blob: ffa4e5d4349bbd71c1c5d54040c29c7933a442a3 [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 Dukhan2e23d2b2020-07-29 16:01:37 -0700494 "src/qs8-requantization/fp32-scalar-lrintf.c",
495 "src/qs8-requantization/fp32-scalar-magic.c",
496 "src/qs8-requantization/precise-scalar-signed64.c",
497 "src/qs8-requantization/precise-scalar-unsigned32.c",
498 "src/qs8-requantization/precise-scalar-unsigned64.c",
499 "src/qs8-requantization/q31-scalar.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -0700500 "src/qu8-avgpool/9p8x-minmax-scalar-c1.c",
501 "src/qu8-avgpool/9x-minmax-scalar-c1.c",
502 "src/qu8-dwconv/up1x9-minmax-scalar.c",
503 "src/qu8-gavgpool/7p7x-minmax-scalar-c1.c",
504 "src/qu8-gavgpool/7x-minmax-scalar-c1.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700505 "src/qu8-gemm/2x2-minmax-scalar.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -0700506 "src/qu8-igemm/2x2-minmax-scalar.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -0700507 "src/qu8-requantization/fp32-scalar-lrintf.c",
508 "src/qu8-requantization/fp32-scalar-magic.c",
509 "src/qu8-requantization/precise-scalar-signed64.c",
510 "src/qu8-requantization/precise-scalar-unsigned32.c",
511 "src/qu8-requantization/precise-scalar-unsigned64.c",
512 "src/qu8-requantization/q31-scalar.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -0700513 "src/qu8-vadd/minmax-scalar.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700514 "src/u8-lut32norm/scalar.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700515 "src/u8-maxpool/9p8x-minmax-scalar-c1.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700516 "src/u8-rmax/scalar.c",
Marat Dukhan6674d692021-05-05 22:27:00 -0700517 "src/u8-vclamp/scalar-x4.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700518 "src/x8-lut/scalar.c",
519 "src/x8-zip/x2-scalar.c",
520 "src/x8-zip/x3-scalar.c",
521 "src/x8-zip/x4-scalar.c",
522 "src/x8-zip/xm-scalar.c",
Marat Dukhanad71b9a2020-11-20 00:01:51 -0800523 "src/x32-depthtospace2d-chw2hwc/scalar.c",
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -0700524 "src/x32-fill/scalar-float.c",
525 "src/x32-fill/scalar-int.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700526 "src/x32-packx/x2-scalar.c",
527 "src/x32-packx/x3-scalar.c",
528 "src/x32-packx/x4-scalar.c",
Marat Dukhan63523d42020-05-22 17:07:33 -0700529 "src/x32-pad/scalar-float.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700530 "src/x32-pad/scalar-int.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700531 "src/x32-unpool/scalar.c",
532 "src/x32-zip/x2-scalar.c",
533 "src/x32-zip/x3-scalar.c",
534 "src/x32-zip/x4-scalar.c",
535 "src/x32-zip/xm-scalar.c",
Marat Dukhan048931b2020-11-24 20:53:54 -0800536 "src/xx-copy/memcpy.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700537]
538
Marat Dukhan436ebe62019-12-04 15:10:12 -0800539WASM_UKERNELS = [
Marat Dukhan99936602020-04-11 16:47:01 -0700540 "src/f32-avgpool/9p8x-minmax-wasm-c1.c",
541 "src/f32-avgpool/9x-minmax-wasm-c1.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700542 "src/f32-dwconv/gen/up1x4-minmax-wasm-acc2.c",
543 "src/f32-dwconv/gen/up1x4-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700544 "src/f32-dwconv/gen/up1x4-wasm-acc2.c",
545 "src/f32-dwconv/gen/up1x4-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700546 "src/f32-dwconv/gen/up1x9-minmax-wasm-acc2.c",
547 "src/f32-dwconv/gen/up1x9-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700548 "src/f32-dwconv/gen/up1x9-wasm-acc2.c",
549 "src/f32-dwconv/gen/up1x9-wasm.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700550 "src/f32-dwconv/gen/up1x25-minmax-wasm-acc2.c",
551 "src/f32-dwconv/gen/up1x25-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700552 "src/f32-dwconv/gen/up1x25-wasm-acc2.c",
553 "src/f32-dwconv/gen/up1x25-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700554 "src/f32-dwconv/gen/up2x4-minmax-wasm-acc2.c",
555 "src/f32-dwconv/gen/up2x4-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700556 "src/f32-dwconv/gen/up2x4-wasm-acc2.c",
557 "src/f32-dwconv/gen/up2x4-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700558 "src/f32-dwconv/gen/up2x9-minmax-wasm-acc2.c",
559 "src/f32-dwconv/gen/up2x9-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700560 "src/f32-dwconv/gen/up2x9-wasm-acc2.c",
561 "src/f32-dwconv/gen/up2x9-wasm.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700562 "src/f32-dwconv/gen/up2x25-minmax-wasm-acc2.c",
563 "src/f32-dwconv/gen/up2x25-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700564 "src/f32-dwconv/gen/up2x25-wasm-acc2.c",
565 "src/f32-dwconv/gen/up2x25-wasm.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700566 "src/f32-gavgpool/7p7x-minmax-wasm-c1.c",
567 "src/f32-gavgpool/7x-minmax-wasm-c1.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700568 "src/f32-gemm/gen-inc/1x4inc-minmax-wasm.c",
569 "src/f32-gemm/gen-inc/2x4inc-minmax-wasm.c",
570 "src/f32-gemm/gen-inc/4x4inc-minmax-wasm.c",
571 "src/f32-gemm/gen/1x4-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700572 "src/f32-gemm/gen/1x4-relu-wasm.c",
573 "src/f32-gemm/gen/1x4-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700574 "src/f32-gemm/gen/2x4-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700575 "src/f32-gemm/gen/2x4-relu-wasm.c",
576 "src/f32-gemm/gen/2x4-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700577 "src/f32-gemm/gen/4x2-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700578 "src/f32-gemm/gen/4x2-relu-wasm.c",
579 "src/f32-gemm/gen/4x2-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700580 "src/f32-gemm/gen/4x4-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700581 "src/f32-gemm/gen/4x4-relu-wasm.c",
582 "src/f32-gemm/gen/4x4-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700583 "src/f32-igemm/gen/1x4-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700584 "src/f32-igemm/gen/1x4-relu-wasm.c",
585 "src/f32-igemm/gen/1x4-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700586 "src/f32-igemm/gen/2x4-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700587 "src/f32-igemm/gen/2x4-relu-wasm.c",
588 "src/f32-igemm/gen/2x4-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700589 "src/f32-igemm/gen/4x2-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700590 "src/f32-igemm/gen/4x2-relu-wasm.c",
591 "src/f32-igemm/gen/4x2-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700592 "src/f32-igemm/gen/4x4-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700593 "src/f32-igemm/gen/4x4-relu-wasm.c",
594 "src/f32-igemm/gen/4x4-wasm.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700595 "src/f32-maxpool/9p8x-minmax-wasm-c1.c",
596 "src/f32-pavgpool/9p8x-minmax-wasm-c1.c",
597 "src/f32-pavgpool/9x-minmax-wasm-c1.c",
Marat Dukhan7c1f8082020-06-25 13:26:20 -0700598 "src/f32-prelu/gen/wasm-2x1.c",
599 "src/f32-prelu/gen/wasm-2x4.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700600 "src/f32-vbinary/gen/vadd-minmax-wasm-x1.c",
601 "src/f32-vbinary/gen/vadd-minmax-wasm-x2.c",
602 "src/f32-vbinary/gen/vadd-minmax-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700603 "src/f32-vbinary/gen/vadd-minmax-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700604 "src/f32-vbinary/gen/vadd-relu-wasm-x1.c",
605 "src/f32-vbinary/gen/vadd-relu-wasm-x2.c",
606 "src/f32-vbinary/gen/vadd-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700607 "src/f32-vbinary/gen/vadd-relu-wasm-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700608 "src/f32-vbinary/gen/vaddc-minmax-wasm-x1.c",
609 "src/f32-vbinary/gen/vaddc-minmax-wasm-x2.c",
610 "src/f32-vbinary/gen/vaddc-minmax-wasm-x4.c",
611 "src/f32-vbinary/gen/vaddc-minmax-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700612 "src/f32-vbinary/gen/vaddc-relu-wasm-x1.c",
613 "src/f32-vbinary/gen/vaddc-relu-wasm-x2.c",
614 "src/f32-vbinary/gen/vaddc-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700615 "src/f32-vbinary/gen/vaddc-relu-wasm-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700616 "src/f32-vbinary/gen/vdiv-minmax-wasm-x1.c",
617 "src/f32-vbinary/gen/vdiv-minmax-wasm-x2.c",
618 "src/f32-vbinary/gen/vdiv-minmax-wasm-x4.c",
619 "src/f32-vbinary/gen/vdiv-minmax-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700620 "src/f32-vbinary/gen/vdiv-relu-wasm-x1.c",
621 "src/f32-vbinary/gen/vdiv-relu-wasm-x2.c",
622 "src/f32-vbinary/gen/vdiv-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700623 "src/f32-vbinary/gen/vdiv-relu-wasm-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700624 "src/f32-vbinary/gen/vdivc-minmax-wasm-x1.c",
625 "src/f32-vbinary/gen/vdivc-minmax-wasm-x2.c",
626 "src/f32-vbinary/gen/vdivc-minmax-wasm-x4.c",
627 "src/f32-vbinary/gen/vdivc-minmax-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700628 "src/f32-vbinary/gen/vdivc-relu-wasm-x1.c",
629 "src/f32-vbinary/gen/vdivc-relu-wasm-x2.c",
630 "src/f32-vbinary/gen/vdivc-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700631 "src/f32-vbinary/gen/vdivc-relu-wasm-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800632 "src/f32-vbinary/gen/vmax-wasm-x1.c",
633 "src/f32-vbinary/gen/vmax-wasm-x2.c",
634 "src/f32-vbinary/gen/vmax-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700635 "src/f32-vbinary/gen/vmax-wasm-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800636 "src/f32-vbinary/gen/vmaxc-wasm-x1.c",
637 "src/f32-vbinary/gen/vmaxc-wasm-x2.c",
638 "src/f32-vbinary/gen/vmaxc-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700639 "src/f32-vbinary/gen/vmaxc-wasm-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800640 "src/f32-vbinary/gen/vmin-wasm-x1.c",
641 "src/f32-vbinary/gen/vmin-wasm-x2.c",
642 "src/f32-vbinary/gen/vmin-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700643 "src/f32-vbinary/gen/vmin-wasm-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800644 "src/f32-vbinary/gen/vminc-wasm-x1.c",
645 "src/f32-vbinary/gen/vminc-wasm-x2.c",
646 "src/f32-vbinary/gen/vminc-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700647 "src/f32-vbinary/gen/vminc-wasm-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700648 "src/f32-vbinary/gen/vmul-minmax-wasm-x1.c",
649 "src/f32-vbinary/gen/vmul-minmax-wasm-x2.c",
650 "src/f32-vbinary/gen/vmul-minmax-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700651 "src/f32-vbinary/gen/vmul-minmax-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700652 "src/f32-vbinary/gen/vmul-relu-wasm-x1.c",
653 "src/f32-vbinary/gen/vmul-relu-wasm-x2.c",
654 "src/f32-vbinary/gen/vmul-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700655 "src/f32-vbinary/gen/vmul-relu-wasm-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700656 "src/f32-vbinary/gen/vmulc-minmax-wasm-x1.c",
657 "src/f32-vbinary/gen/vmulc-minmax-wasm-x2.c",
658 "src/f32-vbinary/gen/vmulc-minmax-wasm-x4.c",
659 "src/f32-vbinary/gen/vmulc-minmax-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700660 "src/f32-vbinary/gen/vmulc-relu-wasm-x1.c",
661 "src/f32-vbinary/gen/vmulc-relu-wasm-x2.c",
662 "src/f32-vbinary/gen/vmulc-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700663 "src/f32-vbinary/gen/vmulc-relu-wasm-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700664 "src/f32-vbinary/gen/vrdivc-minmax-wasm-x1.c",
665 "src/f32-vbinary/gen/vrdivc-minmax-wasm-x2.c",
666 "src/f32-vbinary/gen/vrdivc-minmax-wasm-x4.c",
667 "src/f32-vbinary/gen/vrdivc-minmax-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700668 "src/f32-vbinary/gen/vrdivc-relu-wasm-x1.c",
669 "src/f32-vbinary/gen/vrdivc-relu-wasm-x2.c",
670 "src/f32-vbinary/gen/vrdivc-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700671 "src/f32-vbinary/gen/vrdivc-relu-wasm-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700672 "src/f32-vbinary/gen/vrsubc-minmax-wasm-x1.c",
673 "src/f32-vbinary/gen/vrsubc-minmax-wasm-x2.c",
674 "src/f32-vbinary/gen/vrsubc-minmax-wasm-x4.c",
675 "src/f32-vbinary/gen/vrsubc-minmax-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700676 "src/f32-vbinary/gen/vrsubc-relu-wasm-x1.c",
677 "src/f32-vbinary/gen/vrsubc-relu-wasm-x2.c",
678 "src/f32-vbinary/gen/vrsubc-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700679 "src/f32-vbinary/gen/vrsubc-relu-wasm-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700680 "src/f32-vbinary/gen/vsub-minmax-wasm-x1.c",
681 "src/f32-vbinary/gen/vsub-minmax-wasm-x2.c",
682 "src/f32-vbinary/gen/vsub-minmax-wasm-x4.c",
683 "src/f32-vbinary/gen/vsub-minmax-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700684 "src/f32-vbinary/gen/vsub-relu-wasm-x1.c",
685 "src/f32-vbinary/gen/vsub-relu-wasm-x2.c",
686 "src/f32-vbinary/gen/vsub-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700687 "src/f32-vbinary/gen/vsub-relu-wasm-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700688 "src/f32-vbinary/gen/vsubc-minmax-wasm-x1.c",
689 "src/f32-vbinary/gen/vsubc-minmax-wasm-x2.c",
690 "src/f32-vbinary/gen/vsubc-minmax-wasm-x4.c",
691 "src/f32-vbinary/gen/vsubc-minmax-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700692 "src/f32-vbinary/gen/vsubc-relu-wasm-x1.c",
693 "src/f32-vbinary/gen/vsubc-relu-wasm-x2.c",
694 "src/f32-vbinary/gen/vsubc-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700695 "src/f32-vbinary/gen/vsubc-relu-wasm-x8.c",
Marat Dukhanb1eec082021-05-05 23:24:55 -0700696 "src/f32-vclamp/gen/vclamp-wasm-x1.c",
697 "src/f32-vclamp/gen/vclamp-wasm-x2.c",
698 "src/f32-vclamp/gen/vclamp-wasm-x4.c",
Marat Dukhaned6baaf2020-12-01 15:07:08 -0800699 "src/f32-velu/gen/velu-wasm-rr2-lut16-p3-x1.c",
700 "src/f32-velu/gen/velu-wasm-rr2-lut16-p3-x2.c",
701 "src/f32-velu/gen/velu-wasm-rr2-lut16-p3-x3.c",
702 "src/f32-velu/gen/velu-wasm-rr2-lut16-p3-x4.c",
703 "src/f32-velu/gen/velu-wasm-rr2-lut16-p3-x5.c",
704 "src/f32-velu/gen/velu-wasm-rr2-lut16-p3-x6.c",
705 "src/f32-velu/gen/velu-wasm-rr2-p6-x1.c",
706 "src/f32-velu/gen/velu-wasm-rr2-p6-x2.c",
707 "src/f32-velu/gen/velu-wasm-rr2-p6-x3.c",
708 "src/f32-velu/gen/velu-wasm-rr2-p6-x4.c",
709 "src/f32-velu/gen/velu-wasm-rr2-p6-x5.c",
710 "src/f32-velu/gen/velu-wasm-rr2-p6-x6.c",
Marat Dukhan6674d692021-05-05 22:27:00 -0700711 "src/f32-vhswish/gen/vhswish-wasm-x1.c",
712 "src/f32-vhswish/gen/vhswish-wasm-x2.c",
713 "src/f32-vhswish/gen/vhswish-wasm-x4.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -0700714 "src/f32-vlrelu/gen/vlrelu-wasm-x1.c",
715 "src/f32-vlrelu/gen/vlrelu-wasm-x2.c",
716 "src/f32-vlrelu/gen/vlrelu-wasm-x4.c",
Marat Dukhanb1eec082021-05-05 23:24:55 -0700717 "src/f32-vrelu/gen/vrelu-wasm-x1.c",
718 "src/f32-vrelu/gen/vrelu-wasm-x2.c",
719 "src/f32-vrelu/gen/vrelu-wasm-x4.c",
720 "src/f32-vrelu/gen/vrelu-wasm-x8.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700721 "src/f32-vmulcaddc/gen/c1-minmax-wasm-2x.c",
722 "src/f32-vmulcaddc/gen/c2-minmax-wasm-2x.c",
723 "src/f32-vmulcaddc/gen/c4-minmax-wasm-2x.c",
Marat Dukhan436ebe62019-12-04 15:10:12 -0800724]
725
Marat Dukhan290055c2020-06-09 12:24:29 -0700726WASMSIMD_UKERNELS = [
Marat Dukhan40f05522020-07-16 22:33:12 -0700727 "src/f32-argmaxpool/4x-wasmsimd-c4.c",
728 "src/f32-argmaxpool/9p8x-wasmsimd-c4.c",
729 "src/f32-argmaxpool/9x-wasmsimd-c4.c",
Marat Dukhan3b7432d2020-07-16 17:46:32 -0700730 "src/f32-avgpool/9p8x-minmax-wasmsimd-arm-c4.c",
731 "src/f32-avgpool/9p8x-minmax-wasmsimd-x86-c4.c",
732 "src/f32-avgpool/9x-minmax-wasmsimd-arm-c4.c",
733 "src/f32-avgpool/9x-minmax-wasmsimd-x86-c4.c",
Frank Barchard22136062020-11-24 18:44:46 -0800734 "src/f32-conv-hwc2chw/3x3s2p1c3x4-wasmsimd-2x2.c",
Frank Barchard0725b8d2020-12-07 11:07:35 -0800735 "src/f32-dwconv/gen/up4x4-minmax-wasmsimd-arm-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700736 "src/f32-dwconv/gen/up4x4-minmax-wasmsimd-arm.c",
Frank Barchard0725b8d2020-12-07 11:07:35 -0800737 "src/f32-dwconv/gen/up4x4-minmax-wasmsimd-x86-acc2.c",
Marat Dukhanac014d72020-06-16 08:36:47 -0700738 "src/f32-dwconv/gen/up4x4-minmax-wasmsimd-x86.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700739 "src/f32-dwconv/gen/up4x4-wasmsimd.c",
Frank Barchard0725b8d2020-12-07 11:07:35 -0800740 "src/f32-dwconv/gen/up4x9-minmax-wasmsimd-arm-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700741 "src/f32-dwconv/gen/up4x9-minmax-wasmsimd-arm.c",
Frank Barchard0725b8d2020-12-07 11:07:35 -0800742 "src/f32-dwconv/gen/up4x9-minmax-wasmsimd-x86-acc2.c",
Marat Dukhanac014d72020-06-16 08:36:47 -0700743 "src/f32-dwconv/gen/up4x9-minmax-wasmsimd-x86.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700744 "src/f32-dwconv/gen/up4x9-wasmsimd.c",
Frank Barchard0725b8d2020-12-07 11:07:35 -0800745 "src/f32-dwconv/gen/up4x25-minmax-wasmsimd-arm-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700746 "src/f32-dwconv/gen/up4x25-minmax-wasmsimd-arm.c",
Frank Barchard0725b8d2020-12-07 11:07:35 -0800747 "src/f32-dwconv/gen/up4x25-minmax-wasmsimd-x86-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700748 "src/f32-dwconv/gen/up4x25-minmax-wasmsimd-x86.c",
749 "src/f32-dwconv/gen/up4x25-wasmsimd.c",
Frank Barchard0725b8d2020-12-07 11:07:35 -0800750 "src/f32-dwconv/gen/up8x4-minmax-wasmsimd-arm-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700751 "src/f32-dwconv/gen/up8x4-minmax-wasmsimd-arm.c",
Frank Barchard0725b8d2020-12-07 11:07:35 -0800752 "src/f32-dwconv/gen/up8x4-minmax-wasmsimd-x86-acc2.c",
Marat Dukhanac014d72020-06-16 08:36:47 -0700753 "src/f32-dwconv/gen/up8x4-minmax-wasmsimd-x86.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700754 "src/f32-dwconv/gen/up8x4-wasmsimd.c",
Frank Barchard0725b8d2020-12-07 11:07:35 -0800755 "src/f32-dwconv/gen/up8x9-minmax-wasmsimd-arm-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700756 "src/f32-dwconv/gen/up8x9-minmax-wasmsimd-arm.c",
Frank Barchard0725b8d2020-12-07 11:07:35 -0800757 "src/f32-dwconv/gen/up8x9-minmax-wasmsimd-x86-acc2.c",
Marat Dukhanac014d72020-06-16 08:36:47 -0700758 "src/f32-dwconv/gen/up8x9-minmax-wasmsimd-x86.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700759 "src/f32-dwconv/gen/up8x9-wasmsimd.c",
Frank Barchard0725b8d2020-12-07 11:07:35 -0800760 "src/f32-dwconv/gen/up8x25-minmax-wasmsimd-arm-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700761 "src/f32-dwconv/gen/up8x25-minmax-wasmsimd-arm.c",
Frank Barchard0725b8d2020-12-07 11:07:35 -0800762 "src/f32-dwconv/gen/up8x25-minmax-wasmsimd-x86-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700763 "src/f32-dwconv/gen/up8x25-minmax-wasmsimd-x86.c",
764 "src/f32-dwconv/gen/up8x25-wasmsimd.c",
Frank Barchard412e2f42020-12-11 11:40:50 -0800765 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-loadsplat-1x4-acc2.c",
766 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-loadsplat-1x4-acc3.c",
767 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-loadsplat-1x4-acc4.c",
768 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-loadsplat-1x4.c",
769 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-loadsplat-2x4-acc2.c",
770 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-loadsplat-2x4.c",
771 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-loadsplat-3x4.c",
772 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-loadsplat-4x4.c",
773 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-loadsplat-5x4.c",
774 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-loadsplat-6x4.c",
Frank Barchard02bb4292020-12-15 18:25:32 -0800775 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-splat-1x4-acc2.c",
776 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-splat-1x4-acc3.c",
777 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-splat-1x4-acc4.c",
778 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-splat-1x4.c",
779 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-splat-2x4-acc2.c",
780 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-splat-2x4.c",
781 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-splat-3x4.c",
782 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-splat-4x4.c",
783 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-splat-5x4.c",
784 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-splat-6x4.c",
Frank Barchard412e2f42020-12-11 11:40:50 -0800785 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-loadsplat-1x4-acc2.c",
786 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-loadsplat-1x4-acc3.c",
787 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-loadsplat-1x4-acc4.c",
788 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-loadsplat-1x4.c",
789 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-loadsplat-2x4-acc2.c",
790 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-loadsplat-2x4.c",
791 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-loadsplat-3x4.c",
792 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-loadsplat-4x4.c",
793 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-loadsplat-5x4.c",
794 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-loadsplat-6x4.c",
Frank Barchard02bb4292020-12-15 18:25:32 -0800795 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-splat-1x4-acc2.c",
796 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-splat-1x4-acc3.c",
797 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-splat-1x4-acc4.c",
798 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-splat-1x4.c",
799 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-splat-2x4-acc2.c",
800 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-splat-2x4.c",
801 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-splat-3x4.c",
802 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-splat-4x4.c",
803 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-splat-5x4.c",
804 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-splat-6x4.c",
Frank Barchardc5704bf2020-12-21 23:09:00 -0800805 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-arm-loadsplat-1x4-acc2.c",
806 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-arm-loadsplat-1x4-acc3.c",
807 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-arm-loadsplat-1x4-acc4.c",
808 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-arm-loadsplat-1x4.c",
809 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-arm-loadsplat-2x4-acc2.c",
810 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-arm-loadsplat-2x4.c",
811 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-arm-loadsplat-3x4.c",
812 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-arm-loadsplat-4x4.c",
Frank Barchard412e2f42020-12-11 11:40:50 -0800813 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-arm-splat-1x4-acc2.c",
814 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-arm-splat-1x4-acc3.c",
815 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-arm-splat-1x4-acc4.c",
816 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-arm-splat-1x4.c",
817 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-arm-splat-2x4-acc2.c",
818 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-arm-splat-2x4.c",
819 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-arm-splat-3x4.c",
820 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-arm-splat-4x4.c",
Frank Barchardcadd4222021-01-20 16:27:25 -0800821 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-x86-loadsplat-1x4-acc2.c",
822 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-x86-loadsplat-1x4-acc3.c",
823 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-x86-loadsplat-1x4-acc4.c",
824 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-x86-loadsplat-1x4.c",
825 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-x86-loadsplat-2x4-acc2.c",
826 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-x86-loadsplat-2x4.c",
827 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-x86-loadsplat-3x4.c",
828 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-x86-loadsplat-4x4.c",
Frank Barchard412e2f42020-12-11 11:40:50 -0800829 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-x86-splat-1x4-acc2.c",
830 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-x86-splat-1x4-acc3.c",
831 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-x86-splat-1x4-acc4.c",
832 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-x86-splat-1x4.c",
833 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-x86-splat-2x4-acc2.c",
834 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-x86-splat-2x4.c",
835 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-x86-splat-3x4.c",
836 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-x86-splat-4x4.c",
Frank Barchardb20dcd62020-12-15 16:46:14 -0800837 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-loadsplat-1x4-acc2.c",
838 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-loadsplat-1x4-acc3.c",
839 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-loadsplat-1x4-acc4.c",
840 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-loadsplat-1x4-acc5.c",
841 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-loadsplat-1x4.c",
842 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-loadsplat-2x4-acc2.c",
843 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-loadsplat-2x4-acc3.c",
844 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-loadsplat-2x4.c",
845 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-loadsplat-3x4-acc2.c",
846 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-loadsplat-3x4.c",
847 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-loadsplat-4x4-acc2.c",
848 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-loadsplat-4x4.c",
849 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-loadsplat-5x4.c",
Frank Barchard412e2f42020-12-11 11:40:50 -0800850 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-splat-1x4-acc2.c",
851 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-splat-1x4-acc3.c",
852 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-splat-1x4-acc4.c",
853 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-splat-1x4-acc5.c",
854 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-splat-1x4.c",
855 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-splat-2x4-acc2.c",
856 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-splat-2x4-acc3.c",
857 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-splat-2x4.c",
858 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-splat-3x4-acc2.c",
859 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-splat-3x4.c",
860 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-splat-4x4-acc2.c",
861 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-splat-4x4.c",
862 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-splat-5x4.c",
Frank Barchardb20dcd62020-12-15 16:46:14 -0800863 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-loadsplat-1x4-acc2.c",
864 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-loadsplat-1x4-acc3.c",
865 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-loadsplat-1x4-acc4.c",
866 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-loadsplat-1x4-acc5.c",
867 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-loadsplat-1x4.c",
868 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-loadsplat-2x4-acc2.c",
869 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-loadsplat-2x4-acc3.c",
870 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-loadsplat-2x4.c",
871 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-loadsplat-3x4-acc2.c",
872 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-loadsplat-3x4.c",
873 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-loadsplat-4x4-acc2.c",
874 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-loadsplat-4x4.c",
875 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-loadsplat-5x4.c",
Frank Barchard412e2f42020-12-11 11:40:50 -0800876 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-splat-1x4-acc2.c",
877 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-splat-1x4-acc3.c",
878 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-splat-1x4-acc4.c",
879 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-splat-1x4-acc5.c",
880 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-splat-1x4.c",
881 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-splat-2x4-acc2.c",
882 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-splat-2x4-acc3.c",
883 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-splat-2x4.c",
884 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-splat-3x4-acc2.c",
885 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-splat-3x4.c",
886 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-splat-4x4-acc2.c",
887 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-splat-4x4.c",
888 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-splat-5x4.c",
Frank Barchardc6889b32020-12-21 11:27:22 -0800889 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-arm-loadsplat-1x4-acc2.c",
890 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-arm-loadsplat-1x4-acc3.c",
891 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-arm-loadsplat-1x4-acc4.c",
892 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-arm-loadsplat-1x4-acc5.c",
893 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-arm-loadsplat-1x4.c",
894 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-arm-loadsplat-2x4-acc2.c",
895 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-arm-loadsplat-2x4-acc3.c",
896 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-arm-loadsplat-2x4.c",
897 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-arm-loadsplat-3x4-acc2.c",
898 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-arm-loadsplat-3x4.c",
Frank Barchard412e2f42020-12-11 11:40:50 -0800899 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-arm-splat-1x4-acc2.c",
900 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-arm-splat-1x4-acc3.c",
901 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-arm-splat-1x4-acc4.c",
902 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-arm-splat-1x4-acc5.c",
903 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-arm-splat-1x4.c",
904 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-arm-splat-2x4-acc2.c",
905 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-arm-splat-2x4-acc3.c",
906 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-arm-splat-2x4.c",
907 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-arm-splat-3x4-acc2.c",
908 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-arm-splat-3x4.c",
Frank Barchardc6889b32020-12-21 11:27:22 -0800909 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-x86-loadsplat-1x4-acc2.c",
910 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-x86-loadsplat-1x4-acc3.c",
911 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-x86-loadsplat-1x4-acc4.c",
912 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-x86-loadsplat-1x4-acc5.c",
913 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-x86-loadsplat-1x4.c",
914 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-x86-loadsplat-2x4-acc2.c",
915 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-x86-loadsplat-2x4-acc3.c",
916 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-x86-loadsplat-2x4.c",
917 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-x86-loadsplat-3x4-acc2.c",
918 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-x86-loadsplat-3x4.c",
Frank Barchard412e2f42020-12-11 11:40:50 -0800919 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-x86-splat-1x4-acc2.c",
920 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-x86-splat-1x4-acc3.c",
921 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-x86-splat-1x4-acc4.c",
922 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-x86-splat-1x4-acc5.c",
923 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-x86-splat-1x4.c",
924 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-x86-splat-2x4-acc2.c",
925 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-x86-splat-2x4-acc3.c",
926 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-x86-splat-2x4.c",
927 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-x86-splat-3x4-acc2.c",
928 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-x86-splat-3x4.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700929 "src/f32-gavgpool-cw/wasmsimd-arm-x4.c",
930 "src/f32-gavgpool-cw/wasmsimd-x86-x4.c",
Marat Dukhanc6016802020-07-16 18:51:28 -0700931 "src/f32-gavgpool/7p7x-minmax-wasmsimd-arm-c4.c",
932 "src/f32-gavgpool/7p7x-minmax-wasmsimd-x86-c4.c",
933 "src/f32-gavgpool/7x-minmax-wasmsimd-arm-c4.c",
934 "src/f32-gavgpool/7x-minmax-wasmsimd-x86-c4.c",
Frank Barchard0725b8d2020-12-07 11:07:35 -0800935 "src/f32-gemm/gen-inc/1x8inc-minmax-wasmsimd-arm-loadsplat.c",
936 "src/f32-gemm/gen-inc/1x8inc-minmax-wasmsimd-arm-splat.c",
937 "src/f32-gemm/gen-inc/1x8inc-minmax-wasmsimd-x86-loadsplat.c",
938 "src/f32-gemm/gen-inc/1x8inc-minmax-wasmsimd-x86-splat.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700939 "src/f32-gemm/gen-inc/1x8s4inc-minmax-wasmsimd-arm.c",
940 "src/f32-gemm/gen-inc/1x8s4inc-minmax-wasmsimd-x86.c",
Frank Barchard0725b8d2020-12-07 11:07:35 -0800941 "src/f32-gemm/gen-inc/3x8inc-minmax-wasmsimd-arm-loadsplat.c",
942 "src/f32-gemm/gen-inc/3x8inc-minmax-wasmsimd-arm-splat.c",
943 "src/f32-gemm/gen-inc/3x8inc-minmax-wasmsimd-x86-loadsplat.c",
944 "src/f32-gemm/gen-inc/3x8inc-minmax-wasmsimd-x86-splat.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700945 "src/f32-gemm/gen-inc/3x8s4inc-minmax-wasmsimd-arm.c",
946 "src/f32-gemm/gen-inc/3x8s4inc-minmax-wasmsimd-x86.c",
Frank Barchard0725b8d2020-12-07 11:07:35 -0800947 "src/f32-gemm/gen-inc/4x8inc-minmax-wasmsimd-arm-loadsplat.c",
948 "src/f32-gemm/gen-inc/4x8inc-minmax-wasmsimd-arm-splat.c",
949 "src/f32-gemm/gen-inc/4x8inc-minmax-wasmsimd-x86-loadsplat.c",
950 "src/f32-gemm/gen-inc/4x8inc-minmax-wasmsimd-x86-splat.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700951 "src/f32-gemm/gen-inc/4x8s4inc-minmax-wasmsimd-arm.c",
952 "src/f32-gemm/gen-inc/4x8s4inc-minmax-wasmsimd-x86.c",
Frank Barchard0725b8d2020-12-07 11:07:35 -0800953 "src/f32-gemm/gen-inc/5x8inc-minmax-wasmsimd-arm-loadsplat.c",
954 "src/f32-gemm/gen-inc/5x8inc-minmax-wasmsimd-arm-splat.c",
955 "src/f32-gemm/gen-inc/5x8inc-minmax-wasmsimd-x86-loadsplat.c",
956 "src/f32-gemm/gen-inc/5x8inc-minmax-wasmsimd-x86-splat.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700957 "src/f32-gemm/gen-inc/5x8s4inc-minmax-wasmsimd-arm.c",
958 "src/f32-gemm/gen-inc/5x8s4inc-minmax-wasmsimd-x86.c",
Frank Barchard0725b8d2020-12-07 11:07:35 -0800959 "src/f32-gemm/gen-inc/6x8inc-minmax-wasmsimd-arm-loadsplat.c",
960 "src/f32-gemm/gen-inc/6x8inc-minmax-wasmsimd-arm-splat.c",
961 "src/f32-gemm/gen-inc/6x8inc-minmax-wasmsimd-x86-loadsplat.c",
962 "src/f32-gemm/gen-inc/6x8inc-minmax-wasmsimd-x86-splat.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700963 "src/f32-gemm/gen-inc/6x8s4inc-minmax-wasmsimd-arm.c",
964 "src/f32-gemm/gen-inc/6x8s4inc-minmax-wasmsimd-x86.c",
Frank Barchard0725b8d2020-12-07 11:07:35 -0800965 "src/f32-gemm/gen/1x8-minmax-wasmsimd-arm-loadsplat.c",
966 "src/f32-gemm/gen/1x8-minmax-wasmsimd-arm-splat.c",
967 "src/f32-gemm/gen/1x8-minmax-wasmsimd-x86-loadsplat.c",
968 "src/f32-gemm/gen/1x8-minmax-wasmsimd-x86-splat.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700969 "src/f32-gemm/gen/1x8-relu-wasmsimd-splat.c",
970 "src/f32-gemm/gen/1x8-wasmsimd-splat.c",
971 "src/f32-gemm/gen/1x8s4-minmax-wasmsimd-arm.c",
972 "src/f32-gemm/gen/1x8s4-minmax-wasmsimd-x86.c",
Frank Barchard0725b8d2020-12-07 11:07:35 -0800973 "src/f32-gemm/gen/3x8-minmax-wasmsimd-arm-loadsplat.c",
974 "src/f32-gemm/gen/3x8-minmax-wasmsimd-arm-splat.c",
975 "src/f32-gemm/gen/3x8-minmax-wasmsimd-x86-loadsplat.c",
976 "src/f32-gemm/gen/3x8-minmax-wasmsimd-x86-splat.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700977 "src/f32-gemm/gen/3x8s4-minmax-wasmsimd-arm.c",
978 "src/f32-gemm/gen/3x8s4-minmax-wasmsimd-x86.c",
979 "src/f32-gemm/gen/4x2c4-minmax-wasmsimd-arm.c",
980 "src/f32-gemm/gen/4x2c4-minmax-wasmsimd-x86.c",
981 "src/f32-gemm/gen/4x2c4-relu-wasmsimd.c",
982 "src/f32-gemm/gen/4x2c4-wasmsimd.c",
Frank Barchard0725b8d2020-12-07 11:07:35 -0800983 "src/f32-gemm/gen/4x8-minmax-wasmsimd-arm-loadsplat.c",
984 "src/f32-gemm/gen/4x8-minmax-wasmsimd-arm-splat.c",
985 "src/f32-gemm/gen/4x8-minmax-wasmsimd-x86-loadsplat.c",
986 "src/f32-gemm/gen/4x8-minmax-wasmsimd-x86-splat.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700987 "src/f32-gemm/gen/4x8-relu-wasmsimd-splat.c",
988 "src/f32-gemm/gen/4x8-wasmsimd-splat.c",
989 "src/f32-gemm/gen/4x8s4-minmax-wasmsimd-arm.c",
990 "src/f32-gemm/gen/4x8s4-minmax-wasmsimd-x86.c",
Frank Barchard0725b8d2020-12-07 11:07:35 -0800991 "src/f32-gemm/gen/5x8-minmax-wasmsimd-arm-loadsplat.c",
992 "src/f32-gemm/gen/5x8-minmax-wasmsimd-arm-splat.c",
993 "src/f32-gemm/gen/5x8-minmax-wasmsimd-x86-loadsplat.c",
994 "src/f32-gemm/gen/5x8-minmax-wasmsimd-x86-splat.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700995 "src/f32-gemm/gen/5x8-relu-wasmsimd-splat.c",
996 "src/f32-gemm/gen/5x8-wasmsimd-splat.c",
997 "src/f32-gemm/gen/5x8s4-minmax-wasmsimd-arm.c",
998 "src/f32-gemm/gen/5x8s4-minmax-wasmsimd-x86.c",
Frank Barchard0725b8d2020-12-07 11:07:35 -0800999 "src/f32-gemm/gen/6x8-minmax-wasmsimd-arm-loadsplat.c",
1000 "src/f32-gemm/gen/6x8-minmax-wasmsimd-arm-splat.c",
1001 "src/f32-gemm/gen/6x8-minmax-wasmsimd-x86-loadsplat.c",
1002 "src/f32-gemm/gen/6x8-minmax-wasmsimd-x86-splat.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001003 "src/f32-gemm/gen/6x8s4-minmax-wasmsimd-arm.c",
1004 "src/f32-gemm/gen/6x8s4-minmax-wasmsimd-x86.c",
XNNPACK Team965272b2020-10-23 21:10:15 -07001005 "src/f32-ibilinear-chw/gen/wasmsimd-p4.c",
1006 "src/f32-ibilinear-chw/gen/wasmsimd-p8.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001007 "src/f32-ibilinear/gen/wasmsimd-c4.c",
1008 "src/f32-ibilinear/gen/wasmsimd-c8.c",
Frank Barchard0725b8d2020-12-07 11:07:35 -08001009 "src/f32-igemm/gen/1x8-minmax-wasmsimd-arm-loadsplat.c",
1010 "src/f32-igemm/gen/1x8-minmax-wasmsimd-arm-splat.c",
1011 "src/f32-igemm/gen/1x8-minmax-wasmsimd-x86-loadsplat.c",
1012 "src/f32-igemm/gen/1x8-minmax-wasmsimd-x86-splat.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001013 "src/f32-igemm/gen/1x8-relu-wasmsimd-splat.c",
1014 "src/f32-igemm/gen/1x8-wasmsimd-splat.c",
1015 "src/f32-igemm/gen/1x8s4-minmax-wasmsimd-arm.c",
1016 "src/f32-igemm/gen/1x8s4-minmax-wasmsimd-x86.c",
Frank Barchard0725b8d2020-12-07 11:07:35 -08001017 "src/f32-igemm/gen/3x8-minmax-wasmsimd-arm-loadsplat.c",
1018 "src/f32-igemm/gen/3x8-minmax-wasmsimd-arm-splat.c",
1019 "src/f32-igemm/gen/3x8-minmax-wasmsimd-x86-loadsplat.c",
1020 "src/f32-igemm/gen/3x8-minmax-wasmsimd-x86-splat.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001021 "src/f32-igemm/gen/3x8s4-minmax-wasmsimd-arm.c",
1022 "src/f32-igemm/gen/3x8s4-minmax-wasmsimd-x86.c",
1023 "src/f32-igemm/gen/4x2c4-minmax-wasmsimd-arm.c",
1024 "src/f32-igemm/gen/4x2c4-minmax-wasmsimd-x86.c",
1025 "src/f32-igemm/gen/4x2c4-relu-wasmsimd.c",
1026 "src/f32-igemm/gen/4x2c4-wasmsimd.c",
Frank Barchard0725b8d2020-12-07 11:07:35 -08001027 "src/f32-igemm/gen/4x8-minmax-wasmsimd-arm-loadsplat.c",
1028 "src/f32-igemm/gen/4x8-minmax-wasmsimd-arm-splat.c",
1029 "src/f32-igemm/gen/4x8-minmax-wasmsimd-x86-loadsplat.c",
1030 "src/f32-igemm/gen/4x8-minmax-wasmsimd-x86-splat.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001031 "src/f32-igemm/gen/4x8-relu-wasmsimd-splat.c",
1032 "src/f32-igemm/gen/4x8-wasmsimd-splat.c",
1033 "src/f32-igemm/gen/4x8s4-minmax-wasmsimd-arm.c",
1034 "src/f32-igemm/gen/4x8s4-minmax-wasmsimd-x86.c",
Frank Barchard0725b8d2020-12-07 11:07:35 -08001035 "src/f32-igemm/gen/5x8-minmax-wasmsimd-arm-loadsplat.c",
1036 "src/f32-igemm/gen/5x8-minmax-wasmsimd-arm-splat.c",
1037 "src/f32-igemm/gen/5x8-minmax-wasmsimd-x86-loadsplat.c",
1038 "src/f32-igemm/gen/5x8-minmax-wasmsimd-x86-splat.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001039 "src/f32-igemm/gen/5x8-relu-wasmsimd-splat.c",
1040 "src/f32-igemm/gen/5x8-wasmsimd-splat.c",
1041 "src/f32-igemm/gen/5x8s4-minmax-wasmsimd-arm.c",
1042 "src/f32-igemm/gen/5x8s4-minmax-wasmsimd-x86.c",
Frank Barchard0725b8d2020-12-07 11:07:35 -08001043 "src/f32-igemm/gen/6x8-minmax-wasmsimd-arm-loadsplat.c",
1044 "src/f32-igemm/gen/6x8-minmax-wasmsimd-arm-splat.c",
1045 "src/f32-igemm/gen/6x8-minmax-wasmsimd-x86-loadsplat.c",
1046 "src/f32-igemm/gen/6x8-minmax-wasmsimd-x86-splat.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001047 "src/f32-igemm/gen/6x8s4-minmax-wasmsimd-arm.c",
1048 "src/f32-igemm/gen/6x8s4-minmax-wasmsimd-x86.c",
Marat Dukhanf6e24802020-07-08 22:20:40 -07001049 "src/f32-maxpool/9p8x-minmax-wasmsimd-arm-c4.c",
1050 "src/f32-maxpool/9p8x-minmax-wasmsimd-x86-c4.c",
Marat Dukhan1483c532020-07-16 18:08:19 -07001051 "src/f32-pavgpool/9p8x-minmax-wasmsimd-arm-c4.c",
1052 "src/f32-pavgpool/9p8x-minmax-wasmsimd-x86-c4.c",
1053 "src/f32-pavgpool/9x-minmax-wasmsimd-arm-c4.c",
1054 "src/f32-pavgpool/9x-minmax-wasmsimd-x86-c4.c",
Frank Barchard0725b8d2020-12-07 11:07:35 -08001055 "src/f32-ppmm/gen/4x8-minmax-wasmsimd-arm-splat.c",
1056 "src/f32-ppmm/gen/4x8-minmax-wasmsimd-x86-splat.c",
Frank Barcharda5316982020-07-23 13:19:28 -07001057 "src/f32-prelu/gen/wasmsimd-bitselect-1x4.c",
1058 "src/f32-prelu/gen/wasmsimd-bitselect-1x8.c",
1059 "src/f32-prelu/gen/wasmsimd-bitselect-1x16.c",
Marat Dukhan195f8eb2020-06-25 12:50:57 -07001060 "src/f32-prelu/gen/wasmsimd-bitselect-2x4.c",
1061 "src/f32-prelu/gen/wasmsimd-bitselect-2x8.c",
Frank Barcharda5316982020-07-23 13:19:28 -07001062 "src/f32-prelu/gen/wasmsimd-bitselect-2x16.c",
1063 "src/f32-prelu/gen/wasmsimd-bitselect-4x4.c",
1064 "src/f32-prelu/gen/wasmsimd-bitselect-4x8.c",
1065 "src/f32-prelu/gen/wasmsimd-bitselect-4x16.c",
1066 "src/f32-prelu/gen/wasmsimd-minmax-1x4.c",
1067 "src/f32-prelu/gen/wasmsimd-minmax-1x8.c",
1068 "src/f32-prelu/gen/wasmsimd-minmax-1x16.c",
Marat Dukhan195f8eb2020-06-25 12:50:57 -07001069 "src/f32-prelu/gen/wasmsimd-minmax-2x4.c",
1070 "src/f32-prelu/gen/wasmsimd-minmax-2x8.c",
Frank Barcharda5316982020-07-23 13:19:28 -07001071 "src/f32-prelu/gen/wasmsimd-minmax-2x16.c",
1072 "src/f32-prelu/gen/wasmsimd-minmax-4x4.c",
1073 "src/f32-prelu/gen/wasmsimd-minmax-4x8.c",
1074 "src/f32-prelu/gen/wasmsimd-minmax-4x16.c",
Marat Dukhan52238f02020-07-16 15:30:28 -07001075 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x4.c",
Marat Dukhan52238f02020-07-16 15:30:28 -07001076 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x8-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001077 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x8.c",
Marat Dukhan52238f02020-07-16 15:30:28 -07001078 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x12-acc2.c",
1079 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x12-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001080 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x12.c",
Marat Dukhan52238f02020-07-16 15:30:28 -07001081 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x16-acc2.c",
1082 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x16-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001083 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x16.c",
Marat Dukhan52238f02020-07-16 15:30:28 -07001084 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x20-acc2.c",
1085 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x20-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001086 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x20.c",
Marat Dukhan8c417962020-07-08 12:27:50 -07001087 "src/f32-rmax/wasmsimd-arm.c",
1088 "src/f32-rmax/wasmsimd-x86.c",
Frank Barchard8ef44cd2020-11-03 12:30:23 -08001089 "src/f32-spmm/gen/4x1-minmax-wasmsimd-arm-pipelined-x2.c",
1090 "src/f32-spmm/gen/4x1-minmax-wasmsimd-arm-pipelined.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07001091 "src/f32-spmm/gen/4x1-minmax-wasmsimd-arm-x2.c",
1092 "src/f32-spmm/gen/4x1-minmax-wasmsimd-arm-x4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001093 "src/f32-spmm/gen/4x1-minmax-wasmsimd-arm.c",
Frank Barchard8ef44cd2020-11-03 12:30:23 -08001094 "src/f32-spmm/gen/4x1-minmax-wasmsimd-x86-pipelined-x2.c",
1095 "src/f32-spmm/gen/4x1-minmax-wasmsimd-x86-pipelined.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07001096 "src/f32-spmm/gen/4x1-minmax-wasmsimd-x86-x2.c",
1097 "src/f32-spmm/gen/4x1-minmax-wasmsimd-x86-x4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001098 "src/f32-spmm/gen/4x1-minmax-wasmsimd-x86.c",
Frank Barchard8ef44cd2020-11-03 12:30:23 -08001099 "src/f32-spmm/gen/8x1-minmax-wasmsimd-arm-pipelined-x2.c",
1100 "src/f32-spmm/gen/8x1-minmax-wasmsimd-arm-pipelined.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07001101 "src/f32-spmm/gen/8x1-minmax-wasmsimd-arm-x2.c",
1102 "src/f32-spmm/gen/8x1-minmax-wasmsimd-arm-x4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001103 "src/f32-spmm/gen/8x1-minmax-wasmsimd-arm.c",
Frank Barchard8ef44cd2020-11-03 12:30:23 -08001104 "src/f32-spmm/gen/8x1-minmax-wasmsimd-x86-pipelined-x2.c",
1105 "src/f32-spmm/gen/8x1-minmax-wasmsimd-x86-pipelined.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07001106 "src/f32-spmm/gen/8x1-minmax-wasmsimd-x86-x2.c",
1107 "src/f32-spmm/gen/8x1-minmax-wasmsimd-x86-x4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001108 "src/f32-spmm/gen/8x1-minmax-wasmsimd-x86.c",
Frank Barchard8ef44cd2020-11-03 12:30:23 -08001109 "src/f32-spmm/gen/16x1-minmax-wasmsimd-arm-pipelined-x2.c",
1110 "src/f32-spmm/gen/16x1-minmax-wasmsimd-arm-pipelined.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07001111 "src/f32-spmm/gen/16x1-minmax-wasmsimd-arm-x2.c",
1112 "src/f32-spmm/gen/16x1-minmax-wasmsimd-arm-x4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001113 "src/f32-spmm/gen/16x1-minmax-wasmsimd-arm.c",
Frank Barchard8ef44cd2020-11-03 12:30:23 -08001114 "src/f32-spmm/gen/16x1-minmax-wasmsimd-x86-pipelined-x2.c",
1115 "src/f32-spmm/gen/16x1-minmax-wasmsimd-x86-pipelined.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07001116 "src/f32-spmm/gen/16x1-minmax-wasmsimd-x86-x2.c",
1117 "src/f32-spmm/gen/16x1-minmax-wasmsimd-x86-x4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001118 "src/f32-spmm/gen/16x1-minmax-wasmsimd-x86.c",
Frank Barchard8ef44cd2020-11-03 12:30:23 -08001119 "src/f32-spmm/gen/32x1-minmax-wasmsimd-arm-pipelined-x2.c",
1120 "src/f32-spmm/gen/32x1-minmax-wasmsimd-arm-pipelined.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07001121 "src/f32-spmm/gen/32x1-minmax-wasmsimd-arm-x2.c",
1122 "src/f32-spmm/gen/32x1-minmax-wasmsimd-arm-x4.c",
Frank Barchard846c0c62020-10-26 15:01:39 -07001123 "src/f32-spmm/gen/32x1-minmax-wasmsimd-arm.c",
Frank Barchard8ef44cd2020-11-03 12:30:23 -08001124 "src/f32-spmm/gen/32x1-minmax-wasmsimd-x86-pipelined-x2.c",
1125 "src/f32-spmm/gen/32x1-minmax-wasmsimd-x86-pipelined.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07001126 "src/f32-spmm/gen/32x1-minmax-wasmsimd-x86-x2.c",
1127 "src/f32-spmm/gen/32x1-minmax-wasmsimd-x86-x4.c",
Frank Barchard846c0c62020-10-26 15:01:39 -07001128 "src/f32-spmm/gen/32x1-minmax-wasmsimd-x86.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -07001129 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-arm-x4.c",
1130 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001131 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-arm-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -07001132 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-x86-x4.c",
1133 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001134 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -07001135 "src/f32-vbinary/gen/vadd-relu-wasmsimd-x4.c",
1136 "src/f32-vbinary/gen/vadd-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001137 "src/f32-vbinary/gen/vadd-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07001138 "src/f32-vbinary/gen/vadd-wasmsimd-x4.c",
1139 "src/f32-vbinary/gen/vadd-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001140 "src/f32-vbinary/gen/vadd-wasmsimd-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001141 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-arm-x4.c",
1142 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001143 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-arm-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001144 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-x86-x4.c",
1145 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001146 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -07001147 "src/f32-vbinary/gen/vaddc-relu-wasmsimd-x4.c",
1148 "src/f32-vbinary/gen/vaddc-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001149 "src/f32-vbinary/gen/vaddc-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07001150 "src/f32-vbinary/gen/vaddc-wasmsimd-x4.c",
1151 "src/f32-vbinary/gen/vaddc-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001152 "src/f32-vbinary/gen/vaddc-wasmsimd-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -07001153 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-arm-x4.c",
1154 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001155 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-arm-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -07001156 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-x86-x4.c",
1157 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001158 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -07001159 "src/f32-vbinary/gen/vdiv-relu-wasmsimd-x4.c",
1160 "src/f32-vbinary/gen/vdiv-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001161 "src/f32-vbinary/gen/vdiv-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07001162 "src/f32-vbinary/gen/vdiv-wasmsimd-x4.c",
1163 "src/f32-vbinary/gen/vdiv-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001164 "src/f32-vbinary/gen/vdiv-wasmsimd-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001165 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-arm-x4.c",
1166 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001167 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-arm-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001168 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-x86-x4.c",
1169 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001170 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -07001171 "src/f32-vbinary/gen/vdivc-relu-wasmsimd-x4.c",
1172 "src/f32-vbinary/gen/vdivc-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001173 "src/f32-vbinary/gen/vdivc-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07001174 "src/f32-vbinary/gen/vdivc-wasmsimd-x4.c",
1175 "src/f32-vbinary/gen/vdivc-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001176 "src/f32-vbinary/gen/vdivc-wasmsimd-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -07001177 "src/f32-vbinary/gen/vmax-wasmsimd-arm-x4.c",
1178 "src/f32-vbinary/gen/vmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001179 "src/f32-vbinary/gen/vmax-wasmsimd-arm-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -07001180 "src/f32-vbinary/gen/vmax-wasmsimd-x86-x4.c",
1181 "src/f32-vbinary/gen/vmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001182 "src/f32-vbinary/gen/vmax-wasmsimd-x86-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001183 "src/f32-vbinary/gen/vmaxc-wasmsimd-arm-x4.c",
1184 "src/f32-vbinary/gen/vmaxc-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001185 "src/f32-vbinary/gen/vmaxc-wasmsimd-arm-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001186 "src/f32-vbinary/gen/vmaxc-wasmsimd-x86-x4.c",
1187 "src/f32-vbinary/gen/vmaxc-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001188 "src/f32-vbinary/gen/vmaxc-wasmsimd-x86-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -07001189 "src/f32-vbinary/gen/vmin-wasmsimd-arm-x4.c",
1190 "src/f32-vbinary/gen/vmin-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001191 "src/f32-vbinary/gen/vmin-wasmsimd-arm-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -07001192 "src/f32-vbinary/gen/vmin-wasmsimd-x86-x4.c",
1193 "src/f32-vbinary/gen/vmin-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001194 "src/f32-vbinary/gen/vmin-wasmsimd-x86-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001195 "src/f32-vbinary/gen/vminc-wasmsimd-arm-x4.c",
1196 "src/f32-vbinary/gen/vminc-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001197 "src/f32-vbinary/gen/vminc-wasmsimd-arm-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001198 "src/f32-vbinary/gen/vminc-wasmsimd-x86-x4.c",
1199 "src/f32-vbinary/gen/vminc-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001200 "src/f32-vbinary/gen/vminc-wasmsimd-x86-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -07001201 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-arm-x4.c",
1202 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001203 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-arm-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -07001204 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-x86-x4.c",
1205 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001206 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -07001207 "src/f32-vbinary/gen/vmul-relu-wasmsimd-x4.c",
1208 "src/f32-vbinary/gen/vmul-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001209 "src/f32-vbinary/gen/vmul-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07001210 "src/f32-vbinary/gen/vmul-wasmsimd-x4.c",
1211 "src/f32-vbinary/gen/vmul-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001212 "src/f32-vbinary/gen/vmul-wasmsimd-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001213 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-arm-x4.c",
1214 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001215 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-arm-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001216 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-x86-x4.c",
1217 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001218 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -07001219 "src/f32-vbinary/gen/vmulc-relu-wasmsimd-x4.c",
1220 "src/f32-vbinary/gen/vmulc-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001221 "src/f32-vbinary/gen/vmulc-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07001222 "src/f32-vbinary/gen/vmulc-wasmsimd-x4.c",
1223 "src/f32-vbinary/gen/vmulc-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001224 "src/f32-vbinary/gen/vmulc-wasmsimd-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001225 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-arm-x4.c",
1226 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001227 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-arm-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001228 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-x86-x4.c",
1229 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001230 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -07001231 "src/f32-vbinary/gen/vrdivc-relu-wasmsimd-x4.c",
1232 "src/f32-vbinary/gen/vrdivc-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001233 "src/f32-vbinary/gen/vrdivc-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07001234 "src/f32-vbinary/gen/vrdivc-wasmsimd-x4.c",
1235 "src/f32-vbinary/gen/vrdivc-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001236 "src/f32-vbinary/gen/vrdivc-wasmsimd-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001237 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-arm-x4.c",
1238 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001239 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-arm-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001240 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-x86-x4.c",
1241 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001242 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -07001243 "src/f32-vbinary/gen/vrsubc-relu-wasmsimd-x4.c",
1244 "src/f32-vbinary/gen/vrsubc-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001245 "src/f32-vbinary/gen/vrsubc-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07001246 "src/f32-vbinary/gen/vrsubc-wasmsimd-x4.c",
1247 "src/f32-vbinary/gen/vrsubc-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001248 "src/f32-vbinary/gen/vrsubc-wasmsimd-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -07001249 "src/f32-vbinary/gen/vsqrdiff-wasmsimd-x4.c",
1250 "src/f32-vbinary/gen/vsqrdiff-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001251 "src/f32-vbinary/gen/vsqrdiff-wasmsimd-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001252 "src/f32-vbinary/gen/vsqrdiffc-wasmsimd-x4.c",
1253 "src/f32-vbinary/gen/vsqrdiffc-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001254 "src/f32-vbinary/gen/vsqrdiffc-wasmsimd-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -07001255 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-arm-x4.c",
1256 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001257 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-arm-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -07001258 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-x86-x4.c",
1259 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001260 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -07001261 "src/f32-vbinary/gen/vsub-relu-wasmsimd-x4.c",
1262 "src/f32-vbinary/gen/vsub-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001263 "src/f32-vbinary/gen/vsub-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07001264 "src/f32-vbinary/gen/vsub-wasmsimd-x4.c",
1265 "src/f32-vbinary/gen/vsub-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001266 "src/f32-vbinary/gen/vsub-wasmsimd-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001267 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-arm-x4.c",
1268 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001269 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-arm-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001270 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-x86-x4.c",
1271 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001272 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -07001273 "src/f32-vbinary/gen/vsubc-relu-wasmsimd-x4.c",
1274 "src/f32-vbinary/gen/vsubc-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001275 "src/f32-vbinary/gen/vsubc-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07001276 "src/f32-vbinary/gen/vsubc-wasmsimd-x4.c",
1277 "src/f32-vbinary/gen/vsubc-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001278 "src/f32-vbinary/gen/vsubc-wasmsimd-x16.c",
Marat Dukhanb1eec082021-05-05 23:24:55 -07001279 "src/f32-vclamp/gen/vclamp-wasmsimd-arm-x4.c",
1280 "src/f32-vclamp/gen/vclamp-wasmsimd-arm-x8.c",
1281 "src/f32-vclamp/gen/vclamp-wasmsimd-x86-x4.c",
1282 "src/f32-vclamp/gen/vclamp-wasmsimd-x86-x8.c",
Marat Dukhaned6baaf2020-12-01 15:07:08 -08001283 "src/f32-velu/gen/velu-wasmsimd-arm-rr2-lut16-p3-x4.c",
1284 "src/f32-velu/gen/velu-wasmsimd-arm-rr2-lut16-p3-x8.c",
1285 "src/f32-velu/gen/velu-wasmsimd-arm-rr2-lut16-p3-x12.c",
1286 "src/f32-velu/gen/velu-wasmsimd-arm-rr2-lut16-p3-x16.c",
1287 "src/f32-velu/gen/velu-wasmsimd-arm-rr2-lut16-p3-x20.c",
1288 "src/f32-velu/gen/velu-wasmsimd-arm-rr2-lut16-p3-x24.c",
Marat Dukhaned6baaf2020-12-01 15:07:08 -08001289 "src/f32-velu/gen/velu-wasmsimd-arm-rr2-p6-x4.c",
1290 "src/f32-velu/gen/velu-wasmsimd-arm-rr2-p6-x8.c",
1291 "src/f32-velu/gen/velu-wasmsimd-arm-rr2-p6-x12.c",
1292 "src/f32-velu/gen/velu-wasmsimd-arm-rr2-p6-x16.c",
1293 "src/f32-velu/gen/velu-wasmsimd-arm-rr2-p6-x20.c",
1294 "src/f32-velu/gen/velu-wasmsimd-arm-rr2-p6-x24.c",
Frank Barcharde7223ee2020-12-04 19:04:01 -08001295 "src/f32-velu/gen/velu-wasmsimd-x86-rr2-lut16-p3-x4.c",
1296 "src/f32-velu/gen/velu-wasmsimd-x86-rr2-lut16-p3-x8.c",
1297 "src/f32-velu/gen/velu-wasmsimd-x86-rr2-lut16-p3-x12.c",
1298 "src/f32-velu/gen/velu-wasmsimd-x86-rr2-lut16-p3-x16.c",
1299 "src/f32-velu/gen/velu-wasmsimd-x86-rr2-lut16-p3-x20.c",
1300 "src/f32-velu/gen/velu-wasmsimd-x86-rr2-lut16-p3-x24.c",
Marat Dukhaned6baaf2020-12-01 15:07:08 -08001301 "src/f32-velu/gen/velu-wasmsimd-x86-rr2-p6-x4.c",
1302 "src/f32-velu/gen/velu-wasmsimd-x86-rr2-p6-x8.c",
1303 "src/f32-velu/gen/velu-wasmsimd-x86-rr2-p6-x12.c",
1304 "src/f32-velu/gen/velu-wasmsimd-x86-rr2-p6-x16.c",
1305 "src/f32-velu/gen/velu-wasmsimd-x86-rr2-p6-x20.c",
1306 "src/f32-velu/gen/velu-wasmsimd-x86-rr2-p6-x24.c",
Marat Dukhan6674d692021-05-05 22:27:00 -07001307 "src/f32-vhswish/gen/vhswish-wasmsimd-x4.c",
1308 "src/f32-vhswish/gen/vhswish-wasmsimd-x8.c",
1309 "src/f32-vhswish/gen/vhswish-wasmsimd-x16.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07001310 "src/f32-vlrelu/gen/vlrelu-wasmsimd-bitselect-x4.c",
1311 "src/f32-vlrelu/gen/vlrelu-wasmsimd-bitselect-x8.c",
1312 "src/f32-vlrelu/gen/vlrelu-wasmsimd-minmax-x4.c",
1313 "src/f32-vlrelu/gen/vlrelu-wasmsimd-minmax-x8.c",
Marat Dukhand816f622020-07-15 10:14:39 -07001314 "src/f32-vmulcaddc/gen/c4-minmax-wasmsimd-arm-2x.c",
Marat Dukhand816f622020-07-15 10:14:39 -07001315 "src/f32-vmulcaddc/gen/c4-minmax-wasmsimd-x86-2x.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001316 "src/f32-vmulcaddc/gen/c8-minmax-wasmsimd-arm-2x.c",
Marat Dukhand816f622020-07-15 10:14:39 -07001317 "src/f32-vmulcaddc/gen/c8-minmax-wasmsimd-x86-2x.c",
Marat Dukhanb1eec082021-05-05 23:24:55 -07001318 "src/f32-vrelu/gen/vrelu-wasmsimd-x4.c",
1319 "src/f32-vrelu/gen/vrelu-wasmsimd-x8.c",
1320 "src/f32-vrelu/gen/vrelu-wasmsimd-x16.c",
Marat Dukhanb82b2cd2020-07-16 02:23:42 -07001321 "src/f32-vrnd/gen/vrndd-wasmsimd-addsub-x4.c",
1322 "src/f32-vrnd/gen/vrndd-wasmsimd-addsub-x8.c",
1323 "src/f32-vrnd/gen/vrndd-wasmsimd-cvt-x4.c",
1324 "src/f32-vrnd/gen/vrndd-wasmsimd-cvt-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001325 "src/f32-vrnd/gen/vrndne-wasmsimd-addsub-x4.c",
1326 "src/f32-vrnd/gen/vrndne-wasmsimd-addsub-x8.c",
1327 "src/f32-vrnd/gen/vrndu-wasmsimd-addsub-x4.c",
1328 "src/f32-vrnd/gen/vrndu-wasmsimd-addsub-x8.c",
1329 "src/f32-vrnd/gen/vrndu-wasmsimd-cvt-x4.c",
1330 "src/f32-vrnd/gen/vrndu-wasmsimd-cvt-x8.c",
1331 "src/f32-vrnd/gen/vrndz-wasmsimd-addsub-x4.c",
1332 "src/f32-vrnd/gen/vrndz-wasmsimd-addsub-x8.c",
1333 "src/f32-vrnd/gen/vrndz-wasmsimd-cvt-x4.c",
1334 "src/f32-vrnd/gen/vrndz-wasmsimd-cvt-x8.c",
Marat Dukhanb1eec082021-05-05 23:24:55 -07001335 "src/f32-vsigmoid/gen/vsigmoid-wasmsimd-lut64-p2-div-x4.c",
1336 "src/f32-vsigmoid/gen/vsigmoid-wasmsimd-lut64-p2-div-x8.c",
1337 "src/f32-vsigmoid/gen/vsigmoid-wasmsimd-lut64-p2-div-x12.c",
1338 "src/f32-vsigmoid/gen/vsigmoid-wasmsimd-lut64-p2-div-x16.c",
1339 "src/f32-vsigmoid/gen/vsigmoid-wasmsimd-lut64-p2-div-x20.c",
1340 "src/f32-vsigmoid/gen/vsigmoid-wasmsimd-lut64-p2-div-x24.c",
1341 "src/f32-vsigmoid/gen/vsigmoid-wasmsimd-p5-div-x4.c",
1342 "src/f32-vsigmoid/gen/vsigmoid-wasmsimd-p5-div-x8.c",
1343 "src/f32-vsigmoid/gen/vsigmoid-wasmsimd-p5-div-x12.c",
1344 "src/f32-vsigmoid/gen/vsigmoid-wasmsimd-p5-div-x16.c",
1345 "src/f32-vsigmoid/gen/vsigmoid-wasmsimd-p5-div-x20.c",
1346 "src/f32-vsigmoid/gen/vsigmoid-wasmsimd-p5-div-x24.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07001347 "src/f32-vsqrt/gen/wasmsimd-sqrt-x4.c",
1348 "src/f32-vsqrt/gen/wasmsimd-sqrt-x8.c",
Marat Dukhan37c83512020-06-29 13:25:53 -07001349 "src/f32-vunary/gen/vabs-wasmsimd-x4.c",
1350 "src/f32-vunary/gen/vabs-wasmsimd-x8.c",
1351 "src/f32-vunary/gen/vneg-wasmsimd-x4.c",
1352 "src/f32-vunary/gen/vneg-wasmsimd-x8.c",
1353 "src/f32-vunary/gen/vsqr-wasmsimd-x4.c",
1354 "src/f32-vunary/gen/vsqr-wasmsimd-x8.c",
Marat Dukhande390d42020-11-29 19:32:18 -08001355 "src/math/expm1minus-wasmsimd-rr2-lut16-p3-andnot.c",
1356 "src/math/expm1minus-wasmsimd-rr2-lut16-p3-max.c",
1357 "src/math/expm1minus-wasmsimd-rr2-p6-andnot.c",
1358 "src/math/expm1minus-wasmsimd-rr2-p6-max.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001359 "src/math/roundd-wasmsimd-addsub.c",
1360 "src/math/roundd-wasmsimd-cvt.c",
1361 "src/math/roundne-wasmsimd-addsub.c",
1362 "src/math/roundu-wasmsimd-addsub.c",
1363 "src/math/roundu-wasmsimd-cvt.c",
1364 "src/math/roundz-wasmsimd-addsub.c",
1365 "src/math/roundz-wasmsimd-cvt.c",
1366 "src/math/sigmoid-wasmsimd-rr2-lut64-p2-div.c",
1367 "src/math/sigmoid-wasmsimd-rr2-p5-div.c",
Marat Dukhancc8f34c2020-08-05 16:36:38 -07001368 "src/qs8-dwconv/gen/up8x9-minmax-wasmsimd-mul16.c",
1369 "src/qs8-dwconv/gen/up16x9-minmax-wasmsimd-mul16.c",
1370 "src/qs8-dwconv/gen/up24x9-minmax-wasmsimd-mul16.c",
Marat Dukhanb5e3d172020-08-06 13:29:53 -07001371 "src/qs8-gavgpool/gen/7p7x-minmax-wasmsimd-c8-acc2.c",
1372 "src/qs8-gavgpool/gen/7p7x-minmax-wasmsimd-c16-acc2.c",
1373 "src/qs8-gavgpool/gen/7p7x-minmax-wasmsimd-c24-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001374 "src/qs8-gavgpool/gen/7x-minmax-wasmsimd-c8-acc2.c",
1375 "src/qs8-gavgpool/gen/7x-minmax-wasmsimd-c16-acc2.c",
1376 "src/qs8-gavgpool/gen/7x-minmax-wasmsimd-c24-acc2.c",
Marat Dukhan27203da2020-08-05 15:19:03 -07001377 "src/qs8-gemm/gen/1x4c8-minmax-wasmsimd-ld64.c",
Marat Dukhan27203da2020-08-05 15:19:03 -07001378 "src/qs8-gemm/gen/1x4c8-minmax-wasmsimd-ld128.c",
Marat Dukhan27203da2020-08-05 15:19:03 -07001379 "src/qs8-gemm/gen/1x4c8-xw-minmax-wasmsimd.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001380 "src/qs8-gemm/gen/2x4c8-minmax-wasmsimd-ld64.c",
1381 "src/qs8-gemm/gen/2x4c8-minmax-wasmsimd-ld128.c",
Marat Dukhan27203da2020-08-05 15:19:03 -07001382 "src/qs8-gemm/gen/2x4c8-xw-minmax-wasmsimd.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001383 "src/qs8-gemm/gen/3x4c8-minmax-wasmsimd-ld64.c",
1384 "src/qs8-gemm/gen/3x4c8-minmax-wasmsimd-ld128.c",
Marat Dukhan27203da2020-08-05 15:19:03 -07001385 "src/qs8-gemm/gen/3x4c8-xw-minmax-wasmsimd.c",
1386 "src/qs8-igemm/gen/1x4c8-minmax-wasmsimd-ld64.c",
Marat Dukhan27203da2020-08-05 15:19:03 -07001387 "src/qs8-igemm/gen/1x4c8-minmax-wasmsimd-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001388 "src/qs8-igemm/gen/2x4c8-minmax-wasmsimd-ld64.c",
Marat Dukhan27203da2020-08-05 15:19:03 -07001389 "src/qs8-igemm/gen/2x4c8-minmax-wasmsimd-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001390 "src/qs8-igemm/gen/3x4c8-minmax-wasmsimd-ld64.c",
Marat Dukhan27203da2020-08-05 15:19:03 -07001391 "src/qs8-igemm/gen/3x4c8-minmax-wasmsimd-ld128.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -07001392 "src/qs8-requantization/fp32-wasmsimd.c",
Marat Dukhan138560c2020-08-03 18:57:34 -07001393 "src/qs8-requantization/q31-wasmsimd.c",
Marat Dukhan5df27f82020-09-02 23:59:21 -07001394 "src/qs8-vadd/gen/minmax-wasmsimd-x8.c",
1395 "src/qs8-vadd/gen/minmax-wasmsimd-x16.c",
1396 "src/qs8-vadd/gen/minmax-wasmsimd-x24.c",
1397 "src/qs8-vadd/gen/minmax-wasmsimd-x32.c",
1398 "src/qs8-vaddc/gen/minmax-wasmsimd-x8.c",
1399 "src/qs8-vaddc/gen/minmax-wasmsimd-x16.c",
1400 "src/qs8-vaddc/gen/minmax-wasmsimd-x24.c",
1401 "src/qs8-vaddc/gen/minmax-wasmsimd-x32.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07001402 "src/qu8-requantization/fp32-wasmsimd.c",
Marat Dukhan22de5e72020-08-03 21:53:18 -07001403 "src/qu8-requantization/q31-wasmsimd.c",
Marat Dukhan8ee37012020-07-16 13:17:13 -07001404 "src/x32-fill/wasmsimd.c",
Marat Dukhan66d99e92020-07-16 12:56:21 -07001405 "src/x32-packx/x4-wasmsimd.c",
Marat Dukhan9306ae02020-07-16 15:51:13 -07001406 "src/x32-pad/wasmsimd.c",
Marat Dukhan9d4bfa22020-07-16 19:07:04 -07001407 "src/x32-unpool/wasmsimd.c",
Marat Dukhane3b78762020-07-16 20:02:58 -07001408 "src/x32-zip/x2-wasmsimd.c",
1409 "src/x32-zip/x3-wasmsimd.c",
1410 "src/x32-zip/x4-wasmsimd.c",
1411 "src/x32-zip/xm-wasmsimd.c",
Marat Dukhan290055c2020-06-09 12:24:29 -07001412]
1413
Marat Dukhan08c4a432019-10-03 09:29:21 -07001414# ISA-specific micro-kernels
1415NEON_UKERNELS = [
Marat Dukhanef25c6d2020-07-24 00:59:40 -07001416 "src/f32-argmaxpool/4x-neon-c4.c",
1417 "src/f32-argmaxpool/9p8x-neon-c4.c",
1418 "src/f32-argmaxpool/9x-neon-c4.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001419 "src/f32-avgpool/9p8x-minmax-neon-c4.c",
1420 "src/f32-avgpool/9x-minmax-neon-c4.c",
Marat Dukhan56b10cd2020-05-18 09:35:49 -07001421 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neon-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001422 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neon-2x2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001423 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neon-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001424 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neon-2x2.c",
Marat Dukhan56b10cd2020-05-18 09:35:49 -07001425 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neon-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001426 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neon-2x2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001427 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neon-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001428 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neon-2x2.c",
Marat Dukhanc7634882020-12-07 15:11:12 -08001429 "src/f32-conv-hwc2chw/3x3s2p1c3x4-neon-2x2.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -07001430 "src/f32-dwconv/gen/up4x4-minmax-neon-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001431 "src/f32-dwconv/gen/up4x4-minmax-neon.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001432 "src/f32-dwconv/gen/up4x9-minmax-neon-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001433 "src/f32-dwconv/gen/up4x9-minmax-neon.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -07001434 "src/f32-dwconv/gen/up4x25-minmax-neon-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001435 "src/f32-dwconv/gen/up4x25-minmax-neon.c",
1436 "src/f32-dwconv/gen/up8x4-minmax-neon-acc2.c",
1437 "src/f32-dwconv/gen/up8x4-minmax-neon.c",
1438 "src/f32-dwconv/gen/up8x9-minmax-neon-acc2.c",
1439 "src/f32-dwconv/gen/up8x9-minmax-neon.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -07001440 "src/f32-dwconv/gen/up8x25-minmax-neon-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001441 "src/f32-dwconv/gen/up8x25-minmax-neon.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001442 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-1x4-acc2.c",
1443 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-1x4-acc3.c",
1444 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-1x4-acc4.c",
Marat Dukhanc581e482020-10-24 01:28:11 -07001445 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-1x4.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001446 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-2x4-acc2.c",
Marat Dukhanc581e482020-10-24 01:28:11 -07001447 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-2x4.c",
1448 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-3x4.c",
1449 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-4x4.c",
1450 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-5x4.c",
1451 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-6x4.c",
Marat Dukhan82f0c322020-10-25 19:17:35 -07001452 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neon-1x4-acc2.c",
1453 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neon-1x4-acc3.c",
1454 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neon-1x4-acc4.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001455 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neon-1x4.c",
Marat Dukhan82f0c322020-10-25 19:17:35 -07001456 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neon-2x4-acc2.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001457 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neon-2x4.c",
1458 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neon-3x4.c",
1459 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neon-4x4.c",
Marat Dukhan149f0ea2020-10-26 12:50:33 -07001460 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-1x4-acc2.c",
1461 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-1x4-acc3.c",
1462 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-1x4-acc4.c",
1463 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-1x4-acc5.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001464 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-1x4.c",
Marat Dukhan149f0ea2020-10-26 12:50:33 -07001465 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-2x4-acc2.c",
1466 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-2x4-acc3.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001467 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-2x4.c",
Marat Dukhan149f0ea2020-10-26 12:50:33 -07001468 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-3x4-acc2.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001469 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-3x4.c",
Marat Dukhan149f0ea2020-10-26 12:50:33 -07001470 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-4x4-acc2.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001471 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-4x4.c",
1472 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-5x4.c",
Marat Dukhan30d4b252020-10-29 16:33:22 -07001473 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neon-1x4-acc2.c",
1474 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neon-1x4-acc3.c",
1475 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neon-1x4-acc4.c",
1476 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neon-1x4-acc5.c",
1477 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neon-1x4.c",
1478 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neon-2x4-acc2.c",
1479 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neon-2x4-acc3.c",
1480 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neon-2x4.c",
Marat Dukhan30d4b252020-10-29 16:33:22 -07001481 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neon-3x4-acc2.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07001482 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neon-3x4.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -07001483 "src/f32-gavgpool-cw/neon-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001484 "src/f32-gavgpool/7p7x-minmax-neon-c4.c",
1485 "src/f32-gavgpool/7x-minmax-neon-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001486 "src/f32-gemm/gen-inc/1x8inc-minmax-neon-dup-ld64.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001487 "src/f32-gemm/gen-inc/1x8inc-minmax-neon-lane-ld64.c",
1488 "src/f32-gemm/gen-inc/1x8s4inc-minmax-neon.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001489 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-dup-ld64.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001490 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-dup-ld128.c",
1491 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-lane-ld64.c",
1492 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-lane-ld128.c",
1493 "src/f32-gemm/gen-inc/4x8s4inc-minmax-neon.c",
1494 "src/f32-gemm/gen-inc/5x8inc-minmax-neon-lane-ld64.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001495 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-dup-ld64.c",
1496 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-dup-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001497 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-lane-ld64.c",
1498 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-lane-ld128.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001499 "src/f32-gemm/gen-inc/6x8s4inc-minmax-neon.c",
1500 "src/f32-gemm/gen-inc/8x8s4inc-minmax-neon.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001501 "src/f32-gemm/gen/1x8-minmax-neon-dup-ld64.c",
1502 "src/f32-gemm/gen/1x8-minmax-neon-lane-ld64.c",
1503 "src/f32-gemm/gen/1x8s4-minmax-neon.c",
1504 "src/f32-gemm/gen/4x2-minmax-neon-lane-ld64.c",
1505 "src/f32-gemm/gen/4x8-minmax-neon-dup-ld64.c",
1506 "src/f32-gemm/gen/4x8-minmax-neon-dup-ld128.c",
1507 "src/f32-gemm/gen/4x8-minmax-neon-lane-ld64.c",
1508 "src/f32-gemm/gen/4x8-minmax-neon-lane-ld128.c",
1509 "src/f32-gemm/gen/4x8s4-minmax-neon.c",
1510 "src/f32-gemm/gen/5x8-minmax-neon-lane-ld64.c",
1511 "src/f32-gemm/gen/6x8-minmax-neon-dup-ld64.c",
1512 "src/f32-gemm/gen/6x8-minmax-neon-dup-ld128.c",
1513 "src/f32-gemm/gen/6x8-minmax-neon-lane-ld64.c",
1514 "src/f32-gemm/gen/6x8-minmax-neon-lane-ld128.c",
1515 "src/f32-gemm/gen/6x8s4-minmax-neon.c",
1516 "src/f32-gemm/gen/8x8s4-minmax-neon.c",
Artsiom Ablavatski2202c812021-01-22 14:16:43 -08001517 "src/f32-ibilinear-chw/gen/neon-p4.c",
1518 "src/f32-ibilinear-chw/gen/neon-p8.c",
Frank Barchard8247e212021-02-03 18:12:33 -08001519 "src/f32-ibilinear/gen/neon-c4.c",
1520 "src/f32-ibilinear/gen/neon-c8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001521 "src/f32-igemm/gen/1x8-minmax-neon-dup-ld64.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001522 "src/f32-igemm/gen/1x8-minmax-neon-lane-ld64.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001523 "src/f32-igemm/gen/1x8s4-minmax-neon.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001524 "src/f32-igemm/gen/4x2-minmax-neon-lane-ld64.c",
1525 "src/f32-igemm/gen/4x4-minmax-neon-lane-ld64.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001526 "src/f32-igemm/gen/4x8-minmax-neon-dup-ld64.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001527 "src/f32-igemm/gen/4x8-minmax-neon-dup-ld128.c",
1528 "src/f32-igemm/gen/4x8-minmax-neon-lane-ld64.c",
1529 "src/f32-igemm/gen/4x8-minmax-neon-lane-ld128.c",
1530 "src/f32-igemm/gen/4x8s4-minmax-neon.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001531 "src/f32-igemm/gen/6x8-minmax-neon-dup-ld64.c",
1532 "src/f32-igemm/gen/6x8-minmax-neon-dup-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001533 "src/f32-igemm/gen/6x8-minmax-neon-lane-ld64.c",
1534 "src/f32-igemm/gen/6x8-minmax-neon-lane-ld128.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001535 "src/f32-igemm/gen/6x8s4-minmax-neon.c",
1536 "src/f32-igemm/gen/8x8s4-minmax-neon.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001537 "src/f32-maxpool/9p8x-minmax-neon-c4.c",
1538 "src/f32-pavgpool/9p8x-minmax-neon-c4.c",
1539 "src/f32-pavgpool/9x-minmax-neon-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001540 "src/f32-ppmm/gen/4x8-minmax-neon.c",
1541 "src/f32-ppmm/gen/8x8-minmax-neon.c",
Frank Barcharda5316982020-07-23 13:19:28 -07001542 "src/f32-prelu/gen/neon-1x4.c",
1543 "src/f32-prelu/gen/neon-1x8.c",
1544 "src/f32-prelu/gen/neon-1x16.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001545 "src/f32-prelu/gen/neon-2x4.c",
1546 "src/f32-prelu/gen/neon-2x8.c",
Frank Barcharda5316982020-07-23 13:19:28 -07001547 "src/f32-prelu/gen/neon-2x16.c",
1548 "src/f32-prelu/gen/neon-4x4.c",
1549 "src/f32-prelu/gen/neon-4x8.c",
1550 "src/f32-prelu/gen/neon-4x16.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001551 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x4.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001552 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x8-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001553 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x8.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001554 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12-acc2.c",
1555 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001556 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001557 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16-acc2.c",
1558 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001559 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001560 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20-acc2.c",
1561 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001562 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20.c",
1563 "src/f32-raddstoreexpminusmax/gen/neon-p5-x4.c",
1564 "src/f32-raddstoreexpminusmax/gen/neon-p5-x8-acc2.c",
1565 "src/f32-raddstoreexpminusmax/gen/neon-p5-x8.c",
1566 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12-acc2.c",
1567 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12-acc3.c",
1568 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12.c",
1569 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16-acc2.c",
1570 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16-acc4.c",
1571 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16.c",
1572 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20-acc2.c",
1573 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20-acc5.c",
1574 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001575 "src/f32-rmax/neon.c",
Marat Dukhan5b86c432020-12-06 19:15:03 -08001576 "src/f32-spmm/gen/4x1-minmax-neon-pipelined.c",
1577 "src/f32-spmm/gen/4x1-minmax-neon-x2.c",
1578 "src/f32-spmm/gen/4x1-minmax-neon.c",
1579 "src/f32-spmm/gen/8x1-minmax-neon-pipelined.c",
1580 "src/f32-spmm/gen/8x1-minmax-neon-x2.c",
1581 "src/f32-spmm/gen/8x1-minmax-neon.c",
1582 "src/f32-spmm/gen/12x1-minmax-neon.c",
1583 "src/f32-spmm/gen/16x1-minmax-neon-pipelined.c",
1584 "src/f32-spmm/gen/16x1-minmax-neon-x2.c",
1585 "src/f32-spmm/gen/16x1-minmax-neon.c",
1586 "src/f32-spmm/gen/32x1-minmax-neon-pipelined.c",
1587 "src/f32-spmm/gen/32x1-minmax-neon-x2.c",
1588 "src/f32-spmm/gen/32x1-minmax-neon.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001589 "src/f32-vbinary/gen/vadd-minmax-neon-x4.c",
1590 "src/f32-vbinary/gen/vadd-minmax-neon-x8.c",
1591 "src/f32-vbinary/gen/vaddc-minmax-neon-x4.c",
1592 "src/f32-vbinary/gen/vaddc-minmax-neon-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -08001593 "src/f32-vbinary/gen/vmax-neon-x4.c",
1594 "src/f32-vbinary/gen/vmax-neon-x8.c",
1595 "src/f32-vbinary/gen/vmaxc-neon-x4.c",
1596 "src/f32-vbinary/gen/vmaxc-neon-x8.c",
1597 "src/f32-vbinary/gen/vmin-neon-x4.c",
1598 "src/f32-vbinary/gen/vmin-neon-x8.c",
1599 "src/f32-vbinary/gen/vminc-neon-x4.c",
1600 "src/f32-vbinary/gen/vminc-neon-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001601 "src/f32-vbinary/gen/vmul-minmax-neon-x4.c",
1602 "src/f32-vbinary/gen/vmul-minmax-neon-x8.c",
1603 "src/f32-vbinary/gen/vmulc-minmax-neon-x4.c",
1604 "src/f32-vbinary/gen/vmulc-minmax-neon-x8.c",
1605 "src/f32-vbinary/gen/vrsubc-minmax-neon-x4.c",
1606 "src/f32-vbinary/gen/vrsubc-minmax-neon-x8.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07001607 "src/f32-vbinary/gen/vsqrdiff-neon-x4.c",
1608 "src/f32-vbinary/gen/vsqrdiff-neon-x8.c",
1609 "src/f32-vbinary/gen/vsqrdiffc-neon-x4.c",
1610 "src/f32-vbinary/gen/vsqrdiffc-neon-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001611 "src/f32-vbinary/gen/vsub-minmax-neon-x4.c",
1612 "src/f32-vbinary/gen/vsub-minmax-neon-x8.c",
1613 "src/f32-vbinary/gen/vsubc-minmax-neon-x4.c",
1614 "src/f32-vbinary/gen/vsubc-minmax-neon-x8.c",
Marat Dukhanb1eec082021-05-05 23:24:55 -07001615 "src/f32-vclamp/gen/vclamp-neon-x4.c",
1616 "src/f32-vclamp/gen/vclamp-neon-x8.c",
Marat Dukhaned6baaf2020-12-01 15:07:08 -08001617 "src/f32-velu/gen/velu-neon-rr2-lut16-p3-x4.c",
1618 "src/f32-velu/gen/velu-neon-rr2-lut16-p3-x8.c",
1619 "src/f32-velu/gen/velu-neon-rr2-lut16-p3-x12.c",
1620 "src/f32-velu/gen/velu-neon-rr2-lut16-p3-x16.c",
1621 "src/f32-velu/gen/velu-neon-rr2-lut16-p3-x20.c",
1622 "src/f32-velu/gen/velu-neon-rr2-lut16-p3-x24.c",
1623 "src/f32-velu/gen/velu-neon-rr2-p6-x4.c",
1624 "src/f32-velu/gen/velu-neon-rr2-p6-x8.c",
1625 "src/f32-velu/gen/velu-neon-rr2-p6-x12.c",
1626 "src/f32-velu/gen/velu-neon-rr2-p6-x16.c",
1627 "src/f32-velu/gen/velu-neon-rr2-p6-x20.c",
1628 "src/f32-velu/gen/velu-neon-rr2-p6-x24.c",
Marat Dukhan6674d692021-05-05 22:27:00 -07001629 "src/f32-vhswish/gen/vhswish-neon-x4.c",
1630 "src/f32-vhswish/gen/vhswish-neon-x8.c",
1631 "src/f32-vhswish/gen/vhswish-neon-x16.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07001632 "src/f32-vlrelu/gen/vlrelu-neon-x4.c",
1633 "src/f32-vlrelu/gen/vlrelu-neon-x8.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001634 "src/f32-vmulcaddc/gen/c4-minmax-neon-2x.c",
1635 "src/f32-vmulcaddc/gen/c8-minmax-neon-2x.c",
Marat Dukhanb1eec082021-05-05 23:24:55 -07001636 "src/f32-vrelu/gen/vrelu-neon-x4.c",
1637 "src/f32-vrelu/gen/vrelu-neon-x8.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07001638 "src/f32-vrnd/gen/vrndd-neon-x4.c",
1639 "src/f32-vrnd/gen/vrndd-neon-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001640 "src/f32-vrnd/gen/vrndne-neon-x4.c",
1641 "src/f32-vrnd/gen/vrndne-neon-x8.c",
1642 "src/f32-vrnd/gen/vrndu-neon-x4.c",
1643 "src/f32-vrnd/gen/vrndu-neon-x8.c",
1644 "src/f32-vrnd/gen/vrndz-neon-x4.c",
1645 "src/f32-vrnd/gen/vrndz-neon-x8.c",
Marat Dukhanb1eec082021-05-05 23:24:55 -07001646 "src/f32-vsigmoid/gen/vsigmoid-neon-rr2-lut64-p2-nr2recps-x4.c",
1647 "src/f32-vsigmoid/gen/vsigmoid-neon-rr2-lut64-p2-nr2recps-x8.c",
1648 "src/f32-vsigmoid/gen/vsigmoid-neon-rr2-lut64-p2-nr2recps-x12.c",
1649 "src/f32-vsigmoid/gen/vsigmoid-neon-rr2-lut64-p2-nr2recps-x16.c",
1650 "src/f32-vsigmoid/gen/vsigmoid-neon-rr2-lut64-p2-nr2recps-x20.c",
1651 "src/f32-vsigmoid/gen/vsigmoid-neon-rr2-lut64-p2-nr2recps-x24.c",
1652 "src/f32-vsigmoid/gen/vsigmoid-neon-rr2-lut2048-p1-nr2recps-x4.c",
1653 "src/f32-vsigmoid/gen/vsigmoid-neon-rr2-lut2048-p1-nr2recps-x8.c",
1654 "src/f32-vsigmoid/gen/vsigmoid-neon-rr2-lut2048-p1-nr2recps-x12.c",
1655 "src/f32-vsigmoid/gen/vsigmoid-neon-rr2-lut2048-p1-nr2recps-x16.c",
1656 "src/f32-vsigmoid/gen/vsigmoid-neon-rr2-lut2048-p1-nr2recps-x20.c",
1657 "src/f32-vsigmoid/gen/vsigmoid-neon-rr2-lut2048-p1-nr2recps-x24.c",
1658 "src/f32-vsigmoid/gen/vsigmoid-neon-rr2-p5-nr2recps-x4.c",
1659 "src/f32-vsigmoid/gen/vsigmoid-neon-rr2-p5-nr2recps-x8.c",
1660 "src/f32-vsigmoid/gen/vsigmoid-neon-rr2-p5-nr2recps-x12.c",
1661 "src/f32-vsigmoid/gen/vsigmoid-neon-rr2-p5-nr2recps-x16.c",
1662 "src/f32-vsigmoid/gen/vsigmoid-neon-rr2-p5-nr2recps-x20.c",
1663 "src/f32-vsigmoid/gen/vsigmoid-neon-rr2-p5-nr2recps-x24.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07001664 "src/f32-vunary/gen/vabs-neon-x4.c",
1665 "src/f32-vunary/gen/vabs-neon-x8.c",
1666 "src/f32-vunary/gen/vneg-neon-x4.c",
1667 "src/f32-vunary/gen/vneg-neon-x8.c",
1668 "src/f32-vunary/gen/vsqr-neon-x4.c",
1669 "src/f32-vunary/gen/vsqr-neon-x8.c",
Marat Dukhande390d42020-11-29 19:32:18 -08001670 "src/math/expm1minus-neon-rr2-lut16-p3.c",
1671 "src/math/expm1minus-neon-rr2-p6.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001672 "src/math/roundd-neon-addsub.c",
1673 "src/math/roundd-neon-cvt.c",
1674 "src/math/roundne-neon-addsub.c",
1675 "src/math/roundu-neon-addsub.c",
1676 "src/math/roundu-neon-cvt.c",
1677 "src/math/roundz-neon-addsub.c",
1678 "src/math/roundz-neon-cvt.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001679 "src/math/sigmoid-neon-rr2-lut64-p2-nr2recps.c",
1680 "src/math/sigmoid-neon-rr2-lut2048-p1-nr2recps.c",
1681 "src/math/sigmoid-neon-rr2-p5-nr2recps.c",
1682 "src/math/sqrt-neon-nr1rsqrts.c",
1683 "src/math/sqrt-neon-nr2rsqrts.c",
1684 "src/math/sqrt-neon-nr3rsqrts.c",
Marat Dukhan023bcf92020-08-10 12:40:50 -07001685 "src/qs8-dwconv/gen/up8x9-minmax-neon-mul16.c",
1686 "src/qs8-dwconv/gen/up16x9-minmax-neon-mul16.c",
1687 "src/qs8-dwconv/gen/up24x9-minmax-neon-mul16.c",
1688 "src/qs8-dwconv/gen/up32x9-minmax-neon-mul16.c",
Marat Dukhan281262d2020-08-10 13:23:21 -07001689 "src/qs8-gavgpool/gen/7p7x-minmax-neon-c8-acc2.c",
1690 "src/qs8-gavgpool/gen/7p7x-minmax-neon-c16-acc2.c",
1691 "src/qs8-gavgpool/gen/7p7x-minmax-neon-c24-acc2.c",
1692 "src/qs8-gavgpool/gen/7p7x-minmax-neon-c32-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001693 "src/qs8-gavgpool/gen/7x-minmax-neon-c8-acc2.c",
1694 "src/qs8-gavgpool/gen/7x-minmax-neon-c16-acc2.c",
1695 "src/qs8-gavgpool/gen/7x-minmax-neon-c24-acc2.c",
1696 "src/qs8-gavgpool/gen/7x-minmax-neon-c32-acc2.c",
Frank Barchard1f51d382021-04-20 14:07:14 -07001697 "src/qs8-gemm/gen/1x8-minmax-neon-mlal-lane-prfm.c",
Frank Barchardec0bf142021-01-15 10:14:05 -08001698 "src/qs8-gemm/gen/1x8-minmax-neon-mlal-lane.c",
Frank Barchardec0bf142021-01-15 10:14:05 -08001699 "src/qs8-gemm/gen/1x8-minmax-neon-mull-addw-dup.c",
Frank Barchard8247e212021-02-03 18:12:33 -08001700 "src/qs8-gemm/gen/1x8c2-minmax-neon-mlal-padal-dup.c",
Frank Barchard2302ffd2021-01-22 14:08:02 -08001701 "src/qs8-gemm/gen/1x8c2-minmax-neon-mull-padal-dup.c",
Frank Barchardda78da12021-03-02 14:28:00 -08001702 "src/qs8-gemm/gen/1x8c8-minmax-neon-mlal-padal.c",
Frank Barchard8247e212021-02-03 18:12:33 -08001703 "src/qs8-gemm/gen/1x8c8-minmax-neon-mull-padal.c",
Frank Barchard671d1b02021-03-10 15:42:08 -08001704 "src/qs8-gemm/gen/1x8c16-minmax-neon-mlal-padal.c",
Frank Barchard1f51d382021-04-20 14:07:14 -07001705 "src/qs8-gemm/gen/1x16-minmax-neon-mlal-lane-prfm.c",
Frank Barchardcadd4222021-01-20 16:27:25 -08001706 "src/qs8-gemm/gen/1x16-minmax-neon-mlal-lane.c",
1707 "src/qs8-gemm/gen/1x16-minmax-neon-mull-addw-dup.c",
Frank Barchard8247e212021-02-03 18:12:33 -08001708 "src/qs8-gemm/gen/1x16c2-minmax-neon-mlal-padal-dup.c",
Frank Barchard2302ffd2021-01-22 14:08:02 -08001709 "src/qs8-gemm/gen/1x16c2-minmax-neon-mull-padal-dup.c",
Frank Barchardda78da12021-03-02 14:28:00 -08001710 "src/qs8-gemm/gen/1x16c8-minmax-neon-mlal-padal.c",
Frank Barchard8247e212021-02-03 18:12:33 -08001711 "src/qs8-gemm/gen/1x16c8-minmax-neon-mull-padal.c",
Frank Barchard671d1b02021-03-10 15:42:08 -08001712 "src/qs8-gemm/gen/1x16c16-minmax-neon-mlal-padal.c",
Frank Barchard1f51d382021-04-20 14:07:14 -07001713 "src/qs8-gemm/gen/2x8-minmax-neon-mlal-lane-prfm.c",
Frank Barchardcadd4222021-01-20 16:27:25 -08001714 "src/qs8-gemm/gen/2x8-minmax-neon-mlal-lane.c",
Frank Barchardec0bf142021-01-15 10:14:05 -08001715 "src/qs8-gemm/gen/2x8-minmax-neon-mull-addw-dup.c",
Frank Barchard8247e212021-02-03 18:12:33 -08001716 "src/qs8-gemm/gen/2x8c2-minmax-neon-mlal-padal-dup.c",
Frank Barchard2302ffd2021-01-22 14:08:02 -08001717 "src/qs8-gemm/gen/2x8c2-minmax-neon-mull-padal-dup.c",
Frank Barchardda78da12021-03-02 14:28:00 -08001718 "src/qs8-gemm/gen/2x8c8-minmax-neon-mlal-padal.c",
Frank Barchard8247e212021-02-03 18:12:33 -08001719 "src/qs8-gemm/gen/2x8c8-minmax-neon-mull-padal.c",
Frank Barchard671d1b02021-03-10 15:42:08 -08001720 "src/qs8-gemm/gen/2x8c16-minmax-neon-mlal-padal.c",
Frank Barchard1f51d382021-04-20 14:07:14 -07001721 "src/qs8-gemm/gen/2x16-minmax-neon-mlal-lane-prfm.c",
Frank Barchardcadd4222021-01-20 16:27:25 -08001722 "src/qs8-gemm/gen/2x16-minmax-neon-mlal-lane.c",
1723 "src/qs8-gemm/gen/2x16-minmax-neon-mull-addw-dup.c",
Frank Barchard8247e212021-02-03 18:12:33 -08001724 "src/qs8-gemm/gen/2x16c2-minmax-neon-mlal-padal-dup.c",
Frank Barchard2302ffd2021-01-22 14:08:02 -08001725 "src/qs8-gemm/gen/2x16c2-minmax-neon-mull-padal-dup.c",
Frank Barchardda78da12021-03-02 14:28:00 -08001726 "src/qs8-gemm/gen/2x16c8-minmax-neon-mlal-padal.c",
Frank Barchard8247e212021-02-03 18:12:33 -08001727 "src/qs8-gemm/gen/2x16c8-minmax-neon-mull-padal.c",
Frank Barchard671d1b02021-03-10 15:42:08 -08001728 "src/qs8-gemm/gen/2x16c16-minmax-neon-mlal-padal.c",
Frank Barchard1f51d382021-04-20 14:07:14 -07001729 "src/qs8-gemm/gen/3x8-minmax-neon-mlal-lane-prfm.c",
Frank Barchard2302ffd2021-01-22 14:08:02 -08001730 "src/qs8-gemm/gen/3x8-minmax-neon-mlal-lane.c",
1731 "src/qs8-gemm/gen/3x8-minmax-neon-mull-addw-dup.c",
Frank Barchard8247e212021-02-03 18:12:33 -08001732 "src/qs8-gemm/gen/3x8c2-minmax-neon-mlal-padal-dup.c",
Frank Barchard2302ffd2021-01-22 14:08:02 -08001733 "src/qs8-gemm/gen/3x8c2-minmax-neon-mull-padal-dup.c",
Frank Barchardda78da12021-03-02 14:28:00 -08001734 "src/qs8-gemm/gen/3x8c8-minmax-neon-mlal-padal.c",
Frank Barchard8247e212021-02-03 18:12:33 -08001735 "src/qs8-gemm/gen/3x8c8-minmax-neon-mull-padal.c",
Frank Barchard671d1b02021-03-10 15:42:08 -08001736 "src/qs8-gemm/gen/3x8c16-minmax-neon-mlal-padal.c",
Frank Barchard1f51d382021-04-20 14:07:14 -07001737 "src/qs8-gemm/gen/3x16-minmax-neon-mlal-lane-prfm.c",
Frank Barchard8247e212021-02-03 18:12:33 -08001738 "src/qs8-gemm/gen/3x16-minmax-neon-mlal-lane.c",
1739 "src/qs8-gemm/gen/3x16-minmax-neon-mull-addw-dup.c",
1740 "src/qs8-gemm/gen/3x16c2-minmax-neon-mlal-padal-dup.c",
1741 "src/qs8-gemm/gen/3x16c2-minmax-neon-mull-padal-dup.c",
Frank Barchardda78da12021-03-02 14:28:00 -08001742 "src/qs8-gemm/gen/3x16c8-minmax-neon-mlal-padal.c",
Frank Barchard8247e212021-02-03 18:12:33 -08001743 "src/qs8-gemm/gen/3x16c8-minmax-neon-mull-padal.c",
Frank Barchard671d1b02021-03-10 15:42:08 -08001744 "src/qs8-gemm/gen/3x16c16-minmax-neon-mlal-padal.c",
Frank Barchard1f51d382021-04-20 14:07:14 -07001745 "src/qs8-gemm/gen/4x8-minmax-neon-mlal-lane-prfm.c",
Frank Barchardcadd4222021-01-20 16:27:25 -08001746 "src/qs8-gemm/gen/4x8-minmax-neon-mlal-lane.c",
Frank Barchardec0bf142021-01-15 10:14:05 -08001747 "src/qs8-gemm/gen/4x8-minmax-neon-mull-addw-dup.c",
Frank Barchard8247e212021-02-03 18:12:33 -08001748 "src/qs8-gemm/gen/4x8c2-minmax-neon-mlal-padal-dup.c",
Frank Barchard2302ffd2021-01-22 14:08:02 -08001749 "src/qs8-gemm/gen/4x8c2-minmax-neon-mull-padal-dup.c",
Frank Barchardda78da12021-03-02 14:28:00 -08001750 "src/qs8-gemm/gen/4x8c8-minmax-neon-mlal-padal.c",
Frank Barchard8247e212021-02-03 18:12:33 -08001751 "src/qs8-gemm/gen/4x8c8-minmax-neon-mull-padal.c",
Frank Barchard671d1b02021-03-10 15:42:08 -08001752 "src/qs8-gemm/gen/4x8c16-minmax-neon-mlal-padal.c",
Frank Barchard1f51d382021-04-20 14:07:14 -07001753 "src/qs8-gemm/gen/4x16-minmax-neon-mlal-lane-prfm.c",
Frank Barchardcadd4222021-01-20 16:27:25 -08001754 "src/qs8-gemm/gen/4x16-minmax-neon-mlal-lane.c",
1755 "src/qs8-gemm/gen/4x16-minmax-neon-mull-addw-dup.c",
Frank Barchard8247e212021-02-03 18:12:33 -08001756 "src/qs8-gemm/gen/4x16c2-minmax-neon-mlal-padal-dup.c",
Frank Barchard2302ffd2021-01-22 14:08:02 -08001757 "src/qs8-gemm/gen/4x16c2-minmax-neon-mull-padal-dup.c",
Frank Barchardda78da12021-03-02 14:28:00 -08001758 "src/qs8-gemm/gen/4x16c8-minmax-neon-mlal-padal.c",
Frank Barchard58990122021-01-29 16:09:56 -08001759 "src/qs8-gemm/gen/4x16c8-minmax-neon-mull-padal.c",
Frank Barchard71c4d1a2021-02-19 10:07:23 -08001760 "src/qs8-gemm/gen/4x16c16-minmax-neon-mlal-padal.c",
Frank Barcharddddb38f2021-05-03 16:51:48 -07001761 "src/qs8-gemm/gen/6x8-minmax-neon-mlal-lane-prfm.c",
1762 "src/qs8-gemm/gen/6x8-minmax-neon-mlal-lane.c",
1763 "src/qs8-gemm/gen/6x16-minmax-neon-mlal-lane-prfm.c",
1764 "src/qs8-gemm/gen/6x16-minmax-neon-mlal-lane.c",
Frank Barchard1f51d382021-04-20 14:07:14 -07001765 "src/qs8-igemm/gen/1x8-minmax-neon-mlal-lane-prfm.c",
Frank Barchardec0bf142021-01-15 10:14:05 -08001766 "src/qs8-igemm/gen/1x8-minmax-neon-mlal-lane.c",
Frank Barchard671d1b02021-03-10 15:42:08 -08001767 "src/qs8-igemm/gen/1x8-minmax-neon-mull-addw-dup.c",
1768 "src/qs8-igemm/gen/1x8c2-minmax-neon-mlal-padal-dup.c",
1769 "src/qs8-igemm/gen/1x8c2-minmax-neon-mull-padal-dup.c",
Frank Barchardda78da12021-03-02 14:28:00 -08001770 "src/qs8-igemm/gen/1x8c8-minmax-neon-mlal-padal.c",
1771 "src/qs8-igemm/gen/1x8c8-minmax-neon-mull-padal.c",
Frank Barchard671d1b02021-03-10 15:42:08 -08001772 "src/qs8-igemm/gen/1x8c16-minmax-neon-mlal-padal.c",
Frank Barchard1f51d382021-04-20 14:07:14 -07001773 "src/qs8-igemm/gen/1x16-minmax-neon-mlal-lane-prfm.c",
Frank Barchard671d1b02021-03-10 15:42:08 -08001774 "src/qs8-igemm/gen/1x16-minmax-neon-mlal-lane.c",
1775 "src/qs8-igemm/gen/1x16-minmax-neon-mull-addw-dup.c",
1776 "src/qs8-igemm/gen/1x16c2-minmax-neon-mlal-padal-dup.c",
1777 "src/qs8-igemm/gen/1x16c2-minmax-neon-mull-padal-dup.c",
1778 "src/qs8-igemm/gen/1x16c8-minmax-neon-mlal-padal.c",
1779 "src/qs8-igemm/gen/1x16c8-minmax-neon-mull-padal.c",
1780 "src/qs8-igemm/gen/1x16c16-minmax-neon-mlal-padal.c",
Frank Barchard1f51d382021-04-20 14:07:14 -07001781 "src/qs8-igemm/gen/2x8-minmax-neon-mlal-lane-prfm.c",
Frank Barchard671d1b02021-03-10 15:42:08 -08001782 "src/qs8-igemm/gen/2x8-minmax-neon-mlal-lane.c",
1783 "src/qs8-igemm/gen/2x8-minmax-neon-mull-addw-dup.c",
1784 "src/qs8-igemm/gen/2x8c2-minmax-neon-mlal-padal-dup.c",
1785 "src/qs8-igemm/gen/2x8c2-minmax-neon-mull-padal-dup.c",
Frank Barchardda78da12021-03-02 14:28:00 -08001786 "src/qs8-igemm/gen/2x8c8-minmax-neon-mlal-padal.c",
1787 "src/qs8-igemm/gen/2x8c8-minmax-neon-mull-padal.c",
Frank Barchard671d1b02021-03-10 15:42:08 -08001788 "src/qs8-igemm/gen/2x8c16-minmax-neon-mlal-padal.c",
Frank Barchard1f51d382021-04-20 14:07:14 -07001789 "src/qs8-igemm/gen/2x16-minmax-neon-mlal-lane-prfm.c",
Frank Barchard671d1b02021-03-10 15:42:08 -08001790 "src/qs8-igemm/gen/2x16-minmax-neon-mlal-lane.c",
1791 "src/qs8-igemm/gen/2x16-minmax-neon-mull-addw-dup.c",
1792 "src/qs8-igemm/gen/2x16c2-minmax-neon-mlal-padal-dup.c",
1793 "src/qs8-igemm/gen/2x16c2-minmax-neon-mull-padal-dup.c",
1794 "src/qs8-igemm/gen/2x16c8-minmax-neon-mlal-padal.c",
1795 "src/qs8-igemm/gen/2x16c8-minmax-neon-mull-padal.c",
1796 "src/qs8-igemm/gen/2x16c16-minmax-neon-mlal-padal.c",
Frank Barchard1f51d382021-04-20 14:07:14 -07001797 "src/qs8-igemm/gen/3x8-minmax-neon-mlal-lane-prfm.c",
Frank Barchard671d1b02021-03-10 15:42:08 -08001798 "src/qs8-igemm/gen/3x8-minmax-neon-mlal-lane.c",
1799 "src/qs8-igemm/gen/3x8-minmax-neon-mull-addw-dup.c",
1800 "src/qs8-igemm/gen/3x8c2-minmax-neon-mlal-padal-dup.c",
1801 "src/qs8-igemm/gen/3x8c2-minmax-neon-mull-padal-dup.c",
Frank Barchardda78da12021-03-02 14:28:00 -08001802 "src/qs8-igemm/gen/3x8c8-minmax-neon-mlal-padal.c",
1803 "src/qs8-igemm/gen/3x8c8-minmax-neon-mull-padal.c",
Frank Barchard671d1b02021-03-10 15:42:08 -08001804 "src/qs8-igemm/gen/3x8c16-minmax-neon-mlal-padal.c",
Frank Barchard1f51d382021-04-20 14:07:14 -07001805 "src/qs8-igemm/gen/3x16-minmax-neon-mlal-lane-prfm.c",
Frank Barchard671d1b02021-03-10 15:42:08 -08001806 "src/qs8-igemm/gen/3x16-minmax-neon-mlal-lane.c",
1807 "src/qs8-igemm/gen/3x16-minmax-neon-mull-addw-dup.c",
1808 "src/qs8-igemm/gen/3x16c2-minmax-neon-mlal-padal-dup.c",
1809 "src/qs8-igemm/gen/3x16c2-minmax-neon-mull-padal-dup.c",
1810 "src/qs8-igemm/gen/3x16c8-minmax-neon-mlal-padal.c",
1811 "src/qs8-igemm/gen/3x16c8-minmax-neon-mull-padal.c",
1812 "src/qs8-igemm/gen/3x16c16-minmax-neon-mlal-padal.c",
Frank Barchard1f51d382021-04-20 14:07:14 -07001813 "src/qs8-igemm/gen/4x8-minmax-neon-mlal-lane-prfm.c",
Frank Barchard671d1b02021-03-10 15:42:08 -08001814 "src/qs8-igemm/gen/4x8-minmax-neon-mlal-lane.c",
1815 "src/qs8-igemm/gen/4x8-minmax-neon-mull-addw-dup.c",
1816 "src/qs8-igemm/gen/4x8c2-minmax-neon-mlal-padal-dup.c",
1817 "src/qs8-igemm/gen/4x8c2-minmax-neon-mull-padal-dup.c",
Frank Barchardda78da12021-03-02 14:28:00 -08001818 "src/qs8-igemm/gen/4x8c8-minmax-neon-mlal-padal.c",
1819 "src/qs8-igemm/gen/4x8c8-minmax-neon-mull-padal.c",
Frank Barchardda78da12021-03-02 14:28:00 -08001820 "src/qs8-igemm/gen/4x8c16-minmax-neon-mlal-padal.c",
Frank Barchard1f51d382021-04-20 14:07:14 -07001821 "src/qs8-igemm/gen/4x16-minmax-neon-mlal-lane-prfm.c",
Frank Barchard671d1b02021-03-10 15:42:08 -08001822 "src/qs8-igemm/gen/4x16-minmax-neon-mlal-lane.c",
Frank Barchardda78da12021-03-02 14:28:00 -08001823 "src/qs8-igemm/gen/4x16-minmax-neon-mull-addw-dup.c",
Frank Barchard671d1b02021-03-10 15:42:08 -08001824 "src/qs8-igemm/gen/4x16c2-minmax-neon-mlal-padal-dup.c",
1825 "src/qs8-igemm/gen/4x16c2-minmax-neon-mull-padal-dup.c",
1826 "src/qs8-igemm/gen/4x16c8-minmax-neon-mlal-padal.c",
1827 "src/qs8-igemm/gen/4x16c8-minmax-neon-mull-padal.c",
1828 "src/qs8-igemm/gen/4x16c16-minmax-neon-mlal-padal.c",
Frank Barcharddddb38f2021-05-03 16:51:48 -07001829 "src/qs8-igemm/gen/6x8-minmax-neon-mlal-lane-prfm.c",
1830 "src/qs8-igemm/gen/6x8-minmax-neon-mlal-lane.c",
1831 "src/qs8-igemm/gen/6x16-minmax-neon-mlal-lane-prfm.c",
1832 "src/qs8-igemm/gen/6x16-minmax-neon-mlal-lane.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -07001833 "src/qs8-requantization/fp32-neon.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001834 "src/qs8-requantization/precise-neon.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -07001835 "src/qs8-requantization/q31-neon.c",
Marat Dukhanba7b2792020-09-02 14:26:45 -07001836 "src/qs8-vadd/gen/minmax-neon-ld64-x8.c",
1837 "src/qs8-vadd/gen/minmax-neon-ld64-x16.c",
1838 "src/qs8-vadd/gen/minmax-neon-ld64-x24.c",
1839 "src/qs8-vadd/gen/minmax-neon-ld64-x32.c",
1840 "src/qs8-vaddc/gen/minmax-neon-ld64-x8.c",
1841 "src/qs8-vaddc/gen/minmax-neon-ld64-x16.c",
1842 "src/qs8-vaddc/gen/minmax-neon-ld64-x24.c",
1843 "src/qs8-vaddc/gen/minmax-neon-ld64-x32.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -07001844 "src/qu8-avgpool/9p8x-minmax-neon-c8.c",
1845 "src/qu8-avgpool/9x-minmax-neon-c8.c",
1846 "src/qu8-dwconv/up8x9-minmax-neon.c",
1847 "src/qu8-gavgpool/7p7x-minmax-neon-c8.c",
1848 "src/qu8-gavgpool/7x-minmax-neon-c8.c",
1849 "src/qu8-gemm/4x8-minmax-neon.c",
1850 "src/qu8-gemm/8x8-minmax-neon.c",
1851 "src/qu8-igemm/4x8-minmax-neon.c",
1852 "src/qu8-igemm/8x8-minmax-neon.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07001853 "src/qu8-requantization/fp32-neon.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001854 "src/qu8-requantization/precise-neon.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07001855 "src/qu8-requantization/q31-neon.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -07001856 "src/qu8-vadd/minmax-neon.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001857 "src/u8-maxpool/9p8x-minmax-neon-c16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001858 "src/u8-rmax/neon.c",
Marat Dukhan6674d692021-05-05 22:27:00 -07001859 "src/u8-vclamp/neon-x64.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001860 "src/x8-zip/x2-neon.c",
1861 "src/x8-zip/x3-neon.c",
1862 "src/x8-zip/x4-neon.c",
1863 "src/x8-zip/xm-neon.c",
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -07001864 "src/x32-fill/neon.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001865 "src/x32-packx/x4-neon-st4.c",
Marat Dukhan63523d42020-05-22 17:07:33 -07001866 "src/x32-pad/neon.c",
Marat Dukhan57dccd82020-04-14 00:53:10 -07001867 "src/x32-unpool/neon.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001868 "src/x32-zip/x2-neon.c",
1869 "src/x32-zip/x3-neon.c",
1870 "src/x32-zip/x4-neon.c",
1871 "src/x32-zip/xm-neon.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001872]
1873
1874NEONFMA_UKERNELS = [
Frank Barchard04336c12020-10-22 16:48:55 -07001875 "src/f32-dwconv/gen/up4x4-minmax-neonfma-acc2.c",
1876 "src/f32-dwconv/gen/up4x4-minmax-neonfma.c",
1877 "src/f32-dwconv/gen/up4x9-minmax-neonfma-acc2.c",
1878 "src/f32-dwconv/gen/up4x9-minmax-neonfma.c",
1879 "src/f32-dwconv/gen/up4x25-minmax-neonfma-acc2.c",
1880 "src/f32-dwconv/gen/up4x25-minmax-neonfma.c",
1881 "src/f32-dwconv/gen/up8x4-minmax-neonfma-acc2.c",
1882 "src/f32-dwconv/gen/up8x4-minmax-neonfma.c",
1883 "src/f32-dwconv/gen/up8x9-minmax-neonfma-acc2.c",
1884 "src/f32-dwconv/gen/up8x9-minmax-neonfma.c",
1885 "src/f32-dwconv/gen/up8x25-minmax-neonfma-acc2.c",
1886 "src/f32-dwconv/gen/up8x25-minmax-neonfma.c",
1887 "src/f32-gemm/gen-inc/1x8inc-minmax-neonfma-dup-ld64.c",
1888 "src/f32-gemm/gen-inc/1x8s4inc-minmax-neonfma.c",
1889 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-dup-ld64.c",
1890 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-dup-ld128.c",
1891 "src/f32-gemm/gen-inc/4x8s4inc-minmax-neonfma.c",
1892 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-dup-ld64.c",
1893 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-dup-ld128.c",
1894 "src/f32-gemm/gen-inc/6x8s4inc-minmax-neonfma.c",
1895 "src/f32-gemm/gen-inc/8x8s4inc-minmax-neonfma.c",
1896 "src/f32-gemm/gen/1x8-minmax-neonfma-dup-ld64.c",
1897 "src/f32-gemm/gen/1x8s4-minmax-neonfma.c",
1898 "src/f32-gemm/gen/4x8-minmax-neonfma-dup-ld64.c",
1899 "src/f32-gemm/gen/4x8-minmax-neonfma-dup-ld128.c",
1900 "src/f32-gemm/gen/4x8s4-minmax-neonfma.c",
1901 "src/f32-gemm/gen/6x8-minmax-neonfma-dup-ld64.c",
1902 "src/f32-gemm/gen/6x8-minmax-neonfma-dup-ld128.c",
1903 "src/f32-gemm/gen/6x8s4-minmax-neonfma.c",
1904 "src/f32-gemm/gen/8x8s4-minmax-neonfma.c",
Artsiom Ablavatski2202c812021-01-22 14:16:43 -08001905 "src/f32-ibilinear-chw/gen/neonfma-p4.c",
1906 "src/f32-ibilinear-chw/gen/neonfma-p8.c",
Frank Barchard8247e212021-02-03 18:12:33 -08001907 "src/f32-ibilinear/gen/neonfma-c4.c",
1908 "src/f32-ibilinear/gen/neonfma-c8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001909 "src/f32-igemm/gen/1x8-minmax-neonfma-dup-ld64.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001910 "src/f32-igemm/gen/1x8s4-minmax-neonfma.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001911 "src/f32-igemm/gen/4x8-minmax-neonfma-dup-ld64.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001912 "src/f32-igemm/gen/4x8-minmax-neonfma-dup-ld128.c",
1913 "src/f32-igemm/gen/4x8s4-minmax-neonfma.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001914 "src/f32-igemm/gen/6x8-minmax-neonfma-dup-ld64.c",
1915 "src/f32-igemm/gen/6x8-minmax-neonfma-dup-ld128.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001916 "src/f32-igemm/gen/6x8s4-minmax-neonfma.c",
1917 "src/f32-igemm/gen/8x8s4-minmax-neonfma.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001918 "src/f32-ppmm/gen/4x8-minmax-neonfma.c",
1919 "src/f32-ppmm/gen/8x8-minmax-neonfma.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001920 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x4.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001921 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x8-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001922 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x8.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001923 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12-acc2.c",
1924 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001925 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001926 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16-acc2.c",
1927 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001928 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001929 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20-acc2.c",
1930 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001931 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20.c",
1932 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x4.c",
1933 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x8-acc2.c",
1934 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x8.c",
1935 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12-acc2.c",
1936 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12-acc3.c",
1937 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12.c",
1938 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16-acc2.c",
1939 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16-acc4.c",
1940 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16.c",
1941 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20-acc2.c",
1942 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20-acc5.c",
1943 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20.c",
Marat Dukhan2fa7a0c2020-12-06 19:09:02 -08001944 "src/f32-spmm/gen/4x1-minmax-neonfma-pipelined.c",
1945 "src/f32-spmm/gen/4x1-minmax-neonfma-x2.c",
1946 "src/f32-spmm/gen/4x1-minmax-neonfma.c",
1947 "src/f32-spmm/gen/8x1-minmax-neonfma-pipelined.c",
1948 "src/f32-spmm/gen/8x1-minmax-neonfma-x2.c",
1949 "src/f32-spmm/gen/8x1-minmax-neonfma.c",
1950 "src/f32-spmm/gen/12x1-minmax-neonfma.c",
1951 "src/f32-spmm/gen/16x1-minmax-neonfma-pipelined.c",
1952 "src/f32-spmm/gen/16x1-minmax-neonfma-x2.c",
1953 "src/f32-spmm/gen/16x1-minmax-neonfma.c",
1954 "src/f32-spmm/gen/32x1-minmax-neonfma-pipelined.c",
1955 "src/f32-spmm/gen/32x1-minmax-neonfma-x2.c",
1956 "src/f32-spmm/gen/32x1-minmax-neonfma.c",
Marat Dukhaned6baaf2020-12-01 15:07:08 -08001957 "src/f32-velu/gen/velu-neonfma-rr1-lut16-p3-x4.c",
1958 "src/f32-velu/gen/velu-neonfma-rr1-lut16-p3-x8.c",
1959 "src/f32-velu/gen/velu-neonfma-rr1-lut16-p3-x12.c",
1960 "src/f32-velu/gen/velu-neonfma-rr1-lut16-p3-x16.c",
1961 "src/f32-velu/gen/velu-neonfma-rr1-lut16-p3-x20.c",
1962 "src/f32-velu/gen/velu-neonfma-rr1-lut16-p3-x24.c",
1963 "src/f32-velu/gen/velu-neonfma-rr1-p6-x4.c",
1964 "src/f32-velu/gen/velu-neonfma-rr1-p6-x8.c",
1965 "src/f32-velu/gen/velu-neonfma-rr1-p6-x12.c",
1966 "src/f32-velu/gen/velu-neonfma-rr1-p6-x16.c",
1967 "src/f32-velu/gen/velu-neonfma-rr1-p6-x20.c",
1968 "src/f32-velu/gen/velu-neonfma-rr1-p6-x24.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001969 "src/f32-vmulcaddc/gen/c4-minmax-neonfma-2x.c",
1970 "src/f32-vmulcaddc/gen/c8-minmax-neonfma-2x.c",
Marat Dukhanb1eec082021-05-05 23:24:55 -07001971 "src/f32-vsigmoid/gen/vsigmoid-neonfma-rr1-lut64-p2-nr1recps1fma-x4.c",
1972 "src/f32-vsigmoid/gen/vsigmoid-neonfma-rr1-lut64-p2-nr1recps1fma-x8.c",
1973 "src/f32-vsigmoid/gen/vsigmoid-neonfma-rr1-lut64-p2-nr1recps1fma-x12.c",
1974 "src/f32-vsigmoid/gen/vsigmoid-neonfma-rr1-lut64-p2-nr1recps1fma-x16.c",
1975 "src/f32-vsigmoid/gen/vsigmoid-neonfma-rr1-lut64-p2-nr1recps1fma-x20.c",
1976 "src/f32-vsigmoid/gen/vsigmoid-neonfma-rr1-lut64-p2-nr1recps1fma-x24.c",
1977 "src/f32-vsigmoid/gen/vsigmoid-neonfma-rr1-lut64-p2-nr2fma-x4.c",
1978 "src/f32-vsigmoid/gen/vsigmoid-neonfma-rr1-lut64-p2-nr2fma-x8.c",
1979 "src/f32-vsigmoid/gen/vsigmoid-neonfma-rr1-lut64-p2-nr2fma-x12.c",
1980 "src/f32-vsigmoid/gen/vsigmoid-neonfma-rr1-lut64-p2-nr2fma-x16.c",
1981 "src/f32-vsigmoid/gen/vsigmoid-neonfma-rr1-lut64-p2-nr2fma-x20.c",
1982 "src/f32-vsigmoid/gen/vsigmoid-neonfma-rr1-lut64-p2-nr2fma-x24.c",
1983 "src/f32-vsigmoid/gen/vsigmoid-neonfma-rr1-lut64-p2-nr2recps-x4.c",
1984 "src/f32-vsigmoid/gen/vsigmoid-neonfma-rr1-lut64-p2-nr2recps-x8.c",
1985 "src/f32-vsigmoid/gen/vsigmoid-neonfma-rr1-lut64-p2-nr2recps-x12.c",
1986 "src/f32-vsigmoid/gen/vsigmoid-neonfma-rr1-lut64-p2-nr2recps-x16.c",
1987 "src/f32-vsigmoid/gen/vsigmoid-neonfma-rr1-lut64-p2-nr2recps-x20.c",
1988 "src/f32-vsigmoid/gen/vsigmoid-neonfma-rr1-lut64-p2-nr2recps-x24.c",
1989 "src/f32-vsigmoid/gen/vsigmoid-neonfma-rr1-lut2048-p1-nr1recps1fma-x4.c",
1990 "src/f32-vsigmoid/gen/vsigmoid-neonfma-rr1-lut2048-p1-nr1recps1fma-x8.c",
1991 "src/f32-vsigmoid/gen/vsigmoid-neonfma-rr1-lut2048-p1-nr1recps1fma-x12.c",
1992 "src/f32-vsigmoid/gen/vsigmoid-neonfma-rr1-lut2048-p1-nr1recps1fma-x16.c",
1993 "src/f32-vsigmoid/gen/vsigmoid-neonfma-rr1-lut2048-p1-nr1recps1fma-x20.c",
1994 "src/f32-vsigmoid/gen/vsigmoid-neonfma-rr1-lut2048-p1-nr1recps1fma-x24.c",
1995 "src/f32-vsigmoid/gen/vsigmoid-neonfma-rr1-lut2048-p1-nr2fma-x4.c",
1996 "src/f32-vsigmoid/gen/vsigmoid-neonfma-rr1-lut2048-p1-nr2fma-x8.c",
1997 "src/f32-vsigmoid/gen/vsigmoid-neonfma-rr1-lut2048-p1-nr2fma-x12.c",
1998 "src/f32-vsigmoid/gen/vsigmoid-neonfma-rr1-lut2048-p1-nr2fma-x16.c",
1999 "src/f32-vsigmoid/gen/vsigmoid-neonfma-rr1-lut2048-p1-nr2fma-x20.c",
2000 "src/f32-vsigmoid/gen/vsigmoid-neonfma-rr1-lut2048-p1-nr2fma-x24.c",
2001 "src/f32-vsigmoid/gen/vsigmoid-neonfma-rr1-lut2048-p1-nr2recps-x4.c",
2002 "src/f32-vsigmoid/gen/vsigmoid-neonfma-rr1-lut2048-p1-nr2recps-x8.c",
2003 "src/f32-vsigmoid/gen/vsigmoid-neonfma-rr1-lut2048-p1-nr2recps-x12.c",
2004 "src/f32-vsigmoid/gen/vsigmoid-neonfma-rr1-lut2048-p1-nr2recps-x16.c",
2005 "src/f32-vsigmoid/gen/vsigmoid-neonfma-rr1-lut2048-p1-nr2recps-x20.c",
2006 "src/f32-vsigmoid/gen/vsigmoid-neonfma-rr1-lut2048-p1-nr2recps-x24.c",
2007 "src/f32-vsigmoid/gen/vsigmoid-neonfma-rr1-p5-nr1recps1fma-x4.c",
2008 "src/f32-vsigmoid/gen/vsigmoid-neonfma-rr1-p5-nr1recps1fma-x8.c",
2009 "src/f32-vsigmoid/gen/vsigmoid-neonfma-rr1-p5-nr1recps1fma-x12.c",
2010 "src/f32-vsigmoid/gen/vsigmoid-neonfma-rr1-p5-nr1recps1fma-x16.c",
2011 "src/f32-vsigmoid/gen/vsigmoid-neonfma-rr1-p5-nr1recps1fma-x20.c",
2012 "src/f32-vsigmoid/gen/vsigmoid-neonfma-rr1-p5-nr1recps1fma-x24.c",
2013 "src/f32-vsigmoid/gen/vsigmoid-neonfma-rr1-p5-nr2fma-x4.c",
2014 "src/f32-vsigmoid/gen/vsigmoid-neonfma-rr1-p5-nr2fma-x8.c",
2015 "src/f32-vsigmoid/gen/vsigmoid-neonfma-rr1-p5-nr2fma-x12.c",
2016 "src/f32-vsigmoid/gen/vsigmoid-neonfma-rr1-p5-nr2fma-x16.c",
2017 "src/f32-vsigmoid/gen/vsigmoid-neonfma-rr1-p5-nr2fma-x20.c",
2018 "src/f32-vsigmoid/gen/vsigmoid-neonfma-rr1-p5-nr2fma-x24.c",
2019 "src/f32-vsigmoid/gen/vsigmoid-neonfma-rr1-p5-nr2recps-x4.c",
2020 "src/f32-vsigmoid/gen/vsigmoid-neonfma-rr1-p5-nr2recps-x8.c",
2021 "src/f32-vsigmoid/gen/vsigmoid-neonfma-rr1-p5-nr2recps-x12.c",
2022 "src/f32-vsigmoid/gen/vsigmoid-neonfma-rr1-p5-nr2recps-x16.c",
2023 "src/f32-vsigmoid/gen/vsigmoid-neonfma-rr1-p5-nr2recps-x20.c",
2024 "src/f32-vsigmoid/gen/vsigmoid-neonfma-rr1-p5-nr2recps-x24.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07002025 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x4.c",
2026 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x8.c",
2027 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x12.c",
2028 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x16.c",
2029 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x20.c",
2030 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x24.c",
2031 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x28.c",
2032 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x32.c",
2033 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x36.c",
2034 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x40.c",
2035 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x4.c",
2036 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x8.c",
2037 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x12.c",
2038 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x16.c",
2039 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x20.c",
2040 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x24.c",
2041 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x28.c",
2042 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x32.c",
2043 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x36.c",
2044 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x40.c",
Marat Dukhanb7633f22020-11-20 16:34:56 -08002045 "src/math/exp-neonfma-rr2-lut64-p2.c",
2046 "src/math/exp-neonfma-rr2-p5.c",
Frank Barcharde7223ee2020-12-04 19:04:01 -08002047 "src/math/expm1minus-neonfma-rr1-lut16-p3.c",
2048 "src/math/expm1minus-neonfma-rr1-p6.c",
Marat Dukhan9dd119a2020-11-20 18:20:04 -08002049 "src/math/expminus-neonfma-rr2-lut64-p2.c",
2050 "src/math/expminus-neonfma-rr2-lut2048-p1.c",
2051 "src/math/expminus-neonfma-rr2-p5.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08002052 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr1recps1fma.c",
2053 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr2fma.c",
2054 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr2recps.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002055 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr1recps1fma.c",
2056 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr2fma.c",
2057 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr2recps.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08002058 "src/math/sigmoid-neonfma-rr1-p5-nr1recps1fma.c",
2059 "src/math/sigmoid-neonfma-rr1-p5-nr2fma.c",
2060 "src/math/sigmoid-neonfma-rr1-p5-nr2recps.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08002061 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr1recps1fma.c",
2062 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr2fma.c",
2063 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr2recps.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002064 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr1recps1fma.c",
2065 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr2fma.c",
2066 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr2recps.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08002067 "src/math/sigmoid-neonfma-rr2-p5-nr1recps1fma.c",
2068 "src/math/sigmoid-neonfma-rr2-p5-nr2fma.c",
2069 "src/math/sigmoid-neonfma-rr2-p5-nr2recps.c",
Marat Dukhan84000762020-06-29 18:38:43 -07002070 "src/math/sqrt-neonfma-nr1fma.c",
Marat Dukhan84000762020-06-29 18:38:43 -07002071 "src/math/sqrt-neonfma-nr1rsqrts1fma1adj.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002072 "src/math/sqrt-neonfma-nr2fma.c",
2073 "src/math/sqrt-neonfma-nr2fma1adj.c",
2074 "src/math/sqrt-neonfma-nr3fma.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002075]
2076
2077AARCH64_NEONFMA_UKERNELS = [
Marat Dukhan56b10cd2020-05-18 09:35:49 -07002078 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neonfma-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07002079 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neonfma-2x2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002080 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neonfma-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07002081 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neonfma-2x2.c",
Marat Dukhan56b10cd2020-05-18 09:35:49 -07002082 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neonfma-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07002083 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neonfma-2x2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002084 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neonfma-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07002085 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neonfma-2x2.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -07002086 "src/f32-conv-hwc2chw/3x3s2p1c3x4-neonfma-2x2.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07002087 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-1x4-acc2.c",
2088 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-1x4-acc3.c",
2089 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-1x4-acc4.c",
Marat Dukhan1268a242020-10-24 00:36:32 -07002090 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-1x4.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07002091 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-2x4-acc2.c",
Marat Dukhan1268a242020-10-24 00:36:32 -07002092 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-2x4.c",
2093 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-3x4.c",
2094 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-4x4.c",
2095 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-5x4.c",
2096 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-6x4.c",
Marat Dukhan82f0c322020-10-25 19:17:35 -07002097 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neonfma-1x4-acc2.c",
2098 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neonfma-1x4-acc3.c",
2099 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neonfma-1x4-acc4.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07002100 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neonfma-1x4.c",
Marat Dukhan82f0c322020-10-25 19:17:35 -07002101 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neonfma-2x4-acc2.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07002102 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neonfma-2x4.c",
2103 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neonfma-3x4.c",
2104 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neonfma-4x4.c",
Marat Dukhan149f0ea2020-10-26 12:50:33 -07002105 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-1x4-acc2.c",
2106 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-1x4-acc3.c",
2107 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-1x4-acc4.c",
2108 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-1x4-acc5.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07002109 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-1x4.c",
Marat Dukhan149f0ea2020-10-26 12:50:33 -07002110 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-2x4-acc2.c",
2111 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-2x4-acc3.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07002112 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-2x4.c",
Marat Dukhan149f0ea2020-10-26 12:50:33 -07002113 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-3x4-acc2.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07002114 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-3x4.c",
Marat Dukhan149f0ea2020-10-26 12:50:33 -07002115 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-4x4-acc2.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07002116 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-4x4.c",
2117 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-5x4.c",
Marat Dukhan30d4b252020-10-29 16:33:22 -07002118 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neonfma-1x4-acc2.c",
2119 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neonfma-1x4-acc3.c",
2120 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neonfma-1x4-acc4.c",
2121 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neonfma-1x4-acc5.c",
2122 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neonfma-1x4.c",
2123 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neonfma-2x4-acc2.c",
2124 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neonfma-2x4-acc3.c",
2125 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neonfma-2x4.c",
Marat Dukhan30d4b252020-10-29 16:33:22 -07002126 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neonfma-3x4-acc2.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07002127 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neonfma-3x4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002128 "src/f32-gemm/gen-inc/1x8inc-minmax-neonfma-lane-ld64.c",
2129 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-lane-ld64.c",
2130 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-lane-ld128.c",
2131 "src/f32-gemm/gen-inc/5x8inc-minmax-neonfma-lane-ld64.c",
2132 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-lane-ld64.c",
2133 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-lane-ld128.c",
2134 "src/f32-gemm/gen/1x8-minmax-neonfma-lane-ld64.c",
2135 "src/f32-gemm/gen/4x2-minmax-neonfma-lane-ld64.c",
2136 "src/f32-gemm/gen/4x8-minmax-neonfma-lane-ld64.c",
2137 "src/f32-gemm/gen/4x8-minmax-neonfma-lane-ld128.c",
2138 "src/f32-gemm/gen/5x8-minmax-neonfma-lane-ld64.c",
2139 "src/f32-gemm/gen/6x8-minmax-neonfma-lane-ld64.c",
2140 "src/f32-gemm/gen/6x8-minmax-neonfma-lane-ld128.c",
2141 "src/f32-igemm/gen/1x8-minmax-neonfma-lane-ld64.c",
2142 "src/f32-igemm/gen/4x2-minmax-neonfma-lane-ld64.c",
2143 "src/f32-igemm/gen/4x4-minmax-neonfma-lane-ld64.c",
2144 "src/f32-igemm/gen/4x8-minmax-neonfma-lane-ld64.c",
2145 "src/f32-igemm/gen/4x8-minmax-neonfma-lane-ld128.c",
2146 "src/f32-igemm/gen/6x8-minmax-neonfma-lane-ld64.c",
2147 "src/f32-igemm/gen/6x8-minmax-neonfma-lane-ld128.c",
Marat Dukhan355ab432020-04-09 19:01:52 -07002148 "src/f32-spmm/gen/4x2-minmax-neonfma.c",
2149 "src/f32-spmm/gen/4x4-minmax-neonfma.c",
Marat Dukhan355ab432020-04-09 19:01:52 -07002150 "src/f32-spmm/gen/8x2-minmax-neonfma.c",
2151 "src/f32-spmm/gen/8x4-minmax-neonfma.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002152 "src/f32-spmm/gen/12x2-minmax-neonfma.c",
2153 "src/f32-spmm/gen/12x4-minmax-neonfma.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002154 "src/f32-spmm/gen/16x2-minmax-neonfma.c",
2155 "src/f32-spmm/gen/16x4-minmax-neonfma.c",
Frank Barchard846c0c62020-10-26 15:01:39 -07002156 "src/f32-spmm/gen/32x2-minmax-neonfma.c",
2157 "src/f32-spmm/gen/32x4-minmax-neonfma.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002158 "src/f32-vbinary/gen/vdiv-minmax-neon-x4.c",
2159 "src/f32-vbinary/gen/vdiv-minmax-neon-x8.c",
2160 "src/f32-vbinary/gen/vdivc-minmax-neon-x4.c",
2161 "src/f32-vbinary/gen/vdivc-minmax-neon-x8.c",
2162 "src/f32-vbinary/gen/vrdivc-minmax-neon-x4.c",
2163 "src/f32-vbinary/gen/vrdivc-minmax-neon-x8.c",
Marat Dukhanb1eec082021-05-05 23:24:55 -07002164 "src/f32-vsigmoid/gen/vsigmoid-neonfma-rr1-lut64-p2-div-x4.c",
2165 "src/f32-vsigmoid/gen/vsigmoid-neonfma-rr1-lut64-p2-div-x8.c",
2166 "src/f32-vsigmoid/gen/vsigmoid-neonfma-rr1-lut64-p2-div-x12.c",
2167 "src/f32-vsigmoid/gen/vsigmoid-neonfma-rr1-lut64-p2-div-x16.c",
2168 "src/f32-vsigmoid/gen/vsigmoid-neonfma-rr1-lut64-p2-div-x20.c",
2169 "src/f32-vsigmoid/gen/vsigmoid-neonfma-rr1-lut64-p2-div-x24.c",
2170 "src/f32-vsigmoid/gen/vsigmoid-neonfma-rr1-lut2048-p1-div-x4.c",
2171 "src/f32-vsigmoid/gen/vsigmoid-neonfma-rr1-lut2048-p1-div-x8.c",
2172 "src/f32-vsigmoid/gen/vsigmoid-neonfma-rr1-lut2048-p1-div-x12.c",
2173 "src/f32-vsigmoid/gen/vsigmoid-neonfma-rr1-lut2048-p1-div-x16.c",
2174 "src/f32-vsigmoid/gen/vsigmoid-neonfma-rr1-lut2048-p1-div-x20.c",
2175 "src/f32-vsigmoid/gen/vsigmoid-neonfma-rr1-lut2048-p1-div-x24.c",
2176 "src/f32-vsigmoid/gen/vsigmoid-neonfma-rr1-p5-div-x4.c",
2177 "src/f32-vsigmoid/gen/vsigmoid-neonfma-rr1-p5-div-x8.c",
2178 "src/f32-vsigmoid/gen/vsigmoid-neonfma-rr1-p5-div-x12.c",
2179 "src/f32-vsigmoid/gen/vsigmoid-neonfma-rr1-p5-div-x16.c",
2180 "src/f32-vsigmoid/gen/vsigmoid-neonfma-rr1-p5-div-x20.c",
2181 "src/f32-vsigmoid/gen/vsigmoid-neonfma-rr1-p5-div-x24.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07002182 "src/f32-vsqrt/gen/neon-sqrt-x4.c",
2183 "src/f32-vsqrt/gen/neon-sqrt-x8.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08002184 "src/math/sigmoid-neonfma-rr1-lut64-p2-div.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002185 "src/math/sigmoid-neonfma-rr1-lut2048-p1-div.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08002186 "src/math/sigmoid-neonfma-rr1-p5-div.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08002187 "src/math/sigmoid-neonfma-rr2-lut64-p2-div.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002188 "src/math/sigmoid-neonfma-rr2-lut2048-p1-div.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08002189 "src/math/sigmoid-neonfma-rr2-p5-div.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002190]
2191
Marat Dukhan8853b822020-05-07 12:19:01 -07002192NEONV8_UKERNELS = [
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07002193 "src/f32-vrnd/gen/vrndd-neonv8-x4.c",
2194 "src/f32-vrnd/gen/vrndd-neonv8-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002195 "src/f32-vrnd/gen/vrndne-neonv8-x4.c",
2196 "src/f32-vrnd/gen/vrndne-neonv8-x8.c",
2197 "src/f32-vrnd/gen/vrndu-neonv8-x4.c",
2198 "src/f32-vrnd/gen/vrndu-neonv8-x8.c",
2199 "src/f32-vrnd/gen/vrndz-neonv8-x4.c",
2200 "src/f32-vrnd/gen/vrndz-neonv8-x8.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07002201 "src/math/roundd-neonv8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002202 "src/math/roundne-neonv8.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07002203 "src/math/roundu-neonv8.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07002204 "src/math/roundz-neonv8.c",
Marat Dukhan8853b822020-05-07 12:19:01 -07002205]
2206
Marat Dukhan08c4a432019-10-03 09:29:21 -07002207AARCH64_NEONFP16ARITH_UKERNELS = [
Frank Barchard5a599a62020-06-04 20:12:44 -07002208 "src/f16-dwconv/gen/up8x4-minmax-neonfp16arith-acc2.c",
2209 "src/f16-dwconv/gen/up8x4-minmax-neonfp16arith.c",
2210 "src/f16-dwconv/gen/up8x9-minmax-neonfp16arith-acc2.c",
2211 "src/f16-dwconv/gen/up8x9-minmax-neonfp16arith.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002212 "src/f16-dwconv/gen/up8x25-minmax-neonfp16arith-acc2.c",
2213 "src/f16-dwconv/gen/up8x25-minmax-neonfp16arith.c",
2214 "src/f16-dwconv/gen/up16x4-minmax-neonfp16arith-acc2.c",
2215 "src/f16-dwconv/gen/up16x4-minmax-neonfp16arith.c",
2216 "src/f16-dwconv/gen/up16x9-minmax-neonfp16arith-acc2.c",
2217 "src/f16-dwconv/gen/up16x9-minmax-neonfp16arith.c",
2218 "src/f16-dwconv/gen/up16x25-minmax-neonfp16arith-acc2.c",
2219 "src/f16-dwconv/gen/up16x25-minmax-neonfp16arith.c",
Frank Barchard0bb49a72020-06-04 11:35:11 -07002220 "src/f16-gavgpool/7p7x-minmax-neonfp16arith-c8.c",
2221 "src/f16-gavgpool/7x-minmax-neonfp16arith-c8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002222 "src/f16-gemm/gen-inc/1x8inc-minmax-neonfp16arith-ld64.c",
2223 "src/f16-gemm/gen-inc/1x16inc-minmax-neonfp16arith-ld64.c",
2224 "src/f16-gemm/gen-inc/4x8inc-minmax-neonfp16arith-ld64.c",
2225 "src/f16-gemm/gen-inc/4x16inc-minmax-neonfp16arith-ld64.c",
2226 "src/f16-gemm/gen-inc/6x8inc-minmax-neonfp16arith-ld64.c",
2227 "src/f16-gemm/gen-inc/6x16inc-minmax-neonfp16arith-ld64.c",
2228 "src/f16-gemm/gen-inc/8x8inc-minmax-neonfp16arith-ld64.c",
2229 "src/f16-gemm/gen-inc/8x16inc-minmax-neonfp16arith-ld64.c",
2230 "src/f16-gemm/gen/1x8-minmax-neonfp16arith-ld64.c",
2231 "src/f16-gemm/gen/1x16-minmax-neonfp16arith-ld64.c",
2232 "src/f16-gemm/gen/4x8-minmax-neonfp16arith-ld64.c",
2233 "src/f16-gemm/gen/4x16-minmax-neonfp16arith-ld64.c",
2234 "src/f16-gemm/gen/6x8-minmax-neonfp16arith-ld64.c",
2235 "src/f16-gemm/gen/6x16-minmax-neonfp16arith-ld64.c",
2236 "src/f16-gemm/gen/8x8-minmax-neonfp16arith-ld64.c",
2237 "src/f16-gemm/gen/8x16-minmax-neonfp16arith-ld64.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002238 "src/f16-igemm/gen/1x8-minmax-neonfp16arith-ld64.c",
2239 "src/f16-igemm/gen/1x16-minmax-neonfp16arith-ld64.c",
2240 "src/f16-igemm/gen/4x8-minmax-neonfp16arith-ld64.c",
2241 "src/f16-igemm/gen/4x16-minmax-neonfp16arith-ld64.c",
2242 "src/f16-igemm/gen/6x8-minmax-neonfp16arith-ld64.c",
2243 "src/f16-igemm/gen/6x16-minmax-neonfp16arith-ld64.c",
2244 "src/f16-igemm/gen/8x8-minmax-neonfp16arith-ld64.c",
2245 "src/f16-igemm/gen/8x16-minmax-neonfp16arith-ld64.c",
Frank Barchardb1966592020-05-12 13:47:06 -07002246 "src/f16-prelu/gen/neonfp16arith-2x8.c",
Frank Barcharda5316982020-07-23 13:19:28 -07002247 "src/f16-prelu/gen/neonfp16arith-2x16.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07002248 "src/f16-spmm/gen/8x1-minmax-neonfp16arith-x2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002249 "src/f16-spmm/gen/8x1-minmax-neonfp16arith.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07002250 "src/f16-spmm/gen/16x1-minmax-neonfp16arith-x2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002251 "src/f16-spmm/gen/16x1-minmax-neonfp16arith.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07002252 "src/f16-spmm/gen/24x1-minmax-neonfp16arith-x2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002253 "src/f16-spmm/gen/24x1-minmax-neonfp16arith.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07002254 "src/f16-spmm/gen/32x1-minmax-neonfp16arith-x2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002255 "src/f16-spmm/gen/32x1-minmax-neonfp16arith.c",
2256 "src/f16-vbinary/gen/vadd-minmax-neonfp16arith-x8.c",
2257 "src/f16-vbinary/gen/vadd-minmax-neonfp16arith-x16.c",
2258 "src/f16-vbinary/gen/vaddc-minmax-neonfp16arith-x8.c",
2259 "src/f16-vbinary/gen/vaddc-minmax-neonfp16arith-x16.c",
2260 "src/f16-vbinary/gen/vdiv-minmax-neonfp16arith-x8.c",
2261 "src/f16-vbinary/gen/vdiv-minmax-neonfp16arith-x16.c",
2262 "src/f16-vbinary/gen/vdivc-minmax-neonfp16arith-x8.c",
2263 "src/f16-vbinary/gen/vdivc-minmax-neonfp16arith-x16.c",
2264 "src/f16-vbinary/gen/vmax-neonfp16arith-x8.c",
2265 "src/f16-vbinary/gen/vmax-neonfp16arith-x16.c",
2266 "src/f16-vbinary/gen/vmaxc-neonfp16arith-x8.c",
2267 "src/f16-vbinary/gen/vmaxc-neonfp16arith-x16.c",
2268 "src/f16-vbinary/gen/vmin-neonfp16arith-x8.c",
2269 "src/f16-vbinary/gen/vmin-neonfp16arith-x16.c",
2270 "src/f16-vbinary/gen/vminc-neonfp16arith-x8.c",
2271 "src/f16-vbinary/gen/vminc-neonfp16arith-x16.c",
2272 "src/f16-vbinary/gen/vmul-minmax-neonfp16arith-x8.c",
2273 "src/f16-vbinary/gen/vmul-minmax-neonfp16arith-x16.c",
2274 "src/f16-vbinary/gen/vmulc-minmax-neonfp16arith-x8.c",
2275 "src/f16-vbinary/gen/vmulc-minmax-neonfp16arith-x16.c",
2276 "src/f16-vbinary/gen/vrdivc-minmax-neonfp16arith-x8.c",
2277 "src/f16-vbinary/gen/vrdivc-minmax-neonfp16arith-x16.c",
2278 "src/f16-vbinary/gen/vrsubc-minmax-neonfp16arith-x8.c",
2279 "src/f16-vbinary/gen/vrsubc-minmax-neonfp16arith-x16.c",
2280 "src/f16-vbinary/gen/vsub-minmax-neonfp16arith-x8.c",
2281 "src/f16-vbinary/gen/vsub-minmax-neonfp16arith-x16.c",
2282 "src/f16-vbinary/gen/vsubc-minmax-neonfp16arith-x8.c",
2283 "src/f16-vbinary/gen/vsubc-minmax-neonfp16arith-x16.c",
Marat Dukhanb1eec082021-05-05 23:24:55 -07002284 "src/f16-vclamp/gen/vclamp-neonfp16arith-x8.c",
2285 "src/f16-vclamp/gen/vclamp-neonfp16arith-x16.c",
Marat Dukhan6674d692021-05-05 22:27:00 -07002286 "src/f16-vhswish/gen/vhswish-neonfp16arith-x8.c",
2287 "src/f16-vhswish/gen/vhswish-neonfp16arith-x16.c",
Marat Dukhanb1eec082021-05-05 23:24:55 -07002288 "src/f16-vrelu/gen/vrelu-neonfp16arith-x8.c",
2289 "src/f16-vrelu/gen/vrelu-neonfp16arith-x16.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002290 "src/f16-vmulcaddc/gen/c8-minmax-neonfp16arith-2x.c",
2291 "src/f16-vmulcaddc/gen/c16-minmax-neonfp16arith-2x.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002292]
2293
Benoit Jacoba9644732020-08-13 12:48:55 -07002294NEONDOT_UKERNELS = [
Marat Dukhand4c83032020-09-09 23:20:12 -07002295 "src/qs8-gemm/gen/1x8c4-minmax-neondot.c",
Marat Dukhand4c83032020-09-09 23:20:12 -07002296 "src/qs8-gemm/gen/1x16c4-minmax-neondot.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002297 "src/qs8-gemm/gen/4x8c4-minmax-neondot.c",
Benoit Jacoba9644732020-08-13 12:48:55 -07002298 "src/qs8-gemm/gen/4x16c4-minmax-neondot.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002299 "src/qs8-gemm/gen/6x8c4-minmax-neondot.c",
Frank Barchard2fa17452020-09-22 09:51:46 -07002300 "src/qs8-gemm/gen/6x16c4-minmax-neondot.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002301 "src/qs8-gemm/gen/8x8c4-minmax-neondot.c",
Frank Barcharda48848f2020-09-24 00:59:24 -07002302 "src/qs8-gemm/gen/8x16c4-minmax-neondot.c",
Marat Dukhand4c83032020-09-09 23:20:12 -07002303 "src/qs8-igemm/gen/1x8c4-minmax-neondot.c",
Marat Dukhand4c83032020-09-09 23:20:12 -07002304 "src/qs8-igemm/gen/1x16c4-minmax-neondot.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002305 "src/qs8-igemm/gen/4x8c4-minmax-neondot.c",
Benoit Jacoba05487f2020-08-14 13:37:21 -07002306 "src/qs8-igemm/gen/4x16c4-minmax-neondot.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002307 "src/qs8-igemm/gen/6x8c4-minmax-neondot.c",
Frank Barchard2fa17452020-09-22 09:51:46 -07002308 "src/qs8-igemm/gen/6x16c4-minmax-neondot.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002309 "src/qs8-igemm/gen/8x8c4-minmax-neondot.c",
Frank Barcharda48848f2020-09-24 00:59:24 -07002310 "src/qs8-igemm/gen/8x16c4-minmax-neondot.c",
Benoit Jacoba9644732020-08-13 12:48:55 -07002311]
2312
Marat Dukhan08c4a432019-10-03 09:29:21 -07002313SSE_UKERNELS = [
Marat Dukhan99936602020-04-11 16:47:01 -07002314 "src/f32-avgpool/9p8x-minmax-sse-c4.c",
2315 "src/f32-avgpool/9x-minmax-sse-c4.c",
Erich Elsenb1233402020-06-08 15:53:15 -07002316 "src/f32-conv-hwc2chw/3x3s2p1c3x4-sse-1x1.c",
2317 "src/f32-conv-hwc2chw/3x3s2p1c3x4-sse-2x2.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002318 "src/f32-dwconv/gen/up4x4-minmax-sse-acc2.c",
2319 "src/f32-dwconv/gen/up4x4-minmax-sse.c",
2320 "src/f32-dwconv/gen/up4x9-minmax-sse-acc2.c",
2321 "src/f32-dwconv/gen/up4x9-minmax-sse.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002322 "src/f32-dwconv/gen/up4x25-minmax-sse-acc2.c",
2323 "src/f32-dwconv/gen/up4x25-minmax-sse.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002324 "src/f32-dwconv/gen/up8x4-minmax-sse-acc2.c",
2325 "src/f32-dwconv/gen/up8x4-minmax-sse.c",
2326 "src/f32-dwconv/gen/up8x9-minmax-sse-acc2.c",
2327 "src/f32-dwconv/gen/up8x9-minmax-sse.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002328 "src/f32-dwconv/gen/up8x25-minmax-sse-acc2.c",
2329 "src/f32-dwconv/gen/up8x25-minmax-sse.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07002330 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-1x4-acc2.c",
2331 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-1x4-acc3.c",
2332 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-1x4-acc4.c",
Marat Dukhan470078a2020-10-23 22:36:52 -07002333 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-1x4.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07002334 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-2x4-acc2.c",
Marat Dukhan470078a2020-10-23 22:36:52 -07002335 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-2x4.c",
2336 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-3x4.c",
2337 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-4x4.c",
2338 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-5x4.c",
2339 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-6x4.c",
Marat Dukhan0ff97182020-10-25 19:14:03 -07002340 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-sse-1x4-acc2.c",
2341 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-sse-1x4-acc3.c",
2342 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-sse-1x4-acc4.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07002343 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-sse-1x4.c",
Marat Dukhan0ff97182020-10-25 19:14:03 -07002344 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-sse-2x4-acc2.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07002345 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-sse-2x4.c",
2346 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-sse-3x4.c",
2347 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-sse-4x4.c",
Marat Dukhand0503892020-10-30 08:22:04 -07002348 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-1x4-acc2.c",
2349 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-1x4-acc3.c",
2350 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-1x4-acc4.c",
2351 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-1x4-acc5.c",
2352 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-1x4.c",
2353 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-2x4-acc2.c",
2354 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-2x4-acc3.c",
2355 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-2x4.c",
2356 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-3x4-acc2.c",
2357 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-3x4.c",
2358 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-4x4-acc2.c",
2359 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-4x4.c",
2360 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-5x4.c",
Marat Dukhanccca2142020-10-30 17:32:45 -07002361 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-sse-1x4-acc2.c",
2362 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-sse-1x4-acc3.c",
2363 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-sse-1x4-acc4.c",
2364 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-sse-1x4-acc5.c",
2365 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-sse-1x4.c",
2366 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-sse-2x4-acc2.c",
2367 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-sse-2x4-acc3.c",
2368 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-sse-2x4.c",
Marat Dukhanccca2142020-10-30 17:32:45 -07002369 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-sse-3x4-acc2.c",
Frank Barchard8ef44cd2020-11-03 12:30:23 -08002370 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-sse-3x4.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -07002371 "src/f32-gavgpool-cw/sse-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -07002372 "src/f32-gavgpool/7p7x-minmax-sse-c4.c",
2373 "src/f32-gavgpool/7x-minmax-sse-c4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002374 "src/f32-gemm/gen-inc/1x8inc-minmax-sse-dup.c",
2375 "src/f32-gemm/gen-inc/1x8inc-minmax-sse-load1.c",
2376 "src/f32-gemm/gen-inc/1x8s4inc-minmax-sse.c",
Marat Dukhan802fcae2020-12-11 14:37:25 -08002377 "src/f32-gemm/gen-inc/3x8inc-minmax-sse-dup.c",
2378 "src/f32-gemm/gen-inc/3x8inc-minmax-sse-load1.c",
2379 "src/f32-gemm/gen-inc/3x8s4inc-minmax-sse.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002380 "src/f32-gemm/gen-inc/4x8inc-minmax-sse-dup.c",
2381 "src/f32-gemm/gen-inc/4x8inc-minmax-sse-load1.c",
2382 "src/f32-gemm/gen-inc/4x8s4inc-minmax-sse.c",
Marat Dukhan802fcae2020-12-11 14:37:25 -08002383 "src/f32-gemm/gen-inc/5x8inc-minmax-sse-dup.c",
2384 "src/f32-gemm/gen-inc/5x8inc-minmax-sse-load1.c",
2385 "src/f32-gemm/gen-inc/5x8s4inc-minmax-sse.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002386 "src/f32-gemm/gen/1x8-minmax-sse-dup.c",
2387 "src/f32-gemm/gen/1x8-minmax-sse-load1.c",
2388 "src/f32-gemm/gen/1x8s4-minmax-sse.c",
Marat Dukhan802fcae2020-12-11 14:37:25 -08002389 "src/f32-gemm/gen/3x8-minmax-sse-dup.c",
2390 "src/f32-gemm/gen/3x8-minmax-sse-load1.c",
2391 "src/f32-gemm/gen/3x8s4-minmax-sse.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002392 "src/f32-gemm/gen/4x2c4-minmax-sse.c",
2393 "src/f32-gemm/gen/4x8-minmax-sse-dup.c",
2394 "src/f32-gemm/gen/4x8-minmax-sse-load1.c",
2395 "src/f32-gemm/gen/4x8s4-minmax-sse.c",
Marat Dukhan802fcae2020-12-11 14:37:25 -08002396 "src/f32-gemm/gen/5x8-minmax-sse-dup.c",
2397 "src/f32-gemm/gen/5x8-minmax-sse-load1.c",
2398 "src/f32-gemm/gen/5x8s4-minmax-sse.c",
Artsiom Ablavatskib3ffd582021-03-31 13:00:08 -07002399 "src/f32-ibilinear-chw/gen/sse-p4.c",
2400 "src/f32-ibilinear-chw/gen/sse-p8.c",
Frank Barchard4a352042021-04-13 15:52:08 -07002401 "src/f32-ibilinear/gen/sse-c4.c",
2402 "src/f32-ibilinear/gen/sse-c8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002403 "src/f32-igemm/gen/1x8-minmax-sse-dup.c",
2404 "src/f32-igemm/gen/1x8-minmax-sse-load1.c",
2405 "src/f32-igemm/gen/1x8s4-minmax-sse.c",
Marat Dukhan802fcae2020-12-11 14:37:25 -08002406 "src/f32-igemm/gen/3x8-minmax-sse-dup.c",
2407 "src/f32-igemm/gen/3x8-minmax-sse-load1.c",
2408 "src/f32-igemm/gen/3x8s4-minmax-sse.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002409 "src/f32-igemm/gen/4x2c4-minmax-sse.c",
2410 "src/f32-igemm/gen/4x8-minmax-sse-dup.c",
2411 "src/f32-igemm/gen/4x8-minmax-sse-load1.c",
2412 "src/f32-igemm/gen/4x8s4-minmax-sse.c",
Marat Dukhan802fcae2020-12-11 14:37:25 -08002413 "src/f32-igemm/gen/5x8-minmax-sse-dup.c",
2414 "src/f32-igemm/gen/5x8-minmax-sse-load1.c",
2415 "src/f32-igemm/gen/5x8s4-minmax-sse.c",
Marat Dukhan99936602020-04-11 16:47:01 -07002416 "src/f32-maxpool/9p8x-minmax-sse-c4.c",
2417 "src/f32-pavgpool/9p8x-minmax-sse-c4.c",
2418 "src/f32-pavgpool/9x-minmax-sse-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002419 "src/f32-ppmm/gen/4x8-minmax-sse.c",
Marat Dukhan39b5e942020-06-24 15:03:48 -07002420 "src/f32-prelu/gen/sse-2x4.c",
2421 "src/f32-prelu/gen/sse-2x8.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002422 "src/f32-rmax/sse.c",
Marat Dukhan355ab432020-04-09 19:01:52 -07002423 "src/f32-spmm/gen/4x1-minmax-sse.c",
2424 "src/f32-spmm/gen/8x1-minmax-sse.c",
Erich Elsen6e80fdc2020-06-09 15:35:37 -07002425 "src/f32-spmm/gen/16x1-minmax-sse.c",
Frank Barchard846c0c62020-10-26 15:01:39 -07002426 "src/f32-spmm/gen/32x1-minmax-sse.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002427 "src/f32-vbinary/gen/vadd-minmax-sse-x4.c",
2428 "src/f32-vbinary/gen/vadd-minmax-sse-x8.c",
2429 "src/f32-vbinary/gen/vaddc-minmax-sse-x4.c",
2430 "src/f32-vbinary/gen/vaddc-minmax-sse-x8.c",
2431 "src/f32-vbinary/gen/vdiv-minmax-sse-x4.c",
2432 "src/f32-vbinary/gen/vdiv-minmax-sse-x8.c",
2433 "src/f32-vbinary/gen/vdivc-minmax-sse-x4.c",
2434 "src/f32-vbinary/gen/vdivc-minmax-sse-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -08002435 "src/f32-vbinary/gen/vmax-sse-x4.c",
2436 "src/f32-vbinary/gen/vmax-sse-x8.c",
2437 "src/f32-vbinary/gen/vmaxc-sse-x4.c",
2438 "src/f32-vbinary/gen/vmaxc-sse-x8.c",
2439 "src/f32-vbinary/gen/vmin-sse-x4.c",
2440 "src/f32-vbinary/gen/vmin-sse-x8.c",
2441 "src/f32-vbinary/gen/vminc-sse-x4.c",
2442 "src/f32-vbinary/gen/vminc-sse-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002443 "src/f32-vbinary/gen/vmul-minmax-sse-x4.c",
2444 "src/f32-vbinary/gen/vmul-minmax-sse-x8.c",
2445 "src/f32-vbinary/gen/vmulc-minmax-sse-x4.c",
2446 "src/f32-vbinary/gen/vmulc-minmax-sse-x8.c",
2447 "src/f32-vbinary/gen/vrdivc-minmax-sse-x4.c",
2448 "src/f32-vbinary/gen/vrdivc-minmax-sse-x8.c",
2449 "src/f32-vbinary/gen/vrsubc-minmax-sse-x4.c",
2450 "src/f32-vbinary/gen/vrsubc-minmax-sse-x8.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07002451 "src/f32-vbinary/gen/vsqrdiff-sse-x4.c",
2452 "src/f32-vbinary/gen/vsqrdiff-sse-x8.c",
2453 "src/f32-vbinary/gen/vsqrdiffc-sse-x4.c",
2454 "src/f32-vbinary/gen/vsqrdiffc-sse-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002455 "src/f32-vbinary/gen/vsub-minmax-sse-x4.c",
2456 "src/f32-vbinary/gen/vsub-minmax-sse-x8.c",
2457 "src/f32-vbinary/gen/vsubc-minmax-sse-x4.c",
2458 "src/f32-vbinary/gen/vsubc-minmax-sse-x8.c",
Marat Dukhanb1eec082021-05-05 23:24:55 -07002459 "src/f32-vclamp/gen/vclamp-sse-x4.c",
2460 "src/f32-vclamp/gen/vclamp-sse-x8.c",
Marat Dukhan6674d692021-05-05 22:27:00 -07002461 "src/f32-vhswish/gen/vhswish-sse-x4.c",
2462 "src/f32-vhswish/gen/vhswish-sse-x8.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07002463 "src/f32-vlrelu/gen/vlrelu-sse-x4.c",
2464 "src/f32-vlrelu/gen/vlrelu-sse-x8.c",
Marat Dukhan99936602020-04-11 16:47:01 -07002465 "src/f32-vmulcaddc/gen/c4-minmax-sse-2x.c",
2466 "src/f32-vmulcaddc/gen/c8-minmax-sse-2x.c",
Marat Dukhanb1eec082021-05-05 23:24:55 -07002467 "src/f32-vrelu/gen/vrelu-sse-x4.c",
2468 "src/f32-vrelu/gen/vrelu-sse-x8.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07002469 "src/f32-vsqrt/gen/sse-sqrt-x4.c",
2470 "src/f32-vsqrt/gen/sse-sqrt-x8.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07002471 "src/f32-vunary/gen/vabs-sse-x4.c",
2472 "src/f32-vunary/gen/vabs-sse-x8.c",
2473 "src/f32-vunary/gen/vneg-sse-x4.c",
2474 "src/f32-vunary/gen/vneg-sse-x8.c",
2475 "src/f32-vunary/gen/vsqr-sse-x4.c",
2476 "src/f32-vunary/gen/vsqr-sse-x8.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07002477 "src/math/roundd-sse-addsub.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002478 "src/math/roundne-sse-addsub.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07002479 "src/math/roundu-sse-addsub.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07002480 "src/math/roundz-sse-addsub.c",
Marat Dukhan84000762020-06-29 18:38:43 -07002481 "src/math/sqrt-sse-hh1mac.c",
2482 "src/math/sqrt-sse-nr1mac.c",
2483 "src/math/sqrt-sse-nr2mac.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002484 "src/x32-fill/sse.c",
2485 "src/x32-packx/x4-sse.c",
2486 "src/x32-pad/sse.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002487]
2488
2489SSE2_UKERNELS = [
Marat Dukhan329da642019-11-19 21:44:39 -08002490 "src/f32-argmaxpool/4x-sse2-c4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002491 "src/f32-argmaxpool/9p8x-sse2-c4.c",
Marat Dukhan329da642019-11-19 21:44:39 -08002492 "src/f32-argmaxpool/9x-sse2-c4.c",
Marat Dukhan802fcae2020-12-11 14:37:25 -08002493 "src/f32-gemm/gen-inc/1x8inc-minmax-sse2-dup.c",
2494 "src/f32-gemm/gen-inc/3x8inc-minmax-sse2-dup.c",
2495 "src/f32-gemm/gen-inc/4x8inc-minmax-sse2-dup.c",
2496 "src/f32-gemm/gen-inc/5x8inc-minmax-sse2-dup.c",
2497 "src/f32-gemm/gen/1x8-minmax-sse2-dup.c",
2498 "src/f32-gemm/gen/3x8-minmax-sse2-dup.c",
2499 "src/f32-gemm/gen/4x8-minmax-sse2-dup.c",
2500 "src/f32-gemm/gen/5x8-minmax-sse2-dup.c",
2501 "src/f32-igemm/gen/1x8-minmax-sse2-dup.c",
2502 "src/f32-igemm/gen/3x8-minmax-sse2-dup.c",
2503 "src/f32-igemm/gen/4x8-minmax-sse2-dup.c",
2504 "src/f32-igemm/gen/5x8-minmax-sse2-dup.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08002505 "src/f32-prelu/gen/sse2-2x4.c",
2506 "src/f32-prelu/gen/sse2-2x8.c",
Marat Dukhanb39689d2020-01-24 13:32:20 -08002507 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x4.c",
Marat Dukhanb39689d2020-01-24 13:32:20 -08002508 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x8-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002509 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x8.c",
Marat Dukhanb39689d2020-01-24 13:32:20 -08002510 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12-acc2.c",
2511 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002512 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12.c",
Marat Dukhanb39689d2020-01-24 13:32:20 -08002513 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16-acc2.c",
2514 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002515 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16.c",
Marat Dukhanb39689d2020-01-24 13:32:20 -08002516 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20-acc2.c",
2517 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002518 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20.c",
Marat Dukhaned6baaf2020-12-01 15:07:08 -08002519 "src/f32-velu/gen/velu-sse2-rr2-lut16-p3-x4.c",
2520 "src/f32-velu/gen/velu-sse2-rr2-lut16-p3-x8.c",
2521 "src/f32-velu/gen/velu-sse2-rr2-lut16-p3-x12.c",
2522 "src/f32-velu/gen/velu-sse2-rr2-lut16-p3-x16.c",
2523 "src/f32-velu/gen/velu-sse2-rr2-lut16-p3-x20.c",
2524 "src/f32-velu/gen/velu-sse2-rr2-lut16-p3-x24.c",
2525 "src/f32-velu/gen/velu-sse2-rr2-p6-x4.c",
2526 "src/f32-velu/gen/velu-sse2-rr2-p6-x8.c",
2527 "src/f32-velu/gen/velu-sse2-rr2-p6-x12.c",
2528 "src/f32-velu/gen/velu-sse2-rr2-p6-x16.c",
2529 "src/f32-velu/gen/velu-sse2-rr2-p6-x20.c",
2530 "src/f32-velu/gen/velu-sse2-rr2-p6-x24.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07002531 "src/f32-vlrelu/gen/vlrelu-sse2-x4.c",
2532 "src/f32-vlrelu/gen/vlrelu-sse2-x8.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07002533 "src/f32-vrnd/gen/vrndd-sse2-x4.c",
2534 "src/f32-vrnd/gen/vrndd-sse2-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002535 "src/f32-vrnd/gen/vrndne-sse2-x4.c",
2536 "src/f32-vrnd/gen/vrndne-sse2-x8.c",
2537 "src/f32-vrnd/gen/vrndu-sse2-x4.c",
2538 "src/f32-vrnd/gen/vrndu-sse2-x8.c",
2539 "src/f32-vrnd/gen/vrndz-sse2-x4.c",
2540 "src/f32-vrnd/gen/vrndz-sse2-x8.c",
Marat Dukhanb1eec082021-05-05 23:24:55 -07002541 "src/f32-vsigmoid/gen/vsigmoid-sse2-lut64-p2-div-x4.c",
2542 "src/f32-vsigmoid/gen/vsigmoid-sse2-lut64-p2-div-x8.c",
2543 "src/f32-vsigmoid/gen/vsigmoid-sse2-lut64-p2-div-x12.c",
2544 "src/f32-vsigmoid/gen/vsigmoid-sse2-lut64-p2-div-x16.c",
2545 "src/f32-vsigmoid/gen/vsigmoid-sse2-lut64-p2-div-x20.c",
2546 "src/f32-vsigmoid/gen/vsigmoid-sse2-lut64-p2-div-x24.c",
2547 "src/f32-vsigmoid/gen/vsigmoid-sse2-p5-div-x4.c",
2548 "src/f32-vsigmoid/gen/vsigmoid-sse2-p5-div-x8.c",
2549 "src/f32-vsigmoid/gen/vsigmoid-sse2-p5-div-x12.c",
2550 "src/f32-vsigmoid/gen/vsigmoid-sse2-p5-div-x16.c",
2551 "src/f32-vsigmoid/gen/vsigmoid-sse2-p5-div-x20.c",
2552 "src/f32-vsigmoid/gen/vsigmoid-sse2-p5-div-x24.c",
Marat Dukhanb7633f22020-11-20 16:34:56 -08002553 "src/math/exp-sse2-rr2-lut64-p2.c",
2554 "src/math/exp-sse2-rr2-p5.c",
Marat Dukhande390d42020-11-29 19:32:18 -08002555 "src/math/expm1minus-sse2-rr2-lut16-p3.c",
Marat Dukhana438aca2020-11-20 15:45:01 -08002556 "src/math/expm1minus-sse2-rr2-p6.c",
Frank Barchard3b800452020-11-22 12:12:35 -08002557 "src/math/expminus-sse2-rr2-p5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002558 "src/math/roundd-sse2-cvt.c",
2559 "src/math/roundne-sse2-cvt.c",
2560 "src/math/roundu-sse2-cvt.c",
2561 "src/math/roundz-sse2-cvt.c",
2562 "src/math/sigmoid-sse2-rr2-lut64-p2-div.c",
2563 "src/math/sigmoid-sse2-rr2-lut64-p2-nr1.c",
2564 "src/math/sigmoid-sse2-rr2-lut64-p2-nr2.c",
2565 "src/math/sigmoid-sse2-rr2-p5-div.c",
2566 "src/math/sigmoid-sse2-rr2-p5-nr1.c",
2567 "src/math/sigmoid-sse2-rr2-p5-nr2.c",
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07002568 "src/qs8-dwconv/gen/up8x9-minmax-sse2-mul16.c",
2569 "src/qs8-dwconv/gen/up16x9-minmax-sse2-mul16.c",
2570 "src/qs8-dwconv/gen/up24x9-minmax-sse2-mul16.c",
Marat Dukhan159688f2020-08-06 10:34:29 -07002571 "src/qs8-gavgpool/gen/7p7x-minmax-sse2-c8-acc2.c",
2572 "src/qs8-gavgpool/gen/7p7x-minmax-sse2-c16-acc2.c",
2573 "src/qs8-gavgpool/gen/7p7x-minmax-sse2-c24-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002574 "src/qs8-gavgpool/gen/7x-minmax-sse2-c8-acc2.c",
2575 "src/qs8-gavgpool/gen/7x-minmax-sse2-c16-acc2.c",
2576 "src/qs8-gavgpool/gen/7x-minmax-sse2-c24-acc2.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002577 "src/qs8-gemm/gen/1x4c2-minmax-sse2-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002578 "src/qs8-gemm/gen/1x4c2-minmax-sse2-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002579 "src/qs8-gemm/gen/1x4c2-xw-minmax-sse2.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002580 "src/qs8-gemm/gen/1x4c8-minmax-sse2-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002581 "src/qs8-gemm/gen/1x4c8-minmax-sse2-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002582 "src/qs8-gemm/gen/1x4c8-xw-minmax-sse2.c",
Marat Dukhan2e427872021-04-09 23:40:07 -07002583 "src/qs8-gemm/gen/2x4c2-minmax-sse2-ld64.c",
2584 "src/qs8-gemm/gen/2x4c2-minmax-sse2-ld128.c",
2585 "src/qs8-gemm/gen/2x4c2-xw-minmax-sse2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002586 "src/qs8-gemm/gen/2x4c8-minmax-sse2-ld64.c",
2587 "src/qs8-gemm/gen/2x4c8-minmax-sse2-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002588 "src/qs8-gemm/gen/2x4c8-xw-minmax-sse2.c",
Marat Dukhan2e427872021-04-09 23:40:07 -07002589 "src/qs8-gemm/gen/3x4c2-minmax-sse2-ld64.c",
2590 "src/qs8-gemm/gen/3x4c2-minmax-sse2-ld128.c",
2591 "src/qs8-gemm/gen/3x4c2-xw-minmax-sse2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002592 "src/qs8-gemm/gen/3x4c8-minmax-sse2-ld64.c",
2593 "src/qs8-gemm/gen/3x4c8-minmax-sse2-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002594 "src/qs8-gemm/gen/3x4c8-xw-minmax-sse2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002595 "src/qs8-gemm/gen/4x4c2-minmax-sse2-ld64.c",
2596 "src/qs8-gemm/gen/4x4c2-minmax-sse2-ld128.c",
2597 "src/qs8-gemm/gen/4x4c2-xw-minmax-sse2.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002598 "src/qs8-igemm/gen/1x4c2-minmax-sse2-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002599 "src/qs8-igemm/gen/1x4c2-minmax-sse2-ld128.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002600 "src/qs8-igemm/gen/1x4c8-minmax-sse2-ld64.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002601 "src/qs8-igemm/gen/1x4c8-minmax-sse2-ld128.c",
Marat Dukhan2e427872021-04-09 23:40:07 -07002602 "src/qs8-igemm/gen/2x4c2-minmax-sse2-ld64.c",
2603 "src/qs8-igemm/gen/2x4c2-minmax-sse2-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002604 "src/qs8-igemm/gen/2x4c8-minmax-sse2-ld64.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002605 "src/qs8-igemm/gen/2x4c8-minmax-sse2-ld128.c",
Marat Dukhan2e427872021-04-09 23:40:07 -07002606 "src/qs8-igemm/gen/3x4c2-minmax-sse2-ld64.c",
2607 "src/qs8-igemm/gen/3x4c2-minmax-sse2-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002608 "src/qs8-igemm/gen/3x4c8-minmax-sse2-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002609 "src/qs8-igemm/gen/3x4c8-minmax-sse2-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002610 "src/qs8-igemm/gen/4x4c2-minmax-sse2-ld64.c",
2611 "src/qs8-igemm/gen/4x4c2-minmax-sse2-ld128.c",
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07002612 "src/qs8-requantization/fp32-sse2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002613 "src/qs8-requantization/precise-sse2.c",
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07002614 "src/qs8-requantization/q31-sse2.c",
Marat Dukhand9f3ad42020-08-10 12:30:58 -07002615 "src/qs8-vadd/gen/minmax-sse2-mul16-ld64-x8.c",
2616 "src/qs8-vadd/gen/minmax-sse2-mul16-ld64-x16.c",
2617 "src/qs8-vadd/gen/minmax-sse2-mul16-ld64-x24.c",
2618 "src/qs8-vadd/gen/minmax-sse2-mul16-ld64-x32.c",
Marat Dukhan0270d9f2020-08-11 00:56:46 -07002619 "src/qs8-vaddc/gen/minmax-sse2-mul16-ld64-x8.c",
2620 "src/qs8-vaddc/gen/minmax-sse2-mul16-ld64-x16.c",
2621 "src/qs8-vaddc/gen/minmax-sse2-mul16-ld64-x24.c",
2622 "src/qs8-vaddc/gen/minmax-sse2-mul16-ld64-x32.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -07002623 "src/qu8-avgpool/9p8x-minmax-sse2-c8.c",
2624 "src/qu8-avgpool/9x-minmax-sse2-c8.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -07002625 "src/qu8-dwconv/up8x9-minmax-sse2.c",
2626 "src/qu8-gavgpool/7p7x-minmax-sse2-c8.c",
2627 "src/qu8-gavgpool/7x-minmax-sse2-c8.c",
2628 "src/qu8-gemm/2x4c8-minmax-sse2.c",
2629 "src/qu8-gemm/4x4c2-minmax-sse2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002630 "src/qu8-igemm/4x4c2-minmax-sse2.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07002631 "src/qu8-requantization/fp32-sse2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002632 "src/qu8-requantization/precise-sse2.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07002633 "src/qu8-requantization/q31-sse2.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -07002634 "src/qu8-vadd/minmax-sse2.c",
Marat Dukhan99936602020-04-11 16:47:01 -07002635 "src/u8-maxpool/9p8x-minmax-sse2-c16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002636 "src/u8-rmax/sse2.c",
Marat Dukhan6674d692021-05-05 22:27:00 -07002637 "src/u8-vclamp/sse2-x64.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002638 "src/x8-zip/x2-sse2.c",
2639 "src/x8-zip/x3-sse2.c",
2640 "src/x8-zip/x4-sse2.c",
2641 "src/x8-zip/xm-sse2.c",
Marat Dukhan57dccd82020-04-14 00:53:10 -07002642 "src/x32-unpool/sse2.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002643 "src/x32-zip/x2-sse2.c",
2644 "src/x32-zip/x3-sse2.c",
2645 "src/x32-zip/x4-sse2.c",
2646 "src/x32-zip/xm-sse2.c",
Marat Dukhanfe7acb62020-03-09 19:30:05 -07002647]
2648
2649SSSE3_UKERNELS = [
Frank Barchard35db7d02020-10-26 13:37:34 -07002650 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-1x4-acc2.c",
2651 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-1x4-acc3.c",
2652 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-1x4-acc4.c",
Marat Dukhan98f2eeb2020-10-23 23:13:41 -07002653 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-1x4.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07002654 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-2x4-acc2.c",
Marat Dukhan98f2eeb2020-10-23 23:13:41 -07002655 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-2x4.c",
2656 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-3x4.c",
2657 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-4x4.c",
2658 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-5x4.c",
2659 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-6x4.c",
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07002660 "src/qs8-dwconv/gen/up8x9-minmax-ssse3-mul16.c",
2661 "src/qs8-dwconv/gen/up16x9-minmax-ssse3-mul16.c",
2662 "src/qs8-dwconv/gen/up24x9-minmax-ssse3-mul16.c",
Marat Dukhan159688f2020-08-06 10:34:29 -07002663 "src/qs8-gavgpool/gen/7p7x-minmax-ssse3-c8-acc2.c",
2664 "src/qs8-gavgpool/gen/7p7x-minmax-ssse3-c16-acc2.c",
2665 "src/qs8-gavgpool/gen/7p7x-minmax-ssse3-c24-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002666 "src/qs8-gavgpool/gen/7x-minmax-ssse3-c8-acc2.c",
2667 "src/qs8-gavgpool/gen/7x-minmax-ssse3-c16-acc2.c",
2668 "src/qs8-gavgpool/gen/7x-minmax-ssse3-c24-acc2.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002669 "src/qs8-gemm/gen/1x4c2-minmax-ssse3-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002670 "src/qs8-gemm/gen/1x4c2-minmax-ssse3-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002671 "src/qs8-gemm/gen/1x4c2-xw-minmax-ssse3.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002672 "src/qs8-gemm/gen/1x4c8-minmax-ssse3-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002673 "src/qs8-gemm/gen/1x4c8-minmax-ssse3-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002674 "src/qs8-gemm/gen/1x4c8-xw-minmax-ssse3.c",
Marat Dukhan2e427872021-04-09 23:40:07 -07002675 "src/qs8-gemm/gen/2x4c2-minmax-ssse3-ld64.c",
2676 "src/qs8-gemm/gen/2x4c2-minmax-ssse3-ld128.c",
2677 "src/qs8-gemm/gen/2x4c2-xw-minmax-ssse3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002678 "src/qs8-gemm/gen/2x4c8-minmax-ssse3-ld64.c",
2679 "src/qs8-gemm/gen/2x4c8-minmax-ssse3-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002680 "src/qs8-gemm/gen/2x4c8-xw-minmax-ssse3.c",
Marat Dukhan2e427872021-04-09 23:40:07 -07002681 "src/qs8-gemm/gen/3x4c2-minmax-ssse3-ld64.c",
2682 "src/qs8-gemm/gen/3x4c2-minmax-ssse3-ld128.c",
2683 "src/qs8-gemm/gen/3x4c2-xw-minmax-ssse3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002684 "src/qs8-gemm/gen/3x4c8-minmax-ssse3-ld64.c",
2685 "src/qs8-gemm/gen/3x4c8-minmax-ssse3-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002686 "src/qs8-gemm/gen/3x4c8-xw-minmax-ssse3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002687 "src/qs8-gemm/gen/4x4c2-minmax-ssse3-ld64.c",
2688 "src/qs8-gemm/gen/4x4c2-minmax-ssse3-ld128.c",
2689 "src/qs8-gemm/gen/4x4c2-xw-minmax-ssse3.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002690 "src/qs8-igemm/gen/1x4c2-minmax-ssse3-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002691 "src/qs8-igemm/gen/1x4c2-minmax-ssse3-ld128.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002692 "src/qs8-igemm/gen/1x4c8-minmax-ssse3-ld64.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002693 "src/qs8-igemm/gen/1x4c8-minmax-ssse3-ld128.c",
Marat Dukhan2e427872021-04-09 23:40:07 -07002694 "src/qs8-igemm/gen/2x4c2-minmax-ssse3-ld64.c",
2695 "src/qs8-igemm/gen/2x4c2-minmax-ssse3-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002696 "src/qs8-igemm/gen/2x4c8-minmax-ssse3-ld64.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002697 "src/qs8-igemm/gen/2x4c8-minmax-ssse3-ld128.c",
Marat Dukhan2e427872021-04-09 23:40:07 -07002698 "src/qs8-igemm/gen/3x4c2-minmax-ssse3-ld64.c",
2699 "src/qs8-igemm/gen/3x4c2-minmax-ssse3-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002700 "src/qs8-igemm/gen/3x4c8-minmax-ssse3-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002701 "src/qs8-igemm/gen/3x4c8-minmax-ssse3-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002702 "src/qs8-igemm/gen/4x4c2-minmax-ssse3-ld64.c",
2703 "src/qs8-igemm/gen/4x4c2-minmax-ssse3-ld128.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -07002704 "src/qs8-requantization/precise-ssse3.c",
2705 "src/qs8-requantization/q31-ssse3.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07002706 "src/qu8-requantization/precise-ssse3.c",
2707 "src/qu8-requantization/q31-ssse3.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002708]
2709
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08002710SSE41_UKERNELS = [
Marat Dukhan40a672f2019-11-25 03:08:22 -08002711 "src/f32-prelu/gen/sse41-2x4.c",
2712 "src/f32-prelu/gen/sse41-2x8.c",
Marat Dukhaned6baaf2020-12-01 15:07:08 -08002713 "src/f32-velu/gen/velu-sse41-rr2-lut16-p3-x4.c",
2714 "src/f32-velu/gen/velu-sse41-rr2-lut16-p3-x8.c",
2715 "src/f32-velu/gen/velu-sse41-rr2-lut16-p3-x12.c",
2716 "src/f32-velu/gen/velu-sse41-rr2-lut16-p3-x16.c",
2717 "src/f32-velu/gen/velu-sse41-rr2-lut16-p3-x20.c",
2718 "src/f32-velu/gen/velu-sse41-rr2-lut16-p3-x24.c",
2719 "src/f32-velu/gen/velu-sse41-rr2-p6-x4.c",
2720 "src/f32-velu/gen/velu-sse41-rr2-p6-x8.c",
2721 "src/f32-velu/gen/velu-sse41-rr2-p6-x12.c",
2722 "src/f32-velu/gen/velu-sse41-rr2-p6-x16.c",
2723 "src/f32-velu/gen/velu-sse41-rr2-p6-x20.c",
2724 "src/f32-velu/gen/velu-sse41-rr2-p6-x24.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07002725 "src/f32-vlrelu/gen/vlrelu-sse41-x4.c",
2726 "src/f32-vlrelu/gen/vlrelu-sse41-x8.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07002727 "src/f32-vrnd/gen/vrndd-sse41-x4.c",
2728 "src/f32-vrnd/gen/vrndd-sse41-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002729 "src/f32-vrnd/gen/vrndne-sse41-x4.c",
2730 "src/f32-vrnd/gen/vrndne-sse41-x8.c",
2731 "src/f32-vrnd/gen/vrndu-sse41-x4.c",
2732 "src/f32-vrnd/gen/vrndu-sse41-x8.c",
2733 "src/f32-vrnd/gen/vrndz-sse41-x4.c",
2734 "src/f32-vrnd/gen/vrndz-sse41-x8.c",
Marat Dukhanb1eec082021-05-05 23:24:55 -07002735 "src/f32-vsigmoid/gen/vsigmoid-sse41-lut64-p2-div-x4.c",
2736 "src/f32-vsigmoid/gen/vsigmoid-sse41-lut64-p2-div-x8.c",
2737 "src/f32-vsigmoid/gen/vsigmoid-sse41-lut64-p2-div-x12.c",
2738 "src/f32-vsigmoid/gen/vsigmoid-sse41-lut64-p2-div-x16.c",
2739 "src/f32-vsigmoid/gen/vsigmoid-sse41-lut64-p2-div-x20.c",
2740 "src/f32-vsigmoid/gen/vsigmoid-sse41-lut64-p2-div-x24.c",
2741 "src/f32-vsigmoid/gen/vsigmoid-sse41-p5-div-x4.c",
2742 "src/f32-vsigmoid/gen/vsigmoid-sse41-p5-div-x8.c",
2743 "src/f32-vsigmoid/gen/vsigmoid-sse41-p5-div-x12.c",
2744 "src/f32-vsigmoid/gen/vsigmoid-sse41-p5-div-x16.c",
2745 "src/f32-vsigmoid/gen/vsigmoid-sse41-p5-div-x20.c",
2746 "src/f32-vsigmoid/gen/vsigmoid-sse41-p5-div-x24.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002747 "src/math/roundd-sse41.c",
2748 "src/math/roundne-sse41.c",
2749 "src/math/roundu-sse41.c",
2750 "src/math/roundz-sse41.c",
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07002751 "src/qs8-dwconv/gen/up8x9-minmax-sse41-mul16.c",
Marat Dukhan07feec82021-04-02 22:41:15 -07002752 "src/qs8-dwconv/gen/up8x9-minmax-sse41-mul32.c",
Frank Barchard4a352042021-04-13 15:52:08 -07002753 "src/qs8-dwconv/gen/up16x9-minmax-sse41-mul16.c",
Marat Dukhan07feec82021-04-02 22:41:15 -07002754 "src/qs8-dwconv/gen/up16x9-minmax-sse41-mul32.c",
Frank Barchard4a352042021-04-13 15:52:08 -07002755 "src/qs8-dwconv/gen/up24x9-minmax-sse41-mul16.c",
Marat Dukhan07feec82021-04-02 22:41:15 -07002756 "src/qs8-dwconv/gen/up24x9-minmax-sse41-mul32.c",
Marat Dukhan159688f2020-08-06 10:34:29 -07002757 "src/qs8-gavgpool/gen/7p7x-minmax-sse41-c8-acc2.c",
2758 "src/qs8-gavgpool/gen/7p7x-minmax-sse41-c16-acc2.c",
2759 "src/qs8-gavgpool/gen/7p7x-minmax-sse41-c24-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002760 "src/qs8-gavgpool/gen/7x-minmax-sse41-c8-acc2.c",
2761 "src/qs8-gavgpool/gen/7x-minmax-sse41-c16-acc2.c",
2762 "src/qs8-gavgpool/gen/7x-minmax-sse41-c24-acc2.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002763 "src/qs8-gemm/gen/1x4c2-minmax-sse41-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002764 "src/qs8-gemm/gen/1x4c2-minmax-sse41-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002765 "src/qs8-gemm/gen/1x4c2-xw-minmax-sse41.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002766 "src/qs8-gemm/gen/1x4c8-minmax-sse41-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002767 "src/qs8-gemm/gen/1x4c8-minmax-sse41-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002768 "src/qs8-gemm/gen/1x4c8-xw-minmax-sse41.c",
Marat Dukhan2e427872021-04-09 23:40:07 -07002769 "src/qs8-gemm/gen/2x4c2-minmax-sse41-ld64.c",
2770 "src/qs8-gemm/gen/2x4c2-minmax-sse41-ld128.c",
2771 "src/qs8-gemm/gen/2x4c2-xw-minmax-sse41.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002772 "src/qs8-gemm/gen/2x4c8-minmax-sse41-ld64.c",
2773 "src/qs8-gemm/gen/2x4c8-minmax-sse41-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002774 "src/qs8-gemm/gen/2x4c8-xw-minmax-sse41.c",
Marat Dukhan2e427872021-04-09 23:40:07 -07002775 "src/qs8-gemm/gen/3x4c2-minmax-sse41-ld64.c",
2776 "src/qs8-gemm/gen/3x4c2-minmax-sse41-ld128.c",
2777 "src/qs8-gemm/gen/3x4c2-xw-minmax-sse41.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002778 "src/qs8-gemm/gen/3x4c8-minmax-sse41-ld64.c",
2779 "src/qs8-gemm/gen/3x4c8-minmax-sse41-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002780 "src/qs8-gemm/gen/3x4c8-xw-minmax-sse41.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002781 "src/qs8-gemm/gen/4x4c2-minmax-sse41-ld64.c",
2782 "src/qs8-gemm/gen/4x4c2-minmax-sse41-ld128.c",
2783 "src/qs8-gemm/gen/4x4c2-xw-minmax-sse41.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002784 "src/qs8-igemm/gen/1x4c2-minmax-sse41-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002785 "src/qs8-igemm/gen/1x4c2-minmax-sse41-ld128.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002786 "src/qs8-igemm/gen/1x4c8-minmax-sse41-ld64.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002787 "src/qs8-igemm/gen/1x4c8-minmax-sse41-ld128.c",
Marat Dukhan2e427872021-04-09 23:40:07 -07002788 "src/qs8-igemm/gen/2x4c2-minmax-sse41-ld64.c",
2789 "src/qs8-igemm/gen/2x4c2-minmax-sse41-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002790 "src/qs8-igemm/gen/2x4c8-minmax-sse41-ld64.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002791 "src/qs8-igemm/gen/2x4c8-minmax-sse41-ld128.c",
Marat Dukhan2e427872021-04-09 23:40:07 -07002792 "src/qs8-igemm/gen/3x4c2-minmax-sse41-ld64.c",
2793 "src/qs8-igemm/gen/3x4c2-minmax-sse41-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002794 "src/qs8-igemm/gen/3x4c8-minmax-sse41-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002795 "src/qs8-igemm/gen/3x4c8-minmax-sse41-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002796 "src/qs8-igemm/gen/4x4c2-minmax-sse41-ld64.c",
2797 "src/qs8-igemm/gen/4x4c2-minmax-sse41-ld128.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -07002798 "src/qs8-requantization/fp32-sse4.c",
2799 "src/qs8-requantization/precise-sse4.c",
2800 "src/qs8-requantization/q31-sse4.c",
Marat Dukhand9f3ad42020-08-10 12:30:58 -07002801 "src/qs8-vadd/gen/minmax-sse41-mul16-ld64-x8.c",
2802 "src/qs8-vadd/gen/minmax-sse41-mul16-ld64-x16.c",
2803 "src/qs8-vadd/gen/minmax-sse41-mul16-ld64-x24.c",
2804 "src/qs8-vadd/gen/minmax-sse41-mul16-ld64-x32.c",
Marat Dukhanbb9225e2020-09-06 22:40:56 -07002805 "src/qs8-vadd/gen/minmax-sse41-mul32-ld32-x8.c",
2806 "src/qs8-vadd/gen/minmax-sse41-mul32-ld32-x16.c",
2807 "src/qs8-vadd/gen/minmax-sse41-mul32-ld32-x24.c",
2808 "src/qs8-vadd/gen/minmax-sse41-mul32-ld32-x32.c",
Marat Dukhan0270d9f2020-08-11 00:56:46 -07002809 "src/qs8-vaddc/gen/minmax-sse41-mul16-ld64-x8.c",
2810 "src/qs8-vaddc/gen/minmax-sse41-mul16-ld64-x16.c",
2811 "src/qs8-vaddc/gen/minmax-sse41-mul16-ld64-x24.c",
2812 "src/qs8-vaddc/gen/minmax-sse41-mul16-ld64-x32.c",
Marat Dukhanbb9225e2020-09-06 22:40:56 -07002813 "src/qs8-vaddc/gen/minmax-sse41-mul32-ld32-x8.c",
2814 "src/qs8-vaddc/gen/minmax-sse41-mul32-ld32-x16.c",
2815 "src/qs8-vaddc/gen/minmax-sse41-mul32-ld32-x24.c",
2816 "src/qs8-vaddc/gen/minmax-sse41-mul32-ld32-x32.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07002817 "src/qu8-requantization/precise-sse4.c",
2818 "src/qu8-requantization/q31-sse4.c",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08002819]
2820
Marat Dukhan08c4a432019-10-03 09:29:21 -07002821AVX_UKERNELS = [
Marat Dukhan1c587112020-04-08 20:04:28 -07002822 "src/f32-dwconv/gen/up8x4-minmax-avx-acc2.c",
2823 "src/f32-dwconv/gen/up8x4-minmax-avx.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002824 "src/f32-dwconv/gen/up8x9-minmax-avx-acc2.c",
2825 "src/f32-dwconv/gen/up8x9-minmax-avx.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002826 "src/f32-dwconv/gen/up8x25-minmax-avx-acc2.c",
2827 "src/f32-dwconv/gen/up8x25-minmax-avx.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002828 "src/f32-dwconv/gen/up16x4-minmax-avx-acc2.c",
2829 "src/f32-dwconv/gen/up16x4-minmax-avx.c",
2830 "src/f32-dwconv/gen/up16x9-minmax-avx-acc2.c",
2831 "src/f32-dwconv/gen/up16x9-minmax-avx.c",
2832 "src/f32-dwconv/gen/up16x25-minmax-avx-acc2.c",
2833 "src/f32-dwconv/gen/up16x25-minmax-avx.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002834 "src/f32-gemm/gen-inc/1x8inc-minmax-avx-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002835 "src/f32-gemm/gen-inc/1x16inc-minmax-avx-broadcast.c",
2836 "src/f32-gemm/gen-inc/3x16inc-minmax-avx-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002837 "src/f32-gemm/gen-inc/4x8inc-minmax-avx-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002838 "src/f32-gemm/gen-inc/4x16inc-minmax-avx-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002839 "src/f32-gemm/gen-inc/5x8inc-minmax-avx-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002840 "src/f32-gemm/gen-inc/5x16inc-minmax-avx-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002841 "src/f32-gemm/gen-inc/6x8inc-minmax-avx-broadcast.c",
2842 "src/f32-gemm/gen-inc/7x8inc-minmax-avx-broadcast.c",
2843 "src/f32-gemm/gen/1x8-minmax-avx-broadcast.c",
2844 "src/f32-gemm/gen/1x16-minmax-avx-broadcast.c",
2845 "src/f32-gemm/gen/3x16-minmax-avx-broadcast.c",
2846 "src/f32-gemm/gen/4x8-minmax-avx-broadcast.c",
2847 "src/f32-gemm/gen/4x16-minmax-avx-broadcast.c",
2848 "src/f32-gemm/gen/5x8-minmax-avx-broadcast.c",
2849 "src/f32-gemm/gen/5x16-minmax-avx-broadcast.c",
2850 "src/f32-gemm/gen/6x8-minmax-avx-broadcast.c",
2851 "src/f32-gemm/gen/7x8-minmax-avx-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002852 "src/f32-igemm/gen/1x8-minmax-avx-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002853 "src/f32-igemm/gen/1x16-minmax-avx-broadcast.c",
2854 "src/f32-igemm/gen/3x16-minmax-avx-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002855 "src/f32-igemm/gen/4x8-minmax-avx-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002856 "src/f32-igemm/gen/4x16-minmax-avx-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002857 "src/f32-igemm/gen/5x8-minmax-avx-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002858 "src/f32-igemm/gen/5x16-minmax-avx-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002859 "src/f32-igemm/gen/6x8-minmax-avx-broadcast.c",
2860 "src/f32-igemm/gen/7x8-minmax-avx-broadcast.c",
Marat Dukhan90eca0a2020-03-11 00:52:23 -07002861 "src/f32-prelu/gen/avx-2x8.c",
2862 "src/f32-prelu/gen/avx-2x16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002863 "src/f32-rmax/avx.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002864 "src/f32-vbinary/gen/vadd-minmax-avx-x8.c",
2865 "src/f32-vbinary/gen/vadd-minmax-avx-x16.c",
2866 "src/f32-vbinary/gen/vaddc-minmax-avx-x8.c",
2867 "src/f32-vbinary/gen/vaddc-minmax-avx-x16.c",
2868 "src/f32-vbinary/gen/vdiv-minmax-avx-x8.c",
2869 "src/f32-vbinary/gen/vdiv-minmax-avx-x16.c",
2870 "src/f32-vbinary/gen/vdivc-minmax-avx-x8.c",
2871 "src/f32-vbinary/gen/vdivc-minmax-avx-x16.c",
Marat Dukhan9a88efe2019-12-10 15:54:24 -08002872 "src/f32-vbinary/gen/vmax-avx-x8.c",
2873 "src/f32-vbinary/gen/vmax-avx-x16.c",
2874 "src/f32-vbinary/gen/vmaxc-avx-x8.c",
2875 "src/f32-vbinary/gen/vmaxc-avx-x16.c",
2876 "src/f32-vbinary/gen/vmin-avx-x8.c",
2877 "src/f32-vbinary/gen/vmin-avx-x16.c",
2878 "src/f32-vbinary/gen/vminc-avx-x8.c",
2879 "src/f32-vbinary/gen/vminc-avx-x16.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002880 "src/f32-vbinary/gen/vmul-minmax-avx-x8.c",
2881 "src/f32-vbinary/gen/vmul-minmax-avx-x16.c",
2882 "src/f32-vbinary/gen/vmulc-minmax-avx-x8.c",
2883 "src/f32-vbinary/gen/vmulc-minmax-avx-x16.c",
2884 "src/f32-vbinary/gen/vrdivc-minmax-avx-x8.c",
2885 "src/f32-vbinary/gen/vrdivc-minmax-avx-x16.c",
2886 "src/f32-vbinary/gen/vrsubc-minmax-avx-x8.c",
2887 "src/f32-vbinary/gen/vrsubc-minmax-avx-x16.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07002888 "src/f32-vbinary/gen/vsqrdiff-avx-x8.c",
2889 "src/f32-vbinary/gen/vsqrdiff-avx-x16.c",
2890 "src/f32-vbinary/gen/vsqrdiffc-avx-x8.c",
2891 "src/f32-vbinary/gen/vsqrdiffc-avx-x16.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002892 "src/f32-vbinary/gen/vsub-minmax-avx-x8.c",
2893 "src/f32-vbinary/gen/vsub-minmax-avx-x16.c",
2894 "src/f32-vbinary/gen/vsubc-minmax-avx-x8.c",
2895 "src/f32-vbinary/gen/vsubc-minmax-avx-x16.c",
Marat Dukhanb1eec082021-05-05 23:24:55 -07002896 "src/f32-vclamp/gen/vclamp-avx-x8.c",
2897 "src/f32-vclamp/gen/vclamp-avx-x16.c",
Marat Dukhaned6baaf2020-12-01 15:07:08 -08002898 "src/f32-velu/gen/velu-avx-rr2-lut4-p4-perm-x8.c",
2899 "src/f32-velu/gen/velu-avx-rr2-lut4-p4-perm-x16.c",
2900 "src/f32-velu/gen/velu-avx-rr2-lut4-p4-perm-x24.c",
2901 "src/f32-velu/gen/velu-avx-rr2-lut4-p4-perm-x32.c",
2902 "src/f32-velu/gen/velu-avx-rr2-lut4-p4-perm-x40.c",
2903 "src/f32-velu/gen/velu-avx-rr2-lut4-p4-perm-x48.c",
2904 "src/f32-velu/gen/velu-avx-rr2-lut16-p3-x8.c",
2905 "src/f32-velu/gen/velu-avx-rr2-lut16-p3-x16.c",
2906 "src/f32-velu/gen/velu-avx-rr2-lut16-p3-x24.c",
2907 "src/f32-velu/gen/velu-avx-rr2-lut16-p3-x32.c",
2908 "src/f32-velu/gen/velu-avx-rr2-lut16-p3-x40.c",
2909 "src/f32-velu/gen/velu-avx-rr2-lut16-p3-x48.c",
2910 "src/f32-velu/gen/velu-avx-rr2-p6-x8.c",
2911 "src/f32-velu/gen/velu-avx-rr2-p6-x16.c",
2912 "src/f32-velu/gen/velu-avx-rr2-p6-x24.c",
2913 "src/f32-velu/gen/velu-avx-rr2-p6-x32.c",
2914 "src/f32-velu/gen/velu-avx-rr2-p6-x40.c",
2915 "src/f32-velu/gen/velu-avx-rr2-p6-x48.c",
Marat Dukhan6674d692021-05-05 22:27:00 -07002916 "src/f32-vhswish/gen/vhswish-avx-x8.c",
2917 "src/f32-vhswish/gen/vhswish-avx-x16.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07002918 "src/f32-vlrelu/gen/vlrelu-avx-x8.c",
2919 "src/f32-vlrelu/gen/vlrelu-avx-x16.c",
Marat Dukhanb1eec082021-05-05 23:24:55 -07002920 "src/f32-vrelu/gen/vrelu-avx-x8.c",
2921 "src/f32-vrelu/gen/vrelu-avx-x16.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07002922 "src/f32-vrnd/gen/vrndd-avx-x8.c",
2923 "src/f32-vrnd/gen/vrndd-avx-x16.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002924 "src/f32-vrnd/gen/vrndne-avx-x8.c",
2925 "src/f32-vrnd/gen/vrndne-avx-x16.c",
2926 "src/f32-vrnd/gen/vrndu-avx-x8.c",
2927 "src/f32-vrnd/gen/vrndu-avx-x16.c",
2928 "src/f32-vrnd/gen/vrndz-avx-x8.c",
2929 "src/f32-vrnd/gen/vrndz-avx-x16.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07002930 "src/f32-vscale/avx-x32.c",
Marat Dukhanb1eec082021-05-05 23:24:55 -07002931 "src/f32-vsigmoid/gen/vsigmoid-avx-rr2-p5-div-x8.c",
2932 "src/f32-vsigmoid/gen/vsigmoid-avx-rr2-p5-div-x16.c",
2933 "src/f32-vsigmoid/gen/vsigmoid-avx-rr2-p5-div-x24.c",
2934 "src/f32-vsigmoid/gen/vsigmoid-avx-rr2-p5-div-x32.c",
2935 "src/f32-vsigmoid/gen/vsigmoid-avx-rr2-p5-div-x40.c",
2936 "src/f32-vsigmoid/gen/vsigmoid-avx-rr2-p5-div-x48.c",
2937 "src/f32-vsigmoid/gen/vsigmoid-avx-rr2-p5-div-x56.c",
2938 "src/f32-vsigmoid/gen/vsigmoid-avx-rr2-p5-div-x64.c",
2939 "src/f32-vsigmoid/gen/vsigmoid-avx-rr2-p5-div-x72.c",
2940 "src/f32-vsigmoid/gen/vsigmoid-avx-rr2-p5-div-x80.c",
2941 "src/f32-vsigmoid/gen/vsigmoid-avx-rr2-p5-nr2-x8.c",
2942 "src/f32-vsigmoid/gen/vsigmoid-avx-rr2-p5-nr2-x16.c",
2943 "src/f32-vsigmoid/gen/vsigmoid-avx-rr2-p5-nr2-x24.c",
2944 "src/f32-vsigmoid/gen/vsigmoid-avx-rr2-p5-nr2-x32.c",
2945 "src/f32-vsigmoid/gen/vsigmoid-avx-rr2-p5-nr2-x40.c",
2946 "src/f32-vsigmoid/gen/vsigmoid-avx-rr2-p5-nr2-x48.c",
2947 "src/f32-vsigmoid/gen/vsigmoid-avx-rr2-p5-nr2-x56.c",
2948 "src/f32-vsigmoid/gen/vsigmoid-avx-rr2-p5-nr2-x64.c",
2949 "src/f32-vsigmoid/gen/vsigmoid-avx-rr2-p5-nr2-x72.c",
2950 "src/f32-vsigmoid/gen/vsigmoid-avx-rr2-p5-nr2-x80.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07002951 "src/f32-vsqrt/gen/avx-sqrt-x8.c",
2952 "src/f32-vsqrt/gen/avx-sqrt-x16.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07002953 "src/f32-vunary/gen/vabs-avx-x8.c",
2954 "src/f32-vunary/gen/vabs-avx-x16.c",
2955 "src/f32-vunary/gen/vneg-avx-x8.c",
2956 "src/f32-vunary/gen/vneg-avx-x16.c",
2957 "src/f32-vunary/gen/vsqr-avx-x8.c",
2958 "src/f32-vunary/gen/vsqr-avx-x16.c",
Frank Barchard4a352042021-04-13 15:52:08 -07002959 "src/math/exp-avx-rr2-p5.c",
2960 "src/math/expm1minus-avx-rr2-lut4-p4-perm.c",
2961 "src/math/expm1minus-avx-rr2-lut16-p3.c",
2962 "src/math/expm1minus-avx-rr2-p6.c",
2963 "src/math/sigmoid-avx-rr2-lut64-p2-div.c",
2964 "src/math/sigmoid-avx-rr2-p5-div.c",
2965 "src/math/sigmoid-avx-rr2-p5-nr1.c",
2966 "src/math/sigmoid-avx-rr2-p5-nr2.c",
Marat Dukhanfa0ab852021-04-02 17:30:49 -07002967 "src/qs8-dwconv/gen/up8x9-minmax-avx-mul16.c",
Marat Dukhan07feec82021-04-02 22:41:15 -07002968 "src/qs8-dwconv/gen/up8x9-minmax-avx-mul32.c",
Frank Barchard4a352042021-04-13 15:52:08 -07002969 "src/qs8-dwconv/gen/up16x9-minmax-avx-mul16.c",
Marat Dukhan07feec82021-04-02 22:41:15 -07002970 "src/qs8-dwconv/gen/up16x9-minmax-avx-mul32.c",
Frank Barchard4a352042021-04-13 15:52:08 -07002971 "src/qs8-dwconv/gen/up24x9-minmax-avx-mul16.c",
Marat Dukhan07feec82021-04-02 22:41:15 -07002972 "src/qs8-dwconv/gen/up24x9-minmax-avx-mul32.c",
Marat Dukhana3c16332021-04-02 15:03:27 -07002973 "src/qs8-gemm/gen/1x4c2-minmax-avx-ld64.c",
2974 "src/qs8-gemm/gen/1x4c2-minmax-avx-ld128.c",
2975 "src/qs8-gemm/gen/1x4c2-xw-minmax-avx.c",
2976 "src/qs8-gemm/gen/1x4c8-minmax-avx-ld64.c",
2977 "src/qs8-gemm/gen/1x4c8-minmax-avx-ld128.c",
2978 "src/qs8-gemm/gen/1x4c8-xw-minmax-avx.c",
Marat Dukhan2e427872021-04-09 23:40:07 -07002979 "src/qs8-gemm/gen/2x4c2-minmax-avx-ld64.c",
2980 "src/qs8-gemm/gen/2x4c2-minmax-avx-ld128.c",
2981 "src/qs8-gemm/gen/2x4c2-xw-minmax-avx.c",
Marat Dukhana3c16332021-04-02 15:03:27 -07002982 "src/qs8-gemm/gen/2x4c8-minmax-avx-ld64.c",
2983 "src/qs8-gemm/gen/2x4c8-minmax-avx-ld128.c",
2984 "src/qs8-gemm/gen/2x4c8-xw-minmax-avx.c",
Marat Dukhan2e427872021-04-09 23:40:07 -07002985 "src/qs8-gemm/gen/3x4c2-minmax-avx-ld64.c",
2986 "src/qs8-gemm/gen/3x4c2-minmax-avx-ld128.c",
2987 "src/qs8-gemm/gen/3x4c2-xw-minmax-avx.c",
Marat Dukhana3c16332021-04-02 15:03:27 -07002988 "src/qs8-gemm/gen/3x4c8-minmax-avx-ld64.c",
2989 "src/qs8-gemm/gen/3x4c8-minmax-avx-ld128.c",
2990 "src/qs8-gemm/gen/3x4c8-xw-minmax-avx.c",
2991 "src/qs8-gemm/gen/4x4c2-minmax-avx-ld64.c",
2992 "src/qs8-gemm/gen/4x4c2-minmax-avx-ld128.c",
2993 "src/qs8-gemm/gen/4x4c2-xw-minmax-avx.c",
2994 "src/qs8-igemm/gen/1x4c2-minmax-avx-ld64.c",
2995 "src/qs8-igemm/gen/1x4c2-minmax-avx-ld128.c",
2996 "src/qs8-igemm/gen/1x4c8-minmax-avx-ld64.c",
2997 "src/qs8-igemm/gen/1x4c8-minmax-avx-ld128.c",
Marat Dukhan2e427872021-04-09 23:40:07 -07002998 "src/qs8-igemm/gen/2x4c2-minmax-avx-ld64.c",
2999 "src/qs8-igemm/gen/2x4c2-minmax-avx-ld128.c",
Marat Dukhana3c16332021-04-02 15:03:27 -07003000 "src/qs8-igemm/gen/2x4c8-minmax-avx-ld64.c",
3001 "src/qs8-igemm/gen/2x4c8-minmax-avx-ld128.c",
Marat Dukhan2e427872021-04-09 23:40:07 -07003002 "src/qs8-igemm/gen/3x4c2-minmax-avx-ld64.c",
3003 "src/qs8-igemm/gen/3x4c2-minmax-avx-ld128.c",
Marat Dukhana3c16332021-04-02 15:03:27 -07003004 "src/qs8-igemm/gen/3x4c8-minmax-avx-ld64.c",
3005 "src/qs8-igemm/gen/3x4c8-minmax-avx-ld128.c",
3006 "src/qs8-igemm/gen/4x4c2-minmax-avx-ld64.c",
3007 "src/qs8-igemm/gen/4x4c2-minmax-avx-ld128.c",
Marat Dukhane9c4b962021-04-02 16:56:55 -07003008 "src/qs8-vadd/gen/minmax-avx-mul16-ld64-x8.c",
3009 "src/qs8-vadd/gen/minmax-avx-mul16-ld64-x16.c",
3010 "src/qs8-vadd/gen/minmax-avx-mul16-ld64-x24.c",
3011 "src/qs8-vadd/gen/minmax-avx-mul16-ld64-x32.c",
3012 "src/qs8-vadd/gen/minmax-avx-mul32-ld32-x8.c",
3013 "src/qs8-vadd/gen/minmax-avx-mul32-ld32-x16.c",
3014 "src/qs8-vadd/gen/minmax-avx-mul32-ld32-x24.c",
3015 "src/qs8-vadd/gen/minmax-avx-mul32-ld32-x32.c",
3016 "src/qs8-vaddc/gen/minmax-avx-mul16-ld64-x8.c",
3017 "src/qs8-vaddc/gen/minmax-avx-mul16-ld64-x16.c",
3018 "src/qs8-vaddc/gen/minmax-avx-mul16-ld64-x24.c",
3019 "src/qs8-vaddc/gen/minmax-avx-mul16-ld64-x32.c",
3020 "src/qs8-vaddc/gen/minmax-avx-mul32-ld32-x8.c",
3021 "src/qs8-vaddc/gen/minmax-avx-mul32-ld32-x16.c",
3022 "src/qs8-vaddc/gen/minmax-avx-mul32-ld32-x24.c",
3023 "src/qs8-vaddc/gen/minmax-avx-mul32-ld32-x32.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003024]
3025
Marat Dukhan1566fee2020-08-02 21:55:41 -07003026XOP_UKERNELS = [
Marat Dukhan3fd4e272021-04-10 11:16:42 -07003027 "src/qs8-dwconv/gen/up8x9-minmax-xop-mul32.c",
3028 "src/qs8-dwconv/gen/up16x9-minmax-xop-mul32.c",
3029 "src/qs8-dwconv/gen/up24x9-minmax-xop-mul32.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07003030 "src/qs8-gemm/gen/1x4c2-minmax-xop-ld64.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07003031 "src/qs8-gemm/gen/1x4c2-minmax-xop-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07003032 "src/qs8-gemm/gen/1x4c2-xw-minmax-xop.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07003033 "src/qs8-gemm/gen/1x4c8-minmax-xop-ld64.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07003034 "src/qs8-gemm/gen/1x4c8-minmax-xop-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07003035 "src/qs8-gemm/gen/1x4c8-xw-minmax-xop.c",
Marat Dukhan2e427872021-04-09 23:40:07 -07003036 "src/qs8-gemm/gen/2x4c2-minmax-xop-ld64.c",
3037 "src/qs8-gemm/gen/2x4c2-minmax-xop-ld128.c",
3038 "src/qs8-gemm/gen/2x4c2-xw-minmax-xop.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003039 "src/qs8-gemm/gen/2x4c8-minmax-xop-ld64.c",
3040 "src/qs8-gemm/gen/2x4c8-minmax-xop-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07003041 "src/qs8-gemm/gen/2x4c8-xw-minmax-xop.c",
Marat Dukhan2e427872021-04-09 23:40:07 -07003042 "src/qs8-gemm/gen/3x4c2-minmax-xop-ld64.c",
3043 "src/qs8-gemm/gen/3x4c2-minmax-xop-ld128.c",
3044 "src/qs8-gemm/gen/3x4c2-xw-minmax-xop.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003045 "src/qs8-gemm/gen/3x4c8-minmax-xop-ld64.c",
3046 "src/qs8-gemm/gen/3x4c8-minmax-xop-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07003047 "src/qs8-gemm/gen/3x4c8-xw-minmax-xop.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003048 "src/qs8-gemm/gen/4x4c2-minmax-xop-ld64.c",
3049 "src/qs8-gemm/gen/4x4c2-minmax-xop-ld128.c",
3050 "src/qs8-gemm/gen/4x4c2-xw-minmax-xop.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07003051 "src/qs8-igemm/gen/1x4c2-minmax-xop-ld64.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07003052 "src/qs8-igemm/gen/1x4c2-minmax-xop-ld128.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07003053 "src/qs8-igemm/gen/1x4c8-minmax-xop-ld64.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07003054 "src/qs8-igemm/gen/1x4c8-minmax-xop-ld128.c",
Marat Dukhan2e427872021-04-09 23:40:07 -07003055 "src/qs8-igemm/gen/2x4c2-minmax-xop-ld64.c",
3056 "src/qs8-igemm/gen/2x4c2-minmax-xop-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003057 "src/qs8-igemm/gen/2x4c8-minmax-xop-ld64.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07003058 "src/qs8-igemm/gen/2x4c8-minmax-xop-ld128.c",
Marat Dukhan2e427872021-04-09 23:40:07 -07003059 "src/qs8-igemm/gen/3x4c2-minmax-xop-ld64.c",
3060 "src/qs8-igemm/gen/3x4c2-minmax-xop-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003061 "src/qs8-igemm/gen/3x4c8-minmax-xop-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07003062 "src/qs8-igemm/gen/3x4c8-minmax-xop-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003063 "src/qs8-igemm/gen/4x4c2-minmax-xop-ld64.c",
3064 "src/qs8-igemm/gen/4x4c2-minmax-xop-ld128.c",
Marat Dukhanbb9225e2020-09-06 22:40:56 -07003065 "src/qs8-vadd/gen/minmax-xop-mul32-ld32-x8.c",
3066 "src/qs8-vadd/gen/minmax-xop-mul32-ld32-x16.c",
3067 "src/qs8-vadd/gen/minmax-xop-mul32-ld32-x24.c",
3068 "src/qs8-vadd/gen/minmax-xop-mul32-ld32-x32.c",
3069 "src/qs8-vaddc/gen/minmax-xop-mul32-ld32-x8.c",
3070 "src/qs8-vaddc/gen/minmax-xop-mul32-ld32-x16.c",
3071 "src/qs8-vaddc/gen/minmax-xop-mul32-ld32-x24.c",
3072 "src/qs8-vaddc/gen/minmax-xop-mul32-ld32-x32.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07003073]
3074
Marat Dukhanfda12b82019-11-21 12:27:59 -08003075FMA3_UKERNELS = [
Marat Dukhan1c587112020-04-08 20:04:28 -07003076 "src/f32-dwconv/gen/up8x4-minmax-fma3-acc2.c",
3077 "src/f32-dwconv/gen/up8x4-minmax-fma3.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07003078 "src/f32-dwconv/gen/up8x9-minmax-fma3-acc2.c",
3079 "src/f32-dwconv/gen/up8x9-minmax-fma3.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07003080 "src/f32-dwconv/gen/up8x25-minmax-fma3-acc2.c",
3081 "src/f32-dwconv/gen/up8x25-minmax-fma3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003082 "src/f32-dwconv/gen/up16x4-minmax-fma3-acc2.c",
3083 "src/f32-dwconv/gen/up16x4-minmax-fma3.c",
3084 "src/f32-dwconv/gen/up16x9-minmax-fma3-acc2.c",
3085 "src/f32-dwconv/gen/up16x9-minmax-fma3.c",
3086 "src/f32-dwconv/gen/up16x25-minmax-fma3-acc2.c",
3087 "src/f32-dwconv/gen/up16x25-minmax-fma3.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07003088 "src/f32-gemm/gen-inc/1x8inc-minmax-fma3-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003089 "src/f32-gemm/gen-inc/1x16inc-minmax-fma3-broadcast.c",
3090 "src/f32-gemm/gen-inc/1x16s4inc-minmax-fma3-broadcast.c",
3091 "src/f32-gemm/gen-inc/3x16inc-minmax-fma3-broadcast.c",
3092 "src/f32-gemm/gen-inc/3x16s4inc-minmax-fma3-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07003093 "src/f32-gemm/gen-inc/4x8inc-minmax-fma3-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003094 "src/f32-gemm/gen-inc/4x16inc-minmax-fma3-broadcast.c",
3095 "src/f32-gemm/gen-inc/4x16s4inc-minmax-fma3-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07003096 "src/f32-gemm/gen-inc/5x8inc-minmax-fma3-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003097 "src/f32-gemm/gen-inc/5x16inc-minmax-fma3-broadcast.c",
3098 "src/f32-gemm/gen-inc/5x16s4inc-minmax-fma3-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07003099 "src/f32-gemm/gen-inc/6x8inc-minmax-fma3-broadcast.c",
3100 "src/f32-gemm/gen-inc/7x8inc-minmax-fma3-broadcast.c",
3101 "src/f32-gemm/gen-inc/8x8inc-minmax-fma3-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003102 "src/f32-gemm/gen/1x8-minmax-fma3-broadcast.c",
3103 "src/f32-gemm/gen/1x16-minmax-fma3-broadcast.c",
3104 "src/f32-gemm/gen/1x16s4-minmax-fma3-broadcast.c",
3105 "src/f32-gemm/gen/3x16-minmax-fma3-broadcast.c",
3106 "src/f32-gemm/gen/3x16s4-minmax-fma3-broadcast.c",
3107 "src/f32-gemm/gen/4x8-minmax-fma3-broadcast.c",
3108 "src/f32-gemm/gen/4x16-minmax-fma3-broadcast.c",
3109 "src/f32-gemm/gen/4x16s4-minmax-fma3-broadcast.c",
3110 "src/f32-gemm/gen/5x8-minmax-fma3-broadcast.c",
3111 "src/f32-gemm/gen/5x16-minmax-fma3-broadcast.c",
3112 "src/f32-gemm/gen/5x16s4-minmax-fma3-broadcast.c",
3113 "src/f32-gemm/gen/6x8-minmax-fma3-broadcast.c",
3114 "src/f32-gemm/gen/7x8-minmax-fma3-broadcast.c",
3115 "src/f32-gemm/gen/8x8-minmax-fma3-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07003116 "src/f32-igemm/gen/1x8-minmax-fma3-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003117 "src/f32-igemm/gen/1x16-minmax-fma3-broadcast.c",
3118 "src/f32-igemm/gen/1x16s4-minmax-fma3-broadcast.c",
3119 "src/f32-igemm/gen/3x16-minmax-fma3-broadcast.c",
3120 "src/f32-igemm/gen/3x16s4-minmax-fma3-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07003121 "src/f32-igemm/gen/4x8-minmax-fma3-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003122 "src/f32-igemm/gen/4x16-minmax-fma3-broadcast.c",
3123 "src/f32-igemm/gen/4x16s4-minmax-fma3-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07003124 "src/f32-igemm/gen/5x8-minmax-fma3-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003125 "src/f32-igemm/gen/5x16-minmax-fma3-broadcast.c",
3126 "src/f32-igemm/gen/5x16s4-minmax-fma3-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07003127 "src/f32-igemm/gen/6x8-minmax-fma3-broadcast.c",
3128 "src/f32-igemm/gen/7x8-minmax-fma3-broadcast.c",
3129 "src/f32-igemm/gen/8x8-minmax-fma3-broadcast.c",
Marat Dukhan6674d692021-05-05 22:27:00 -07003130 "src/f32-vhswish/gen/vhswish-fma3-x8.c",
3131 "src/f32-vhswish/gen/vhswish-fma3-x16.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07003132 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x8.c",
3133 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x16.c",
3134 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x24.c",
3135 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x32.c",
3136 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x40.c",
3137 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x48.c",
3138 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x56.c",
3139 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x64.c",
Marat Dukhan84000762020-06-29 18:38:43 -07003140 "src/math/sqrt-fma3-nr1fma.c",
Marat Dukhan84000762020-06-29 18:38:43 -07003141 "src/math/sqrt-fma3-nr1fma1adj.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003142 "src/math/sqrt-fma3-nr2fma.c",
Marat Dukhanfda12b82019-11-21 12:27:59 -08003143]
3144
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003145AVX2_UKERNELS = [
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003146 "src/f32-raddexpminusmax/gen/avx2-p5-x64-acc2.c",
3147 "src/f32-raddexpminusmax/gen/avx2-p5-x64-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003148 "src/f32-raddexpminusmax/gen/avx2-p5-x64.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003149 "src/f32-raddexpminusmax/gen/avx2-p5-x72-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003150 "src/f32-raddexpminusmax/gen/avx2-p5-x72.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003151 "src/f32-raddexpminusmax/gen/avx2-p5-x80-acc2.c",
3152 "src/f32-raddexpminusmax/gen/avx2-p5-x80-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003153 "src/f32-raddexpminusmax/gen/avx2-p5-x80.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003154 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc2.c",
3155 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc3.c",
3156 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc6.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003157 "src/f32-raddexpminusmax/gen/avx2-p5-x96.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003158 "src/f32-raddextexp/gen/avx2-p5-x64-acc2.c",
3159 "src/f32-raddextexp/gen/avx2-p5-x64-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003160 "src/f32-raddextexp/gen/avx2-p5-x64.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003161 "src/f32-raddextexp/gen/avx2-p5-x72-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003162 "src/f32-raddextexp/gen/avx2-p5-x72.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003163 "src/f32-raddextexp/gen/avx2-p5-x80-acc2.c",
3164 "src/f32-raddextexp/gen/avx2-p5-x80-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003165 "src/f32-raddextexp/gen/avx2-p5-x80.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003166 "src/f32-raddextexp/gen/avx2-p5-x96-acc2.c",
3167 "src/f32-raddextexp/gen/avx2-p5-x96-acc3.c",
3168 "src/f32-raddextexp/gen/avx2-p5-x96-acc6.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003169 "src/f32-raddextexp/gen/avx2-p5-x96.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003170 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64-acc2.c",
3171 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003172 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003173 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x72-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003174 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x72.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003175 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80-acc2.c",
3176 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003177 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003178 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc2.c",
3179 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc3.c",
3180 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc6.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003181 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96.c",
Marat Dukhaned6baaf2020-12-01 15:07:08 -08003182 "src/f32-velu/gen/velu-avx2-rr1-lut4-p4-perm-x8.c",
3183 "src/f32-velu/gen/velu-avx2-rr1-lut4-p4-perm-x16.c",
3184 "src/f32-velu/gen/velu-avx2-rr1-lut4-p4-perm-x24.c",
3185 "src/f32-velu/gen/velu-avx2-rr1-lut4-p4-perm-x32.c",
3186 "src/f32-velu/gen/velu-avx2-rr1-lut4-p4-perm-x40.c",
3187 "src/f32-velu/gen/velu-avx2-rr1-lut4-p4-perm-x48.c",
3188 "src/f32-velu/gen/velu-avx2-rr1-lut4-p4-perm-x56.c",
3189 "src/f32-velu/gen/velu-avx2-rr1-lut4-p4-perm-x64.c",
3190 "src/f32-velu/gen/velu-avx2-rr1-lut4-p4-perm-x72.c",
3191 "src/f32-velu/gen/velu-avx2-rr1-lut4-p4-perm-x80.c",
3192 "src/f32-velu/gen/velu-avx2-rr1-lut8-p4-perm-x8.c",
3193 "src/f32-velu/gen/velu-avx2-rr1-lut8-p4-perm-x16.c",
3194 "src/f32-velu/gen/velu-avx2-rr1-lut8-p4-perm-x24.c",
3195 "src/f32-velu/gen/velu-avx2-rr1-lut8-p4-perm-x32.c",
3196 "src/f32-velu/gen/velu-avx2-rr1-lut8-p4-perm-x40.c",
3197 "src/f32-velu/gen/velu-avx2-rr1-lut8-p4-perm-x48.c",
3198 "src/f32-velu/gen/velu-avx2-rr1-lut8-p4-perm-x56.c",
3199 "src/f32-velu/gen/velu-avx2-rr1-lut8-p4-perm-x64.c",
3200 "src/f32-velu/gen/velu-avx2-rr1-lut8-p4-perm-x72.c",
3201 "src/f32-velu/gen/velu-avx2-rr1-lut8-p4-perm-x80.c",
3202 "src/f32-velu/gen/velu-avx2-rr1-lut16-p3-gather-x8.c",
3203 "src/f32-velu/gen/velu-avx2-rr1-lut16-p3-gather-x16.c",
3204 "src/f32-velu/gen/velu-avx2-rr1-lut16-p3-gather-x24.c",
3205 "src/f32-velu/gen/velu-avx2-rr1-lut16-p3-gather-x32.c",
3206 "src/f32-velu/gen/velu-avx2-rr1-lut16-p3-gather-x40.c",
3207 "src/f32-velu/gen/velu-avx2-rr1-lut16-p3-gather-x48.c",
3208 "src/f32-velu/gen/velu-avx2-rr1-lut16-p3-gather-x56.c",
3209 "src/f32-velu/gen/velu-avx2-rr1-lut16-p3-gather-x64.c",
3210 "src/f32-velu/gen/velu-avx2-rr1-lut16-p3-gather-x72.c",
3211 "src/f32-velu/gen/velu-avx2-rr1-lut16-p3-gather-x80.c",
3212 "src/f32-velu/gen/velu-avx2-rr1-p6-x8.c",
3213 "src/f32-velu/gen/velu-avx2-rr1-p6-x16.c",
3214 "src/f32-velu/gen/velu-avx2-rr1-p6-x24.c",
3215 "src/f32-velu/gen/velu-avx2-rr1-p6-x32.c",
3216 "src/f32-velu/gen/velu-avx2-rr1-p6-x40.c",
3217 "src/f32-velu/gen/velu-avx2-rr1-p6-x48.c",
3218 "src/f32-velu/gen/velu-avx2-rr1-p6-x56.c",
3219 "src/f32-velu/gen/velu-avx2-rr1-p6-x64.c",
3220 "src/f32-velu/gen/velu-avx2-rr1-p6-x72.c",
3221 "src/f32-velu/gen/velu-avx2-rr1-p6-x80.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003222 "src/f32-vscaleexpminusmax/gen/avx2-p5-x8.c",
3223 "src/f32-vscaleexpminusmax/gen/avx2-p5-x16.c",
3224 "src/f32-vscaleexpminusmax/gen/avx2-p5-x24.c",
3225 "src/f32-vscaleexpminusmax/gen/avx2-p5-x32.c",
3226 "src/f32-vscaleexpminusmax/gen/avx2-p5-x40.c",
3227 "src/f32-vscaleexpminusmax/gen/avx2-p5-x48.c",
3228 "src/f32-vscaleexpminusmax/gen/avx2-p5-x56.c",
3229 "src/f32-vscaleexpminusmax/gen/avx2-p5-x64.c",
3230 "src/f32-vscaleexpminusmax/gen/avx2-p5-x72.c",
3231 "src/f32-vscaleexpminusmax/gen/avx2-p5-x80.c",
3232 "src/f32-vscaleexpminusmax/gen/avx2-p5-x88.c",
3233 "src/f32-vscaleexpminusmax/gen/avx2-p5-x96.c",
3234 "src/f32-vscaleextexp/gen/avx2-p5-x8.c",
3235 "src/f32-vscaleextexp/gen/avx2-p5-x16.c",
3236 "src/f32-vscaleextexp/gen/avx2-p5-x24.c",
3237 "src/f32-vscaleextexp/gen/avx2-p5-x32.c",
3238 "src/f32-vscaleextexp/gen/avx2-p5-x40.c",
3239 "src/f32-vscaleextexp/gen/avx2-p5-x48.c",
3240 "src/f32-vscaleextexp/gen/avx2-p5-x56.c",
3241 "src/f32-vscaleextexp/gen/avx2-p5-x64.c",
3242 "src/f32-vscaleextexp/gen/avx2-p5-x72.c",
3243 "src/f32-vscaleextexp/gen/avx2-p5-x80.c",
3244 "src/f32-vscaleextexp/gen/avx2-p5-x88.c",
3245 "src/f32-vscaleextexp/gen/avx2-p5-x96.c",
Marat Dukhanb1eec082021-05-05 23:24:55 -07003246 "src/f32-vsigmoid/gen/vsigmoid-avx2-rr1-p5-div-x8.c",
3247 "src/f32-vsigmoid/gen/vsigmoid-avx2-rr1-p5-div-x16.c",
3248 "src/f32-vsigmoid/gen/vsigmoid-avx2-rr1-p5-div-x24.c",
3249 "src/f32-vsigmoid/gen/vsigmoid-avx2-rr1-p5-div-x32.c",
3250 "src/f32-vsigmoid/gen/vsigmoid-avx2-rr1-p5-div-x40.c",
3251 "src/f32-vsigmoid/gen/vsigmoid-avx2-rr1-p5-div-x48.c",
3252 "src/f32-vsigmoid/gen/vsigmoid-avx2-rr1-p5-div-x56.c",
3253 "src/f32-vsigmoid/gen/vsigmoid-avx2-rr1-p5-div-x64.c",
3254 "src/f32-vsigmoid/gen/vsigmoid-avx2-rr1-p5-div-x72.c",
3255 "src/f32-vsigmoid/gen/vsigmoid-avx2-rr1-p5-div-x80.c",
3256 "src/f32-vsigmoid/gen/vsigmoid-avx2-rr1-p5-nr1fma-x8.c",
3257 "src/f32-vsigmoid/gen/vsigmoid-avx2-rr1-p5-nr1fma-x16.c",
3258 "src/f32-vsigmoid/gen/vsigmoid-avx2-rr1-p5-nr1fma-x24.c",
3259 "src/f32-vsigmoid/gen/vsigmoid-avx2-rr1-p5-nr1fma-x32.c",
3260 "src/f32-vsigmoid/gen/vsigmoid-avx2-rr1-p5-nr1fma-x40.c",
3261 "src/f32-vsigmoid/gen/vsigmoid-avx2-rr1-p5-nr1fma-x48.c",
3262 "src/f32-vsigmoid/gen/vsigmoid-avx2-rr1-p5-nr1fma-x56.c",
3263 "src/f32-vsigmoid/gen/vsigmoid-avx2-rr1-p5-nr1fma-x64.c",
3264 "src/f32-vsigmoid/gen/vsigmoid-avx2-rr1-p5-nr1fma-x72.c",
3265 "src/f32-vsigmoid/gen/vsigmoid-avx2-rr1-p5-nr1fma-x80.c",
3266 "src/f32-vsigmoid/gen/vsigmoid-avx2-rr1-p5-nr2fma-x8.c",
3267 "src/f32-vsigmoid/gen/vsigmoid-avx2-rr1-p5-nr2fma-x16.c",
3268 "src/f32-vsigmoid/gen/vsigmoid-avx2-rr1-p5-nr2fma-x24.c",
3269 "src/f32-vsigmoid/gen/vsigmoid-avx2-rr1-p5-nr2fma-x32.c",
3270 "src/f32-vsigmoid/gen/vsigmoid-avx2-rr1-p5-nr2fma-x40.c",
3271 "src/f32-vsigmoid/gen/vsigmoid-avx2-rr1-p5-nr2fma-x48.c",
3272 "src/f32-vsigmoid/gen/vsigmoid-avx2-rr1-p5-nr2fma-x56.c",
3273 "src/f32-vsigmoid/gen/vsigmoid-avx2-rr1-p5-nr2fma-x64.c",
3274 "src/f32-vsigmoid/gen/vsigmoid-avx2-rr1-p5-nr2fma-x72.c",
3275 "src/f32-vsigmoid/gen/vsigmoid-avx2-rr1-p5-nr2fma-x80.c",
Marat Dukhanb7633f22020-11-20 16:34:56 -08003276 "src/math/exp-avx2-rr2-lut8-p3-perm.c",
3277 "src/math/exp-avx2-rr2-lut8-p4-perm.c",
3278 "src/math/exp-avx2-rr2-p5.c",
Marat Dukhande390d42020-11-29 19:32:18 -08003279 "src/math/expm1minus-avx2-rr1-lut4-p4-perm.c",
3280 "src/math/expm1minus-avx2-rr1-lut8-p4-perm.c",
3281 "src/math/expm1minus-avx2-rr1-lut16-p3-gather.c",
3282 "src/math/expm1minus-avx2-rr1-p6.c",
Frank Barcharde7223ee2020-12-04 19:04:01 -08003283 "src/math/expminus-avx2-rr2-p5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003284 "src/math/extexp-avx2-p5.c",
3285 "src/math/sigmoid-avx2-rr1-lut64-p2-gather-div.c",
3286 "src/math/sigmoid-avx2-rr1-lut64-p2-gather-nr1fma.c",
3287 "src/math/sigmoid-avx2-rr1-lut64-p2-gather-nr2fma.c",
3288 "src/math/sigmoid-avx2-rr1-lut64-p2-gather-nr2fma1adj.c",
3289 "src/math/sigmoid-avx2-rr1-p5-div.c",
3290 "src/math/sigmoid-avx2-rr1-p5-nr1fma.c",
3291 "src/math/sigmoid-avx2-rr1-p5-nr2fma.c",
3292 "src/math/sigmoid-avx2-rr2-lut64-p2-gather-div.c",
3293 "src/math/sigmoid-avx2-rr2-lut64-p2-gather-nr1fma.c",
3294 "src/math/sigmoid-avx2-rr2-lut64-p2-gather-nr2fma.c",
3295 "src/math/sigmoid-avx2-rr2-lut64-p2-gather-nr2fma1adj.c",
3296 "src/math/sigmoid-avx2-rr2-p5-div.c",
3297 "src/math/sigmoid-avx2-rr2-p5-nr1fma.c",
3298 "src/math/sigmoid-avx2-rr2-p5-nr2fma.c",
Marat Dukhand65a1522020-08-04 19:28:18 -07003299 "src/qs8-dwconv/gen/up8x9-minmax-avx2-mul32.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003300 "src/qs8-dwconv/gen/up16x9-minmax-avx2-mul16.c",
Marat Dukhand65a1522020-08-04 19:28:18 -07003301 "src/qs8-dwconv/gen/up16x9-minmax-avx2-mul32.c",
3302 "src/qs8-dwconv/gen/up24x9-minmax-avx2-mul32.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003303 "src/qs8-dwconv/gen/up32x9-minmax-avx2-mul16.c",
Marat Dukhand65a1522020-08-04 19:28:18 -07003304 "src/qs8-dwconv/gen/up32x9-minmax-avx2-mul32.c",
Marat Dukhan12809522020-08-02 22:23:51 -07003305 "src/qs8-gemm/gen/1x8c8-minmax-avx2.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07003306 "src/qs8-gemm/gen/1x8c8-xw-minmax-avx2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003307 "src/qs8-gemm/gen/2x8c8-minmax-avx2.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07003308 "src/qs8-gemm/gen/2x8c8-xw-minmax-avx2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003309 "src/qs8-gemm/gen/3x8c8-minmax-avx2.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07003310 "src/qs8-gemm/gen/3x8c8-xw-minmax-avx2.c",
Marat Dukhan12809522020-08-02 22:23:51 -07003311 "src/qs8-igemm/gen/1x8c8-minmax-avx2.c",
3312 "src/qs8-igemm/gen/2x8c8-minmax-avx2.c",
3313 "src/qs8-igemm/gen/3x8c8-minmax-avx2.c",
Marat Dukhane6dc0b62020-09-08 23:57:14 -07003314 "src/qs8-vadd/gen/minmax-avx2-mul32-ld64-x8.c",
3315 "src/qs8-vadd/gen/minmax-avx2-mul32-ld64-x16.c",
3316 "src/qs8-vadd/gen/minmax-avx2-mul32-ld64-x24.c",
3317 "src/qs8-vadd/gen/minmax-avx2-mul32-ld64-x32.c",
3318 "src/qs8-vaddc/gen/minmax-avx2-mul32-ld64-x8.c",
3319 "src/qs8-vaddc/gen/minmax-avx2-mul32-ld64-x16.c",
3320 "src/qs8-vaddc/gen/minmax-avx2-mul32-ld64-x24.c",
3321 "src/qs8-vaddc/gen/minmax-avx2-mul32-ld64-x32.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003322]
3323
Marat Dukhan08c4a432019-10-03 09:29:21 -07003324AVX512F_UKERNELS = [
Marat Dukhan1c587112020-04-08 20:04:28 -07003325 "src/f32-dwconv/gen/up16x4-minmax-avx512f-acc2.c",
3326 "src/f32-dwconv/gen/up16x4-minmax-avx512f.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07003327 "src/f32-dwconv/gen/up16x9-minmax-avx512f-acc2.c",
3328 "src/f32-dwconv/gen/up16x9-minmax-avx512f.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07003329 "src/f32-dwconv/gen/up16x25-minmax-avx512f-acc2.c",
3330 "src/f32-dwconv/gen/up16x25-minmax-avx512f.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003331 "src/f32-dwconv/gen/up32x4-minmax-avx512f-acc2.c",
3332 "src/f32-dwconv/gen/up32x4-minmax-avx512f.c",
3333 "src/f32-dwconv/gen/up32x9-minmax-avx512f-acc2.c",
3334 "src/f32-dwconv/gen/up32x9-minmax-avx512f.c",
3335 "src/f32-dwconv/gen/up32x25-minmax-avx512f-acc2.c",
3336 "src/f32-dwconv/gen/up32x25-minmax-avx512f.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07003337 "src/f32-gemm/gen-inc/1x16inc-minmax-avx512f-broadcast.c",
3338 "src/f32-gemm/gen-inc/4x16inc-minmax-avx512f-broadcast.c",
3339 "src/f32-gemm/gen-inc/5x16inc-minmax-avx512f-broadcast.c",
3340 "src/f32-gemm/gen-inc/6x16inc-minmax-avx512f-broadcast.c",
3341 "src/f32-gemm/gen-inc/7x16inc-minmax-avx512f-broadcast.c",
3342 "src/f32-gemm/gen-inc/8x16inc-minmax-avx512f-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003343 "src/f32-gemm/gen/1x16-minmax-avx512f-broadcast.c",
3344 "src/f32-gemm/gen/4x16-minmax-avx512f-broadcast.c",
3345 "src/f32-gemm/gen/5x16-minmax-avx512f-broadcast.c",
3346 "src/f32-gemm/gen/6x16-minmax-avx512f-broadcast.c",
3347 "src/f32-gemm/gen/7x16-minmax-avx512f-broadcast.c",
3348 "src/f32-gemm/gen/8x16-minmax-avx512f-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07003349 "src/f32-igemm/gen/1x16-minmax-avx512f-broadcast.c",
3350 "src/f32-igemm/gen/4x16-minmax-avx512f-broadcast.c",
3351 "src/f32-igemm/gen/5x16-minmax-avx512f-broadcast.c",
3352 "src/f32-igemm/gen/6x16-minmax-avx512f-broadcast.c",
3353 "src/f32-igemm/gen/7x16-minmax-avx512f-broadcast.c",
3354 "src/f32-igemm/gen/8x16-minmax-avx512f-broadcast.c",
Marat Dukhan90eca0a2020-03-11 00:52:23 -07003355 "src/f32-prelu/gen/avx512f-2x16.c",
3356 "src/f32-prelu/gen/avx512f-2x32.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003357 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128-acc2.c",
3358 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003359 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003360 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x144-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003361 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x144.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003362 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160-acc2.c",
3363 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003364 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003365 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc2.c",
3366 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc3.c",
3367 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc6.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003368 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003369 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128-acc2.c",
3370 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003371 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003372 "src/f32-raddextexp/gen/avx512f-p5-scalef-x144-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003373 "src/f32-raddextexp/gen/avx512f-p5-scalef-x144.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003374 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160-acc2.c",
3375 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003376 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003377 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc2.c",
3378 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc3.c",
3379 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc6.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003380 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003381 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128-acc2.c",
3382 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003383 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003384 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x144-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003385 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x144.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003386 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160-acc2.c",
3387 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003388 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003389 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc2.c",
3390 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc3.c",
3391 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc6.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003392 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003393 "src/f32-rmax/avx512f.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07003394 "src/f32-vbinary/gen/vadd-minmax-avx512f-x16.c",
3395 "src/f32-vbinary/gen/vadd-minmax-avx512f-x32.c",
3396 "src/f32-vbinary/gen/vaddc-minmax-avx512f-x16.c",
3397 "src/f32-vbinary/gen/vaddc-minmax-avx512f-x32.c",
3398 "src/f32-vbinary/gen/vdiv-minmax-avx512f-x16.c",
3399 "src/f32-vbinary/gen/vdiv-minmax-avx512f-x32.c",
3400 "src/f32-vbinary/gen/vdivc-minmax-avx512f-x16.c",
3401 "src/f32-vbinary/gen/vdivc-minmax-avx512f-x32.c",
Marat Dukhan9a88efe2019-12-10 15:54:24 -08003402 "src/f32-vbinary/gen/vmax-avx512f-x16.c",
3403 "src/f32-vbinary/gen/vmax-avx512f-x32.c",
3404 "src/f32-vbinary/gen/vmaxc-avx512f-x16.c",
3405 "src/f32-vbinary/gen/vmaxc-avx512f-x32.c",
3406 "src/f32-vbinary/gen/vmin-avx512f-x16.c",
3407 "src/f32-vbinary/gen/vmin-avx512f-x32.c",
3408 "src/f32-vbinary/gen/vminc-avx512f-x16.c",
3409 "src/f32-vbinary/gen/vminc-avx512f-x32.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07003410 "src/f32-vbinary/gen/vmul-minmax-avx512f-x16.c",
3411 "src/f32-vbinary/gen/vmul-minmax-avx512f-x32.c",
3412 "src/f32-vbinary/gen/vmulc-minmax-avx512f-x16.c",
3413 "src/f32-vbinary/gen/vmulc-minmax-avx512f-x32.c",
3414 "src/f32-vbinary/gen/vrdivc-minmax-avx512f-x16.c",
3415 "src/f32-vbinary/gen/vrdivc-minmax-avx512f-x32.c",
3416 "src/f32-vbinary/gen/vrsubc-minmax-avx512f-x16.c",
3417 "src/f32-vbinary/gen/vrsubc-minmax-avx512f-x32.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07003418 "src/f32-vbinary/gen/vsqrdiff-avx512f-x16.c",
3419 "src/f32-vbinary/gen/vsqrdiff-avx512f-x32.c",
3420 "src/f32-vbinary/gen/vsqrdiffc-avx512f-x16.c",
3421 "src/f32-vbinary/gen/vsqrdiffc-avx512f-x32.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07003422 "src/f32-vbinary/gen/vsub-minmax-avx512f-x16.c",
3423 "src/f32-vbinary/gen/vsub-minmax-avx512f-x32.c",
3424 "src/f32-vbinary/gen/vsubc-minmax-avx512f-x16.c",
3425 "src/f32-vbinary/gen/vsubc-minmax-avx512f-x32.c",
Marat Dukhanb1eec082021-05-05 23:24:55 -07003426 "src/f32-vclamp/gen/vclamp-avx512f-x16.c",
3427 "src/f32-vclamp/gen/vclamp-avx512f-x32.c",
Marat Dukhaned6baaf2020-12-01 15:07:08 -08003428 "src/f32-velu/gen/velu-avx512f-rr1-lut16-p3-perm-x16.c",
3429 "src/f32-velu/gen/velu-avx512f-rr1-lut16-p3-perm-x32.c",
3430 "src/f32-velu/gen/velu-avx512f-rr1-lut16-p3-perm-x48.c",
3431 "src/f32-velu/gen/velu-avx512f-rr1-lut16-p3-perm-x64.c",
3432 "src/f32-velu/gen/velu-avx512f-rr1-lut16-p3-perm-x80.c",
3433 "src/f32-velu/gen/velu-avx512f-rr1-lut16-p3-perm-x96.c",
3434 "src/f32-velu/gen/velu-avx512f-rr1-lut16-p3-perm-x112.c",
3435 "src/f32-velu/gen/velu-avx512f-rr1-lut16-p3-perm-x128.c",
3436 "src/f32-velu/gen/velu-avx512f-rr1-p6-x16.c",
3437 "src/f32-velu/gen/velu-avx512f-rr1-p6-x32.c",
3438 "src/f32-velu/gen/velu-avx512f-rr1-p6-x48.c",
3439 "src/f32-velu/gen/velu-avx512f-rr1-p6-x64.c",
3440 "src/f32-velu/gen/velu-avx512f-rr1-p6-x80.c",
3441 "src/f32-velu/gen/velu-avx512f-rr1-p6-x96.c",
3442 "src/f32-velu/gen/velu-avx512f-rr1-p6-x112.c",
3443 "src/f32-velu/gen/velu-avx512f-rr1-p6-x128.c",
Marat Dukhan6674d692021-05-05 22:27:00 -07003444 "src/f32-vhswish/gen/vhswish-avx512f-x16.c",
3445 "src/f32-vhswish/gen/vhswish-avx512f-x32.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07003446 "src/f32-vlrelu/gen/vlrelu-avx512f-x16.c",
3447 "src/f32-vlrelu/gen/vlrelu-avx512f-x32.c",
Marat Dukhanb1eec082021-05-05 23:24:55 -07003448 "src/f32-vrelu/gen/vrelu-avx512f-x16.c",
3449 "src/f32-vrelu/gen/vrelu-avx512f-x32.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003450 "src/f32-vrnd/gen/vrndd-avx512f-x16.c",
3451 "src/f32-vrnd/gen/vrndd-avx512f-x32.c",
3452 "src/f32-vrnd/gen/vrndne-avx512f-x16.c",
3453 "src/f32-vrnd/gen/vrndne-avx512f-x32.c",
3454 "src/f32-vrnd/gen/vrndu-avx512f-x16.c",
3455 "src/f32-vrnd/gen/vrndu-avx512f-x32.c",
3456 "src/f32-vrnd/gen/vrndz-avx512f-x16.c",
3457 "src/f32-vrnd/gen/vrndz-avx512f-x32.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07003458 "src/f32-vscale/avx512f-x64.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003459 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x16.c",
3460 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x32.c",
3461 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x48.c",
3462 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x64.c",
3463 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x80.c",
3464 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x96.c",
3465 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x112.c",
3466 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x128.c",
3467 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x144.c",
3468 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x160.c",
3469 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x176.c",
3470 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x192.c",
3471 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x16.c",
3472 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x32.c",
3473 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x48.c",
3474 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x64.c",
3475 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x80.c",
3476 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x96.c",
3477 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x112.c",
3478 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x128.c",
3479 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x144.c",
3480 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x160.c",
3481 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x176.c",
3482 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x192.c",
Marat Dukhanb1eec082021-05-05 23:24:55 -07003483 "src/f32-vsigmoid/gen/vsigmoid-avx512f-rr1-lut16-p3-perm-scalef-div-x16.c",
3484 "src/f32-vsigmoid/gen/vsigmoid-avx512f-rr1-lut16-p3-perm-scalef-div-x32.c",
3485 "src/f32-vsigmoid/gen/vsigmoid-avx512f-rr1-lut16-p3-perm-scalef-div-x48.c",
3486 "src/f32-vsigmoid/gen/vsigmoid-avx512f-rr1-lut16-p3-perm-scalef-div-x64.c",
3487 "src/f32-vsigmoid/gen/vsigmoid-avx512f-rr1-lut16-p3-perm-scalef-div-x80.c",
3488 "src/f32-vsigmoid/gen/vsigmoid-avx512f-rr1-lut16-p3-perm-scalef-div-x96.c",
3489 "src/f32-vsigmoid/gen/vsigmoid-avx512f-rr1-lut16-p3-perm-scalef-div-x112.c",
3490 "src/f32-vsigmoid/gen/vsigmoid-avx512f-rr1-lut16-p3-perm-scalef-div-x128.c",
3491 "src/f32-vsigmoid/gen/vsigmoid-avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x16.c",
3492 "src/f32-vsigmoid/gen/vsigmoid-avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x32.c",
3493 "src/f32-vsigmoid/gen/vsigmoid-avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x48.c",
3494 "src/f32-vsigmoid/gen/vsigmoid-avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x64.c",
3495 "src/f32-vsigmoid/gen/vsigmoid-avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x80.c",
3496 "src/f32-vsigmoid/gen/vsigmoid-avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x96.c",
3497 "src/f32-vsigmoid/gen/vsigmoid-avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x112.c",
3498 "src/f32-vsigmoid/gen/vsigmoid-avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x128.c",
3499 "src/f32-vsigmoid/gen/vsigmoid-avx512f-rr1-p5-scalef-div-x16.c",
3500 "src/f32-vsigmoid/gen/vsigmoid-avx512f-rr1-p5-scalef-div-x32.c",
3501 "src/f32-vsigmoid/gen/vsigmoid-avx512f-rr1-p5-scalef-div-x48.c",
3502 "src/f32-vsigmoid/gen/vsigmoid-avx512f-rr1-p5-scalef-div-x64.c",
3503 "src/f32-vsigmoid/gen/vsigmoid-avx512f-rr1-p5-scalef-div-x80.c",
3504 "src/f32-vsigmoid/gen/vsigmoid-avx512f-rr1-p5-scalef-div-x96.c",
3505 "src/f32-vsigmoid/gen/vsigmoid-avx512f-rr1-p5-scalef-div-x112.c",
3506 "src/f32-vsigmoid/gen/vsigmoid-avx512f-rr1-p5-scalef-div-x128.c",
3507 "src/f32-vsigmoid/gen/vsigmoid-avx512f-rr1-p5-scalef-nr1fma-x16.c",
3508 "src/f32-vsigmoid/gen/vsigmoid-avx512f-rr1-p5-scalef-nr1fma-x32.c",
3509 "src/f32-vsigmoid/gen/vsigmoid-avx512f-rr1-p5-scalef-nr1fma-x48.c",
3510 "src/f32-vsigmoid/gen/vsigmoid-avx512f-rr1-p5-scalef-nr1fma-x64.c",
3511 "src/f32-vsigmoid/gen/vsigmoid-avx512f-rr1-p5-scalef-nr1fma-x80.c",
3512 "src/f32-vsigmoid/gen/vsigmoid-avx512f-rr1-p5-scalef-nr1fma-x96.c",
3513 "src/f32-vsigmoid/gen/vsigmoid-avx512f-rr1-p5-scalef-nr1fma-x112.c",
3514 "src/f32-vsigmoid/gen/vsigmoid-avx512f-rr1-p5-scalef-nr1fma-x128.c",
3515 "src/f32-vsigmoid/gen/vsigmoid-avx512f-rr2-lut32-p2-perm2-scalef-div-x16.c",
3516 "src/f32-vsigmoid/gen/vsigmoid-avx512f-rr2-lut32-p2-perm2-scalef-div-x32.c",
3517 "src/f32-vsigmoid/gen/vsigmoid-avx512f-rr2-lut32-p2-perm2-scalef-div-x48.c",
3518 "src/f32-vsigmoid/gen/vsigmoid-avx512f-rr2-lut32-p2-perm2-scalef-div-x64.c",
3519 "src/f32-vsigmoid/gen/vsigmoid-avx512f-rr2-lut32-p2-perm2-scalef-div-x80.c",
3520 "src/f32-vsigmoid/gen/vsigmoid-avx512f-rr2-lut32-p2-perm2-scalef-div-x96.c",
3521 "src/f32-vsigmoid/gen/vsigmoid-avx512f-rr2-lut32-p2-perm2-scalef-div-x112.c",
3522 "src/f32-vsigmoid/gen/vsigmoid-avx512f-rr2-lut32-p2-perm2-scalef-div-x128.c",
3523 "src/f32-vsigmoid/gen/vsigmoid-avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x16.c",
3524 "src/f32-vsigmoid/gen/vsigmoid-avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x32.c",
3525 "src/f32-vsigmoid/gen/vsigmoid-avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x48.c",
3526 "src/f32-vsigmoid/gen/vsigmoid-avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x64.c",
3527 "src/f32-vsigmoid/gen/vsigmoid-avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x80.c",
3528 "src/f32-vsigmoid/gen/vsigmoid-avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x96.c",
3529 "src/f32-vsigmoid/gen/vsigmoid-avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x112.c",
3530 "src/f32-vsigmoid/gen/vsigmoid-avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x128.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07003531 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x16.c",
3532 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x32.c",
3533 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x48.c",
3534 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x64.c",
3535 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x80.c",
3536 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x96.c",
3537 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x112.c",
3538 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x128.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07003539 "src/f32-vunary/gen/vabs-avx512f-x16.c",
3540 "src/f32-vunary/gen/vabs-avx512f-x32.c",
3541 "src/f32-vunary/gen/vneg-avx512f-x16.c",
3542 "src/f32-vunary/gen/vneg-avx512f-x32.c",
3543 "src/f32-vunary/gen/vsqr-avx512f-x16.c",
3544 "src/f32-vunary/gen/vsqr-avx512f-x32.c",
Marat Dukhanb7633f22020-11-20 16:34:56 -08003545 "src/math/exp-avx512f-rr2-lut16-p3-perm-scalef.c",
3546 "src/math/exp-avx512f-rr2-lut16-p3-perm.c",
3547 "src/math/exp-avx512f-rr2-lut32-p2-perm2-scalef.c",
3548 "src/math/exp-avx512f-rr2-lut32-p2-perm2.c",
3549 "src/math/exp-avx512f-rr2-p5-scalef.c",
3550 "src/math/exp-avx512f-rr2-p5.c",
Marat Dukhande390d42020-11-29 19:32:18 -08003551 "src/math/expm1minus-avx512f-rr1-lut16-p3-perm.c",
3552 "src/math/expm1minus-avx512f-rr1-p6.c",
Marat Dukhan98ba4412019-10-23 02:14:28 -07003553 "src/math/extexp-avx512f-p5.c",
Marat Dukhan6a34c5f2020-09-22 21:44:15 -07003554 "src/math/sigmoid-avx512f-rr1-lut16-p3-perm-scalef-div.c",
Marat Dukhan6a34c5f2020-09-22 21:44:15 -07003555 "src/math/sigmoid-avx512f-rr1-lut16-p3-perm-scalef-nr1fma.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003556 "src/math/sigmoid-avx512f-rr1-lut16-p3-perm-scalef-nr1fma1adj.c",
Marat Dukhan6a34c5f2020-09-22 21:44:15 -07003557 "src/math/sigmoid-avx512f-rr1-lut32-p2-perm2-scalef-div.c",
Marat Dukhan6a34c5f2020-09-22 21:44:15 -07003558 "src/math/sigmoid-avx512f-rr1-lut32-p2-perm2-scalef-nr1fma.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003559 "src/math/sigmoid-avx512f-rr1-lut32-p2-perm2-scalef-nr1fma1adj.c",
Marat Dukhan36173d22020-10-15 17:14:26 -07003560 "src/math/sigmoid-avx512f-rr1-lut64-p2-gather-scalef-div.c",
Marat Dukhan36173d22020-10-15 17:14:26 -07003561 "src/math/sigmoid-avx512f-rr1-lut64-p2-gather-scalef-nr1fma.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003562 "src/math/sigmoid-avx512f-rr1-lut64-p2-gather-scalef-nr1fma1adj.c",
3563 "src/math/sigmoid-avx512f-rr1-p5-scalef-div.c",
3564 "src/math/sigmoid-avx512f-rr1-p5-scalef-nr1fma.c",
3565 "src/math/sigmoid-avx512f-rr1-p5-scalef-nr1fma1adj.c",
3566 "src/math/sigmoid-avx512f-rr2-lut16-p3-perm-scalef-div.c",
3567 "src/math/sigmoid-avx512f-rr2-lut16-p3-perm-scalef-nr1fma.c",
3568 "src/math/sigmoid-avx512f-rr2-lut16-p3-perm-scalef-nr1fma1adj.c",
3569 "src/math/sigmoid-avx512f-rr2-lut32-p2-perm2-scalef-div.c",
3570 "src/math/sigmoid-avx512f-rr2-lut32-p2-perm2-scalef-nr1fma.c",
3571 "src/math/sigmoid-avx512f-rr2-lut32-p2-perm2-scalef-nr1fma1adj.c",
Marat Dukhan36173d22020-10-15 17:14:26 -07003572 "src/math/sigmoid-avx512f-rr2-lut64-p2-gather-scalef-div.c",
Marat Dukhan36173d22020-10-15 17:14:26 -07003573 "src/math/sigmoid-avx512f-rr2-lut64-p2-gather-scalef-nr1fma.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003574 "src/math/sigmoid-avx512f-rr2-lut64-p2-gather-scalef-nr1fma1adj.c",
3575 "src/math/sigmoid-avx512f-rr2-p5-scalef-div.c",
3576 "src/math/sigmoid-avx512f-rr2-p5-scalef-nr1fma.c",
3577 "src/math/sigmoid-avx512f-rr2-p5-scalef-nr1fma1adj.c",
Marat Dukhan84000762020-06-29 18:38:43 -07003578 "src/math/sqrt-avx512f-nr1fma.c",
Marat Dukhan84000762020-06-29 18:38:43 -07003579 "src/math/sqrt-avx512f-nr1fma1adj.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003580 "src/math/sqrt-avx512f-nr2fma.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003581]
3582
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07003583AVX512SKX_UKERNELS = [
Marat Dukhan2ffc5e62020-09-06 22:33:38 -07003584 "src/qs8-dwconv/gen/up16x9-minmax-avx512skx-mul32.c",
3585 "src/qs8-dwconv/gen/up32x9-minmax-avx512skx-mul32.c",
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07003586 "src/qs8-gemm/gen/1x16c8-minmax-avx512skx.c",
3587 "src/qs8-gemm/gen/2x16c8-minmax-avx512skx.c",
3588 "src/qs8-gemm/gen/3x16c8-minmax-avx512skx.c",
3589 "src/qs8-gemm/gen/4x16c8-minmax-avx512skx.c",
3590 "src/qs8-igemm/gen/1x16c8-minmax-avx512skx.c",
3591 "src/qs8-igemm/gen/2x16c8-minmax-avx512skx.c",
3592 "src/qs8-igemm/gen/3x16c8-minmax-avx512skx.c",
3593 "src/qs8-igemm/gen/4x16c8-minmax-avx512skx.c",
3594]
3595
Frank Barchardbcedc082020-08-17 18:00:51 -07003596WASM32_ASM_UKERNELS = [
Marat Dukhan6674d692021-05-05 22:27:00 -07003597 "src/f32-vrelu/wasm_shr_x1.S",
3598 "src/f32-vrelu/wasm_shr_x2.S",
3599 "src/f32-vrelu/wasm_shr_x4.S",
Frank Barchardbcedc082020-08-17 18:00:51 -07003600]
3601
Marat Dukhan08c4a432019-10-03 09:29:21 -07003602AARCH32_ASM_UKERNELS = [
Marat Dukhan32f93812020-05-17 20:31:21 -07003603 "src/f32-gemm/4x4-aarch32-vfp-ld64.S",
Marat Dukhan3b98f6b2020-05-17 10:09:22 -07003604 "src/f32-gemm/4x4-minmax-aarch32-vfp-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003605 "src/f32-gemm/4x8-minmax-aarch32-neon-cortex-a53.S",
3606 "src/f32-gemm/4x8-minmax-aarch32-neon-cortex-a55.S",
Frank Barchard490febe2020-07-16 18:42:17 -07003607 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-cortex-a7.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003608 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-cortex-a75.S",
Frank Barchard569561d2020-06-17 13:11:12 -07003609 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-ld64.S",
Frank Barchard490febe2020-07-16 18:42:17 -07003610 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-pld-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003611 "src/f32-igemm/4x8-minmax-aarch32-neon-cortex-a53.S",
3612 "src/f32-igemm/4x8-minmax-aarch32-neon-cortex-a55.S",
Frank Barchard490febe2020-07-16 18:42:17 -07003613 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-cortex-a7.S",
3614 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-cortex-a75.S",
3615 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-ld64.S",
3616 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-pld-cortex-a75.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003617]
3618
3619AARCH64_ASM_UKERNELS = [
Frank Barchardbddfbcd2020-04-15 12:32:41 -07003620 "src/f16-gemm/gen-inc/1x8inc-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchard04336c12020-10-22 16:48:55 -07003621 "src/f16-gemm/gen-inc/1x16inc-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchardbddfbcd2020-04-15 12:32:41 -07003622 "src/f16-gemm/gen-inc/4x8inc-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchard04336c12020-10-22 16:48:55 -07003623 "src/f16-gemm/gen-inc/4x16inc-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchardbddfbcd2020-04-15 12:32:41 -07003624 "src/f16-gemm/gen-inc/6x8inc-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchard04336c12020-10-22 16:48:55 -07003625 "src/f16-gemm/gen-inc/6x16inc-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchard3b8e5662020-04-20 12:12:53 -07003626 "src/f16-gemm/gen-inc/8x8inc-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchard04336c12020-10-22 16:48:55 -07003627 "src/f16-gemm/gen/1x8-minmax-aarch64-neonfp16arith-ld64.S",
3628 "src/f16-gemm/gen/1x16-minmax-aarch64-neonfp16arith-ld32.S",
3629 "src/f16-gemm/gen/4x8-minmax-aarch64-neonfp16arith-ld64.S",
3630 "src/f16-gemm/gen/4x16-minmax-aarch64-neonfp16arith-ld32.S",
3631 "src/f16-gemm/gen/6x8-minmax-aarch64-neonfp16arith-ld64.S",
3632 "src/f16-gemm/gen/6x16-minmax-aarch64-neonfp16arith-ld32.S",
3633 "src/f16-gemm/gen/8x8-minmax-aarch64-neonfp16arith-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003634 "src/f32-dwconv/up4x9-minmax-aarch64-neonfma-cortex-a55.S",
3635 "src/f32-dwconv/up4x9-minmax-aarch64-neonfma.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003636 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a53.S",
3637 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a57.S",
3638 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a75.S",
Frank Barchard04336c12020-10-22 16:48:55 -07003639 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-ld64.S",
3640 "src/f32-gemm/gen-inc/1x12inc-minmax-aarch64-neonfma-cortex-a53.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003641 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a53.S",
3642 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a55.S",
3643 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a57.S",
3644 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003645 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-ld64.S",
Frank Barchard04336c12020-10-22 16:48:55 -07003646 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-ld128.S",
3647 "src/f32-gemm/gen-inc/4x12inc-minmax-aarch64-neonfma-cortex-a53.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003648 "src/f32-gemm/gen-inc/5x8inc-minmax-aarch64-neonfma-cortex-a57.S",
3649 "src/f32-gemm/gen-inc/5x8inc-minmax-aarch64-neonfma-cortex-a75.S",
3650 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a53.S",
3651 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a55.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003652 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a57.S",
Frank Barchard04336c12020-10-22 16:48:55 -07003653 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a73.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003654 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003655 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-ld64.S",
Frank Barchard04336c12020-10-22 16:48:55 -07003656 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-ld128.S",
3657 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a53.S",
3658 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a57.S",
3659 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a75.S",
3660 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-ld64.S",
3661 "src/f32-gemm/gen/1x12-minmax-aarch64-neonfma-cortex-a53.S",
3662 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a53.S",
3663 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a55.S",
3664 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a57.S",
3665 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a75.S",
3666 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-ld64.S",
3667 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-ld128.S",
3668 "src/f32-gemm/gen/4x12-minmax-aarch64-neonfma-cortex-a53.S",
3669 "src/f32-gemm/gen/5x8-minmax-aarch64-neonfma-cortex-a57.S",
3670 "src/f32-gemm/gen/5x8-minmax-aarch64-neonfma-cortex-a75.S",
3671 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a53.S",
3672 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a55.S",
3673 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a57.S",
3674 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a73.S",
3675 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a75.S",
3676 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-ld64.S",
3677 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-ld128.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003678 "src/f32-igemm/1x8-minmax-aarch64-neonfma-cortex-a53.S",
Frank Barchard04336c12020-10-22 16:48:55 -07003679 "src/f32-igemm/1x12-minmax-aarch64-neonfma-cortex-a53.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003680 "src/f32-igemm/4x8-minmax-aarch64-neonfma-cortex-a53.S",
3681 "src/f32-igemm/4x8-minmax-aarch64-neonfma-cortex-a55.S",
Frank Barchard04336c12020-10-22 16:48:55 -07003682 "src/f32-igemm/4x12-minmax-aarch64-neonfma-cortex-a53.S",
3683 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a53.S",
3684 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a55.S",
3685 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a73.S",
3686 "src/f32-igemm/gen/1x8-minmax-aarch64-neonfma-cortex-a57.S",
3687 "src/f32-igemm/gen/1x8-minmax-aarch64-neonfma-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003688 "src/f32-igemm/gen/4x8-minmax-aarch64-neonfma-cortex-a57.S",
3689 "src/f32-igemm/gen/4x8-minmax-aarch64-neonfma-cortex-a75.S",
3690 "src/f32-igemm/gen/5x8-minmax-aarch64-neonfma-cortex-a57.S",
3691 "src/f32-igemm/gen/5x8-minmax-aarch64-neonfma-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003692 "src/f32-igemm/gen/6x8-minmax-aarch64-neonfma-cortex-a57.S",
3693 "src/f32-igemm/gen/6x8-minmax-aarch64-neonfma-cortex-a75.S",
Frank Barchard4a352042021-04-13 15:52:08 -07003694 "src/qs8-gemm/1x8c8-aarch64-neon-mlal-padal.S",
Frank Barchard4a4be4e2021-03-01 15:00:28 -08003695 "src/qs8-gemm/1x16c4-aarch64-neondot-ld32.S",
Frank Barchardf1fd89e2020-10-06 23:27:11 -07003696 "src/qs8-gemm/1x16c4-aarch64-neondot-ld64.S",
Frank Barchard1dc9fef2021-03-10 23:50:19 -08003697 "src/qs8-gemm/2x8c8-aarch64-neon-mull-padal.S",
Frank Barchard5655cb72021-03-09 20:57:14 -08003698 "src/qs8-gemm/2x8c16-aarch64-neon-mlal-padal.S",
Frank Barchard671d1b02021-03-10 15:42:08 -08003699 "src/qs8-gemm/4x16c4-aarch64-neondot-cortex-a55.S",
Frank Barchardda78da12021-03-02 14:28:00 -08003700 "src/qs8-gemm/4x16c4-aarch64-neondot-ld32.S",
Frank Barchard671d1b02021-03-10 15:42:08 -08003701 "src/qs8-gemm/4x16c4-aarch64-neondot-ld64.S",
Frank Barcharddddb38f2021-05-03 16:51:48 -07003702 "src/qs8-gemm/gen/1x8c8-aarch64-neon-mlal-padal-cortex-a53.S",
3703 "src/qs8-gemm/gen/1x8c8-aarch64-neon-mlal-padal-prfm-cortex-a53.S",
3704 "src/qs8-gemm/gen/2x8c8-aarch64-neon-mlal-padal-cortex-a53.S",
3705 "src/qs8-gemm/gen/2x8c8-aarch64-neon-mlal-padal-prfm-cortex-a53.S",
Frank Barchard4a352042021-04-13 15:52:08 -07003706 "src/qs8-gemm/gen/2x8c8-aarch64-neon-mlal-padal-prfm.S",
3707 "src/qs8-gemm/gen/2x8c8-aarch64-neon-mlal-padal.S",
Frank Barchard6e35de52021-03-25 04:33:21 -07003708 "src/qs8-igemm/1x8c8-aarch64-neon-mlal-padal.S",
Frank Barchard7ca54df2021-03-12 00:09:02 -08003709 "src/qs8-igemm/2x8c16-aarch64-neon-mlal-padal.S",
Frank Barchard89e12f82021-03-08 11:42:44 -08003710 "src/qs8-igemm/4x16c4-aarch64-neondot-cortex-a55.S",
Frank Barchard671d1b02021-03-10 15:42:08 -08003711 "src/qs8-igemm/4x16c4-aarch64-neondot-ld64.S",
Frank Barcharddddb38f2021-05-03 16:51:48 -07003712 "src/qs8-igemm/gen/1x8c8-aarch64-neon-mlal-padal-cortex-a53.S",
3713 "src/qs8-igemm/gen/1x8c8-aarch64-neon-mlal-padal-prfm-cortex-a53.S",
3714 "src/qs8-igemm/gen/2x8c8-aarch64-neon-mlal-padal-cortex-a53.S",
3715 "src/qs8-igemm/gen/2x8c8-aarch64-neon-mlal-padal-prfm-cortex-a53.S",
Frank Barchard4a352042021-04-13 15:52:08 -07003716 "src/qs8-igemm/gen/2x8c8-aarch64-neon-mlal-padal-prfm.S",
3717 "src/qs8-igemm/gen/2x8c8-aarch64-neon-mlal-padal.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003718]
3719
Marat Dukhan1b354632020-03-23 12:50:22 -07003720INTERNAL_MICROKERNEL_HDRS = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003721 "src/xnnpack/argmaxpool.h",
3722 "src/xnnpack/avgpool.h",
3723 "src/xnnpack/clamp.h",
3724 "src/xnnpack/common.h",
3725 "src/xnnpack/conv.h",
Yury Kartynnike7841862020-11-04 18:22:18 -08003726 "src/xnnpack/depthtospace.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003727 "src/xnnpack/dwconv.h",
Frank Barchard04336c12020-10-22 16:48:55 -07003728 "src/xnnpack/fill.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003729 "src/xnnpack/gavgpool.h",
3730 "src/xnnpack/gemm.h",
3731 "src/xnnpack/hswish.h",
Marat Dukhan660fd192020-03-10 04:55:30 -07003732 "src/xnnpack/ibilinear.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003733 "src/xnnpack/igemm.h",
Marat Dukhancfb31342019-12-05 10:42:57 -08003734 "src/xnnpack/intrinsics-polyfill.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003735 "src/xnnpack/lut.h",
3736 "src/xnnpack/math.h",
3737 "src/xnnpack/maxpool.h",
3738 "src/xnnpack/packx.h",
3739 "src/xnnpack/pad.h",
3740 "src/xnnpack/params.h",
3741 "src/xnnpack/pavgpool.h",
3742 "src/xnnpack/ppmm.h",
3743 "src/xnnpack/prelu.h",
Marat Dukhan97579532019-10-18 16:40:39 -07003744 "src/xnnpack/raddexpminusmax.h",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07003745 "src/xnnpack/raddextexp.h",
Marat Dukhan97579532019-10-18 16:40:39 -07003746 "src/xnnpack/raddstoreexpminusmax.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003747 "src/xnnpack/rmax.h",
3748 "src/xnnpack/scalar-utils.h",
3749 "src/xnnpack/spmm.h",
3750 "src/xnnpack/unpool.h",
3751 "src/xnnpack/vadd.h",
Marat Dukhan1e782c42019-11-21 17:02:40 -08003752 "src/xnnpack/vbinary.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003753 "src/xnnpack/vmulcaddc.h",
Marat Dukhan05ac8e32019-10-21 15:39:33 -07003754 "src/xnnpack/vscale.h",
Marat Dukhan97579532019-10-18 16:40:39 -07003755 "src/xnnpack/vscaleexpminusmax.h",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07003756 "src/xnnpack/vscaleextexp.h",
Marat Dukhan1e782c42019-11-21 17:02:40 -08003757 "src/xnnpack/vunary.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003758 "src/xnnpack/zip.h",
Marat Dukhan1b354632020-03-23 12:50:22 -07003759]
3760
3761INTERNAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003762 "include/xnnpack.h",
3763 "src/xnnpack/allocator.h",
3764 "src/xnnpack/compute.h",
3765 "src/xnnpack/im2col.h",
3766 "src/xnnpack/indirection.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003767 "src/xnnpack/math-stubs.h",
Chao Mei6ddfc602020-05-13 22:29:36 -07003768 "src/xnnpack/memory-planner.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003769 "src/xnnpack/operator.h",
3770 "src/xnnpack/pack.h",
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07003771 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003772 "src/xnnpack/requantization-stubs.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003773 "src/xnnpack/requantization.h",
Marat Dukhan1d75a542020-02-03 12:23:01 -08003774 "src/xnnpack/subgraph.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003775]
3776
Marat Dukhan1b354632020-03-23 12:50:22 -07003777ACCURACY_EVAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003778 "src/xnnpack/math-stubs.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003779]
3780
Marat Dukhan1b354632020-03-23 12:50:22 -07003781MICROKERNEL_BENCHMARK_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003782 "include/xnnpack.h",
Frank Barchard35db7d02020-10-26 13:37:34 -07003783 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003784]
3785
Marat Dukhan1b354632020-03-23 12:50:22 -07003786MICROKERNEL_TEST_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Frank Barchard35db7d02020-10-26 13:37:34 -07003787 "include/xnnpack.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003788 "src/xnnpack/isa-checks.h",
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07003789 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003790 "src/xnnpack/requantization.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003791]
3792
3793OPERATOR_TEST_PARAMS_HDRS = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003794 "src/xnnpack/common.h",
Frank Barchard04336c12020-10-22 16:48:55 -07003795 "src/xnnpack/params.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003796]
3797
3798WEIGHTS_PACK_HDRS = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003799 "src/xnnpack/compute.h",
Frank Barchard04336c12020-10-22 16:48:55 -07003800 "src/xnnpack/operator.h",
3801 "src/xnnpack/pack.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003802]
3803
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07003804LOGGING_COPTS = select({
3805 # No logging in optimized mode
3806 ":optimized_build": ["-DXNN_LOG_LEVEL=0"],
3807 # Full logging in debug mode
3808 ":debug_build": ["-DXNN_LOG_LEVEL=5"],
3809 # Error-only logging in default (fastbuild) mode
3810 "//conditions:default": ["-DXNN_LOG_LEVEL=2"],
3811})
3812
Marat Dukhan3b59de22020-06-03 20:15:19 -07003813LOGGING_SRCS = select({
3814 # No logging in optimized mode
3815 ":optimized_build": [],
3816 "//conditions:default": [
Marat Dukhanccd3a1d2021-03-29 16:03:12 -07003817 "src/datatype-strings.c",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003818 "src/operator-strings.c",
3819 "src/subgraph-strings.c",
3820 ],
3821})
3822
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07003823LOGGING_HDRS = [
3824 "src/xnnpack/log.h",
3825]
3826
Marat Dukhan08c4a432019-10-03 09:29:21 -07003827xnnpack_cc_library(
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003828 name = "tables",
3829 srcs = TABLE_SRCS,
3830 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003831 gcc_copts = xnnpack_gcc_std_copts(),
3832 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003833)
3834
3835xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003836 name = "scalar_ukernels",
3837 srcs = SCALAR_UKERNELS,
3838 hdrs = INTERNAL_HDRS,
3839 aarch32_copts = ["-marm"],
Marat Dukhan10a38082020-04-17 03:58:35 -07003840 gcc_copts = xnnpack_gcc_std_copts(),
3841 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003842 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003843 ":tables",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003844 "@FP16",
3845 "@FXdiv",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003846 "@pthreadpool",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003847 ],
3848)
3849
3850xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003851 name = "scalar_ukernels_test_mode",
3852 srcs = SCALAR_UKERNELS,
3853 hdrs = INTERNAL_HDRS,
3854 aarch32_copts = ["-marm"],
3855 copts = [
3856 "-UNDEBUG",
3857 "-DXNN_TEST_MODE=1",
3858 ],
3859 gcc_copts = xnnpack_gcc_std_copts(),
3860 msvc_copts = xnnpack_msvc_std_copts(),
3861 deps = [
3862 ":tables",
3863 "@FP16",
3864 "@FXdiv",
3865 "@pthreadpool",
3866 ],
3867)
3868
3869xnnpack_cc_library(
Marat Dukhan436ebe62019-12-04 15:10:12 -08003870 name = "wasm_ukernels",
3871 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003872 gcc_copts = xnnpack_gcc_std_copts(),
3873 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan436ebe62019-12-04 15:10:12 -08003874 wasm_srcs = WASM_UKERNELS,
Marat Dukhan1c6e3892020-06-25 23:56:10 -07003875 wasmsimd_srcs = WASM_UKERNELS + WASMSIMD_UKERNELS,
Marat Dukhan436ebe62019-12-04 15:10:12 -08003876 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003877 ":tables",
Marat Dukhan436ebe62019-12-04 15:10:12 -08003878 "@FP16",
3879 "@FXdiv",
3880 "@pthreadpool",
3881 ],
3882)
3883
3884xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003885 name = "wasm_ukernels_test_mode",
3886 hdrs = INTERNAL_HDRS,
3887 copts = [
3888 "-UNDEBUG",
3889 "-DXNN_TEST_MODE=1",
3890 ],
3891 gcc_copts = xnnpack_gcc_std_copts(),
3892 msvc_copts = xnnpack_msvc_std_copts(),
3893 wasm_srcs = WASM_UKERNELS,
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07003894 wasmsimd_srcs = WASM_UKERNELS + WASMSIMD_UKERNELS,
Marat Dukhan33fcf782020-05-24 14:27:15 -07003895 deps = [
3896 ":tables",
3897 "@FP16",
3898 "@FXdiv",
3899 "@pthreadpool",
3900 ],
3901)
3902
3903xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003904 name = "neon_ukernels",
3905 hdrs = INTERNAL_HDRS,
3906 aarch32_copts = [
3907 "-marm",
Marat Dukhan8853b822020-05-07 12:19:01 -07003908 "-march=armv7-a",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003909 "-mfpu=neon",
3910 ],
3911 aarch32_srcs = NEON_UKERNELS,
3912 aarch64_srcs = NEON_UKERNELS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003913 gcc_copts = xnnpack_gcc_std_copts(),
3914 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08003915 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003916 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003917 "@FP16",
3918 "@pthreadpool",
3919 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003920)
3921
3922xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003923 name = "neon_ukernels_test_mode",
3924 hdrs = INTERNAL_HDRS,
3925 aarch32_copts = [
3926 "-marm",
3927 "-march=armv7-a",
3928 "-mfpu=neon",
3929 ],
3930 aarch32_srcs = NEON_UKERNELS,
3931 aarch64_srcs = NEON_UKERNELS,
3932 copts = [
3933 "-UNDEBUG",
3934 "-DXNN_TEST_MODE=1",
3935 ],
3936 gcc_copts = xnnpack_gcc_std_copts(),
3937 msvc_copts = xnnpack_msvc_std_copts(),
3938 deps = [
3939 ":tables",
3940 "@FP16",
3941 "@pthreadpool",
3942 ],
3943)
3944
3945xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003946 name = "neonfma_ukernels",
3947 hdrs = INTERNAL_HDRS,
3948 aarch32_copts = [
3949 "-marm",
Marat Dukhan8853b822020-05-07 12:19:01 -07003950 "-march=armv7-a",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003951 "-mfpu=neon-vfpv4",
3952 ],
3953 aarch32_srcs = NEONFMA_UKERNELS,
3954 aarch64_srcs = NEONFMA_UKERNELS + AARCH64_NEONFMA_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07003955 apple_aarch32_copts = [
3956 "-mcpu=swift",
3957 "-mtune=generic",
3958 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07003959 gcc_copts = xnnpack_gcc_std_copts(),
3960 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08003961 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003962 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003963 "@FP16",
3964 "@pthreadpool",
3965 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003966)
3967
3968xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003969 name = "neonfma_ukernels_test_mode",
3970 hdrs = INTERNAL_HDRS,
3971 aarch32_copts = [
3972 "-marm",
3973 "-march=armv7-a",
3974 "-mfpu=neon-vfpv4",
3975 ],
3976 aarch32_srcs = NEONFMA_UKERNELS,
3977 aarch64_srcs = NEONFMA_UKERNELS + AARCH64_NEONFMA_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07003978 apple_aarch32_copts = [
3979 "-mcpu=swift",
3980 "-mtune=generic",
3981 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07003982 copts = [
3983 "-UNDEBUG",
3984 "-DXNN_TEST_MODE=1",
3985 ],
3986 gcc_copts = xnnpack_gcc_std_copts(),
3987 msvc_copts = xnnpack_msvc_std_copts(),
3988 deps = [
3989 ":tables",
3990 "@FP16",
3991 "@pthreadpool",
3992 ],
3993)
3994
3995xnnpack_cc_library(
Marat Dukhan8853b822020-05-07 12:19:01 -07003996 name = "neonv8_ukernels",
3997 hdrs = INTERNAL_HDRS,
3998 aarch32_copts = [
3999 "-marm",
4000 "-march=armv8-a",
4001 "-mfpu=neon-fp-armv8",
4002 ],
4003 aarch32_srcs = NEONV8_UKERNELS,
4004 aarch64_srcs = NEONV8_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07004005 apple_aarch32_copts = [
4006 "-mcpu=cyclone",
4007 "-mtune=generic",
4008 ],
Marat Dukhan8853b822020-05-07 12:19:01 -07004009 gcc_copts = xnnpack_gcc_std_copts(),
4010 msvc_copts = xnnpack_msvc_std_copts(),
4011 deps = [
4012 ":tables",
4013 "@FP16",
4014 "@pthreadpool",
4015 ],
4016)
4017
4018xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07004019 name = "neonv8_ukernels_test_mode",
4020 hdrs = INTERNAL_HDRS,
4021 aarch32_copts = [
4022 "-marm",
4023 "-march=armv8-a",
4024 "-mfpu=neon-fp-armv8",
4025 ],
4026 aarch32_srcs = NEONV8_UKERNELS,
4027 aarch64_srcs = NEONV8_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07004028 apple_aarch32_copts = [
4029 "-mcpu=cyclone",
4030 "-mtune=generic",
4031 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07004032 copts = [
4033 "-UNDEBUG",
4034 "-DXNN_TEST_MODE=1",
4035 ],
4036 gcc_copts = xnnpack_gcc_std_copts(),
4037 msvc_copts = xnnpack_msvc_std_copts(),
4038 deps = [
4039 ":tables",
4040 "@FP16",
4041 "@pthreadpool",
4042 ],
4043)
4044
4045xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004046 name = "neonfp16arith_ukernels",
4047 hdrs = INTERNAL_HDRS,
4048 aarch64_copts = ["-march=armv8.2-a+fp16"],
4049 aarch64_srcs = AARCH64_NEONFP16ARITH_UKERNELS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004050 gcc_copts = xnnpack_gcc_std_copts(),
4051 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08004052 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08004053 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08004054 "@FP16",
4055 "@pthreadpool",
4056 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004057)
4058
4059xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07004060 name = "neonfp16arith_ukernels_test_mode",
4061 hdrs = INTERNAL_HDRS,
4062 aarch64_copts = ["-march=armv8.2-a+fp16"],
4063 aarch64_srcs = AARCH64_NEONFP16ARITH_UKERNELS,
4064 copts = [
4065 "-UNDEBUG",
4066 "-DXNN_TEST_MODE=1",
4067 ],
4068 gcc_copts = xnnpack_gcc_std_copts(),
4069 msvc_copts = xnnpack_msvc_std_copts(),
4070 deps = [
4071 ":tables",
4072 "@FP16",
4073 "@pthreadpool",
4074 ],
4075)
4076
4077xnnpack_cc_library(
Benoit Jacoba9644732020-08-13 12:48:55 -07004078 name = "neondot_ukernels",
4079 hdrs = INTERNAL_HDRS,
Marat Dukhan799ac752020-08-13 16:08:03 -07004080 aarch32_copts = [
4081 "-marm",
4082 "-march=armv8.2-a+dotprod",
4083 "-mfpu=neon-fp-armv8",
4084 ],
Benoit Jacoba9644732020-08-13 12:48:55 -07004085 aarch32_srcs = NEONDOT_UKERNELS,
4086 aarch64_copts = ["-march=armv8.2-a+dotprod"],
4087 aarch64_srcs = NEONDOT_UKERNELS,
4088 gcc_copts = xnnpack_gcc_std_copts(),
4089 msvc_copts = xnnpack_msvc_std_copts(),
4090 deps = [
4091 ":tables",
4092 "@FP16",
4093 "@pthreadpool",
4094 ],
4095)
4096
4097xnnpack_cc_library(
4098 name = "neondot_ukernels_test_mode",
4099 hdrs = INTERNAL_HDRS,
Marat Dukhan799ac752020-08-13 16:08:03 -07004100 aarch32_copts = [
4101 "-marm",
4102 "-march=armv8.2-a+dotprod",
4103 "-mfpu=neon-fp-armv8",
4104 ],
Benoit Jacoba9644732020-08-13 12:48:55 -07004105 aarch32_srcs = NEONDOT_UKERNELS,
4106 aarch64_copts = ["-march=armv8.2-a+dotprod"],
4107 aarch64_srcs = NEONDOT_UKERNELS,
4108 copts = [
4109 "-UNDEBUG",
4110 "-DXNN_TEST_MODE=1",
4111 ],
4112 gcc_copts = xnnpack_gcc_std_copts(),
4113 msvc_copts = xnnpack_msvc_std_copts(),
4114 deps = [
4115 ":tables",
4116 "@FP16",
4117 "@pthreadpool",
4118 ],
4119)
4120
4121xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004122 name = "sse2_ukernels",
4123 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004124 gcc_copts = xnnpack_gcc_std_copts(),
4125 gcc_x86_copts = ["-msse2"],
4126 msvc_copts = xnnpack_msvc_std_copts(),
4127 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004128 x86_srcs = SSE_UKERNELS + SSE2_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08004129 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08004130 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08004131 "@FP16",
4132 "@pthreadpool",
4133 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004134)
4135
4136xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07004137 name = "sse2_ukernels_test_mode",
4138 hdrs = INTERNAL_HDRS,
4139 copts = [
4140 "-UNDEBUG",
4141 "-DXNN_TEST_MODE=1",
4142 ],
4143 gcc_copts = xnnpack_gcc_std_copts(),
4144 gcc_x86_copts = ["-msse2"],
4145 msvc_copts = xnnpack_msvc_std_copts(),
4146 msvc_x86_32_copts = ["/arch:SSE2"],
4147 x86_srcs = SSE_UKERNELS + SSE2_UKERNELS,
4148 deps = [
4149 ":tables",
4150 "@FP16",
4151 "@pthreadpool",
4152 ],
4153)
4154
4155xnnpack_cc_library(
Marat Dukhanfe7acb62020-03-09 19:30:05 -07004156 name = "ssse3_ukernels",
4157 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004158 gcc_copts = xnnpack_gcc_std_copts(),
4159 gcc_x86_copts = ["-mssse3"],
4160 msvc_copts = xnnpack_msvc_std_copts(),
4161 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhanfe7acb62020-03-09 19:30:05 -07004162 x86_srcs = SSSE3_UKERNELS,
4163 deps = [
4164 ":tables",
4165 "@FP16",
4166 "@pthreadpool",
4167 ],
4168)
4169
4170xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07004171 name = "ssse3_ukernels_test_mode",
4172 hdrs = INTERNAL_HDRS,
4173 copts = [
4174 "-UNDEBUG",
4175 "-DXNN_TEST_MODE=1",
4176 ],
4177 gcc_copts = xnnpack_gcc_std_copts(),
4178 gcc_x86_copts = ["-mssse3"],
4179 msvc_copts = xnnpack_msvc_std_copts(),
4180 msvc_x86_32_copts = ["/arch:SSE2"],
4181 x86_srcs = SSSE3_UKERNELS,
4182 deps = [
4183 ":tables",
4184 "@FP16",
4185 "@pthreadpool",
4186 ],
4187)
4188
4189xnnpack_cc_library(
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08004190 name = "sse41_ukernels",
4191 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004192 gcc_copts = xnnpack_gcc_std_copts(),
4193 gcc_x86_copts = ["-msse4.1"],
4194 msvc_copts = xnnpack_msvc_std_copts(),
4195 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08004196 x86_srcs = SSE41_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08004197 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08004198 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08004199 "@FP16",
4200 "@pthreadpool",
4201 ],
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08004202)
4203
4204xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07004205 name = "sse41_ukernels_test_mode",
4206 hdrs = INTERNAL_HDRS,
4207 copts = [
4208 "-UNDEBUG",
4209 "-DXNN_TEST_MODE=1",
4210 ],
4211 gcc_copts = xnnpack_gcc_std_copts(),
4212 gcc_x86_copts = ["-msse4.1"],
4213 msvc_copts = xnnpack_msvc_std_copts(),
4214 msvc_x86_32_copts = ["/arch:SSE2"],
4215 x86_srcs = SSE41_UKERNELS,
4216 deps = [
4217 ":tables",
4218 "@FP16",
4219 "@pthreadpool",
4220 ],
4221)
4222
4223xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004224 name = "avx_ukernels",
4225 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004226 gcc_copts = xnnpack_gcc_std_copts(),
4227 gcc_x86_copts = ["-mavx"],
4228 msvc_copts = xnnpack_msvc_std_copts(),
4229 msvc_x86_32_copts = ["/arch:AVX"],
4230 msvc_x86_64_copts = ["/arch:AVX"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004231 x86_srcs = AVX_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08004232 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08004233 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08004234 "@FP16",
4235 "@pthreadpool",
4236 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004237)
4238
4239xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07004240 name = "avx_ukernels_test_mode",
4241 hdrs = INTERNAL_HDRS,
4242 copts = [
4243 "-UNDEBUG",
4244 "-DXNN_TEST_MODE=1",
4245 ],
4246 gcc_copts = xnnpack_gcc_std_copts(),
4247 gcc_x86_copts = ["-mavx"],
4248 msvc_copts = xnnpack_msvc_std_copts(),
4249 msvc_x86_32_copts = ["/arch:AVX"],
4250 msvc_x86_64_copts = ["/arch:AVX"],
4251 x86_srcs = AVX_UKERNELS,
4252 deps = [
4253 ":tables",
4254 "@FP16",
4255 "@pthreadpool",
4256 ],
4257)
4258
4259xnnpack_cc_library(
Marat Dukhan1566fee2020-08-02 21:55:41 -07004260 name = "xop_ukernels",
4261 hdrs = INTERNAL_HDRS,
4262 gcc_copts = xnnpack_gcc_std_copts(),
4263 gcc_x86_copts = ["-mxop"],
4264 msvc_copts = xnnpack_msvc_std_copts(),
4265 msvc_x86_32_copts = ["/arch:AVX"],
4266 msvc_x86_64_copts = ["/arch:AVX"],
4267 x86_srcs = XOP_UKERNELS,
4268 deps = [
4269 ":tables",
4270 "@FP16",
4271 "@pthreadpool",
4272 ],
4273)
4274
4275xnnpack_cc_library(
4276 name = "xop_ukernels_test_mode",
4277 hdrs = INTERNAL_HDRS,
4278 copts = [
4279 "-UNDEBUG",
4280 "-DXNN_TEST_MODE=1",
4281 ],
4282 gcc_copts = xnnpack_gcc_std_copts(),
4283 gcc_x86_copts = ["-mxop"],
4284 msvc_copts = xnnpack_msvc_std_copts(),
4285 msvc_x86_32_copts = ["/arch:AVX"],
4286 msvc_x86_64_copts = ["/arch:AVX"],
4287 x86_srcs = XOP_UKERNELS,
4288 deps = [
4289 ":tables",
4290 "@FP16",
4291 "@pthreadpool",
4292 ],
4293)
4294
4295xnnpack_cc_library(
Marat Dukhanfda12b82019-11-21 12:27:59 -08004296 name = "fma3_ukernels",
4297 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004298 gcc_copts = xnnpack_gcc_std_copts(),
4299 gcc_x86_copts = ["-mfma"],
4300 msvc_copts = xnnpack_msvc_std_copts(),
4301 msvc_x86_32_copts = ["/arch:AVX"],
4302 msvc_x86_64_copts = ["/arch:AVX"],
Marat Dukhanfda12b82019-11-21 12:27:59 -08004303 x86_srcs = FMA3_UKERNELS,
4304 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08004305 ":tables",
Marat Dukhanfda12b82019-11-21 12:27:59 -08004306 "@FP16",
4307 "@pthreadpool",
4308 ],
4309)
4310
4311xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07004312 name = "fma3_ukernels_test_mode",
4313 hdrs = INTERNAL_HDRS,
4314 copts = [
4315 "-UNDEBUG",
4316 "-DXNN_TEST_MODE=1",
4317 ],
4318 gcc_copts = xnnpack_gcc_std_copts(),
4319 gcc_x86_copts = ["-mfma"],
4320 msvc_copts = xnnpack_msvc_std_copts(),
4321 msvc_x86_32_copts = ["/arch:AVX"],
4322 msvc_x86_64_copts = ["/arch:AVX"],
4323 x86_srcs = FMA3_UKERNELS,
4324 deps = [
4325 ":tables",
4326 "@FP16",
4327 "@pthreadpool",
4328 ],
4329)
4330
4331xnnpack_cc_library(
Marat Dukhan6adff4e2019-10-14 18:32:07 -07004332 name = "avx2_ukernels",
4333 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004334 gcc_copts = xnnpack_gcc_std_copts(),
4335 gcc_x86_copts = [
Marat Dukhan6adff4e2019-10-14 18:32:07 -07004336 "-mfma",
4337 "-mavx2",
4338 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07004339 msvc_copts = xnnpack_msvc_std_copts(),
4340 msvc_x86_32_copts = ["/arch:AVX2"],
4341 msvc_x86_64_copts = ["/arch:AVX2"],
Marat Dukhan6adff4e2019-10-14 18:32:07 -07004342 x86_srcs = AVX2_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08004343 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08004344 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08004345 "@FP16",
4346 "@pthreadpool",
4347 ],
Marat Dukhan6adff4e2019-10-14 18:32:07 -07004348)
4349
4350xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07004351 name = "avx2_ukernels_test_mode",
4352 hdrs = INTERNAL_HDRS,
4353 copts = [
4354 "-UNDEBUG",
4355 "-DXNN_TEST_MODE=1",
4356 ],
4357 gcc_copts = xnnpack_gcc_std_copts(),
4358 gcc_x86_copts = [
4359 "-mfma",
4360 "-mavx2",
4361 ],
4362 msvc_copts = xnnpack_msvc_std_copts(),
4363 msvc_x86_32_copts = ["/arch:AVX2"],
4364 msvc_x86_64_copts = ["/arch:AVX2"],
4365 x86_srcs = AVX2_UKERNELS,
4366 deps = [
4367 ":tables",
4368 "@FP16",
4369 "@pthreadpool",
4370 ],
4371)
4372
4373xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004374 name = "avx512f_ukernels",
4375 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004376 gcc_copts = xnnpack_gcc_std_copts(),
4377 gcc_x86_copts = ["-mavx512f"],
4378 mingw_copts = ["-fno-asynchronous-unwind-tables"],
4379 msvc_copts = xnnpack_msvc_std_copts(),
4380 msvc_x86_32_copts = ["/arch:AVX512"],
4381 msvc_x86_64_copts = ["/arch:AVX512"],
4382 msys_copts = ["-fno-asynchronous-unwind-tables"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004383 x86_srcs = AVX512F_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08004384 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08004385 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08004386 "@FP16",
4387 "@pthreadpool",
4388 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004389)
4390
4391xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07004392 name = "avx512f_ukernels_test_mode",
4393 hdrs = INTERNAL_HDRS,
4394 copts = [
4395 "-UNDEBUG",
4396 "-DXNN_TEST_MODE=1",
4397 ],
4398 gcc_copts = xnnpack_gcc_std_copts(),
4399 gcc_x86_copts = ["-mavx512f"],
4400 mingw_copts = ["-fno-asynchronous-unwind-tables"],
4401 msvc_copts = xnnpack_msvc_std_copts(),
4402 msvc_x86_32_copts = ["/arch:AVX512"],
4403 msvc_x86_64_copts = ["/arch:AVX512"],
4404 msys_copts = ["-fno-asynchronous-unwind-tables"],
4405 x86_srcs = AVX512F_UKERNELS,
4406 deps = [
4407 ":tables",
4408 "@FP16",
4409 "@pthreadpool",
4410 ],
4411)
4412
4413xnnpack_cc_library(
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07004414 name = "avx512skx_ukernels",
4415 hdrs = INTERNAL_HDRS,
4416 gcc_copts = xnnpack_gcc_std_copts(),
4417 gcc_x86_copts = [
4418 "-mavx512f",
4419 "-mavx512cd",
4420 "-mavx512bw",
4421 "-mavx512dq",
4422 "-mavx512vl",
4423 ],
4424 mingw_copts = ["-fno-asynchronous-unwind-tables"],
4425 msvc_copts = xnnpack_msvc_std_copts(),
4426 msvc_x86_32_copts = ["/arch:AVX512"],
4427 msvc_x86_64_copts = ["/arch:AVX512"],
4428 msys_copts = ["-fno-asynchronous-unwind-tables"],
4429 x86_srcs = AVX512SKX_UKERNELS,
4430 deps = [
4431 ":tables",
4432 "@FP16",
4433 "@pthreadpool",
4434 ],
4435)
4436
4437xnnpack_cc_library(
4438 name = "avx512skx_ukernels_test_mode",
4439 hdrs = INTERNAL_HDRS,
4440 copts = [
4441 "-UNDEBUG",
4442 "-DXNN_TEST_MODE=1",
4443 ],
4444 gcc_copts = xnnpack_gcc_std_copts(),
4445 gcc_x86_copts = [
4446 "-mavx512f",
4447 "-mavx512cd",
4448 "-mavx512bw",
4449 "-mavx512dq",
4450 "-mavx512vl",
4451 ],
4452 mingw_copts = ["-fno-asynchronous-unwind-tables"],
4453 msvc_copts = xnnpack_msvc_std_copts(),
4454 msvc_x86_32_copts = ["/arch:AVX512"],
4455 msvc_x86_64_copts = ["/arch:AVX512"],
4456 msys_copts = ["-fno-asynchronous-unwind-tables"],
4457 x86_srcs = AVX512SKX_UKERNELS,
4458 deps = [
4459 ":tables",
4460 "@FP16",
4461 "@pthreadpool",
4462 ],
4463)
4464
4465xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004466 name = "asm_ukernels",
4467 hdrs = ["src/xnnpack/assembly.h"],
4468 aarch32_srcs = AARCH32_ASM_UKERNELS,
Frank Barchard31bb45b2020-10-06 00:26:33 -07004469 aarch64_copts = ["-march=armv8.2-a+fp16+dotprod"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004470 aarch64_srcs = AARCH64_ASM_UKERNELS,
Frank Barchardbcedc082020-08-17 18:00:51 -07004471 wasm_srcs = WASM32_ASM_UKERNELS,
4472 wasmsimd_srcs = WASM32_ASM_UKERNELS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07004473)
4474
Marat Dukhan3b59de22020-06-03 20:15:19 -07004475xnnpack_cc_library(
4476 name = "logging_utils",
4477 srcs = LOGGING_SRCS,
4478 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
4479 copts = LOGGING_COPTS + [
4480 "-Isrc",
4481 "-Iinclude",
4482 ] + select({
4483 ":debug_build": [],
4484 "//conditions:default": xnnpack_min_size_copts(),
4485 }),
4486 gcc_copts = xnnpack_gcc_std_copts(),
4487 msvc_copts = xnnpack_msvc_std_copts(),
4488 visibility = xnnpack_visibility(),
4489 deps = [
4490 "@FP16",
4491 "@clog",
4492 "@pthreadpool",
4493 ],
4494)
4495
Marat Dukhan08c4a432019-10-03 09:29:21 -07004496xnnpack_aggregate_library(
4497 name = "ukernels",
Marat Dukhan6e8c0ce2021-04-13 14:35:08 -07004498 aarch32_ios_deps = [
4499 ":neon_ukernels",
4500 ":neonfma_ukernels",
4501 ":neonv8_ukernels",
4502 ":asm_ukernels",
4503 ],
4504 aarch32_nonios_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07004505 ":neon_ukernels",
4506 ":neonfma_ukernels",
Marat Dukhan8853b822020-05-07 12:19:01 -07004507 ":neonv8_ukernels",
Benoit Jacoba9644732020-08-13 12:48:55 -07004508 ":neondot_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004509 ":asm_ukernels",
4510 ],
4511 aarch64_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07004512 ":neon_ukernels",
4513 ":neonfma_ukernels",
Marat Dukhan8853b822020-05-07 12:19:01 -07004514 ":neonv8_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004515 ":neonfp16arith_ukernels",
Benoit Jacoba9644732020-08-13 12:48:55 -07004516 ":neondot_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004517 ":asm_ukernels",
4518 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07004519 generic_deps = [
4520 ":scalar_ukernels",
4521 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07004522 wasm_deps = [
4523 ":wasm_ukernels",
Frank Barchardbcedc082020-08-17 18:00:51 -07004524 ":asm_ukernels",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004525 ],
4526 wasmsimd_deps = [
4527 ":wasm_ukernels",
Frank Barchardbcedc082020-08-17 18:00:51 -07004528 ":asm_ukernels",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004529 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004530 x86_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07004531 ":sse2_ukernels",
Marat Dukhanfe7acb62020-03-09 19:30:05 -07004532 ":ssse3_ukernels",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08004533 ":sse41_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004534 ":avx_ukernels",
Marat Dukhan1566fee2020-08-02 21:55:41 -07004535 ":xop_ukernels",
Marat Dukhanfda12b82019-11-21 12:27:59 -08004536 ":fma3_ukernels",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07004537 ":avx2_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004538 ":avx512f_ukernels",
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07004539 ":avx512skx_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004540 ],
4541)
4542
Marat Dukhan33fcf782020-05-24 14:27:15 -07004543xnnpack_aggregate_library(
4544 name = "ukernels_test_mode",
Marat Dukhan6e8c0ce2021-04-13 14:35:08 -07004545 aarch32_ios_deps = [
4546 ":neon_ukernels_test_mode",
4547 ":neonfma_ukernels_test_mode",
4548 ":neonv8_ukernels_test_mode",
4549 ":asm_ukernels",
4550 ],
4551 aarch32_nonios_deps = [
Marat Dukhan33fcf782020-05-24 14:27:15 -07004552 ":neon_ukernels_test_mode",
4553 ":neonfma_ukernels_test_mode",
4554 ":neonv8_ukernels_test_mode",
Benoit Jacoba9644732020-08-13 12:48:55 -07004555 ":neondot_ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004556 ":asm_ukernels",
4557 ],
4558 aarch64_deps = [
4559 ":neon_ukernels_test_mode",
4560 ":neonfma_ukernels_test_mode",
4561 ":neonv8_ukernels_test_mode",
4562 ":neonfp16arith_ukernels_test_mode",
Benoit Jacoba9644732020-08-13 12:48:55 -07004563 ":neondot_ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004564 ":asm_ukernels",
4565 ],
4566 generic_deps = [
4567 ":scalar_ukernels_test_mode",
4568 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07004569 wasm_deps = [
4570 ":wasm_ukernels_test_mode",
Frank Barchardbcedc082020-08-17 18:00:51 -07004571 ":asm_ukernels",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004572 ],
4573 wasmsimd_deps = [
4574 ":wasm_ukernels_test_mode",
Frank Barchardbcedc082020-08-17 18:00:51 -07004575 ":asm_ukernels",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004576 ],
4577 x86_deps = [
4578 ":sse2_ukernels_test_mode",
4579 ":ssse3_ukernels_test_mode",
4580 ":sse41_ukernels_test_mode",
4581 ":avx_ukernels_test_mode",
Marat Dukhan1566fee2020-08-02 21:55:41 -07004582 ":xop_ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004583 ":fma3_ukernels_test_mode",
4584 ":avx2_ukernels_test_mode",
4585 ":avx512f_ukernels_test_mode",
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07004586 ":avx512skx_ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004587 ],
4588)
4589
Marat Dukhan08c4a432019-10-03 09:29:21 -07004590xnnpack_cc_library(
4591 name = "im2col",
4592 srcs = ["src/im2col.c"],
4593 hdrs = [
4594 "src/xnnpack/common.h",
4595 "src/xnnpack/im2col.h",
4596 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07004597 gcc_copts = xnnpack_gcc_std_copts(),
4598 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004599)
4600
4601xnnpack_cc_library(
4602 name = "indirection",
4603 srcs = ["src/indirection.c"],
4604 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004605 gcc_copts = xnnpack_gcc_std_copts(),
4606 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004607 deps = [
4608 "@FP16",
4609 "@FXdiv",
4610 "@pthreadpool",
4611 ],
4612)
4613
4614xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07004615 name = "indirection_test_mode",
4616 srcs = ["src/indirection.c"],
4617 hdrs = INTERNAL_HDRS,
4618 copts = [
4619 "-UNDEBUG",
4620 "-DXNN_TEST_MODE=1",
4621 ],
4622 gcc_copts = xnnpack_gcc_std_copts(),
4623 msvc_copts = xnnpack_msvc_std_copts(),
4624 deps = [
4625 "@FP16",
4626 "@FXdiv",
4627 "@pthreadpool",
4628 ],
4629)
4630
4631xnnpack_cc_library(
Marat Dukhanab582382020-07-06 13:32:08 -07004632 name = "packing",
4633 srcs = ["src/packing.c"],
4634 hdrs = INTERNAL_HDRS,
4635 gcc_copts = xnnpack_gcc_std_copts(),
4636 msvc_copts = xnnpack_msvc_std_copts(),
4637 deps = [
4638 "@FP16",
4639 "@FXdiv",
4640 "@pthreadpool",
4641 ],
4642)
4643
4644xnnpack_cc_library(
4645 name = "packing_test_mode",
4646 srcs = ["src/packing.c"],
4647 hdrs = INTERNAL_HDRS,
4648 copts = [
4649 "-UNDEBUG",
4650 "-DXNN_TEST_MODE=1",
4651 ],
4652 gcc_copts = xnnpack_gcc_std_copts(),
4653 msvc_copts = xnnpack_msvc_std_copts(),
4654 deps = [
4655 "@FP16",
4656 "@FXdiv",
4657 "@pthreadpool",
4658 ],
4659)
4660
4661xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004662 name = "operator_run",
4663 srcs = ["src/operator-run.c"],
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07004664 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004665 copts = LOGGING_COPTS + select({
Marat Dukhan05702cf2020-03-26 15:41:33 -07004666 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4667 "//conditions:default": [],
4668 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07004669 gcc_copts = xnnpack_gcc_std_copts(),
4670 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004671 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07004672 ":logging_utils",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004673 "@FP16",
4674 "@FXdiv",
4675 "@clog",
4676 "@pthreadpool",
4677 ],
4678)
4679
Chao Mei6ddfc602020-05-13 22:29:36 -07004680xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07004681 name = "operator_run_test_mode",
4682 srcs = ["src/operator-run.c"],
4683 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
4684 copts = LOGGING_COPTS + [
4685 "-UNDEBUG",
4686 "-DXNN_TEST_MODE=1",
4687 ] + select({
4688 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4689 "//conditions:default": [],
4690 }),
4691 gcc_copts = xnnpack_gcc_std_copts(),
4692 msvc_copts = xnnpack_msvc_std_copts(),
4693 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07004694 ":logging_utils",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004695 "@FP16",
4696 "@FXdiv",
4697 "@clog",
4698 "@pthreadpool",
4699 ],
4700)
4701
4702xnnpack_cc_library(
Chao Mei6ddfc602020-05-13 22:29:36 -07004703 name = "memory_planner",
4704 srcs = ["src/memory-planner.c"],
4705 hdrs = INTERNAL_HDRS,
4706 defines = select({
4707 ":xnn_enable_memopt_explicit_true": ["XNN_ENABLE_MEMOPT=1"],
4708 ":xnn_enable_memopt_explicit_false": ["XNN_ENABLE_MEMOPT=0"],
4709 "//conditions:default": ["XNN_ENABLE_MEMOPT=1"],
4710 }),
4711 gcc_copts = xnnpack_gcc_std_copts(),
4712 msvc_copts = xnnpack_msvc_std_copts(),
4713 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07004714 ":logging_utils",
Chao Mei6ddfc602020-05-13 22:29:36 -07004715 "@pthreadpool",
4716 ],
4717)
4718
Marat Dukhan33fcf782020-05-24 14:27:15 -07004719xnnpack_cc_library(
4720 name = "memory_planner_test_mode",
4721 srcs = ["src/memory-planner.c"],
4722 hdrs = INTERNAL_HDRS,
4723 copts = [
4724 "-UNDEBUG",
4725 "-DXNN_TEST_MODE=1",
4726 ],
4727 defines = select({
4728 ":xnn_enable_memopt_explicit_true": ["XNN_ENABLE_MEMOPT=1"],
4729 ":xnn_enable_memopt_explicit_false": ["XNN_ENABLE_MEMOPT=0"],
4730 "//conditions:default": ["XNN_ENABLE_MEMOPT=1"],
4731 }),
4732 gcc_copts = xnnpack_gcc_std_copts(),
4733 msvc_copts = xnnpack_msvc_std_copts(),
4734 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07004735 ":logging_utils",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004736 "@pthreadpool",
4737 ],
4738)
4739
Marat Dukhan08c4a432019-10-03 09:29:21 -07004740cc_library(
4741 name = "enable_assembly",
4742 defines = select({
4743 ":xnn_enable_assembly_explicit_true": ["XNN_ENABLE_ASSEMBLY=1"],
4744 ":xnn_enable_assembly_explicit_false": ["XNN_ENABLE_ASSEMBLY=0"],
Frank Barchard810171d2019-10-10 10:34:51 -07004745 "//conditions:default": ["XNN_ENABLE_ASSEMBLY=1"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004746 }),
4747)
4748
Marat Dukhan9de90e02020-06-18 16:04:12 -07004749cc_library(
4750 name = "enable_sparse",
4751 defines = select({
4752 ":xnn_enable_sparse_explicit_true": ["XNN_ENABLE_SPARSE=1"],
4753 ":xnn_enable_sparse_explicit_false": ["XNN_ENABLE_SPARSE=0"],
Marat Dukhanb36582b2020-12-08 11:16:28 -08004754 "//conditions:default": ["XNN_ENABLE_SPARSE=1"],
Marat Dukhan9de90e02020-06-18 16:04:12 -07004755 }),
4756)
4757
Marat Dukhancf056b22019-10-07 10:26:29 -07004758xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004759 name = "operators",
4760 srcs = OPERATOR_SRCS + [
Marat Dukhan496389f2021-04-07 15:47:12 -07004761 "src/allocator.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004762 "src/operator-delete.c",
Marat Dukhancf056b22019-10-07 10:26:29 -07004763 ],
4764 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004765 copts = LOGGING_COPTS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07004766 "-Isrc",
4767 "-Iinclude",
4768 ] + select({
4769 ":debug_build": [],
4770 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07004771 }) + select({
4772 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4773 "//conditions:default": [],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004774 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07004775 gcc_copts = xnnpack_gcc_std_copts(),
4776 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004777 deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07004778 ":indirection",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004779 ":logging_utils",
Marat Dukhanab582382020-07-06 13:32:08 -07004780 ":packing",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004781 "@FP16",
4782 "@FXdiv",
4783 "@clog",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004784 "@pthreadpool",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004785 ],
4786)
4787
Marat Dukhan10a38082020-04-17 03:58:35 -07004788xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07004789 name = "operators_test_mode",
4790 srcs = OPERATOR_SRCS + [
Marat Dukhan496389f2021-04-07 15:47:12 -07004791 "src/allocator.c",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004792 "src/operator-delete.c",
4793 ],
4794 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
4795 copts = LOGGING_COPTS + [
4796 "-Isrc",
4797 "-Iinclude",
4798 "-UNDEBUG",
4799 "-DXNN_TEST_MODE=1",
4800 ] + select({
4801 ":debug_build": [],
4802 "//conditions:default": xnnpack_min_size_copts(),
4803 }) + select({
4804 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4805 "//conditions:default": [],
4806 }),
4807 gcc_copts = xnnpack_gcc_std_copts(),
4808 msvc_copts = xnnpack_msvc_std_copts(),
4809 deps = [
4810 ":indirection_test_mode",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004811 ":logging_utils",
Marat Dukhanab582382020-07-06 13:32:08 -07004812 ":packing_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004813 "@FP16",
4814 "@FXdiv",
4815 "@clog",
4816 "@pthreadpool",
4817 ],
4818)
4819
4820xnnpack_cc_library(
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004821 name = "XNNPACK",
4822 srcs = [
4823 "src/init.c",
Marat Dukhanccfdbd12020-02-03 14:27:45 -08004824 "src/runtime.c",
4825 "src/subgraph.c",
4826 "src/tensor.c",
Marat Dukhanf03da0d2020-06-10 16:00:20 -07004827 ] + SUBGRAPH_SRCS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004828 hdrs = ["include/xnnpack.h"],
4829 copts = LOGGING_COPTS + [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004830 "-Isrc",
4831 "-Iinclude",
4832 ] + select({
4833 ":debug_build": [],
4834 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07004835 }) + select({
4836 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4837 "//conditions:default": [],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004838 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07004839 gcc_copts = xnnpack_gcc_std_copts(),
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004840 includes = ["include"],
Marat Dukhan10a38082020-04-17 03:58:35 -07004841 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004842 visibility = xnnpack_visibility(),
4843 deps = [
4844 ":enable_assembly",
Marat Dukhan9de90e02020-06-18 16:04:12 -07004845 ":enable_sparse",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004846 ":logging_utils",
Chao Mei6ddfc602020-05-13 22:29:36 -07004847 ":memory_planner",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004848 ":operator_run",
4849 ":operators",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004850 ":ukernels",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004851 "@clog",
Marat Dukhanab2946c2020-05-21 20:04:13 -07004852 "@FP16",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004853 "@pthreadpool",
Marat Dukhand343c222019-10-07 09:22:14 -07004854 ] + select({
4855 ":emscripten": [],
4856 "//conditions:default": ["@cpuinfo"],
4857 }),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004858)
4859
Marat Dukhan10a38082020-04-17 03:58:35 -07004860xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07004861 name = "XNNPACK_test_mode",
4862 srcs = [
4863 "src/init.c",
4864 "src/runtime.c",
4865 "src/subgraph.c",
4866 "src/tensor.c",
Marat Dukhanf03da0d2020-06-10 16:00:20 -07004867 ] + SUBGRAPH_SRCS,
Marat Dukhan33fcf782020-05-24 14:27:15 -07004868 hdrs = ["include/xnnpack.h"],
4869 copts = LOGGING_COPTS + [
4870 "-Isrc",
4871 "-Iinclude",
4872 "-UNDEBUG",
4873 "-DXNN_TEST_MODE=1",
4874 ] + select({
4875 ":debug_build": [],
4876 "//conditions:default": xnnpack_min_size_copts(),
4877 }) + select({
4878 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4879 "//conditions:default": [],
4880 }),
4881 gcc_copts = xnnpack_gcc_std_copts(),
4882 includes = ["include"],
4883 msvc_copts = xnnpack_msvc_std_copts(),
4884 visibility = xnnpack_visibility(),
4885 deps = [
4886 ":enable_assembly",
Marat Dukhan9de90e02020-06-18 16:04:12 -07004887 ":enable_sparse",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004888 ":logging_utils",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004889 ":memory_planner_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004890 ":operator_run_test_mode",
4891 ":operators_test_mode",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004892 ":ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004893 "@clog",
4894 "@FP16",
4895 "@pthreadpool",
4896 ] + select({
4897 ":emscripten": [],
4898 "//conditions:default": ["@cpuinfo"],
4899 }),
4900)
4901
Marat Dukhanf0cb70a2021-03-30 15:45:15 -07004902# Specialized XNNPACK version for TensorFlow Lite. Excludes operators currently
4903# not used by the TensorFlow Lite XNNPACK delegate to minimize code size.
Marat Dukhanae046f52020-06-15 13:16:14 -07004904xnnpack_cc_library(
Marat Dukhanf0cb70a2021-03-30 15:45:15 -07004905 name = "xnnpack_for_tflite",
4906 srcs = [
4907 "src/init.c",
4908 "src/runtime.c",
4909 "src/subgraph.c",
4910 "src/tensor.c",
4911 ] + SUBGRAPH_SRCS,
4912 hdrs = ["include/xnnpack.h"],
4913 copts = LOGGING_COPTS + [
4914 "-Isrc",
4915 "-Iinclude",
4916 ] + select({
4917 ":debug_build": [],
4918 "//conditions:default": xnnpack_min_size_copts(),
4919 }) + select({
4920 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4921 "//conditions:default": [],
4922 }),
4923 defines = [
Marat Dukhanf0cb70a2021-03-30 15:45:15 -07004924 "XNN_NO_QU8_OPERATORS",
4925 "XNN_NO_U8_OPERATORS",
4926 "XNN_NO_X8_OPERATORS",
4927 "XNN_NO_F16_OPERATORS",
4928 "XNN_NO_X16_OPERATORS",
Marat Dukhanb939cdb2021-03-30 18:51:51 -07004929 ] + select({
4930 ":xnn_enable_qs8_explicit_true": [],
4931 ":xnn_enable_qs8_explicit_false": ["XNN_NO_QS8_OPERATORS"],
4932 "//conditions:default": ["XNN_NO_QS8_OPERATORS"],
4933 }),
Marat Dukhanf0cb70a2021-03-30 15:45:15 -07004934 gcc_copts = xnnpack_gcc_std_copts(),
4935 includes = ["include"],
4936 msvc_copts = xnnpack_msvc_std_copts(),
4937 visibility = xnnpack_visibility(),
4938 deps = [
4939 ":enable_assembly",
4940 ":enable_sparse",
4941 ":logging_utils",
4942 ":memory_planner",
4943 ":operator_run",
4944 ":operators",
Marat Dukhand09ca262021-03-30 16:17:12 -07004945 ":ukernels",
Marat Dukhanf0cb70a2021-03-30 15:45:15 -07004946 "@clog",
4947 "@FP16",
4948 "@pthreadpool",
4949 ] + select({
4950 ":emscripten": [],
4951 "//conditions:default": ["@cpuinfo"],
4952 }),
4953)
4954
4955# Specialized XNNPACK version for TensorFlow.js. Excludes operators currently
4956# not used by the TensorFlow.js WebAssembly backend to minimize code size.
4957xnnpack_cc_library(
4958 name = "xnnpack_for_tfjs",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004959 srcs = [
4960 "src/init.c",
4961 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07004962 hdrs = ["include/xnnpack.h"],
4963 copts = LOGGING_COPTS + [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004964 "-Isrc",
4965 "-Iinclude",
4966 ] + select({
4967 ":debug_build": [],
4968 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07004969 }) + select({
4970 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4971 "//conditions:default": [],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004972 }),
4973 defines = [
Marat Dukhan16f1e1a2020-08-04 16:38:22 -07004974 "XNN_NO_QS8_OPERATORS",
Marat Dukhan08b7a972020-07-14 18:17:29 -07004975 "XNN_NO_QU8_OPERATORS",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004976 "XNN_NO_U8_OPERATORS",
4977 "XNN_NO_X8_OPERATORS",
Marat Dukhanefc47b82019-11-18 09:25:38 -08004978 "XNN_NO_NCHW_OPERATORS",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004979 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07004980 gcc_copts = xnnpack_gcc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004981 includes = ["include"],
Marat Dukhan10a38082020-04-17 03:58:35 -07004982 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004983 visibility = xnnpack_visibility(),
4984 deps = [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004985 ":enable_assembly",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004986 ":logging_utils",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004987 ":operator_run",
4988 ":operators",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004989 ":ukernels",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004990 "@clog",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004991 "@pthreadpool",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004992 ] + select({
4993 ":emscripten": [],
4994 "//conditions:default": ["@cpuinfo"],
4995 }),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004996)
4997
Marat Dukhancf056b22019-10-07 10:26:29 -07004998xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004999 name = "bench_utils",
5000 srcs = ["bench/utils.cc"],
5001 hdrs = ["bench/utils.h"],
Marat Dukhanbad48fe2019-11-04 10:35:22 -08005002 deps = [
5003 "@com_google_benchmark//:benchmark",
5004 "@cpuinfo",
5005 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005006)
5007
Frank Barchard7e955972019-10-11 10:34:25 -07005008######################### Benchmarks for micro-kernels #########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07005009
5010xnnpack_benchmark(
Marat Dukhan595e1702020-07-31 10:12:52 -07005011 name = "qs8_gemm_bench",
5012 srcs = [
5013 "bench/gemm.h",
5014 "bench/qs8-gemm.cc",
5015 "src/xnnpack/AlignedAllocator.h",
5016 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Frank Barchard31328cb2020-10-12 11:55:18 -07005017 copts = xnnpack_optional_ruy_copts() + xnnpack_optional_gemmlowp_copts(),
5018 deps = MICROKERNEL_BENCHMARK_DEPS + [":packing"] + xnnpack_optional_ruy_deps() + xnnpack_optional_gemmlowp_deps(),
Marat Dukhan595e1702020-07-31 10:12:52 -07005019)
5020
5021xnnpack_benchmark(
Marat Dukhan56bdd4a2020-08-03 19:47:04 -07005022 name = "qs8_requantization_bench",
5023 srcs = [
5024 "bench/qs8-requantization.cc",
Marat Dukhan56bdd4a2020-08-03 19:47:04 -07005025 "src/xnnpack/AlignedAllocator.h",
Frank Barchard04336c12020-10-22 16:48:55 -07005026 "src/xnnpack/requantization-stubs.h",
Marat Dukhan56bdd4a2020-08-03 19:47:04 -07005027 ] + MICROKERNEL_BENCHMARK_HDRS,
5028 deps = MICROKERNEL_BENCHMARK_DEPS,
5029)
5030
5031xnnpack_benchmark(
Marat Dukhan08b7a972020-07-14 18:17:29 -07005032 name = "qu8_gemm_bench",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005033 srcs = [
5034 "bench/gemm.h",
Marat Dukhan08b7a972020-07-14 18:17:29 -07005035 "bench/qu8-gemm.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005036 "src/xnnpack/AlignedAllocator.h",
5037 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07005038 copts = xnnpack_optional_ruy_copts() + xnnpack_optional_gemmlowp_copts(),
Marat Dukhanab582382020-07-06 13:32:08 -07005039 deps = MICROKERNEL_BENCHMARK_DEPS + [":packing"] + xnnpack_optional_ruy_deps() + xnnpack_optional_gemmlowp_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07005040)
5041
5042xnnpack_benchmark(
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07005043 name = "qu8_requantization_bench",
5044 srcs = [
5045 "bench/qu8-requantization.cc",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07005046 "src/xnnpack/AlignedAllocator.h",
Frank Barchard04336c12020-10-22 16:48:55 -07005047 "src/xnnpack/requantization-stubs.h",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07005048 ] + MICROKERNEL_BENCHMARK_HDRS,
5049 deps = MICROKERNEL_BENCHMARK_DEPS,
5050)
5051
5052xnnpack_benchmark(
Frank Barchard40d20fe2020-05-05 00:37:45 -07005053 name = "f16_igemm_bench",
5054 srcs = [
5055 "bench/f16-igemm.cc",
5056 "bench/conv.h",
5057 "bench/google/conv.h",
5058 "src/xnnpack/AlignedAllocator.h",
5059 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005060 deps = MICROKERNEL_BENCHMARK_DEPS + [
5061 ":indirection",
5062 ":packing",
5063 ],
Frank Barchard40d20fe2020-05-05 00:37:45 -07005064)
5065
5066xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005067 name = "f16_gemm_bench",
5068 srcs = [
5069 "bench/f16-gemm.cc",
5070 "bench/gemm.h",
5071 "src/xnnpack/AlignedAllocator.h",
5072 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005073 deps = MICROKERNEL_BENCHMARK_DEPS + [
5074 ":packing",
5075 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005076)
5077
5078xnnpack_benchmark(
Marat Dukhanbdb56f52020-02-05 21:42:49 -08005079 name = "f16_spmm_bench",
5080 srcs = [
5081 "bench/f16-spmm.cc",
Marat Dukhan1631e3e2020-12-06 19:29:31 -08005082 "bench/spmm.h",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08005083 "src/xnnpack/AlignedAllocator.h",
5084 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanbdb56f52020-02-05 21:42:49 -08005085 deps = MICROKERNEL_BENCHMARK_DEPS,
5086)
5087
5088xnnpack_benchmark(
Marat Dukhan6674d692021-05-05 22:27:00 -07005089 name = "f16_vrelu_bench",
5090 srcs = [
5091 "bench/f16-vrelu.cc",
5092 "src/xnnpack/AlignedAllocator.h",
5093 ] + MICROKERNEL_BENCHMARK_HDRS,
5094 deps = MICROKERNEL_BENCHMARK_DEPS,
5095)
5096
5097xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005098 name = "f32_igemm_bench",
5099 srcs = [
5100 "bench/f32-igemm.cc",
5101 "bench/conv.h",
5102 "src/xnnpack/AlignedAllocator.h",
5103 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005104 deps = MICROKERNEL_BENCHMARK_DEPS + [
5105 ":indirection",
5106 ":packing",
5107 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005108)
5109
5110xnnpack_benchmark(
5111 name = "f32_conv_hwc_bench",
5112 srcs = [
5113 "bench/f32-conv-hwc.cc",
5114 "bench/dconv.h",
5115 "src/xnnpack/AlignedAllocator.h",
5116 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005117 deps = MICROKERNEL_BENCHMARK_DEPS + [
5118 ":packing",
5119 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005120)
5121
5122xnnpack_benchmark(
Marat Dukhan1f29b802020-05-15 23:46:39 -07005123 name = "f32_conv_hwc2chw_bench",
Erich Elsen563df5f2019-10-23 08:02:21 -07005124 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07005125 "bench/f32-conv-hwc2chw.cc",
Erich Elsen563df5f2019-10-23 08:02:21 -07005126 "bench/dconv.h",
5127 "src/xnnpack/AlignedAllocator.h",
5128 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005129 deps = MICROKERNEL_BENCHMARK_DEPS + [
5130 ":packing",
5131 ],
Erich Elsen563df5f2019-10-23 08:02:21 -07005132)
5133
5134xnnpack_benchmark(
Frank Barchard5a599a62020-06-04 20:12:44 -07005135 name = "f16_dwconv_bench",
5136 srcs = [
5137 "bench/f16-dwconv.cc",
5138 "bench/dwconv.h",
5139 "bench/google/dwconv.h",
5140 "src/xnnpack/AlignedAllocator.h",
5141 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005142 deps = MICROKERNEL_BENCHMARK_DEPS + [
5143 ":indirection",
5144 ":packing",
5145 ],
Frank Barchard5a599a62020-06-04 20:12:44 -07005146)
5147
5148xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005149 name = "f32_dwconv_bench",
5150 srcs = [
5151 "bench/f32-dwconv.cc",
5152 "bench/dwconv.h",
5153 "src/xnnpack/AlignedAllocator.h",
5154 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005155 deps = MICROKERNEL_BENCHMARK_DEPS + [
5156 ":indirection",
5157 ":packing",
5158 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005159)
5160
5161xnnpack_benchmark(
Marat Dukhanbf715f92020-10-23 20:17:00 -07005162 name = "f32_dwconv2d_chw_bench",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005163 srcs = [
Marat Dukhanbf715f92020-10-23 20:17:00 -07005164 "bench/f32-dwconv2d-chw.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005165 "bench/dwconv.h",
5166 "src/xnnpack/AlignedAllocator.h",
5167 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005168 deps = MICROKERNEL_BENCHMARK_DEPS + [
5169 ":indirection",
5170 ":packing",
5171 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005172)
5173
5174xnnpack_benchmark(
5175 name = "f32_gemm_bench",
5176 srcs = [
5177 "bench/f32-gemm.cc",
5178 "bench/gemm.h",
5179 "src/xnnpack/AlignedAllocator.h",
5180 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07005181 copts = xnnpack_optional_ruy_copts(),
Marat Dukhanab582382020-07-06 13:32:08 -07005182 deps = MICROKERNEL_BENCHMARK_DEPS + [":packing"] + xnnpack_optional_ruy_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07005183)
5184
5185xnnpack_benchmark(
Marat Dukhan4c4eb002019-12-08 21:27:49 -08005186 name = "f32_raddexpminusmax_bench",
5187 srcs = [
5188 "bench/f32-raddexpminusmax.cc",
5189 "src/xnnpack/AlignedAllocator.h",
5190 ] + MICROKERNEL_BENCHMARK_HDRS,
5191 deps = MICROKERNEL_BENCHMARK_DEPS,
5192)
5193
5194xnnpack_benchmark(
5195 name = "f32_raddextexp_bench",
5196 srcs = [
5197 "bench/f32-raddextexp.cc",
5198 "src/xnnpack/AlignedAllocator.h",
5199 ] + MICROKERNEL_BENCHMARK_HDRS,
5200 deps = MICROKERNEL_BENCHMARK_DEPS,
5201)
5202
5203xnnpack_benchmark(
5204 name = "f32_raddstoreexpminusmax_bench",
5205 srcs = [
5206 "bench/f32-raddstoreexpminusmax.cc",
5207 "src/xnnpack/AlignedAllocator.h",
5208 ] + MICROKERNEL_BENCHMARK_HDRS,
5209 deps = MICROKERNEL_BENCHMARK_DEPS,
5210)
5211
5212xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005213 name = "f32_rmax_bench",
5214 srcs = [
5215 "bench/f32-rmax.cc",
5216 "src/xnnpack/AlignedAllocator.h",
5217 ] + MICROKERNEL_BENCHMARK_HDRS,
5218 deps = MICROKERNEL_BENCHMARK_DEPS,
5219)
5220
5221xnnpack_benchmark(
5222 name = "f32_spmm_bench",
5223 srcs = [
5224 "bench/f32-spmm.cc",
Marat Dukhan1631e3e2020-12-06 19:29:31 -08005225 "bench/spmm.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005226 "src/xnnpack/AlignedAllocator.h",
5227 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005228 deps = MICROKERNEL_BENCHMARK_DEPS,
5229)
5230
5231xnnpack_benchmark(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08005232 name = "f32_softmax_bench",
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07005233 srcs = [
Marat Dukhanfd8e6892020-01-27 15:25:25 -08005234 "bench/f32-softmax.cc",
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07005235 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07005236 copts = xnnpack_optional_dnnl_copts(),
Marat Dukhan8d3c6932020-03-06 20:27:27 -08005237 deps = MICROKERNEL_BENCHMARK_DEPS + xnnpack_optional_dnnl_deps(),
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07005238)
5239
5240xnnpack_benchmark(
Marat Dukhaned6baaf2020-12-01 15:07:08 -08005241 name = "f32_velu_bench",
5242 srcs = [
5243 "bench/f32-velu.cc",
5244 "src/xnnpack/AlignedAllocator.h",
5245 ] + MICROKERNEL_BENCHMARK_HDRS,
5246 deps = MICROKERNEL_BENCHMARK_DEPS,
5247)
5248
5249xnnpack_benchmark(
Marat Dukhan6674d692021-05-05 22:27:00 -07005250 name = "f32_vhswish_bench",
5251 srcs = [
5252 "bench/f32-vhswish.cc",
5253 "src/xnnpack/AlignedAllocator.h",
5254 ] + MICROKERNEL_BENCHMARK_HDRS,
5255 deps = MICROKERNEL_BENCHMARK_DEPS,
5256)
5257
5258xnnpack_benchmark(
5259 name = "f32_vrelu_bench",
5260 srcs = [
5261 "bench/f32-vrelu.cc",
5262 "src/xnnpack/AlignedAllocator.h",
5263 ] + MICROKERNEL_BENCHMARK_HDRS,
5264 deps = MICROKERNEL_BENCHMARK_DEPS,
5265)
5266
5267xnnpack_benchmark(
Marat Dukhan4c4eb002019-12-08 21:27:49 -08005268 name = "f32_vscaleexpminusmax_bench",
5269 srcs = [
5270 "bench/f32-vscaleexpminusmax.cc",
5271 "src/xnnpack/AlignedAllocator.h",
5272 ] + MICROKERNEL_BENCHMARK_HDRS,
5273 deps = MICROKERNEL_BENCHMARK_DEPS,
5274)
5275
5276xnnpack_benchmark(
5277 name = "f32_vscaleextexp_bench",
5278 srcs = [
5279 "bench/f32-vscaleextexp.cc",
5280 "src/xnnpack/AlignedAllocator.h",
5281 ] + MICROKERNEL_BENCHMARK_HDRS,
5282 deps = MICROKERNEL_BENCHMARK_DEPS,
5283)
5284
5285xnnpack_benchmark(
Marat Dukhan6674d692021-05-05 22:27:00 -07005286 name = "f32_vsigmoid_bench",
5287 srcs = [
5288 "bench/f32-vsigmoid.cc",
5289 "src/xnnpack/AlignedAllocator.h",
5290 ] + MICROKERNEL_BENCHMARK_HDRS,
5291 deps = MICROKERNEL_BENCHMARK_DEPS,
5292)
5293
5294xnnpack_benchmark(
Marat Dukhanf4db2f32020-06-30 10:55:30 -07005295 name = "f32_vsqrt_bench",
5296 srcs = [
5297 "bench/f32-vsqrt.cc",
5298 "src/xnnpack/AlignedAllocator.h",
5299 ] + MICROKERNEL_BENCHMARK_HDRS,
5300 deps = MICROKERNEL_BENCHMARK_DEPS,
5301)
5302
5303xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005304 name = "f32_im2col_gemm_bench",
5305 srcs = [
5306 "bench/f32-im2col-gemm.cc",
5307 "bench/conv.h",
5308 "src/xnnpack/AlignedAllocator.h",
5309 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005310 deps = MICROKERNEL_BENCHMARK_DEPS + [
5311 ":im2col",
5312 ":packing",
5313 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005314)
5315
Marat Dukhanfe7acb62020-03-09 19:30:05 -07005316xnnpack_benchmark(
Marat Dukhanffbf96a2020-05-14 02:59:08 -07005317 name = "rounding_bench",
5318 srcs = [
5319 "bench/rounding.cc",
Marat Dukhanffbf96a2020-05-14 02:59:08 -07005320 "src/xnnpack/AlignedAllocator.h",
Frank Barchard04336c12020-10-22 16:48:55 -07005321 "src/xnnpack/math-stubs.h",
Marat Dukhanffbf96a2020-05-14 02:59:08 -07005322 ] + MICROKERNEL_BENCHMARK_HDRS,
5323 deps = MICROKERNEL_BENCHMARK_DEPS,
5324)
5325
Marat Dukhan08c4a432019-10-03 09:29:21 -07005326########################### Benchmarks for operators ###########################
5327
5328xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005329 name = "average_pooling_bench",
5330 srcs = ["bench/average-pooling.cc"],
Marat Dukhan7a16d8b2020-03-11 04:22:44 -07005331 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07005332 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07005333 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07005334)
5335
5336xnnpack_benchmark(
Marat Dukhan87727142020-06-24 15:24:10 -07005337 name = "bankers_rounding_bench",
5338 srcs = ["bench/bankers-rounding.cc"],
5339 copts = xnnpack_optional_tflite_copts(),
5340 tags = ["nowin32"],
5341 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
5342)
5343
5344xnnpack_benchmark(
5345 name = "ceiling_bench",
5346 srcs = ["bench/ceiling.cc"],
5347 copts = xnnpack_optional_tflite_copts(),
5348 tags = ["nowin32"],
5349 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
5350)
5351
5352xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005353 name = "channel_shuffle_bench",
5354 srcs = ["bench/channel-shuffle.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07005355 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005356)
5357
5358xnnpack_benchmark(
5359 name = "convolution_bench",
5360 srcs = ["bench/convolution.cc"],
5361 copts = xnnpack_optional_tflite_copts() + xnnpack_optional_armcl_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07005362 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07005363 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps() + xnnpack_optional_armcl_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07005364)
5365
5366xnnpack_benchmark(
5367 name = "deconvolution_bench",
5368 srcs = ["bench/deconvolution.cc"],
5369 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07005370 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07005371 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07005372)
5373
5374xnnpack_benchmark(
Marat Dukhanb6bd4bc2020-12-01 17:01:40 -08005375 name = "elu_bench",
5376 srcs = ["bench/elu.cc"],
5377 copts = xnnpack_optional_tflite_copts(),
5378 tags = ["nowin32"],
5379 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
5380)
5381
5382xnnpack_benchmark(
Marat Dukhan87727142020-06-24 15:24:10 -07005383 name = "floor_bench",
5384 srcs = ["bench/floor.cc"],
5385 copts = xnnpack_optional_tflite_copts(),
5386 tags = ["nowin32"],
5387 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
5388)
5389
5390xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005391 name = "global_average_pooling_bench",
5392 srcs = ["bench/global-average-pooling.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07005393 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005394)
5395
5396xnnpack_benchmark(
Marat Dukhanad352602020-06-25 21:50:54 -07005397 name = "hardswish_bench",
5398 srcs = ["bench/hardswish.cc"],
5399 copts = xnnpack_optional_tflite_copts(),
5400 tags = ["nowin32"],
5401 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
5402)
5403
5404xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005405 name = "max_pooling_bench",
5406 srcs = ["bench/max-pooling.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07005407 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005408)
5409
5410xnnpack_benchmark(
5411 name = "sigmoid_bench",
5412 srcs = ["bench/sigmoid.cc"],
Marat Dukhanc3b9e862019-11-17 13:18:54 -08005413 copts = xnnpack_optional_tflite_copts(),
Marat Dukhanca2ba702020-04-24 01:31:47 -07005414 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07005415 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07005416)
5417
5418xnnpack_benchmark(
Marat Dukhan95b22432019-10-30 16:30:14 -07005419 name = "prelu_bench",
5420 srcs = ["bench/prelu.cc"],
5421 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07005422 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07005423 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan95b22432019-10-30 16:30:14 -07005424)
5425
5426xnnpack_benchmark(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08005427 name = "softmax_bench",
5428 srcs = ["bench/softmax.cc"],
Marat Dukhan9c0db962020-01-28 12:30:14 -08005429 copts = xnnpack_optional_tflite_copts(),
Marat Dukhanca2ba702020-04-24 01:31:47 -07005430 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07005431 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07005432)
5433
Marat Dukhan87727142020-06-24 15:24:10 -07005434xnnpack_benchmark(
Marat Dukhan6804bbd2020-06-30 19:26:11 -07005435 name = "square_root_bench",
5436 srcs = ["bench/square-root.cc"],
5437 copts = xnnpack_optional_tflite_copts(),
5438 tags = ["nowin32"],
5439 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
5440)
5441
5442xnnpack_benchmark(
Marat Dukhan87727142020-06-24 15:24:10 -07005443 name = "truncation_bench",
5444 srcs = ["bench/truncation.cc"],
5445 deps = OPERATOR_BENCHMARK_DEPS,
5446)
5447
Marat Dukhanc068bb62019-10-04 13:24:39 -07005448############################# End-to-end benchmarks ############################
5449
5450cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07005451 name = "fp32_mobilenet_v1",
5452 srcs = ["models/fp32-mobilenet-v1.cc"],
Marat Dukhanc068bb62019-10-04 13:24:39 -07005453 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08005454 copts = xnnpack_std_cxxopts(),
Marat Dukhanc068bb62019-10-04 13:24:39 -07005455 linkstatic = True,
5456 deps = [
5457 ":XNNPACK",
5458 "@pthreadpool",
5459 ],
5460)
5461
5462cc_library(
Marat Dukhan4cea2322021-03-09 09:35:36 -08005463 name = "fp32_sparse_mobilenet_v1",
5464 srcs = ["models/fp32-sparse-mobilenet-v1.cc"],
5465 hdrs = ["models/models.h"],
5466 copts = xnnpack_std_cxxopts(),
5467 linkstatic = True,
5468 deps = [
5469 ":XNNPACK",
5470 "@pthreadpool",
5471 ],
5472)
5473
5474cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07005475 name = "fp16_mobilenet_v1",
5476 srcs = ["models/fp16-mobilenet-v1.cc"],
5477 hdrs = ["models/models.h"],
5478 copts = xnnpack_std_cxxopts(),
5479 linkstatic = True,
5480 deps = [
5481 ":XNNPACK",
5482 "@FP16",
5483 "@pthreadpool",
5484 ],
5485)
5486
5487cc_library(
Marat Dukhan0743cdf2020-08-04 18:52:07 -07005488 name = "qs8_mobilenet_v1",
5489 srcs = ["models/qs8-mobilenet-v1.cc"],
5490 hdrs = ["models/models.h"],
5491 copts = xnnpack_std_cxxopts(),
5492 linkstatic = True,
5493 deps = [
5494 ":XNNPACK",
5495 "@pthreadpool",
5496 ],
5497)
5498
5499cc_library(
Marat Dukhan70a96182020-09-03 17:13:58 -07005500 name = "qs8_mobilenet_v2",
5501 srcs = ["models/qs8-mobilenet-v2.cc"],
5502 hdrs = ["models/models.h"],
5503 copts = xnnpack_std_cxxopts(),
5504 linkstatic = True,
5505 deps = [
5506 ":XNNPACK",
5507 "@pthreadpool",
5508 ],
5509)
5510
5511cc_library(
Marat Dukhan12a23bb2021-03-08 08:13:21 -08005512 name = "qu8_mobilenet_v1",
5513 srcs = ["models/qu8-mobilenet-v1.cc"],
5514 hdrs = ["models/models.h"],
5515 copts = xnnpack_std_cxxopts(),
5516 linkstatic = True,
5517 deps = [
5518 ":XNNPACK",
5519 "@pthreadpool",
5520 ],
5521)
5522
5523cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07005524 name = "fp32_mobilenet_v2",
5525 srcs = ["models/fp32-mobilenet-v2.cc"],
Marat Dukhanc068bb62019-10-04 13:24:39 -07005526 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08005527 copts = xnnpack_std_cxxopts(),
Marat Dukhanc068bb62019-10-04 13:24:39 -07005528 linkstatic = True,
5529 deps = [
5530 ":XNNPACK",
5531 "@pthreadpool",
5532 ],
5533)
5534
Marat Dukhanc08cdf52019-12-09 09:17:51 -08005535cc_library(
Marat Dukhan4cea2322021-03-09 09:35:36 -08005536 name = "fp32_sparse_mobilenet_v2",
5537 srcs = ["models/fp32-sparse-mobilenet-v2.cc"],
5538 hdrs = ["models/models.h"],
5539 copts = xnnpack_std_cxxopts(),
5540 linkstatic = True,
5541 deps = [
5542 ":XNNPACK",
5543 "@pthreadpool",
5544 ],
5545)
5546
5547cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07005548 name = "fp16_mobilenet_v2",
5549 srcs = ["models/fp16-mobilenet-v2.cc"],
5550 hdrs = ["models/models.h"],
5551 copts = xnnpack_std_cxxopts(),
5552 linkstatic = True,
5553 deps = [
5554 ":XNNPACK",
5555 "@FP16",
5556 "@pthreadpool",
5557 ],
5558)
5559
5560cc_library(
5561 name = "fp32_mobilenet_v3_large",
5562 srcs = ["models/fp32-mobilenet-v3-large.cc"],
Marat Dukhanc08cdf52019-12-09 09:17:51 -08005563 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08005564 copts = xnnpack_std_cxxopts(),
Marat Dukhanc08cdf52019-12-09 09:17:51 -08005565 linkstatic = True,
5566 deps = [
5567 ":XNNPACK",
5568 "@pthreadpool",
5569 ],
5570)
5571
5572cc_library(
Marat Dukhan4cea2322021-03-09 09:35:36 -08005573 name = "fp32_sparse_mobilenet_v3_large",
5574 srcs = ["models/fp32-sparse-mobilenet-v3-large.cc"],
5575 hdrs = ["models/models.h"],
5576 copts = xnnpack_std_cxxopts(),
5577 linkstatic = True,
5578 deps = [
5579 ":XNNPACK",
5580 "@pthreadpool",
5581 ],
5582)
5583
5584cc_library(
Marat Dukhan66f3ccd2020-09-14 16:09:38 -07005585 name = "fp16_mobilenet_v3_large",
5586 srcs = ["models/fp16-mobilenet-v3-large.cc"],
5587 hdrs = ["models/models.h"],
5588 copts = xnnpack_std_cxxopts(),
5589 linkstatic = True,
5590 deps = [
5591 ":XNNPACK",
5592 "@FP16",
5593 "@pthreadpool",
5594 ],
5595)
5596
5597cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07005598 name = "fp32_mobilenet_v3_small",
5599 srcs = ["models/fp32-mobilenet-v3-small.cc"],
Marat Dukhanc08cdf52019-12-09 09:17:51 -08005600 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08005601 copts = xnnpack_std_cxxopts(),
Marat Dukhanc08cdf52019-12-09 09:17:51 -08005602 linkstatic = True,
5603 deps = [
5604 ":XNNPACK",
5605 "@pthreadpool",
5606 ],
5607)
5608
Marat Dukhan66f3ccd2020-09-14 16:09:38 -07005609cc_library(
Marat Dukhan4cea2322021-03-09 09:35:36 -08005610 name = "fp32_sparse_mobilenet_v3_small",
5611 srcs = ["models/fp32-sparse-mobilenet-v3-small.cc"],
5612 hdrs = ["models/models.h"],
5613 copts = xnnpack_std_cxxopts(),
5614 linkstatic = True,
5615 deps = [
5616 ":XNNPACK",
5617 "@pthreadpool",
5618 ],
5619)
5620
5621cc_library(
Marat Dukhan66f3ccd2020-09-14 16:09:38 -07005622 name = "fp16_mobilenet_v3_small",
5623 srcs = ["models/fp16-mobilenet-v3-small.cc"],
5624 hdrs = ["models/models.h"],
5625 copts = xnnpack_std_cxxopts(),
5626 linkstatic = True,
5627 deps = [
5628 ":XNNPACK",
5629 "@FP16",
5630 "@pthreadpool",
5631 ],
5632)
5633
Marat Dukhanc068bb62019-10-04 13:24:39 -07005634xnnpack_benchmark(
Marat Dukhanef4416e2019-10-31 13:44:40 -07005635 name = "f32_dwconv_e2e_bench",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08005636 srcs = [
5637 "bench/f32-dwconv-e2e.cc",
5638 "bench/end2end.h",
5639 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanef4416e2019-10-31 13:44:40 -07005640 deps = MICROKERNEL_BENCHMARK_DEPS + [
5641 ":XNNPACK",
Marat Dukhan270a2c42020-06-26 16:45:52 -07005642 ":fp32_mobilenet_v1",
5643 ":fp32_mobilenet_v2",
5644 ":fp32_mobilenet_v3_large",
5645 ":fp32_mobilenet_v3_small",
Marat Dukhanef4416e2019-10-31 13:44:40 -07005646 ],
5647)
5648
5649xnnpack_benchmark(
Marat Dukhan5f18d262019-10-31 10:24:14 -07005650 name = "f32_gemm_e2e_bench",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08005651 srcs = [
5652 "bench/f32-gemm-e2e.cc",
5653 "bench/end2end.h",
5654 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan5f18d262019-10-31 10:24:14 -07005655 deps = MICROKERNEL_BENCHMARK_DEPS + [
5656 ":XNNPACK",
Marat Dukhan270a2c42020-06-26 16:45:52 -07005657 ":fp32_mobilenet_v1",
5658 ":fp32_mobilenet_v2",
5659 ":fp32_mobilenet_v3_large",
5660 ":fp32_mobilenet_v3_small",
Marat Dukhan5f18d262019-10-31 10:24:14 -07005661 ],
5662)
5663
5664xnnpack_benchmark(
Frank Barcharddc909cb2021-02-08 13:59:31 -08005665 name = "qs8_gemm_e2e_bench",
5666 srcs = [
5667 "bench/qs8-gemm-e2e.cc",
5668 "bench/end2end.h",
5669 ] + MICROKERNEL_BENCHMARK_HDRS,
5670 deps = MICROKERNEL_BENCHMARK_DEPS + [
5671 ":XNNPACK",
5672 ":qs8_mobilenet_v1",
5673 ":qs8_mobilenet_v2",
5674 ],
5675)
5676
5677xnnpack_benchmark(
Marat Dukhanc068bb62019-10-04 13:24:39 -07005678 name = "end2end_bench",
5679 srcs = ["bench/end2end.cc"],
5680 deps = [
5681 ":XNNPACK",
Frank Barchardc712fa42019-10-31 14:00:21 -07005682 ":bench_utils",
Marat Dukhan270a2c42020-06-26 16:45:52 -07005683 ":fp16_mobilenet_v1",
5684 ":fp16_mobilenet_v2",
Marat Dukhan66f3ccd2020-09-14 16:09:38 -07005685 ":fp16_mobilenet_v3_large",
5686 ":fp16_mobilenet_v3_small",
Marat Dukhan270a2c42020-06-26 16:45:52 -07005687 ":fp32_mobilenet_v1",
5688 ":fp32_mobilenet_v2",
5689 ":fp32_mobilenet_v3_large",
5690 ":fp32_mobilenet_v3_small",
Marat Dukhan4cea2322021-03-09 09:35:36 -08005691 ":fp32_sparse_mobilenet_v1",
5692 ":fp32_sparse_mobilenet_v2",
5693 ":fp32_sparse_mobilenet_v3_large",
5694 ":fp32_sparse_mobilenet_v3_small",
Marat Dukhan0743cdf2020-08-04 18:52:07 -07005695 ":qs8_mobilenet_v1",
Marat Dukhan70a96182020-09-03 17:13:58 -07005696 ":qs8_mobilenet_v2",
Marat Dukhan12a23bb2021-03-08 08:13:21 -08005697 ":qu8_mobilenet_v1",
Marat Dukhanc068bb62019-10-04 13:24:39 -07005698 "@pthreadpool",
5699 ],
5700)
5701
Marat Dukhan6adff4e2019-10-14 18:32:07 -07005702#################### Accuracy evaluation for math functions ####################
5703
5704xnnpack_benchmark(
Marat Dukhancf67ec62020-12-14 09:56:37 -08005705 name = "f32_exp_ulp_eval",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07005706 srcs = [
Marat Dukhancf67ec62020-12-14 09:56:37 -08005707 "eval/f32-exp-ulp.cc",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07005708 "src/xnnpack/AlignedAllocator.h",
5709 ] + ACCURACY_EVAL_HDRS,
Marat Dukhan3a305212020-12-06 19:24:27 -08005710 deps = ACCURACY_EVAL_DEPS + [
5711 ":bench_utils",
5712 "@cpuinfo",
5713 ],
Marat Dukhan6adff4e2019-10-14 18:32:07 -07005714)
5715
Marat Dukhan515c9772019-10-17 18:07:57 -07005716xnnpack_benchmark(
Marat Dukhancf67ec62020-12-14 09:56:37 -08005717 name = "f32_expminus_ulp_eval",
Marat Dukhan515c9772019-10-17 18:07:57 -07005718 srcs = [
Marat Dukhancf67ec62020-12-14 09:56:37 -08005719 "eval/f32-expminus-ulp.cc",
Marat Dukhan515c9772019-10-17 18:07:57 -07005720 "src/xnnpack/AlignedAllocator.h",
5721 ] + ACCURACY_EVAL_HDRS,
Marat Dukhan3a305212020-12-06 19:24:27 -08005722 deps = ACCURACY_EVAL_DEPS + [
5723 ":bench_utils",
5724 "@cpuinfo",
5725 ],
Marat Dukhan515c9772019-10-17 18:07:57 -07005726)
5727
Marat Dukhan98ba4412019-10-23 02:14:28 -07005728xnnpack_benchmark(
Marat Dukhancf67ec62020-12-14 09:56:37 -08005729 name = "f32_expm1minus_ulp_eval",
Marat Dukhana438aca2020-11-20 15:45:01 -08005730 srcs = [
Marat Dukhancf67ec62020-12-14 09:56:37 -08005731 "eval/f32-expm1minus-ulp.cc",
Marat Dukhana438aca2020-11-20 15:45:01 -08005732 "src/xnnpack/AlignedAllocator.h",
5733 ] + ACCURACY_EVAL_HDRS,
Marat Dukhande390d42020-11-29 19:32:18 -08005734 deps = ACCURACY_EVAL_DEPS + [
5735 ":bench_utils",
Marat Dukhan3a305212020-12-06 19:24:27 -08005736 "@cpuinfo",
Marat Dukhande390d42020-11-29 19:32:18 -08005737 ],
Marat Dukhana438aca2020-11-20 15:45:01 -08005738)
5739
5740xnnpack_benchmark(
Marat Dukhancf67ec62020-12-14 09:56:37 -08005741 name = "f32_extexp_ulp_eval",
Marat Dukhan98ba4412019-10-23 02:14:28 -07005742 srcs = [
Marat Dukhancf67ec62020-12-14 09:56:37 -08005743 "eval/f32-extexp-ulp.cc",
Marat Dukhan98ba4412019-10-23 02:14:28 -07005744 "src/xnnpack/AlignedAllocator.h",
5745 ] + ACCURACY_EVAL_HDRS,
Marat Dukhan3a305212020-12-06 19:24:27 -08005746 deps = ACCURACY_EVAL_DEPS + [
5747 ":bench_utils",
5748 "@cpuinfo",
5749 ],
Marat Dukhan98ba4412019-10-23 02:14:28 -07005750)
5751
Marat Dukhanf44f0222020-12-14 11:53:27 -08005752xnnpack_benchmark(
5753 name = "f32_sigmoid_ulp_eval",
5754 srcs = [
5755 "eval/f32-sigmoid-ulp.cc",
5756 "src/xnnpack/AlignedAllocator.h",
5757 ] + ACCURACY_EVAL_HDRS,
5758 deps = ACCURACY_EVAL_DEPS + [
5759 ":bench_utils",
5760 "@cpuinfo",
5761 ],
5762)
5763
5764xnnpack_benchmark(
5765 name = "f32_sqrt_ulp_eval",
5766 srcs = [
5767 "eval/f32-sqrt-ulp.cc",
5768 "src/xnnpack/AlignedAllocator.h",
5769 ] + ACCURACY_EVAL_HDRS,
5770 deps = ACCURACY_EVAL_DEPS + [
5771 ":bench_utils",
5772 "@cpuinfo",
5773 ],
5774)
5775
5776################### Accuracy verification for math functions ##################
5777
5778xnnpack_unit_test(
Marat Dukhanf7291fc2020-12-15 11:02:50 -08005779 name = "f32_exp_eval",
5780 srcs = [
5781 "eval/f32-exp.cc",
5782 "src/xnnpack/AlignedAllocator.h",
5783 "src/xnnpack/math-stubs.h",
5784 ] + MICROKERNEL_TEST_HDRS,
5785 automatic = False,
5786 deps = MICROKERNEL_TEST_DEPS,
5787)
5788
5789xnnpack_unit_test(
Marat Dukhanf44f0222020-12-14 11:53:27 -08005790 name = "f32_expm1minus_eval",
5791 srcs = [
5792 "eval/f32-expm1minus.cc",
5793 "src/xnnpack/AlignedAllocator.h",
5794 "src/xnnpack/math-stubs.h",
5795 ] + MICROKERNEL_TEST_HDRS,
5796 automatic = False,
5797 deps = MICROKERNEL_TEST_DEPS,
5798)
5799
Marat Dukhan8853b822020-05-07 12:19:01 -07005800xnnpack_unit_test(
Marat Dukhand28a5a22020-12-14 15:27:22 -08005801 name = "f32_expminus_eval",
5802 srcs = [
5803 "eval/f32-expminus.cc",
5804 "src/xnnpack/AlignedAllocator.h",
5805 "src/xnnpack/math-stubs.h",
5806 ] + MICROKERNEL_TEST_HDRS,
5807 automatic = False,
5808 deps = MICROKERNEL_TEST_DEPS,
5809)
5810
5811xnnpack_unit_test(
Marat Dukhan8853b822020-05-07 12:19:01 -07005812 name = "f32_roundne_eval",
5813 srcs = [
5814 "eval/f32-roundne.cc",
5815 "src/xnnpack/AlignedAllocator.h",
5816 "src/xnnpack/math-stubs.h",
5817 ] + MICROKERNEL_TEST_HDRS,
Marat Dukhan22eed3d2020-05-11 20:13:37 -07005818 automatic = False,
Marat Dukhan8853b822020-05-07 12:19:01 -07005819 deps = MICROKERNEL_TEST_DEPS,
5820)
5821
Marat Dukhan2dbb9442020-05-12 20:43:43 -07005822xnnpack_unit_test(
Marat Dukhanc9852ba2020-05-13 17:21:29 -07005823 name = "f32_roundd_eval",
5824 srcs = [
5825 "eval/f32-roundd.cc",
5826 "src/xnnpack/AlignedAllocator.h",
5827 "src/xnnpack/math-stubs.h",
5828 ] + MICROKERNEL_TEST_HDRS,
5829 automatic = False,
5830 deps = MICROKERNEL_TEST_DEPS,
5831)
5832
5833xnnpack_unit_test(
5834 name = "f32_roundu_eval",
5835 srcs = [
5836 "eval/f32-roundu.cc",
5837 "src/xnnpack/AlignedAllocator.h",
5838 "src/xnnpack/math-stubs.h",
5839 ] + MICROKERNEL_TEST_HDRS,
5840 automatic = False,
5841 deps = MICROKERNEL_TEST_DEPS,
5842)
5843
5844xnnpack_unit_test(
Marat Dukhan2dbb9442020-05-12 20:43:43 -07005845 name = "f32_roundz_eval",
5846 srcs = [
5847 "eval/f32-roundz.cc",
5848 "src/xnnpack/AlignedAllocator.h",
5849 "src/xnnpack/math-stubs.h",
5850 ] + MICROKERNEL_TEST_HDRS,
Marat Dukhanc9852ba2020-05-13 17:21:29 -07005851 automatic = False,
Marat Dukhan2dbb9442020-05-12 20:43:43 -07005852 deps = MICROKERNEL_TEST_DEPS,
5853)
5854
Marat Dukhan08c4a432019-10-03 09:29:21 -07005855######################### Unit tests for micro-kernels #########################
5856
5857xnnpack_unit_test(
Marat Dukhan6674d692021-05-05 22:27:00 -07005858 name = "f16_dwconv_minmax_test",
5859 srcs = [
5860 "test/f16-dwconv-minmax.cc",
5861 "test/dwconv-microkernel-tester.h",
5862 "src/xnnpack/AlignedAllocator.h",
5863 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
5864 deps = MICROKERNEL_TEST_DEPS + [":packing"],
5865)
5866
5867xnnpack_unit_test(
5868 name = "f16_gavgpool_minmax_test",
5869 srcs = [
5870 "test/f16-gavgpool-minmax.cc",
5871 "test/gavgpool-microkernel-tester.h",
5872 "src/xnnpack/AlignedAllocator.h",
5873 ] + MICROKERNEL_TEST_HDRS,
5874 deps = MICROKERNEL_TEST_DEPS,
5875)
5876
5877xnnpack_unit_test(
Marat Dukhande06f492020-04-09 00:19:31 -07005878 name = "f16_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005879 srcs = [
Marat Dukhande06f492020-04-09 00:19:31 -07005880 "test/f16-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005881 "test/gemm-microkernel-tester.h",
5882 "src/xnnpack/AlignedAllocator.h",
5883 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005884 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005885)
5886
5887xnnpack_unit_test(
Marat Dukhan6674d692021-05-05 22:27:00 -07005888 name = "f16_igemm_minmax_test",
5889 srcs = [
5890 "test/f16-igemm-minmax.cc",
5891 "test/gemm-microkernel-tester.h",
5892 "src/xnnpack/AlignedAllocator.h",
5893 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
5894 deps = MICROKERNEL_TEST_DEPS + [":packing"],
5895)
5896
5897xnnpack_unit_test(
Marat Dukhan355ab432020-04-09 19:01:52 -07005898 name = "f16_spmm_minmax_test",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08005899 srcs = [
Marat Dukhan355ab432020-04-09 19:01:52 -07005900 "test/f16-spmm-minmax.cc",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08005901 "test/spmm-microkernel-tester.h",
5902 "src/xnnpack/AlignedAllocator.h",
5903 ] + MICROKERNEL_TEST_HDRS,
5904 deps = MICROKERNEL_TEST_DEPS,
5905)
5906
5907xnnpack_unit_test(
Marat Dukhan6674d692021-05-05 22:27:00 -07005908 name = "f16_vadd_minmax_test",
5909 srcs = [
5910 "test/f16-vadd-minmax.cc",
5911 "test/vbinary-microkernel-tester.h",
5912 ] + MICROKERNEL_TEST_HDRS,
5913 deps = MICROKERNEL_TEST_DEPS,
5914)
5915
5916xnnpack_unit_test(
5917 name = "f16_vaddc_minmax_test",
5918 srcs = [
5919 "test/f16-vaddc-minmax.cc",
5920 "test/vbinaryc-microkernel-tester.h",
5921 ] + MICROKERNEL_TEST_HDRS,
5922 deps = MICROKERNEL_TEST_DEPS,
5923)
5924
5925xnnpack_unit_test(
5926 name = "f16_vclamp_test",
5927 srcs = [
5928 "test/f16-vclamp.cc",
5929 "test/vclamp-microkernel-tester.h",
5930 ] + MICROKERNEL_TEST_HDRS,
5931 deps = MICROKERNEL_TEST_DEPS,
5932)
5933
5934xnnpack_unit_test(
5935 name = "f16_vdiv_minmax_test",
5936 srcs = [
5937 "test/f16-vdiv-minmax.cc",
5938 "test/vbinary-microkernel-tester.h",
5939 ] + MICROKERNEL_TEST_HDRS,
5940 deps = MICROKERNEL_TEST_DEPS,
5941)
5942
5943xnnpack_unit_test(
5944 name = "f16_vdivc_minmax_test",
5945 srcs = [
5946 "test/f16-vdivc-minmax.cc",
5947 "test/vbinaryc-microkernel-tester.h",
5948 ] + MICROKERNEL_TEST_HDRS,
5949 deps = MICROKERNEL_TEST_DEPS,
5950)
5951
5952xnnpack_unit_test(
5953 name = "f16_vrdivc_minmax_test",
5954 srcs = [
5955 "test/f16-vrdivc-minmax.cc",
5956 "test/vbinaryc-microkernel-tester.h",
5957 ] + MICROKERNEL_TEST_HDRS,
5958 deps = MICROKERNEL_TEST_DEPS,
5959)
5960
5961xnnpack_unit_test(
5962 name = "f16_vhswish_test",
5963 srcs = [
5964 "test/f16-vhswish.cc",
5965 "test/vhswish-microkernel-tester.h",
5966 ] + MICROKERNEL_TEST_HDRS,
5967 deps = MICROKERNEL_TEST_DEPS,
5968)
5969
5970xnnpack_unit_test(
5971 name = "f16_vmax_test",
5972 srcs = [
5973 "test/f16-vmax.cc",
5974 "test/vbinary-microkernel-tester.h",
5975 ] + MICROKERNEL_TEST_HDRS,
5976 deps = MICROKERNEL_TEST_DEPS,
5977)
5978
5979xnnpack_unit_test(
5980 name = "f16_vmaxc_test",
5981 srcs = [
5982 "test/f16-vmaxc.cc",
5983 "test/vbinaryc-microkernel-tester.h",
5984 ] + MICROKERNEL_TEST_HDRS,
5985 deps = MICROKERNEL_TEST_DEPS,
5986)
5987
5988xnnpack_unit_test(
5989 name = "f16_vmin_test",
5990 srcs = [
5991 "test/f16-vmin.cc",
5992 "test/vbinary-microkernel-tester.h",
5993 ] + MICROKERNEL_TEST_HDRS,
5994 deps = MICROKERNEL_TEST_DEPS,
5995)
5996
5997xnnpack_unit_test(
5998 name = "f16_vminc_test",
5999 srcs = [
6000 "test/f16-vminc.cc",
6001 "test/vbinaryc-microkernel-tester.h",
6002 ] + MICROKERNEL_TEST_HDRS,
6003 deps = MICROKERNEL_TEST_DEPS,
6004)
6005
6006xnnpack_unit_test(
6007 name = "f16_vmul_minmax_test",
6008 srcs = [
6009 "test/f16-vmul-minmax.cc",
6010 "test/vbinary-microkernel-tester.h",
6011 ] + MICROKERNEL_TEST_HDRS,
6012 deps = MICROKERNEL_TEST_DEPS,
6013)
6014
6015xnnpack_unit_test(
6016 name = "f16_vmulc_minmax_test",
6017 srcs = [
6018 "test/f16-vmulc-minmax.cc",
6019 "test/vbinaryc-microkernel-tester.h",
6020 ] + MICROKERNEL_TEST_HDRS,
6021 deps = MICROKERNEL_TEST_DEPS,
6022)
6023
6024xnnpack_unit_test(
6025 name = "f16_vmulcaddc_minmax_test",
6026 srcs = [
6027 "test/f16-vmulcaddc-minmax.cc",
6028 "test/vmulcaddc-microkernel-tester.h",
6029 "src/xnnpack/AlignedAllocator.h",
6030 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
6031 deps = MICROKERNEL_TEST_DEPS + [":packing"],
6032)
6033
6034xnnpack_unit_test(
6035 name = "f16_vsub_minmax_test",
6036 srcs = [
6037 "test/f16-vsub-minmax.cc",
6038 "test/vbinary-microkernel-tester.h",
6039 ] + MICROKERNEL_TEST_HDRS,
6040 deps = MICROKERNEL_TEST_DEPS,
6041)
6042
6043xnnpack_unit_test(
6044 name = "f16_vsubc_minmax_test",
6045 srcs = [
6046 "test/f16-vsubc-minmax.cc",
6047 "test/vbinaryc-microkernel-tester.h",
6048 ] + MICROKERNEL_TEST_HDRS,
6049 deps = MICROKERNEL_TEST_DEPS,
6050)
6051
6052xnnpack_unit_test(
6053 name = "f16_vrsubc_minmax_test",
6054 srcs = [
6055 "test/f16-vrsubc-minmax.cc",
6056 "test/vbinaryc-microkernel-tester.h",
6057 ] + MICROKERNEL_TEST_HDRS,
6058 deps = MICROKERNEL_TEST_DEPS,
6059)
6060
6061xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07006062 name = "f32_argmaxpool_test",
6063 srcs = [
6064 "test/f32-argmaxpool.cc",
6065 "test/argmaxpool-microkernel-tester.h",
6066 "src/xnnpack/AlignedAllocator.h",
6067 ] + MICROKERNEL_TEST_HDRS,
6068 deps = MICROKERNEL_TEST_DEPS,
6069)
6070
6071xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07006072 name = "f32_avgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006073 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07006074 "test/f32-avgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006075 "test/avgpool-microkernel-tester.h",
6076 "src/xnnpack/AlignedAllocator.h",
6077 ] + MICROKERNEL_TEST_HDRS,
6078 deps = MICROKERNEL_TEST_DEPS,
6079)
6080
6081xnnpack_unit_test(
Marat Dukhan660fd192020-03-10 04:55:30 -07006082 name = "f32_ibilinear_test",
Marat Dukhan35dacfb2019-11-07 19:18:16 -08006083 srcs = [
Marat Dukhan660fd192020-03-10 04:55:30 -07006084 "test/f32-ibilinear.cc",
6085 "test/ibilinear-microkernel-tester.h",
Marat Dukhan35dacfb2019-11-07 19:18:16 -08006086 "src/xnnpack/AlignedAllocator.h",
6087 ] + MICROKERNEL_TEST_HDRS,
6088 deps = MICROKERNEL_TEST_DEPS,
6089)
6090
6091xnnpack_unit_test(
XNNPACK Team21432672020-10-19 19:58:48 -07006092 name = "f32_ibilinear_chw_test",
6093 srcs = [
6094 "test/f32-ibilinear-chw.cc",
XNNPACK Team6be46b22020-10-22 23:34:54 -07006095 "test/ibilinear-microkernel-tester.h",
XNNPACK Team21432672020-10-19 19:58:48 -07006096 "src/xnnpack/AlignedAllocator.h",
6097 ] + MICROKERNEL_TEST_HDRS,
6098 deps = MICROKERNEL_TEST_DEPS,
6099)
6100
6101xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07006102 name = "f32_igemm_test",
6103 srcs = [
6104 "test/f32-igemm.cc",
6105 "test/gemm-microkernel-tester.h",
6106 "src/xnnpack/AlignedAllocator.h",
6107 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07006108 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan163a7e62020-04-09 04:19:26 -07006109)
6110
6111xnnpack_unit_test(
Marat Dukhan467f6362020-05-22 23:21:55 -07006112 name = "f32_igemm_relu_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006113 srcs = [
Marat Dukhan467f6362020-05-22 23:21:55 -07006114 "test/f32-igemm-relu.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006115 "test/gemm-microkernel-tester.h",
6116 "src/xnnpack/AlignedAllocator.h",
6117 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07006118 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07006119)
6120
6121xnnpack_unit_test(
Marat Dukhane207b7b2020-05-28 16:27:42 -07006122 name = "f32_igemm_minmax_test",
6123 srcs = [
6124 "test/f32-igemm-minmax.cc",
6125 "test/gemm-microkernel-tester.h",
6126 "src/xnnpack/AlignedAllocator.h",
6127 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07006128 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhane207b7b2020-05-28 16:27:42 -07006129)
6130
6131xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07006132 name = "f32_conv_hwc_test",
6133 srcs = [
6134 "test/f32-conv-hwc.cc",
6135 "test/conv-hwc-microkernel-tester.h",
6136 "src/xnnpack/AlignedAllocator.h",
6137 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07006138 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07006139)
6140
6141xnnpack_unit_test(
Marat Dukhan1f29b802020-05-15 23:46:39 -07006142 name = "f32_conv_hwc2chw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006143 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07006144 "test/f32-conv-hwc2chw.cc",
6145 "test/conv-hwc2chw-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006146 "src/xnnpack/AlignedAllocator.h",
6147 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07006148 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07006149)
6150
6151xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07006152 name = "f32_dwconv_test",
6153 srcs = [
6154 "test/f32-dwconv.cc",
6155 "test/dwconv-microkernel-tester.h",
6156 "src/xnnpack/AlignedAllocator.h",
6157 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07006158 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan163a7e62020-04-09 04:19:26 -07006159)
6160
6161xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07006162 name = "f32_dwconv_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006163 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07006164 "test/f32-dwconv-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006165 "test/dwconv-microkernel-tester.h",
6166 "src/xnnpack/AlignedAllocator.h",
6167 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07006168 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07006169)
6170
6171xnnpack_unit_test(
Marat Dukhanbf715f92020-10-23 20:17:00 -07006172 name = "f32_dwconv2d_chw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006173 srcs = [
Marat Dukhanbf715f92020-10-23 20:17:00 -07006174 "test/f32-dwconv2d-chw.cc",
6175 "test/dwconv2d-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006176 "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 Dukhan08c4a432019-10-03 09:29:21 -07006179)
6180
6181xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07006182 name = "f32_gavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006183 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07006184 "test/f32-gavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006185 "test/gavgpool-microkernel-tester.h",
6186 "src/xnnpack/AlignedAllocator.h",
6187 ] + MICROKERNEL_TEST_HDRS,
6188 deps = MICROKERNEL_TEST_DEPS,
6189)
6190
6191xnnpack_unit_test(
Marat Dukhan1f29b802020-05-15 23:46:39 -07006192 name = "f32_gavgpool_cw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006193 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07006194 "test/f32-gavgpool-cw.cc",
6195 "test/gavgpool-cw-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006196 "src/xnnpack/AlignedAllocator.h",
6197 ] + MICROKERNEL_TEST_HDRS,
6198 deps = MICROKERNEL_TEST_DEPS,
6199)
6200
6201xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07006202 name = "f32_gemm_test",
6203 srcs = [
6204 "test/f32-gemm.cc",
6205 "test/gemm-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 Dukhan163a7e62020-04-09 04:19:26 -07006209)
6210
6211xnnpack_unit_test(
Marat Dukhan467f6362020-05-22 23:21:55 -07006212 name = "f32_gemm_relu_test",
6213 srcs = [
6214 "test/f32-gemm-relu.cc",
6215 "test/gemm-microkernel-tester.h",
6216 "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 Dukhan467f6362020-05-22 23:21:55 -07006219)
6220
6221xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07006222 name = "f32_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006223 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07006224 "test/f32-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006225 "test/gemm-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 Dukhan08c4a432019-10-03 09:29:21 -07006229)
6230
6231xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07006232 name = "f32_gemminc_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006233 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07006234 "test/f32-gemminc-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006235 "test/gemm-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 Dukhan6674d692021-05-05 22:27:00 -07006242 name = "f32_vhswish_test",
Frank Barchardb1966592020-05-12 13:47:06 -07006243 srcs = [
Marat Dukhan6674d692021-05-05 22:27:00 -07006244 "test/f32-vhswish.cc",
6245 "test/vhswish-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006246 ] + MICROKERNEL_TEST_HDRS,
6247 deps = MICROKERNEL_TEST_DEPS,
6248)
6249
6250xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07006251 name = "f32_maxpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006252 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07006253 "test/f32-maxpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006254 "test/maxpool-microkernel-tester.h",
6255 ] + MICROKERNEL_TEST_HDRS,
6256 deps = MICROKERNEL_TEST_DEPS,
6257)
6258
6259xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07006260 name = "f32_pavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006261 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07006262 "test/f32-pavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006263 "test/avgpool-microkernel-tester.h",
6264 "src/xnnpack/AlignedAllocator.h",
6265 ] + MICROKERNEL_TEST_HDRS,
6266 deps = MICROKERNEL_TEST_DEPS,
6267)
6268
6269xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07006270 name = "f32_ppmm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006271 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07006272 "test/f32-ppmm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006273 "test/gemm-microkernel-tester.h",
6274 "src/xnnpack/AlignedAllocator.h",
6275 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07006276 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07006277)
6278
6279xnnpack_unit_test(
Frank Barchardb1966592020-05-12 13:47:06 -07006280 name = "f16_prelu_test",
6281 srcs = [
6282 "test/f16-prelu.cc",
6283 "test/prelu-microkernel-tester.h",
6284 "src/xnnpack/AlignedAllocator.h",
6285 ] + MICROKERNEL_TEST_HDRS,
6286 deps = MICROKERNEL_TEST_DEPS,
6287)
6288
6289xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07006290 name = "f32_prelu_test",
6291 srcs = [
6292 "test/f32-prelu.cc",
6293 "test/prelu-microkernel-tester.h",
6294 "src/xnnpack/AlignedAllocator.h",
6295 ] + MICROKERNEL_TEST_HDRS,
6296 deps = MICROKERNEL_TEST_DEPS,
6297)
6298
6299xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07006300 name = "f32_raddexpminusmax_test",
6301 srcs = [
6302 "test/f32-raddexpminusmax.cc",
6303 "test/raddexpminusmax-microkernel-tester.h",
6304 ] + MICROKERNEL_TEST_HDRS,
6305 deps = MICROKERNEL_TEST_DEPS,
6306)
6307
6308xnnpack_unit_test(
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07006309 name = "f32_raddextexp_test",
6310 srcs = [
6311 "test/f32-raddextexp.cc",
6312 "test/raddextexp-microkernel-tester.h",
6313 ] + MICROKERNEL_TEST_HDRS,
6314 deps = MICROKERNEL_TEST_DEPS,
6315)
6316
6317xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07006318 name = "f32_raddstoreexpminusmax_test",
6319 srcs = [
6320 "test/f32-raddstoreexpminusmax.cc",
6321 "test/raddstoreexpminusmax-microkernel-tester.h",
6322 ] + MICROKERNEL_TEST_HDRS,
6323 deps = MICROKERNEL_TEST_DEPS,
6324)
6325
6326xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07006327 name = "f32_rmax_test",
6328 srcs = [
6329 "test/f32-rmax.cc",
6330 "test/rmax-microkernel-tester.h",
6331 ] + MICROKERNEL_TEST_HDRS,
6332 deps = MICROKERNEL_TEST_DEPS,
6333)
6334
6335xnnpack_unit_test(
Marat Dukhan355ab432020-04-09 19:01:52 -07006336 name = "f32_spmm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006337 srcs = [
Marat Dukhan355ab432020-04-09 19:01:52 -07006338 "test/f32-spmm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006339 "test/spmm-microkernel-tester.h",
6340 "src/xnnpack/AlignedAllocator.h",
6341 ] + MICROKERNEL_TEST_HDRS,
6342 deps = MICROKERNEL_TEST_DEPS,
6343)
6344
6345xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07006346 name = "f32_vabs_test",
6347 srcs = [
6348 "test/f32-vabs.cc",
6349 "test/vunary-microkernel-tester.h",
6350 ] + MICROKERNEL_TEST_HDRS,
6351 deps = MICROKERNEL_TEST_DEPS,
6352)
6353
6354xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07006355 name = "f32_vadd_test",
6356 srcs = [
6357 "test/f32-vadd.cc",
6358 "test/vbinary-microkernel-tester.h",
6359 ] + MICROKERNEL_TEST_HDRS,
6360 deps = MICROKERNEL_TEST_DEPS,
6361)
6362
6363xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006364 name = "f32_vadd_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006365 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006366 "test/f32-vadd-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08006367 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08006368 ] + MICROKERNEL_TEST_HDRS,
6369 deps = MICROKERNEL_TEST_DEPS,
6370)
6371
6372xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07006373 name = "f32_vadd_relu_test",
6374 srcs = [
6375 "test/f32-vadd-relu.cc",
6376 "test/vbinary-microkernel-tester.h",
6377 ] + MICROKERNEL_TEST_HDRS,
6378 deps = MICROKERNEL_TEST_DEPS,
6379)
6380
6381xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07006382 name = "f32_vaddc_test",
6383 srcs = [
6384 "test/f32-vaddc.cc",
6385 "test/vbinaryc-microkernel-tester.h",
6386 ] + MICROKERNEL_TEST_HDRS,
6387 deps = MICROKERNEL_TEST_DEPS,
6388)
6389
6390xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006391 name = "f32_vaddc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08006392 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006393 "test/f32-vaddc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08006394 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006395 ] + MICROKERNEL_TEST_HDRS,
6396 deps = MICROKERNEL_TEST_DEPS,
6397)
6398
6399xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07006400 name = "f32_vaddc_relu_test",
6401 srcs = [
6402 "test/f32-vaddc-relu.cc",
6403 "test/vbinaryc-microkernel-tester.h",
6404 ] + MICROKERNEL_TEST_HDRS,
6405 deps = MICROKERNEL_TEST_DEPS,
6406)
6407
6408xnnpack_unit_test(
Marat Dukhan6674d692021-05-05 22:27:00 -07006409 name = "f32_vclamp_test",
6410 srcs = [
6411 "test/f32-vclamp.cc",
6412 "test/vclamp-microkernel-tester.h",
6413 ] + MICROKERNEL_TEST_HDRS,
6414 deps = MICROKERNEL_TEST_DEPS,
6415)
6416
6417xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07006418 name = "f32_vdiv_test",
6419 srcs = [
6420 "test/f32-vdiv.cc",
6421 "test/vbinary-microkernel-tester.h",
6422 ] + MICROKERNEL_TEST_HDRS,
6423 deps = MICROKERNEL_TEST_DEPS,
6424)
6425
6426xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006427 name = "f32_vdiv_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08006428 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006429 "test/f32-vdiv-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08006430 "test/vbinary-microkernel-tester.h",
6431 ] + MICROKERNEL_TEST_HDRS,
6432 deps = MICROKERNEL_TEST_DEPS,
6433)
6434
6435xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07006436 name = "f32_vdiv_relu_test",
6437 srcs = [
6438 "test/f32-vdiv-relu.cc",
6439 "test/vbinary-microkernel-tester.h",
6440 ] + MICROKERNEL_TEST_HDRS,
6441 deps = MICROKERNEL_TEST_DEPS,
6442)
6443
6444xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07006445 name = "f32_vdivc_test",
6446 srcs = [
6447 "test/f32-vdivc.cc",
6448 "test/vbinaryc-microkernel-tester.h",
6449 ] + MICROKERNEL_TEST_HDRS,
6450 deps = MICROKERNEL_TEST_DEPS,
6451)
6452
6453xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006454 name = "f32_vdivc_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08006455 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006456 "test/f32-vdivc-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08006457 "test/vbinaryc-microkernel-tester.h",
6458 ] + MICROKERNEL_TEST_HDRS,
6459 deps = MICROKERNEL_TEST_DEPS,
6460)
6461
6462xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07006463 name = "f32_vdivc_relu_test",
6464 srcs = [
6465 "test/f32-vdivc-relu.cc",
6466 "test/vbinaryc-microkernel-tester.h",
6467 ] + MICROKERNEL_TEST_HDRS,
6468 deps = MICROKERNEL_TEST_DEPS,
6469)
6470
6471xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07006472 name = "f32_vrdivc_test",
6473 srcs = [
6474 "test/f32-vrdivc.cc",
6475 "test/vbinaryc-microkernel-tester.h",
6476 ] + MICROKERNEL_TEST_HDRS,
6477 deps = MICROKERNEL_TEST_DEPS,
6478)
6479
6480xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006481 name = "f32_vrdivc_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08006482 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006483 "test/f32-vrdivc-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08006484 "test/vbinaryc-microkernel-tester.h",
6485 ] + MICROKERNEL_TEST_HDRS,
6486 deps = MICROKERNEL_TEST_DEPS,
6487)
6488
6489xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07006490 name = "f32_vrdivc_relu_test",
6491 srcs = [
6492 "test/f32-vrdivc-relu.cc",
6493 "test/vbinaryc-microkernel-tester.h",
6494 ] + MICROKERNEL_TEST_HDRS,
6495 deps = MICROKERNEL_TEST_DEPS,
6496)
6497
6498xnnpack_unit_test(
Marat Dukhaned6baaf2020-12-01 15:07:08 -08006499 name = "f32_velu_test",
6500 srcs = [
6501 "test/f32-velu.cc",
6502 "test/vunary-microkernel-tester.h",
6503 ] + MICROKERNEL_TEST_HDRS,
6504 deps = MICROKERNEL_TEST_DEPS,
6505)
6506
6507xnnpack_unit_test(
Marat Dukhan403b7d42019-12-05 12:49:11 -08006508 name = "f32_vmax_test",
6509 srcs = [
6510 "test/f32-vmax.cc",
6511 "test/vbinary-microkernel-tester.h",
6512 ] + MICROKERNEL_TEST_HDRS,
6513 deps = MICROKERNEL_TEST_DEPS,
6514)
6515
6516xnnpack_unit_test(
6517 name = "f32_vmaxc_test",
6518 srcs = [
6519 "test/f32-vmaxc.cc",
6520 "test/vbinaryc-microkernel-tester.h",
6521 ] + MICROKERNEL_TEST_HDRS,
6522 deps = MICROKERNEL_TEST_DEPS,
6523)
6524
6525xnnpack_unit_test(
6526 name = "f32_vmin_test",
6527 srcs = [
6528 "test/f32-vmin.cc",
6529 "test/vbinary-microkernel-tester.h",
6530 ] + MICROKERNEL_TEST_HDRS,
6531 deps = MICROKERNEL_TEST_DEPS,
6532)
6533
6534xnnpack_unit_test(
6535 name = "f32_vminc_test",
6536 srcs = [
6537 "test/f32-vminc.cc",
6538 "test/vbinaryc-microkernel-tester.h",
6539 ] + MICROKERNEL_TEST_HDRS,
6540 deps = MICROKERNEL_TEST_DEPS,
6541)
6542
6543xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07006544 name = "f32_vmul_test",
6545 srcs = [
6546 "test/f32-vmul.cc",
6547 "test/vbinary-microkernel-tester.h",
6548 ] + MICROKERNEL_TEST_HDRS,
6549 deps = MICROKERNEL_TEST_DEPS,
6550)
6551
6552xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006553 name = "f32_vmul_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006554 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006555 "test/f32-vmul-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08006556 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08006557 ] + MICROKERNEL_TEST_HDRS,
6558 deps = MICROKERNEL_TEST_DEPS,
6559)
6560
6561xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07006562 name = "f32_vmul_relu_test",
6563 srcs = [
6564 "test/f32-vmul-relu.cc",
6565 "test/vbinary-microkernel-tester.h",
6566 ] + MICROKERNEL_TEST_HDRS,
6567 deps = MICROKERNEL_TEST_DEPS,
6568)
6569
6570xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07006571 name = "f32_vmulc_test",
6572 srcs = [
6573 "test/f32-vmulc.cc",
6574 "test/vbinaryc-microkernel-tester.h",
6575 ] + MICROKERNEL_TEST_HDRS,
6576 deps = MICROKERNEL_TEST_DEPS,
6577)
6578
6579xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006580 name = "f32_vmulc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08006581 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006582 "test/f32-vmulc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08006583 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006584 ] + MICROKERNEL_TEST_HDRS,
6585 deps = MICROKERNEL_TEST_DEPS,
6586)
6587
6588xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07006589 name = "f32_vmulc_relu_test",
6590 srcs = [
6591 "test/f32-vmulc-relu.cc",
6592 "test/vbinaryc-microkernel-tester.h",
6593 ] + MICROKERNEL_TEST_HDRS,
6594 deps = MICROKERNEL_TEST_DEPS,
6595)
6596
6597xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07006598 name = "f32_vmulcaddc_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006599 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07006600 "test/f32-vmulcaddc-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006601 "test/vmulcaddc-microkernel-tester.h",
6602 "src/xnnpack/AlignedAllocator.h",
6603 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07006604 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07006605)
6606
6607xnnpack_unit_test(
Marat Dukhan8cc7efe2020-06-10 16:24:27 -07006608 name = "f32_vlrelu_test",
6609 srcs = [
6610 "test/f32-vlrelu.cc",
6611 "test/vunary-microkernel-tester.h",
6612 ] + MICROKERNEL_TEST_HDRS,
6613 deps = MICROKERNEL_TEST_DEPS,
6614)
6615
6616xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07006617 name = "f32_vneg_test",
6618 srcs = [
6619 "test/f32-vneg.cc",
6620 "test/vunary-microkernel-tester.h",
6621 ] + MICROKERNEL_TEST_HDRS,
6622 deps = MICROKERNEL_TEST_DEPS,
6623)
6624
6625xnnpack_unit_test(
Marat Dukhan6674d692021-05-05 22:27:00 -07006626 name = "f32_vrelu_test",
6627 srcs = [
6628 "test/f32-vrelu.cc",
6629 "test/vunary-microkernel-tester.h",
6630 ] + MICROKERNEL_TEST_HDRS,
6631 deps = MICROKERNEL_TEST_DEPS,
6632)
6633
6634xnnpack_unit_test(
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07006635 name = "f32_vrndne_test",
6636 srcs = [
6637 "test/f32-vrndne.cc",
6638 "test/vunary-microkernel-tester.h",
6639 ] + MICROKERNEL_TEST_HDRS,
6640 deps = MICROKERNEL_TEST_DEPS,
6641)
6642
6643xnnpack_unit_test(
6644 name = "f32_vrndz_test",
6645 srcs = [
6646 "test/f32-vrndz.cc",
6647 "test/vunary-microkernel-tester.h",
6648 ] + MICROKERNEL_TEST_HDRS,
6649 deps = MICROKERNEL_TEST_DEPS,
6650)
6651
6652xnnpack_unit_test(
6653 name = "f32_vrndu_test",
6654 srcs = [
6655 "test/f32-vrndu.cc",
6656 "test/vunary-microkernel-tester.h",
6657 ] + MICROKERNEL_TEST_HDRS,
6658 deps = MICROKERNEL_TEST_DEPS,
6659)
6660
6661xnnpack_unit_test(
6662 name = "f32_vrndd_test",
6663 srcs = [
6664 "test/f32-vrndd.cc",
6665 "test/vunary-microkernel-tester.h",
6666 ] + MICROKERNEL_TEST_HDRS,
6667 deps = MICROKERNEL_TEST_DEPS,
6668)
6669
6670xnnpack_unit_test(
Marat Dukhan05ac8e32019-10-21 15:39:33 -07006671 name = "f32_vscale_test",
6672 srcs = [
6673 "test/f32-vscale.cc",
6674 "test/vscale-microkernel-tester.h",
6675 ] + MICROKERNEL_TEST_HDRS,
6676 deps = MICROKERNEL_TEST_DEPS,
6677)
6678
6679xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07006680 name = "f32_vscaleexpminusmax_test",
6681 srcs = [
6682 "test/f32-vscaleexpminusmax.cc",
6683 "test/vscaleexpminusmax-microkernel-tester.h",
6684 ] + MICROKERNEL_TEST_HDRS,
6685 deps = MICROKERNEL_TEST_DEPS,
6686)
6687
6688xnnpack_unit_test(
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07006689 name = "f32_vscaleextexp_test",
6690 srcs = [
6691 "test/f32-vscaleextexp.cc",
6692 "test/vscaleextexp-microkernel-tester.h",
6693 ] + MICROKERNEL_TEST_HDRS,
6694 deps = MICROKERNEL_TEST_DEPS,
6695)
6696
6697xnnpack_unit_test(
Marat Dukhan6674d692021-05-05 22:27:00 -07006698 name = "f32_vsigmoid_test",
6699 srcs = [
6700 "test/f32-vsigmoid.cc",
6701 "test/vunary-microkernel-tester.h",
6702 ] + MICROKERNEL_TEST_HDRS,
6703 deps = MICROKERNEL_TEST_DEPS,
6704)
6705
6706xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07006707 name = "f32_vsqr_test",
6708 srcs = [
6709 "test/f32-vsqr.cc",
6710 "test/vunary-microkernel-tester.h",
6711 ] + MICROKERNEL_TEST_HDRS,
6712 deps = MICROKERNEL_TEST_DEPS,
6713)
6714
6715xnnpack_unit_test(
Marat Dukhan13bafb02020-06-05 00:43:11 -07006716 name = "f32_vsqrdiff_test",
6717 srcs = [
6718 "test/f32-vsqrdiff.cc",
6719 "test/vbinary-microkernel-tester.h",
6720 ] + MICROKERNEL_TEST_HDRS,
6721 deps = MICROKERNEL_TEST_DEPS,
6722)
6723
6724xnnpack_unit_test(
6725 name = "f32_vsqrdiffc_test",
6726 srcs = [
6727 "test/f32-vsqrdiffc.cc",
6728 "test/vbinaryc-microkernel-tester.h",
6729 ] + MICROKERNEL_TEST_HDRS,
6730 deps = MICROKERNEL_TEST_DEPS,
6731)
6732
6733xnnpack_unit_test(
Marat Dukhanf4db2f32020-06-30 10:55:30 -07006734 name = "f32_vsqrt_test",
6735 srcs = [
6736 "test/f32-vsqrt.cc",
6737 "test/vunary-microkernel-tester.h",
6738 ] + MICROKERNEL_TEST_HDRS,
6739 deps = MICROKERNEL_TEST_DEPS,
6740)
6741
6742xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07006743 name = "f32_vsub_test",
6744 srcs = [
6745 "test/f32-vsub.cc",
6746 "test/vbinary-microkernel-tester.h",
6747 ] + MICROKERNEL_TEST_HDRS,
6748 deps = MICROKERNEL_TEST_DEPS,
6749)
6750
6751xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006752 name = "f32_vsub_minmax_test",
Marat Dukhan97579532019-10-18 16:40:39 -07006753 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006754 "test/f32-vsub-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08006755 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08006756 ] + MICROKERNEL_TEST_HDRS,
6757 deps = MICROKERNEL_TEST_DEPS,
6758)
6759
6760xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07006761 name = "f32_vsub_relu_test",
6762 srcs = [
6763 "test/f32-vsub-relu.cc",
6764 "test/vbinary-microkernel-tester.h",
6765 ] + MICROKERNEL_TEST_HDRS,
6766 deps = MICROKERNEL_TEST_DEPS,
6767)
6768
6769xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07006770 name = "f32_vsubc_test",
6771 srcs = [
6772 "test/f32-vsubc.cc",
6773 "test/vbinaryc-microkernel-tester.h",
6774 ] + MICROKERNEL_TEST_HDRS,
6775 deps = MICROKERNEL_TEST_DEPS,
6776)
6777
6778xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006779 name = "f32_vsubc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08006780 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006781 "test/f32-vsubc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08006782 "test/vbinaryc-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08006783 ] + MICROKERNEL_TEST_HDRS,
6784 deps = MICROKERNEL_TEST_DEPS,
6785)
6786
6787xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07006788 name = "f32_vsubc_relu_test",
6789 srcs = [
6790 "test/f32-vsubc-relu.cc",
6791 "test/vbinaryc-microkernel-tester.h",
6792 ] + MICROKERNEL_TEST_HDRS,
6793 deps = MICROKERNEL_TEST_DEPS,
6794)
6795
6796xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07006797 name = "f32_vrsubc_test",
6798 srcs = [
6799 "test/f32-vrsubc.cc",
6800 "test/vbinaryc-microkernel-tester.h",
6801 ] + MICROKERNEL_TEST_HDRS,
6802 deps = MICROKERNEL_TEST_DEPS,
6803)
6804
6805xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006806 name = "f32_vrsubc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08006807 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006808 "test/f32-vrsubc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08006809 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan97579532019-10-18 16:40:39 -07006810 ] + MICROKERNEL_TEST_HDRS,
6811 deps = MICROKERNEL_TEST_DEPS,
6812)
6813
6814xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07006815 name = "f32_vrsubc_relu_test",
6816 srcs = [
6817 "test/f32-vrsubc-relu.cc",
6818 "test/vbinaryc-microkernel-tester.h",
6819 ] + MICROKERNEL_TEST_HDRS,
6820 deps = MICROKERNEL_TEST_DEPS,
6821)
6822
6823xnnpack_unit_test(
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07006824 name = "qs8_dwconv_minmax_test",
6825 srcs = [
6826 "test/qs8-dwconv-minmax.cc",
6827 "test/dwconv-microkernel-tester.h",
6828 "src/xnnpack/AlignedAllocator.h",
6829 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
6830 deps = MICROKERNEL_TEST_DEPS + [":packing"],
6831)
6832
6833xnnpack_unit_test(
Marat Dukhan4ed53f42020-08-06 01:12:55 -07006834 name = "qs8_gavgpool_minmax_test",
6835 srcs = [
6836 "test/qs8-gavgpool-minmax.cc",
6837 "test/gavgpool-microkernel-tester.h",
6838 "src/xnnpack/AlignedAllocator.h",
6839 ] + MICROKERNEL_TEST_HDRS,
6840 deps = MICROKERNEL_TEST_DEPS,
6841)
6842
6843xnnpack_unit_test(
Marat Dukhan595e1702020-07-31 10:12:52 -07006844 name = "qs8_gemm_minmax_test",
6845 srcs = [
6846 "test/qs8-gemm-minmax.cc",
6847 "test/gemm-microkernel-tester.h",
6848 "src/xnnpack/AlignedAllocator.h",
6849 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
6850 deps = MICROKERNEL_TEST_DEPS + [":packing"],
6851)
6852
6853xnnpack_unit_test(
Marat Dukhanf9480682020-07-31 14:50:24 -07006854 name = "qs8_igemm_minmax_test",
6855 srcs = [
6856 "test/qs8-igemm-minmax.cc",
6857 "test/gemm-microkernel-tester.h",
6858 "src/xnnpack/AlignedAllocator.h",
6859 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
6860 deps = MICROKERNEL_TEST_DEPS + [":packing"],
6861)
6862
6863xnnpack_unit_test(
6864 name = "qs8_requantization_test",
6865 srcs = [
6866 "src/xnnpack/requantization-stubs.h",
6867 "test/qs8-requantization.cc",
6868 "test/requantization-tester.h",
6869 ] + MICROKERNEL_TEST_HDRS,
6870 deps = MICROKERNEL_TEST_DEPS,
6871)
6872
6873xnnpack_unit_test(
Marat Dukhand9f3ad42020-08-10 12:30:58 -07006874 name = "qs8_vadd_minmax_test",
6875 srcs = [
6876 "test/qs8-vadd-minmax.cc",
6877 "test/vadd-microkernel-tester.h",
6878 ] + MICROKERNEL_TEST_HDRS,
6879 deps = MICROKERNEL_TEST_DEPS,
6880)
6881
6882xnnpack_unit_test(
Marat Dukhan0270d9f2020-08-11 00:56:46 -07006883 name = "qs8_vaddc_minmax_test",
6884 srcs = [
6885 "test/qs8-vaddc-minmax.cc",
6886 "test/vaddc-microkernel-tester.h",
6887 ] + MICROKERNEL_TEST_HDRS,
6888 deps = MICROKERNEL_TEST_DEPS,
6889)
6890
6891xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07006892 name = "qu8_avgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006893 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07006894 "test/qu8-avgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006895 "test/avgpool-microkernel-tester.h",
6896 "src/xnnpack/AlignedAllocator.h",
6897 ] + MICROKERNEL_TEST_HDRS,
6898 deps = MICROKERNEL_TEST_DEPS,
6899)
6900
6901xnnpack_unit_test(
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07006902 name = "qu8_dwconv_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006903 srcs = [
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07006904 "test/qu8-dwconv-minmax.cc",
6905 "test/dwconv-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006906 "src/xnnpack/AlignedAllocator.h",
6907 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07006908 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07006909)
6910
6911xnnpack_unit_test(
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07006912 name = "qu8_igemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006913 srcs = [
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07006914 "test/qu8-igemm-minmax.cc",
6915 "test/gemm-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006916 "src/xnnpack/AlignedAllocator.h",
6917 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07006918 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07006919)
6920
6921xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07006922 name = "qu8_gavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006923 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07006924 "test/qu8-gavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006925 "test/gavgpool-microkernel-tester.h",
6926 "src/xnnpack/AlignedAllocator.h",
6927 ] + MICROKERNEL_TEST_HDRS,
6928 deps = MICROKERNEL_TEST_DEPS,
6929)
6930
6931xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07006932 name = "qu8_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006933 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07006934 "test/qu8-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006935 "test/gemm-microkernel-tester.h",
6936 "src/xnnpack/AlignedAllocator.h",
6937 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07006938 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07006939)
6940
6941xnnpack_unit_test(
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07006942 name = "qu8_requantization_test",
6943 srcs = [
6944 "src/xnnpack/requantization-stubs.h",
6945 "test/qu8-requantization.cc",
6946 "test/requantization-tester.h",
6947 ] + MICROKERNEL_TEST_HDRS,
6948 deps = MICROKERNEL_TEST_DEPS,
6949)
6950
6951xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07006952 name = "qu8_vadd_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006953 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07006954 "test/qu8-vadd-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006955 "test/vadd-microkernel-tester.h",
6956 ] + MICROKERNEL_TEST_HDRS,
6957 deps = MICROKERNEL_TEST_DEPS,
6958)
6959
6960xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07006961 name = "u8_lut32norm_test",
6962 srcs = [
6963 "test/u8-lut32norm.cc",
6964 "test/lut-norm-microkernel-tester.h",
6965 ] + MICROKERNEL_TEST_HDRS,
6966 deps = MICROKERNEL_TEST_DEPS,
6967)
6968
6969xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07006970 name = "u8_maxpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006971 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07006972 "test/u8-maxpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006973 "test/maxpool-microkernel-tester.h",
6974 ] + MICROKERNEL_TEST_HDRS,
6975 deps = MICROKERNEL_TEST_DEPS,
6976)
6977
6978xnnpack_unit_test(
6979 name = "u8_rmax_test",
6980 srcs = [
6981 "test/u8-rmax.cc",
6982 "test/rmax-microkernel-tester.h",
6983 ] + MICROKERNEL_TEST_HDRS,
6984 deps = MICROKERNEL_TEST_DEPS,
6985)
6986
6987xnnpack_unit_test(
Marat Dukhan6674d692021-05-05 22:27:00 -07006988 name = "u8_vclamp_test",
6989 srcs = [
6990 "test/u8-vclamp.cc",
6991 "test/vclamp-microkernel-tester.h",
6992 ] + MICROKERNEL_TEST_HDRS,
6993 deps = MICROKERNEL_TEST_DEPS,
6994)
6995
6996xnnpack_unit_test(
Marat Dukhanad71b9a2020-11-20 00:01:51 -08006997 name = "x32_depthtospace2d_chw2hwc_test",
Yury Kartynnike7841862020-11-04 18:22:18 -08006998 srcs = [
Marat Dukhanad71b9a2020-11-20 00:01:51 -08006999 "test/x32-depthtospace2d-chw2hwc.cc",
7000 "test/depthtospace-microkernel-tester.h",
Yury Kartynnike7841862020-11-04 18:22:18 -08007001 ] + MICROKERNEL_TEST_HDRS,
7002 deps = MICROKERNEL_TEST_DEPS,
7003)
7004
7005xnnpack_unit_test(
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -07007006 name = "x32_fill_test",
7007 srcs = [
7008 "test/x32-fill.cc",
7009 "test/fill-microkernel-tester.h",
7010 ] + MICROKERNEL_TEST_HDRS,
7011 deps = MICROKERNEL_TEST_DEPS,
7012)
7013
7014xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07007015 name = "x32_packx_test",
7016 srcs = [
7017 "test/x32-packx.cc",
7018 "test/pack-microkernel-tester.h",
7019 "src/xnnpack/AlignedAllocator.h",
7020 ] + MICROKERNEL_TEST_HDRS,
7021 deps = MICROKERNEL_TEST_DEPS,
7022)
7023
7024xnnpack_unit_test(
7025 name = "x32_pad_test",
7026 srcs = [
7027 "test/x32-pad.cc",
7028 "test/pad-microkernel-tester.h",
7029 ] + MICROKERNEL_TEST_HDRS,
7030 deps = MICROKERNEL_TEST_DEPS,
7031)
7032
7033xnnpack_unit_test(
7034 name = "x32_unpool_test",
7035 srcs = [
7036 "test/x32-unpool.cc",
7037 "test/unpool-microkernel-tester.h",
7038 ] + MICROKERNEL_TEST_HDRS,
7039 deps = MICROKERNEL_TEST_DEPS,
7040)
7041
7042xnnpack_unit_test(
7043 name = "x32_zip_test",
7044 srcs = [
7045 "test/x32-zip.cc",
7046 "test/zip-microkernel-tester.h",
7047 ] + MICROKERNEL_TEST_HDRS,
7048 deps = MICROKERNEL_TEST_DEPS,
7049)
7050
7051xnnpack_unit_test(
7052 name = "x8_lut_test",
7053 srcs = [
7054 "test/x8-lut.cc",
7055 "test/lut-microkernel-tester.h",
7056 ] + MICROKERNEL_TEST_HDRS,
7057 deps = MICROKERNEL_TEST_DEPS,
7058)
7059
7060xnnpack_unit_test(
7061 name = "x8_zip_test",
7062 srcs = [
7063 "test/x8-zip.cc",
7064 "test/zip-microkernel-tester.h",
7065 ] + MICROKERNEL_TEST_HDRS,
7066 deps = MICROKERNEL_TEST_DEPS,
7067)
7068
Marat Dukhan20c3b922020-03-10 03:45:06 -07007069########################## Size tests for the library #########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07007070
7071xnnpack_binary(
Marat Dukhan20c3b922020-03-10 03:45:06 -07007072 name = "operator_size_test",
7073 srcs = ["test/operator-size.c"],
Marat Dukhanf0cb70a2021-03-30 15:45:15 -07007074 deps = [":xnnpack_for_tfjs"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07007075)
7076
Marat Dukhan20c3b922020-03-10 03:45:06 -07007077xnnpack_binary(
7078 name = "subgraph_size_test",
7079 srcs = ["test/subgraph-size.c"],
7080 deps = [":XNNPACK"],
7081)
7082
7083########################### Unit tests for operators ##########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07007084
7085xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07007086 name = "abs_nc_test",
7087 srcs = [
7088 "test/abs-nc.cc",
7089 "test/abs-operator-tester.h",
7090 ],
7091 deps = OPERATOR_TEST_DEPS,
7092)
7093
7094xnnpack_unit_test(
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08007095 name = "add_nd_test",
Artsiom Ablavatski2202c812021-01-22 14:16:43 -08007096 timeout = "moderate",
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08007097 srcs = [
7098 "test/add-nd.cc",
7099 "test/binary-elementwise-operator-tester.h",
7100 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07007101 deps = OPERATOR_TEST_DEPS,
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08007102)
7103
7104xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08007105 name = "argmax_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007106 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08007107 "test/argmax-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007108 "test/argmax-pooling-operator-tester.h",
7109 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07007110 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07007111)
7112
7113xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08007114 name = "average_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007115 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08007116 "test/average-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007117 "test/average-pooling-operator-tester.h",
7118 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07007119 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07007120)
7121
7122xnnpack_unit_test(
Marat Dukhan64e52512020-06-09 13:41:16 -07007123 name = "bankers_rounding_nc_test",
7124 srcs = [
7125 "test/bankers-rounding-nc.cc",
7126 "test/bankers-rounding-operator-tester.h",
7127 ],
7128 deps = OPERATOR_TEST_DEPS,
7129)
7130
7131xnnpack_unit_test(
7132 name = "ceiling_nc_test",
7133 srcs = [
7134 "test/ceiling-nc.cc",
7135 "test/ceiling-operator-tester.h",
7136 ],
7137 deps = OPERATOR_TEST_DEPS,
7138)
7139
7140xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08007141 name = "channel_shuffle_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007142 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08007143 "test/channel-shuffle-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007144 "test/channel-shuffle-operator-tester.h",
7145 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07007146 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07007147)
7148
7149xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08007150 name = "clamp_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007151 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08007152 "test/clamp-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007153 "test/clamp-operator-tester.h",
7154 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07007155 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07007156)
7157
7158xnnpack_unit_test(
Marat Dukhan065b11e2020-05-22 09:49:41 -07007159 name = "constant_pad_nd_test",
7160 srcs = [
7161 "test/constant-pad-nd.cc",
7162 "test/constant-pad-operator-tester.h",
7163 ],
7164 deps = OPERATOR_TEST_DEPS,
7165)
7166
7167xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08007168 name = "convolution_nhwc_test",
Artsiom Ablavatski2202c812021-01-22 14:16:43 -08007169 timeout = "moderate",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007170 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08007171 "test/convolution-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007172 "test/convolution-operator-tester.h",
7173 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07007174 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07007175)
7176
7177xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08007178 name = "convolution_nchw_test",
Artsiom Ablavatski2202c812021-01-22 14:16:43 -08007179 timeout = "moderate",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007180 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08007181 "test/convolution-nchw.cc",
7182 "test/convolution-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007183 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07007184 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07007185)
7186
7187xnnpack_unit_test(
Marat Dukhan4e21b272020-06-04 18:45:01 -07007188 name = "copy_nc_test",
7189 srcs = [
7190 "test/copy-nc.cc",
7191 "test/copy-operator-tester.h",
7192 ],
7193 deps = OPERATOR_TEST_DEPS,
7194)
7195
7196xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08007197 name = "deconvolution_nhwc_test",
Artsiom Ablavatskic1aa2972020-12-08 11:23:34 -08007198 timeout = "moderate",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007199 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08007200 "test/deconvolution-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007201 "test/deconvolution-operator-tester.h",
7202 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07007203 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07007204)
7205
7206xnnpack_unit_test(
Marat Dukhan188d1042020-11-24 23:39:40 -08007207 name = "depth_to_space_nchw2nhwc_test",
Artsiom Ablavatski0f1dc182020-11-05 19:21:50 -08007208 srcs = [
Marat Dukhan188d1042020-11-24 23:39:40 -08007209 "test/depth-to-space-nchw2nhwc.cc",
Artsiom Ablavatski0f1dc182020-11-05 19:21:50 -08007210 "test/depth-to-space-operator-tester.h",
7211 ] + OPERATOR_TEST_PARAMS_HDRS,
7212 deps = OPERATOR_TEST_DEPS,
7213)
7214
7215xnnpack_unit_test(
Marat Dukhan0e521172020-11-25 13:10:04 -08007216 name = "depth_to_space_nhwc_test",
7217 srcs = [
7218 "test/depth-to-space-nhwc.cc",
7219 "test/depth-to-space-operator-tester.h",
7220 ] + OPERATOR_TEST_PARAMS_HDRS,
7221 deps = OPERATOR_TEST_DEPS,
7222)
7223
7224xnnpack_unit_test(
Marat Dukhan69180502019-12-06 15:00:31 -08007225 name = "divide_nd_test",
7226 srcs = [
7227 "test/binary-elementwise-operator-tester.h",
7228 "test/divide-nd.cc",
7229 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07007230 deps = OPERATOR_TEST_DEPS,
Marat Dukhan69180502019-12-06 15:00:31 -08007231)
7232
7233xnnpack_unit_test(
Marat Dukhanb6bd4bc2020-12-01 17:01:40 -08007234 name = "elu_nc_test",
7235 srcs = [
7236 "test/elu-nc.cc",
7237 "test/elu-operator-tester.h",
7238 ],
7239 deps = OPERATOR_TEST_DEPS,
7240)
7241
7242xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08007243 name = "fully_connected_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007244 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08007245 "test/fully-connected-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007246 "test/fully-connected-operator-tester.h",
7247 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07007248 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07007249)
7250
7251xnnpack_unit_test(
Marat Dukhan64e52512020-06-09 13:41:16 -07007252 name = "floor_nc_test",
7253 srcs = [
7254 "test/floor-nc.cc",
7255 "test/floor-operator-tester.h",
7256 ],
7257 deps = OPERATOR_TEST_DEPS,
7258)
7259
7260xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08007261 name = "global_average_pooling_nwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007262 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08007263 "test/global-average-pooling-nwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007264 "test/global-average-pooling-operator-tester.h",
Marat Dukhanef61d022020-06-19 13:54:49 -07007265 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07007266 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07007267)
7268
7269xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08007270 name = "global_average_pooling_ncw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007271 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08007272 "test/global-average-pooling-ncw.cc",
7273 "test/global-average-pooling-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007274 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07007275 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07007276)
7277
7278xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08007279 name = "hardswish_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007280 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08007281 "test/hardswish-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007282 "test/hardswish-operator-tester.h",
7283 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07007284 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07007285)
7286
7287xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08007288 name = "leaky_relu_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007289 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08007290 "test/leaky-relu-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007291 "test/leaky-relu-operator-tester.h",
7292 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07007293 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07007294)
7295
7296xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08007297 name = "max_pooling_nhwc_test",
Artsiom Ablavatski2202c812021-01-22 14:16:43 -08007298 timeout = "moderate",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007299 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08007300 "test/max-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007301 "test/max-pooling-operator-tester.h",
7302 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07007303 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07007304)
7305
7306xnnpack_unit_test(
Marat Dukhan79e7f842019-12-05 14:35:50 -08007307 name = "maximum_nd_test",
7308 srcs = [
7309 "test/binary-elementwise-operator-tester.h",
7310 "test/maximum-nd.cc",
7311 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07007312 deps = OPERATOR_TEST_DEPS,
Marat Dukhan79e7f842019-12-05 14:35:50 -08007313)
7314
7315xnnpack_unit_test(
7316 name = "minimum_nd_test",
7317 srcs = [
7318 "test/binary-elementwise-operator-tester.h",
7319 "test/minimum-nd.cc",
7320 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07007321 deps = OPERATOR_TEST_DEPS,
Marat Dukhan79e7f842019-12-05 14:35:50 -08007322)
7323
7324xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08007325 name = "multiply_nd_test",
Marat Dukhanca2733c2019-11-15 23:21:17 -08007326 srcs = [
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08007327 "test/binary-elementwise-operator-tester.h",
Marat Dukhanefc47b82019-11-18 09:25:38 -08007328 "test/multiply-nd.cc",
Marat Dukhanca2733c2019-11-15 23:21:17 -08007329 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07007330 deps = OPERATOR_TEST_DEPS,
Marat Dukhanca2733c2019-11-15 23:21:17 -08007331)
7332
7333xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07007334 name = "negate_nc_test",
7335 srcs = [
7336 "test/negate-nc.cc",
7337 "test/negate-operator-tester.h",
7338 ],
7339 deps = OPERATOR_TEST_DEPS,
7340)
7341
7342xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08007343 name = "prelu_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007344 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08007345 "test/prelu-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007346 "test/prelu-operator-tester.h",
7347 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07007348 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07007349)
7350
7351xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08007352 name = "resize_bilinear_nhwc_test",
Marat Dukhan69722492019-11-11 19:55:50 -08007353 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08007354 "test/resize-bilinear-nhwc.cc",
Marat Dukhan69722492019-11-11 19:55:50 -08007355 "test/resize-bilinear-operator-tester.h",
7356 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07007357 deps = OPERATOR_TEST_DEPS,
Marat Dukhan69722492019-11-11 19:55:50 -08007358)
7359
7360xnnpack_unit_test(
Artsiom Ablavatski97918102020-10-27 15:52:59 -07007361 name = "resize_bilinear_nchw_test",
7362 srcs = [
7363 "test/resize-bilinear-nchw.cc",
7364 "test/resize-bilinear-operator-tester.h",
7365 ] + OPERATOR_TEST_PARAMS_HDRS,
7366 deps = OPERATOR_TEST_DEPS,
7367)
7368
7369xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08007370 name = "sigmoid_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007371 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08007372 "test/sigmoid-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007373 "test/sigmoid-operator-tester.h",
7374 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07007375 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07007376)
7377
7378xnnpack_unit_test(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08007379 name = "softmax_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007380 srcs = [
Marat Dukhanfd8e6892020-01-27 15:25:25 -08007381 "test/softmax-nc.cc",
7382 "test/softmax-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007383 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07007384 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07007385)
7386
7387xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07007388 name = "square_nc_test",
7389 srcs = [
7390 "test/square-nc.cc",
7391 "test/square-operator-tester.h",
7392 ],
7393 deps = OPERATOR_TEST_DEPS,
7394)
7395
7396xnnpack_unit_test(
Marat Dukhan6804bbd2020-06-30 19:26:11 -07007397 name = "square_root_nc_test",
7398 srcs = [
7399 "test/square-root-nc.cc",
7400 "test/square-root-operator-tester.h",
7401 ],
7402 deps = OPERATOR_TEST_DEPS,
7403)
7404
7405xnnpack_unit_test(
Marat Dukhanf7399262020-06-05 10:58:44 -07007406 name = "squared_difference_nd_test",
7407 srcs = [
7408 "test/binary-elementwise-operator-tester.h",
7409 "test/squared-difference-nd.cc",
7410 ],
7411 deps = OPERATOR_TEST_DEPS,
7412)
7413
7414xnnpack_unit_test(
Marat Dukhan05f3f6d2019-12-03 15:13:53 -08007415 name = "subtract_nd_test",
7416 srcs = [
7417 "test/binary-elementwise-operator-tester.h",
7418 "test/subtract-nd.cc",
7419 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07007420 deps = OPERATOR_TEST_DEPS,
Marat Dukhan05f3f6d2019-12-03 15:13:53 -08007421)
7422
7423xnnpack_unit_test(
Marat Dukhan64e52512020-06-09 13:41:16 -07007424 name = "truncation_nc_test",
7425 srcs = [
7426 "test/truncation-nc.cc",
7427 "test/truncation-operator-tester.h",
7428 ],
7429 deps = OPERATOR_TEST_DEPS,
7430)
7431
7432xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08007433 name = "unpooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007434 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08007435 "test/unpooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007436 "test/unpooling-operator-tester.h",
7437 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07007438 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07007439)
7440
Chao Mei6ddfc602020-05-13 22:29:36 -07007441############################### Misc unit tests ###############################
7442
7443xnnpack_unit_test(
7444 name = "memory_planner_test",
7445 srcs = [
7446 "test/memory-planner-test.cc",
7447 ],
7448 deps = [
7449 ":XNNPACK",
7450 ":memory_planner",
7451 ],
7452)
7453
XNNPACK Teamab8c4c82020-10-09 08:05:51 -07007454xnnpack_unit_test(
7455 name = "subgraph_nchw_test",
7456 srcs = [
7457 "src/xnnpack/subgraph.h",
7458 "test/subgraph-nchw.cc",
7459 "test/subgraph-tester.h",
7460 ],
7461 deps = [
7462 ":XNNPACK",
7463 ],
7464)
7465
Marat Dukhan08c4a432019-10-03 09:29:21 -07007466############################# Build configurations #############################
7467
Marat Dukhanb8642352019-10-30 15:43:02 -07007468# Enables usage of assembly kernels.
Marat Dukhan08c4a432019-10-03 09:29:21 -07007469config_setting(
Marat Dukhanb8642352019-10-30 15:43:02 -07007470 name = "xnn_enable_assembly_explicit_true",
7471 define_values = {"xnn_enable_assembly": "true"},
7472)
7473
7474# Disables usage of assembly kernels.
7475config_setting(
7476 name = "xnn_enable_assembly_explicit_false",
7477 define_values = {"xnn_enable_assembly": "false"},
7478)
7479
Marat Dukhan9de90e02020-06-18 16:04:12 -07007480# Enables usage of sparse inference.
7481config_setting(
7482 name = "xnn_enable_sparse_explicit_true",
7483 define_values = {"xnn_enable_sparse": "true"},
7484)
7485
7486# Disables usage of sparse inference.
7487config_setting(
7488 name = "xnn_enable_sparse_explicit_false",
7489 define_values = {"xnn_enable_sparse": "false"},
7490)
7491
Marat Dukhan05702cf2020-03-26 15:41:33 -07007492# Disables usage of HMP-aware optimizations.
7493config_setting(
7494 name = "xnn_enable_hmp_explicit_false",
7495 define_values = {"xnn_enable_hmp": "false"},
7496)
7497
Chao Mei6ddfc602020-05-13 22:29:36 -07007498# Enable usage of optimized memory allocation
7499config_setting(
7500 name = "xnn_enable_memopt_explicit_true",
Marat Dukhan03f46212021-03-30 21:29:49 -07007501 define_values = {"xnn_enable_memopt": "true"},
Chao Mei6ddfc602020-05-13 22:29:36 -07007502)
7503
7504# Disable usage of optimized memory allocation
7505config_setting(
7506 name = "xnn_enable_memopt_explicit_false",
Marat Dukhan03f46212021-03-30 21:29:49 -07007507 define_values = {"xnn_enable_memopt": "false"},
Chao Mei6ddfc602020-05-13 22:29:36 -07007508)
7509
Marat Dukhanb939cdb2021-03-30 18:51:51 -07007510# Enable QS8 inference in TFLite-specific version
7511config_setting(
7512 name = "xnn_enable_qs8_explicit_true",
7513 define_values = {"xnn_enable_qs8": "true"},
7514)
7515
7516# Disable QS8 inference in TFLite-specific version
7517config_setting(
7518 name = "xnn_enable_qs8_explicit_false",
7519 define_values = {"xnn_enable_qs8": "false"},
7520)
7521
Marat Dukhanb8642352019-10-30 15:43:02 -07007522# Builds with -c dbg
7523config_setting(
7524 name = "debug_build",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007525 values = {
Marat Dukhanb8642352019-10-30 15:43:02 -07007526 "compilation_mode": "dbg",
7527 },
7528)
7529
7530# Builds with -c opt
7531config_setting(
7532 name = "optimized_build",
7533 values = {
7534 "compilation_mode": "opt",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007535 },
7536)
7537
7538config_setting(
Marat Dukhanb8642352019-10-30 15:43:02 -07007539 name = "linux_k8",
7540 values = {"cpu": "k8"},
7541)
7542
7543config_setting(
Marat Dukhan582094e2020-04-30 17:21:25 -07007544 name = "linux_arm",
7545 values = {"cpu": "arm"},
Marat Dukhan4e45e662019-10-03 15:40:24 -07007546)
7547
7548config_setting(
Marat Dukhanf0bd4de2020-06-15 15:53:19 -07007549 name = "linux_armeabi",
7550 values = {"cpu": "armeabi"},
7551)
7552
7553config_setting(
Terry Heo68eef3f2020-04-13 22:53:52 -07007554 name = "linux_armhf",
7555 values = {"cpu": "armhf"},
7556)
7557
7558config_setting(
Marat Dukhana720e932020-06-10 13:01:11 -07007559 name = "linux_armv7a",
7560 values = {"cpu": "armv7a"},
7561)
7562
7563config_setting(
Marat Dukhan582094e2020-04-30 17:21:25 -07007564 name = "linux_aarch64",
7565 values = {"cpu": "aarch64"},
7566)
7567
7568config_setting(
Marat Dukhan08c4a432019-10-03 09:29:21 -07007569 name = "android",
7570 values = {"crosstool_top": "//external:android/crosstool"},
7571)
7572
7573config_setting(
7574 name = "android_armv7",
7575 values = {
7576 "crosstool_top": "//external:android/crosstool",
7577 "cpu": "armeabi-v7a",
7578 },
7579)
7580
7581config_setting(
7582 name = "android_arm64",
7583 values = {
7584 "crosstool_top": "//external:android/crosstool",
7585 "cpu": "arm64-v8a",
7586 },
7587)
7588
7589config_setting(
7590 name = "android_x86",
7591 values = {
7592 "crosstool_top": "//external:android/crosstool",
7593 "cpu": "x86",
7594 },
7595)
7596
7597config_setting(
7598 name = "android_x86_64",
7599 values = {
7600 "crosstool_top": "//external:android/crosstool",
7601 "cpu": "x86_64",
7602 },
7603)
7604
7605config_setting(
Marat Dukhan10a38082020-04-17 03:58:35 -07007606 name = "windows_x86_64",
7607 values = {"cpu": "x64_windows"},
Marat Dukhan9fe932e2020-04-11 17:14:15 -07007608)
7609
7610config_setting(
Marat Dukhan10a38082020-04-17 03:58:35 -07007611 name = "windows_x86_64_clang",
7612 values = {
7613 "compiler": "clang-cl",
7614 "cpu": "x64_windows",
7615 },
7616)
7617
7618config_setting(
7619 name = "windows_x86_64_mingw",
7620 values = {
7621 "compiler": "mingw-gcc",
7622 "cpu": "x64_windows",
7623 },
7624)
7625
7626config_setting(
7627 name = "windows_x86_64_msys",
7628 values = {
7629 "compiler": "msys-gcc",
7630 "cpu": "x64_windows",
7631 },
Marat Dukhan9fe932e2020-04-11 17:14:15 -07007632)
7633
7634config_setting(
Marat Dukhan885ca242019-10-07 09:17:32 -07007635 name = "macos_x86_64",
7636 values = {
7637 "apple_platform_type": "macos",
7638 "cpu": "darwin",
7639 },
7640)
7641
7642config_setting(
Simon Maurerae33ab82021-03-03 23:38:22 +01007643 name = "macos_arm64",
7644 values = {
7645 "apple_platform_type": "macos",
7646 "cpu": "darwin_arm64",
7647 },
7648)
7649
7650config_setting(
Marat Dukhan08c4a432019-10-03 09:29:21 -07007651 name = "emscripten",
XNNPACK Team3bfbdaf2021-03-29 15:26:23 -07007652 values = {"crosstool_top": "@emsdk//emscripten_toolchain:everything"},
Marat Dukhan08c4a432019-10-03 09:29:21 -07007653)
7654
7655config_setting(
7656 name = "emscripten_wasm",
7657 values = {
XNNPACK Team3bfbdaf2021-03-29 15:26:23 -07007658 "crosstool_top": "@emsdk//emscripten_toolchain:everything",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007659 "cpu": "wasm",
7660 },
7661)
7662
7663config_setting(
7664 name = "emscripten_wasmsimd",
7665 values = {
XNNPACK Team3bfbdaf2021-03-29 15:26:23 -07007666 "crosstool_top": "@emsdk//emscripten_toolchain:everything",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007667 "cpu": "wasm",
Marat Dukhan81c62602020-05-29 13:22:49 -07007668 "copt": "-msimd128",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007669 },
7670)
7671
7672config_setting(
Marat Dukhan1498d1d2020-02-11 20:00:05 -08007673 name = "ios_armv7",
7674 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08007675 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08007676 "cpu": "ios_armv7",
7677 },
7678)
7679
7680config_setting(
7681 name = "ios_arm64",
7682 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08007683 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08007684 "cpu": "ios_arm64",
7685 },
7686)
7687
7688config_setting(
7689 name = "ios_arm64e",
7690 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08007691 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08007692 "cpu": "ios_arm64e",
7693 },
7694)
7695
7696config_setting(
7697 name = "ios_x86",
7698 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08007699 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08007700 "cpu": "ios_i386",
7701 },
7702)
7703
7704config_setting(
7705 name = "ios_x86_64",
7706 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08007707 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08007708 "cpu": "ios_x86_64",
7709 },
7710)
7711
7712config_setting(
7713 name = "watchos_armv7k",
7714 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08007715 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08007716 "cpu": "watchos_armv7k",
7717 },
7718)
7719
7720config_setting(
7721 name = "watchos_arm64_32",
7722 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08007723 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08007724 "cpu": "watchos_arm64_32",
7725 },
7726)
7727
7728config_setting(
7729 name = "watchos_x86",
7730 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08007731 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08007732 "cpu": "watchos_i386",
7733 },
7734)
7735
7736config_setting(
7737 name = "watchos_x86_64",
7738 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08007739 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08007740 "cpu": "watchos_x86_64",
7741 },
7742)
7743
7744config_setting(
7745 name = "tvos_arm64",
7746 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08007747 "apple_platform_type": "tvos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08007748 "cpu": "tvos_arm64",
7749 },
7750)
7751
7752config_setting(
7753 name = "tvos_x86_64",
7754 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08007755 "apple_platform_type": "tvos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08007756 "cpu": "tvos_x86_64",
7757 },
7758)