ["bench/floor.cc"],
5455 copts = xnnpack_optional_tflite_copts(),
5456 tags = ["nowin32"],
5457 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
5458)
5459
5460xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005461 name = "global_average_pooling_bench",
5462 srcs = ["bench/global-average-pooling.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07005463 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005464)
5465
5466xnnpack_benchmark(
Marat Dukhanad352602020-06-25 21:50:54 -07005467 name = "hardswish_bench",
5468 srcs = ["bench/hardswish.cc"],
5469 copts = xnnpack_optional_tflite_copts(),
5470 tags = ["nowin32"],
5471 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
5472)
5473
5474xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005475 name = "max_pooling_bench",
5476 srcs = ["bench/max-pooling.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07005477 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005478)
5479
5480xnnpack_benchmark(
5481 name = "sigmoid_bench",
5482 srcs = ["bench/sigmoid.cc"],
Marat Dukhanc3b9e862019-11-17 13:18:54 -08005483 copts = xnnpack_optional_tflite_copts(),
Marat Dukhanca2ba702020-04-24 01:31:47 -07005484 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07005485 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07005486)
5487
5488xnnpack_benchmark(
Marat Dukhan95b22432019-10-30 16:30:14 -07005489 name = "prelu_bench",
5490 srcs = ["bench/prelu.cc"],
5491 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07005492 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07005493 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan95b22432019-10-30 16:30:14 -07005494)
5495
5496xnnpack_benchmark(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08005497 name = "softmax_bench",
5498 srcs = ["bench/softmax.cc"],
Marat Dukhan9c0db962020-01-28 12:30:14 -08005499 copts = xnnpack_optional_tflite_copts(),
Marat Dukhanca2ba702020-04-24 01:31:47 -07005500 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07005501 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07005502)
5503
Marat Dukhan87727142020-06-24 15:24:10 -07005504xnnpack_benchmark(
Marat Dukhan6804bbd2020-06-30 19:26:11 -07005505 name = "square_root_bench",
5506 srcs = ["bench/square-root.cc"],
5507 copts = xnnpack_optional_tflite_copts(),
5508 tags = ["nowin32"],
5509 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
5510)
5511
5512xnnpack_benchmark(
Marat Dukhan87727142020-06-24 15:24:10 -07005513 name = "truncation_bench",
5514 srcs = ["bench/truncation.cc"],
5515 deps = OPERATOR_BENCHMARK_DEPS,
5516)
5517
Marat Dukhanc068bb62019-10-04 13:24:39 -07005518############################# End-to-end benchmarks ############################
5519
5520cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07005521 name = "fp32_mobilenet_v1",
5522 srcs = ["models/fp32-mobilenet-v1.cc"],
Marat Dukhanc068bb62019-10-04 13:24:39 -07005523 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08005524 copts = xnnpack_std_cxxopts(),
Marat Dukhanc068bb62019-10-04 13:24:39 -07005525 linkstatic = True,
5526 deps = [
5527 ":XNNPACK",
5528 "@pthreadpool",
5529 ],
5530)
5531
5532cc_library(
Marat Dukhan4cea2322021-03-09 09:35:36 -08005533 name = "fp32_sparse_mobilenet_v1",
5534 srcs = ["models/fp32-sparse-mobilenet-v1.cc"],
5535 hdrs = ["models/models.h"],
5536 copts = xnnpack_std_cxxopts(),
5537 linkstatic = True,
5538 deps = [
5539 ":XNNPACK",
5540 "@pthreadpool",
5541 ],
5542)
5543
5544cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07005545 name = "fp16_mobilenet_v1",
5546 srcs = ["models/fp16-mobilenet-v1.cc"],
5547 hdrs = ["models/models.h"],
5548 copts = xnnpack_std_cxxopts(),
5549 linkstatic = True,
5550 deps = [
5551 ":XNNPACK",
5552 "@FP16",
5553 "@pthreadpool",
5554 ],
5555)
5556
5557cc_library(
Marat Dukhan0743cdf2020-08-04 18:52:07 -07005558 name = "qs8_mobilenet_v1",
5559 srcs = ["models/qs8-mobilenet-v1.cc"],
5560 hdrs = ["models/models.h"],
5561 copts = xnnpack_std_cxxopts(),
5562 linkstatic = True,
5563 deps = [
5564 ":XNNPACK",
5565 "@pthreadpool",
5566 ],
5567)
5568
5569cc_library(
Marat Dukhan70a96182020-09-03 17:13:58 -07005570 name = "qs8_mobilenet_v2",
5571 srcs = ["models/qs8-mobilenet-v2.cc"],
5572 hdrs = ["models/models.h"],
5573 copts = xnnpack_std_cxxopts(),
5574 linkstatic = True,
5575 deps = [
5576 ":XNNPACK",
5577 "@pthreadpool",
5578 ],
5579)
5580
5581cc_library(
Marat Dukhan12a23bb2021-03-08 08:13:21 -08005582 name = "qu8_mobilenet_v1",
5583 srcs = ["models/qu8-mobilenet-v1.cc"],
5584 hdrs = ["models/models.h"],
5585 copts = xnnpack_std_cxxopts(),
5586 linkstatic = True,
5587 deps = [
5588 ":XNNPACK",
5589 "@pthreadpool",
5590 ],
5591)
5592
5593cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07005594 name = "fp32_mobilenet_v2",
5595 srcs = ["models/fp32-mobilenet-v2.cc"],
Marat Dukhanc068bb62019-10-04 13:24:39 -07005596 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08005597 copts = xnnpack_std_cxxopts(),
Marat Dukhanc068bb62019-10-04 13:24:39 -07005598 linkstatic = True,
5599 deps = [
5600 ":XNNPACK",
5601 "@pthreadpool",
5602 ],
5603)
5604
Marat Dukhanc08cdf52019-12-09 09:17:51 -08005605cc_library(
Marat Dukhan4cea2322021-03-09 09:35:36 -08005606 name = "fp32_sparse_mobilenet_v2",
5607 srcs = ["models/fp32-sparse-mobilenet-v2.cc"],
5608 hdrs = ["models/models.h"],
5609 copts = xnnpack_std_cxxopts(),
5610 linkstatic = True,
5611 deps = [
5612 ":XNNPACK",
5613 "@pthreadpool",
5614 ],
5615)
5616
5617cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07005618 name = "fp16_mobilenet_v2",
5619 srcs = ["models/fp16-mobilenet-v2.cc"],
5620 hdrs = ["models/models.h"],
5621 copts = xnnpack_std_cxxopts(),
5622 linkstatic = True,
5623 deps = [
5624 ":XNNPACK",
5625 "@FP16",
5626 "@pthreadpool",
5627 ],
5628)
5629
5630cc_library(
5631 name = "fp32_mobilenet_v3_large",
5632 srcs = ["models/fp32-mobilenet-v3-large.cc"],
Marat Dukhanc08cdf52019-12-09 09:17:51 -08005633 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08005634 copts = xnnpack_std_cxxopts(),
Marat Dukhanc08cdf52019-12-09 09:17:51 -08005635 linkstatic = True,
5636 deps = [
5637 ":XNNPACK",
5638 "@pthreadpool",
5639 ],
5640)
5641
5642cc_library(
Marat Dukhan4cea2322021-03-09 09:35:36 -08005643 name = "fp32_sparse_mobilenet_v3_large",
5644 srcs = ["models/fp32-sparse-mobilenet-v3-large.cc"],
5645 hdrs = ["models/models.h"],
5646 copts = xnnpack_std_cxxopts(),
5647 linkstatic = True,
5648 deps = [
5649 ":XNNPACK",
5650 "@pthreadpool",
5651 ],
5652)
5653
5654cc_library(
Marat Dukhan66f3ccd2020-09-14 16:09:38 -07005655 name = "fp16_mobilenet_v3_large",
5656 srcs = ["models/fp16-mobilenet-v3-large.cc"],
5657 hdrs = ["models/models.h"],
5658 copts = xnnpack_std_cxxopts(),
5659 linkstatic = True,
5660 deps = [
5661 ":XNNPACK",
5662 "@FP16",
5663 "@pthreadpool",
5664 ],
5665)
5666
5667cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07005668 name = "fp32_mobilenet_v3_small",
5669 srcs = ["models/fp32-mobilenet-v3-small.cc"],
Marat Dukhanc08cdf52019-12-09 09:17:51 -08005670 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08005671 copts = xnnpack_std_cxxopts(),
Marat Dukhanc08cdf52019-12-09 09:17:51 -08005672 linkstatic = True,
5673 deps = [
5674 ":XNNPACK",
5675 "@pthreadpool",
5676 ],
5677)
5678
Marat Dukhan66f3ccd2020-09-14 16:09:38 -07005679cc_library(
Marat Dukhan4cea2322021-03-09 09:35:36 -08005680 name = "fp32_sparse_mobilenet_v3_small",
5681 srcs = ["models/fp32-sparse-mobilenet-v3-small.cc"],
5682 hdrs = ["models/models.h"],
5683 copts = xnnpack_std_cxxopts(),
5684 linkstatic = True,
5685 deps = [
5686 ":XNNPACK",
5687 "@pthreadpool",
5688 ],
5689)
5690
5691cc_library(
Marat Dukhan66f3ccd2020-09-14 16:09:38 -07005692 name = "fp16_mobilenet_v3_small",
5693 srcs = ["models/fp16-mobilenet-v3-small.cc"],
5694 hdrs = ["models/models.h"],
5695 copts = xnnpack_std_cxxopts(),
5696 linkstatic = True,
5697 deps = [
5698 ":XNNPACK",
5699 "@FP16",
5700 "@pthreadpool",
5701 ],
5702)
5703
Marat Dukhanc068bb62019-10-04 13:24:39 -07005704xnnpack_benchmark(
Marat Dukhanef4416e2019-10-31 13:44:40 -07005705 name = "f32_dwconv_e2e_bench",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08005706 srcs = [
5707 "bench/f32-dwconv-e2e.cc",
5708 "bench/end2end.h",
5709 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanef4416e2019-10-31 13:44:40 -07005710 deps = MICROKERNEL_BENCHMARK_DEPS + [
5711 ":XNNPACK",
Marat Dukhan270a2c42020-06-26 16:45:52 -07005712 ":fp32_mobilenet_v1",
5713 ":fp32_mobilenet_v2",
5714 ":fp32_mobilenet_v3_large",
5715 ":fp32_mobilenet_v3_small",
Marat Dukhanef4416e2019-10-31 13:44:40 -07005716 ],
5717)
5718
5719xnnpack_benchmark(
Marat Dukhan5f18d262019-10-31 10:24:14 -07005720 name = "f32_gemm_e2e_bench",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08005721 srcs = [
5722 "bench/f32-gemm-e2e.cc",
5723 "bench/end2end.h",
5724 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan5f18d262019-10-31 10:24:14 -07005725 deps = MICROKERNEL_BENCHMARK_DEPS + [
5726 ":XNNPACK",
Marat Dukhan270a2c42020-06-26 16:45:52 -07005727 ":fp32_mobilenet_v1",
5728 ":fp32_mobilenet_v2",
5729 ":fp32_mobilenet_v3_large",
5730 ":fp32_mobilenet_v3_small",
Marat Dukhan5f18d262019-10-31 10:24:14 -07005731 ],
5732)
5733
5734xnnpack_benchmark(
Frank Barcharddc909cb2021-02-08 13:59:31 -08005735 name = "qs8_gemm_e2e_bench",
5736 srcs = [
5737 "bench/qs8-gemm-e2e.cc",
5738 "bench/end2end.h",
5739 ] + MICROKERNEL_BENCHMARK_HDRS,
5740 deps = MICROKERNEL_BENCHMARK_DEPS + [
5741 ":XNNPACK",
5742 ":qs8_mobilenet_v1",
5743 ":qs8_mobilenet_v2",
5744 ],
5745)
5746
5747xnnpack_benchmark(
Marat Dukhanc068bb62019-10-04 13:24:39 -07005748 name = "end2end_bench",
5749 srcs = ["bench/end2end.cc"],
5750 deps = [
5751 ":XNNPACK",
Frank Barchardc712fa42019-10-31 14:00:21 -07005752 ":bench_utils",
Marat Dukhan270a2c42020-06-26 16:45:52 -07005753 ":fp16_mobilenet_v1",
5754 ":fp16_mobilenet_v2",
Marat Dukhan66f3ccd2020-09-14 16:09:38 -07005755 ":fp16_mobilenet_v3_large",
5756 ":fp16_mobilenet_v3_small",
Marat Dukhan270a2c42020-06-26 16:45:52 -07005757 ":fp32_mobilenet_v1",
5758 ":fp32_mobilenet_v2",
5759 ":fp32_mobilenet_v3_large",
5760 ":fp32_mobilenet_v3_small",
Marat Dukhan4cea2322021-03-09 09:35:36 -08005761 ":fp32_sparse_mobilenet_v1",
5762 ":fp32_sparse_mobilenet_v2",
5763 ":fp32_sparse_mobilenet_v3_large",
5764 ":fp32_sparse_mobilenet_v3_small",
Marat Dukhan0743cdf2020-08-04 18:52:07 -07005765 ":qs8_mobilenet_v1",
Marat Dukhan70a96182020-09-03 17:13:58 -07005766 ":qs8_mobilenet_v2",
Marat Dukhan12a23bb2021-03-08 08:13:21 -08005767 ":qu8_mobilenet_v1",
Marat Dukhanc068bb62019-10-04 13:24:39 -07005768 "@pthreadpool",
5769 ],
5770)
5771
Marat Dukhan6adff4e2019-10-14 18:32:07 -07005772#################### Accuracy evaluation for math functions ####################
5773
5774xnnpack_benchmark(
Marat Dukhancf67ec62020-12-14 09:56:37 -08005775 name = "f32_exp_ulp_eval",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07005776 srcs = [
Marat Dukhancf67ec62020-12-14 09:56:37 -08005777 "eval/f32-exp-ulp.cc",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07005778 "src/xnnpack/AlignedAllocator.h",
5779 ] + ACCURACY_EVAL_HDRS,
Marat Dukhan3a305212020-12-06 19:24:27 -08005780 deps = ACCURACY_EVAL_DEPS + [
5781 ":bench_utils",
5782 "@cpuinfo",
5783 ],
Marat Dukhan6adff4e2019-10-14 18:32:07 -07005784)
5785
Marat Dukhan515c9772019-10-17 18:07:57 -07005786xnnpack_benchmark(
Marat Dukhancf67ec62020-12-14 09:56:37 -08005787 name = "f32_expminus_ulp_eval",
Marat Dukhan515c9772019-10-17 18:07:57 -07005788 srcs = [
Marat Dukhancf67ec62020-12-14 09:56:37 -08005789 "eval/f32-expminus-ulp.cc",
Marat Dukhan515c9772019-10-17 18:07:57 -07005790 "src/xnnpack/AlignedAllocator.h",
5791 ] + ACCURACY_EVAL_HDRS,
Marat Dukhan3a305212020-12-06 19:24:27 -08005792 deps = ACCURACY_EVAL_DEPS + [
5793 ":bench_utils",
5794 "@cpuinfo",
5795 ],
Marat Dukhan515c9772019-10-17 18:07:57 -07005796)
5797
Marat Dukhan98ba4412019-10-23 02:14:28 -07005798xnnpack_benchmark(
Marat Dukhancf67ec62020-12-14 09:56:37 -08005799 name = "f32_expm1minus_ulp_eval",
Marat Dukhana438aca2020-11-20 15:45:01 -08005800 srcs = [
Marat Dukhancf67ec62020-12-14 09:56:37 -08005801 "eval/f32-expm1minus-ulp.cc",
Marat Dukhana438aca2020-11-20 15:45:01 -08005802 "src/xnnpack/AlignedAllocator.h",
5803 ] + ACCURACY_EVAL_HDRS,
Marat Dukhande390d42020-11-29 19:32:18 -08005804 deps = ACCURACY_EVAL_DEPS + [
5805 ":bench_utils",
Marat Dukhan3a305212020-12-06 19:24:27 -08005806 "@cpuinfo",
Marat Dukhande390d42020-11-29 19:32:18 -08005807 ],
Marat Dukhana438aca2020-11-20 15:45:01 -08005808)
5809
5810xnnpack_benchmark(
Marat Dukhancf67ec62020-12-14 09:56:37 -08005811 name = "f32_extexp_ulp_eval",
Marat Dukhan98ba4412019-10-23 02:14:28 -07005812 srcs = [
Marat Dukhancf67ec62020-12-14 09:56:37 -08005813 "eval/f32-extexp-ulp.cc",
Marat Dukhan98ba4412019-10-23 02:14:28 -07005814 "src/xnnpack/AlignedAllocator.h",
5815 ] + ACCURACY_EVAL_HDRS,
Marat Dukhan3a305212020-12-06 19:24:27 -08005816 deps = ACCURACY_EVAL_DEPS + [
5817 ":bench_utils",
5818 "@cpuinfo",
5819 ],
Marat Dukhan98ba4412019-10-23 02:14:28 -07005820)
5821
Marat Dukhanf44f0222020-12-14 11:53:27 -08005822xnnpack_benchmark(
5823 name = "f32_sigmoid_ulp_eval",
5824 srcs = [
5825 "eval/f32-sigmoid-ulp.cc",
5826 "src/xnnpack/AlignedAllocator.h",
5827 ] + ACCURACY_EVAL_HDRS,
5828 deps = ACCURACY_EVAL_DEPS + [
5829 ":bench_utils",
5830 "@cpuinfo",
5831 ],
5832)
5833
5834xnnpack_benchmark(
5835 name = "f32_sqrt_ulp_eval",
5836 srcs = [
5837 "eval/f32-sqrt-ulp.cc",
5838 "src/xnnpack/AlignedAllocator.h",
5839 ] + ACCURACY_EVAL_HDRS,
5840 deps = ACCURACY_EVAL_DEPS + [
5841 ":bench_utils",
5842 "@cpuinfo",
5843 ],
5844)
5845
5846################### Accuracy verification for math functions ##################
5847
5848xnnpack_unit_test(
Marat Dukhanf7291fc2020-12-15 11:02:50 -08005849 name = "f32_exp_eval",
5850 srcs = [
5851 "eval/f32-exp.cc",
5852 "src/xnnpack/AlignedAllocator.h",
5853 "src/xnnpack/math-stubs.h",
5854 ] + MICROKERNEL_TEST_HDRS,
5855 automatic = False,
5856 deps = MICROKERNEL_TEST_DEPS,
5857)
5858
5859xnnpack_unit_test(
Marat Dukhanf44f0222020-12-14 11:53:27 -08005860 name = "f32_expm1minus_eval",
5861 srcs = [
5862 "eval/f32-expm1minus.cc",
5863 "src/xnnpack/AlignedAllocator.h",
5864 "src/xnnpack/math-stubs.h",
5865 ] + MICROKERNEL_TEST_HDRS,
5866 automatic = False,
5867 deps = MICROKERNEL_TEST_DEPS,
5868)
5869
Marat Dukhan8853b822020-05-07 12:19:01 -07005870xnnpack_unit_test(
Marat Dukhand28a5a22020-12-14 15:27:22 -08005871 name = "f32_expminus_eval",
5872 srcs = [
5873 "eval/f32-expminus.cc",
5874 "src/xnnpack/AlignedAllocator.h",
5875 "src/xnnpack/math-stubs.h",
5876 ] + MICROKERNEL_TEST_HDRS,
5877 automatic = False,
5878 deps = MICROKERNEL_TEST_DEPS,
5879)
5880
5881xnnpack_unit_test(
Marat Dukhan8853b822020-05-07 12:19:01 -07005882 name = "f32_roundne_eval",
5883 srcs = [
5884 "eval/f32-roundne.cc",
5885 "src/xnnpack/AlignedAllocator.h",
5886 "src/xnnpack/math-stubs.h",
5887 ] + MICROKERNEL_TEST_HDRS,
Marat Dukhan22eed3d2020-05-11 20:13:37 -07005888 automatic = False,
Marat Dukhan8853b822020-05-07 12:19:01 -07005889 deps = MICROKERNEL_TEST_DEPS,
5890)
5891
Marat Dukhan2dbb9442020-05-12 20:43:43 -07005892xnnpack_unit_test(
Marat Dukhanc9852ba2020-05-13 17:21:29 -07005893 name = "f32_roundd_eval",
5894 srcs = [
5895 "eval/f32-roundd.cc",
5896 "src/xnnpack/AlignedAllocator.h",
5897 "src/xnnpack/math-stubs.h",
5898 ] + MICROKERNEL_TEST_HDRS,
5899 automatic = False,
5900 deps = MICROKERNEL_TEST_DEPS,
5901)
5902
5903xnnpack_unit_test(
5904 name = "f32_roundu_eval",
5905 srcs = [
5906 "eval/f32-roundu.cc",
5907 "src/xnnpack/AlignedAllocator.h",
5908 "src/xnnpack/math-stubs.h",
5909 ] + MICROKERNEL_TEST_HDRS,
5910 automatic = False,
5911 deps = MICROKERNEL_TEST_DEPS,
5912)
5913
5914xnnpack_unit_test(
Marat Dukhan2dbb9442020-05-12 20:43:43 -07005915 name = "f32_roundz_eval",
5916 srcs = [
5917 "eval/f32-roundz.cc",
5918 "src/xnnpack/AlignedAllocator.h",
5919 "src/xnnpack/math-stubs.h",
5920 ] + MICROKERNEL_TEST_HDRS,
Marat Dukhanc9852ba2020-05-13 17:21:29 -07005921 automatic = False,
Marat Dukhan2dbb9442020-05-12 20:43:43 -07005922 deps = MICROKERNEL_TEST_DEPS,
5923)
5924
Marat Dukhan08c4a432019-10-03 09:29:21 -07005925######################### Unit tests for micro-kernels #########################
5926
5927xnnpack_unit_test(
Marat Dukhan6674d692021-05-05 22:27:00 -07005928 name = "f16_dwconv_minmax_test",
5929 srcs = [
5930 "test/f16-dwconv-minmax.cc",
5931 "test/dwconv-microkernel-tester.h",
5932 "src/xnnpack/AlignedAllocator.h",
5933 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
5934 deps = MICROKERNEL_TEST_DEPS + [":packing"],
5935)
5936
5937xnnpack_unit_test(
5938 name = "f16_gavgpool_minmax_test",
5939 srcs = [
5940 "test/f16-gavgpool-minmax.cc",
5941 "test/gavgpool-microkernel-tester.h",
5942 "src/xnnpack/AlignedAllocator.h",
5943 ] + MICROKERNEL_TEST_HDRS,
5944 deps = MICROKERNEL_TEST_DEPS,
5945)
5946
5947xnnpack_unit_test(
Marat Dukhande06f492020-04-09 00:19:31 -07005948 name = "f16_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005949 srcs = [
Marat Dukhande06f492020-04-09 00:19:31 -07005950 "test/f16-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005951 "test/gemm-microkernel-tester.h",
5952 "src/xnnpack/AlignedAllocator.h",
5953 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005954 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005955)
5956
5957xnnpack_unit_test(
Marat Dukhan6674d692021-05-05 22:27:00 -07005958 name = "f16_igemm_minmax_test",
5959 srcs = [
5960 "test/f16-igemm-minmax.cc",
5961 "test/gemm-microkernel-tester.h",
5962 "src/xnnpack/AlignedAllocator.h",
5963 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
5964 deps = MICROKERNEL_TEST_DEPS + [":packing"],
5965)
5966
5967xnnpack_unit_test(
Marat Dukhan355ab432020-04-09 19:01:52 -07005968 name = "f16_spmm_minmax_test",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08005969 srcs = [
Marat Dukhan355ab432020-04-09 19:01:52 -07005970 "test/f16-spmm-minmax.cc",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08005971 "test/spmm-microkernel-tester.h",
5972 "src/xnnpack/AlignedAllocator.h",
5973 ] + MICROKERNEL_TEST_HDRS,
5974 deps = MICROKERNEL_TEST_DEPS,
5975)
5976
5977xnnpack_unit_test(
Marat Dukhan6674d692021-05-05 22:27:00 -07005978 name = "f16_vadd_minmax_test",
5979 srcs = [
5980 "test/f16-vadd-minmax.cc",
5981 "test/vbinary-microkernel-tester.h",
5982 ] + MICROKERNEL_TEST_HDRS,
5983 deps = MICROKERNEL_TEST_DEPS,
5984)
5985
5986xnnpack_unit_test(
5987 name = "f16_vaddc_minmax_test",
5988 srcs = [
5989 "test/f16-vaddc-minmax.cc",
5990 "test/vbinaryc-microkernel-tester.h",
5991 ] + MICROKERNEL_TEST_HDRS,
5992 deps = MICROKERNEL_TEST_DEPS,
5993)
5994
5995xnnpack_unit_test(
5996 name = "f16_vclamp_test",
5997 srcs = [
5998 "test/f16-vclamp.cc",
5999 "test/vclamp-microkernel-tester.h",
6000 ] + MICROKERNEL_TEST_HDRS,
6001 deps = MICROKERNEL_TEST_DEPS,
6002)
6003
6004xnnpack_unit_test(
6005 name = "f16_vdiv_minmax_test",
6006 srcs = [
6007 "test/f16-vdiv-minmax.cc",
6008 "test/vbinary-microkernel-tester.h",
6009 ] + MICROKERNEL_TEST_HDRS,
6010 deps = MICROKERNEL_TEST_DEPS,
6011)
6012
6013xnnpack_unit_test(
6014 name = "f16_vdivc_minmax_test",
6015 srcs = [
6016 "test/f16-vdivc-minmax.cc",
6017 "test/vbinaryc-microkernel-tester.h",
6018 ] + MICROKERNEL_TEST_HDRS,
6019 deps = MICROKERNEL_TEST_DEPS,
6020)
6021
6022xnnpack_unit_test(
6023 name = "f16_vrdivc_minmax_test",
6024 srcs = [
6025 "test/f16-vrdivc-minmax.cc",
6026 "test/vbinaryc-microkernel-tester.h",
6027 ] + MICROKERNEL_TEST_HDRS,
6028 deps = MICROKERNEL_TEST_DEPS,
6029)
6030
6031xnnpack_unit_test(
6032 name = "f16_vhswish_test",
6033 srcs = [
6034 "test/f16-vhswish.cc",
6035 "test/vhswish-microkernel-tester.h",
6036 ] + MICROKERNEL_TEST_HDRS,
6037 deps = MICROKERNEL_TEST_DEPS,
6038)
6039
6040xnnpack_unit_test(
6041 name = "f16_vmax_test",
6042 srcs = [
6043 "test/f16-vmax.cc",
6044 "test/vbinary-microkernel-tester.h",
6045 ] + MICROKERNEL_TEST_HDRS,
6046 deps = MICROKERNEL_TEST_DEPS,
6047)
6048
6049xnnpack_unit_test(
6050 name = "f16_vmaxc_test",
6051 srcs = [
6052 "test/f16-vmaxc.cc",
6053 "test/vbinaryc-microkernel-tester.h",
6054 ] + MICROKERNEL_TEST_HDRS,
6055 deps = MICROKERNEL_TEST_DEPS,
6056)
6057
6058xnnpack_unit_test(
6059 name = "f16_vmin_test",
6060 srcs = [
6061 "test/f16-vmin.cc",
6062 "test/vbinary-microkernel-tester.h",
6063 ] + MICROKERNEL_TEST_HDRS,
6064 deps = MICROKERNEL_TEST_DEPS,
6065)
6066
6067xnnpack_unit_test(
6068 name = "f16_vminc_test",
6069 srcs = [
6070 "test/f16-vminc.cc",
6071 "test/vbinaryc-microkernel-tester.h",
6072 ] + MICROKERNEL_TEST_HDRS,
6073 deps = MICROKERNEL_TEST_DEPS,
6074)
6075
6076xnnpack_unit_test(
6077 name = "f16_vmul_minmax_test",
6078 srcs = [
6079 "test/f16-vmul-minmax.cc",
6080 "test/vbinary-microkernel-tester.h",
6081 ] + MICROKERNEL_TEST_HDRS,
6082 deps = MICROKERNEL_TEST_DEPS,
6083)
6084
6085xnnpack_unit_test(
6086 name = "f16_vmulc_minmax_test",
6087 srcs = [
6088 "test/f16-vmulc-minmax.cc",
6089 "test/vbinaryc-microkernel-tester.h",
6090 ] + MICROKERNEL_TEST_HDRS,
6091 deps = MICROKERNEL_TEST_DEPS,
6092)
6093
6094xnnpack_unit_test(
6095 name = "f16_vmulcaddc_minmax_test",
6096 srcs = [
6097 "test/f16-vmulcaddc-minmax.cc",
6098 "test/vmulcaddc-microkernel-tester.h",
6099 "src/xnnpack/AlignedAllocator.h",
6100 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
6101 deps = MICROKERNEL_TEST_DEPS + [":packing"],
6102)
6103
6104xnnpack_unit_test(
6105 name = "f16_vsub_minmax_test",
6106 srcs = [
6107 "test/f16-vsub-minmax.cc",
6108 "test/vbinary-microkernel-tester.h",
6109 ] + MICROKERNEL_TEST_HDRS,
6110 deps = MICROKERNEL_TEST_DEPS,
6111)
6112
6113xnnpack_unit_test(
6114 name = "f16_vsubc_minmax_test",
6115 srcs = [
6116 "test/f16-vsubc-minmax.cc",
6117 "test/vbinaryc-microkernel-tester.h",
6118 ] + MICROKERNEL_TEST_HDRS,
6119 deps = MICROKERNEL_TEST_DEPS,
6120)
6121
6122xnnpack_unit_test(
6123 name = "f16_vrsubc_minmax_test",
6124 srcs = [
6125 "test/f16-vrsubc-minmax.cc",
6126 "test/vbinaryc-microkernel-tester.h",
6127 ] + MICROKERNEL_TEST_HDRS,
6128 deps = MICROKERNEL_TEST_DEPS,
6129)
6130
6131xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07006132 name = "f32_argmaxpool_test",
6133 srcs = [
6134 "test/f32-argmaxpool.cc",
6135 "test/argmaxpool-microkernel-tester.h",
6136 "src/xnnpack/AlignedAllocator.h",
6137 ] + MICROKERNEL_TEST_HDRS,
6138 deps = MICROKERNEL_TEST_DEPS,
6139)
6140
6141xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07006142 name = "f32_avgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006143 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07006144 "test/f32-avgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006145 "test/avgpool-microkernel-tester.h",
6146 "src/xnnpack/AlignedAllocator.h",
6147 ] + MICROKERNEL_TEST_HDRS,
6148 deps = MICROKERNEL_TEST_DEPS,
6149)
6150
6151xnnpack_unit_test(
Marat Dukhan660fd192020-03-10 04:55:30 -07006152 name = "f32_ibilinear_test",
Marat Dukhan35dacfb2019-11-07 19:18:16 -08006153 srcs = [
Marat Dukhan660fd192020-03-10 04:55:30 -07006154 "test/f32-ibilinear.cc",
6155 "test/ibilinear-microkernel-tester.h",
Marat Dukhan35dacfb2019-11-07 19:18:16 -08006156 "src/xnnpack/AlignedAllocator.h",
6157 ] + MICROKERNEL_TEST_HDRS,
6158 deps = MICROKERNEL_TEST_DEPS,
6159)
6160
6161xnnpack_unit_test(
XNNPACK Team21432672020-10-19 19:58:48 -07006162 name = "f32_ibilinear_chw_test",
6163 srcs = [
6164 "test/f32-ibilinear-chw.cc",
XNNPACK Team6be46b22020-10-22 23:34:54 -07006165 "test/ibilinear-microkernel-tester.h",
XNNPACK Team21432672020-10-19 19:58:48 -07006166 "src/xnnpack/AlignedAllocator.h",
6167 ] + MICROKERNEL_TEST_HDRS,
6168 deps = MICROKERNEL_TEST_DEPS,
6169)
6170
6171xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07006172 name = "f32_igemm_test",
6173 srcs = [
6174 "test/f32-igemm.cc",
6175 "test/gemm-microkernel-tester.h",
6176 "src/xnnpack/AlignedAllocator.h",
6177 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07006178 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan163a7e62020-04-09 04:19:26 -07006179)
6180
6181xnnpack_unit_test(
Marat Dukhan467f6362020-05-22 23:21:55 -07006182 name = "f32_igemm_relu_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006183 srcs = [
Marat Dukhan467f6362020-05-22 23:21:55 -07006184 "test/f32-igemm-relu.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006185 "test/gemm-microkernel-tester.h",
6186 "src/xnnpack/AlignedAllocator.h",
6187 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07006188 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07006189)
6190
6191xnnpack_unit_test(
Marat Dukhane207b7b2020-05-28 16:27:42 -07006192 name = "f32_igemm_minmax_test",
6193 srcs = [
6194 "test/f32-igemm-minmax.cc",
6195 "test/gemm-microkernel-tester.h",
6196 "src/xnnpack/AlignedAllocator.h",
6197 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07006198 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhane207b7b2020-05-28 16:27:42 -07006199)
6200
6201xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07006202 name = "f32_conv_hwc_test",
6203 srcs = [
6204 "test/f32-conv-hwc.cc",
6205 "test/conv-hwc-microkernel-tester.h",
6206 "src/xnnpack/AlignedAllocator.h",
6207 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07006208 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07006209)
6210
6211xnnpack_unit_test(
Marat Dukhan1f29b802020-05-15 23:46:39 -07006212 name = "f32_conv_hwc2chw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006213 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07006214 "test/f32-conv-hwc2chw.cc",
6215 "test/conv-hwc2chw-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006216 "src/xnnpack/AlignedAllocator.h",
6217 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07006218 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07006219)
6220
6221xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07006222 name = "f32_dwconv_test",
6223 srcs = [
6224 "test/f32-dwconv.cc",
6225 "test/dwconv-microkernel-tester.h",
6226 "src/xnnpack/AlignedAllocator.h",
6227 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07006228 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan163a7e62020-04-09 04:19:26 -07006229)
6230
6231xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07006232 name = "f32_dwconv_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006233 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07006234 "test/f32-dwconv-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006235 "test/dwconv-microkernel-tester.h",
6236 "src/xnnpack/AlignedAllocator.h",
6237 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07006238 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07006239)
6240
6241xnnpack_unit_test(
Marat Dukhanbf715f92020-10-23 20:17:00 -07006242 name = "f32_dwconv2d_chw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006243 srcs = [
Marat Dukhanbf715f92020-10-23 20:17:00 -07006244 "test/f32-dwconv2d-chw.cc",
6245 "test/dwconv2d-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006246 "src/xnnpack/AlignedAllocator.h",
6247 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07006248 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07006249)
6250
6251xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07006252 name = "f32_gavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006253 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07006254 "test/f32-gavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006255 "test/gavgpool-microkernel-tester.h",
6256 "src/xnnpack/AlignedAllocator.h",
6257 ] + MICROKERNEL_TEST_HDRS,
6258 deps = MICROKERNEL_TEST_DEPS,
6259)
6260
6261xnnpack_unit_test(
Marat Dukhan1f29b802020-05-15 23:46:39 -07006262 name = "f32_gavgpool_cw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006263 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07006264 "test/f32-gavgpool-cw.cc",
6265 "test/gavgpool-cw-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006266 "src/xnnpack/AlignedAllocator.h",
6267 ] + MICROKERNEL_TEST_HDRS,
6268 deps = MICROKERNEL_TEST_DEPS,
6269)
6270
6271xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07006272 name = "f32_gemm_test",
6273 srcs = [
6274 "test/f32-gemm.cc",
6275 "test/gemm-microkernel-tester.h",
6276 "src/xnnpack/AlignedAllocator.h",
6277 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07006278 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan163a7e62020-04-09 04:19:26 -07006279)
6280
6281xnnpack_unit_test(
Marat Dukhan467f6362020-05-22 23:21:55 -07006282 name = "f32_gemm_relu_test",
6283 srcs = [
6284 "test/f32-gemm-relu.cc",
6285 "test/gemm-microkernel-tester.h",
6286 "src/xnnpack/AlignedAllocator.h",
6287 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07006288 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan467f6362020-05-22 23:21:55 -07006289)
6290
6291xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07006292 name = "f32_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006293 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07006294 "test/f32-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006295 "test/gemm-microkernel-tester.h",
6296 "src/xnnpack/AlignedAllocator.h",
6297 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07006298 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07006299)
6300
6301xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07006302 name = "f32_gemminc_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006303 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07006304 "test/f32-gemminc-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006305 "test/gemm-microkernel-tester.h",
6306 "src/xnnpack/AlignedAllocator.h",
6307 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07006308 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07006309)
6310
6311xnnpack_unit_test(
Marat Dukhan6674d692021-05-05 22:27:00 -07006312 name = "f32_vhswish_test",
Frank Barchardb1966592020-05-12 13:47:06 -07006313 srcs = [
Marat Dukhan6674d692021-05-05 22:27:00 -07006314 "test/f32-vhswish.cc",
Marat Dukhan949b6e72021-05-13 11:21:06 -07006315 "test/vunary-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006316 ] + MICROKERNEL_TEST_HDRS,
6317 deps = MICROKERNEL_TEST_DEPS,
6318)
6319
6320xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07006321 name = "f32_maxpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006322 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07006323 "test/f32-maxpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006324 "test/maxpool-microkernel-tester.h",
6325 ] + MICROKERNEL_TEST_HDRS,
6326 deps = MICROKERNEL_TEST_DEPS,
6327)
6328
6329xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07006330 name = "f32_pavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006331 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07006332 "test/f32-pavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006333 "test/avgpool-microkernel-tester.h",
6334 "src/xnnpack/AlignedAllocator.h",
6335 ] + MICROKERNEL_TEST_HDRS,
6336 deps = MICROKERNEL_TEST_DEPS,
6337)
6338
6339xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07006340 name = "f32_ppmm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006341 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07006342 "test/f32-ppmm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006343 "test/gemm-microkernel-tester.h",
6344 "src/xnnpack/AlignedAllocator.h",
6345 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07006346 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07006347)
6348
6349xnnpack_unit_test(
Frank Barchardb1966592020-05-12 13:47:06 -07006350 name = "f16_prelu_test",
6351 srcs = [
6352 "test/f16-prelu.cc",
6353 "test/prelu-microkernel-tester.h",
6354 "src/xnnpack/AlignedAllocator.h",
6355 ] + MICROKERNEL_TEST_HDRS,
6356 deps = MICROKERNEL_TEST_DEPS,
6357)
6358
6359xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07006360 name = "f32_prelu_test",
6361 srcs = [
6362 "test/f32-prelu.cc",
6363 "test/prelu-microkernel-tester.h",
6364 "src/xnnpack/AlignedAllocator.h",
6365 ] + MICROKERNEL_TEST_HDRS,
6366 deps = MICROKERNEL_TEST_DEPS,
6367)
6368
6369xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07006370 name = "f32_raddexpminusmax_test",
6371 srcs = [
6372 "test/f32-raddexpminusmax.cc",
6373 "test/raddexpminusmax-microkernel-tester.h",
6374 ] + MICROKERNEL_TEST_HDRS,
6375 deps = MICROKERNEL_TEST_DEPS,
6376)
6377
6378xnnpack_unit_test(
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07006379 name = "f32_raddextexp_test",
6380 srcs = [
6381 "test/f32-raddextexp.cc",
6382 "test/raddextexp-microkernel-tester.h",
6383 ] + MICROKERNEL_TEST_HDRS,
6384 deps = MICROKERNEL_TEST_DEPS,
6385)
6386
6387xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07006388 name = "f32_raddstoreexpminusmax_test",
6389 srcs = [
6390 "test/f32-raddstoreexpminusmax.cc",
6391 "test/raddstoreexpminusmax-microkernel-tester.h",
6392 ] + MICROKERNEL_TEST_HDRS,
6393 deps = MICROKERNEL_TEST_DEPS,
6394)
6395
6396xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07006397 name = "f32_rmax_test",
6398 srcs = [
6399 "test/f32-rmax.cc",
6400 "test/rmax-microkernel-tester.h",
6401 ] + MICROKERNEL_TEST_HDRS,
6402 deps = MICROKERNEL_TEST_DEPS,
6403)
6404
6405xnnpack_unit_test(
Marat Dukhan355ab432020-04-09 19:01:52 -07006406 name = "f32_spmm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006407 srcs = [
Marat Dukhan355ab432020-04-09 19:01:52 -07006408 "test/f32-spmm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006409 "test/spmm-microkernel-tester.h",
6410 "src/xnnpack/AlignedAllocator.h",
6411 ] + MICROKERNEL_TEST_HDRS,
6412 deps = MICROKERNEL_TEST_DEPS,
6413)
6414
6415xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07006416 name = "f32_vabs_test",
6417 srcs = [
6418 "test/f32-vabs.cc",
6419 "test/vunary-microkernel-tester.h",
6420 ] + MICROKERNEL_TEST_HDRS,
6421 deps = MICROKERNEL_TEST_DEPS,
6422)
6423
6424xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07006425 name = "f32_vadd_test",
6426 srcs = [
6427 "test/f32-vadd.cc",
6428 "test/vbinary-microkernel-tester.h",
6429 ] + MICROKERNEL_TEST_HDRS,
6430 deps = MICROKERNEL_TEST_DEPS,
6431)
6432
6433xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006434 name = "f32_vadd_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006435 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006436 "test/f32-vadd-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08006437 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08006438 ] + MICROKERNEL_TEST_HDRS,
6439 deps = MICROKERNEL_TEST_DEPS,
6440)
6441
6442xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07006443 name = "f32_vadd_relu_test",
6444 srcs = [
6445 "test/f32-vadd-relu.cc",
6446 "test/vbinary-microkernel-tester.h",
6447 ] + MICROKERNEL_TEST_HDRS,
6448 deps = MICROKERNEL_TEST_DEPS,
6449)
6450
6451xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07006452 name = "f32_vaddc_test",
6453 srcs = [
6454 "test/f32-vaddc.cc",
6455 "test/vbinaryc-microkernel-tester.h",
6456 ] + MICROKERNEL_TEST_HDRS,
6457 deps = MICROKERNEL_TEST_DEPS,
6458)
6459
6460xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006461 name = "f32_vaddc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08006462 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006463 "test/f32-vaddc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08006464 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006465 ] + MICROKERNEL_TEST_HDRS,
6466 deps = MICROKERNEL_TEST_DEPS,
6467)
6468
6469xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07006470 name = "f32_vaddc_relu_test",
6471 srcs = [
6472 "test/f32-vaddc-relu.cc",
6473 "test/vbinaryc-microkernel-tester.h",
6474 ] + MICROKERNEL_TEST_HDRS,
6475 deps = MICROKERNEL_TEST_DEPS,
6476)
6477
6478xnnpack_unit_test(
Marat Dukhan6674d692021-05-05 22:27:00 -07006479 name = "f32_vclamp_test",
6480 srcs = [
6481 "test/f32-vclamp.cc",
Marat Dukhan60d3f242021-05-13 11:59:02 -07006482 "test/vunary-microkernel-tester.h",
Marat Dukhan6674d692021-05-05 22:27:00 -07006483 ] + MICROKERNEL_TEST_HDRS,
6484 deps = MICROKERNEL_TEST_DEPS,
6485)
6486
6487xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07006488 name = "f32_vdiv_test",
6489 srcs = [
6490 "test/f32-vdiv.cc",
6491 "test/vbinary-microkernel-tester.h",
6492 ] + MICROKERNEL_TEST_HDRS,
6493 deps = MICROKERNEL_TEST_DEPS,
6494)
6495
6496xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006497 name = "f32_vdiv_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08006498 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006499 "test/f32-vdiv-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08006500 "test/vbinary-microkernel-tester.h",
6501 ] + MICROKERNEL_TEST_HDRS,
6502 deps = MICROKERNEL_TEST_DEPS,
6503)
6504
6505xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07006506 name = "f32_vdiv_relu_test",
6507 srcs = [
6508 "test/f32-vdiv-relu.cc",
6509 "test/vbinary-microkernel-tester.h",
6510 ] + MICROKERNEL_TEST_HDRS,
6511 deps = MICROKERNEL_TEST_DEPS,
6512)
6513
6514xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07006515 name = "f32_vdivc_test",
6516 srcs = [
6517 "test/f32-vdivc.cc",
6518 "test/vbinaryc-microkernel-tester.h",
6519 ] + MICROKERNEL_TEST_HDRS,
6520 deps = MICROKERNEL_TEST_DEPS,
6521)
6522
6523xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006524 name = "f32_vdivc_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08006525 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006526 "test/f32-vdivc-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08006527 "test/vbinaryc-microkernel-tester.h",
6528 ] + MICROKERNEL_TEST_HDRS,
6529 deps = MICROKERNEL_TEST_DEPS,
6530)
6531
6532xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07006533 name = "f32_vdivc_relu_test",
6534 srcs = [
6535 "test/f32-vdivc-relu.cc",
6536 "test/vbinaryc-microkernel-tester.h",
6537 ] + MICROKERNEL_TEST_HDRS,
6538 deps = MICROKERNEL_TEST_DEPS,
6539)
6540
6541xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07006542 name = "f32_vrdivc_test",
6543 srcs = [
6544 "test/f32-vrdivc.cc",
6545 "test/vbinaryc-microkernel-tester.h",
6546 ] + MICROKERNEL_TEST_HDRS,
6547 deps = MICROKERNEL_TEST_DEPS,
6548)
6549
6550xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006551 name = "f32_vrdivc_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08006552 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006553 "test/f32-vrdivc-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08006554 "test/vbinaryc-microkernel-tester.h",
6555 ] + MICROKERNEL_TEST_HDRS,
6556 deps = MICROKERNEL_TEST_DEPS,
6557)
6558
6559xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07006560 name = "f32_vrdivc_relu_test",
6561 srcs = [
6562 "test/f32-vrdivc-relu.cc",
6563 "test/vbinaryc-microkernel-tester.h",
6564 ] + MICROKERNEL_TEST_HDRS,
6565 deps = MICROKERNEL_TEST_DEPS,
6566)
6567
6568xnnpack_unit_test(
Marat Dukhaned6baaf2020-12-01 15:07:08 -08006569 name = "f32_velu_test",
6570 srcs = [
6571 "test/f32-velu.cc",
6572 "test/vunary-microkernel-tester.h",
6573 ] + MICROKERNEL_TEST_HDRS,
6574 deps = MICROKERNEL_TEST_DEPS,
6575)
6576
6577xnnpack_unit_test(
Marat Dukhan403b7d42019-12-05 12:49:11 -08006578 name = "f32_vmax_test",
6579 srcs = [
6580 "test/f32-vmax.cc",
6581 "test/vbinary-microkernel-tester.h",
6582 ] + MICROKERNEL_TEST_HDRS,
6583 deps = MICROKERNEL_TEST_DEPS,
6584)
6585
6586xnnpack_unit_test(
6587 name = "f32_vmaxc_test",
6588 srcs = [
6589 "test/f32-vmaxc.cc",
6590 "test/vbinaryc-microkernel-tester.h",
6591 ] + MICROKERNEL_TEST_HDRS,
6592 deps = MICROKERNEL_TEST_DEPS,
6593)
6594
6595xnnpack_unit_test(
6596 name = "f32_vmin_test",
6597 srcs = [
6598 "test/f32-vmin.cc",
6599 "test/vbinary-microkernel-tester.h",
6600 ] + MICROKERNEL_TEST_HDRS,
6601 deps = MICROKERNEL_TEST_DEPS,
6602)
6603
6604xnnpack_unit_test(
6605 name = "f32_vminc_test",
6606 srcs = [
6607 "test/f32-vminc.cc",
6608 "test/vbinaryc-microkernel-tester.h",
6609 ] + MICROKERNEL_TEST_HDRS,
6610 deps = MICROKERNEL_TEST_DEPS,
6611)
6612
6613xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07006614 name = "f32_vmul_test",
6615 srcs = [
6616 "test/f32-vmul.cc",
6617 "test/vbinary-microkernel-tester.h",
6618 ] + MICROKERNEL_TEST_HDRS,
6619 deps = MICROKERNEL_TEST_DEPS,
6620)
6621
6622xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006623 name = "f32_vmul_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006624 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006625 "test/f32-vmul-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08006626 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08006627 ] + MICROKERNEL_TEST_HDRS,
6628 deps = MICROKERNEL_TEST_DEPS,
6629)
6630
6631xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07006632 name = "f32_vmul_relu_test",
6633 srcs = [
6634 "test/f32-vmul-relu.cc",
6635 "test/vbinary-microkernel-tester.h",
6636 ] + MICROKERNEL_TEST_HDRS,
6637 deps = MICROKERNEL_TEST_DEPS,
6638)
6639
6640xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07006641 name = "f32_vmulc_test",
6642 srcs = [
6643 "test/f32-vmulc.cc",
6644 "test/vbinaryc-microkernel-tester.h",
6645 ] + MICROKERNEL_TEST_HDRS,
6646 deps = MICROKERNEL_TEST_DEPS,
6647)
6648
6649xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006650 name = "f32_vmulc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08006651 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006652 "test/f32-vmulc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08006653 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006654 ] + MICROKERNEL_TEST_HDRS,
6655 deps = MICROKERNEL_TEST_DEPS,
6656)
6657
6658xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07006659 name = "f32_vmulc_relu_test",
6660 srcs = [
6661 "test/f32-vmulc-relu.cc",
6662 "test/vbinaryc-microkernel-tester.h",
6663 ] + MICROKERNEL_TEST_HDRS,
6664 deps = MICROKERNEL_TEST_DEPS,
6665)
6666
6667xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07006668 name = "f32_vmulcaddc_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006669 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07006670 "test/f32-vmulcaddc-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006671 "test/vmulcaddc-microkernel-tester.h",
6672 "src/xnnpack/AlignedAllocator.h",
6673 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07006674 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07006675)
6676
6677xnnpack_unit_test(
Marat Dukhan8cc7efe2020-06-10 16:24:27 -07006678 name = "f32_vlrelu_test",
6679 srcs = [
6680 "test/f32-vlrelu.cc",
6681 "test/vunary-microkernel-tester.h",
6682 ] + MICROKERNEL_TEST_HDRS,
6683 deps = MICROKERNEL_TEST_DEPS,
6684)
6685
6686xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07006687 name = "f32_vneg_test",
6688 srcs = [
6689 "test/f32-vneg.cc",
6690 "test/vunary-microkernel-tester.h",
6691 ] + MICROKERNEL_TEST_HDRS,
6692 deps = MICROKERNEL_TEST_DEPS,
6693)
6694
6695xnnpack_unit_test(
Marat Dukhan6674d692021-05-05 22:27:00 -07006696 name = "f32_vrelu_test",
6697 srcs = [
6698 "test/f32-vrelu.cc",
6699 "test/vunary-microkernel-tester.h",
6700 ] + MICROKERNEL_TEST_HDRS,
6701 deps = MICROKERNEL_TEST_DEPS,
6702)
6703
6704xnnpack_unit_test(
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07006705 name = "f32_vrndne_test",
6706 srcs = [
6707 "test/f32-vrndne.cc",
6708 "test/vunary-microkernel-tester.h",
6709 ] + MICROKERNEL_TEST_HDRS,
6710 deps = MICROKERNEL_TEST_DEPS,
6711)
6712
6713xnnpack_unit_test(
6714 name = "f32_vrndz_test",
6715 srcs = [
6716 "test/f32-vrndz.cc",
6717 "test/vunary-microkernel-tester.h",
6718 ] + MICROKERNEL_TEST_HDRS,
6719 deps = MICROKERNEL_TEST_DEPS,
6720)
6721
6722xnnpack_unit_test(
6723 name = "f32_vrndu_test",
6724 srcs = [
6725 "test/f32-vrndu.cc",
6726 "test/vunary-microkernel-tester.h",
6727 ] + MICROKERNEL_TEST_HDRS,
6728 deps = MICROKERNEL_TEST_DEPS,
6729)
6730
6731xnnpack_unit_test(
6732 name = "f32_vrndd_test",
6733 srcs = [
6734 "test/f32-vrndd.cc",
6735 "test/vunary-microkernel-tester.h",
6736 ] + MICROKERNEL_TEST_HDRS,
6737 deps = MICROKERNEL_TEST_DEPS,
6738)
6739
6740xnnpack_unit_test(
Marat Dukhan05ac8e32019-10-21 15:39:33 -07006741 name = "f32_vscale_test",
6742 srcs = [
6743 "test/f32-vscale.cc",
6744 "test/vscale-microkernel-tester.h",
6745 ] + MICROKERNEL_TEST_HDRS,
6746 deps = MICROKERNEL_TEST_DEPS,
6747)
6748
6749xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07006750 name = "f32_vscaleexpminusmax_test",
6751 srcs = [
6752 "test/f32-vscaleexpminusmax.cc",
6753 "test/vscaleexpminusmax-microkernel-tester.h",
6754 ] + MICROKERNEL_TEST_HDRS,
6755 deps = MICROKERNEL_TEST_DEPS,
6756)
6757
6758xnnpack_unit_test(
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07006759 name = "f32_vscaleextexp_test",
6760 srcs = [
6761 "test/f32-vscaleextexp.cc",
6762 "test/vscaleextexp-microkernel-tester.h",
6763 ] + MICROKERNEL_TEST_HDRS,
6764 deps = MICROKERNEL_TEST_DEPS,
6765)
6766
6767xnnpack_unit_test(
Marat Dukhan6674d692021-05-05 22:27:00 -07006768 name = "f32_vsigmoid_test",
6769 srcs = [
6770 "test/f32-vsigmoid.cc",
6771 "test/vunary-microkernel-tester.h",
6772 ] + MICROKERNEL_TEST_HDRS,
6773 deps = MICROKERNEL_TEST_DEPS,
6774)
6775
6776xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07006777 name = "f32_vsqr_test",
6778 srcs = [
6779 "test/f32-vsqr.cc",
6780 "test/vunary-microkernel-tester.h",
6781 ] + MICROKERNEL_TEST_HDRS,
6782 deps = MICROKERNEL_TEST_DEPS,
6783)
6784
6785xnnpack_unit_test(
Marat Dukhan13bafb02020-06-05 00:43:11 -07006786 name = "f32_vsqrdiff_test",
6787 srcs = [
6788 "test/f32-vsqrdiff.cc",
6789 "test/vbinary-microkernel-tester.h",
6790 ] + MICROKERNEL_TEST_HDRS,
6791 deps = MICROKERNEL_TEST_DEPS,
6792)
6793
6794xnnpack_unit_test(
6795 name = "f32_vsqrdiffc_test",
6796 srcs = [
6797 "test/f32-vsqrdiffc.cc",
6798 "test/vbinaryc-microkernel-tester.h",
6799 ] + MICROKERNEL_TEST_HDRS,
6800 deps = MICROKERNEL_TEST_DEPS,
6801)
6802
6803xnnpack_unit_test(
Marat Dukhanf4db2f32020-06-30 10:55:30 -07006804 name = "f32_vsqrt_test",
6805 srcs = [
6806 "test/f32-vsqrt.cc",
6807 "test/vunary-microkernel-tester.h",
6808 ] + MICROKERNEL_TEST_HDRS,
6809 deps = MICROKERNEL_TEST_DEPS,
6810)
6811
6812xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07006813 name = "f32_vsub_test",
6814 srcs = [
6815 "test/f32-vsub.cc",
6816 "test/vbinary-microkernel-tester.h",
6817 ] + MICROKERNEL_TEST_HDRS,
6818 deps = MICROKERNEL_TEST_DEPS,
6819)
6820
6821xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006822 name = "f32_vsub_minmax_test",
Marat Dukhan97579532019-10-18 16:40:39 -07006823 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006824 "test/f32-vsub-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08006825 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08006826 ] + MICROKERNEL_TEST_HDRS,
6827 deps = MICROKERNEL_TEST_DEPS,
6828)
6829
6830xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07006831 name = "f32_vsub_relu_test",
6832 srcs = [
6833 "test/f32-vsub-relu.cc",
6834 "test/vbinary-microkernel-tester.h",
6835 ] + MICROKERNEL_TEST_HDRS,
6836 deps = MICROKERNEL_TEST_DEPS,
6837)
6838
6839xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07006840 name = "f32_vsubc_test",
6841 srcs = [
6842 "test/f32-vsubc.cc",
6843 "test/vbinaryc-microkernel-tester.h",
6844 ] + MICROKERNEL_TEST_HDRS,
6845 deps = MICROKERNEL_TEST_DEPS,
6846)
6847
6848xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006849 name = "f32_vsubc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08006850 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006851 "test/f32-vsubc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08006852 "test/vbinaryc-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08006853 ] + MICROKERNEL_TEST_HDRS,
6854 deps = MICROKERNEL_TEST_DEPS,
6855)
6856
6857xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07006858 name = "f32_vsubc_relu_test",
6859 srcs = [
6860 "test/f32-vsubc-relu.cc",
6861 "test/vbinaryc-microkernel-tester.h",
6862 ] + MICROKERNEL_TEST_HDRS,
6863 deps = MICROKERNEL_TEST_DEPS,
6864)
6865
6866xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07006867 name = "f32_vrsubc_test",
6868 srcs = [
6869 "test/f32-vrsubc.cc",
6870 "test/vbinaryc-microkernel-tester.h",
6871 ] + MICROKERNEL_TEST_HDRS,
6872 deps = MICROKERNEL_TEST_DEPS,
6873)
6874
6875xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006876 name = "f32_vrsubc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08006877 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006878 "test/f32-vrsubc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08006879 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan97579532019-10-18 16:40:39 -07006880 ] + MICROKERNEL_TEST_HDRS,
6881 deps = MICROKERNEL_TEST_DEPS,
6882)
6883
6884xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07006885 name = "f32_vrsubc_relu_test",
6886 srcs = [
6887 "test/f32-vrsubc-relu.cc",
6888 "test/vbinaryc-microkernel-tester.h",
6889 ] + MICROKERNEL_TEST_HDRS,
6890 deps = MICROKERNEL_TEST_DEPS,
6891)
6892
6893xnnpack_unit_test(
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07006894 name = "qs8_dwconv_minmax_test",
6895 srcs = [
6896 "test/qs8-dwconv-minmax.cc",
6897 "test/dwconv-microkernel-tester.h",
6898 "src/xnnpack/AlignedAllocator.h",
6899 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
6900 deps = MICROKERNEL_TEST_DEPS + [":packing"],
6901)
6902
6903xnnpack_unit_test(
Marat Dukhan4ed53f42020-08-06 01:12:55 -07006904 name = "qs8_gavgpool_minmax_test",
6905 srcs = [
6906 "test/qs8-gavgpool-minmax.cc",
6907 "test/gavgpool-microkernel-tester.h",
6908 "src/xnnpack/AlignedAllocator.h",
6909 ] + MICROKERNEL_TEST_HDRS,
6910 deps = MICROKERNEL_TEST_DEPS,
6911)
6912
6913xnnpack_unit_test(
Marat Dukhan595e1702020-07-31 10:12:52 -07006914 name = "qs8_gemm_minmax_test",
6915 srcs = [
6916 "test/qs8-gemm-minmax.cc",
6917 "test/gemm-microkernel-tester.h",
6918 "src/xnnpack/AlignedAllocator.h",
6919 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
6920 deps = MICROKERNEL_TEST_DEPS + [":packing"],
6921)
6922
6923xnnpack_unit_test(
Marat Dukhanf9480682020-07-31 14:50:24 -07006924 name = "qs8_igemm_minmax_test",
6925 srcs = [
6926 "test/qs8-igemm-minmax.cc",
6927 "test/gemm-microkernel-tester.h",
6928 "src/xnnpack/AlignedAllocator.h",
6929 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
6930 deps = MICROKERNEL_TEST_DEPS + [":packing"],
6931)
6932
6933xnnpack_unit_test(
6934 name = "qs8_requantization_test",
6935 srcs = [
6936 "src/xnnpack/requantization-stubs.h",
6937 "test/qs8-requantization.cc",
6938 "test/requantization-tester.h",
6939 ] + MICROKERNEL_TEST_HDRS,
6940 deps = MICROKERNEL_TEST_DEPS,
6941)
6942
6943xnnpack_unit_test(
Marat Dukhand9f3ad42020-08-10 12:30:58 -07006944 name = "qs8_vadd_minmax_test",
6945 srcs = [
6946 "test/qs8-vadd-minmax.cc",
6947 "test/vadd-microkernel-tester.h",
6948 ] + MICROKERNEL_TEST_HDRS,
6949 deps = MICROKERNEL_TEST_DEPS,
6950)
6951
6952xnnpack_unit_test(
Marat Dukhan0270d9f2020-08-11 00:56:46 -07006953 name = "qs8_vaddc_minmax_test",
6954 srcs = [
6955 "test/qs8-vaddc-minmax.cc",
6956 "test/vaddc-microkernel-tester.h",
6957 ] + MICROKERNEL_TEST_HDRS,
6958 deps = MICROKERNEL_TEST_DEPS,
6959)
6960
6961xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07006962 name = "qu8_avgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006963 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07006964 "test/qu8-avgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006965 "test/avgpool-microkernel-tester.h",
6966 "src/xnnpack/AlignedAllocator.h",
6967 ] + MICROKERNEL_TEST_HDRS,
6968 deps = MICROKERNEL_TEST_DEPS,
6969)
6970
6971xnnpack_unit_test(
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07006972 name = "qu8_dwconv_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006973 srcs = [
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07006974 "test/qu8-dwconv-minmax.cc",
6975 "test/dwconv-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006976 "src/xnnpack/AlignedAllocator.h",
6977 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07006978 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07006979)
6980
6981xnnpack_unit_test(
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07006982 name = "qu8_igemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006983 srcs = [
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07006984 "test/qu8-igemm-minmax.cc",
6985 "test/gemm-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006986 "src/xnnpack/AlignedAllocator.h",
6987 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07006988 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07006989)
6990
6991xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07006992 name = "qu8_gavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006993 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07006994 "test/qu8-gavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006995 "test/gavgpool-microkernel-tester.h",
6996 "src/xnnpack/AlignedAllocator.h",
6997 ] + MICROKERNEL_TEST_HDRS,
6998 deps = MICROKERNEL_TEST_DEPS,
6999)
7000
7001xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07007002 name = "qu8_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007003 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07007004 "test/qu8-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007005 "test/gemm-microkernel-tester.h",
7006 "src/xnnpack/AlignedAllocator.h",
7007 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07007008 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07007009)
7010
7011xnnpack_unit_test(
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07007012 name = "qu8_requantization_test",
7013 srcs = [
7014 "src/xnnpack/requantization-stubs.h",
7015 "test/qu8-requantization.cc",
7016 "test/requantization-tester.h",
7017 ] + MICROKERNEL_TEST_HDRS,
7018 deps = MICROKERNEL_TEST_DEPS,
7019)
7020
7021xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07007022 name = "qu8_vadd_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007023 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07007024 "test/qu8-vadd-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007025 "test/vadd-microkernel-tester.h",
7026 ] + MICROKERNEL_TEST_HDRS,
7027 deps = MICROKERNEL_TEST_DEPS,
7028)
7029
7030xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07007031 name = "u8_lut32norm_test",
7032 srcs = [
7033 "test/u8-lut32norm.cc",
7034 "test/lut-norm-microkernel-tester.h",
7035 ] + MICROKERNEL_TEST_HDRS,
7036 deps = MICROKERNEL_TEST_DEPS,
7037)
7038
7039xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07007040 name = "u8_maxpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007041 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07007042 "test/u8-maxpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007043 "test/maxpool-microkernel-tester.h",
7044 ] + MICROKERNEL_TEST_HDRS,
7045 deps = MICROKERNEL_TEST_DEPS,
7046)
7047
7048xnnpack_unit_test(
7049 name = "u8_rmax_test",
7050 srcs = [
7051 "test/u8-rmax.cc",
7052 "test/rmax-microkernel-tester.h",
7053 ] + MICROKERNEL_TEST_HDRS,
7054 deps = MICROKERNEL_TEST_DEPS,
7055)
7056
7057xnnpack_unit_test(
Marat Dukhan6674d692021-05-05 22:27:00 -07007058 name = "u8_vclamp_test",
7059 srcs = [
7060 "test/u8-vclamp.cc",
7061 "test/vclamp-microkernel-tester.h",
7062 ] + MICROKERNEL_TEST_HDRS,
7063 deps = MICROKERNEL_TEST_DEPS,
7064)
7065
7066xnnpack_unit_test(
Marat Dukhanad71b9a2020-11-20 00:01:51 -08007067 name = "x32_depthtospace2d_chw2hwc_test",
Yury Kartynnike7841862020-11-04 18:22:18 -08007068 srcs = [
Marat Dukhanad71b9a2020-11-20 00:01:51 -08007069 "test/x32-depthtospace2d-chw2hwc.cc",
7070 "test/depthtospace-microkernel-tester.h",
Yury Kartynnike7841862020-11-04 18:22:18 -08007071 ] + MICROKERNEL_TEST_HDRS,
7072 deps = MICROKERNEL_TEST_DEPS,
7073)
7074
7075xnnpack_unit_test(
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -07007076 name = "x32_fill_test",
7077 srcs = [
7078 "test/x32-fill.cc",
7079 "test/fill-microkernel-tester.h",
7080 ] + MICROKERNEL_TEST_HDRS,
7081 deps = MICROKERNEL_TEST_DEPS,
7082)
7083
7084xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07007085 name = "x32_packx_test",
7086 srcs = [
7087 "test/x32-packx.cc",
7088 "test/pack-microkernel-tester.h",
7089 "src/xnnpack/AlignedAllocator.h",
7090 ] + MICROKERNEL_TEST_HDRS,
7091 deps = MICROKERNEL_TEST_DEPS,
7092)
7093
7094xnnpack_unit_test(
7095 name = "x32_pad_test",
7096 srcs = [
7097 "test/x32-pad.cc",
7098 "test/pad-microkernel-tester.h",
7099 ] + MICROKERNEL_TEST_HDRS,
7100 deps = MICROKERNEL_TEST_DEPS,
7101)
7102
7103xnnpack_unit_test(
7104 name = "x32_unpool_test",
7105 srcs = [
7106 "test/x32-unpool.cc",
7107 "test/unpool-microkernel-tester.h",
7108 ] + MICROKERNEL_TEST_HDRS,
7109 deps = MICROKERNEL_TEST_DEPS,
7110)
7111
7112xnnpack_unit_test(
7113 name = "x32_zip_test",
7114 srcs = [
7115 "test/x32-zip.cc",
7116 "test/zip-microkernel-tester.h",
7117 ] + MICROKERNEL_TEST_HDRS,
7118 deps = MICROKERNEL_TEST_DEPS,
7119)
7120
7121xnnpack_unit_test(
7122 name = "x8_lut_test",
7123 srcs = [
7124 "test/x8-lut.cc",
7125 "test/lut-microkernel-tester.h",
7126 ] + MICROKERNEL_TEST_HDRS,
7127 deps = MICROKERNEL_TEST_DEPS,
7128)
7129
7130xnnpack_unit_test(
7131 name = "x8_zip_test",
7132 srcs = [
7133 "test/x8-zip.cc",
7134 "test/zip-microkernel-tester.h",
7135 ] + MICROKERNEL_TEST_HDRS,
7136 deps = MICROKERNEL_TEST_DEPS,
7137)
7138
Marat Dukhan20c3b922020-03-10 03:45:06 -07007139########################## Size tests for the library #########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07007140
7141xnnpack_binary(
Marat Dukhan20c3b922020-03-10 03:45:06 -07007142 name = "operator_size_test",
7143 srcs = ["test/operator-size.c"],
Marat Dukhanf0cb70a2021-03-30 15:45:15 -07007144 deps = [":xnnpack_for_tfjs"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07007145)
7146
Marat Dukhan20c3b922020-03-10 03:45:06 -07007147xnnpack_binary(
7148 name = "subgraph_size_test",
7149 srcs = ["test/subgraph-size.c"],
7150 deps = [":XNNPACK"],
7151)
7152
7153########################### Unit tests for operators ##########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07007154
7155xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07007156 name = "abs_nc_test",
7157 srcs = [
7158 "test/abs-nc.cc",
7159 "test/abs-operator-tester.h",
7160 ],
7161 deps = OPERATOR_TEST_DEPS,
7162)
7163
7164xnnpack_unit_test(
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08007165 name = "add_nd_test",
Artsiom Ablavatski2202c812021-01-22 14:16:43 -08007166 timeout = "moderate",
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08007167 srcs = [
7168 "test/add-nd.cc",
7169 "test/binary-elementwise-operator-tester.h",
7170 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07007171 deps = OPERATOR_TEST_DEPS,
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08007172)
7173
7174xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08007175 name = "argmax_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007176 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08007177 "test/argmax-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007178 "test/argmax-pooling-operator-tester.h",
7179 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07007180 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07007181)
7182
7183xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08007184 name = "average_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007185 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08007186 "test/average-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007187 "test/average-pooling-operator-tester.h",
7188 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07007189 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07007190)
7191
7192xnnpack_unit_test(
Marat Dukhan64e52512020-06-09 13:41:16 -07007193 name = "bankers_rounding_nc_test",
7194 srcs = [
7195 "test/bankers-rounding-nc.cc",
7196 "test/bankers-rounding-operator-tester.h",
7197 ],
7198 deps = OPERATOR_TEST_DEPS,
7199)
7200
7201xnnpack_unit_test(
7202 name = "ceiling_nc_test",
7203 srcs = [
7204 "test/ceiling-nc.cc",
7205 "test/ceiling-operator-tester.h",
7206 ],
7207 deps = OPERATOR_TEST_DEPS,
7208)
7209
7210xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08007211 name = "channel_shuffle_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007212 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08007213 "test/channel-shuffle-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007214 "test/channel-shuffle-operator-tester.h",
7215 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07007216 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07007217)
7218
7219xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08007220 name = "clamp_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007221 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08007222 "test/clamp-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007223 "test/clamp-operator-tester.h",
7224 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07007225 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07007226)
7227
7228xnnpack_unit_test(
Marat Dukhan065b11e2020-05-22 09:49:41 -07007229 name = "constant_pad_nd_test",
7230 srcs = [
7231 "test/constant-pad-nd.cc",
7232 "test/constant-pad-operator-tester.h",
7233 ],
7234 deps = OPERATOR_TEST_DEPS,
7235)
7236
7237xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08007238 name = "convolution_nhwc_test",
Artsiom Ablavatski2202c812021-01-22 14:16:43 -08007239 timeout = "moderate",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007240 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08007241 "test/convolution-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007242 "test/convolution-operator-tester.h",
7243 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07007244 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07007245)
7246
7247xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08007248 name = "convolution_nchw_test",
Artsiom Ablavatski2202c812021-01-22 14:16:43 -08007249 timeout = "moderate",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007250 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08007251 "test/convolution-nchw.cc",
7252 "test/convolution-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007253 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07007254 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07007255)
7256
7257xnnpack_unit_test(
Marat Dukhan4e21b272020-06-04 18:45:01 -07007258 name = "copy_nc_test",
7259 srcs = [
7260 "test/copy-nc.cc",
7261 "test/copy-operator-tester.h",
7262 ],
7263 deps = OPERATOR_TEST_DEPS,
7264)
7265
7266xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08007267 name = "deconvolution_nhwc_test",
Artsiom Ablavatskic1aa2972020-12-08 11:23:34 -08007268 timeout = "moderate",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007269 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08007270 "test/deconvolution-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007271 "test/deconvolution-operator-tester.h",
7272 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07007273 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07007274)
7275
7276xnnpack_unit_test(
Marat Dukhan188d1042020-11-24 23:39:40 -08007277 name = "depth_to_space_nchw2nhwc_test",
Artsiom Ablavatski0f1dc182020-11-05 19:21:50 -08007278 srcs = [
Marat Dukhan188d1042020-11-24 23:39:40 -08007279 "test/depth-to-space-nchw2nhwc.cc",
Artsiom Ablavatski0f1dc182020-11-05 19:21:50 -08007280 "test/depth-to-space-operator-tester.h",
7281 ] + OPERATOR_TEST_PARAMS_HDRS,
7282 deps = OPERATOR_TEST_DEPS,
7283)
7284
7285xnnpack_unit_test(
Marat Dukhan0e521172020-11-25 13:10:04 -08007286 name = "depth_to_space_nhwc_test",
7287 srcs = [
7288 "test/depth-to-space-nhwc.cc",
7289 "test/depth-to-space-operator-tester.h",
7290 ] + OPERATOR_TEST_PARAMS_HDRS,
7291 deps = OPERATOR_TEST_DEPS,
7292)
7293
7294xnnpack_unit_test(
Marat Dukhan69180502019-12-06 15:00:31 -08007295 name = "divide_nd_test",
7296 srcs = [
7297 "test/binary-elementwise-operator-tester.h",
7298 "test/divide-nd.cc",
7299 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07007300 deps = OPERATOR_TEST_DEPS,
Marat Dukhan69180502019-12-06 15:00:31 -08007301)
7302
7303xnnpack_unit_test(
Marat Dukhanb6bd4bc2020-12-01 17:01:40 -08007304 name = "elu_nc_test",
7305 srcs = [
7306 "test/elu-nc.cc",
7307 "test/elu-operator-tester.h",
7308 ],
7309 deps = OPERATOR_TEST_DEPS,
7310)
7311
7312xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08007313 name = "fully_connected_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007314 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08007315 "test/fully-connected-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007316 "test/fully-connected-operator-tester.h",
7317 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07007318 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07007319)
7320
7321xnnpack_unit_test(
Marat Dukhan64e52512020-06-09 13:41:16 -07007322 name = "floor_nc_test",
7323 srcs = [
7324 "test/floor-nc.cc",
7325 "test/floor-operator-tester.h",
7326 ],
7327 deps = OPERATOR_TEST_DEPS,
7328)
7329
7330xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08007331 name = "global_average_pooling_nwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007332 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08007333 "test/global-average-pooling-nwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007334 "test/global-average-pooling-operator-tester.h",
Marat Dukhanef61d022020-06-19 13:54:49 -07007335 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07007336 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07007337)
7338
7339xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08007340 name = "global_average_pooling_ncw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007341 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08007342 "test/global-average-pooling-ncw.cc",
7343 "test/global-average-pooling-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007344 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07007345 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07007346)
7347
7348xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08007349 name = "hardswish_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007350 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08007351 "test/hardswish-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007352 "test/hardswish-operator-tester.h",
7353 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07007354 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07007355)
7356
7357xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08007358 name = "leaky_relu_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007359 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08007360 "test/leaky-relu-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007361 "test/leaky-relu-operator-tester.h",
7362 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07007363 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07007364)
7365
7366xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08007367 name = "max_pooling_nhwc_test",
Artsiom Ablavatski2202c812021-01-22 14:16:43 -08007368 timeout = "moderate",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007369 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08007370 "test/max-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007371 "test/max-pooling-operator-tester.h",
7372 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07007373 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07007374)
7375
7376xnnpack_unit_test(
Marat Dukhan79e7f842019-12-05 14:35:50 -08007377 name = "maximum_nd_test",
7378 srcs = [
7379 "test/binary-elementwise-operator-tester.h",
7380 "test/maximum-nd.cc",
7381 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07007382 deps = OPERATOR_TEST_DEPS,
Marat Dukhan79e7f842019-12-05 14:35:50 -08007383)
7384
7385xnnpack_unit_test(
7386 name = "minimum_nd_test",
7387 srcs = [
7388 "test/binary-elementwise-operator-tester.h",
7389 "test/minimum-nd.cc",
7390 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07007391 deps = OPERATOR_TEST_DEPS,
Marat Dukhan79e7f842019-12-05 14:35:50 -08007392)
7393
7394xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08007395 name = "multiply_nd_test",
Marat Dukhanca2733c2019-11-15 23:21:17 -08007396 srcs = [
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08007397 "test/binary-elementwise-operator-tester.h",
Marat Dukhanefc47b82019-11-18 09:25:38 -08007398 "test/multiply-nd.cc",
Marat Dukhanca2733c2019-11-15 23:21:17 -08007399 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07007400 deps = OPERATOR_TEST_DEPS,
Marat Dukhanca2733c2019-11-15 23:21:17 -08007401)
7402
7403xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07007404 name = "negate_nc_test",
7405 srcs = [
7406 "test/negate-nc.cc",
7407 "test/negate-operator-tester.h",
7408 ],
7409 deps = OPERATOR_TEST_DEPS,
7410)
7411
7412xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08007413 name = "prelu_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007414 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08007415 "test/prelu-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007416 "test/prelu-operator-tester.h",
7417 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07007418 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07007419)
7420
7421xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08007422 name = "resize_bilinear_nhwc_test",
Marat Dukhan69722492019-11-11 19:55:50 -08007423 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08007424 "test/resize-bilinear-nhwc.cc",
Marat Dukhan69722492019-11-11 19:55:50 -08007425 "test/resize-bilinear-operator-tester.h",
7426 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07007427 deps = OPERATOR_TEST_DEPS,
Marat Dukhan69722492019-11-11 19:55:50 -08007428)
7429
7430xnnpack_unit_test(
Artsiom Ablavatski97918102020-10-27 15:52:59 -07007431 name = "resize_bilinear_nchw_test",
7432 srcs = [
7433 "test/resize-bilinear-nchw.cc",
7434 "test/resize-bilinear-operator-tester.h",
7435 ] + OPERATOR_TEST_PARAMS_HDRS,
7436 deps = OPERATOR_TEST_DEPS,
7437)
7438
7439xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08007440 name = "sigmoid_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007441 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08007442 "test/sigmoid-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007443 "test/sigmoid-operator-tester.h",
7444 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07007445 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07007446)
7447
7448xnnpack_unit_test(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08007449 name = "softmax_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007450 srcs = [
Marat Dukhanfd8e6892020-01-27 15:25:25 -08007451 "test/softmax-nc.cc",
7452 "test/softmax-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007453 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07007454 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07007455)
7456
7457xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07007458 name = "square_nc_test",
7459 srcs = [
7460 "test/square-nc.cc",
7461 "test/square-operator-tester.h",
7462 ],
7463 deps = OPERATOR_TEST_DEPS,
7464)
7465
7466xnnpack_unit_test(
Marat Dukhan6804bbd2020-06-30 19:26:11 -07007467 name = "square_root_nc_test",
7468 srcs = [
7469 "test/square-root-nc.cc",
7470 "test/square-root-operator-tester.h",
7471 ],
7472 deps = OPERATOR_TEST_DEPS,
7473)
7474
7475xnnpack_unit_test(
Marat Dukhanf7399262020-06-05 10:58:44 -07007476 name = "squared_difference_nd_test",
7477 srcs = [
7478 "test/binary-elementwise-operator-tester.h",
7479 "test/squared-difference-nd.cc",
7480 ],
7481 deps = OPERATOR_TEST_DEPS,
7482)
7483
7484xnnpack_unit_test(
Marat Dukhan05f3f6d2019-12-03 15:13:53 -08007485 name = "subtract_nd_test",
7486 srcs = [
7487 "test/binary-elementwise-operator-tester.h",
7488 "test/subtract-nd.cc",
7489 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07007490 deps = OPERATOR_TEST_DEPS,
Marat Dukhan05f3f6d2019-12-03 15:13:53 -08007491)
7492
7493xnnpack_unit_test(
Marat Dukhan64e52512020-06-09 13:41:16 -07007494 name = "truncation_nc_test",
7495 srcs = [
7496 "test/truncation-nc.cc",
7497 "test/truncation-operator-tester.h",
7498 ],
7499 deps = OPERATOR_TEST_DEPS,
7500)
7501
7502xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08007503 name = "unpooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007504 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08007505 "test/unpooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007506 "test/unpooling-operator-tester.h",
7507 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07007508 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07007509)
7510
Chao Mei6ddfc602020-05-13 22:29:36 -07007511############################### Misc unit tests ###############################
7512
7513xnnpack_unit_test(
7514 name = "memory_planner_test",
7515 srcs = [
7516 "test/memory-planner-test.cc",
7517 ],
7518 deps = [
7519 ":XNNPACK",
7520 ":memory_planner",
7521 ],
7522)
7523
XNNPACK Teamab8c4c82020-10-09 08:05:51 -07007524xnnpack_unit_test(
7525 name = "subgraph_nchw_test",
7526 srcs = [
7527 "src/xnnpack/subgraph.h",
7528 "test/subgraph-nchw.cc",
7529 "test/subgraph-tester.h",
7530 ],
7531 deps = [
7532 ":XNNPACK",
7533 ],
7534)
7535
Marat Dukhan08c4a432019-10-03 09:29:21 -07007536############################# Build configurations #############################
7537
Marat Dukhanb8642352019-10-30 15:43:02 -07007538# Enables usage of assembly kernels.
Marat Dukhan08c4a432019-10-03 09:29:21 -07007539config_setting(
Marat Dukhanb8642352019-10-30 15:43:02 -07007540 name = "xnn_enable_assembly_explicit_true",
7541 define_values = {"xnn_enable_assembly": "true"},
7542)
7543
7544# Disables usage of assembly kernels.
7545config_setting(
7546 name = "xnn_enable_assembly_explicit_false",
7547 define_values = {"xnn_enable_assembly": "false"},
7548)
7549
Marat Dukhan9de90e02020-06-18 16:04:12 -07007550# Enables usage of sparse inference.
7551config_setting(
7552 name = "xnn_enable_sparse_explicit_true",
7553 define_values = {"xnn_enable_sparse": "true"},
7554)
7555
7556# Disables usage of sparse inference.
7557config_setting(
7558 name = "xnn_enable_sparse_explicit_false",
7559 define_values = {"xnn_enable_sparse": "false"},
7560)
7561
Marat Dukhan05702cf2020-03-26 15:41:33 -07007562# Disables usage of HMP-aware optimizations.
7563config_setting(
7564 name = "xnn_enable_hmp_explicit_false",
7565 define_values = {"xnn_enable_hmp": "false"},
7566)
7567
Chao Mei6ddfc602020-05-13 22:29:36 -07007568# Enable usage of optimized memory allocation
7569config_setting(
7570 name = "xnn_enable_memopt_explicit_true",
Marat Dukhan03f46212021-03-30 21:29:49 -07007571 define_values = {"xnn_enable_memopt": "true"},
Chao Mei6ddfc602020-05-13 22:29:36 -07007572)
7573
7574# Disable usage of optimized memory allocation
7575config_setting(
7576 name = "xnn_enable_memopt_explicit_false",
Marat Dukhan03f46212021-03-30 21:29:49 -07007577 define_values = {"xnn_enable_memopt": "false"},
Chao Mei6ddfc602020-05-13 22:29:36 -07007578)
7579
Marat Dukhanb939cdb2021-03-30 18:51:51 -07007580# Enable QS8 inference in TFLite-specific version
7581config_setting(
7582 name = "xnn_enable_qs8_explicit_true",
7583 define_values = {"xnn_enable_qs8": "true"},
7584)
7585
7586# Disable QS8 inference in TFLite-specific version
7587config_setting(
7588 name = "xnn_enable_qs8_explicit_false",
7589 define_values = {"xnn_enable_qs8": "false"},
7590)
7591
Marat Dukhanb8642352019-10-30 15:43:02 -07007592# Builds with -c dbg
7593config_setting(
7594 name = "debug_build",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007595 values = {
Marat Dukhanb8642352019-10-30 15:43:02 -07007596 "compilation_mode": "dbg",
7597 },
7598)
7599
7600# Builds with -c opt
7601config_setting(
7602 name = "optimized_build",
7603 values = {
7604 "compilation_mode": "opt",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007605 },
7606)
7607
7608config_setting(
Marat Dukhanb8642352019-10-30 15:43:02 -07007609 name = "linux_k8",
7610 values = {"cpu": "k8"},
7611)
7612
7613config_setting(
Marat Dukhan582094e2020-04-30 17:21:25 -07007614 name = "linux_arm",
7615 values = {"cpu": "arm"},
Marat Dukhan4e45e662019-10-03 15:40:24 -07007616)
7617
7618config_setting(
Marat Dukhanf0bd4de2020-06-15 15:53:19 -07007619 name = "linux_armeabi",
7620 values = {"cpu": "armeabi"},
7621)
7622
7623config_setting(
Terry Heo68eef3f2020-04-13 22:53:52 -07007624 name = "linux_armhf",
7625 values = {"cpu": "armhf"},
7626)
7627
7628config_setting(
Marat Dukhana720e932020-06-10 13:01:11 -07007629 name = "linux_armv7a",
7630 values = {"cpu": "armv7a"},
7631)
7632
7633config_setting(
Marat Dukhan582094e2020-04-30 17:21:25 -07007634 name = "linux_aarch64",
7635 values = {"cpu": "aarch64"},
7636)
7637
7638config_setting(
Marat Dukhan08c4a432019-10-03 09:29:21 -07007639 name = "android",
7640 values = {"crosstool_top": "//external:android/crosstool"},
7641)
7642
7643config_setting(
7644 name = "android_armv7",
7645 values = {
7646 "crosstool_top": "//external:android/crosstool",
7647 "cpu": "armeabi-v7a",
7648 },
7649)
7650
7651config_setting(
7652 name = "android_arm64",
7653 values = {
7654 "crosstool_top": "//external:android/crosstool",
7655 "cpu": "arm64-v8a",
7656 },
7657)
7658
7659config_setting(
7660 name = "android_x86",
7661 values = {
7662 "crosstool_top": "//external:android/crosstool",
7663 "cpu": "x86",
7664 },
7665)
7666
7667config_setting(
7668 name = "android_x86_64",
7669 values = {
7670 "crosstool_top": "//external:android/crosstool",
7671 "cpu": "x86_64",
7672 },
7673)
7674
7675config_setting(
Marat Dukhan10a38082020-04-17 03:58:35 -07007676 name = "windows_x86_64",
7677 values = {"cpu": "x64_windows"},
Marat Dukhan9fe932e2020-04-11 17:14:15 -07007678)
7679
7680config_setting(
Marat Dukhan10a38082020-04-17 03:58:35 -07007681 name = "windows_x86_64_clang",
7682 values = {
7683 "compiler": "clang-cl",
7684 "cpu": "x64_windows",
7685 },
7686)
7687
7688config_setting(
7689 name = "windows_x86_64_mingw",
7690 values = {
7691 "compiler": "mingw-gcc",
7692 "cpu": "x64_windows",
7693 },
7694)
7695
7696config_setting(
7697 name = "windows_x86_64_msys",
7698 values = {
7699 "compiler": "msys-gcc",
7700 "cpu": "x64_windows",
7701 },
Marat Dukhan9fe932e2020-04-11 17:14:15 -07007702)
7703
7704config_setting(
Marat Dukhan885ca242019-10-07 09:17:32 -07007705 name = "macos_x86_64",
7706 values = {
7707 "apple_platform_type": "macos",
7708 "cpu": "darwin",
7709 },
7710)
7711
7712config_setting(
Simon Maurerae33ab82021-03-03 23:38:22 +01007713 name = "macos_arm64",
7714 values = {
7715 "apple_platform_type": "macos",
7716 "cpu": "darwin_arm64",
7717 },
7718)
7719
7720config_setting(
Marat Dukhan08c4a432019-10-03 09:29:21 -07007721 name = "emscripten",
XNNPACK Team3bfbdaf2021-03-29 15:26:23 -07007722 values = {"crosstool_top": "@emsdk//emscripten_toolchain:everything"},
Marat Dukhan08c4a432019-10-03 09:29:21 -07007723)
7724
7725config_setting(
7726 name = "emscripten_wasm",
7727 values = {
XNNPACK Team3bfbdaf2021-03-29 15:26:23 -07007728 "crosstool_top": "@emsdk//emscripten_toolchain:everything",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007729 "cpu": "wasm",
7730 },
7731)
7732
7733config_setting(
7734 name = "emscripten_wasmsimd",
7735 values = {
XNNPACK Team3bfbdaf2021-03-29 15:26:23 -07007736 "crosstool_top": "@emsdk//emscripten_toolchain:everything",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007737 "cpu": "wasm",
Marat Dukhan81c62602020-05-29 13:22:49 -07007738 "copt": "-msimd128",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007739 },
7740)
7741
7742config_setting(
Marat Dukhan1498d1d2020-02-11 20:00:05 -08007743 name = "ios_armv7",
7744 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08007745 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08007746 "cpu": "ios_armv7",
7747 },
7748)
7749
7750config_setting(
7751 name = "ios_arm64",
7752 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08007753 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08007754 "cpu": "ios_arm64",
7755 },
7756)
7757
7758config_setting(
7759 name = "ios_arm64e",
7760 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08007761 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08007762 "cpu": "ios_arm64e",
7763 },
7764)
7765
7766config_setting(
7767 name = "ios_x86",
7768 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08007769 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08007770 "cpu": "ios_i386",
7771 },
7772)
7773
7774config_setting(
7775 name = "ios_x86_64",
7776 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08007777 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08007778 "cpu": "ios_x86_64",
7779 },
7780)
7781
7782config_setting(
7783 name = "watchos_armv7k",
7784 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08007785 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08007786 "cpu": "watchos_armv7k",
7787 },
7788)
7789
7790config_setting(
7791 name = "watchos_arm64_32",
7792 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08007793 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08007794 "cpu": "watchos_arm64_32",
7795 },
7796)
7797
7798config_setting(
7799 name = "watchos_x86",
7800 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08007801 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08007802 "cpu": "watchos_i386",
7803 },
7804)
7805
7806config_setting(
7807 name = "watchos_x86_64",
7808 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08007809 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08007810 "cpu": "watchos_x86_64",
7811 },
7812)
7813
7814config_setting(
7815 name = "tvos_arm64",
7816 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08007817 "apple_platform_type": "tvos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08007818 "cpu": "tvos_arm64",
7819 },
7820)
7821
7822config_setting(
7823 name = "tvos_x86_64",
7824 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08007825 "apple_platform_type": "tvos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08007826 "cpu": "tvos_x86_64",
7827 },
7828)