blob: c2572cd2f9396cff61d22bd92cfdc483588ff645 [file] [log] [blame]
Frank Barchard1a953052020-11-16 18:44:58 -08001# Copyright 2020 Google LLC
Marat Dukhan08c4a432019-10-03 09:29:21 -07002#
3# This source code is licensed under the BSD-style license found in the
4# LICENSE file in the root directory of this source tree.
5#
6# Description:
7# XNNPACK - optimized floating-point neural network operators library
8
Marat Dukhan10a38082020-04-17 03:58:35 -07009load(":build_defs.bzl", "xnnpack_aggregate_library", "xnnpack_benchmark", "xnnpack_binary", "xnnpack_cc_library", "xnnpack_gcc_std_copts", "xnnpack_min_size_copts", "xnnpack_msvc_std_copts", "xnnpack_optional_armcl_copts", "xnnpack_optional_armcl_deps", "xnnpack_optional_dnnl_copts", "xnnpack_optional_dnnl_deps", "xnnpack_optional_gemmlowp_copts", "xnnpack_optional_gemmlowp_deps", "xnnpack_optional_ruy_copts", "xnnpack_optional_ruy_deps", "xnnpack_optional_tflite_copts", "xnnpack_optional_tflite_deps", "xnnpack_std_cxxopts", "xnnpack_unit_test", "xnnpack_visibility")
Marat Dukhan69c3f2c2019-11-06 12:30:01 -080010
Marat Dukhan08c4a432019-10-03 09:29:21 -070011licenses(["notice"])
12
13exports_files(["LICENSE"])
14
Marat Dukhan1b354632020-03-23 12:50:22 -070015OPERATOR_BENCHMARK_DEPS = [
16 ":XNNPACK",
17 ":bench_utils",
18 "@cpuinfo",
Frank Barchard0c849732020-06-12 13:31:32 -070019 "@FP16",
Marat Dukhan1b354632020-03-23 12:50:22 -070020 "@pthreadpool",
21]
22
Marat Dukhan08c4a432019-10-03 09:29:21 -070023MICROKERNEL_BENCHMARK_DEPS = [
24 ":ukernels",
25 ":bench_utils",
Frank Barchard7e955972019-10-11 10:34:25 -070026 ":enable_assembly",
Marat Dukhan08c4a432019-10-03 09:29:21 -070027 "@cpuinfo",
28 "@FP16",
29 "@pthreadpool",
30]
31
Marat Dukhan6adff4e2019-10-14 18:32:07 -070032ACCURACY_EVAL_DEPS = [
33 ":XNNPACK",
34 ":ukernels",
35 "@FP16",
36 "@pthreadpool",
37]
38
Marat Dukhan08c4a432019-10-03 09:29:21 -070039MICROKERNEL_TEST_DEPS = [
Marat Dukhan33fcf782020-05-24 14:27:15 -070040 ":ukernels_test_mode",
Frank Barchard7e955972019-10-11 10:34:25 -070041 ":enable_assembly",
Marat Dukhan08c4a432019-10-03 09:29:21 -070042 "@cpuinfo",
43 "@FP16",
44 "@pthreadpool",
45]
46
Marat Dukhan1b354632020-03-23 12:50:22 -070047OPERATOR_TEST_DEPS = [
Marat Dukhan33fcf782020-05-24 14:27:15 -070048 ":XNNPACK_test_mode",
Marat Dukhan1b354632020-03-23 12:50:22 -070049 "@pthreadpool",
50 "@FP16",
51]
52
Marat Dukhan08c4a432019-10-03 09:29:21 -070053OPERATOR_SRCS = [
Marat Dukhane8265432020-04-28 18:42:59 -070054 "src/operators/argmax-pooling-nhwc.c",
55 "src/operators/average-pooling-nhwc.c",
56 "src/operators/binary-elementwise-nd.c",
Marat Dukhane8265432020-04-28 18:42:59 -070057 "src/operators/channel-shuffle-nc.c",
Marat Dukhan065b11e2020-05-22 09:49:41 -070058 "src/operators/constant-pad-nd.c",
Marat Dukhane8265432020-04-28 18:42:59 -070059 "src/operators/convolution-nchw.c",
60 "src/operators/convolution-nhwc.c",
61 "src/operators/deconvolution-nhwc.c",
Marat Dukhan13b68f22020-11-12 11:55:19 -080062 "src/operators/depth-to-space-nchw2nhwc.c",
Marat Dukhan0e521172020-11-25 13:10:04 -080063 "src/operators/depth-to-space-nhwc.c",
Marat Dukhane8265432020-04-28 18:42:59 -070064 "src/operators/fully-connected-nc.c",
65 "src/operators/global-average-pooling-ncw.c",
66 "src/operators/global-average-pooling-nwc.c",
Marat Dukhane8265432020-04-28 18:42:59 -070067 "src/operators/leaky-relu-nc.c",
68 "src/operators/max-pooling-nhwc.c",
69 "src/operators/prelu-nc.c",
Artsiom Ablavatski97918102020-10-27 15:52:59 -070070 "src/operators/resize-bilinear-nchw.c",
Marat Dukhane8265432020-04-28 18:42:59 -070071 "src/operators/resize-bilinear-nhwc.c",
72 "src/operators/sigmoid-nc.c",
73 "src/operators/softmax-nc.c",
Marat Dukhanc3065f52020-06-04 13:33:32 -070074 "src/operators/unary-elementwise-nc.c",
Marat Dukhane8265432020-04-28 18:42:59 -070075 "src/operators/unpooling-nhwc.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -070076]
77
Marat Dukhan95e8b7a2020-06-03 12:46:26 -070078SUBGRAPH_SRCS = [
Marat Dukhan5fab4092020-06-10 01:28:28 -070079 "src/subgraph/abs.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -070080 "src/subgraph/add2.c",
81 "src/subgraph/argmax-pooling-2d.c",
82 "src/subgraph/average-pooling-2d.c",
Marat Dukhan5fab4092020-06-10 01:28:28 -070083 "src/subgraph/bankers-rounding.c",
84 "src/subgraph/ceiling.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -070085 "src/subgraph/clamp.c",
86 "src/subgraph/convolution-2d.c",
87 "src/subgraph/deconvolution-2d.c",
Artsiom Ablavatskibbe85062020-11-05 14:07:37 -080088 "src/subgraph/depth-to-space.c",
Frank Barchard9cef5ea2020-11-18 14:52:08 -080089 "src/subgraph/depthwise-convolution-2d.c",
Marat Dukhan9d3a4592020-06-05 16:52:42 -070090 "src/subgraph/divide.c",
Marat Dukhana1600202020-12-01 22:17:16 -080091 "src/subgraph/elu.c",
Marat Dukhan5fab4092020-06-10 01:28:28 -070092 "src/subgraph/floor.c",
Frank Barchard04336c12020-10-22 16:48:55 -070093 "src/subgraph/fully-connected.c",
Marat Dukhana059b7d2020-06-11 11:41:27 -070094 "src/subgraph/global-average-pooling-2d.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -070095 "src/subgraph/hardswish.c",
Marat Dukhan5bbebac2020-06-10 19:42:15 -070096 "src/subgraph/leaky-relu.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -070097 "src/subgraph/max-pooling-2d.c",
Marat Dukhan9d3a4592020-06-05 16:52:42 -070098 "src/subgraph/maximum2.c",
99 "src/subgraph/minimum2.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -0700100 "src/subgraph/multiply2.c",
Marat Dukhan5fab4092020-06-10 01:28:28 -0700101 "src/subgraph/negate.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -0700102 "src/subgraph/prelu.c",
103 "src/subgraph/sigmoid.c",
104 "src/subgraph/softmax.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700105 "src/subgraph/square-root.c",
106 "src/subgraph/square.c",
107 "src/subgraph/squared-difference.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -0700108 "src/subgraph/static-constant-pad.c",
Marat Dukhand27202d2020-07-09 23:43:40 -0700109 "src/subgraph/static-reshape.c",
Marat Dukhanaff24e22020-07-23 01:43:58 -0700110 "src/subgraph/static-resize-bilinear-2d.c",
Marat Dukhan9d3a4592020-06-05 16:52:42 -0700111 "src/subgraph/subtract.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -0700112 "src/subgraph/unpooling-2d.c",
113]
114
Marat Dukhan3a77ea72019-12-23 12:10:24 -0800115TABLE_SRCS = [
116 "src/tables/exp2-k-over-64.c",
117 "src/tables/exp2-k-over-2048.c",
Marat Dukhande390d42020-11-29 19:32:18 -0800118 "src/tables/exp2minus-k-over-4.c",
119 "src/tables/exp2minus-k-over-8.c",
Marat Dukhanc60742b2020-11-23 12:33:27 -0800120 "src/tables/exp2minus-k-over-16.c",
Marat Dukhan1f256fc2020-09-24 21:27:14 -0700121 "src/tables/exp2minus-k-over-64.c",
122 "src/tables/exp2minus-k-over-2048.c",
Marat Dukhan3a77ea72019-12-23 12:10:24 -0800123]
124
Marat Dukhan08c4a432019-10-03 09:29:21 -0700125SCALAR_UKERNELS = [
Marat Dukhan329da642019-11-19 21:44:39 -0800126 "src/f32-argmaxpool/4x-scalar-c1.c",
Marat Dukhan1e782c42019-11-21 17:02:40 -0800127 "src/f32-argmaxpool/9p8x-scalar-c1.c",
Marat Dukhan329da642019-11-19 21:44:39 -0800128 "src/f32-argmaxpool/9x-scalar-c1.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700129 "src/f32-avgpool/9p8x-minmax-scalar-c1.c",
130 "src/f32-avgpool/9x-minmax-scalar-c1.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700131 "src/f32-conv-hwc/3x3s2p0p1c3x4-scalar-1x1.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700132 "src/f32-conv-hwc/3x3s2p1c3x4-scalar-1x1.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -0700133 "src/f32-conv-hwc2chw/3x3s2p1c3x4-scalar-1x1.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700134 "src/f32-dwconv/gen/up1x4-minmax-scalar-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700135 "src/f32-dwconv/gen/up1x4-minmax-scalar.c",
136 "src/f32-dwconv/gen/up1x4-scalar-acc2.c",
137 "src/f32-dwconv/gen/up1x4-scalar.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700138 "src/f32-dwconv/gen/up1x9-minmax-scalar-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700139 "src/f32-dwconv/gen/up1x9-minmax-scalar.c",
140 "src/f32-dwconv/gen/up1x9-scalar-acc2.c",
141 "src/f32-dwconv/gen/up1x9-scalar.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700142 "src/f32-dwconv/gen/up1x25-minmax-scalar-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700143 "src/f32-dwconv/gen/up1x25-minmax-scalar.c",
144 "src/f32-dwconv/gen/up1x25-scalar-acc2.c",
145 "src/f32-dwconv/gen/up1x25-scalar.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700146 "src/f32-dwconv/gen/up2x4-minmax-scalar-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700147 "src/f32-dwconv/gen/up2x4-minmax-scalar.c",
148 "src/f32-dwconv/gen/up2x4-scalar-acc2.c",
149 "src/f32-dwconv/gen/up2x4-scalar.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700150 "src/f32-dwconv/gen/up2x9-minmax-scalar-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700151 "src/f32-dwconv/gen/up2x9-minmax-scalar.c",
152 "src/f32-dwconv/gen/up2x9-scalar-acc2.c",
153 "src/f32-dwconv/gen/up2x9-scalar.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700154 "src/f32-dwconv/gen/up2x25-minmax-scalar-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700155 "src/f32-dwconv/gen/up2x25-minmax-scalar.c",
156 "src/f32-dwconv/gen/up2x25-scalar-acc2.c",
157 "src/f32-dwconv/gen/up2x25-scalar.c",
Frank Barchard35db7d02020-10-26 13:37:34 -0700158 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-scalar-1x1-acc2.c",
159 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-scalar-1x1-acc3.c",
160 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-scalar-1x1-acc4.c",
Marat Dukhan91249d22020-10-24 12:02:51 -0700161 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-scalar-1x1.c",
Frank Barchard35db7d02020-10-26 13:37:34 -0700162 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-scalar-2x1-acc2.c",
Marat Dukhan91249d22020-10-24 12:02:51 -0700163 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-scalar-2x1.c",
164 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-scalar-3x1.c",
165 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-scalar-4x1.c",
166 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-scalar-5x1.c",
167 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-scalar-6x1.c",
Marat Dukhancf5b3c32020-10-25 19:21:10 -0700168 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-scalar-1x1-acc2.c",
169 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-scalar-1x1-acc3.c",
170 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-scalar-1x1-acc4.c",
Frank Barchard35db7d02020-10-26 13:37:34 -0700171 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-scalar-1x1.c",
Marat Dukhancf5b3c32020-10-25 19:21:10 -0700172 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-scalar-2x1-acc2.c",
Frank Barchard35db7d02020-10-26 13:37:34 -0700173 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-scalar-2x1.c",
174 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-scalar-3x1.c",
175 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-scalar-4x1.c",
Marat Dukhanc4efb002020-10-25 23:14:47 -0700176 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-scalar-1x1-acc2.c",
177 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-scalar-1x1-acc3.c",
178 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-scalar-1x1-acc4.c",
179 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-scalar-1x1-acc5.c",
Frank Barchard35db7d02020-10-26 13:37:34 -0700180 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-scalar-1x1.c",
Marat Dukhanc4efb002020-10-25 23:14:47 -0700181 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-scalar-2x1-acc2.c",
182 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-scalar-2x1-acc3.c",
Frank Barchard35db7d02020-10-26 13:37:34 -0700183 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-scalar-2x1.c",
Marat Dukhanc4efb002020-10-25 23:14:47 -0700184 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-scalar-3x1-acc2.c",
Frank Barchard35db7d02020-10-26 13:37:34 -0700185 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-scalar-3x1.c",
Marat Dukhan29c0c332020-10-28 22:11:00 -0700186 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-scalar-1x1-acc2.c",
187 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-scalar-1x1-acc3.c",
188 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-scalar-1x1-acc4.c",
189 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-scalar-1x1-acc5.c",
190 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-scalar-1x1.c",
191 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-scalar-2x1-acc2.c",
192 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-scalar-2x1-acc3.c",
193 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-scalar-2x1.c",
194 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-scalar-3x1-acc2.c",
195 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-scalar-3x1.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -0700196 "src/f32-gavgpool-cw/scalar-x1.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700197 "src/f32-gavgpool/7p7x-minmax-scalar-c1.c",
198 "src/f32-gavgpool/7x-minmax-scalar-c1.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700199 "src/f32-gemm/gen-inc/1x4inc-minmax-scalar.c",
200 "src/f32-gemm/gen-inc/2x4inc-minmax-scalar.c",
201 "src/f32-gemm/gen-inc/4x4inc-minmax-scalar.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700202 "src/f32-gemm/gen/1x4-minmax-scalar.c",
203 "src/f32-gemm/gen/1x4-relu-scalar.c",
204 "src/f32-gemm/gen/1x4-scalar.c",
205 "src/f32-gemm/gen/2x4-minmax-scalar.c",
206 "src/f32-gemm/gen/2x4-relu-scalar.c",
207 "src/f32-gemm/gen/2x4-scalar.c",
208 "src/f32-gemm/gen/4x2-minmax-scalar.c",
209 "src/f32-gemm/gen/4x2-relu-scalar.c",
210 "src/f32-gemm/gen/4x2-scalar.c",
211 "src/f32-gemm/gen/4x4-minmax-scalar.c",
212 "src/f32-gemm/gen/4x4-relu-scalar.c",
213 "src/f32-gemm/gen/4x4-scalar.c",
XNNPACK Team21432672020-10-19 19:58:48 -0700214 "src/f32-ibilinear-chw/gen/scalar-p1.c",
215 "src/f32-ibilinear-chw/gen/scalar-p2.c",
216 "src/f32-ibilinear-chw/gen/scalar-p4.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700217 "src/f32-ibilinear/gen/scalar-c1.c",
218 "src/f32-ibilinear/gen/scalar-c2.c",
219 "src/f32-ibilinear/gen/scalar-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700220 "src/f32-igemm/gen/1x4-minmax-scalar.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700221 "src/f32-igemm/gen/1x4-relu-scalar.c",
222 "src/f32-igemm/gen/1x4-scalar.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700223 "src/f32-igemm/gen/2x4-minmax-scalar.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700224 "src/f32-igemm/gen/2x4-relu-scalar.c",
225 "src/f32-igemm/gen/2x4-scalar.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700226 "src/f32-igemm/gen/4x2-minmax-scalar.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700227 "src/f32-igemm/gen/4x2-relu-scalar.c",
228 "src/f32-igemm/gen/4x2-scalar.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700229 "src/f32-igemm/gen/4x4-minmax-scalar.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700230 "src/f32-igemm/gen/4x4-relu-scalar.c",
231 "src/f32-igemm/gen/4x4-scalar.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700232 "src/f32-maxpool/9p8x-minmax-scalar-c1.c",
233 "src/f32-pavgpool/9p8x-minmax-scalar-c1.c",
234 "src/f32-pavgpool/9x-minmax-scalar-c1.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700235 "src/f32-ppmm/gen/2x4-minmax-scalar.c",
236 "src/f32-ppmm/gen/3x3-minmax-scalar.c",
237 "src/f32-ppmm/gen/4x2-minmax-scalar.c",
238 "src/f32-ppmm/gen/4x4-minmax-scalar.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800239 "src/f32-prelu/gen/scalar-2x1.c",
240 "src/f32-prelu/gen/scalar-2x4.c",
Marat Dukhanf46f6752020-01-21 11:03:49 -0800241 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x1.c",
Marat Dukhanf46f6752020-01-21 11:03:49 -0800242 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x2-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700243 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x2.c",
Marat Dukhanf46f6752020-01-21 11:03:49 -0800244 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x4-acc2.c",
245 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x4-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700246 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x4.c",
Marat Dukhanf46f6752020-01-21 11:03:49 -0800247 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x1.c",
Marat Dukhanf46f6752020-01-21 11:03:49 -0800248 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x2-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700249 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x2.c",
Marat Dukhanf46f6752020-01-21 11:03:49 -0800250 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x4-acc2.c",
251 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x4-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700252 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x4.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700253 "src/f32-rmax/scalar.c",
Marat Dukhan355ab432020-04-09 19:01:52 -0700254 "src/f32-spmm/gen/1x1-minmax-scalar-pipelined.c",
255 "src/f32-spmm/gen/1x1-minmax-scalar.c",
256 "src/f32-spmm/gen/2x1-minmax-scalar-pipelined.c",
257 "src/f32-spmm/gen/2x1-minmax-scalar.c",
258 "src/f32-spmm/gen/4x1-minmax-scalar-pipelined.c",
259 "src/f32-spmm/gen/4x1-minmax-scalar.c",
260 "src/f32-spmm/gen/8x1-minmax-scalar-pipelined.c",
261 "src/f32-spmm/gen/8x1-minmax-scalar.c",
262 "src/f32-spmm/gen/8x2-minmax-scalar.c",
263 "src/f32-spmm/gen/8x4-minmax-scalar.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700264 "src/f32-vbinary/gen/vadd-minmax-scalar-x1.c",
265 "src/f32-vbinary/gen/vadd-minmax-scalar-x2.c",
266 "src/f32-vbinary/gen/vadd-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700267 "src/f32-vbinary/gen/vadd-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700268 "src/f32-vbinary/gen/vadd-relu-scalar-x1.c",
269 "src/f32-vbinary/gen/vadd-relu-scalar-x2.c",
270 "src/f32-vbinary/gen/vadd-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700271 "src/f32-vbinary/gen/vadd-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700272 "src/f32-vbinary/gen/vadd-scalar-x1.c",
273 "src/f32-vbinary/gen/vadd-scalar-x2.c",
274 "src/f32-vbinary/gen/vadd-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700275 "src/f32-vbinary/gen/vadd-scalar-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700276 "src/f32-vbinary/gen/vaddc-minmax-scalar-x1.c",
277 "src/f32-vbinary/gen/vaddc-minmax-scalar-x2.c",
278 "src/f32-vbinary/gen/vaddc-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700279 "src/f32-vbinary/gen/vaddc-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700280 "src/f32-vbinary/gen/vaddc-relu-scalar-x1.c",
281 "src/f32-vbinary/gen/vaddc-relu-scalar-x2.c",
282 "src/f32-vbinary/gen/vaddc-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700283 "src/f32-vbinary/gen/vaddc-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700284 "src/f32-vbinary/gen/vaddc-scalar-x1.c",
285 "src/f32-vbinary/gen/vaddc-scalar-x2.c",
286 "src/f32-vbinary/gen/vaddc-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700287 "src/f32-vbinary/gen/vaddc-scalar-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700288 "src/f32-vbinary/gen/vdiv-minmax-scalar-x1.c",
289 "src/f32-vbinary/gen/vdiv-minmax-scalar-x2.c",
290 "src/f32-vbinary/gen/vdiv-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700291 "src/f32-vbinary/gen/vdiv-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700292 "src/f32-vbinary/gen/vdiv-relu-scalar-x1.c",
293 "src/f32-vbinary/gen/vdiv-relu-scalar-x2.c",
294 "src/f32-vbinary/gen/vdiv-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700295 "src/f32-vbinary/gen/vdiv-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700296 "src/f32-vbinary/gen/vdiv-scalar-x1.c",
297 "src/f32-vbinary/gen/vdiv-scalar-x2.c",
298 "src/f32-vbinary/gen/vdiv-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700299 "src/f32-vbinary/gen/vdiv-scalar-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700300 "src/f32-vbinary/gen/vdivc-minmax-scalar-x1.c",
301 "src/f32-vbinary/gen/vdivc-minmax-scalar-x2.c",
302 "src/f32-vbinary/gen/vdivc-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700303 "src/f32-vbinary/gen/vdivc-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700304 "src/f32-vbinary/gen/vdivc-relu-scalar-x1.c",
305 "src/f32-vbinary/gen/vdivc-relu-scalar-x2.c",
306 "src/f32-vbinary/gen/vdivc-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700307 "src/f32-vbinary/gen/vdivc-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700308 "src/f32-vbinary/gen/vdivc-scalar-x1.c",
309 "src/f32-vbinary/gen/vdivc-scalar-x2.c",
310 "src/f32-vbinary/gen/vdivc-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700311 "src/f32-vbinary/gen/vdivc-scalar-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800312 "src/f32-vbinary/gen/vmax-scalar-x1.c",
313 "src/f32-vbinary/gen/vmax-scalar-x2.c",
314 "src/f32-vbinary/gen/vmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700315 "src/f32-vbinary/gen/vmax-scalar-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800316 "src/f32-vbinary/gen/vmaxc-scalar-x1.c",
317 "src/f32-vbinary/gen/vmaxc-scalar-x2.c",
318 "src/f32-vbinary/gen/vmaxc-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700319 "src/f32-vbinary/gen/vmaxc-scalar-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800320 "src/f32-vbinary/gen/vmin-scalar-x1.c",
321 "src/f32-vbinary/gen/vmin-scalar-x2.c",
322 "src/f32-vbinary/gen/vmin-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700323 "src/f32-vbinary/gen/vmin-scalar-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800324 "src/f32-vbinary/gen/vminc-scalar-x1.c",
325 "src/f32-vbinary/gen/vminc-scalar-x2.c",
326 "src/f32-vbinary/gen/vminc-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700327 "src/f32-vbinary/gen/vminc-scalar-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700328 "src/f32-vbinary/gen/vmul-minmax-scalar-x1.c",
329 "src/f32-vbinary/gen/vmul-minmax-scalar-x2.c",
330 "src/f32-vbinary/gen/vmul-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700331 "src/f32-vbinary/gen/vmul-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700332 "src/f32-vbinary/gen/vmul-relu-scalar-x1.c",
333 "src/f32-vbinary/gen/vmul-relu-scalar-x2.c",
334 "src/f32-vbinary/gen/vmul-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700335 "src/f32-vbinary/gen/vmul-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700336 "src/f32-vbinary/gen/vmul-scalar-x1.c",
337 "src/f32-vbinary/gen/vmul-scalar-x2.c",
338 "src/f32-vbinary/gen/vmul-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700339 "src/f32-vbinary/gen/vmul-scalar-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700340 "src/f32-vbinary/gen/vmulc-minmax-scalar-x1.c",
341 "src/f32-vbinary/gen/vmulc-minmax-scalar-x2.c",
342 "src/f32-vbinary/gen/vmulc-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700343 "src/f32-vbinary/gen/vmulc-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700344 "src/f32-vbinary/gen/vmulc-relu-scalar-x1.c",
345 "src/f32-vbinary/gen/vmulc-relu-scalar-x2.c",
346 "src/f32-vbinary/gen/vmulc-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700347 "src/f32-vbinary/gen/vmulc-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700348 "src/f32-vbinary/gen/vmulc-scalar-x1.c",
349 "src/f32-vbinary/gen/vmulc-scalar-x2.c",
350 "src/f32-vbinary/gen/vmulc-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700351 "src/f32-vbinary/gen/vmulc-scalar-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700352 "src/f32-vbinary/gen/vrdivc-minmax-scalar-x1.c",
353 "src/f32-vbinary/gen/vrdivc-minmax-scalar-x2.c",
354 "src/f32-vbinary/gen/vrdivc-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700355 "src/f32-vbinary/gen/vrdivc-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700356 "src/f32-vbinary/gen/vrdivc-relu-scalar-x1.c",
357 "src/f32-vbinary/gen/vrdivc-relu-scalar-x2.c",
358 "src/f32-vbinary/gen/vrdivc-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700359 "src/f32-vbinary/gen/vrdivc-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700360 "src/f32-vbinary/gen/vrdivc-scalar-x1.c",
361 "src/f32-vbinary/gen/vrdivc-scalar-x2.c",
362 "src/f32-vbinary/gen/vrdivc-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700363 "src/f32-vbinary/gen/vrdivc-scalar-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700364 "src/f32-vbinary/gen/vrsubc-minmax-scalar-x1.c",
365 "src/f32-vbinary/gen/vrsubc-minmax-scalar-x2.c",
366 "src/f32-vbinary/gen/vrsubc-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700367 "src/f32-vbinary/gen/vrsubc-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700368 "src/f32-vbinary/gen/vrsubc-relu-scalar-x1.c",
369 "src/f32-vbinary/gen/vrsubc-relu-scalar-x2.c",
370 "src/f32-vbinary/gen/vrsubc-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700371 "src/f32-vbinary/gen/vrsubc-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700372 "src/f32-vbinary/gen/vrsubc-scalar-x1.c",
373 "src/f32-vbinary/gen/vrsubc-scalar-x2.c",
374 "src/f32-vbinary/gen/vrsubc-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700375 "src/f32-vbinary/gen/vrsubc-scalar-x8.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -0700376 "src/f32-vbinary/gen/vsqrdiff-scalar-x1.c",
377 "src/f32-vbinary/gen/vsqrdiff-scalar-x2.c",
378 "src/f32-vbinary/gen/vsqrdiff-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700379 "src/f32-vbinary/gen/vsqrdiff-scalar-x8.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -0700380 "src/f32-vbinary/gen/vsqrdiffc-scalar-x1.c",
381 "src/f32-vbinary/gen/vsqrdiffc-scalar-x2.c",
382 "src/f32-vbinary/gen/vsqrdiffc-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700383 "src/f32-vbinary/gen/vsqrdiffc-scalar-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700384 "src/f32-vbinary/gen/vsub-minmax-scalar-x1.c",
385 "src/f32-vbinary/gen/vsub-minmax-scalar-x2.c",
386 "src/f32-vbinary/gen/vsub-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700387 "src/f32-vbinary/gen/vsub-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700388 "src/f32-vbinary/gen/vsub-relu-scalar-x1.c",
389 "src/f32-vbinary/gen/vsub-relu-scalar-x2.c",
390 "src/f32-vbinary/gen/vsub-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700391 "src/f32-vbinary/gen/vsub-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700392 "src/f32-vbinary/gen/vsub-scalar-x1.c",
393 "src/f32-vbinary/gen/vsub-scalar-x2.c",
394 "src/f32-vbinary/gen/vsub-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700395 "src/f32-vbinary/gen/vsub-scalar-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700396 "src/f32-vbinary/gen/vsubc-minmax-scalar-x1.c",
397 "src/f32-vbinary/gen/vsubc-minmax-scalar-x2.c",
398 "src/f32-vbinary/gen/vsubc-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700399 "src/f32-vbinary/gen/vsubc-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700400 "src/f32-vbinary/gen/vsubc-relu-scalar-x1.c",
401 "src/f32-vbinary/gen/vsubc-relu-scalar-x2.c",
402 "src/f32-vbinary/gen/vsubc-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700403 "src/f32-vbinary/gen/vsubc-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700404 "src/f32-vbinary/gen/vsubc-scalar-x1.c",
405 "src/f32-vbinary/gen/vsubc-scalar-x2.c",
406 "src/f32-vbinary/gen/vsubc-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700407 "src/f32-vbinary/gen/vsubc-scalar-x8.c",
Marat Dukhanb1eec082021-05-05 23:24:55 -0700408 "src/f32-vclamp/gen/vclamp-scalar-x1.c",
409 "src/f32-vclamp/gen/vclamp-scalar-x2.c",
410 "src/f32-vclamp/gen/vclamp-scalar-x4.c",
Marat Dukhaned6baaf2020-12-01 15:07:08 -0800411 "src/f32-velu/gen/velu-scalar-rr2-lut16-p3-x1.c",
412 "src/f32-velu/gen/velu-scalar-rr2-lut16-p3-x2.c",
413 "src/f32-velu/gen/velu-scalar-rr2-lut16-p3-x3.c",
414 "src/f32-velu/gen/velu-scalar-rr2-lut16-p3-x4.c",
415 "src/f32-velu/gen/velu-scalar-rr2-lut16-p3-x5.c",
416 "src/f32-velu/gen/velu-scalar-rr2-lut16-p3-x6.c",
417 "src/f32-velu/gen/velu-scalar-rr2-p6-x1.c",
418 "src/f32-velu/gen/velu-scalar-rr2-p6-x2.c",
419 "src/f32-velu/gen/velu-scalar-rr2-p6-x3.c",
420 "src/f32-velu/gen/velu-scalar-rr2-p6-x4.c",
421 "src/f32-velu/gen/velu-scalar-rr2-p6-x5.c",
422 "src/f32-velu/gen/velu-scalar-rr2-p6-x6.c",
Marat Dukhan6674d692021-05-05 22:27:00 -0700423 "src/f32-vhswish/gen/vhswish-scalar-x1.c",
424 "src/f32-vhswish/gen/vhswish-scalar-x2.c",
425 "src/f32-vhswish/gen/vhswish-scalar-x4.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700426 "src/f32-vlrelu/gen/vlrelu-scalar-x1.c",
427 "src/f32-vlrelu/gen/vlrelu-scalar-x2.c",
428 "src/f32-vlrelu/gen/vlrelu-scalar-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700429 "src/f32-vmulcaddc/gen/c1-minmax-scalar-2x.c",
430 "src/f32-vmulcaddc/gen/c2-minmax-scalar-2x.c",
431 "src/f32-vmulcaddc/gen/c4-minmax-scalar-2x.c",
Marat Dukhanb1eec082021-05-05 23:24:55 -0700432 "src/f32-vrelu/gen/vrelu-scalar-x1.c",
433 "src/f32-vrelu/gen/vrelu-scalar-x2.c",
434 "src/f32-vrelu/gen/vrelu-scalar-x4.c",
435 "src/f32-vrelu/gen/vrelu-scalar-x8.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -0700436 "src/f32-vrnd/gen/vrndd-scalar-libm-x1.c",
437 "src/f32-vrnd/gen/vrndd-scalar-libm-x2.c",
438 "src/f32-vrnd/gen/vrndd-scalar-libm-x4.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700439 "src/f32-vrnd/gen/vrndne-scalar-libm-x1.c",
440 "src/f32-vrnd/gen/vrndne-scalar-libm-x2.c",
441 "src/f32-vrnd/gen/vrndne-scalar-libm-x4.c",
442 "src/f32-vrnd/gen/vrndu-scalar-libm-x1.c",
443 "src/f32-vrnd/gen/vrndu-scalar-libm-x2.c",
444 "src/f32-vrnd/gen/vrndu-scalar-libm-x4.c",
445 "src/f32-vrnd/gen/vrndz-scalar-libm-x1.c",
446 "src/f32-vrnd/gen/vrndz-scalar-libm-x2.c",
447 "src/f32-vrnd/gen/vrndz-scalar-libm-x4.c",
Marat Dukhanb1eec082021-05-05 23:24:55 -0700448 "src/f32-vsigmoid/gen/vsigmoid-scalar-lut64-p2-div-x1.c",
449 "src/f32-vsigmoid/gen/vsigmoid-scalar-lut64-p2-div-x2.c",
450 "src/f32-vsigmoid/gen/vsigmoid-scalar-lut64-p2-div-x4.c",
451 "src/f32-vsigmoid/gen/vsigmoid-scalar-lut2048-p1-div-x1.c",
452 "src/f32-vsigmoid/gen/vsigmoid-scalar-lut2048-p1-div-x2.c",
453 "src/f32-vsigmoid/gen/vsigmoid-scalar-lut2048-p1-div-x4.c",
454 "src/f32-vsigmoid/gen/vsigmoid-scalar-p5-div-x1.c",
455 "src/f32-vsigmoid/gen/vsigmoid-scalar-p5-div-x2.c",
456 "src/f32-vsigmoid/gen/vsigmoid-scalar-p5-div-x4.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -0700457 "src/f32-vsqrt/gen/scalar-sqrt-x1.c",
458 "src/f32-vsqrt/gen/scalar-sqrt-x2.c",
459 "src/f32-vsqrt/gen/scalar-sqrt-x4.c",
Marat Dukhan5020b962020-06-08 13:30:10 -0700460 "src/f32-vunary/gen/vabs-scalar-x1.c",
461 "src/f32-vunary/gen/vabs-scalar-x2.c",
462 "src/f32-vunary/gen/vabs-scalar-x4.c",
463 "src/f32-vunary/gen/vneg-scalar-x1.c",
464 "src/f32-vunary/gen/vneg-scalar-x2.c",
465 "src/f32-vunary/gen/vneg-scalar-x4.c",
466 "src/f32-vunary/gen/vsqr-scalar-x1.c",
467 "src/f32-vunary/gen/vsqr-scalar-x2.c",
468 "src/f32-vunary/gen/vsqr-scalar-x4.c",
Marat Dukhande390d42020-11-29 19:32:18 -0800469 "src/math/expm1minus-scalar-rr2-lut4-p4.c",
470 "src/math/expm1minus-scalar-rr2-lut8-p3.c",
471 "src/math/expm1minus-scalar-rr2-lut8-p4.c",
Marat Dukhanc60742b2020-11-23 12:33:27 -0800472 "src/math/expm1minus-scalar-rr2-lut16-p3.c",
473 "src/math/expm1minus-scalar-rr2-lut16-p4.c",
474 "src/math/expm1minus-scalar-rr2-p5.c",
475 "src/math/expm1minus-scalar-rr2-p6.c",
Frank Barchard22136062020-11-24 18:44:46 -0800476 "src/math/expminus-scalar-rr2-lut64-p2.c",
477 "src/math/expminus-scalar-rr2-lut2048-p1.c",
478 "src/math/expminus-scalar-rr2-p5.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -0700479 "src/math/roundd-scalar-addsub.c",
480 "src/math/roundd-scalar-cvt.c",
Marat Dukhanffbf96a2020-05-14 02:59:08 -0700481 "src/math/roundd-scalar-floor.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700482 "src/math/roundne-scalar-addsub.c",
483 "src/math/roundne-scalar-nearbyint.c",
484 "src/math/roundne-scalar-rint.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -0700485 "src/math/roundu-scalar-addsub.c",
Marat Dukhanffbf96a2020-05-14 02:59:08 -0700486 "src/math/roundu-scalar-ceil.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -0700487 "src/math/roundu-scalar-cvt.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -0700488 "src/math/roundz-scalar-addsub.c",
489 "src/math/roundz-scalar-cvt.c",
Marat Dukhanffbf96a2020-05-14 02:59:08 -0700490 "src/math/roundz-scalar-trunc.c",
Marat Dukhanf8475d62020-09-17 15:01:43 -0700491 "src/math/sigmoid-scalar-rr2-lut64-p2-div.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700492 "src/math/sigmoid-scalar-rr2-lut2048-p1-div.c",
Marat Dukhanf8475d62020-09-17 15:01:43 -0700493 "src/math/sigmoid-scalar-rr2-p5-div.c",
Marat Dukhan44542882021-05-10 09:21:48 -0700494 "src/qs8-dwconv/gen/up1x9-minmax-scalar.c",
495 "src/qs8-dwconv/gen/up2x9-minmax-scalar.c",
496 "src/qs8-dwconv/gen/up4x9-minmax-scalar.c",
Marat Dukhana1a4e782021-05-07 17:49:47 -0700497 "src/qs8-gemm/gen/1x2-minmax-scalar.c",
498 "src/qs8-gemm/gen/2x2-minmax-scalar.c",
499 "src/qs8-gemm/gen/3x2-minmax-scalar.c",
500 "src/qs8-gemm/gen/4x2-minmax-scalar.c",
501 "src/qs8-gemm/gen/1x4-minmax-scalar.c",
502 "src/qs8-gemm/gen/2x4-minmax-scalar.c",
503 "src/qs8-gemm/gen/3x4-minmax-scalar.c",
504 "src/qs8-gemm/gen/4x4-minmax-scalar.c",
505 "src/qs8-igemm/gen/1x2-minmax-scalar.c",
506 "src/qs8-igemm/gen/2x2-minmax-scalar.c",
507 "src/qs8-igemm/gen/3x2-minmax-scalar.c",
508 "src/qs8-igemm/gen/4x2-minmax-scalar.c",
509 "src/qs8-igemm/gen/1x4-minmax-scalar.c",
510 "src/qs8-igemm/gen/2x4-minmax-scalar.c",
511 "src/qs8-igemm/gen/3x4-minmax-scalar.c",
512 "src/qs8-igemm/gen/4x4-minmax-scalar.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -0700513 "src/qs8-requantization/fp32-scalar-lrintf.c",
514 "src/qs8-requantization/fp32-scalar-magic.c",
515 "src/qs8-requantization/precise-scalar-signed64.c",
516 "src/qs8-requantization/precise-scalar-unsigned32.c",
517 "src/qs8-requantization/precise-scalar-unsigned64.c",
518 "src/qs8-requantization/q31-scalar.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -0700519 "src/qu8-avgpool/9p8x-minmax-scalar-c1.c",
520 "src/qu8-avgpool/9x-minmax-scalar-c1.c",
521 "src/qu8-dwconv/up1x9-minmax-scalar.c",
522 "src/qu8-gavgpool/7p7x-minmax-scalar-c1.c",
523 "src/qu8-gavgpool/7x-minmax-scalar-c1.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700524 "src/qu8-gemm/2x2-minmax-scalar.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -0700525 "src/qu8-igemm/2x2-minmax-scalar.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -0700526 "src/qu8-requantization/fp32-scalar-lrintf.c",
527 "src/qu8-requantization/fp32-scalar-magic.c",
528 "src/qu8-requantization/precise-scalar-signed64.c",
529 "src/qu8-requantization/precise-scalar-unsigned32.c",
530 "src/qu8-requantization/precise-scalar-unsigned64.c",
531 "src/qu8-requantization/q31-scalar.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -0700532 "src/qu8-vadd/minmax-scalar.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700533 "src/u8-lut32norm/scalar.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700534 "src/u8-maxpool/9p8x-minmax-scalar-c1.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700535 "src/u8-rmax/scalar.c",
Marat Dukhan6674d692021-05-05 22:27:00 -0700536 "src/u8-vclamp/scalar-x4.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700537 "src/x8-lut/scalar.c",
538 "src/x8-zip/x2-scalar.c",
539 "src/x8-zip/x3-scalar.c",
540 "src/x8-zip/x4-scalar.c",
541 "src/x8-zip/xm-scalar.c",
Marat Dukhanad71b9a2020-11-20 00:01:51 -0800542 "src/x32-depthtospace2d-chw2hwc/scalar.c",
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -0700543 "src/x32-fill/scalar-float.c",
544 "src/x32-fill/scalar-int.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700545 "src/x32-packx/x2-scalar.c",
546 "src/x32-packx/x3-scalar.c",
547 "src/x32-packx/x4-scalar.c",
Marat Dukhan63523d42020-05-22 17:07:33 -0700548 "src/x32-pad/scalar-float.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700549 "src/x32-pad/scalar-int.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700550 "src/x32-unpool/scalar.c",
551 "src/x32-zip/x2-scalar.c",
552 "src/x32-zip/x3-scalar.c",
553 "src/x32-zip/x4-scalar.c",
554 "src/x32-zip/xm-scalar.c",
Marat Dukhan048931b2020-11-24 20:53:54 -0800555 "src/xx-copy/memcpy.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700556]
557
Marat Dukhan436ebe62019-12-04 15:10:12 -0800558WASM_UKERNELS = [
Marat Dukhan99936602020-04-11 16:47:01 -0700559 "src/f32-avgpool/9p8x-minmax-wasm-c1.c",
560 "src/f32-avgpool/9x-minmax-wasm-c1.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700561 "src/f32-dwconv/gen/up1x4-minmax-wasm-acc2.c",
562 "src/f32-dwconv/gen/up1x4-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700563 "src/f32-dwconv/gen/up1x4-wasm-acc2.c",
564 "src/f32-dwconv/gen/up1x4-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700565 "src/f32-dwconv/gen/up1x9-minmax-wasm-acc2.c",
566 "src/f32-dwconv/gen/up1x9-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700567 "src/f32-dwconv/gen/up1x9-wasm-acc2.c",
568 "src/f32-dwconv/gen/up1x9-wasm.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700569 "src/f32-dwconv/gen/up1x25-minmax-wasm-acc2.c",
570 "src/f32-dwconv/gen/up1x25-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700571 "src/f32-dwconv/gen/up1x25-wasm-acc2.c",
572 "src/f32-dwconv/gen/up1x25-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700573 "src/f32-dwconv/gen/up2x4-minmax-wasm-acc2.c",
574 "src/f32-dwconv/gen/up2x4-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700575 "src/f32-dwconv/gen/up2x4-wasm-acc2.c",
576 "src/f32-dwconv/gen/up2x4-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700577 "src/f32-dwconv/gen/up2x9-minmax-wasm-acc2.c",
578 "src/f32-dwconv/gen/up2x9-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700579 "src/f32-dwconv/gen/up2x9-wasm-acc2.c",
580 "src/f32-dwconv/gen/up2x9-wasm.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700581 "src/f32-dwconv/gen/up2x25-minmax-wasm-acc2.c",
582 "src/f32-dwconv/gen/up2x25-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700583 "src/f32-dwconv/gen/up2x25-wasm-acc2.c",
584 "src/f32-dwconv/gen/up2x25-wasm.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700585 "src/f32-gavgpool/7p7x-minmax-wasm-c1.c",
586 "src/f32-gavgpool/7x-minmax-wasm-c1.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700587 "src/f32-gemm/gen-inc/1x4inc-minmax-wasm.c",
588 "src/f32-gemm/gen-inc/2x4inc-minmax-wasm.c",
589 "src/f32-gemm/gen-inc/4x4inc-minmax-wasm.c",
590 "src/f32-gemm/gen/1x4-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700591 "src/f32-gemm/gen/1x4-relu-wasm.c",
592 "src/f32-gemm/gen/1x4-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700593 "src/f32-gemm/gen/2x4-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700594 "src/f32-gemm/gen/2x4-relu-wasm.c",
595 "src/f32-gemm/gen/2x4-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700596 "src/f32-gemm/gen/4x2-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700597 "src/f32-gemm/gen/4x2-relu-wasm.c",
598 "src/f32-gemm/gen/4x2-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700599 "src/f32-gemm/gen/4x4-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700600 "src/f32-gemm/gen/4x4-relu-wasm.c",
601 "src/f32-gemm/gen/4x4-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700602 "src/f32-igemm/gen/1x4-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700603 "src/f32-igemm/gen/1x4-relu-wasm.c",
604 "src/f32-igemm/gen/1x4-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700605 "src/f32-igemm/gen/2x4-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700606 "src/f32-igemm/gen/2x4-relu-wasm.c",
607 "src/f32-igemm/gen/2x4-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700608 "src/f32-igemm/gen/4x2-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700609 "src/f32-igemm/gen/4x2-relu-wasm.c",
610 "src/f32-igemm/gen/4x2-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700611 "src/f32-igemm/gen/4x4-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700612 "src/f32-igemm/gen/4x4-relu-wasm.c",
613 "src/f32-igemm/gen/4x4-wasm.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700614 "src/f32-maxpool/9p8x-minmax-wasm-c1.c",
615 "src/f32-pavgpool/9p8x-minmax-wasm-c1.c",
616 "src/f32-pavgpool/9x-minmax-wasm-c1.c",
Marat Dukhan7c1f8082020-06-25 13:26:20 -0700617 "src/f32-prelu/gen/wasm-2x1.c",
618 "src/f32-prelu/gen/wasm-2x4.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700619 "src/f32-vbinary/gen/vadd-minmax-wasm-x1.c",
620 "src/f32-vbinary/gen/vadd-minmax-wasm-x2.c",
621 "src/f32-vbinary/gen/vadd-minmax-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700622 "src/f32-vbinary/gen/vadd-minmax-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700623 "src/f32-vbinary/gen/vadd-relu-wasm-x1.c",
624 "src/f32-vbinary/gen/vadd-relu-wasm-x2.c",
625 "src/f32-vbinary/gen/vadd-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700626 "src/f32-vbinary/gen/vadd-relu-wasm-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700627 "src/f32-vbinary/gen/vaddc-minmax-wasm-x1.c",
628 "src/f32-vbinary/gen/vaddc-minmax-wasm-x2.c",
629 "src/f32-vbinary/gen/vaddc-minmax-wasm-x4.c",
630 "src/f32-vbinary/gen/vaddc-minmax-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700631 "src/f32-vbinary/gen/vaddc-relu-wasm-x1.c",
632 "src/f32-vbinary/gen/vaddc-relu-wasm-x2.c",
633 "src/f32-vbinary/gen/vaddc-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700634 "src/f32-vbinary/gen/vaddc-relu-wasm-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700635 "src/f32-vbinary/gen/vdiv-minmax-wasm-x1.c",
636 "src/f32-vbinary/gen/vdiv-minmax-wasm-x2.c",
637 "src/f32-vbinary/gen/vdiv-minmax-wasm-x4.c",
638 "src/f32-vbinary/gen/vdiv-minmax-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700639 "src/f32-vbinary/gen/vdiv-relu-wasm-x1.c",
640 "src/f32-vbinary/gen/vdiv-relu-wasm-x2.c",
641 "src/f32-vbinary/gen/vdiv-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700642 "src/f32-vbinary/gen/vdiv-relu-wasm-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700643 "src/f32-vbinary/gen/vdivc-minmax-wasm-x1.c",
644 "src/f32-vbinary/gen/vdivc-minmax-wasm-x2.c",
645 "src/f32-vbinary/gen/vdivc-minmax-wasm-x4.c",
646 "src/f32-vbinary/gen/vdivc-minmax-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700647 "src/f32-vbinary/gen/vdivc-relu-wasm-x1.c",
648 "src/f32-vbinary/gen/vdivc-relu-wasm-x2.c",
649 "src/f32-vbinary/gen/vdivc-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700650 "src/f32-vbinary/gen/vdivc-relu-wasm-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800651 "src/f32-vbinary/gen/vmax-wasm-x1.c",
652 "src/f32-vbinary/gen/vmax-wasm-x2.c",
653 "src/f32-vbinary/gen/vmax-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700654 "src/f32-vbinary/gen/vmax-wasm-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800655 "src/f32-vbinary/gen/vmaxc-wasm-x1.c",
656 "src/f32-vbinary/gen/vmaxc-wasm-x2.c",
657 "src/f32-vbinary/gen/vmaxc-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700658 "src/f32-vbinary/gen/vmaxc-wasm-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800659 "src/f32-vbinary/gen/vmin-wasm-x1.c",
660 "src/f32-vbinary/gen/vmin-wasm-x2.c",
661 "src/f32-vbinary/gen/vmin-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700662 "src/f32-vbinary/gen/vmin-wasm-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800663 "src/f32-vbinary/gen/vminc-wasm-x1.c",
664 "src/f32-vbinary/gen/vminc-wasm-x2.c",
665 "src/f32-vbinary/gen/vminc-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700666 "src/f32-vbinary/gen/vminc-wasm-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700667 "src/f32-vbinary/gen/vmul-minmax-wasm-x1.c",
668 "src/f32-vbinary/gen/vmul-minmax-wasm-x2.c",
669 "src/f32-vbinary/gen/vmul-minmax-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700670 "src/f32-vbinary/gen/vmul-minmax-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700671 "src/f32-vbinary/gen/vmul-relu-wasm-x1.c",
672 "src/f32-vbinary/gen/vmul-relu-wasm-x2.c",
673 "src/f32-vbinary/gen/vmul-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700674 "src/f32-vbinary/gen/vmul-relu-wasm-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700675 "src/f32-vbinary/gen/vmulc-minmax-wasm-x1.c",
676 "src/f32-vbinary/gen/vmulc-minmax-wasm-x2.c",
677 "src/f32-vbinary/gen/vmulc-minmax-wasm-x4.c",
678 "src/f32-vbinary/gen/vmulc-minmax-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700679 "src/f32-vbinary/gen/vmulc-relu-wasm-x1.c",
680 "src/f32-vbinary/gen/vmulc-relu-wasm-x2.c",
681 "src/f32-vbinary/gen/vmulc-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700682 "src/f32-vbinary/gen/vmulc-relu-wasm-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700683 "src/f32-vbinary/gen/vrdivc-minmax-wasm-x1.c",
684 "src/f32-vbinary/gen/vrdivc-minmax-wasm-x2.c",
685 "src/f32-vbinary/gen/vrdivc-minmax-wasm-x4.c",
686 "src/f32-vbinary/gen/vrdivc-minmax-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700687 "src/f32-vbinary/gen/vrdivc-relu-wasm-x1.c",
688 "src/f32-vbinary/gen/vrdivc-relu-wasm-x2.c",
689 "src/f32-vbinary/gen/vrdivc-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700690 "src/f32-vbinary/gen/vrdivc-relu-wasm-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700691 "src/f32-vbinary/gen/vrsubc-minmax-wasm-x1.c",
692 "src/f32-vbinary/gen/vrsubc-minmax-wasm-x2.c",
693 "src/f32-vbinary/gen/vrsubc-minmax-wasm-x4.c",
694 "src/f32-vbinary/gen/vrsubc-minmax-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700695 "src/f32-vbinary/gen/vrsubc-relu-wasm-x1.c",
696 "src/f32-vbinary/gen/vrsubc-relu-wasm-x2.c",
697 "src/f32-vbinary/gen/vrsubc-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700698 "src/f32-vbinary/gen/vrsubc-relu-wasm-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700699 "src/f32-vbinary/gen/vsub-minmax-wasm-x1.c",
700 "src/f32-vbinary/gen/vsub-minmax-wasm-x2.c",
701 "src/f32-vbinary/gen/vsub-minmax-wasm-x4.c",
702 "src/f32-vbinary/gen/vsub-minmax-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700703 "src/f32-vbinary/gen/vsub-relu-wasm-x1.c",
704 "src/f32-vbinary/gen/vsub-relu-wasm-x2.c",
705 "src/f32-vbinary/gen/vsub-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700706 "src/f32-vbinary/gen/vsub-relu-wasm-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700707 "src/f32-vbinary/gen/vsubc-minmax-wasm-x1.c",
708 "src/f32-vbinary/gen/vsubc-minmax-wasm-x2.c",
709 "src/f32-vbinary/gen/vsubc-minmax-wasm-x4.c",
710 "src/f32-vbinary/gen/vsubc-minmax-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700711 "src/f32-vbinary/gen/vsubc-relu-wasm-x1.c",
712 "src/f32-vbinary/gen/vsubc-relu-wasm-x2.c",
713 "src/f32-vbinary/gen/vsubc-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700714 "src/f32-vbinary/gen/vsubc-relu-wasm-x8.c",
Marat Dukhanb1eec082021-05-05 23:24:55 -0700715 "src/f32-vclamp/gen/vclamp-wasm-x1.c",
716 "src/f32-vclamp/gen/vclamp-wasm-x2.c",
717 "src/f32-vclamp/gen/vclamp-wasm-x4.c",
Marat Dukhaned6baaf2020-12-01 15:07:08 -0800718 "src/f32-velu/gen/velu-wasm-rr2-lut16-p3-x1.c",
719 "src/f32-velu/gen/velu-wasm-rr2-lut16-p3-x2.c",
720 "src/f32-velu/gen/velu-wasm-rr2-lut16-p3-x3.c",
721 "src/f32-velu/gen/velu-wasm-rr2-lut16-p3-x4.c",
722 "src/f32-velu/gen/velu-wasm-rr2-lut16-p3-x5.c",
723 "src/f32-velu/gen/velu-wasm-rr2-lut16-p3-x6.c",
724 "src/f32-velu/gen/velu-wasm-rr2-p6-x1.c",
725 "src/f32-velu/gen/velu-wasm-rr2-p6-x2.c",
726 "src/f32-velu/gen/velu-wasm-rr2-p6-x3.c",
727 "src/f32-velu/gen/velu-wasm-rr2-p6-x4.c",
728 "src/f32-velu/gen/velu-wasm-rr2-p6-x5.c",
729 "src/f32-velu/gen/velu-wasm-rr2-p6-x6.c",
Marat Dukhan6674d692021-05-05 22:27:00 -0700730 "src/f32-vhswish/gen/vhswish-wasm-x1.c",
731 "src/f32-vhswish/gen/vhswish-wasm-x2.c",
732 "src/f32-vhswish/gen/vhswish-wasm-x4.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -0700733 "src/f32-vlrelu/gen/vlrelu-wasm-x1.c",
734 "src/f32-vlrelu/gen/vlrelu-wasm-x2.c",
735 "src/f32-vlrelu/gen/vlrelu-wasm-x4.c",
Marat Dukhanb1eec082021-05-05 23:24:55 -0700736 "src/f32-vrelu/gen/vrelu-wasm-x1.c",
737 "src/f32-vrelu/gen/vrelu-wasm-x2.c",
738 "src/f32-vrelu/gen/vrelu-wasm-x4.c",
739 "src/f32-vrelu/gen/vrelu-wasm-x8.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700740 "src/f32-vmulcaddc/gen/c1-minmax-wasm-2x.c",
741 "src/f32-vmulcaddc/gen/c2-minmax-wasm-2x.c",
742 "src/f32-vmulcaddc/gen/c4-minmax-wasm-2x.c",
Marat Dukhan436ebe62019-12-04 15:10:12 -0800743]
744
Marat Dukhan290055c2020-06-09 12:24:29 -0700745WASMSIMD_UKERNELS = [
Marat Dukhan40f05522020-07-16 22:33:12 -0700746 "src/f32-argmaxpool/4x-wasmsimd-c4.c",
747 "src/f32-argmaxpool/9p8x-wasmsimd-c4.c",
748 "src/f32-argmaxpool/9x-wasmsimd-c4.c",
Marat Dukhan3b7432d2020-07-16 17:46:32 -0700749 "src/f32-avgpool/9p8x-minmax-wasmsimd-arm-c4.c",
750 "src/f32-avgpool/9p8x-minmax-wasmsimd-x86-c4.c",
751 "src/f32-avgpool/9x-minmax-wasmsimd-arm-c4.c",
752 "src/f32-avgpool/9x-minmax-wasmsimd-x86-c4.c",
Frank Barchard22136062020-11-24 18:44:46 -0800753 "src/f32-conv-hwc2chw/3x3s2p1c3x4-wasmsimd-2x2.c",
Frank Barchard0725b8d2020-12-07 11:07:35 -0800754 "src/f32-dwconv/gen/up4x4-minmax-wasmsimd-arm-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700755 "src/f32-dwconv/gen/up4x4-minmax-wasmsimd-arm.c",
Frank Barchard0725b8d2020-12-07 11:07:35 -0800756 "src/f32-dwconv/gen/up4x4-minmax-wasmsimd-x86-acc2.c",
Marat Dukhanac014d72020-06-16 08:36:47 -0700757 "src/f32-dwconv/gen/up4x4-minmax-wasmsimd-x86.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700758 "src/f32-dwconv/gen/up4x4-wasmsimd.c",
Frank Barchard0725b8d2020-12-07 11:07:35 -0800759 "src/f32-dwconv/gen/up4x9-minmax-wasmsimd-arm-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700760 "src/f32-dwconv/gen/up4x9-minmax-wasmsimd-arm.c",
Frank Barchard0725b8d2020-12-07 11:07:35 -0800761 "src/f32-dwconv/gen/up4x9-minmax-wasmsimd-x86-acc2.c",
Marat Dukhanac014d72020-06-16 08:36:47 -0700762 "src/f32-dwconv/gen/up4x9-minmax-wasmsimd-x86.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700763 "src/f32-dwconv/gen/up4x9-wasmsimd.c",
Frank Barchard0725b8d2020-12-07 11:07:35 -0800764 "src/f32-dwconv/gen/up4x25-minmax-wasmsimd-arm-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700765 "src/f32-dwconv/gen/up4x25-minmax-wasmsimd-arm.c",
Frank Barchard0725b8d2020-12-07 11:07:35 -0800766 "src/f32-dwconv/gen/up4x25-minmax-wasmsimd-x86-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700767 "src/f32-dwconv/gen/up4x25-minmax-wasmsimd-x86.c",
768 "src/f32-dwconv/gen/up4x25-wasmsimd.c",
Frank Barchard0725b8d2020-12-07 11:07:35 -0800769 "src/f32-dwconv/gen/up8x4-minmax-wasmsimd-arm-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700770 "src/f32-dwconv/gen/up8x4-minmax-wasmsimd-arm.c",
Frank Barchard0725b8d2020-12-07 11:07:35 -0800771 "src/f32-dwconv/gen/up8x4-minmax-wasmsimd-x86-acc2.c",
Marat Dukhanac014d72020-06-16 08:36:47 -0700772 "src/f32-dwconv/gen/up8x4-minmax-wasmsimd-x86.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700773 "src/f32-dwconv/gen/up8x4-wasmsimd.c",
Frank Barchard0725b8d2020-12-07 11:07:35 -0800774 "src/f32-dwconv/gen/up8x9-minmax-wasmsimd-arm-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700775 "src/f32-dwconv/gen/up8x9-minmax-wasmsimd-arm.c",
Frank Barchard0725b8d2020-12-07 11:07:35 -0800776 "src/f32-dwconv/gen/up8x9-minmax-wasmsimd-x86-acc2.c",
Marat Dukhanac014d72020-06-16 08:36:47 -0700777 "src/f32-dwconv/gen/up8x9-minmax-wasmsimd-x86.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700778 "src/f32-dwconv/gen/up8x9-wasmsimd.c",
Frank Barchard0725b8d2020-12-07 11:07:35 -0800779 "src/f32-dwconv/gen/up8x25-minmax-wasmsimd-arm-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700780 "src/f32-dwconv/gen/up8x25-minmax-wasmsimd-arm.c",
Frank Barchard0725b8d2020-12-07 11:07:35 -0800781 "src/f32-dwconv/gen/up8x25-minmax-wasmsimd-x86-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700782 "src/f32-dwconv/gen/up8x25-minmax-wasmsimd-x86.c",
783 "src/f32-dwconv/gen/up8x25-wasmsimd.c",
Frank Barchard412e2f42020-12-11 11:40:50 -0800784 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-loadsplat-1x4-acc2.c",
785 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-loadsplat-1x4-acc3.c",
786 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-loadsplat-1x4-acc4.c",
787 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-loadsplat-1x4.c",
788 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-loadsplat-2x4-acc2.c",
789 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-loadsplat-2x4.c",
790 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-loadsplat-3x4.c",
791 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-loadsplat-4x4.c",
792 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-loadsplat-5x4.c",
793 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-loadsplat-6x4.c",
Frank Barchard02bb4292020-12-15 18:25:32 -0800794 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-splat-1x4-acc2.c",
795 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-splat-1x4-acc3.c",
796 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-splat-1x4-acc4.c",
797 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-splat-1x4.c",
798 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-splat-2x4-acc2.c",
799 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-splat-2x4.c",
800 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-splat-3x4.c",
801 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-splat-4x4.c",
802 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-splat-5x4.c",
803 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-splat-6x4.c",
Frank Barchard412e2f42020-12-11 11:40:50 -0800804 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-loadsplat-1x4-acc2.c",
805 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-loadsplat-1x4-acc3.c",
806 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-loadsplat-1x4-acc4.c",
807 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-loadsplat-1x4.c",
808 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-loadsplat-2x4-acc2.c",
809 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-loadsplat-2x4.c",
810 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-loadsplat-3x4.c",
811 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-loadsplat-4x4.c",
812 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-loadsplat-5x4.c",
813 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-loadsplat-6x4.c",
Frank Barchard02bb4292020-12-15 18:25:32 -0800814 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-splat-1x4-acc2.c",
815 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-splat-1x4-acc3.c",
816 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-splat-1x4-acc4.c",
817 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-splat-1x4.c",
818 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-splat-2x4-acc2.c",
819 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-splat-2x4.c",
820 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-splat-3x4.c",
821 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-splat-4x4.c",
822 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-splat-5x4.c",
823 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-splat-6x4.c",
Frank Barchardc5704bf2020-12-21 23:09:00 -0800824 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-arm-loadsplat-1x4-acc2.c",
825 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-arm-loadsplat-1x4-acc3.c",
826 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-arm-loadsplat-1x4-acc4.c",
827 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-arm-loadsplat-1x4.c",
828 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-arm-loadsplat-2x4-acc2.c",
829 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-arm-loadsplat-2x4.c",
830 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-arm-loadsplat-3x4.c",
831 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-arm-loadsplat-4x4.c",
Frank Barchard412e2f42020-12-11 11:40:50 -0800832 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-arm-splat-1x4-acc2.c",
833 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-arm-splat-1x4-acc3.c",
834 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-arm-splat-1x4-acc4.c",
835 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-arm-splat-1x4.c",
836 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-arm-splat-2x4-acc2.c",
837 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-arm-splat-2x4.c",
838 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-arm-splat-3x4.c",
839 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-arm-splat-4x4.c",
Frank Barchardcadd4222021-01-20 16:27:25 -0800840 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-x86-loadsplat-1x4-acc2.c",
841 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-x86-loadsplat-1x4-acc3.c",
842 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-x86-loadsplat-1x4-acc4.c",
843 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-x86-loadsplat-1x4.c",
844 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-x86-loadsplat-2x4-acc2.c",
845 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-x86-loadsplat-2x4.c",
846 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-x86-loadsplat-3x4.c",
847 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-x86-loadsplat-4x4.c",
Frank Barchard412e2f42020-12-11 11:40:50 -0800848 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-x86-splat-1x4-acc2.c",
849 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-x86-splat-1x4-acc3.c",
850 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-x86-splat-1x4-acc4.c",
851 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-x86-splat-1x4.c",
852 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-x86-splat-2x4-acc2.c",
853 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-x86-splat-2x4.c",
854 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-x86-splat-3x4.c",
855 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-x86-splat-4x4.c",
Frank Barchardb20dcd62020-12-15 16:46:14 -0800856 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-loadsplat-1x4-acc2.c",
857 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-loadsplat-1x4-acc3.c",
858 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-loadsplat-1x4-acc4.c",
859 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-loadsplat-1x4-acc5.c",
860 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-loadsplat-1x4.c",
861 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-loadsplat-2x4-acc2.c",
862 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-loadsplat-2x4-acc3.c",
863 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-loadsplat-2x4.c",
864 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-loadsplat-3x4-acc2.c",
865 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-loadsplat-3x4.c",
866 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-loadsplat-4x4-acc2.c",
867 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-loadsplat-4x4.c",
868 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-loadsplat-5x4.c",
Frank Barchard412e2f42020-12-11 11:40:50 -0800869 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-splat-1x4-acc2.c",
870 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-splat-1x4-acc3.c",
871 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-splat-1x4-acc4.c",
872 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-splat-1x4-acc5.c",
873 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-splat-1x4.c",
874 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-splat-2x4-acc2.c",
875 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-splat-2x4-acc3.c",
876 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-splat-2x4.c",
877 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-splat-3x4-acc2.c",
878 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-splat-3x4.c",
879 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-splat-4x4-acc2.c",
880 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-splat-4x4.c",
881 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-splat-5x4.c",
Frank Barchardb20dcd62020-12-15 16:46:14 -0800882 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-loadsplat-1x4-acc2.c",
883 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-loadsplat-1x4-acc3.c",
884 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-loadsplat-1x4-acc4.c",
885 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-loadsplat-1x4-acc5.c",
886 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-loadsplat-1x4.c",
887 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-loadsplat-2x4-acc2.c",
888 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-loadsplat-2x4-acc3.c",
889 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-loadsplat-2x4.c",
890 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-loadsplat-3x4-acc2.c",
891 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-loadsplat-3x4.c",
892 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-loadsplat-4x4-acc2.c",
893 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-loadsplat-4x4.c",
894 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-loadsplat-5x4.c",
Frank Barchard412e2f42020-12-11 11:40:50 -0800895 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-splat-1x4-acc2.c",
896 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-splat-1x4-acc3.c",
897 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-splat-1x4-acc4.c",
898 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-splat-1x4-acc5.c",
899 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-splat-1x4.c",
900 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-splat-2x4-acc2.c",
901 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-splat-2x4-acc3.c",
902 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-splat-2x4.c",
903 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-splat-3x4-acc2.c",
904 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-splat-3x4.c",
905 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-splat-4x4-acc2.c",
906 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-splat-4x4.c",
907 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-splat-5x4.c",
Frank Barchardc6889b32020-12-21 11:27:22 -0800908 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-arm-loadsplat-1x4-acc2.c",
909 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-arm-loadsplat-1x4-acc3.c",
910 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-arm-loadsplat-1x4-acc4.c",
911 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-arm-loadsplat-1x4-acc5.c",
912 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-arm-loadsplat-1x4.c",
913 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-arm-loadsplat-2x4-acc2.c",
914 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-arm-loadsplat-2x4-acc3.c",
915 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-arm-loadsplat-2x4.c",
916 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-arm-loadsplat-3x4-acc2.c",
917 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-arm-loadsplat-3x4.c",
Frank Barchard412e2f42020-12-11 11:40:50 -0800918 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-arm-splat-1x4-acc2.c",
919 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-arm-splat-1x4-acc3.c",
920 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-arm-splat-1x4-acc4.c",
921 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-arm-splat-1x4-acc5.c",
922 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-arm-splat-1x4.c",
923 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-arm-splat-2x4-acc2.c",
924 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-arm-splat-2x4-acc3.c",
925 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-arm-splat-2x4.c",
926 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-arm-splat-3x4-acc2.c",
927 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-arm-splat-3x4.c",
Frank Barchardc6889b32020-12-21 11:27:22 -0800928 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-x86-loadsplat-1x4-acc2.c",
929 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-x86-loadsplat-1x4-acc3.c",
930 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-x86-loadsplat-1x4-acc4.c",
931 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-x86-loadsplat-1x4-acc5.c",
932 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-x86-loadsplat-1x4.c",
933 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-x86-loadsplat-2x4-acc2.c",
934 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-x86-loadsplat-2x4-acc3.c",
935 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-x86-loadsplat-2x4.c",
936 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-x86-loadsplat-3x4-acc2.c",
937 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-x86-loadsplat-3x4.c",
Frank Barchard412e2f42020-12-11 11:40:50 -0800938 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-x86-splat-1x4-acc2.c",
939 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-x86-splat-1x4-acc3.c",
940 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-x86-splat-1x4-acc4.c",
941 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-x86-splat-1x4-acc5.c",
942 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-x86-splat-1x4.c",
943 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-x86-splat-2x4-acc2.c",
944 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-x86-splat-2x4-acc3.c",
945 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-x86-splat-2x4.c",
946 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-x86-splat-3x4-acc2.c",
947 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-x86-splat-3x4.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700948 "src/f32-gavgpool-cw/wasmsimd-arm-x4.c",
949 "src/f32-gavgpool-cw/wasmsimd-x86-x4.c",
Marat Dukhanc6016802020-07-16 18:51:28 -0700950 "src/f32-gavgpool/7p7x-minmax-wasmsimd-arm-c4.c",
951 "src/f32-gavgpool/7p7x-minmax-wasmsimd-x86-c4.c",
952 "src/f32-gavgpool/7x-minmax-wasmsimd-arm-c4.c",
953 "src/f32-gavgpool/7x-minmax-wasmsimd-x86-c4.c",
Frank Barchard0725b8d2020-12-07 11:07:35 -0800954 "src/f32-gemm/gen-inc/1x8inc-minmax-wasmsimd-arm-loadsplat.c",
955 "src/f32-gemm/gen-inc/1x8inc-minmax-wasmsimd-arm-splat.c",
956 "src/f32-gemm/gen-inc/1x8inc-minmax-wasmsimd-x86-loadsplat.c",
957 "src/f32-gemm/gen-inc/1x8inc-minmax-wasmsimd-x86-splat.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700958 "src/f32-gemm/gen-inc/1x8s4inc-minmax-wasmsimd-arm.c",
959 "src/f32-gemm/gen-inc/1x8s4inc-minmax-wasmsimd-x86.c",
Frank Barchard0725b8d2020-12-07 11:07:35 -0800960 "src/f32-gemm/gen-inc/3x8inc-minmax-wasmsimd-arm-loadsplat.c",
961 "src/f32-gemm/gen-inc/3x8inc-minmax-wasmsimd-arm-splat.c",
962 "src/f32-gemm/gen-inc/3x8inc-minmax-wasmsimd-x86-loadsplat.c",
963 "src/f32-gemm/gen-inc/3x8inc-minmax-wasmsimd-x86-splat.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700964 "src/f32-gemm/gen-inc/3x8s4inc-minmax-wasmsimd-arm.c",
965 "src/f32-gemm/gen-inc/3x8s4inc-minmax-wasmsimd-x86.c",
Frank Barchard0725b8d2020-12-07 11:07:35 -0800966 "src/f32-gemm/gen-inc/4x8inc-minmax-wasmsimd-arm-loadsplat.c",
967 "src/f32-gemm/gen-inc/4x8inc-minmax-wasmsimd-arm-splat.c",
968 "src/f32-gemm/gen-inc/4x8inc-minmax-wasmsimd-x86-loadsplat.c",
969 "src/f32-gemm/gen-inc/4x8inc-minmax-wasmsimd-x86-splat.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700970 "src/f32-gemm/gen-inc/4x8s4inc-minmax-wasmsimd-arm.c",
971 "src/f32-gemm/gen-inc/4x8s4inc-minmax-wasmsimd-x86.c",
Frank Barchard0725b8d2020-12-07 11:07:35 -0800972 "src/f32-gemm/gen-inc/5x8inc-minmax-wasmsimd-arm-loadsplat.c",
973 "src/f32-gemm/gen-inc/5x8inc-minmax-wasmsimd-arm-splat.c",
974 "src/f32-gemm/gen-inc/5x8inc-minmax-wasmsimd-x86-loadsplat.c",
975 "src/f32-gemm/gen-inc/5x8inc-minmax-wasmsimd-x86-splat.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700976 "src/f32-gemm/gen-inc/5x8s4inc-minmax-wasmsimd-arm.c",
977 "src/f32-gemm/gen-inc/5x8s4inc-minmax-wasmsimd-x86.c",
Frank Barchard0725b8d2020-12-07 11:07:35 -0800978 "src/f32-gemm/gen-inc/6x8inc-minmax-wasmsimd-arm-loadsplat.c",
979 "src/f32-gemm/gen-inc/6x8inc-minmax-wasmsimd-arm-splat.c",
980 "src/f32-gemm/gen-inc/6x8inc-minmax-wasmsimd-x86-loadsplat.c",
981 "src/f32-gemm/gen-inc/6x8inc-minmax-wasmsimd-x86-splat.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700982 "src/f32-gemm/gen-inc/6x8s4inc-minmax-wasmsimd-arm.c",
983 "src/f32-gemm/gen-inc/6x8s4inc-minmax-wasmsimd-x86.c",
Frank Barchard0725b8d2020-12-07 11:07:35 -0800984 "src/f32-gemm/gen/1x8-minmax-wasmsimd-arm-loadsplat.c",
985 "src/f32-gemm/gen/1x8-minmax-wasmsimd-arm-splat.c",
986 "src/f32-gemm/gen/1x8-minmax-wasmsimd-x86-loadsplat.c",
987 "src/f32-gemm/gen/1x8-minmax-wasmsimd-x86-splat.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700988 "src/f32-gemm/gen/1x8-relu-wasmsimd-splat.c",
989 "src/f32-gemm/gen/1x8-wasmsimd-splat.c",
990 "src/f32-gemm/gen/1x8s4-minmax-wasmsimd-arm.c",
991 "src/f32-gemm/gen/1x8s4-minmax-wasmsimd-x86.c",
Frank Barchard0725b8d2020-12-07 11:07:35 -0800992 "src/f32-gemm/gen/3x8-minmax-wasmsimd-arm-loadsplat.c",
993 "src/f32-gemm/gen/3x8-minmax-wasmsimd-arm-splat.c",
994 "src/f32-gemm/gen/3x8-minmax-wasmsimd-x86-loadsplat.c",
995 "src/f32-gemm/gen/3x8-minmax-wasmsimd-x86-splat.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700996 "src/f32-gemm/gen/3x8s4-minmax-wasmsimd-arm.c",
997 "src/f32-gemm/gen/3x8s4-minmax-wasmsimd-x86.c",
998 "src/f32-gemm/gen/4x2c4-minmax-wasmsimd-arm.c",
999 "src/f32-gemm/gen/4x2c4-minmax-wasmsimd-x86.c",
1000 "src/f32-gemm/gen/4x2c4-relu-wasmsimd.c",
1001 "src/f32-gemm/gen/4x2c4-wasmsimd.c",
Frank Barchard0725b8d2020-12-07 11:07:35 -08001002 "src/f32-gemm/gen/4x8-minmax-wasmsimd-arm-loadsplat.c",
1003 "src/f32-gemm/gen/4x8-minmax-wasmsimd-arm-splat.c",
1004 "src/f32-gemm/gen/4x8-minmax-wasmsimd-x86-loadsplat.c",
1005 "src/f32-gemm/gen/4x8-minmax-wasmsimd-x86-splat.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001006 "src/f32-gemm/gen/4x8-relu-wasmsimd-splat.c",
1007 "src/f32-gemm/gen/4x8-wasmsimd-splat.c",
1008 "src/f32-gemm/gen/4x8s4-minmax-wasmsimd-arm.c",
1009 "src/f32-gemm/gen/4x8s4-minmax-wasmsimd-x86.c",
Frank Barchard0725b8d2020-12-07 11:07:35 -08001010 "src/f32-gemm/gen/5x8-minmax-wasmsimd-arm-loadsplat.c",
1011 "src/f32-gemm/gen/5x8-minmax-wasmsimd-arm-splat.c",
1012 "src/f32-gemm/gen/5x8-minmax-wasmsimd-x86-loadsplat.c",
1013 "src/f32-gemm/gen/5x8-minmax-wasmsimd-x86-splat.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001014 "src/f32-gemm/gen/5x8-relu-wasmsimd-splat.c",
1015 "src/f32-gemm/gen/5x8-wasmsimd-splat.c",
1016 "src/f32-gemm/gen/5x8s4-minmax-wasmsimd-arm.c",
1017 "src/f32-gemm/gen/5x8s4-minmax-wasmsimd-x86.c",
Frank Barchard0725b8d2020-12-07 11:07:35 -08001018 "src/f32-gemm/gen/6x8-minmax-wasmsimd-arm-loadsplat.c",
1019 "src/f32-gemm/gen/6x8-minmax-wasmsimd-arm-splat.c",
1020 "src/f32-gemm/gen/6x8-minmax-wasmsimd-x86-loadsplat.c",
1021 "src/f32-gemm/gen/6x8-minmax-wasmsimd-x86-splat.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001022 "src/f32-gemm/gen/6x8s4-minmax-wasmsimd-arm.c",
1023 "src/f32-gemm/gen/6x8s4-minmax-wasmsimd-x86.c",
XNNPACK Team965272b2020-10-23 21:10:15 -07001024 "src/f32-ibilinear-chw/gen/wasmsimd-p4.c",
1025 "src/f32-ibilinear-chw/gen/wasmsimd-p8.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001026 "src/f32-ibilinear/gen/wasmsimd-c4.c",
1027 "src/f32-ibilinear/gen/wasmsimd-c8.c",
Frank Barchard0725b8d2020-12-07 11:07:35 -08001028 "src/f32-igemm/gen/1x8-minmax-wasmsimd-arm-loadsplat.c",
1029 "src/f32-igemm/gen/1x8-minmax-wasmsimd-arm-splat.c",
1030 "src/f32-igemm/gen/1x8-minmax-wasmsimd-x86-loadsplat.c",
1031 "src/f32-igemm/gen/1x8-minmax-wasmsimd-x86-splat.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001032 "src/f32-igemm/gen/1x8-relu-wasmsimd-splat.c",
1033 "src/f32-igemm/gen/1x8-wasmsimd-splat.c",
1034 "src/f32-igemm/gen/1x8s4-minmax-wasmsimd-arm.c",
1035 "src/f32-igemm/gen/1x8s4-minmax-wasmsimd-x86.c",
Frank Barchard0725b8d2020-12-07 11:07:35 -08001036 "src/f32-igemm/gen/3x8-minmax-wasmsimd-arm-loadsplat.c",
1037 "src/f32-igemm/gen/3x8-minmax-wasmsimd-arm-splat.c",
1038 "src/f32-igemm/gen/3x8-minmax-wasmsimd-x86-loadsplat.c",
1039 "src/f32-igemm/gen/3x8-minmax-wasmsimd-x86-splat.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001040 "src/f32-igemm/gen/3x8s4-minmax-wasmsimd-arm.c",
1041 "src/f32-igemm/gen/3x8s4-minmax-wasmsimd-x86.c",
1042 "src/f32-igemm/gen/4x2c4-minmax-wasmsimd-arm.c",
1043 "src/f32-igemm/gen/4x2c4-minmax-wasmsimd-x86.c",
1044 "src/f32-igemm/gen/4x2c4-relu-wasmsimd.c",
1045 "src/f32-igemm/gen/4x2c4-wasmsimd.c",
Frank Barchard0725b8d2020-12-07 11:07:35 -08001046 "src/f32-igemm/gen/4x8-minmax-wasmsimd-arm-loadsplat.c",
1047 "src/f32-igemm/gen/4x8-minmax-wasmsimd-arm-splat.c",
1048 "src/f32-igemm/gen/4x8-minmax-wasmsimd-x86-loadsplat.c",
1049 "src/f32-igemm/gen/4x8-minmax-wasmsimd-x86-splat.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001050 "src/f32-igemm/gen/4x8-relu-wasmsimd-splat.c",
1051 "src/f32-igemm/gen/4x8-wasmsimd-splat.c",
1052 "src/f32-igemm/gen/4x8s4-minmax-wasmsimd-arm.c",
1053 "src/f32-igemm/gen/4x8s4-minmax-wasmsimd-x86.c",
Frank Barchard0725b8d2020-12-07 11:07:35 -08001054 "src/f32-igemm/gen/5x8-minmax-wasmsimd-arm-loadsplat.c",
1055 "src/f32-igemm/gen/5x8-minmax-wasmsimd-arm-splat.c",
1056 "src/f32-igemm/gen/5x8-minmax-wasmsimd-x86-loadsplat.c",
1057 "src/f32-igemm/gen/5x8-minmax-wasmsimd-x86-splat.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001058 "src/f32-igemm/gen/5x8-relu-wasmsimd-splat.c",
1059 "src/f32-igemm/gen/5x8-wasmsimd-splat.c",
1060 "src/f32-igemm/gen/5x8s4-minmax-wasmsimd-arm.c",
1061 "src/f32-igemm/gen/5x8s4-minmax-wasmsimd-x86.c",
Frank Barchard0725b8d2020-12-07 11:07:35 -08001062 "src/f32-igemm/gen/6x8-minmax-wasmsimd-arm-loadsplat.c",
1063 "src/f32-igemm/gen/6x8-minmax-wasmsimd-arm-splat.c",
1064 "src/f32-igemm/gen/6x8-minmax-wasmsimd-x86-loadsplat.c",
1065 "src/f32-igemm/gen/6x8-minmax-wasmsimd-x86-splat.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001066 "src/f32-igemm/gen/6x8s4-minmax-wasmsimd-arm.c",
1067 "src/f32-igemm/gen/6x8s4-minmax-wasmsimd-x86.c",
Marat Dukhanf6e24802020-07-08 22:20:40 -07001068 "src/f32-maxpool/9p8x-minmax-wasmsimd-arm-c4.c",
1069 "src/f32-maxpool/9p8x-minmax-wasmsimd-x86-c4.c",
Marat Dukhan1483c532020-07-16 18:08:19 -07001070 "src/f32-pavgpool/9p8x-minmax-wasmsimd-arm-c4.c",
1071 "src/f32-pavgpool/9p8x-minmax-wasmsimd-x86-c4.c",
1072 "src/f32-pavgpool/9x-minmax-wasmsimd-arm-c4.c",
1073 "src/f32-pavgpool/9x-minmax-wasmsimd-x86-c4.c",
Frank Barchard0725b8d2020-12-07 11:07:35 -08001074 "src/f32-ppmm/gen/4x8-minmax-wasmsimd-arm-splat.c",
1075 "src/f32-ppmm/gen/4x8-minmax-wasmsimd-x86-splat.c",
Frank Barcharda5316982020-07-23 13:19:28 -07001076 "src/f32-prelu/gen/wasmsimd-bitselect-1x4.c",
1077 "src/f32-prelu/gen/wasmsimd-bitselect-1x8.c",
1078 "src/f32-prelu/gen/wasmsimd-bitselect-1x16.c",
Marat Dukhan195f8eb2020-06-25 12:50:57 -07001079 "src/f32-prelu/gen/wasmsimd-bitselect-2x4.c",
1080 "src/f32-prelu/gen/wasmsimd-bitselect-2x8.c",
Frank Barcharda5316982020-07-23 13:19:28 -07001081 "src/f32-prelu/gen/wasmsimd-bitselect-2x16.c",
1082 "src/f32-prelu/gen/wasmsimd-bitselect-4x4.c",
1083 "src/f32-prelu/gen/wasmsimd-bitselect-4x8.c",
1084 "src/f32-prelu/gen/wasmsimd-bitselect-4x16.c",
1085 "src/f32-prelu/gen/wasmsimd-minmax-1x4.c",
1086 "src/f32-prelu/gen/wasmsimd-minmax-1x8.c",
1087 "src/f32-prelu/gen/wasmsimd-minmax-1x16.c",
Marat Dukhan195f8eb2020-06-25 12:50:57 -07001088 "src/f32-prelu/gen/wasmsimd-minmax-2x4.c",
1089 "src/f32-prelu/gen/wasmsimd-minmax-2x8.c",
Frank Barcharda5316982020-07-23 13:19:28 -07001090 "src/f32-prelu/gen/wasmsimd-minmax-2x16.c",
1091 "src/f32-prelu/gen/wasmsimd-minmax-4x4.c",
1092 "src/f32-prelu/gen/wasmsimd-minmax-4x8.c",
1093 "src/f32-prelu/gen/wasmsimd-minmax-4x16.c",
Marat Dukhan52238f02020-07-16 15:30:28 -07001094 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x4.c",
Marat Dukhan52238f02020-07-16 15:30:28 -07001095 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x8-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001096 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x8.c",
Marat Dukhan52238f02020-07-16 15:30:28 -07001097 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x12-acc2.c",
1098 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x12-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001099 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x12.c",
Marat Dukhan52238f02020-07-16 15:30:28 -07001100 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x16-acc2.c",
1101 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x16-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001102 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x16.c",
Marat Dukhan52238f02020-07-16 15:30:28 -07001103 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x20-acc2.c",
1104 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x20-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001105 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x20.c",
Marat Dukhan8c417962020-07-08 12:27:50 -07001106 "src/f32-rmax/wasmsimd-arm.c",
1107 "src/f32-rmax/wasmsimd-x86.c",
Frank Barchard8ef44cd2020-11-03 12:30:23 -08001108 "src/f32-spmm/gen/4x1-minmax-wasmsimd-arm-pipelined-x2.c",
1109 "src/f32-spmm/gen/4x1-minmax-wasmsimd-arm-pipelined.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07001110 "src/f32-spmm/gen/4x1-minmax-wasmsimd-arm-x2.c",
1111 "src/f32-spmm/gen/4x1-minmax-wasmsimd-arm-x4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001112 "src/f32-spmm/gen/4x1-minmax-wasmsimd-arm.c",
Frank Barchard8ef44cd2020-11-03 12:30:23 -08001113 "src/f32-spmm/gen/4x1-minmax-wasmsimd-x86-pipelined-x2.c",
1114 "src/f32-spmm/gen/4x1-minmax-wasmsimd-x86-pipelined.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07001115 "src/f32-spmm/gen/4x1-minmax-wasmsimd-x86-x2.c",
1116 "src/f32-spmm/gen/4x1-minmax-wasmsimd-x86-x4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001117 "src/f32-spmm/gen/4x1-minmax-wasmsimd-x86.c",
Frank Barchard8ef44cd2020-11-03 12:30:23 -08001118 "src/f32-spmm/gen/8x1-minmax-wasmsimd-arm-pipelined-x2.c",
1119 "src/f32-spmm/gen/8x1-minmax-wasmsimd-arm-pipelined.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07001120 "src/f32-spmm/gen/8x1-minmax-wasmsimd-arm-x2.c",
1121 "src/f32-spmm/gen/8x1-minmax-wasmsimd-arm-x4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001122 "src/f32-spmm/gen/8x1-minmax-wasmsimd-arm.c",
Frank Barchard8ef44cd2020-11-03 12:30:23 -08001123 "src/f32-spmm/gen/8x1-minmax-wasmsimd-x86-pipelined-x2.c",
1124 "src/f32-spmm/gen/8x1-minmax-wasmsimd-x86-pipelined.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07001125 "src/f32-spmm/gen/8x1-minmax-wasmsimd-x86-x2.c",
1126 "src/f32-spmm/gen/8x1-minmax-wasmsimd-x86-x4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001127 "src/f32-spmm/gen/8x1-minmax-wasmsimd-x86.c",
Frank Barchard8ef44cd2020-11-03 12:30:23 -08001128 "src/f32-spmm/gen/16x1-minmax-wasmsimd-arm-pipelined-x2.c",
1129 "src/f32-spmm/gen/16x1-minmax-wasmsimd-arm-pipelined.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07001130 "src/f32-spmm/gen/16x1-minmax-wasmsimd-arm-x2.c",
1131 "src/f32-spmm/gen/16x1-minmax-wasmsimd-arm-x4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001132 "src/f32-spmm/gen/16x1-minmax-wasmsimd-arm.c",
Frank Barchard8ef44cd2020-11-03 12:30:23 -08001133 "src/f32-spmm/gen/16x1-minmax-wasmsimd-x86-pipelined-x2.c",
1134 "src/f32-spmm/gen/16x1-minmax-wasmsimd-x86-pipelined.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07001135 "src/f32-spmm/gen/16x1-minmax-wasmsimd-x86-x2.c",
1136 "src/f32-spmm/gen/16x1-minmax-wasmsimd-x86-x4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001137 "src/f32-spmm/gen/16x1-minmax-wasmsimd-x86.c",
Frank Barchard8ef44cd2020-11-03 12:30:23 -08001138 "src/f32-spmm/gen/32x1-minmax-wasmsimd-arm-pipelined-x2.c",
1139 "src/f32-spmm/gen/32x1-minmax-wasmsimd-arm-pipelined.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07001140 "src/f32-spmm/gen/32x1-minmax-wasmsimd-arm-x2.c",
1141 "src/f32-spmm/gen/32x1-minmax-wasmsimd-arm-x4.c",
Frank Barchard846c0c62020-10-26 15:01:39 -07001142 "src/f32-spmm/gen/32x1-minmax-wasmsimd-arm.c",
Frank Barchard8ef44cd2020-11-03 12:30:23 -08001143 "src/f32-spmm/gen/32x1-minmax-wasmsimd-x86-pipelined-x2.c",
1144 "src/f32-spmm/gen/32x1-minmax-wasmsimd-x86-pipelined.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07001145 "src/f32-spmm/gen/32x1-minmax-wasmsimd-x86-x2.c",
1146 "src/f32-spmm/gen/32x1-minmax-wasmsimd-x86-x4.c",
Frank Barchard846c0c62020-10-26 15:01:39 -07001147 "src/f32-spmm/gen/32x1-minmax-wasmsimd-x86.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -07001148 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-arm-x4.c",
1149 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001150 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-arm-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -07001151 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-x86-x4.c",
1152 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001153 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -07001154 "src/f32-vbinary/gen/vadd-relu-wasmsimd-x4.c",
1155 "src/f32-vbinary/gen/vadd-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001156 "src/f32-vbinary/gen/vadd-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07001157 "src/f32-vbinary/gen/vadd-wasmsimd-x4.c",
1158 "src/f32-vbinary/gen/vadd-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001159 "src/f32-vbinary/gen/vadd-wasmsimd-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001160 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-arm-x4.c",
1161 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001162 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-arm-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001163 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-x86-x4.c",
1164 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001165 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -07001166 "src/f32-vbinary/gen/vaddc-relu-wasmsimd-x4.c",
1167 "src/f32-vbinary/gen/vaddc-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001168 "src/f32-vbinary/gen/vaddc-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07001169 "src/f32-vbinary/gen/vaddc-wasmsimd-x4.c",
1170 "src/f32-vbinary/gen/vaddc-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001171 "src/f32-vbinary/gen/vaddc-wasmsimd-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -07001172 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-arm-x4.c",
1173 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001174 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-arm-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -07001175 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-x86-x4.c",
1176 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001177 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -07001178 "src/f32-vbinary/gen/vdiv-relu-wasmsimd-x4.c",
1179 "src/f32-vbinary/gen/vdiv-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001180 "src/f32-vbinary/gen/vdiv-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07001181 "src/f32-vbinary/gen/vdiv-wasmsimd-x4.c",
1182 "src/f32-vbinary/gen/vdiv-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001183 "src/f32-vbinary/gen/vdiv-wasmsimd-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001184 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-arm-x4.c",
1185 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001186 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-arm-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001187 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-x86-x4.c",
1188 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001189 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -07001190 "src/f32-vbinary/gen/vdivc-relu-wasmsimd-x4.c",
1191 "src/f32-vbinary/gen/vdivc-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001192 "src/f32-vbinary/gen/vdivc-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07001193 "src/f32-vbinary/gen/vdivc-wasmsimd-x4.c",
1194 "src/f32-vbinary/gen/vdivc-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001195 "src/f32-vbinary/gen/vdivc-wasmsimd-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -07001196 "src/f32-vbinary/gen/vmax-wasmsimd-arm-x4.c",
1197 "src/f32-vbinary/gen/vmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001198 "src/f32-vbinary/gen/vmax-wasmsimd-arm-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -07001199 "src/f32-vbinary/gen/vmax-wasmsimd-x86-x4.c",
1200 "src/f32-vbinary/gen/vmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001201 "src/f32-vbinary/gen/vmax-wasmsimd-x86-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001202 "src/f32-vbinary/gen/vmaxc-wasmsimd-arm-x4.c",
1203 "src/f32-vbinary/gen/vmaxc-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001204 "src/f32-vbinary/gen/vmaxc-wasmsimd-arm-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001205 "src/f32-vbinary/gen/vmaxc-wasmsimd-x86-x4.c",
1206 "src/f32-vbinary/gen/vmaxc-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001207 "src/f32-vbinary/gen/vmaxc-wasmsimd-x86-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -07001208 "src/f32-vbinary/gen/vmin-wasmsimd-arm-x4.c",
1209 "src/f32-vbinary/gen/vmin-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001210 "src/f32-vbinary/gen/vmin-wasmsimd-arm-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -07001211 "src/f32-vbinary/gen/vmin-wasmsimd-x86-x4.c",
1212 "src/f32-vbinary/gen/vmin-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001213 "src/f32-vbinary/gen/vmin-wasmsimd-x86-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001214 "src/f32-vbinary/gen/vminc-wasmsimd-arm-x4.c",
1215 "src/f32-vbinary/gen/vminc-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001216 "src/f32-vbinary/gen/vminc-wasmsimd-arm-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001217 "src/f32-vbinary/gen/vminc-wasmsimd-x86-x4.c",
1218 "src/f32-vbinary/gen/vminc-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001219 "src/f32-vbinary/gen/vminc-wasmsimd-x86-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -07001220 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-arm-x4.c",
1221 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001222 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-arm-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -07001223 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-x86-x4.c",
1224 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001225 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -07001226 "src/f32-vbinary/gen/vmul-relu-wasmsimd-x4.c",
1227 "src/f32-vbinary/gen/vmul-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001228 "src/f32-vbinary/gen/vmul-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07001229 "src/f32-vbinary/gen/vmul-wasmsimd-x4.c",
1230 "src/f32-vbinary/gen/vmul-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001231 "src/f32-vbinary/gen/vmul-wasmsimd-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001232 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-arm-x4.c",
1233 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001234 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-arm-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001235 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-x86-x4.c",
1236 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001237 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -07001238 "src/f32-vbinary/gen/vmulc-relu-wasmsimd-x4.c",
1239 "src/f32-vbinary/gen/vmulc-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001240 "src/f32-vbinary/gen/vmulc-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07001241 "src/f32-vbinary/gen/vmulc-wasmsimd-x4.c",
1242 "src/f32-vbinary/gen/vmulc-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001243 "src/f32-vbinary/gen/vmulc-wasmsimd-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001244 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-arm-x4.c",
1245 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001246 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-arm-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001247 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-x86-x4.c",
1248 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001249 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -07001250 "src/f32-vbinary/gen/vrdivc-relu-wasmsimd-x4.c",
1251 "src/f32-vbinary/gen/vrdivc-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001252 "src/f32-vbinary/gen/vrdivc-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07001253 "src/f32-vbinary/gen/vrdivc-wasmsimd-x4.c",
1254 "src/f32-vbinary/gen/vrdivc-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001255 "src/f32-vbinary/gen/vrdivc-wasmsimd-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001256 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-arm-x4.c",
1257 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001258 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-arm-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001259 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-x86-x4.c",
1260 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001261 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -07001262 "src/f32-vbinary/gen/vrsubc-relu-wasmsimd-x4.c",
1263 "src/f32-vbinary/gen/vrsubc-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001264 "src/f32-vbinary/gen/vrsubc-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07001265 "src/f32-vbinary/gen/vrsubc-wasmsimd-x4.c",
1266 "src/f32-vbinary/gen/vrsubc-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001267 "src/f32-vbinary/gen/vrsubc-wasmsimd-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -07001268 "src/f32-vbinary/gen/vsqrdiff-wasmsimd-x4.c",
1269 "src/f32-vbinary/gen/vsqrdiff-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001270 "src/f32-vbinary/gen/vsqrdiff-wasmsimd-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001271 "src/f32-vbinary/gen/vsqrdiffc-wasmsimd-x4.c",
1272 "src/f32-vbinary/gen/vsqrdiffc-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001273 "src/f32-vbinary/gen/vsqrdiffc-wasmsimd-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -07001274 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-arm-x4.c",
1275 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001276 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-arm-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -07001277 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-x86-x4.c",
1278 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001279 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -07001280 "src/f32-vbinary/gen/vsub-relu-wasmsimd-x4.c",
1281 "src/f32-vbinary/gen/vsub-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001282 "src/f32-vbinary/gen/vsub-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07001283 "src/f32-vbinary/gen/vsub-wasmsimd-x4.c",
1284 "src/f32-vbinary/gen/vsub-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001285 "src/f32-vbinary/gen/vsub-wasmsimd-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001286 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-arm-x4.c",
1287 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001288 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-arm-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001289 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-x86-x4.c",
1290 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001291 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -07001292 "src/f32-vbinary/gen/vsubc-relu-wasmsimd-x4.c",
1293 "src/f32-vbinary/gen/vsubc-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001294 "src/f32-vbinary/gen/vsubc-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07001295 "src/f32-vbinary/gen/vsubc-wasmsimd-x4.c",
1296 "src/f32-vbinary/gen/vsubc-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001297 "src/f32-vbinary/gen/vsubc-wasmsimd-x16.c",
Marat Dukhanb1eec082021-05-05 23:24:55 -07001298 "src/f32-vclamp/gen/vclamp-wasmsimd-arm-x4.c",
1299 "src/f32-vclamp/gen/vclamp-wasmsimd-arm-x8.c",
1300 "src/f32-vclamp/gen/vclamp-wasmsimd-x86-x4.c",
1301 "src/f32-vclamp/gen/vclamp-wasmsimd-x86-x8.c",
Marat Dukhaned6baaf2020-12-01 15:07:08 -08001302 "src/f32-velu/gen/velu-wasmsimd-arm-rr2-lut16-p3-x4.c",
1303 "src/f32-velu/gen/velu-wasmsimd-arm-rr2-lut16-p3-x8.c",
1304 "src/f32-velu/gen/velu-wasmsimd-arm-rr2-lut16-p3-x12.c",
1305 "src/f32-velu/gen/velu-wasmsimd-arm-rr2-lut16-p3-x16.c",
1306 "src/f32-velu/gen/velu-wasmsimd-arm-rr2-lut16-p3-x20.c",
1307 "src/f32-velu/gen/velu-wasmsimd-arm-rr2-lut16-p3-x24.c",
Marat Dukhaned6baaf2020-12-01 15:07:08 -08001308 "src/f32-velu/gen/velu-wasmsimd-arm-rr2-p6-x4.c",
1309 "src/f32-velu/gen/velu-wasmsimd-arm-rr2-p6-x8.c",
1310 "src/f32-velu/gen/velu-wasmsimd-arm-rr2-p6-x12.c",
1311 "src/f32-velu/gen/velu-wasmsimd-arm-rr2-p6-x16.c",
1312 "src/f32-velu/gen/velu-wasmsimd-arm-rr2-p6-x20.c",
1313 "src/f32-velu/gen/velu-wasmsimd-arm-rr2-p6-x24.c",
Frank Barcharde7223ee2020-12-04 19:04:01 -08001314 "src/f32-velu/gen/velu-wasmsimd-x86-rr2-lut16-p3-x4.c",
1315 "src/f32-velu/gen/velu-wasmsimd-x86-rr2-lut16-p3-x8.c",
1316 "src/f32-velu/gen/velu-wasmsimd-x86-rr2-lut16-p3-x12.c",
1317 "src/f32-velu/gen/velu-wasmsimd-x86-rr2-lut16-p3-x16.c",
1318 "src/f32-velu/gen/velu-wasmsimd-x86-rr2-lut16-p3-x20.c",
1319 "src/f32-velu/gen/velu-wasmsimd-x86-rr2-lut16-p3-x24.c",
Marat Dukhaned6baaf2020-12-01 15:07:08 -08001320 "src/f32-velu/gen/velu-wasmsimd-x86-rr2-p6-x4.c",
1321 "src/f32-velu/gen/velu-wasmsimd-x86-rr2-p6-x8.c",
1322 "src/f32-velu/gen/velu-wasmsimd-x86-rr2-p6-x12.c",
1323 "src/f32-velu/gen/velu-wasmsimd-x86-rr2-p6-x16.c",
1324 "src/f32-velu/gen/velu-wasmsimd-x86-rr2-p6-x20.c",
1325 "src/f32-velu/gen/velu-wasmsimd-x86-rr2-p6-x24.c",
Marat Dukhan6674d692021-05-05 22:27:00 -07001326 "src/f32-vhswish/gen/vhswish-wasmsimd-x4.c",
1327 "src/f32-vhswish/gen/vhswish-wasmsimd-x8.c",
1328 "src/f32-vhswish/gen/vhswish-wasmsimd-x16.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07001329 "src/f32-vlrelu/gen/vlrelu-wasmsimd-bitselect-x4.c",
1330 "src/f32-vlrelu/gen/vlrelu-wasmsimd-bitselect-x8.c",
1331 "src/f32-vlrelu/gen/vlrelu-wasmsimd-minmax-x4.c",
1332 "src/f32-vlrelu/gen/vlrelu-wasmsimd-minmax-x8.c",
Marat Dukhand816f622020-07-15 10:14:39 -07001333 "src/f32-vmulcaddc/gen/c4-minmax-wasmsimd-arm-2x.c",
Marat Dukhand816f622020-07-15 10:14:39 -07001334 "src/f32-vmulcaddc/gen/c4-minmax-wasmsimd-x86-2x.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001335 "src/f32-vmulcaddc/gen/c8-minmax-wasmsimd-arm-2x.c",
Marat Dukhand816f622020-07-15 10:14:39 -07001336 "src/f32-vmulcaddc/gen/c8-minmax-wasmsimd-x86-2x.c",
Marat Dukhanb1eec082021-05-05 23:24:55 -07001337 "src/f32-vrelu/gen/vrelu-wasmsimd-x4.c",
1338 "src/f32-vrelu/gen/vrelu-wasmsimd-x8.c",
1339 "src/f32-vrelu/gen/vrelu-wasmsimd-x16.c",
Marat Dukhanb82b2cd2020-07-16 02:23:42 -07001340 "src/f32-vrnd/gen/vrndd-wasmsimd-addsub-x4.c",
1341 "src/f32-vrnd/gen/vrndd-wasmsimd-addsub-x8.c",
1342 "src/f32-vrnd/gen/vrndd-wasmsimd-cvt-x4.c",
1343 "src/f32-vrnd/gen/vrndd-wasmsimd-cvt-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001344 "src/f32-vrnd/gen/vrndne-wasmsimd-addsub-x4.c",
1345 "src/f32-vrnd/gen/vrndne-wasmsimd-addsub-x8.c",
1346 "src/f32-vrnd/gen/vrndu-wasmsimd-addsub-x4.c",
1347 "src/f32-vrnd/gen/vrndu-wasmsimd-addsub-x8.c",
1348 "src/f32-vrnd/gen/vrndu-wasmsimd-cvt-x4.c",
1349 "src/f32-vrnd/gen/vrndu-wasmsimd-cvt-x8.c",
1350 "src/f32-vrnd/gen/vrndz-wasmsimd-addsub-x4.c",
1351 "src/f32-vrnd/gen/vrndz-wasmsimd-addsub-x8.c",
1352 "src/f32-vrnd/gen/vrndz-wasmsimd-cvt-x4.c",
1353 "src/f32-vrnd/gen/vrndz-wasmsimd-cvt-x8.c",
Marat Dukhanb1eec082021-05-05 23:24:55 -07001354 "src/f32-vsigmoid/gen/vsigmoid-wasmsimd-lut64-p2-div-x4.c",
1355 "src/f32-vsigmoid/gen/vsigmoid-wasmsimd-lut64-p2-div-x8.c",
1356 "src/f32-vsigmoid/gen/vsigmoid-wasmsimd-lut64-p2-div-x12.c",
1357 "src/f32-vsigmoid/gen/vsigmoid-wasmsimd-lut64-p2-div-x16.c",
1358 "src/f32-vsigmoid/gen/vsigmoid-wasmsimd-lut64-p2-div-x20.c",
1359 "src/f32-vsigmoid/gen/vsigmoid-wasmsimd-lut64-p2-div-x24.c",
1360 "src/f32-vsigmoid/gen/vsigmoid-wasmsimd-p5-div-x4.c",
1361 "src/f32-vsigmoid/gen/vsigmoid-wasmsimd-p5-div-x8.c",
1362 "src/f32-vsigmoid/gen/vsigmoid-wasmsimd-p5-div-x12.c",
1363 "src/f32-vsigmoid/gen/vsigmoid-wasmsimd-p5-div-x16.c",
1364 "src/f32-vsigmoid/gen/vsigmoid-wasmsimd-p5-div-x20.c",
1365 "src/f32-vsigmoid/gen/vsigmoid-wasmsimd-p5-div-x24.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07001366 "src/f32-vsqrt/gen/wasmsimd-sqrt-x4.c",
1367 "src/f32-vsqrt/gen/wasmsimd-sqrt-x8.c",
Marat Dukhan37c83512020-06-29 13:25:53 -07001368 "src/f32-vunary/gen/vabs-wasmsimd-x4.c",
1369 "src/f32-vunary/gen/vabs-wasmsimd-x8.c",
1370 "src/f32-vunary/gen/vneg-wasmsimd-x4.c",
1371 "src/f32-vunary/gen/vneg-wasmsimd-x8.c",
1372 "src/f32-vunary/gen/vsqr-wasmsimd-x4.c",
1373 "src/f32-vunary/gen/vsqr-wasmsimd-x8.c",
Marat Dukhande390d42020-11-29 19:32:18 -08001374 "src/math/expm1minus-wasmsimd-rr2-lut16-p3-andnot.c",
1375 "src/math/expm1minus-wasmsimd-rr2-lut16-p3-max.c",
1376 "src/math/expm1minus-wasmsimd-rr2-p6-andnot.c",
1377 "src/math/expm1minus-wasmsimd-rr2-p6-max.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001378 "src/math/roundd-wasmsimd-addsub.c",
1379 "src/math/roundd-wasmsimd-cvt.c",
1380 "src/math/roundne-wasmsimd-addsub.c",
1381 "src/math/roundu-wasmsimd-addsub.c",
1382 "src/math/roundu-wasmsimd-cvt.c",
1383 "src/math/roundz-wasmsimd-addsub.c",
1384 "src/math/roundz-wasmsimd-cvt.c",
1385 "src/math/sigmoid-wasmsimd-rr2-lut64-p2-div.c",
1386 "src/math/sigmoid-wasmsimd-rr2-p5-div.c",
Marat Dukhancc8f34c2020-08-05 16:36:38 -07001387 "src/qs8-dwconv/gen/up8x9-minmax-wasmsimd-mul16.c",
1388 "src/qs8-dwconv/gen/up16x9-minmax-wasmsimd-mul16.c",
1389 "src/qs8-dwconv/gen/up24x9-minmax-wasmsimd-mul16.c",
Marat Dukhanb5e3d172020-08-06 13:29:53 -07001390 "src/qs8-gavgpool/gen/7p7x-minmax-wasmsimd-c8-acc2.c",
1391 "src/qs8-gavgpool/gen/7p7x-minmax-wasmsimd-c16-acc2.c",
1392 "src/qs8-gavgpool/gen/7p7x-minmax-wasmsimd-c24-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001393 "src/qs8-gavgpool/gen/7x-minmax-wasmsimd-c8-acc2.c",
1394 "src/qs8-gavgpool/gen/7x-minmax-wasmsimd-c16-acc2.c",
1395 "src/qs8-gavgpool/gen/7x-minmax-wasmsimd-c24-acc2.c",
Marat Dukhan27203da2020-08-05 15:19:03 -07001396 "src/qs8-gemm/gen/1x4c8-minmax-wasmsimd-ld64.c",
Marat Dukhan27203da2020-08-05 15:19:03 -07001397 "src/qs8-gemm/gen/1x4c8-minmax-wasmsimd-ld128.c",
Marat Dukhan27203da2020-08-05 15:19:03 -07001398 "src/qs8-gemm/gen/1x4c8-xw-minmax-wasmsimd.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001399 "src/qs8-gemm/gen/2x4c8-minmax-wasmsimd-ld64.c",
1400 "src/qs8-gemm/gen/2x4c8-minmax-wasmsimd-ld128.c",
Marat Dukhan27203da2020-08-05 15:19:03 -07001401 "src/qs8-gemm/gen/2x4c8-xw-minmax-wasmsimd.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001402 "src/qs8-gemm/gen/3x4c8-minmax-wasmsimd-ld64.c",
1403 "src/qs8-gemm/gen/3x4c8-minmax-wasmsimd-ld128.c",
Marat Dukhan27203da2020-08-05 15:19:03 -07001404 "src/qs8-gemm/gen/3x4c8-xw-minmax-wasmsimd.c",
1405 "src/qs8-igemm/gen/1x4c8-minmax-wasmsimd-ld64.c",
Marat Dukhan27203da2020-08-05 15:19:03 -07001406 "src/qs8-igemm/gen/1x4c8-minmax-wasmsimd-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001407 "src/qs8-igemm/gen/2x4c8-minmax-wasmsimd-ld64.c",
Marat Dukhan27203da2020-08-05 15:19:03 -07001408 "src/qs8-igemm/gen/2x4c8-minmax-wasmsimd-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001409 "src/qs8-igemm/gen/3x4c8-minmax-wasmsimd-ld64.c",
Marat Dukhan27203da2020-08-05 15:19:03 -07001410 "src/qs8-igemm/gen/3x4c8-minmax-wasmsimd-ld128.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -07001411 "src/qs8-requantization/fp32-wasmsimd.c",
Marat Dukhan138560c2020-08-03 18:57:34 -07001412 "src/qs8-requantization/q31-wasmsimd.c",
Marat Dukhan5df27f82020-09-02 23:59:21 -07001413 "src/qs8-vadd/gen/minmax-wasmsimd-x8.c",
1414 "src/qs8-vadd/gen/minmax-wasmsimd-x16.c",
1415 "src/qs8-vadd/gen/minmax-wasmsimd-x24.c",
1416 "src/qs8-vadd/gen/minmax-wasmsimd-x32.c",
1417 "src/qs8-vaddc/gen/minmax-wasmsimd-x8.c",
1418 "src/qs8-vaddc/gen/minmax-wasmsimd-x16.c",
1419 "src/qs8-vaddc/gen/minmax-wasmsimd-x24.c",
1420 "src/qs8-vaddc/gen/minmax-wasmsimd-x32.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07001421 "src/qu8-requantization/fp32-wasmsimd.c",
Marat Dukhan22de5e72020-08-03 21:53:18 -07001422 "src/qu8-requantization/q31-wasmsimd.c",
Marat Dukhan8ee37012020-07-16 13:17:13 -07001423 "src/x32-fill/wasmsimd.c",
Marat Dukhan66d99e92020-07-16 12:56:21 -07001424 "src/x32-packx/x4-wasmsimd.c",
Marat Dukhan9306ae02020-07-16 15:51:13 -07001425 "src/x32-pad/wasmsimd.c",
Marat Dukhan9d4bfa22020-07-16 19:07:04 -07001426 "src/x32-unpool/wasmsimd.c",
Marat Dukhane3b78762020-07-16 20:02:58 -07001427 "src/x32-zip/x2-wasmsimd.c",
1428 "src/x32-zip/x3-wasmsimd.c",
1429 "src/x32-zip/x4-wasmsimd.c",
1430 "src/x32-zip/xm-wasmsimd.c",
Marat Dukhan290055c2020-06-09 12:24:29 -07001431]
1432
Marat Dukhan08c4a432019-10-03 09:29:21 -07001433# ISA-specific micro-kernels
1434NEON_UKERNELS = [
Marat Dukhanef25c6d2020-07-24 00:59:40 -07001435 "src/f32-argmaxpool/4x-neon-c4.c",
1436 "src/f32-argmaxpool/9p8x-neon-c4.c",
1437 "src/f32-argmaxpool/9x-neon-c4.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001438 "src/f32-avgpool/9p8x-minmax-neon-c4.c",
1439 "src/f32-avgpool/9x-minmax-neon-c4.c",
Marat Dukhan56b10cd2020-05-18 09:35:49 -07001440 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neon-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001441 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neon-2x2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001442 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neon-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001443 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neon-2x2.c",
Marat Dukhan56b10cd2020-05-18 09:35:49 -07001444 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neon-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001445 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neon-2x2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001446 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neon-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001447 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neon-2x2.c",
Marat Dukhanc7634882020-12-07 15:11:12 -08001448 "src/f32-conv-hwc2chw/3x3s2p1c3x4-neon-2x2.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -07001449 "src/f32-dwconv/gen/up4x4-minmax-neon-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001450 "src/f32-dwconv/gen/up4x4-minmax-neon.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001451 "src/f32-dwconv/gen/up4x9-minmax-neon-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001452 "src/f32-dwconv/gen/up4x9-minmax-neon.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -07001453 "src/f32-dwconv/gen/up4x25-minmax-neon-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001454 "src/f32-dwconv/gen/up4x25-minmax-neon.c",
1455 "src/f32-dwconv/gen/up8x4-minmax-neon-acc2.c",
1456 "src/f32-dwconv/gen/up8x4-minmax-neon.c",
1457 "src/f32-dwconv/gen/up8x9-minmax-neon-acc2.c",
1458 "src/f32-dwconv/gen/up8x9-minmax-neon.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -07001459 "src/f32-dwconv/gen/up8x25-minmax-neon-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001460 "src/f32-dwconv/gen/up8x25-minmax-neon.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001461 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-1x4-acc2.c",
1462 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-1x4-acc3.c",
1463 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-1x4-acc4.c",
Marat Dukhanc581e482020-10-24 01:28:11 -07001464 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-1x4.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001465 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-2x4-acc2.c",
Marat Dukhanc581e482020-10-24 01:28:11 -07001466 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-2x4.c",
1467 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-3x4.c",
1468 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-4x4.c",
1469 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-5x4.c",
1470 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-6x4.c",
Marat Dukhan82f0c322020-10-25 19:17:35 -07001471 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neon-1x4-acc2.c",
1472 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neon-1x4-acc3.c",
1473 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neon-1x4-acc4.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001474 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neon-1x4.c",
Marat Dukhan82f0c322020-10-25 19:17:35 -07001475 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neon-2x4-acc2.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001476 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neon-2x4.c",
1477 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neon-3x4.c",
1478 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neon-4x4.c",
Marat Dukhan149f0ea2020-10-26 12:50:33 -07001479 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-1x4-acc2.c",
1480 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-1x4-acc3.c",
1481 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-1x4-acc4.c",
1482 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-1x4-acc5.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001483 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-1x4.c",
Marat Dukhan149f0ea2020-10-26 12:50:33 -07001484 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-2x4-acc2.c",
1485 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-2x4-acc3.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001486 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-2x4.c",
Marat Dukhan149f0ea2020-10-26 12:50:33 -07001487 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-3x4-acc2.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001488 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-3x4.c",
Marat Dukhan149f0ea2020-10-26 12:50:33 -07001489 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-4x4-acc2.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001490 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-4x4.c",
1491 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-5x4.c",
Marat Dukhan30d4b252020-10-29 16:33:22 -07001492 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neon-1x4-acc2.c",
1493 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neon-1x4-acc3.c",
1494 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neon-1x4-acc4.c",
1495 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neon-1x4-acc5.c",
1496 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neon-1x4.c",
1497 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neon-2x4-acc2.c",
1498 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neon-2x4-acc3.c",
1499 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neon-2x4.c",
Marat Dukhan30d4b252020-10-29 16:33:22 -07001500 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neon-3x4-acc2.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07001501 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neon-3x4.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -07001502 "src/f32-gavgpool-cw/neon-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001503 "src/f32-gavgpool/7p7x-minmax-neon-c4.c",
1504 "src/f32-gavgpool/7x-minmax-neon-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001505 "src/f32-gemm/gen-inc/1x8inc-minmax-neon-dup-ld64.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001506 "src/f32-gemm/gen-inc/1x8inc-minmax-neon-lane-ld64.c",
1507 "src/f32-gemm/gen-inc/1x8s4inc-minmax-neon.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001508 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-dup-ld64.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001509 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-dup-ld128.c",
1510 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-lane-ld64.c",
1511 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-lane-ld128.c",
1512 "src/f32-gemm/gen-inc/4x8s4inc-minmax-neon.c",
1513 "src/f32-gemm/gen-inc/5x8inc-minmax-neon-lane-ld64.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001514 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-dup-ld64.c",
1515 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-dup-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001516 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-lane-ld64.c",
1517 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-lane-ld128.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001518 "src/f32-gemm/gen-inc/6x8s4inc-minmax-neon.c",
1519 "src/f32-gemm/gen-inc/8x8s4inc-minmax-neon.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001520 "src/f32-gemm/gen/1x8-minmax-neon-dup-ld64.c",
1521 "src/f32-gemm/gen/1x8-minmax-neon-lane-ld64.c",
1522 "src/f32-gemm/gen/1x8s4-minmax-neon.c",
1523 "src/f32-gemm/gen/4x2-minmax-neon-lane-ld64.c",
1524 "src/f32-gemm/gen/4x8-minmax-neon-dup-ld64.c",
1525 "src/f32-gemm/gen/4x8-minmax-neon-dup-ld128.c",
1526 "src/f32-gemm/gen/4x8-minmax-neon-lane-ld64.c",
1527 "src/f32-gemm/gen/4x8-minmax-neon-lane-ld128.c",
1528 "src/f32-gemm/gen/4x8s4-minmax-neon.c",
1529 "src/f32-gemm/gen/5x8-minmax-neon-lane-ld64.c",
1530 "src/f32-gemm/gen/6x8-minmax-neon-dup-ld64.c",
1531 "src/f32-gemm/gen/6x8-minmax-neon-dup-ld128.c",
1532 "src/f32-gemm/gen/6x8-minmax-neon-lane-ld64.c",
1533 "src/f32-gemm/gen/6x8-minmax-neon-lane-ld128.c",
1534 "src/f32-gemm/gen/6x8s4-minmax-neon.c",
1535 "src/f32-gemm/gen/8x8s4-minmax-neon.c",
Artsiom Ablavatski2202c812021-01-22 14:16:43 -08001536 "src/f32-ibilinear-chw/gen/neon-p4.c",
1537 "src/f32-ibilinear-chw/gen/neon-p8.c",
Frank Barchard8247e212021-02-03 18:12:33 -08001538 "src/f32-ibilinear/gen/neon-c4.c",
1539 "src/f32-ibilinear/gen/neon-c8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001540 "src/f32-igemm/gen/1x8-minmax-neon-dup-ld64.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001541 "src/f32-igemm/gen/1x8-minmax-neon-lane-ld64.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001542 "src/f32-igemm/gen/1x8s4-minmax-neon.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001543 "src/f32-igemm/gen/4x2-minmax-neon-lane-ld64.c",
1544 "src/f32-igemm/gen/4x4-minmax-neon-lane-ld64.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001545 "src/f32-igemm/gen/4x8-minmax-neon-dup-ld64.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001546 "src/f32-igemm/gen/4x8-minmax-neon-dup-ld128.c",
1547 "src/f32-igemm/gen/4x8-minmax-neon-lane-ld64.c",
1548 "src/f32-igemm/gen/4x8-minmax-neon-lane-ld128.c",
1549 "src/f32-igemm/gen/4x8s4-minmax-neon.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001550 "src/f32-igemm/gen/6x8-minmax-neon-dup-ld64.c",
1551 "src/f32-igemm/gen/6x8-minmax-neon-dup-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001552 "src/f32-igemm/gen/6x8-minmax-neon-lane-ld64.c",
1553 "src/f32-igemm/gen/6x8-minmax-neon-lane-ld128.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001554 "src/f32-igemm/gen/6x8s4-minmax-neon.c",
1555 "src/f32-igemm/gen/8x8s4-minmax-neon.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001556 "src/f32-maxpool/9p8x-minmax-neon-c4.c",
1557 "src/f32-pavgpool/9p8x-minmax-neon-c4.c",
1558 "src/f32-pavgpool/9x-minmax-neon-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001559 "src/f32-ppmm/gen/4x8-minmax-neon.c",
1560 "src/f32-ppmm/gen/8x8-minmax-neon.c",
Frank Barcharda5316982020-07-23 13:19:28 -07001561 "src/f32-prelu/gen/neon-1x4.c",
1562 "src/f32-prelu/gen/neon-1x8.c",
1563 "src/f32-prelu/gen/neon-1x16.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001564 "src/f32-prelu/gen/neon-2x4.c",
1565 "src/f32-prelu/gen/neon-2x8.c",
Frank Barcharda5316982020-07-23 13:19:28 -07001566 "src/f32-prelu/gen/neon-2x16.c",
1567 "src/f32-prelu/gen/neon-4x4.c",
1568 "src/f32-prelu/gen/neon-4x8.c",
1569 "src/f32-prelu/gen/neon-4x16.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001570 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x4.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001571 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x8-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001572 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x8.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001573 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12-acc2.c",
1574 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001575 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001576 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16-acc2.c",
1577 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001578 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001579 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20-acc2.c",
1580 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001581 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20.c",
1582 "src/f32-raddstoreexpminusmax/gen/neon-p5-x4.c",
1583 "src/f32-raddstoreexpminusmax/gen/neon-p5-x8-acc2.c",
1584 "src/f32-raddstoreexpminusmax/gen/neon-p5-x8.c",
1585 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12-acc2.c",
1586 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12-acc3.c",
1587 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12.c",
1588 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16-acc2.c",
1589 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16-acc4.c",
1590 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16.c",
1591 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20-acc2.c",
1592 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20-acc5.c",
1593 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001594 "src/f32-rmax/neon.c",
Marat Dukhan5b86c432020-12-06 19:15:03 -08001595 "src/f32-spmm/gen/4x1-minmax-neon-pipelined.c",
1596 "src/f32-spmm/gen/4x1-minmax-neon-x2.c",
1597 "src/f32-spmm/gen/4x1-minmax-neon.c",
1598 "src/f32-spmm/gen/8x1-minmax-neon-pipelined.c",
1599 "src/f32-spmm/gen/8x1-minmax-neon-x2.c",
1600 "src/f32-spmm/gen/8x1-minmax-neon.c",
1601 "src/f32-spmm/gen/12x1-minmax-neon.c",
1602 "src/f32-spmm/gen/16x1-minmax-neon-pipelined.c",
1603 "src/f32-spmm/gen/16x1-minmax-neon-x2.c",
1604 "src/f32-spmm/gen/16x1-minmax-neon.c",
1605 "src/f32-spmm/gen/32x1-minmax-neon-pipelined.c",
1606 "src/f32-spmm/gen/32x1-minmax-neon-x2.c",
1607 "src/f32-spmm/gen/32x1-minmax-neon.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001608 "src/f32-vbinary/gen/vadd-minmax-neon-x4.c",
1609 "src/f32-vbinary/gen/vadd-minmax-neon-x8.c",
1610 "src/f32-vbinary/gen/vaddc-minmax-neon-x4.c",
1611 "src/f32-vbinary/gen/vaddc-minmax-neon-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -08001612 "src/f32-vbinary/gen/vmax-neon-x4.c",
1613 "src/f32-vbinary/gen/vmax-neon-x8.c",
1614 "src/f32-vbinary/gen/vmaxc-neon-x4.c",
1615 "src/f32-vbinary/gen/vmaxc-neon-x8.c",
1616 "src/f32-vbinary/gen/vmin-neon-x4.c",
1617 "src/f32-vbinary/gen/vmin-neon-x8.c",
1618 "src/f32-vbinary/gen/vminc-neon-x4.c",
1619 "src/f32-vbinary/gen/vminc-neon-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001620 "src/f32-vbinary/gen/vmul-minmax-neon-x4.c",
1621 "src/f32-vbinary/gen/vmul-minmax-neon-x8.c",
1622 "src/f32-vbinary/gen/vmulc-minmax-neon-x4.c",
1623 "src/f32-vbinary/gen/vmulc-minmax-neon-x8.c",
1624 "src/f32-vbinary/gen/vrsubc-minmax-neon-x4.c",
1625 "src/f32-vbinary/gen/vrsubc-minmax-neon-x8.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07001626 "src/f32-vbinary/gen/vsqrdiff-neon-x4.c",
1627 "src/f32-vbinary/gen/vsqrdiff-neon-x8.c",
1628 "src/f32-vbinary/gen/vsqrdiffc-neon-x4.c",
1629 "src/f32-vbinary/gen/vsqrdiffc-neon-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001630 "src/f32-vbinary/gen/vsub-minmax-neon-x4.c",
1631 "src/f32-vbinary/gen/vsub-minmax-neon-x8.c",
1632 "src/f32-vbinary/gen/vsubc-minmax-neon-x4.c",
1633 "src/f32-vbinary/gen/vsubc-minmax-neon-x8.c",
Marat Dukhanb1eec082021-05-05 23:24:55 -07001634 "src/f32-vclamp/gen/vclamp-neon-x4.c",
1635 "src/f32-vclamp/gen/vclamp-neon-x8.c",
Marat Dukhaned6baaf2020-12-01 15:07:08 -08001636 "src/f32-velu/gen/velu-neon-rr2-lut16-p3-x4.c",
1637 "src/f32-velu/gen/velu-neon-rr2-lut16-p3-x8.c",
1638 "src/f32-velu/gen/velu-neon-rr2-lut16-p3-x12.c",
1639 "src/f32-velu/gen/velu-neon-rr2-lut16-p3-x16.c",
1640 "src/f32-velu/gen/velu-neon-rr2-lut16-p3-x20.c",
1641 "src/f32-velu/gen/velu-neon-rr2-lut16-p3-x24.c",
1642 "src/f32-velu/gen/velu-neon-rr2-p6-x4.c",
1643 "src/f32-velu/gen/velu-neon-rr2-p6-x8.c",
1644 "src/f32-velu/gen/velu-neon-rr2-p6-x12.c",
1645 "src/f32-velu/gen/velu-neon-rr2-p6-x16.c",
1646 "src/f32-velu/gen/velu-neon-rr2-p6-x20.c",
1647 "src/f32-velu/gen/velu-neon-rr2-p6-x24.c",
Marat Dukhan6674d692021-05-05 22:27:00 -07001648 "src/f32-vhswish/gen/vhswish-neon-x4.c",
1649 "src/f32-vhswish/gen/vhswish-neon-x8.c",
1650 "src/f32-vhswish/gen/vhswish-neon-x16.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07001651 "src/f32-vlrelu/gen/vlrelu-neon-x4.c",
1652 "src/f32-vlrelu/gen/vlrelu-neon-x8.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001653 "src/f32-vmulcaddc/gen/c4-minmax-neon-2x.c",
1654 "src/f32-vmulcaddc/gen/c8-minmax-neon-2x.c",
Marat Dukhanb1eec082021-05-05 23:24:55 -07001655 "src/f32-vrelu/gen/vrelu-neon-x4.c",
1656 "src/f32-vrelu/gen/vrelu-neon-x8.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07001657 "src/f32-vrnd/gen/vrndd-neon-x4.c",
1658 "src/f32-vrnd/gen/vrndd-neon-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001659 "src/f32-vrnd/gen/vrndne-neon-x4.c",
1660 "src/f32-vrnd/gen/vrndne-neon-x8.c",
1661 "src/f32-vrnd/gen/vrndu-neon-x4.c",
1662 "src/f32-vrnd/gen/vrndu-neon-x8.c",
1663 "src/f32-vrnd/gen/vrndz-neon-x4.c",
1664 "src/f32-vrnd/gen/vrndz-neon-x8.c",
Marat Dukhanb1eec082021-05-05 23:24:55 -07001665 "src/f32-vsigmoid/gen/vsigmoid-neon-rr2-lut64-p2-nr2recps-x4.c",
1666 "src/f32-vsigmoid/gen/vsigmoid-neon-rr2-lut64-p2-nr2recps-x8.c",
1667 "src/f32-vsigmoid/gen/vsigmoid-neon-rr2-lut64-p2-nr2recps-x12.c",
1668 "src/f32-vsigmoid/gen/vsigmoid-neon-rr2-lut64-p2-nr2recps-x16.c",
1669 "src/f32-vsigmoid/gen/vsigmoid-neon-rr2-lut64-p2-nr2recps-x20.c",
1670 "src/f32-vsigmoid/gen/vsigmoid-neon-rr2-lut64-p2-nr2recps-x24.c",
1671 "src/f32-vsigmoid/gen/vsigmoid-neon-rr2-lut2048-p1-nr2recps-x4.c",
1672 "src/f32-vsigmoid/gen/vsigmoid-neon-rr2-lut2048-p1-nr2recps-x8.c",
1673 "src/f32-vsigmoid/gen/vsigmoid-neon-rr2-lut2048-p1-nr2recps-x12.c",
1674 "src/f32-vsigmoid/gen/vsigmoid-neon-rr2-lut2048-p1-nr2recps-x16.c",
1675 "src/f32-vsigmoid/gen/vsigmoid-neon-rr2-lut2048-p1-nr2recps-x20.c",
1676 "src/f32-vsigmoid/gen/vsigmoid-neon-rr2-lut2048-p1-nr2recps-x24.c",
1677 "src/f32-vsigmoid/gen/vsigmoid-neon-rr2-p5-nr2recps-x4.c",
1678 "src/f32-vsigmoid/gen/vsigmoid-neon-rr2-p5-nr2recps-x8.c",
1679 "src/f32-vsigmoid/gen/vsigmoid-neon-rr2-p5-nr2recps-x12.c",
1680 "src/f32-vsigmoid/gen/vsigmoid-neon-rr2-p5-nr2recps-x16.c",
1681 "src/f32-vsigmoid/gen/vsigmoid-neon-rr2-p5-nr2recps-x20.c",
1682 "src/f32-vsigmoid/gen/vsigmoid-neon-rr2-p5-nr2recps-x24.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07001683 "src/f32-vunary/gen/vabs-neon-x4.c",
1684 "src/f32-vunary/gen/vabs-neon-x8.c",
1685 "src/f32-vunary/gen/vneg-neon-x4.c",
1686 "src/f32-vunary/gen/vneg-neon-x8.c",
1687 "src/f32-vunary/gen/vsqr-neon-x4.c",
1688 "src/f32-vunary/gen/vsqr-neon-x8.c",
Marat Dukhande390d42020-11-29 19:32:18 -08001689 "src/math/expm1minus-neon-rr2-lut16-p3.c",
1690 "src/math/expm1minus-neon-rr2-p6.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001691 "src/math/roundd-neon-addsub.c",
1692 "src/math/roundd-neon-cvt.c",
1693 "src/math/roundne-neon-addsub.c",
1694 "src/math/roundu-neon-addsub.c",
1695 "src/math/roundu-neon-cvt.c",
1696 "src/math/roundz-neon-addsub.c",
1697 "src/math/roundz-neon-cvt.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001698 "src/math/sigmoid-neon-rr2-lut64-p2-nr2recps.c",
1699 "src/math/sigmoid-neon-rr2-lut2048-p1-nr2recps.c",
1700 "src/math/sigmoid-neon-rr2-p5-nr2recps.c",
1701 "src/math/sqrt-neon-nr1rsqrts.c",
1702 "src/math/sqrt-neon-nr2rsqrts.c",
1703 "src/math/sqrt-neon-nr3rsqrts.c",
Marat Dukhan023bcf92020-08-10 12:40:50 -07001704 "src/qs8-dwconv/gen/up8x9-minmax-neon-mul16.c",
1705 "src/qs8-dwconv/gen/up16x9-minmax-neon-mul16.c",
1706 "src/qs8-dwconv/gen/up24x9-minmax-neon-mul16.c",
1707 "src/qs8-dwconv/gen/up32x9-minmax-neon-mul16.c",
Marat Dukhan281262d2020-08-10 13:23:21 -07001708 "src/qs8-gavgpool/gen/7p7x-minmax-neon-c8-acc2.c",
1709 "src/qs8-gavgpool/gen/7p7x-minmax-neon-c16-acc2.c",
1710 "src/qs8-gavgpool/gen/7p7x-minmax-neon-c24-acc2.c",
1711 "src/qs8-gavgpool/gen/7p7x-minmax-neon-c32-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001712 "src/qs8-gavgpool/gen/7x-minmax-neon-c8-acc2.c",
1713 "src/qs8-gavgpool/gen/7x-minmax-neon-c16-acc2.c",
1714 "src/qs8-gavgpool/gen/7x-minmax-neon-c24-acc2.c",
1715 "src/qs8-gavgpool/gen/7x-minmax-neon-c32-acc2.c",
Frank Barchard1f51d382021-04-20 14:07:14 -07001716 "src/qs8-gemm/gen/1x8-minmax-neon-mlal-lane-prfm.c",
Frank Barchardec0bf142021-01-15 10:14:05 -08001717 "src/qs8-gemm/gen/1x8-minmax-neon-mlal-lane.c",
Frank Barchardec0bf142021-01-15 10:14:05 -08001718 "src/qs8-gemm/gen/1x8-minmax-neon-mull-addw-dup.c",
Frank Barchard8247e212021-02-03 18:12:33 -08001719 "src/qs8-gemm/gen/1x8c2-minmax-neon-mlal-padal-dup.c",
Frank Barchard2302ffd2021-01-22 14:08:02 -08001720 "src/qs8-gemm/gen/1x8c2-minmax-neon-mull-padal-dup.c",
Frank Barchardda78da12021-03-02 14:28:00 -08001721 "src/qs8-gemm/gen/1x8c8-minmax-neon-mlal-padal.c",
Frank Barchard8247e212021-02-03 18:12:33 -08001722 "src/qs8-gemm/gen/1x8c8-minmax-neon-mull-padal.c",
Frank Barchard671d1b02021-03-10 15:42:08 -08001723 "src/qs8-gemm/gen/1x8c16-minmax-neon-mlal-padal.c",
Frank Barchard1f51d382021-04-20 14:07:14 -07001724 "src/qs8-gemm/gen/1x16-minmax-neon-mlal-lane-prfm.c",
Frank Barchardcadd4222021-01-20 16:27:25 -08001725 "src/qs8-gemm/gen/1x16-minmax-neon-mlal-lane.c",
1726 "src/qs8-gemm/gen/1x16-minmax-neon-mull-addw-dup.c",
Frank Barchard8247e212021-02-03 18:12:33 -08001727 "src/qs8-gemm/gen/1x16c2-minmax-neon-mlal-padal-dup.c",
Frank Barchard2302ffd2021-01-22 14:08:02 -08001728 "src/qs8-gemm/gen/1x16c2-minmax-neon-mull-padal-dup.c",
Frank Barchardda78da12021-03-02 14:28:00 -08001729 "src/qs8-gemm/gen/1x16c8-minmax-neon-mlal-padal.c",
Frank Barchard8247e212021-02-03 18:12:33 -08001730 "src/qs8-gemm/gen/1x16c8-minmax-neon-mull-padal.c",
Frank Barchard671d1b02021-03-10 15:42:08 -08001731 "src/qs8-gemm/gen/1x16c16-minmax-neon-mlal-padal.c",
Frank Barchard1f51d382021-04-20 14:07:14 -07001732 "src/qs8-gemm/gen/2x8-minmax-neon-mlal-lane-prfm.c",
Frank Barchardcadd4222021-01-20 16:27:25 -08001733 "src/qs8-gemm/gen/2x8-minmax-neon-mlal-lane.c",
Frank Barchardec0bf142021-01-15 10:14:05 -08001734 "src/qs8-gemm/gen/2x8-minmax-neon-mull-addw-dup.c",
Frank Barchard8247e212021-02-03 18:12:33 -08001735 "src/qs8-gemm/gen/2x8c2-minmax-neon-mlal-padal-dup.c",
Frank Barchard2302ffd2021-01-22 14:08:02 -08001736 "src/qs8-gemm/gen/2x8c2-minmax-neon-mull-padal-dup.c",
Frank Barchardda78da12021-03-02 14:28:00 -08001737 "src/qs8-gemm/gen/2x8c8-minmax-neon-mlal-padal.c",
Frank Barchard8247e212021-02-03 18:12:33 -08001738 "src/qs8-gemm/gen/2x8c8-minmax-neon-mull-padal.c",
Frank Barchard671d1b02021-03-10 15:42:08 -08001739 "src/qs8-gemm/gen/2x8c16-minmax-neon-mlal-padal.c",
Frank Barchard1f51d382021-04-20 14:07:14 -07001740 "src/qs8-gemm/gen/2x16-minmax-neon-mlal-lane-prfm.c",
Frank Barchardcadd4222021-01-20 16:27:25 -08001741 "src/qs8-gemm/gen/2x16-minmax-neon-mlal-lane.c",
1742 "src/qs8-gemm/gen/2x16-minmax-neon-mull-addw-dup.c",
Frank Barchard8247e212021-02-03 18:12:33 -08001743 "src/qs8-gemm/gen/2x16c2-minmax-neon-mlal-padal-dup.c",
Frank Barchard2302ffd2021-01-22 14:08:02 -08001744 "src/qs8-gemm/gen/2x16c2-minmax-neon-mull-padal-dup.c",
Frank Barchardda78da12021-03-02 14:28:00 -08001745 "src/qs8-gemm/gen/2x16c8-minmax-neon-mlal-padal.c",
Frank Barchard8247e212021-02-03 18:12:33 -08001746 "src/qs8-gemm/gen/2x16c8-minmax-neon-mull-padal.c",
Frank Barchard671d1b02021-03-10 15:42:08 -08001747 "src/qs8-gemm/gen/2x16c16-minmax-neon-mlal-padal.c",
Frank Barchard1f51d382021-04-20 14:07:14 -07001748 "src/qs8-gemm/gen/3x8-minmax-neon-mlal-lane-prfm.c",
Frank Barchard2302ffd2021-01-22 14:08:02 -08001749 "src/qs8-gemm/gen/3x8-minmax-neon-mlal-lane.c",
1750 "src/qs8-gemm/gen/3x8-minmax-neon-mull-addw-dup.c",
Frank Barchard8247e212021-02-03 18:12:33 -08001751 "src/qs8-gemm/gen/3x8c2-minmax-neon-mlal-padal-dup.c",
Frank Barchard2302ffd2021-01-22 14:08:02 -08001752 "src/qs8-gemm/gen/3x8c2-minmax-neon-mull-padal-dup.c",
Frank Barchardda78da12021-03-02 14:28:00 -08001753 "src/qs8-gemm/gen/3x8c8-minmax-neon-mlal-padal.c",
Frank Barchard8247e212021-02-03 18:12:33 -08001754 "src/qs8-gemm/gen/3x8c8-minmax-neon-mull-padal.c",
Frank Barchard671d1b02021-03-10 15:42:08 -08001755 "src/qs8-gemm/gen/3x8c16-minmax-neon-mlal-padal.c",
Frank Barchard1f51d382021-04-20 14:07:14 -07001756 "src/qs8-gemm/gen/3x16-minmax-neon-mlal-lane-prfm.c",
Frank Barchard8247e212021-02-03 18:12:33 -08001757 "src/qs8-gemm/gen/3x16-minmax-neon-mlal-lane.c",
1758 "src/qs8-gemm/gen/3x16-minmax-neon-mull-addw-dup.c",
1759 "src/qs8-gemm/gen/3x16c2-minmax-neon-mlal-padal-dup.c",
1760 "src/qs8-gemm/gen/3x16c2-minmax-neon-mull-padal-dup.c",
Frank Barchardda78da12021-03-02 14:28:00 -08001761 "src/qs8-gemm/gen/3x16c8-minmax-neon-mlal-padal.c",
Frank Barchard8247e212021-02-03 18:12:33 -08001762 "src/qs8-gemm/gen/3x16c8-minmax-neon-mull-padal.c",
Frank Barchard671d1b02021-03-10 15:42:08 -08001763 "src/qs8-gemm/gen/3x16c16-minmax-neon-mlal-padal.c",
Frank Barchard1f51d382021-04-20 14:07:14 -07001764 "src/qs8-gemm/gen/4x8-minmax-neon-mlal-lane-prfm.c",
Frank Barchardcadd4222021-01-20 16:27:25 -08001765 "src/qs8-gemm/gen/4x8-minmax-neon-mlal-lane.c",
Frank Barchardec0bf142021-01-15 10:14:05 -08001766 "src/qs8-gemm/gen/4x8-minmax-neon-mull-addw-dup.c",
Frank Barchard8247e212021-02-03 18:12:33 -08001767 "src/qs8-gemm/gen/4x8c2-minmax-neon-mlal-padal-dup.c",
Frank Barchard2302ffd2021-01-22 14:08:02 -08001768 "src/qs8-gemm/gen/4x8c2-minmax-neon-mull-padal-dup.c",
Frank Barchardda78da12021-03-02 14:28:00 -08001769 "src/qs8-gemm/gen/4x8c8-minmax-neon-mlal-padal.c",
Frank Barchard8247e212021-02-03 18:12:33 -08001770 "src/qs8-gemm/gen/4x8c8-minmax-neon-mull-padal.c",
Frank Barchard671d1b02021-03-10 15:42:08 -08001771 "src/qs8-gemm/gen/4x8c16-minmax-neon-mlal-padal.c",
Frank Barchard1f51d382021-04-20 14:07:14 -07001772 "src/qs8-gemm/gen/4x16-minmax-neon-mlal-lane-prfm.c",
Frank Barchardcadd4222021-01-20 16:27:25 -08001773 "src/qs8-gemm/gen/4x16-minmax-neon-mlal-lane.c",
1774 "src/qs8-gemm/gen/4x16-minmax-neon-mull-addw-dup.c",
Frank Barchard8247e212021-02-03 18:12:33 -08001775 "src/qs8-gemm/gen/4x16c2-minmax-neon-mlal-padal-dup.c",
Frank Barchard2302ffd2021-01-22 14:08:02 -08001776 "src/qs8-gemm/gen/4x16c2-minmax-neon-mull-padal-dup.c",
Frank Barchardda78da12021-03-02 14:28:00 -08001777 "src/qs8-gemm/gen/4x16c8-minmax-neon-mlal-padal.c",
Frank Barchard58990122021-01-29 16:09:56 -08001778 "src/qs8-gemm/gen/4x16c8-minmax-neon-mull-padal.c",
Frank Barchard71c4d1a2021-02-19 10:07:23 -08001779 "src/qs8-gemm/gen/4x16c16-minmax-neon-mlal-padal.c",
Frank Barcharddddb38f2021-05-03 16:51:48 -07001780 "src/qs8-gemm/gen/6x8-minmax-neon-mlal-lane-prfm.c",
1781 "src/qs8-gemm/gen/6x8-minmax-neon-mlal-lane.c",
1782 "src/qs8-gemm/gen/6x16-minmax-neon-mlal-lane-prfm.c",
1783 "src/qs8-gemm/gen/6x16-minmax-neon-mlal-lane.c",
Frank Barchard1f51d382021-04-20 14:07:14 -07001784 "src/qs8-igemm/gen/1x8-minmax-neon-mlal-lane-prfm.c",
Frank Barchardec0bf142021-01-15 10:14:05 -08001785 "src/qs8-igemm/gen/1x8-minmax-neon-mlal-lane.c",
Frank Barchard671d1b02021-03-10 15:42:08 -08001786 "src/qs8-igemm/gen/1x8-minmax-neon-mull-addw-dup.c",
1787 "src/qs8-igemm/gen/1x8c2-minmax-neon-mlal-padal-dup.c",
1788 "src/qs8-igemm/gen/1x8c2-minmax-neon-mull-padal-dup.c",
Frank Barchardda78da12021-03-02 14:28:00 -08001789 "src/qs8-igemm/gen/1x8c8-minmax-neon-mlal-padal.c",
1790 "src/qs8-igemm/gen/1x8c8-minmax-neon-mull-padal.c",
Frank Barchard671d1b02021-03-10 15:42:08 -08001791 "src/qs8-igemm/gen/1x8c16-minmax-neon-mlal-padal.c",
Frank Barchard1f51d382021-04-20 14:07:14 -07001792 "src/qs8-igemm/gen/1x16-minmax-neon-mlal-lane-prfm.c",
Frank Barchard671d1b02021-03-10 15:42:08 -08001793 "src/qs8-igemm/gen/1x16-minmax-neon-mlal-lane.c",
1794 "src/qs8-igemm/gen/1x16-minmax-neon-mull-addw-dup.c",
1795 "src/qs8-igemm/gen/1x16c2-minmax-neon-mlal-padal-dup.c",
1796 "src/qs8-igemm/gen/1x16c2-minmax-neon-mull-padal-dup.c",
1797 "src/qs8-igemm/gen/1x16c8-minmax-neon-mlal-padal.c",
1798 "src/qs8-igemm/gen/1x16c8-minmax-neon-mull-padal.c",
1799 "src/qs8-igemm/gen/1x16c16-minmax-neon-mlal-padal.c",
Frank Barchard1f51d382021-04-20 14:07:14 -07001800 "src/qs8-igemm/gen/2x8-minmax-neon-mlal-lane-prfm.c",
Frank Barchard671d1b02021-03-10 15:42:08 -08001801 "src/qs8-igemm/gen/2x8-minmax-neon-mlal-lane.c",
1802 "src/qs8-igemm/gen/2x8-minmax-neon-mull-addw-dup.c",
1803 "src/qs8-igemm/gen/2x8c2-minmax-neon-mlal-padal-dup.c",
1804 "src/qs8-igemm/gen/2x8c2-minmax-neon-mull-padal-dup.c",
Frank Barchardda78da12021-03-02 14:28:00 -08001805 "src/qs8-igemm/gen/2x8c8-minmax-neon-mlal-padal.c",
1806 "src/qs8-igemm/gen/2x8c8-minmax-neon-mull-padal.c",
Frank Barchard671d1b02021-03-10 15:42:08 -08001807 "src/qs8-igemm/gen/2x8c16-minmax-neon-mlal-padal.c",
Frank Barchard1f51d382021-04-20 14:07:14 -07001808 "src/qs8-igemm/gen/2x16-minmax-neon-mlal-lane-prfm.c",
Frank Barchard671d1b02021-03-10 15:42:08 -08001809 "src/qs8-igemm/gen/2x16-minmax-neon-mlal-lane.c",
1810 "src/qs8-igemm/gen/2x16-minmax-neon-mull-addw-dup.c",
1811 "src/qs8-igemm/gen/2x16c2-minmax-neon-mlal-padal-dup.c",
1812 "src/qs8-igemm/gen/2x16c2-minmax-neon-mull-padal-dup.c",
1813 "src/qs8-igemm/gen/2x16c8-minmax-neon-mlal-padal.c",
1814 "src/qs8-igemm/gen/2x16c8-minmax-neon-mull-padal.c",
1815 "src/qs8-igemm/gen/2x16c16-minmax-neon-mlal-padal.c",
Frank Barchard1f51d382021-04-20 14:07:14 -07001816 "src/qs8-igemm/gen/3x8-minmax-neon-mlal-lane-prfm.c",
Frank Barchard671d1b02021-03-10 15:42:08 -08001817 "src/qs8-igemm/gen/3x8-minmax-neon-mlal-lane.c",
1818 "src/qs8-igemm/gen/3x8-minmax-neon-mull-addw-dup.c",
1819 "src/qs8-igemm/gen/3x8c2-minmax-neon-mlal-padal-dup.c",
1820 "src/qs8-igemm/gen/3x8c2-minmax-neon-mull-padal-dup.c",
Frank Barchardda78da12021-03-02 14:28:00 -08001821 "src/qs8-igemm/gen/3x8c8-minmax-neon-mlal-padal.c",
1822 "src/qs8-igemm/gen/3x8c8-minmax-neon-mull-padal.c",
Frank Barchard671d1b02021-03-10 15:42:08 -08001823 "src/qs8-igemm/gen/3x8c16-minmax-neon-mlal-padal.c",
Frank Barchard1f51d382021-04-20 14:07:14 -07001824 "src/qs8-igemm/gen/3x16-minmax-neon-mlal-lane-prfm.c",
Frank Barchard671d1b02021-03-10 15:42:08 -08001825 "src/qs8-igemm/gen/3x16-minmax-neon-mlal-lane.c",
1826 "src/qs8-igemm/gen/3x16-minmax-neon-mull-addw-dup.c",
1827 "src/qs8-igemm/gen/3x16c2-minmax-neon-mlal-padal-dup.c",
1828 "src/qs8-igemm/gen/3x16c2-minmax-neon-mull-padal-dup.c",
1829 "src/qs8-igemm/gen/3x16c8-minmax-neon-mlal-padal.c",
1830 "src/qs8-igemm/gen/3x16c8-minmax-neon-mull-padal.c",
1831 "src/qs8-igemm/gen/3x16c16-minmax-neon-mlal-padal.c",
Frank Barchard1f51d382021-04-20 14:07:14 -07001832 "src/qs8-igemm/gen/4x8-minmax-neon-mlal-lane-prfm.c",
Frank Barchard671d1b02021-03-10 15:42:08 -08001833 "src/qs8-igemm/gen/4x8-minmax-neon-mlal-lane.c",
1834 "src/qs8-igemm/gen/4x8-minmax-neon-mull-addw-dup.c",
1835 "src/qs8-igemm/gen/4x8c2-minmax-neon-mlal-padal-dup.c",
1836 "src/qs8-igemm/gen/4x8c2-minmax-neon-mull-padal-dup.c",
Frank Barchardda78da12021-03-02 14:28:00 -08001837 "src/qs8-igemm/gen/4x8c8-minmax-neon-mlal-padal.c",
1838 "src/qs8-igemm/gen/4x8c8-minmax-neon-mull-padal.c",
Frank Barchardda78da12021-03-02 14:28:00 -08001839 "src/qs8-igemm/gen/4x8c16-minmax-neon-mlal-padal.c",
Frank Barchard1f51d382021-04-20 14:07:14 -07001840 "src/qs8-igemm/gen/4x16-minmax-neon-mlal-lane-prfm.c",
Frank Barchard671d1b02021-03-10 15:42:08 -08001841 "src/qs8-igemm/gen/4x16-minmax-neon-mlal-lane.c",
Frank Barchardda78da12021-03-02 14:28:00 -08001842 "src/qs8-igemm/gen/4x16-minmax-neon-mull-addw-dup.c",
Frank Barchard671d1b02021-03-10 15:42:08 -08001843 "src/qs8-igemm/gen/4x16c2-minmax-neon-mlal-padal-dup.c",
1844 "src/qs8-igemm/gen/4x16c2-minmax-neon-mull-padal-dup.c",
1845 "src/qs8-igemm/gen/4x16c8-minmax-neon-mlal-padal.c",
1846 "src/qs8-igemm/gen/4x16c8-minmax-neon-mull-padal.c",
1847 "src/qs8-igemm/gen/4x16c16-minmax-neon-mlal-padal.c",
Frank Barcharddddb38f2021-05-03 16:51:48 -07001848 "src/qs8-igemm/gen/6x8-minmax-neon-mlal-lane-prfm.c",
1849 "src/qs8-igemm/gen/6x8-minmax-neon-mlal-lane.c",
1850 "src/qs8-igemm/gen/6x16-minmax-neon-mlal-lane-prfm.c",
1851 "src/qs8-igemm/gen/6x16-minmax-neon-mlal-lane.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -07001852 "src/qs8-requantization/fp32-neon.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001853 "src/qs8-requantization/precise-neon.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -07001854 "src/qs8-requantization/q31-neon.c",
Marat Dukhanba7b2792020-09-02 14:26:45 -07001855 "src/qs8-vadd/gen/minmax-neon-ld64-x8.c",
1856 "src/qs8-vadd/gen/minmax-neon-ld64-x16.c",
1857 "src/qs8-vadd/gen/minmax-neon-ld64-x24.c",
1858 "src/qs8-vadd/gen/minmax-neon-ld64-x32.c",
1859 "src/qs8-vaddc/gen/minmax-neon-ld64-x8.c",
1860 "src/qs8-vaddc/gen/minmax-neon-ld64-x16.c",
1861 "src/qs8-vaddc/gen/minmax-neon-ld64-x24.c",
1862 "src/qs8-vaddc/gen/minmax-neon-ld64-x32.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -07001863 "src/qu8-avgpool/9p8x-minmax-neon-c8.c",
1864 "src/qu8-avgpool/9x-minmax-neon-c8.c",
1865 "src/qu8-dwconv/up8x9-minmax-neon.c",
1866 "src/qu8-gavgpool/7p7x-minmax-neon-c8.c",
1867 "src/qu8-gavgpool/7x-minmax-neon-c8.c",
1868 "src/qu8-gemm/4x8-minmax-neon.c",
1869 "src/qu8-gemm/8x8-minmax-neon.c",
1870 "src/qu8-igemm/4x8-minmax-neon.c",
1871 "src/qu8-igemm/8x8-minmax-neon.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07001872 "src/qu8-requantization/fp32-neon.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001873 "src/qu8-requantization/precise-neon.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07001874 "src/qu8-requantization/q31-neon.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -07001875 "src/qu8-vadd/minmax-neon.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001876 "src/u8-maxpool/9p8x-minmax-neon-c16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001877 "src/u8-rmax/neon.c",
Marat Dukhan6674d692021-05-05 22:27:00 -07001878 "src/u8-vclamp/neon-x64.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001879 "src/x8-zip/x2-neon.c",
1880 "src/x8-zip/x3-neon.c",
1881 "src/x8-zip/x4-neon.c",
1882 "src/x8-zip/xm-neon.c",
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -07001883 "src/x32-fill/neon.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001884 "src/x32-packx/x4-neon-st4.c",
Marat Dukhan63523d42020-05-22 17:07:33 -07001885 "src/x32-pad/neon.c",
Marat Dukhan57dccd82020-04-14 00:53:10 -07001886 "src/x32-unpool/neon.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001887 "src/x32-zip/x2-neon.c",
1888 "src/x32-zip/x3-neon.c",
1889 "src/x32-zip/x4-neon.c",
1890 "src/x32-zip/xm-neon.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001891]
1892
1893NEONFMA_UKERNELS = [
Frank Barchard04336c12020-10-22 16:48:55 -07001894 "src/f32-dwconv/gen/up4x4-minmax-neonfma-acc2.c",
1895 "src/f32-dwconv/gen/up4x4-minmax-neonfma.c",
1896 "src/f32-dwconv/gen/up4x9-minmax-neonfma-acc2.c",
1897 "src/f32-dwconv/gen/up4x9-minmax-neonfma.c",
1898 "src/f32-dwconv/gen/up4x25-minmax-neonfma-acc2.c",
1899 "src/f32-dwconv/gen/up4x25-minmax-neonfma.c",
1900 "src/f32-dwconv/gen/up8x4-minmax-neonfma-acc2.c",
1901 "src/f32-dwconv/gen/up8x4-minmax-neonfma.c",
1902 "src/f32-dwconv/gen/up8x9-minmax-neonfma-acc2.c",
1903 "src/f32-dwconv/gen/up8x9-minmax-neonfma.c",
1904 "src/f32-dwconv/gen/up8x25-minmax-neonfma-acc2.c",
1905 "src/f32-dwconv/gen/up8x25-minmax-neonfma.c",
1906 "src/f32-gemm/gen-inc/1x8inc-minmax-neonfma-dup-ld64.c",
1907 "src/f32-gemm/gen-inc/1x8s4inc-minmax-neonfma.c",
1908 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-dup-ld64.c",
1909 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-dup-ld128.c",
1910 "src/f32-gemm/gen-inc/4x8s4inc-minmax-neonfma.c",
1911 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-dup-ld64.c",
1912 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-dup-ld128.c",
1913 "src/f32-gemm/gen-inc/6x8s4inc-minmax-neonfma.c",
1914 "src/f32-gemm/gen-inc/8x8s4inc-minmax-neonfma.c",
1915 "src/f32-gemm/gen/1x8-minmax-neonfma-dup-ld64.c",
1916 "src/f32-gemm/gen/1x8s4-minmax-neonfma.c",
1917 "src/f32-gemm/gen/4x8-minmax-neonfma-dup-ld64.c",
1918 "src/f32-gemm/gen/4x8-minmax-neonfma-dup-ld128.c",
1919 "src/f32-gemm/gen/4x8s4-minmax-neonfma.c",
1920 "src/f32-gemm/gen/6x8-minmax-neonfma-dup-ld64.c",
1921 "src/f32-gemm/gen/6x8-minmax-neonfma-dup-ld128.c",
1922 "src/f32-gemm/gen/6x8s4-minmax-neonfma.c",
1923 "src/f32-gemm/gen/8x8s4-minmax-neonfma.c",
Artsiom Ablavatski2202c812021-01-22 14:16:43 -08001924 "src/f32-ibilinear-chw/gen/neonfma-p4.c",
1925 "src/f32-ibilinear-chw/gen/neonfma-p8.c",
Frank Barchard8247e212021-02-03 18:12:33 -08001926 "src/f32-ibilinear/gen/neonfma-c4.c",
1927 "src/f32-ibilinear/gen/neonfma-c8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001928 "src/f32-igemm/gen/1x8-minmax-neonfma-dup-ld64.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001929 "src/f32-igemm/gen/1x8s4-minmax-neonfma.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001930 "src/f32-igemm/gen/4x8-minmax-neonfma-dup-ld64.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001931 "src/f32-igemm/gen/4x8-minmax-neonfma-dup-ld128.c",
1932 "src/f32-igemm/gen/4x8s4-minmax-neonfma.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001933 "src/f32-igemm/gen/6x8-minmax-neonfma-dup-ld64.c",
1934 "src/f32-igemm/gen/6x8-minmax-neonfma-dup-ld128.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001935 "src/f32-igemm/gen/6x8s4-minmax-neonfma.c",
1936 "src/f32-igemm/gen/8x8s4-minmax-neonfma.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001937 "src/f32-ppmm/gen/4x8-minmax-neonfma.c",
1938 "src/f32-ppmm/gen/8x8-minmax-neonfma.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001939 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x4.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001940 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x8-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001941 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x8.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001942 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12-acc2.c",
1943 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001944 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001945 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16-acc2.c",
1946 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001947 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001948 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20-acc2.c",
1949 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001950 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20.c",
1951 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x4.c",
1952 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x8-acc2.c",
1953 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x8.c",
1954 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12-acc2.c",
1955 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12-acc3.c",
1956 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12.c",
1957 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16-acc2.c",
1958 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16-acc4.c",
1959 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16.c",
1960 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20-acc2.c",
1961 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20-acc5.c",
1962 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20.c",
Marat Dukhan2fa7a0c2020-12-06 19:09:02 -08001963 "src/f32-spmm/gen/4x1-minmax-neonfma-pipelined.c",
1964 "src/f32-spmm/gen/4x1-minmax-neonfma-x2.c",
1965 "src/f32-spmm/gen/4x1-minmax-neonfma.c",
1966 "src/f32-spmm/gen/8x1-minmax-neonfma-pipelined.c",
1967 "src/f32-spmm/gen/8x1-minmax-neonfma-x2.c",
1968 "src/f32-spmm/gen/8x1-minmax-neonfma.c",
1969 "src/f32-spmm/gen/12x1-minmax-neonfma.c",
1970 "src/f32-spmm/gen/16x1-minmax-neonfma-pipelined.c",
1971 "src/f32-spmm/gen/16x1-minmax-neonfma-x2.c",
1972 "src/f32-spmm/gen/16x1-minmax-neonfma.c",
1973 "src/f32-spmm/gen/32x1-minmax-neonfma-pipelined.c",
1974 "src/f32-spmm/gen/32x1-minmax-neonfma-x2.c",
1975 "src/f32-spmm/gen/32x1-minmax-neonfma.c",
Marat Dukhaned6baaf2020-12-01 15:07:08 -08001976 "src/f32-velu/gen/velu-neonfma-rr1-lut16-p3-x4.c",
1977 "src/f32-velu/gen/velu-neonfma-rr1-lut16-p3-x8.c",
1978 "src/f32-velu/gen/velu-neonfma-rr1-lut16-p3-x12.c",
1979 "src/f32-velu/gen/velu-neonfma-rr1-lut16-p3-x16.c",
1980 "src/f32-velu/gen/velu-neonfma-rr1-lut16-p3-x20.c",
1981 "src/f32-velu/gen/velu-neonfma-rr1-lut16-p3-x24.c",
1982 "src/f32-velu/gen/velu-neonfma-rr1-p6-x4.c",
1983 "src/f32-velu/gen/velu-neonfma-rr1-p6-x8.c",
1984 "src/f32-velu/gen/velu-neonfma-rr1-p6-x12.c",
1985 "src/f32-velu/gen/velu-neonfma-rr1-p6-x16.c",
1986 "src/f32-velu/gen/velu-neonfma-rr1-p6-x20.c",
1987 "src/f32-velu/gen/velu-neonfma-rr1-p6-x24.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001988 "src/f32-vmulcaddc/gen/c4-minmax-neonfma-2x.c",
1989 "src/f32-vmulcaddc/gen/c8-minmax-neonfma-2x.c",
Marat Dukhanb1eec082021-05-05 23:24:55 -07001990 "src/f32-vsigmoid/gen/vsigmoid-neonfma-rr1-lut64-p2-nr1recps1fma-x4.c",
1991 "src/f32-vsigmoid/gen/vsigmoid-neonfma-rr1-lut64-p2-nr1recps1fma-x8.c",
1992 "src/f32-vsigmoid/gen/vsigmoid-neonfma-rr1-lut64-p2-nr1recps1fma-x12.c",
1993 "src/f32-vsigmoid/gen/vsigmoid-neonfma-rr1-lut64-p2-nr1recps1fma-x16.c",
1994 "src/f32-vsigmoid/gen/vsigmoid-neonfma-rr1-lut64-p2-nr1recps1fma-x20.c",
1995 "src/f32-vsigmoid/gen/vsigmoid-neonfma-rr1-lut64-p2-nr1recps1fma-x24.c",
1996 "src/f32-vsigmoid/gen/vsigmoid-neonfma-rr1-lut64-p2-nr2fma-x4.c",
1997 "src/f32-vsigmoid/gen/vsigmoid-neonfma-rr1-lut64-p2-nr2fma-x8.c",
1998 "src/f32-vsigmoid/gen/vsigmoid-neonfma-rr1-lut64-p2-nr2fma-x12.c",
1999 "src/f32-vsigmoid/gen/vsigmoid-neonfma-rr1-lut64-p2-nr2fma-x16.c",
2000 "src/f32-vsigmoid/gen/vsigmoid-neonfma-rr1-lut64-p2-nr2fma-x20.c",
2001 "src/f32-vsigmoid/gen/vsigmoid-neonfma-rr1-lut64-p2-nr2fma-x24.c",
2002 "src/f32-vsigmoid/gen/vsigmoid-neonfma-rr1-lut64-p2-nr2recps-x4.c",
2003 "src/f32-vsigmoid/gen/vsigmoid-neonfma-rr1-lut64-p2-nr2recps-x8.c",
2004 "src/f32-vsigmoid/gen/vsigmoid-neonfma-rr1-lut64-p2-nr2recps-x12.c",
2005 "src/f32-vsigmoid/gen/vsigmoid-neonfma-rr1-lut64-p2-nr2recps-x16.c",
2006 "src/f32-vsigmoid/gen/vsigmoid-neonfma-rr1-lut64-p2-nr2recps-x20.c",
2007 "src/f32-vsigmoid/gen/vsigmoid-neonfma-rr1-lut64-p2-nr2recps-x24.c",
2008 "src/f32-vsigmoid/gen/vsigmoid-neonfma-rr1-lut2048-p1-nr1recps1fma-x4.c",
2009 "src/f32-vsigmoid/gen/vsigmoid-neonfma-rr1-lut2048-p1-nr1recps1fma-x8.c",
2010 "src/f32-vsigmoid/gen/vsigmoid-neonfma-rr1-lut2048-p1-nr1recps1fma-x12.c",
2011 "src/f32-vsigmoid/gen/vsigmoid-neonfma-rr1-lut2048-p1-nr1recps1fma-x16.c",
2012 "src/f32-vsigmoid/gen/vsigmoid-neonfma-rr1-lut2048-p1-nr1recps1fma-x20.c",
2013 "src/f32-vsigmoid/gen/vsigmoid-neonfma-rr1-lut2048-p1-nr1recps1fma-x24.c",
2014 "src/f32-vsigmoid/gen/vsigmoid-neonfma-rr1-lut2048-p1-nr2fma-x4.c",
2015 "src/f32-vsigmoid/gen/vsigmoid-neonfma-rr1-lut2048-p1-nr2fma-x8.c",
2016 "src/f32-vsigmoid/gen/vsigmoid-neonfma-rr1-lut2048-p1-nr2fma-x12.c",
2017 "src/f32-vsigmoid/gen/vsigmoid-neonfma-rr1-lut2048-p1-nr2fma-x16.c",
2018 "src/f32-vsigmoid/gen/vsigmoid-neonfma-rr1-lut2048-p1-nr2fma-x20.c",
2019 "src/f32-vsigmoid/gen/vsigmoid-neonfma-rr1-lut2048-p1-nr2fma-x24.c",
2020 "src/f32-vsigmoid/gen/vsigmoid-neonfma-rr1-lut2048-p1-nr2recps-x4.c",
2021 "src/f32-vsigmoid/gen/vsigmoid-neonfma-rr1-lut2048-p1-nr2recps-x8.c",
2022 "src/f32-vsigmoid/gen/vsigmoid-neonfma-rr1-lut2048-p1-nr2recps-x12.c",
2023 "src/f32-vsigmoid/gen/vsigmoid-neonfma-rr1-lut2048-p1-nr2recps-x16.c",
2024 "src/f32-vsigmoid/gen/vsigmoid-neonfma-rr1-lut2048-p1-nr2recps-x20.c",
2025 "src/f32-vsigmoid/gen/vsigmoid-neonfma-rr1-lut2048-p1-nr2recps-x24.c",
2026 "src/f32-vsigmoid/gen/vsigmoid-neonfma-rr1-p5-nr1recps1fma-x4.c",
2027 "src/f32-vsigmoid/gen/vsigmoid-neonfma-rr1-p5-nr1recps1fma-x8.c",
2028 "src/f32-vsigmoid/gen/vsigmoid-neonfma-rr1-p5-nr1recps1fma-x12.c",
2029 "src/f32-vsigmoid/gen/vsigmoid-neonfma-rr1-p5-nr1recps1fma-x16.c",
2030 "src/f32-vsigmoid/gen/vsigmoid-neonfma-rr1-p5-nr1recps1fma-x20.c",
2031 "src/f32-vsigmoid/gen/vsigmoid-neonfma-rr1-p5-nr1recps1fma-x24.c",
2032 "src/f32-vsigmoid/gen/vsigmoid-neonfma-rr1-p5-nr2fma-x4.c",
2033 "src/f32-vsigmoid/gen/vsigmoid-neonfma-rr1-p5-nr2fma-x8.c",
2034 "src/f32-vsigmoid/gen/vsigmoid-neonfma-rr1-p5-nr2fma-x12.c",
2035 "src/f32-vsigmoid/gen/vsigmoid-neonfma-rr1-p5-nr2fma-x16.c",
2036 "src/f32-vsigmoid/gen/vsigmoid-neonfma-rr1-p5-nr2fma-x20.c",
2037 "src/f32-vsigmoid/gen/vsigmoid-neonfma-rr1-p5-nr2fma-x24.c",
2038 "src/f32-vsigmoid/gen/vsigmoid-neonfma-rr1-p5-nr2recps-x4.c",
2039 "src/f32-vsigmoid/gen/vsigmoid-neonfma-rr1-p5-nr2recps-x8.c",
2040 "src/f32-vsigmoid/gen/vsigmoid-neonfma-rr1-p5-nr2recps-x12.c",
2041 "src/f32-vsigmoid/gen/vsigmoid-neonfma-rr1-p5-nr2recps-x16.c",
2042 "src/f32-vsigmoid/gen/vsigmoid-neonfma-rr1-p5-nr2recps-x20.c",
2043 "src/f32-vsigmoid/gen/vsigmoid-neonfma-rr1-p5-nr2recps-x24.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07002044 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x4.c",
2045 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x8.c",
2046 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x12.c",
2047 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x16.c",
2048 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x20.c",
2049 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x24.c",
2050 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x28.c",
2051 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x32.c",
2052 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x36.c",
2053 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x40.c",
2054 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x4.c",
2055 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x8.c",
2056 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x12.c",
2057 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x16.c",
2058 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x20.c",
2059 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x24.c",
2060 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x28.c",
2061 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x32.c",
2062 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x36.c",
2063 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x40.c",
Marat Dukhanb7633f22020-11-20 16:34:56 -08002064 "src/math/exp-neonfma-rr2-lut64-p2.c",
2065 "src/math/exp-neonfma-rr2-p5.c",
Frank Barcharde7223ee2020-12-04 19:04:01 -08002066 "src/math/expm1minus-neonfma-rr1-lut16-p3.c",
2067 "src/math/expm1minus-neonfma-rr1-p6.c",
Marat Dukhan9dd119a2020-11-20 18:20:04 -08002068 "src/math/expminus-neonfma-rr2-lut64-p2.c",
2069 "src/math/expminus-neonfma-rr2-lut2048-p1.c",
2070 "src/math/expminus-neonfma-rr2-p5.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08002071 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr1recps1fma.c",
2072 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr2fma.c",
2073 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr2recps.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002074 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr1recps1fma.c",
2075 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr2fma.c",
2076 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr2recps.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08002077 "src/math/sigmoid-neonfma-rr1-p5-nr1recps1fma.c",
2078 "src/math/sigmoid-neonfma-rr1-p5-nr2fma.c",
2079 "src/math/sigmoid-neonfma-rr1-p5-nr2recps.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08002080 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr1recps1fma.c",
2081 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr2fma.c",
2082 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr2recps.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002083 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr1recps1fma.c",
2084 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr2fma.c",
2085 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr2recps.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08002086 "src/math/sigmoid-neonfma-rr2-p5-nr1recps1fma.c",
2087 "src/math/sigmoid-neonfma-rr2-p5-nr2fma.c",
2088 "src/math/sigmoid-neonfma-rr2-p5-nr2recps.c",
Marat Dukhan84000762020-06-29 18:38:43 -07002089 "src/math/sqrt-neonfma-nr1fma.c",
Marat Dukhan84000762020-06-29 18:38:43 -07002090 "src/math/sqrt-neonfma-nr1rsqrts1fma1adj.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002091 "src/math/sqrt-neonfma-nr2fma.c",
2092 "src/math/sqrt-neonfma-nr2fma1adj.c",
2093 "src/math/sqrt-neonfma-nr3fma.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002094]
2095
2096AARCH64_NEONFMA_UKERNELS = [
Marat Dukhan56b10cd2020-05-18 09:35:49 -07002097 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neonfma-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07002098 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neonfma-2x2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002099 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neonfma-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07002100 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neonfma-2x2.c",
Marat Dukhan56b10cd2020-05-18 09:35:49 -07002101 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neonfma-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07002102 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neonfma-2x2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002103 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neonfma-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07002104 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neonfma-2x2.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -07002105 "src/f32-conv-hwc2chw/3x3s2p1c3x4-neonfma-2x2.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07002106 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-1x4-acc2.c",
2107 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-1x4-acc3.c",
2108 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-1x4-acc4.c",
Marat Dukhan1268a242020-10-24 00:36:32 -07002109 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-1x4.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07002110 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-2x4-acc2.c",
Marat Dukhan1268a242020-10-24 00:36:32 -07002111 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-2x4.c",
2112 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-3x4.c",
2113 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-4x4.c",
2114 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-5x4.c",
2115 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-6x4.c",
Marat Dukhan82f0c322020-10-25 19:17:35 -07002116 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neonfma-1x4-acc2.c",
2117 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neonfma-1x4-acc3.c",
2118 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neonfma-1x4-acc4.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07002119 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neonfma-1x4.c",
Marat Dukhan82f0c322020-10-25 19:17:35 -07002120 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neonfma-2x4-acc2.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07002121 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neonfma-2x4.c",
2122 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neonfma-3x4.c",
2123 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neonfma-4x4.c",
Marat Dukhan149f0ea2020-10-26 12:50:33 -07002124 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-1x4-acc2.c",
2125 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-1x4-acc3.c",
2126 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-1x4-acc4.c",
2127 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-1x4-acc5.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07002128 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-1x4.c",
Marat Dukhan149f0ea2020-10-26 12:50:33 -07002129 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-2x4-acc2.c",
2130 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-2x4-acc3.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07002131 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-2x4.c",
Marat Dukhan149f0ea2020-10-26 12:50:33 -07002132 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-3x4-acc2.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07002133 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-3x4.c",
Marat Dukhan149f0ea2020-10-26 12:50:33 -07002134 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-4x4-acc2.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07002135 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-4x4.c",
2136 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-5x4.c",
Marat Dukhan30d4b252020-10-29 16:33:22 -07002137 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neonfma-1x4-acc2.c",
2138 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neonfma-1x4-acc3.c",
2139 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neonfma-1x4-acc4.c",
2140 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neonfma-1x4-acc5.c",
2141 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neonfma-1x4.c",
2142 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neonfma-2x4-acc2.c",
2143 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neonfma-2x4-acc3.c",
2144 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neonfma-2x4.c",
Marat Dukhan30d4b252020-10-29 16:33:22 -07002145 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neonfma-3x4-acc2.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07002146 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neonfma-3x4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002147 "src/f32-gemm/gen-inc/1x8inc-minmax-neonfma-lane-ld64.c",
2148 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-lane-ld64.c",
2149 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-lane-ld128.c",
2150 "src/f32-gemm/gen-inc/5x8inc-minmax-neonfma-lane-ld64.c",
2151 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-lane-ld64.c",
2152 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-lane-ld128.c",
2153 "src/f32-gemm/gen/1x8-minmax-neonfma-lane-ld64.c",
2154 "src/f32-gemm/gen/4x2-minmax-neonfma-lane-ld64.c",
2155 "src/f32-gemm/gen/4x8-minmax-neonfma-lane-ld64.c",
2156 "src/f32-gemm/gen/4x8-minmax-neonfma-lane-ld128.c",
2157 "src/f32-gemm/gen/5x8-minmax-neonfma-lane-ld64.c",
2158 "src/f32-gemm/gen/6x8-minmax-neonfma-lane-ld64.c",
2159 "src/f32-gemm/gen/6x8-minmax-neonfma-lane-ld128.c",
2160 "src/f32-igemm/gen/1x8-minmax-neonfma-lane-ld64.c",
2161 "src/f32-igemm/gen/4x2-minmax-neonfma-lane-ld64.c",
2162 "src/f32-igemm/gen/4x4-minmax-neonfma-lane-ld64.c",
2163 "src/f32-igemm/gen/4x8-minmax-neonfma-lane-ld64.c",
2164 "src/f32-igemm/gen/4x8-minmax-neonfma-lane-ld128.c",
2165 "src/f32-igemm/gen/6x8-minmax-neonfma-lane-ld64.c",
2166 "src/f32-igemm/gen/6x8-minmax-neonfma-lane-ld128.c",
Marat Dukhan355ab432020-04-09 19:01:52 -07002167 "src/f32-spmm/gen/4x2-minmax-neonfma.c",
2168 "src/f32-spmm/gen/4x4-minmax-neonfma.c",
Marat Dukhan355ab432020-04-09 19:01:52 -07002169 "src/f32-spmm/gen/8x2-minmax-neonfma.c",
2170 "src/f32-spmm/gen/8x4-minmax-neonfma.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002171 "src/f32-spmm/gen/12x2-minmax-neonfma.c",
2172 "src/f32-spmm/gen/12x4-minmax-neonfma.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002173 "src/f32-spmm/gen/16x2-minmax-neonfma.c",
2174 "src/f32-spmm/gen/16x4-minmax-neonfma.c",
Frank Barchard846c0c62020-10-26 15:01:39 -07002175 "src/f32-spmm/gen/32x2-minmax-neonfma.c",
2176 "src/f32-spmm/gen/32x4-minmax-neonfma.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002177 "src/f32-vbinary/gen/vdiv-minmax-neon-x4.c",
2178 "src/f32-vbinary/gen/vdiv-minmax-neon-x8.c",
2179 "src/f32-vbinary/gen/vdivc-minmax-neon-x4.c",
2180 "src/f32-vbinary/gen/vdivc-minmax-neon-x8.c",
2181 "src/f32-vbinary/gen/vrdivc-minmax-neon-x4.c",
2182 "src/f32-vbinary/gen/vrdivc-minmax-neon-x8.c",
Marat Dukhanb1eec082021-05-05 23:24:55 -07002183 "src/f32-vsigmoid/gen/vsigmoid-neonfma-rr1-lut64-p2-div-x4.c",
2184 "src/f32-vsigmoid/gen/vsigmoid-neonfma-rr1-lut64-p2-div-x8.c",
2185 "src/f32-vsigmoid/gen/vsigmoid-neonfma-rr1-lut64-p2-div-x12.c",
2186 "src/f32-vsigmoid/gen/vsigmoid-neonfma-rr1-lut64-p2-div-x16.c",
2187 "src/f32-vsigmoid/gen/vsigmoid-neonfma-rr1-lut64-p2-div-x20.c",
2188 "src/f32-vsigmoid/gen/vsigmoid-neonfma-rr1-lut64-p2-div-x24.c",
2189 "src/f32-vsigmoid/gen/vsigmoid-neonfma-rr1-lut2048-p1-div-x4.c",
2190 "src/f32-vsigmoid/gen/vsigmoid-neonfma-rr1-lut2048-p1-div-x8.c",
2191 "src/f32-vsigmoid/gen/vsigmoid-neonfma-rr1-lut2048-p1-div-x12.c",
2192 "src/f32-vsigmoid/gen/vsigmoid-neonfma-rr1-lut2048-p1-div-x16.c",
2193 "src/f32-vsigmoid/gen/vsigmoid-neonfma-rr1-lut2048-p1-div-x20.c",
2194 "src/f32-vsigmoid/gen/vsigmoid-neonfma-rr1-lut2048-p1-div-x24.c",
2195 "src/f32-vsigmoid/gen/vsigmoid-neonfma-rr1-p5-div-x4.c",
2196 "src/f32-vsigmoid/gen/vsigmoid-neonfma-rr1-p5-div-x8.c",
2197 "src/f32-vsigmoid/gen/vsigmoid-neonfma-rr1-p5-div-x12.c",
2198 "src/f32-vsigmoid/gen/vsigmoid-neonfma-rr1-p5-div-x16.c",
2199 "src/f32-vsigmoid/gen/vsigmoid-neonfma-rr1-p5-div-x20.c",
2200 "src/f32-vsigmoid/gen/vsigmoid-neonfma-rr1-p5-div-x24.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07002201 "src/f32-vsqrt/gen/neon-sqrt-x4.c",
2202 "src/f32-vsqrt/gen/neon-sqrt-x8.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08002203 "src/math/sigmoid-neonfma-rr1-lut64-p2-div.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002204 "src/math/sigmoid-neonfma-rr1-lut2048-p1-div.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08002205 "src/math/sigmoid-neonfma-rr1-p5-div.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08002206 "src/math/sigmoid-neonfma-rr2-lut64-p2-div.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002207 "src/math/sigmoid-neonfma-rr2-lut2048-p1-div.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08002208 "src/math/sigmoid-neonfma-rr2-p5-div.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002209]
2210
Marat Dukhan8853b822020-05-07 12:19:01 -07002211NEONV8_UKERNELS = [
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07002212 "src/f32-vrnd/gen/vrndd-neonv8-x4.c",
2213 "src/f32-vrnd/gen/vrndd-neonv8-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002214 "src/f32-vrnd/gen/vrndne-neonv8-x4.c",
2215 "src/f32-vrnd/gen/vrndne-neonv8-x8.c",
2216 "src/f32-vrnd/gen/vrndu-neonv8-x4.c",
2217 "src/f32-vrnd/gen/vrndu-neonv8-x8.c",
2218 "src/f32-vrnd/gen/vrndz-neonv8-x4.c",
2219 "src/f32-vrnd/gen/vrndz-neonv8-x8.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07002220 "src/math/roundd-neonv8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002221 "src/math/roundne-neonv8.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07002222 "src/math/roundu-neonv8.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07002223 "src/math/roundz-neonv8.c",
Marat Dukhan8853b822020-05-07 12:19:01 -07002224]
2225
Marat Dukhan08c4a432019-10-03 09:29:21 -07002226AARCH64_NEONFP16ARITH_UKERNELS = [
Frank Barchard5a599a62020-06-04 20:12:44 -07002227 "src/f16-dwconv/gen/up8x4-minmax-neonfp16arith-acc2.c",
2228 "src/f16-dwconv/gen/up8x4-minmax-neonfp16arith.c",
2229 "src/f16-dwconv/gen/up8x9-minmax-neonfp16arith-acc2.c",
2230 "src/f16-dwconv/gen/up8x9-minmax-neonfp16arith.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002231 "src/f16-dwconv/gen/up8x25-minmax-neonfp16arith-acc2.c",
2232 "src/f16-dwconv/gen/up8x25-minmax-neonfp16arith.c",
2233 "src/f16-dwconv/gen/up16x4-minmax-neonfp16arith-acc2.c",
2234 "src/f16-dwconv/gen/up16x4-minmax-neonfp16arith.c",
2235 "src/f16-dwconv/gen/up16x9-minmax-neonfp16arith-acc2.c",
2236 "src/f16-dwconv/gen/up16x9-minmax-neonfp16arith.c",
2237 "src/f16-dwconv/gen/up16x25-minmax-neonfp16arith-acc2.c",
2238 "src/f16-dwconv/gen/up16x25-minmax-neonfp16arith.c",
Frank Barchard0bb49a72020-06-04 11:35:11 -07002239 "src/f16-gavgpool/7p7x-minmax-neonfp16arith-c8.c",
2240 "src/f16-gavgpool/7x-minmax-neonfp16arith-c8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002241 "src/f16-gemm/gen-inc/1x8inc-minmax-neonfp16arith-ld64.c",
2242 "src/f16-gemm/gen-inc/1x16inc-minmax-neonfp16arith-ld64.c",
2243 "src/f16-gemm/gen-inc/4x8inc-minmax-neonfp16arith-ld64.c",
2244 "src/f16-gemm/gen-inc/4x16inc-minmax-neonfp16arith-ld64.c",
2245 "src/f16-gemm/gen-inc/6x8inc-minmax-neonfp16arith-ld64.c",
2246 "src/f16-gemm/gen-inc/6x16inc-minmax-neonfp16arith-ld64.c",
2247 "src/f16-gemm/gen-inc/8x8inc-minmax-neonfp16arith-ld64.c",
2248 "src/f16-gemm/gen-inc/8x16inc-minmax-neonfp16arith-ld64.c",
2249 "src/f16-gemm/gen/1x8-minmax-neonfp16arith-ld64.c",
2250 "src/f16-gemm/gen/1x16-minmax-neonfp16arith-ld64.c",
2251 "src/f16-gemm/gen/4x8-minmax-neonfp16arith-ld64.c",
2252 "src/f16-gemm/gen/4x16-minmax-neonfp16arith-ld64.c",
2253 "src/f16-gemm/gen/6x8-minmax-neonfp16arith-ld64.c",
2254 "src/f16-gemm/gen/6x16-minmax-neonfp16arith-ld64.c",
2255 "src/f16-gemm/gen/8x8-minmax-neonfp16arith-ld64.c",
2256 "src/f16-gemm/gen/8x16-minmax-neonfp16arith-ld64.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002257 "src/f16-igemm/gen/1x8-minmax-neonfp16arith-ld64.c",
2258 "src/f16-igemm/gen/1x16-minmax-neonfp16arith-ld64.c",
2259 "src/f16-igemm/gen/4x8-minmax-neonfp16arith-ld64.c",
2260 "src/f16-igemm/gen/4x16-minmax-neonfp16arith-ld64.c",
2261 "src/f16-igemm/gen/6x8-minmax-neonfp16arith-ld64.c",
2262 "src/f16-igemm/gen/6x16-minmax-neonfp16arith-ld64.c",
2263 "src/f16-igemm/gen/8x8-minmax-neonfp16arith-ld64.c",
2264 "src/f16-igemm/gen/8x16-minmax-neonfp16arith-ld64.c",
Frank Barchardb1966592020-05-12 13:47:06 -07002265 "src/f16-prelu/gen/neonfp16arith-2x8.c",
Frank Barcharda5316982020-07-23 13:19:28 -07002266 "src/f16-prelu/gen/neonfp16arith-2x16.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07002267 "src/f16-spmm/gen/8x1-minmax-neonfp16arith-x2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002268 "src/f16-spmm/gen/8x1-minmax-neonfp16arith.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07002269 "src/f16-spmm/gen/16x1-minmax-neonfp16arith-x2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002270 "src/f16-spmm/gen/16x1-minmax-neonfp16arith.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07002271 "src/f16-spmm/gen/24x1-minmax-neonfp16arith-x2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002272 "src/f16-spmm/gen/24x1-minmax-neonfp16arith.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07002273 "src/f16-spmm/gen/32x1-minmax-neonfp16arith-x2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002274 "src/f16-spmm/gen/32x1-minmax-neonfp16arith.c",
2275 "src/f16-vbinary/gen/vadd-minmax-neonfp16arith-x8.c",
2276 "src/f16-vbinary/gen/vadd-minmax-neonfp16arith-x16.c",
2277 "src/f16-vbinary/gen/vaddc-minmax-neonfp16arith-x8.c",
2278 "src/f16-vbinary/gen/vaddc-minmax-neonfp16arith-x16.c",
2279 "src/f16-vbinary/gen/vdiv-minmax-neonfp16arith-x8.c",
2280 "src/f16-vbinary/gen/vdiv-minmax-neonfp16arith-x16.c",
2281 "src/f16-vbinary/gen/vdivc-minmax-neonfp16arith-x8.c",
2282 "src/f16-vbinary/gen/vdivc-minmax-neonfp16arith-x16.c",
2283 "src/f16-vbinary/gen/vmax-neonfp16arith-x8.c",
2284 "src/f16-vbinary/gen/vmax-neonfp16arith-x16.c",
2285 "src/f16-vbinary/gen/vmaxc-neonfp16arith-x8.c",
2286 "src/f16-vbinary/gen/vmaxc-neonfp16arith-x16.c",
2287 "src/f16-vbinary/gen/vmin-neonfp16arith-x8.c",
2288 "src/f16-vbinary/gen/vmin-neonfp16arith-x16.c",
2289 "src/f16-vbinary/gen/vminc-neonfp16arith-x8.c",
2290 "src/f16-vbinary/gen/vminc-neonfp16arith-x16.c",
2291 "src/f16-vbinary/gen/vmul-minmax-neonfp16arith-x8.c",
2292 "src/f16-vbinary/gen/vmul-minmax-neonfp16arith-x16.c",
2293 "src/f16-vbinary/gen/vmulc-minmax-neonfp16arith-x8.c",
2294 "src/f16-vbinary/gen/vmulc-minmax-neonfp16arith-x16.c",
2295 "src/f16-vbinary/gen/vrdivc-minmax-neonfp16arith-x8.c",
2296 "src/f16-vbinary/gen/vrdivc-minmax-neonfp16arith-x16.c",
2297 "src/f16-vbinary/gen/vrsubc-minmax-neonfp16arith-x8.c",
2298 "src/f16-vbinary/gen/vrsubc-minmax-neonfp16arith-x16.c",
2299 "src/f16-vbinary/gen/vsub-minmax-neonfp16arith-x8.c",
2300 "src/f16-vbinary/gen/vsub-minmax-neonfp16arith-x16.c",
2301 "src/f16-vbinary/gen/vsubc-minmax-neonfp16arith-x8.c",
2302 "src/f16-vbinary/gen/vsubc-minmax-neonfp16arith-x16.c",
Marat Dukhanb1eec082021-05-05 23:24:55 -07002303 "src/f16-vclamp/gen/vclamp-neonfp16arith-x8.c",
2304 "src/f16-vclamp/gen/vclamp-neonfp16arith-x16.c",
Marat Dukhan6674d692021-05-05 22:27:00 -07002305 "src/f16-vhswish/gen/vhswish-neonfp16arith-x8.c",
2306 "src/f16-vhswish/gen/vhswish-neonfp16arith-x16.c",
Marat Dukhanb1eec082021-05-05 23:24:55 -07002307 "src/f16-vrelu/gen/vrelu-neonfp16arith-x8.c",
2308 "src/f16-vrelu/gen/vrelu-neonfp16arith-x16.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002309 "src/f16-vmulcaddc/gen/c8-minmax-neonfp16arith-2x.c",
2310 "src/f16-vmulcaddc/gen/c16-minmax-neonfp16arith-2x.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002311]
2312
Benoit Jacoba9644732020-08-13 12:48:55 -07002313NEONDOT_UKERNELS = [
Marat Dukhand4c83032020-09-09 23:20:12 -07002314 "src/qs8-gemm/gen/1x8c4-minmax-neondot.c",
Marat Dukhand4c83032020-09-09 23:20:12 -07002315 "src/qs8-gemm/gen/1x16c4-minmax-neondot.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002316 "src/qs8-gemm/gen/4x8c4-minmax-neondot.c",
Benoit Jacoba9644732020-08-13 12:48:55 -07002317 "src/qs8-gemm/gen/4x16c4-minmax-neondot.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002318 "src/qs8-gemm/gen/6x8c4-minmax-neondot.c",
Frank Barchard2fa17452020-09-22 09:51:46 -07002319 "src/qs8-gemm/gen/6x16c4-minmax-neondot.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002320 "src/qs8-gemm/gen/8x8c4-minmax-neondot.c",
Frank Barcharda48848f2020-09-24 00:59:24 -07002321 "src/qs8-gemm/gen/8x16c4-minmax-neondot.c",
Marat Dukhand4c83032020-09-09 23:20:12 -07002322 "src/qs8-igemm/gen/1x8c4-minmax-neondot.c",
Marat Dukhand4c83032020-09-09 23:20:12 -07002323 "src/qs8-igemm/gen/1x16c4-minmax-neondot.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002324 "src/qs8-igemm/gen/4x8c4-minmax-neondot.c",
Benoit Jacoba05487f2020-08-14 13:37:21 -07002325 "src/qs8-igemm/gen/4x16c4-minmax-neondot.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002326 "src/qs8-igemm/gen/6x8c4-minmax-neondot.c",
Frank Barchard2fa17452020-09-22 09:51:46 -07002327 "src/qs8-igemm/gen/6x16c4-minmax-neondot.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002328 "src/qs8-igemm/gen/8x8c4-minmax-neondot.c",
Frank Barcharda48848f2020-09-24 00:59:24 -07002329 "src/qs8-igemm/gen/8x16c4-minmax-neondot.c",
Benoit Jacoba9644732020-08-13 12:48:55 -07002330]
2331
Marat Dukhan08c4a432019-10-03 09:29:21 -07002332SSE_UKERNELS = [
Marat Dukhan99936602020-04-11 16:47:01 -07002333 "src/f32-avgpool/9p8x-minmax-sse-c4.c",
2334 "src/f32-avgpool/9x-minmax-sse-c4.c",
Erich Elsenb1233402020-06-08 15:53:15 -07002335 "src/f32-conv-hwc2chw/3x3s2p1c3x4-sse-1x1.c",
2336 "src/f32-conv-hwc2chw/3x3s2p1c3x4-sse-2x2.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002337 "src/f32-dwconv/gen/up4x4-minmax-sse-acc2.c",
2338 "src/f32-dwconv/gen/up4x4-minmax-sse.c",
2339 "src/f32-dwconv/gen/up4x9-minmax-sse-acc2.c",
2340 "src/f32-dwconv/gen/up4x9-minmax-sse.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002341 "src/f32-dwconv/gen/up4x25-minmax-sse-acc2.c",
2342 "src/f32-dwconv/gen/up4x25-minmax-sse.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002343 "src/f32-dwconv/gen/up8x4-minmax-sse-acc2.c",
2344 "src/f32-dwconv/gen/up8x4-minmax-sse.c",
2345 "src/f32-dwconv/gen/up8x9-minmax-sse-acc2.c",
2346 "src/f32-dwconv/gen/up8x9-minmax-sse.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002347 "src/f32-dwconv/gen/up8x25-minmax-sse-acc2.c",
2348 "src/f32-dwconv/gen/up8x25-minmax-sse.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07002349 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-1x4-acc2.c",
2350 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-1x4-acc3.c",
2351 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-1x4-acc4.c",
Marat Dukhan470078a2020-10-23 22:36:52 -07002352 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-1x4.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07002353 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-2x4-acc2.c",
Marat Dukhan470078a2020-10-23 22:36:52 -07002354 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-2x4.c",
2355 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-3x4.c",
2356 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-4x4.c",
2357 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-5x4.c",
2358 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-6x4.c",
Marat Dukhan0ff97182020-10-25 19:14:03 -07002359 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-sse-1x4-acc2.c",
2360 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-sse-1x4-acc3.c",
2361 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-sse-1x4-acc4.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07002362 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-sse-1x4.c",
Marat Dukhan0ff97182020-10-25 19:14:03 -07002363 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-sse-2x4-acc2.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07002364 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-sse-2x4.c",
2365 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-sse-3x4.c",
2366 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-sse-4x4.c",
Marat Dukhand0503892020-10-30 08:22:04 -07002367 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-1x4-acc2.c",
2368 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-1x4-acc3.c",
2369 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-1x4-acc4.c",
2370 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-1x4-acc5.c",
2371 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-1x4.c",
2372 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-2x4-acc2.c",
2373 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-2x4-acc3.c",
2374 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-2x4.c",
2375 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-3x4-acc2.c",
2376 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-3x4.c",
2377 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-4x4-acc2.c",
2378 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-4x4.c",
2379 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-5x4.c",
Marat Dukhanccca2142020-10-30 17:32:45 -07002380 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-sse-1x4-acc2.c",
2381 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-sse-1x4-acc3.c",
2382 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-sse-1x4-acc4.c",
2383 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-sse-1x4-acc5.c",
2384 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-sse-1x4.c",
2385 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-sse-2x4-acc2.c",
2386 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-sse-2x4-acc3.c",
2387 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-sse-2x4.c",
Marat Dukhanccca2142020-10-30 17:32:45 -07002388 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-sse-3x4-acc2.c",
Frank Barchard8ef44cd2020-11-03 12:30:23 -08002389 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-sse-3x4.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -07002390 "src/f32-gavgpool-cw/sse-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -07002391 "src/f32-gavgpool/7p7x-minmax-sse-c4.c",
2392 "src/f32-gavgpool/7x-minmax-sse-c4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002393 "src/f32-gemm/gen-inc/1x8inc-minmax-sse-dup.c",
2394 "src/f32-gemm/gen-inc/1x8inc-minmax-sse-load1.c",
2395 "src/f32-gemm/gen-inc/1x8s4inc-minmax-sse.c",
Marat Dukhan802fcae2020-12-11 14:37:25 -08002396 "src/f32-gemm/gen-inc/3x8inc-minmax-sse-dup.c",
2397 "src/f32-gemm/gen-inc/3x8inc-minmax-sse-load1.c",
2398 "src/f32-gemm/gen-inc/3x8s4inc-minmax-sse.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002399 "src/f32-gemm/gen-inc/4x8inc-minmax-sse-dup.c",
2400 "src/f32-gemm/gen-inc/4x8inc-minmax-sse-load1.c",
2401 "src/f32-gemm/gen-inc/4x8s4inc-minmax-sse.c",
Marat Dukhan802fcae2020-12-11 14:37:25 -08002402 "src/f32-gemm/gen-inc/5x8inc-minmax-sse-dup.c",
2403 "src/f32-gemm/gen-inc/5x8inc-minmax-sse-load1.c",
2404 "src/f32-gemm/gen-inc/5x8s4inc-minmax-sse.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002405 "src/f32-gemm/gen/1x8-minmax-sse-dup.c",
2406 "src/f32-gemm/gen/1x8-minmax-sse-load1.c",
2407 "src/f32-gemm/gen/1x8s4-minmax-sse.c",
Marat Dukhan802fcae2020-12-11 14:37:25 -08002408 "src/f32-gemm/gen/3x8-minmax-sse-dup.c",
2409 "src/f32-gemm/gen/3x8-minmax-sse-load1.c",
2410 "src/f32-gemm/gen/3x8s4-minmax-sse.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002411 "src/f32-gemm/gen/4x2c4-minmax-sse.c",
2412 "src/f32-gemm/gen/4x8-minmax-sse-dup.c",
2413 "src/f32-gemm/gen/4x8-minmax-sse-load1.c",
2414 "src/f32-gemm/gen/4x8s4-minmax-sse.c",
Marat Dukhan802fcae2020-12-11 14:37:25 -08002415 "src/f32-gemm/gen/5x8-minmax-sse-dup.c",
2416 "src/f32-gemm/gen/5x8-minmax-sse-load1.c",
2417 "src/f32-gemm/gen/5x8s4-minmax-sse.c",
Artsiom Ablavatskib3ffd582021-03-31 13:00:08 -07002418 "src/f32-ibilinear-chw/gen/sse-p4.c",
2419 "src/f32-ibilinear-chw/gen/sse-p8.c",
Frank Barchard4a352042021-04-13 15:52:08 -07002420 "src/f32-ibilinear/gen/sse-c4.c",
2421 "src/f32-ibilinear/gen/sse-c8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002422 "src/f32-igemm/gen/1x8-minmax-sse-dup.c",
2423 "src/f32-igemm/gen/1x8-minmax-sse-load1.c",
2424 "src/f32-igemm/gen/1x8s4-minmax-sse.c",
Marat Dukhan802fcae2020-12-11 14:37:25 -08002425 "src/f32-igemm/gen/3x8-minmax-sse-dup.c",
2426 "src/f32-igemm/gen/3x8-minmax-sse-load1.c",
2427 "src/f32-igemm/gen/3x8s4-minmax-sse.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002428 "src/f32-igemm/gen/4x2c4-minmax-sse.c",
2429 "src/f32-igemm/gen/4x8-minmax-sse-dup.c",
2430 "src/f32-igemm/gen/4x8-minmax-sse-load1.c",
2431 "src/f32-igemm/gen/4x8s4-minmax-sse.c",
Marat Dukhan802fcae2020-12-11 14:37:25 -08002432 "src/f32-igemm/gen/5x8-minmax-sse-dup.c",
2433 "src/f32-igemm/gen/5x8-minmax-sse-load1.c",
2434 "src/f32-igemm/gen/5x8s4-minmax-sse.c",
Marat Dukhan99936602020-04-11 16:47:01 -07002435 "src/f32-maxpool/9p8x-minmax-sse-c4.c",
2436 "src/f32-pavgpool/9p8x-minmax-sse-c4.c",
2437 "src/f32-pavgpool/9x-minmax-sse-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002438 "src/f32-ppmm/gen/4x8-minmax-sse.c",
Marat Dukhan39b5e942020-06-24 15:03:48 -07002439 "src/f32-prelu/gen/sse-2x4.c",
2440 "src/f32-prelu/gen/sse-2x8.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002441 "src/f32-rmax/sse.c",
Marat Dukhan355ab432020-04-09 19:01:52 -07002442 "src/f32-spmm/gen/4x1-minmax-sse.c",
2443 "src/f32-spmm/gen/8x1-minmax-sse.c",
Erich Elsen6e80fdc2020-06-09 15:35:37 -07002444 "src/f32-spmm/gen/16x1-minmax-sse.c",
Frank Barchard846c0c62020-10-26 15:01:39 -07002445 "src/f32-spmm/gen/32x1-minmax-sse.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002446 "src/f32-vbinary/gen/vadd-minmax-sse-x4.c",
2447 "src/f32-vbinary/gen/vadd-minmax-sse-x8.c",
2448 "src/f32-vbinary/gen/vaddc-minmax-sse-x4.c",
2449 "src/f32-vbinary/gen/vaddc-minmax-sse-x8.c",
2450 "src/f32-vbinary/gen/vdiv-minmax-sse-x4.c",
2451 "src/f32-vbinary/gen/vdiv-minmax-sse-x8.c",
2452 "src/f32-vbinary/gen/vdivc-minmax-sse-x4.c",
2453 "src/f32-vbinary/gen/vdivc-minmax-sse-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -08002454 "src/f32-vbinary/gen/vmax-sse-x4.c",
2455 "src/f32-vbinary/gen/vmax-sse-x8.c",
2456 "src/f32-vbinary/gen/vmaxc-sse-x4.c",
2457 "src/f32-vbinary/gen/vmaxc-sse-x8.c",
2458 "src/f32-vbinary/gen/vmin-sse-x4.c",
2459 "src/f32-vbinary/gen/vmin-sse-x8.c",
2460 "src/f32-vbinary/gen/vminc-sse-x4.c",
2461 "src/f32-vbinary/gen/vminc-sse-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002462 "src/f32-vbinary/gen/vmul-minmax-sse-x4.c",
2463 "src/f32-vbinary/gen/vmul-minmax-sse-x8.c",
2464 "src/f32-vbinary/gen/vmulc-minmax-sse-x4.c",
2465 "src/f32-vbinary/gen/vmulc-minmax-sse-x8.c",
2466 "src/f32-vbinary/gen/vrdivc-minmax-sse-x4.c",
2467 "src/f32-vbinary/gen/vrdivc-minmax-sse-x8.c",
2468 "src/f32-vbinary/gen/vrsubc-minmax-sse-x4.c",
2469 "src/f32-vbinary/gen/vrsubc-minmax-sse-x8.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07002470 "src/f32-vbinary/gen/vsqrdiff-sse-x4.c",
2471 "src/f32-vbinary/gen/vsqrdiff-sse-x8.c",
2472 "src/f32-vbinary/gen/vsqrdiffc-sse-x4.c",
2473 "src/f32-vbinary/gen/vsqrdiffc-sse-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002474 "src/f32-vbinary/gen/vsub-minmax-sse-x4.c",
2475 "src/f32-vbinary/gen/vsub-minmax-sse-x8.c",
2476 "src/f32-vbinary/gen/vsubc-minmax-sse-x4.c",
2477 "src/f32-vbinary/gen/vsubc-minmax-sse-x8.c",
Marat Dukhanb1eec082021-05-05 23:24:55 -07002478 "src/f32-vclamp/gen/vclamp-sse-x4.c",
2479 "src/f32-vclamp/gen/vclamp-sse-x8.c",
Marat Dukhan6674d692021-05-05 22:27:00 -07002480 "src/f32-vhswish/gen/vhswish-sse-x4.c",
2481 "src/f32-vhswish/gen/vhswish-sse-x8.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07002482 "src/f32-vlrelu/gen/vlrelu-sse-x4.c",
2483 "src/f32-vlrelu/gen/vlrelu-sse-x8.c",
Marat Dukhan99936602020-04-11 16:47:01 -07002484 "src/f32-vmulcaddc/gen/c4-minmax-sse-2x.c",
2485 "src/f32-vmulcaddc/gen/c8-minmax-sse-2x.c",
Marat Dukhanb1eec082021-05-05 23:24:55 -07002486 "src/f32-vrelu/gen/vrelu-sse-x4.c",
2487 "src/f32-vrelu/gen/vrelu-sse-x8.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07002488 "src/f32-vsqrt/gen/sse-sqrt-x4.c",
2489 "src/f32-vsqrt/gen/sse-sqrt-x8.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07002490 "src/f32-vunary/gen/vabs-sse-x4.c",
2491 "src/f32-vunary/gen/vabs-sse-x8.c",
2492 "src/f32-vunary/gen/vneg-sse-x4.c",
2493 "src/f32-vunary/gen/vneg-sse-x8.c",
2494 "src/f32-vunary/gen/vsqr-sse-x4.c",
2495 "src/f32-vunary/gen/vsqr-sse-x8.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07002496 "src/math/roundd-sse-addsub.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002497 "src/math/roundne-sse-addsub.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07002498 "src/math/roundu-sse-addsub.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07002499 "src/math/roundz-sse-addsub.c",
Marat Dukhan84000762020-06-29 18:38:43 -07002500 "src/math/sqrt-sse-hh1mac.c",
2501 "src/math/sqrt-sse-nr1mac.c",
2502 "src/math/sqrt-sse-nr2mac.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002503 "src/x32-fill/sse.c",
2504 "src/x32-packx/x4-sse.c",
2505 "src/x32-pad/sse.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002506]
2507
2508SSE2_UKERNELS = [
Marat Dukhan329da642019-11-19 21:44:39 -08002509 "src/f32-argmaxpool/4x-sse2-c4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002510 "src/f32-argmaxpool/9p8x-sse2-c4.c",
Marat Dukhan329da642019-11-19 21:44:39 -08002511 "src/f32-argmaxpool/9x-sse2-c4.c",
Marat Dukhan802fcae2020-12-11 14:37:25 -08002512 "src/f32-gemm/gen-inc/1x8inc-minmax-sse2-dup.c",
2513 "src/f32-gemm/gen-inc/3x8inc-minmax-sse2-dup.c",
2514 "src/f32-gemm/gen-inc/4x8inc-minmax-sse2-dup.c",
2515 "src/f32-gemm/gen-inc/5x8inc-minmax-sse2-dup.c",
2516 "src/f32-gemm/gen/1x8-minmax-sse2-dup.c",
2517 "src/f32-gemm/gen/3x8-minmax-sse2-dup.c",
2518 "src/f32-gemm/gen/4x8-minmax-sse2-dup.c",
2519 "src/f32-gemm/gen/5x8-minmax-sse2-dup.c",
2520 "src/f32-igemm/gen/1x8-minmax-sse2-dup.c",
2521 "src/f32-igemm/gen/3x8-minmax-sse2-dup.c",
2522 "src/f32-igemm/gen/4x8-minmax-sse2-dup.c",
2523 "src/f32-igemm/gen/5x8-minmax-sse2-dup.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08002524 "src/f32-prelu/gen/sse2-2x4.c",
2525 "src/f32-prelu/gen/sse2-2x8.c",
Marat Dukhanb39689d2020-01-24 13:32:20 -08002526 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x4.c",
Marat Dukhanb39689d2020-01-24 13:32:20 -08002527 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x8-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002528 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x8.c",
Marat Dukhanb39689d2020-01-24 13:32:20 -08002529 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12-acc2.c",
2530 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002531 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12.c",
Marat Dukhanb39689d2020-01-24 13:32:20 -08002532 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16-acc2.c",
2533 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002534 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16.c",
Marat Dukhanb39689d2020-01-24 13:32:20 -08002535 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20-acc2.c",
2536 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002537 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20.c",
Marat Dukhaned6baaf2020-12-01 15:07:08 -08002538 "src/f32-velu/gen/velu-sse2-rr2-lut16-p3-x4.c",
2539 "src/f32-velu/gen/velu-sse2-rr2-lut16-p3-x8.c",
2540 "src/f32-velu/gen/velu-sse2-rr2-lut16-p3-x12.c",
2541 "src/f32-velu/gen/velu-sse2-rr2-lut16-p3-x16.c",
2542 "src/f32-velu/gen/velu-sse2-rr2-lut16-p3-x20.c",
2543 "src/f32-velu/gen/velu-sse2-rr2-lut16-p3-x24.c",
2544 "src/f32-velu/gen/velu-sse2-rr2-p6-x4.c",
2545 "src/f32-velu/gen/velu-sse2-rr2-p6-x8.c",
2546 "src/f32-velu/gen/velu-sse2-rr2-p6-x12.c",
2547 "src/f32-velu/gen/velu-sse2-rr2-p6-x16.c",
2548 "src/f32-velu/gen/velu-sse2-rr2-p6-x20.c",
2549 "src/f32-velu/gen/velu-sse2-rr2-p6-x24.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07002550 "src/f32-vlrelu/gen/vlrelu-sse2-x4.c",
2551 "src/f32-vlrelu/gen/vlrelu-sse2-x8.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07002552 "src/f32-vrnd/gen/vrndd-sse2-x4.c",
2553 "src/f32-vrnd/gen/vrndd-sse2-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002554 "src/f32-vrnd/gen/vrndne-sse2-x4.c",
2555 "src/f32-vrnd/gen/vrndne-sse2-x8.c",
2556 "src/f32-vrnd/gen/vrndu-sse2-x4.c",
2557 "src/f32-vrnd/gen/vrndu-sse2-x8.c",
2558 "src/f32-vrnd/gen/vrndz-sse2-x4.c",
2559 "src/f32-vrnd/gen/vrndz-sse2-x8.c",
Marat Dukhanb1eec082021-05-05 23:24:55 -07002560 "src/f32-vsigmoid/gen/vsigmoid-sse2-lut64-p2-div-x4.c",
2561 "src/f32-vsigmoid/gen/vsigmoid-sse2-lut64-p2-div-x8.c",
2562 "src/f32-vsigmoid/gen/vsigmoid-sse2-lut64-p2-div-x12.c",
2563 "src/f32-vsigmoid/gen/vsigmoid-sse2-lut64-p2-div-x16.c",
2564 "src/f32-vsigmoid/gen/vsigmoid-sse2-lut64-p2-div-x20.c",
2565 "src/f32-vsigmoid/gen/vsigmoid-sse2-lut64-p2-div-x24.c",
2566 "src/f32-vsigmoid/gen/vsigmoid-sse2-p5-div-x4.c",
2567 "src/f32-vsigmoid/gen/vsigmoid-sse2-p5-div-x8.c",
2568 "src/f32-vsigmoid/gen/vsigmoid-sse2-p5-div-x12.c",
2569 "src/f32-vsigmoid/gen/vsigmoid-sse2-p5-div-x16.c",
2570 "src/f32-vsigmoid/gen/vsigmoid-sse2-p5-div-x20.c",
2571 "src/f32-vsigmoid/gen/vsigmoid-sse2-p5-div-x24.c",
Marat Dukhanb7633f22020-11-20 16:34:56 -08002572 "src/math/exp-sse2-rr2-lut64-p2.c",
2573 "src/math/exp-sse2-rr2-p5.c",
Marat Dukhande390d42020-11-29 19:32:18 -08002574 "src/math/expm1minus-sse2-rr2-lut16-p3.c",
Marat Dukhana438aca2020-11-20 15:45:01 -08002575 "src/math/expm1minus-sse2-rr2-p6.c",
Frank Barchard3b800452020-11-22 12:12:35 -08002576 "src/math/expminus-sse2-rr2-p5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002577 "src/math/roundd-sse2-cvt.c",
2578 "src/math/roundne-sse2-cvt.c",
2579 "src/math/roundu-sse2-cvt.c",
2580 "src/math/roundz-sse2-cvt.c",
2581 "src/math/sigmoid-sse2-rr2-lut64-p2-div.c",
2582 "src/math/sigmoid-sse2-rr2-lut64-p2-nr1.c",
2583 "src/math/sigmoid-sse2-rr2-lut64-p2-nr2.c",
2584 "src/math/sigmoid-sse2-rr2-p5-div.c",
2585 "src/math/sigmoid-sse2-rr2-p5-nr1.c",
2586 "src/math/sigmoid-sse2-rr2-p5-nr2.c",
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07002587 "src/qs8-dwconv/gen/up8x9-minmax-sse2-mul16.c",
2588 "src/qs8-dwconv/gen/up16x9-minmax-sse2-mul16.c",
2589 "src/qs8-dwconv/gen/up24x9-minmax-sse2-mul16.c",
Marat Dukhan159688f2020-08-06 10:34:29 -07002590 "src/qs8-gavgpool/gen/7p7x-minmax-sse2-c8-acc2.c",
2591 "src/qs8-gavgpool/gen/7p7x-minmax-sse2-c16-acc2.c",
2592 "src/qs8-gavgpool/gen/7p7x-minmax-sse2-c24-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002593 "src/qs8-gavgpool/gen/7x-minmax-sse2-c8-acc2.c",
2594 "src/qs8-gavgpool/gen/7x-minmax-sse2-c16-acc2.c",
2595 "src/qs8-gavgpool/gen/7x-minmax-sse2-c24-acc2.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002596 "src/qs8-gemm/gen/1x4c2-minmax-sse2-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002597 "src/qs8-gemm/gen/1x4c2-minmax-sse2-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002598 "src/qs8-gemm/gen/1x4c2-xw-minmax-sse2.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002599 "src/qs8-gemm/gen/1x4c8-minmax-sse2-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002600 "src/qs8-gemm/gen/1x4c8-minmax-sse2-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002601 "src/qs8-gemm/gen/1x4c8-xw-minmax-sse2.c",
Marat Dukhan2e427872021-04-09 23:40:07 -07002602 "src/qs8-gemm/gen/2x4c2-minmax-sse2-ld64.c",
2603 "src/qs8-gemm/gen/2x4c2-minmax-sse2-ld128.c",
2604 "src/qs8-gemm/gen/2x4c2-xw-minmax-sse2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002605 "src/qs8-gemm/gen/2x4c8-minmax-sse2-ld64.c",
2606 "src/qs8-gemm/gen/2x4c8-minmax-sse2-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002607 "src/qs8-gemm/gen/2x4c8-xw-minmax-sse2.c",
Marat Dukhan2e427872021-04-09 23:40:07 -07002608 "src/qs8-gemm/gen/3x4c2-minmax-sse2-ld64.c",
2609 "src/qs8-gemm/gen/3x4c2-minmax-sse2-ld128.c",
2610 "src/qs8-gemm/gen/3x4c2-xw-minmax-sse2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002611 "src/qs8-gemm/gen/3x4c8-minmax-sse2-ld64.c",
2612 "src/qs8-gemm/gen/3x4c8-minmax-sse2-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002613 "src/qs8-gemm/gen/3x4c8-xw-minmax-sse2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002614 "src/qs8-gemm/gen/4x4c2-minmax-sse2-ld64.c",
2615 "src/qs8-gemm/gen/4x4c2-minmax-sse2-ld128.c",
2616 "src/qs8-gemm/gen/4x4c2-xw-minmax-sse2.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002617 "src/qs8-igemm/gen/1x4c2-minmax-sse2-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002618 "src/qs8-igemm/gen/1x4c2-minmax-sse2-ld128.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002619 "src/qs8-igemm/gen/1x4c8-minmax-sse2-ld64.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002620 "src/qs8-igemm/gen/1x4c8-minmax-sse2-ld128.c",
Marat Dukhan2e427872021-04-09 23:40:07 -07002621 "src/qs8-igemm/gen/2x4c2-minmax-sse2-ld64.c",
2622 "src/qs8-igemm/gen/2x4c2-minmax-sse2-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002623 "src/qs8-igemm/gen/2x4c8-minmax-sse2-ld64.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002624 "src/qs8-igemm/gen/2x4c8-minmax-sse2-ld128.c",
Marat Dukhan2e427872021-04-09 23:40:07 -07002625 "src/qs8-igemm/gen/3x4c2-minmax-sse2-ld64.c",
2626 "src/qs8-igemm/gen/3x4c2-minmax-sse2-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002627 "src/qs8-igemm/gen/3x4c8-minmax-sse2-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002628 "src/qs8-igemm/gen/3x4c8-minmax-sse2-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002629 "src/qs8-igemm/gen/4x4c2-minmax-sse2-ld64.c",
2630 "src/qs8-igemm/gen/4x4c2-minmax-sse2-ld128.c",
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07002631 "src/qs8-requantization/fp32-sse2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002632 "src/qs8-requantization/precise-sse2.c",
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07002633 "src/qs8-requantization/q31-sse2.c",
Marat Dukhand9f3ad42020-08-10 12:30:58 -07002634 "src/qs8-vadd/gen/minmax-sse2-mul16-ld64-x8.c",
2635 "src/qs8-vadd/gen/minmax-sse2-mul16-ld64-x16.c",
2636 "src/qs8-vadd/gen/minmax-sse2-mul16-ld64-x24.c",
2637 "src/qs8-vadd/gen/minmax-sse2-mul16-ld64-x32.c",
Marat Dukhan0270d9f2020-08-11 00:56:46 -07002638 "src/qs8-vaddc/gen/minmax-sse2-mul16-ld64-x8.c",
2639 "src/qs8-vaddc/gen/minmax-sse2-mul16-ld64-x16.c",
2640 "src/qs8-vaddc/gen/minmax-sse2-mul16-ld64-x24.c",
2641 "src/qs8-vaddc/gen/minmax-sse2-mul16-ld64-x32.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -07002642 "src/qu8-avgpool/9p8x-minmax-sse2-c8.c",
2643 "src/qu8-avgpool/9x-minmax-sse2-c8.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -07002644 "src/qu8-dwconv/up8x9-minmax-sse2.c",
2645 "src/qu8-gavgpool/7p7x-minmax-sse2-c8.c",
2646 "src/qu8-gavgpool/7x-minmax-sse2-c8.c",
2647 "src/qu8-gemm/2x4c8-minmax-sse2.c",
2648 "src/qu8-gemm/4x4c2-minmax-sse2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002649 "src/qu8-igemm/4x4c2-minmax-sse2.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07002650 "src/qu8-requantization/fp32-sse2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002651 "src/qu8-requantization/precise-sse2.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07002652 "src/qu8-requantization/q31-sse2.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -07002653 "src/qu8-vadd/minmax-sse2.c",
Marat Dukhan99936602020-04-11 16:47:01 -07002654 "src/u8-maxpool/9p8x-minmax-sse2-c16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002655 "src/u8-rmax/sse2.c",
Marat Dukhan6674d692021-05-05 22:27:00 -07002656 "src/u8-vclamp/sse2-x64.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002657 "src/x8-zip/x2-sse2.c",
2658 "src/x8-zip/x3-sse2.c",
2659 "src/x8-zip/x4-sse2.c",
2660 "src/x8-zip/xm-sse2.c",
Marat Dukhan57dccd82020-04-14 00:53:10 -07002661 "src/x32-unpool/sse2.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002662 "src/x32-zip/x2-sse2.c",
2663 "src/x32-zip/x3-sse2.c",
2664 "src/x32-zip/x4-sse2.c",
2665 "src/x32-zip/xm-sse2.c",
Marat Dukhanfe7acb62020-03-09 19:30:05 -07002666]
2667
2668SSSE3_UKERNELS = [
Frank Barchard35db7d02020-10-26 13:37:34 -07002669 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-1x4-acc2.c",
2670 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-1x4-acc3.c",
2671 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-1x4-acc4.c",
Marat Dukhan98f2eeb2020-10-23 23:13:41 -07002672 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-1x4.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07002673 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-2x4-acc2.c",
Marat Dukhan98f2eeb2020-10-23 23:13:41 -07002674 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-2x4.c",
2675 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-3x4.c",
2676 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-4x4.c",
2677 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-5x4.c",
2678 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-6x4.c",
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07002679 "src/qs8-dwconv/gen/up8x9-minmax-ssse3-mul16.c",
2680 "src/qs8-dwconv/gen/up16x9-minmax-ssse3-mul16.c",
2681 "src/qs8-dwconv/gen/up24x9-minmax-ssse3-mul16.c",
Marat Dukhan159688f2020-08-06 10:34:29 -07002682 "src/qs8-gavgpool/gen/7p7x-minmax-ssse3-c8-acc2.c",
2683 "src/qs8-gavgpool/gen/7p7x-minmax-ssse3-c16-acc2.c",
2684 "src/qs8-gavgpool/gen/7p7x-minmax-ssse3-c24-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002685 "src/qs8-gavgpool/gen/7x-minmax-ssse3-c8-acc2.c",
2686 "src/qs8-gavgpool/gen/7x-minmax-ssse3-c16-acc2.c",
2687 "src/qs8-gavgpool/gen/7x-minmax-ssse3-c24-acc2.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002688 "src/qs8-gemm/gen/1x4c2-minmax-ssse3-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002689 "src/qs8-gemm/gen/1x4c2-minmax-ssse3-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002690 "src/qs8-gemm/gen/1x4c2-xw-minmax-ssse3.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002691 "src/qs8-gemm/gen/1x4c8-minmax-ssse3-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002692 "src/qs8-gemm/gen/1x4c8-minmax-ssse3-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002693 "src/qs8-gemm/gen/1x4c8-xw-minmax-ssse3.c",
Marat Dukhan2e427872021-04-09 23:40:07 -07002694 "src/qs8-gemm/gen/2x4c2-minmax-ssse3-ld64.c",
2695 "src/qs8-gemm/gen/2x4c2-minmax-ssse3-ld128.c",
2696 "src/qs8-gemm/gen/2x4c2-xw-minmax-ssse3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002697 "src/qs8-gemm/gen/2x4c8-minmax-ssse3-ld64.c",
2698 "src/qs8-gemm/gen/2x4c8-minmax-ssse3-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002699 "src/qs8-gemm/gen/2x4c8-xw-minmax-ssse3.c",
Marat Dukhan2e427872021-04-09 23:40:07 -07002700 "src/qs8-gemm/gen/3x4c2-minmax-ssse3-ld64.c",
2701 "src/qs8-gemm/gen/3x4c2-minmax-ssse3-ld128.c",
2702 "src/qs8-gemm/gen/3x4c2-xw-minmax-ssse3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002703 "src/qs8-gemm/gen/3x4c8-minmax-ssse3-ld64.c",
2704 "src/qs8-gemm/gen/3x4c8-minmax-ssse3-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002705 "src/qs8-gemm/gen/3x4c8-xw-minmax-ssse3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002706 "src/qs8-gemm/gen/4x4c2-minmax-ssse3-ld64.c",
2707 "src/qs8-gemm/gen/4x4c2-minmax-ssse3-ld128.c",
2708 "src/qs8-gemm/gen/4x4c2-xw-minmax-ssse3.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002709 "src/qs8-igemm/gen/1x4c2-minmax-ssse3-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002710 "src/qs8-igemm/gen/1x4c2-minmax-ssse3-ld128.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002711 "src/qs8-igemm/gen/1x4c8-minmax-ssse3-ld64.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002712 "src/qs8-igemm/gen/1x4c8-minmax-ssse3-ld128.c",
Marat Dukhan2e427872021-04-09 23:40:07 -07002713 "src/qs8-igemm/gen/2x4c2-minmax-ssse3-ld64.c",
2714 "src/qs8-igemm/gen/2x4c2-minmax-ssse3-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002715 "src/qs8-igemm/gen/2x4c8-minmax-ssse3-ld64.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002716 "src/qs8-igemm/gen/2x4c8-minmax-ssse3-ld128.c",
Marat Dukhan2e427872021-04-09 23:40:07 -07002717 "src/qs8-igemm/gen/3x4c2-minmax-ssse3-ld64.c",
2718 "src/qs8-igemm/gen/3x4c2-minmax-ssse3-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002719 "src/qs8-igemm/gen/3x4c8-minmax-ssse3-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002720 "src/qs8-igemm/gen/3x4c8-minmax-ssse3-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002721 "src/qs8-igemm/gen/4x4c2-minmax-ssse3-ld64.c",
2722 "src/qs8-igemm/gen/4x4c2-minmax-ssse3-ld128.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -07002723 "src/qs8-requantization/precise-ssse3.c",
2724 "src/qs8-requantization/q31-ssse3.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07002725 "src/qu8-requantization/precise-ssse3.c",
2726 "src/qu8-requantization/q31-ssse3.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002727]
2728
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08002729SSE41_UKERNELS = [
Marat Dukhan40a672f2019-11-25 03:08:22 -08002730 "src/f32-prelu/gen/sse41-2x4.c",
2731 "src/f32-prelu/gen/sse41-2x8.c",
Marat Dukhaned6baaf2020-12-01 15:07:08 -08002732 "src/f32-velu/gen/velu-sse41-rr2-lut16-p3-x4.c",
2733 "src/f32-velu/gen/velu-sse41-rr2-lut16-p3-x8.c",
2734 "src/f32-velu/gen/velu-sse41-rr2-lut16-p3-x12.c",
2735 "src/f32-velu/gen/velu-sse41-rr2-lut16-p3-x16.c",
2736 "src/f32-velu/gen/velu-sse41-rr2-lut16-p3-x20.c",
2737 "src/f32-velu/gen/velu-sse41-rr2-lut16-p3-x24.c",
2738 "src/f32-velu/gen/velu-sse41-rr2-p6-x4.c",
2739 "src/f32-velu/gen/velu-sse41-rr2-p6-x8.c",
2740 "src/f32-velu/gen/velu-sse41-rr2-p6-x12.c",
2741 "src/f32-velu/gen/velu-sse41-rr2-p6-x16.c",
2742 "src/f32-velu/gen/velu-sse41-rr2-p6-x20.c",
2743 "src/f32-velu/gen/velu-sse41-rr2-p6-x24.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07002744 "src/f32-vlrelu/gen/vlrelu-sse41-x4.c",
2745 "src/f32-vlrelu/gen/vlrelu-sse41-x8.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07002746 "src/f32-vrnd/gen/vrndd-sse41-x4.c",
2747 "src/f32-vrnd/gen/vrndd-sse41-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002748 "src/f32-vrnd/gen/vrndne-sse41-x4.c",
2749 "src/f32-vrnd/gen/vrndne-sse41-x8.c",
2750 "src/f32-vrnd/gen/vrndu-sse41-x4.c",
2751 "src/f32-vrnd/gen/vrndu-sse41-x8.c",
2752 "src/f32-vrnd/gen/vrndz-sse41-x4.c",
2753 "src/f32-vrnd/gen/vrndz-sse41-x8.c",
Marat Dukhanb1eec082021-05-05 23:24:55 -07002754 "src/f32-vsigmoid/gen/vsigmoid-sse41-lut64-p2-div-x4.c",
2755 "src/f32-vsigmoid/gen/vsigmoid-sse41-lut64-p2-div-x8.c",
2756 "src/f32-vsigmoid/gen/vsigmoid-sse41-lut64-p2-div-x12.c",
2757 "src/f32-vsigmoid/gen/vsigmoid-sse41-lut64-p2-div-x16.c",
2758 "src/f32-vsigmoid/gen/vsigmoid-sse41-lut64-p2-div-x20.c",
2759 "src/f32-vsigmoid/gen/vsigmoid-sse41-lut64-p2-div-x24.c",
2760 "src/f32-vsigmoid/gen/vsigmoid-sse41-p5-div-x4.c",
2761 "src/f32-vsigmoid/gen/vsigmoid-sse41-p5-div-x8.c",
2762 "src/f32-vsigmoid/gen/vsigmoid-sse41-p5-div-x12.c",
2763 "src/f32-vsigmoid/gen/vsigmoid-sse41-p5-div-x16.c",
2764 "src/f32-vsigmoid/gen/vsigmoid-sse41-p5-div-x20.c",
2765 "src/f32-vsigmoid/gen/vsigmoid-sse41-p5-div-x24.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002766 "src/math/roundd-sse41.c",
2767 "src/math/roundne-sse41.c",
2768 "src/math/roundu-sse41.c",
2769 "src/math/roundz-sse41.c",
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07002770 "src/qs8-dwconv/gen/up8x9-minmax-sse41-mul16.c",
Marat Dukhan07feec82021-04-02 22:41:15 -07002771 "src/qs8-dwconv/gen/up8x9-minmax-sse41-mul32.c",
Frank Barchard4a352042021-04-13 15:52:08 -07002772 "src/qs8-dwconv/gen/up16x9-minmax-sse41-mul16.c",
Marat Dukhan07feec82021-04-02 22:41:15 -07002773 "src/qs8-dwconv/gen/up16x9-minmax-sse41-mul32.c",
Frank Barchard4a352042021-04-13 15:52:08 -07002774 "src/qs8-dwconv/gen/up24x9-minmax-sse41-mul16.c",
Marat Dukhan07feec82021-04-02 22:41:15 -07002775 "src/qs8-dwconv/gen/up24x9-minmax-sse41-mul32.c",
Marat Dukhan159688f2020-08-06 10:34:29 -07002776 "src/qs8-gavgpool/gen/7p7x-minmax-sse41-c8-acc2.c",
2777 "src/qs8-gavgpool/gen/7p7x-minmax-sse41-c16-acc2.c",
2778 "src/qs8-gavgpool/gen/7p7x-minmax-sse41-c24-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002779 "src/qs8-gavgpool/gen/7x-minmax-sse41-c8-acc2.c",
2780 "src/qs8-gavgpool/gen/7x-minmax-sse41-c16-acc2.c",
2781 "src/qs8-gavgpool/gen/7x-minmax-sse41-c24-acc2.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002782 "src/qs8-gemm/gen/1x4c2-minmax-sse41-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002783 "src/qs8-gemm/gen/1x4c2-minmax-sse41-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002784 "src/qs8-gemm/gen/1x4c2-xw-minmax-sse41.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002785 "src/qs8-gemm/gen/1x4c8-minmax-sse41-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002786 "src/qs8-gemm/gen/1x4c8-minmax-sse41-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002787 "src/qs8-gemm/gen/1x4c8-xw-minmax-sse41.c",
Marat Dukhan2e427872021-04-09 23:40:07 -07002788 "src/qs8-gemm/gen/2x4c2-minmax-sse41-ld64.c",
2789 "src/qs8-gemm/gen/2x4c2-minmax-sse41-ld128.c",
2790 "src/qs8-gemm/gen/2x4c2-xw-minmax-sse41.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002791 "src/qs8-gemm/gen/2x4c8-minmax-sse41-ld64.c",
2792 "src/qs8-gemm/gen/2x4c8-minmax-sse41-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002793 "src/qs8-gemm/gen/2x4c8-xw-minmax-sse41.c",
Marat Dukhan2e427872021-04-09 23:40:07 -07002794 "src/qs8-gemm/gen/3x4c2-minmax-sse41-ld64.c",
2795 "src/qs8-gemm/gen/3x4c2-minmax-sse41-ld128.c",
2796 "src/qs8-gemm/gen/3x4c2-xw-minmax-sse41.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002797 "src/qs8-gemm/gen/3x4c8-minmax-sse41-ld64.c",
2798 "src/qs8-gemm/gen/3x4c8-minmax-sse41-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002799 "src/qs8-gemm/gen/3x4c8-xw-minmax-sse41.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002800 "src/qs8-gemm/gen/4x4c2-minmax-sse41-ld64.c",
2801 "src/qs8-gemm/gen/4x4c2-minmax-sse41-ld128.c",
2802 "src/qs8-gemm/gen/4x4c2-xw-minmax-sse41.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002803 "src/qs8-igemm/gen/1x4c2-minmax-sse41-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002804 "src/qs8-igemm/gen/1x4c2-minmax-sse41-ld128.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002805 "src/qs8-igemm/gen/1x4c8-minmax-sse41-ld64.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002806 "src/qs8-igemm/gen/1x4c8-minmax-sse41-ld128.c",
Marat Dukhan2e427872021-04-09 23:40:07 -07002807 "src/qs8-igemm/gen/2x4c2-minmax-sse41-ld64.c",
2808 "src/qs8-igemm/gen/2x4c2-minmax-sse41-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002809 "src/qs8-igemm/gen/2x4c8-minmax-sse41-ld64.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002810 "src/qs8-igemm/gen/2x4c8-minmax-sse41-ld128.c",
Marat Dukhan2e427872021-04-09 23:40:07 -07002811 "src/qs8-igemm/gen/3x4c2-minmax-sse41-ld64.c",
2812 "src/qs8-igemm/gen/3x4c2-minmax-sse41-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002813 "src/qs8-igemm/gen/3x4c8-minmax-sse41-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002814 "src/qs8-igemm/gen/3x4c8-minmax-sse41-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002815 "src/qs8-igemm/gen/4x4c2-minmax-sse41-ld64.c",
2816 "src/qs8-igemm/gen/4x4c2-minmax-sse41-ld128.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -07002817 "src/qs8-requantization/fp32-sse4.c",
2818 "src/qs8-requantization/precise-sse4.c",
2819 "src/qs8-requantization/q31-sse4.c",
Marat Dukhand9f3ad42020-08-10 12:30:58 -07002820 "src/qs8-vadd/gen/minmax-sse41-mul16-ld64-x8.c",
2821 "src/qs8-vadd/gen/minmax-sse41-mul16-ld64-x16.c",
2822 "src/qs8-vadd/gen/minmax-sse41-mul16-ld64-x24.c",
2823 "src/qs8-vadd/gen/minmax-sse41-mul16-ld64-x32.c",
Marat Dukhanbb9225e2020-09-06 22:40:56 -07002824 "src/qs8-vadd/gen/minmax-sse41-mul32-ld32-x8.c",
2825 "src/qs8-vadd/gen/minmax-sse41-mul32-ld32-x16.c",
2826 "src/qs8-vadd/gen/minmax-sse41-mul32-ld32-x24.c",
2827 "src/qs8-vadd/gen/minmax-sse41-mul32-ld32-x32.c",
Marat Dukhan0270d9f2020-08-11 00:56:46 -07002828 "src/qs8-vaddc/gen/minmax-sse41-mul16-ld64-x8.c",
2829 "src/qs8-vaddc/gen/minmax-sse41-mul16-ld64-x16.c",
2830 "src/qs8-vaddc/gen/minmax-sse41-mul16-ld64-x24.c",
2831 "src/qs8-vaddc/gen/minmax-sse41-mul16-ld64-x32.c",
Marat Dukhanbb9225e2020-09-06 22:40:56 -07002832 "src/qs8-vaddc/gen/minmax-sse41-mul32-ld32-x8.c",
2833 "src/qs8-vaddc/gen/minmax-sse41-mul32-ld32-x16.c",
2834 "src/qs8-vaddc/gen/minmax-sse41-mul32-ld32-x24.c",
2835 "src/qs8-vaddc/gen/minmax-sse41-mul32-ld32-x32.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07002836 "src/qu8-requantization/precise-sse4.c",
2837 "src/qu8-requantization/q31-sse4.c",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08002838]
2839
Marat Dukhan08c4a432019-10-03 09:29:21 -07002840AVX_UKERNELS = [
Marat Dukhan1c587112020-04-08 20:04:28 -07002841 "src/f32-dwconv/gen/up8x4-minmax-avx-acc2.c",
2842 "src/f32-dwconv/gen/up8x4-minmax-avx.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002843 "src/f32-dwconv/gen/up8x9-minmax-avx-acc2.c",
2844 "src/f32-dwconv/gen/up8x9-minmax-avx.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002845 "src/f32-dwconv/gen/up8x25-minmax-avx-acc2.c",
2846 "src/f32-dwconv/gen/up8x25-minmax-avx.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002847 "src/f32-dwconv/gen/up16x4-minmax-avx-acc2.c",
2848 "src/f32-dwconv/gen/up16x4-minmax-avx.c",
2849 "src/f32-dwconv/gen/up16x9-minmax-avx-acc2.c",
2850 "src/f32-dwconv/gen/up16x9-minmax-avx.c",
2851 "src/f32-dwconv/gen/up16x25-minmax-avx-acc2.c",
2852 "src/f32-dwconv/gen/up16x25-minmax-avx.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002853 "src/f32-gemm/gen-inc/1x8inc-minmax-avx-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002854 "src/f32-gemm/gen-inc/1x16inc-minmax-avx-broadcast.c",
2855 "src/f32-gemm/gen-inc/3x16inc-minmax-avx-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002856 "src/f32-gemm/gen-inc/4x8inc-minmax-avx-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002857 "src/f32-gemm/gen-inc/4x16inc-minmax-avx-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002858 "src/f32-gemm/gen-inc/5x8inc-minmax-avx-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002859 "src/f32-gemm/gen-inc/5x16inc-minmax-avx-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002860 "src/f32-gemm/gen-inc/6x8inc-minmax-avx-broadcast.c",
2861 "src/f32-gemm/gen-inc/7x8inc-minmax-avx-broadcast.c",
2862 "src/f32-gemm/gen/1x8-minmax-avx-broadcast.c",
2863 "src/f32-gemm/gen/1x16-minmax-avx-broadcast.c",
2864 "src/f32-gemm/gen/3x16-minmax-avx-broadcast.c",
2865 "src/f32-gemm/gen/4x8-minmax-avx-broadcast.c",
2866 "src/f32-gemm/gen/4x16-minmax-avx-broadcast.c",
2867 "src/f32-gemm/gen/5x8-minmax-avx-broadcast.c",
2868 "src/f32-gemm/gen/5x16-minmax-avx-broadcast.c",
2869 "src/f32-gemm/gen/6x8-minmax-avx-broadcast.c",
2870 "src/f32-gemm/gen/7x8-minmax-avx-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002871 "src/f32-igemm/gen/1x8-minmax-avx-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002872 "src/f32-igemm/gen/1x16-minmax-avx-broadcast.c",
2873 "src/f32-igemm/gen/3x16-minmax-avx-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002874 "src/f32-igemm/gen/4x8-minmax-avx-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002875 "src/f32-igemm/gen/4x16-minmax-avx-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002876 "src/f32-igemm/gen/5x8-minmax-avx-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002877 "src/f32-igemm/gen/5x16-minmax-avx-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002878 "src/f32-igemm/gen/6x8-minmax-avx-broadcast.c",
2879 "src/f32-igemm/gen/7x8-minmax-avx-broadcast.c",
Marat Dukhan90eca0a2020-03-11 00:52:23 -07002880 "src/f32-prelu/gen/avx-2x8.c",
2881 "src/f32-prelu/gen/avx-2x16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002882 "src/f32-rmax/avx.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002883 "src/f32-vbinary/gen/vadd-minmax-avx-x8.c",
2884 "src/f32-vbinary/gen/vadd-minmax-avx-x16.c",
2885 "src/f32-vbinary/gen/vaddc-minmax-avx-x8.c",
2886 "src/f32-vbinary/gen/vaddc-minmax-avx-x16.c",
2887 "src/f32-vbinary/gen/vdiv-minmax-avx-x8.c",
2888 "src/f32-vbinary/gen/vdiv-minmax-avx-x16.c",
2889 "src/f32-vbinary/gen/vdivc-minmax-avx-x8.c",
2890 "src/f32-vbinary/gen/vdivc-minmax-avx-x16.c",
Marat Dukhan9a88efe2019-12-10 15:54:24 -08002891 "src/f32-vbinary/gen/vmax-avx-x8.c",
2892 "src/f32-vbinary/gen/vmax-avx-x16.c",
2893 "src/f32-vbinary/gen/vmaxc-avx-x8.c",
2894 "src/f32-vbinary/gen/vmaxc-avx-x16.c",
2895 "src/f32-vbinary/gen/vmin-avx-x8.c",
2896 "src/f32-vbinary/gen/vmin-avx-x16.c",
2897 "src/f32-vbinary/gen/vminc-avx-x8.c",
2898 "src/f32-vbinary/gen/vminc-avx-x16.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002899 "src/f32-vbinary/gen/vmul-minmax-avx-x8.c",
2900 "src/f32-vbinary/gen/vmul-minmax-avx-x16.c",
2901 "src/f32-vbinary/gen/vmulc-minmax-avx-x8.c",
2902 "src/f32-vbinary/gen/vmulc-minmax-avx-x16.c",
2903 "src/f32-vbinary/gen/vrdivc-minmax-avx-x8.c",
2904 "src/f32-vbinary/gen/vrdivc-minmax-avx-x16.c",
2905 "src/f32-vbinary/gen/vrsubc-minmax-avx-x8.c",
2906 "src/f32-vbinary/gen/vrsubc-minmax-avx-x16.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07002907 "src/f32-vbinary/gen/vsqrdiff-avx-x8.c",
2908 "src/f32-vbinary/gen/vsqrdiff-avx-x16.c",
2909 "src/f32-vbinary/gen/vsqrdiffc-avx-x8.c",
2910 "src/f32-vbinary/gen/vsqrdiffc-avx-x16.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002911 "src/f32-vbinary/gen/vsub-minmax-avx-x8.c",
2912 "src/f32-vbinary/gen/vsub-minmax-avx-x16.c",
2913 "src/f32-vbinary/gen/vsubc-minmax-avx-x8.c",
2914 "src/f32-vbinary/gen/vsubc-minmax-avx-x16.c",
Marat Dukhanb1eec082021-05-05 23:24:55 -07002915 "src/f32-vclamp/gen/vclamp-avx-x8.c",
2916 "src/f32-vclamp/gen/vclamp-avx-x16.c",
Marat Dukhaned6baaf2020-12-01 15:07:08 -08002917 "src/f32-velu/gen/velu-avx-rr2-lut4-p4-perm-x8.c",
2918 "src/f32-velu/gen/velu-avx-rr2-lut4-p4-perm-x16.c",
2919 "src/f32-velu/gen/velu-avx-rr2-lut4-p4-perm-x24.c",
2920 "src/f32-velu/gen/velu-avx-rr2-lut4-p4-perm-x32.c",
2921 "src/f32-velu/gen/velu-avx-rr2-lut4-p4-perm-x40.c",
2922 "src/f32-velu/gen/velu-avx-rr2-lut4-p4-perm-x48.c",
2923 "src/f32-velu/gen/velu-avx-rr2-lut16-p3-x8.c",
2924 "src/f32-velu/gen/velu-avx-rr2-lut16-p3-x16.c",
2925 "src/f32-velu/gen/velu-avx-rr2-lut16-p3-x24.c",
2926 "src/f32-velu/gen/velu-avx-rr2-lut16-p3-x32.c",
2927 "src/f32-velu/gen/velu-avx-rr2-lut16-p3-x40.c",
2928 "src/f32-velu/gen/velu-avx-rr2-lut16-p3-x48.c",
2929 "src/f32-velu/gen/velu-avx-rr2-p6-x8.c",
2930 "src/f32-velu/gen/velu-avx-rr2-p6-x16.c",
2931 "src/f32-velu/gen/velu-avx-rr2-p6-x24.c",
2932 "src/f32-velu/gen/velu-avx-rr2-p6-x32.c",
2933 "src/f32-velu/gen/velu-avx-rr2-p6-x40.c",
2934 "src/f32-velu/gen/velu-avx-rr2-p6-x48.c",
Marat Dukhan6674d692021-05-05 22:27:00 -07002935 "src/f32-vhswish/gen/vhswish-avx-x8.c",
2936 "src/f32-vhswish/gen/vhswish-avx-x16.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07002937 "src/f32-vlrelu/gen/vlrelu-avx-x8.c",
2938 "src/f32-vlrelu/gen/vlrelu-avx-x16.c",
Marat Dukhanb1eec082021-05-05 23:24:55 -07002939 "src/f32-vrelu/gen/vrelu-avx-x8.c",
2940 "src/f32-vrelu/gen/vrelu-avx-x16.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07002941 "src/f32-vrnd/gen/vrndd-avx-x8.c",
2942 "src/f32-vrnd/gen/vrndd-avx-x16.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002943 "src/f32-vrnd/gen/vrndne-avx-x8.c",
2944 "src/f32-vrnd/gen/vrndne-avx-x16.c",
2945 "src/f32-vrnd/gen/vrndu-avx-x8.c",
2946 "src/f32-vrnd/gen/vrndu-avx-x16.c",
2947 "src/f32-vrnd/gen/vrndz-avx-x8.c",
2948 "src/f32-vrnd/gen/vrndz-avx-x16.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07002949 "src/f32-vscale/avx-x32.c",
Marat Dukhanb1eec082021-05-05 23:24:55 -07002950 "src/f32-vsigmoid/gen/vsigmoid-avx-rr2-p5-div-x8.c",
2951 "src/f32-vsigmoid/gen/vsigmoid-avx-rr2-p5-div-x16.c",
2952 "src/f32-vsigmoid/gen/vsigmoid-avx-rr2-p5-div-x24.c",
2953 "src/f32-vsigmoid/gen/vsigmoid-avx-rr2-p5-div-x32.c",
2954 "src/f32-vsigmoid/gen/vsigmoid-avx-rr2-p5-div-x40.c",
2955 "src/f32-vsigmoid/gen/vsigmoid-avx-rr2-p5-div-x48.c",
2956 "src/f32-vsigmoid/gen/vsigmoid-avx-rr2-p5-div-x56.c",
2957 "src/f32-vsigmoid/gen/vsigmoid-avx-rr2-p5-div-x64.c",
2958 "src/f32-vsigmoid/gen/vsigmoid-avx-rr2-p5-div-x72.c",
2959 "src/f32-vsigmoid/gen/vsigmoid-avx-rr2-p5-div-x80.c",
2960 "src/f32-vsigmoid/gen/vsigmoid-avx-rr2-p5-nr2-x8.c",
2961 "src/f32-vsigmoid/gen/vsigmoid-avx-rr2-p5-nr2-x16.c",
2962 "src/f32-vsigmoid/gen/vsigmoid-avx-rr2-p5-nr2-x24.c",
2963 "src/f32-vsigmoid/gen/vsigmoid-avx-rr2-p5-nr2-x32.c",
2964 "src/f32-vsigmoid/gen/vsigmoid-avx-rr2-p5-nr2-x40.c",
2965 "src/f32-vsigmoid/gen/vsigmoid-avx-rr2-p5-nr2-x48.c",
2966 "src/f32-vsigmoid/gen/vsigmoid-avx-rr2-p5-nr2-x56.c",
2967 "src/f32-vsigmoid/gen/vsigmoid-avx-rr2-p5-nr2-x64.c",
2968 "src/f32-vsigmoid/gen/vsigmoid-avx-rr2-p5-nr2-x72.c",
2969 "src/f32-vsigmoid/gen/vsigmoid-avx-rr2-p5-nr2-x80.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07002970 "src/f32-vsqrt/gen/avx-sqrt-x8.c",
2971 "src/f32-vsqrt/gen/avx-sqrt-x16.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07002972 "src/f32-vunary/gen/vabs-avx-x8.c",
2973 "src/f32-vunary/gen/vabs-avx-x16.c",
2974 "src/f32-vunary/gen/vneg-avx-x8.c",
2975 "src/f32-vunary/gen/vneg-avx-x16.c",
2976 "src/f32-vunary/gen/vsqr-avx-x8.c",
2977 "src/f32-vunary/gen/vsqr-avx-x16.c",
Frank Barchard4a352042021-04-13 15:52:08 -07002978 "src/math/exp-avx-rr2-p5.c",
2979 "src/math/expm1minus-avx-rr2-lut4-p4-perm.c",
2980 "src/math/expm1minus-avx-rr2-lut16-p3.c",
2981 "src/math/expm1minus-avx-rr2-p6.c",
2982 "src/math/sigmoid-avx-rr2-lut64-p2-div.c",
2983 "src/math/sigmoid-avx-rr2-p5-div.c",
2984 "src/math/sigmoid-avx-rr2-p5-nr1.c",
2985 "src/math/sigmoid-avx-rr2-p5-nr2.c",
Marat Dukhanfa0ab852021-04-02 17:30:49 -07002986 "src/qs8-dwconv/gen/up8x9-minmax-avx-mul16.c",
Marat Dukhan07feec82021-04-02 22:41:15 -07002987 "src/qs8-dwconv/gen/up8x9-minmax-avx-mul32.c",
Frank Barchard4a352042021-04-13 15:52:08 -07002988 "src/qs8-dwconv/gen/up16x9-minmax-avx-mul16.c",
Marat Dukhan07feec82021-04-02 22:41:15 -07002989 "src/qs8-dwconv/gen/up16x9-minmax-avx-mul32.c",
Frank Barchard4a352042021-04-13 15:52:08 -07002990 "src/qs8-dwconv/gen/up24x9-minmax-avx-mul16.c",
Marat Dukhan07feec82021-04-02 22:41:15 -07002991 "src/qs8-dwconv/gen/up24x9-minmax-avx-mul32.c",
Marat Dukhana3c16332021-04-02 15:03:27 -07002992 "src/qs8-gemm/gen/1x4c2-minmax-avx-ld64.c",
2993 "src/qs8-gemm/gen/1x4c2-minmax-avx-ld128.c",
2994 "src/qs8-gemm/gen/1x4c2-xw-minmax-avx.c",
2995 "src/qs8-gemm/gen/1x4c8-minmax-avx-ld64.c",
2996 "src/qs8-gemm/gen/1x4c8-minmax-avx-ld128.c",
2997 "src/qs8-gemm/gen/1x4c8-xw-minmax-avx.c",
Marat Dukhan2e427872021-04-09 23:40:07 -07002998 "src/qs8-gemm/gen/2x4c2-minmax-avx-ld64.c",
2999 "src/qs8-gemm/gen/2x4c2-minmax-avx-ld128.c",
3000 "src/qs8-gemm/gen/2x4c2-xw-minmax-avx.c",
Marat Dukhana3c16332021-04-02 15:03:27 -07003001 "src/qs8-gemm/gen/2x4c8-minmax-avx-ld64.c",
3002 "src/qs8-gemm/gen/2x4c8-minmax-avx-ld128.c",
3003 "src/qs8-gemm/gen/2x4c8-xw-minmax-avx.c",
Marat Dukhan2e427872021-04-09 23:40:07 -07003004 "src/qs8-gemm/gen/3x4c2-minmax-avx-ld64.c",
3005 "src/qs8-gemm/gen/3x4c2-minmax-avx-ld128.c",
3006 "src/qs8-gemm/gen/3x4c2-xw-minmax-avx.c",
Marat Dukhana3c16332021-04-02 15:03:27 -07003007 "src/qs8-gemm/gen/3x4c8-minmax-avx-ld64.c",
3008 "src/qs8-gemm/gen/3x4c8-minmax-avx-ld128.c",
3009 "src/qs8-gemm/gen/3x4c8-xw-minmax-avx.c",
3010 "src/qs8-gemm/gen/4x4c2-minmax-avx-ld64.c",
3011 "src/qs8-gemm/gen/4x4c2-minmax-avx-ld128.c",
3012 "src/qs8-gemm/gen/4x4c2-xw-minmax-avx.c",
3013 "src/qs8-igemm/gen/1x4c2-minmax-avx-ld64.c",
3014 "src/qs8-igemm/gen/1x4c2-minmax-avx-ld128.c",
3015 "src/qs8-igemm/gen/1x4c8-minmax-avx-ld64.c",
3016 "src/qs8-igemm/gen/1x4c8-minmax-avx-ld128.c",
Marat Dukhan2e427872021-04-09 23:40:07 -07003017 "src/qs8-igemm/gen/2x4c2-minmax-avx-ld64.c",
3018 "src/qs8-igemm/gen/2x4c2-minmax-avx-ld128.c",
Marat Dukhana3c16332021-04-02 15:03:27 -07003019 "src/qs8-igemm/gen/2x4c8-minmax-avx-ld64.c",
3020 "src/qs8-igemm/gen/2x4c8-minmax-avx-ld128.c",
Marat Dukhan2e427872021-04-09 23:40:07 -07003021 "src/qs8-igemm/gen/3x4c2-minmax-avx-ld64.c",
3022 "src/qs8-igemm/gen/3x4c2-minmax-avx-ld128.c",
Marat Dukhana3c16332021-04-02 15:03:27 -07003023 "src/qs8-igemm/gen/3x4c8-minmax-avx-ld64.c",
3024 "src/qs8-igemm/gen/3x4c8-minmax-avx-ld128.c",
3025 "src/qs8-igemm/gen/4x4c2-minmax-avx-ld64.c",
3026 "src/qs8-igemm/gen/4x4c2-minmax-avx-ld128.c",
Marat Dukhane9c4b962021-04-02 16:56:55 -07003027 "src/qs8-vadd/gen/minmax-avx-mul16-ld64-x8.c",
3028 "src/qs8-vadd/gen/minmax-avx-mul16-ld64-x16.c",
3029 "src/qs8-vadd/gen/minmax-avx-mul16-ld64-x24.c",
3030 "src/qs8-vadd/gen/minmax-avx-mul16-ld64-x32.c",
3031 "src/qs8-vadd/gen/minmax-avx-mul32-ld32-x8.c",
3032 "src/qs8-vadd/gen/minmax-avx-mul32-ld32-x16.c",
3033 "src/qs8-vadd/gen/minmax-avx-mul32-ld32-x24.c",
3034 "src/qs8-vadd/gen/minmax-avx-mul32-ld32-x32.c",
3035 "src/qs8-vaddc/gen/minmax-avx-mul16-ld64-x8.c",
3036 "src/qs8-vaddc/gen/minmax-avx-mul16-ld64-x16.c",
3037 "src/qs8-vaddc/gen/minmax-avx-mul16-ld64-x24.c",
3038 "src/qs8-vaddc/gen/minmax-avx-mul16-ld64-x32.c",
3039 "src/qs8-vaddc/gen/minmax-avx-mul32-ld32-x8.c",
3040 "src/qs8-vaddc/gen/minmax-avx-mul32-ld32-x16.c",
3041 "src/qs8-vaddc/gen/minmax-avx-mul32-ld32-x24.c",
3042 "src/qs8-vaddc/gen/minmax-avx-mul32-ld32-x32.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003043]
3044
Marat Dukhan1566fee2020-08-02 21:55:41 -07003045XOP_UKERNELS = [
Marat Dukhan3fd4e272021-04-10 11:16:42 -07003046 "src/qs8-dwconv/gen/up8x9-minmax-xop-mul32.c",
3047 "src/qs8-dwconv/gen/up16x9-minmax-xop-mul32.c",
3048 "src/qs8-dwconv/gen/up24x9-minmax-xop-mul32.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07003049 "src/qs8-gemm/gen/1x4c2-minmax-xop-ld64.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07003050 "src/qs8-gemm/gen/1x4c2-minmax-xop-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07003051 "src/qs8-gemm/gen/1x4c2-xw-minmax-xop.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07003052 "src/qs8-gemm/gen/1x4c8-minmax-xop-ld64.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07003053 "src/qs8-gemm/gen/1x4c8-minmax-xop-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07003054 "src/qs8-gemm/gen/1x4c8-xw-minmax-xop.c",
Marat Dukhan2e427872021-04-09 23:40:07 -07003055 "src/qs8-gemm/gen/2x4c2-minmax-xop-ld64.c",
3056 "src/qs8-gemm/gen/2x4c2-minmax-xop-ld128.c",
3057 "src/qs8-gemm/gen/2x4c2-xw-minmax-xop.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003058 "src/qs8-gemm/gen/2x4c8-minmax-xop-ld64.c",
3059 "src/qs8-gemm/gen/2x4c8-minmax-xop-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07003060 "src/qs8-gemm/gen/2x4c8-xw-minmax-xop.c",
Marat Dukhan2e427872021-04-09 23:40:07 -07003061 "src/qs8-gemm/gen/3x4c2-minmax-xop-ld64.c",
3062 "src/qs8-gemm/gen/3x4c2-minmax-xop-ld128.c",
3063 "src/qs8-gemm/gen/3x4c2-xw-minmax-xop.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003064 "src/qs8-gemm/gen/3x4c8-minmax-xop-ld64.c",
3065 "src/qs8-gemm/gen/3x4c8-minmax-xop-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07003066 "src/qs8-gemm/gen/3x4c8-xw-minmax-xop.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003067 "src/qs8-gemm/gen/4x4c2-minmax-xop-ld64.c",
3068 "src/qs8-gemm/gen/4x4c2-minmax-xop-ld128.c",
3069 "src/qs8-gemm/gen/4x4c2-xw-minmax-xop.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07003070 "src/qs8-igemm/gen/1x4c2-minmax-xop-ld64.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07003071 "src/qs8-igemm/gen/1x4c2-minmax-xop-ld128.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07003072 "src/qs8-igemm/gen/1x4c8-minmax-xop-ld64.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07003073 "src/qs8-igemm/gen/1x4c8-minmax-xop-ld128.c",
Marat Dukhan2e427872021-04-09 23:40:07 -07003074 "src/qs8-igemm/gen/2x4c2-minmax-xop-ld64.c",
3075 "src/qs8-igemm/gen/2x4c2-minmax-xop-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003076 "src/qs8-igemm/gen/2x4c8-minmax-xop-ld64.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07003077 "src/qs8-igemm/gen/2x4c8-minmax-xop-ld128.c",
Marat Dukhan2e427872021-04-09 23:40:07 -07003078 "src/qs8-igemm/gen/3x4c2-minmax-xop-ld64.c",
3079 "src/qs8-igemm/gen/3x4c2-minmax-xop-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003080 "src/qs8-igemm/gen/3x4c8-minmax-xop-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07003081 "src/qs8-igemm/gen/3x4c8-minmax-xop-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003082 "src/qs8-igemm/gen/4x4c2-minmax-xop-ld64.c",
3083 "src/qs8-igemm/gen/4x4c2-minmax-xop-ld128.c",
Marat Dukhanbb9225e2020-09-06 22:40:56 -07003084 "src/qs8-vadd/gen/minmax-xop-mul32-ld32-x8.c",
3085 "src/qs8-vadd/gen/minmax-xop-mul32-ld32-x16.c",
3086 "src/qs8-vadd/gen/minmax-xop-mul32-ld32-x24.c",
3087 "src/qs8-vadd/gen/minmax-xop-mul32-ld32-x32.c",
3088 "src/qs8-vaddc/gen/minmax-xop-mul32-ld32-x8.c",
3089 "src/qs8-vaddc/gen/minmax-xop-mul32-ld32-x16.c",
3090 "src/qs8-vaddc/gen/minmax-xop-mul32-ld32-x24.c",
3091 "src/qs8-vaddc/gen/minmax-xop-mul32-ld32-x32.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07003092]
3093
Marat Dukhanfda12b82019-11-21 12:27:59 -08003094FMA3_UKERNELS = [
Marat Dukhan1c587112020-04-08 20:04:28 -07003095 "src/f32-dwconv/gen/up8x4-minmax-fma3-acc2.c",
3096 "src/f32-dwconv/gen/up8x4-minmax-fma3.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07003097 "src/f32-dwconv/gen/up8x9-minmax-fma3-acc2.c",
3098 "src/f32-dwconv/gen/up8x9-minmax-fma3.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07003099 "src/f32-dwconv/gen/up8x25-minmax-fma3-acc2.c",
3100 "src/f32-dwconv/gen/up8x25-minmax-fma3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003101 "src/f32-dwconv/gen/up16x4-minmax-fma3-acc2.c",
3102 "src/f32-dwconv/gen/up16x4-minmax-fma3.c",
3103 "src/f32-dwconv/gen/up16x9-minmax-fma3-acc2.c",
3104 "src/f32-dwconv/gen/up16x9-minmax-fma3.c",
3105 "src/f32-dwconv/gen/up16x25-minmax-fma3-acc2.c",
3106 "src/f32-dwconv/gen/up16x25-minmax-fma3.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07003107 "src/f32-gemm/gen-inc/1x8inc-minmax-fma3-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003108 "src/f32-gemm/gen-inc/1x16inc-minmax-fma3-broadcast.c",
3109 "src/f32-gemm/gen-inc/1x16s4inc-minmax-fma3-broadcast.c",
3110 "src/f32-gemm/gen-inc/3x16inc-minmax-fma3-broadcast.c",
3111 "src/f32-gemm/gen-inc/3x16s4inc-minmax-fma3-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07003112 "src/f32-gemm/gen-inc/4x8inc-minmax-fma3-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003113 "src/f32-gemm/gen-inc/4x16inc-minmax-fma3-broadcast.c",
3114 "src/f32-gemm/gen-inc/4x16s4inc-minmax-fma3-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07003115 "src/f32-gemm/gen-inc/5x8inc-minmax-fma3-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003116 "src/f32-gemm/gen-inc/5x16inc-minmax-fma3-broadcast.c",
3117 "src/f32-gemm/gen-inc/5x16s4inc-minmax-fma3-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07003118 "src/f32-gemm/gen-inc/6x8inc-minmax-fma3-broadcast.c",
3119 "src/f32-gemm/gen-inc/7x8inc-minmax-fma3-broadcast.c",
3120 "src/f32-gemm/gen-inc/8x8inc-minmax-fma3-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003121 "src/f32-gemm/gen/1x8-minmax-fma3-broadcast.c",
3122 "src/f32-gemm/gen/1x16-minmax-fma3-broadcast.c",
3123 "src/f32-gemm/gen/1x16s4-minmax-fma3-broadcast.c",
3124 "src/f32-gemm/gen/3x16-minmax-fma3-broadcast.c",
3125 "src/f32-gemm/gen/3x16s4-minmax-fma3-broadcast.c",
3126 "src/f32-gemm/gen/4x8-minmax-fma3-broadcast.c",
3127 "src/f32-gemm/gen/4x16-minmax-fma3-broadcast.c",
3128 "src/f32-gemm/gen/4x16s4-minmax-fma3-broadcast.c",
3129 "src/f32-gemm/gen/5x8-minmax-fma3-broadcast.c",
3130 "src/f32-gemm/gen/5x16-minmax-fma3-broadcast.c",
3131 "src/f32-gemm/gen/5x16s4-minmax-fma3-broadcast.c",
3132 "src/f32-gemm/gen/6x8-minmax-fma3-broadcast.c",
3133 "src/f32-gemm/gen/7x8-minmax-fma3-broadcast.c",
3134 "src/f32-gemm/gen/8x8-minmax-fma3-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07003135 "src/f32-igemm/gen/1x8-minmax-fma3-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003136 "src/f32-igemm/gen/1x16-minmax-fma3-broadcast.c",
3137 "src/f32-igemm/gen/1x16s4-minmax-fma3-broadcast.c",
3138 "src/f32-igemm/gen/3x16-minmax-fma3-broadcast.c",
3139 "src/f32-igemm/gen/3x16s4-minmax-fma3-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07003140 "src/f32-igemm/gen/4x8-minmax-fma3-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003141 "src/f32-igemm/gen/4x16-minmax-fma3-broadcast.c",
3142 "src/f32-igemm/gen/4x16s4-minmax-fma3-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07003143 "src/f32-igemm/gen/5x8-minmax-fma3-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003144 "src/f32-igemm/gen/5x16-minmax-fma3-broadcast.c",
3145 "src/f32-igemm/gen/5x16s4-minmax-fma3-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07003146 "src/f32-igemm/gen/6x8-minmax-fma3-broadcast.c",
3147 "src/f32-igemm/gen/7x8-minmax-fma3-broadcast.c",
3148 "src/f32-igemm/gen/8x8-minmax-fma3-broadcast.c",
Marat Dukhan6674d692021-05-05 22:27:00 -07003149 "src/f32-vhswish/gen/vhswish-fma3-x8.c",
3150 "src/f32-vhswish/gen/vhswish-fma3-x16.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07003151 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x8.c",
3152 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x16.c",
3153 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x24.c",
3154 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x32.c",
3155 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x40.c",
3156 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x48.c",
3157 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x56.c",
3158 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x64.c",
Marat Dukhan84000762020-06-29 18:38:43 -07003159 "src/math/sqrt-fma3-nr1fma.c",
Marat Dukhan84000762020-06-29 18:38:43 -07003160 "src/math/sqrt-fma3-nr1fma1adj.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003161 "src/math/sqrt-fma3-nr2fma.c",
Marat Dukhanfda12b82019-11-21 12:27:59 -08003162]
3163
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003164AVX2_UKERNELS = [
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003165 "src/f32-raddexpminusmax/gen/avx2-p5-x64-acc2.c",
3166 "src/f32-raddexpminusmax/gen/avx2-p5-x64-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003167 "src/f32-raddexpminusmax/gen/avx2-p5-x64.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003168 "src/f32-raddexpminusmax/gen/avx2-p5-x72-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003169 "src/f32-raddexpminusmax/gen/avx2-p5-x72.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003170 "src/f32-raddexpminusmax/gen/avx2-p5-x80-acc2.c",
3171 "src/f32-raddexpminusmax/gen/avx2-p5-x80-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003172 "src/f32-raddexpminusmax/gen/avx2-p5-x80.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003173 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc2.c",
3174 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc3.c",
3175 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc6.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003176 "src/f32-raddexpminusmax/gen/avx2-p5-x96.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003177 "src/f32-raddextexp/gen/avx2-p5-x64-acc2.c",
3178 "src/f32-raddextexp/gen/avx2-p5-x64-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003179 "src/f32-raddextexp/gen/avx2-p5-x64.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003180 "src/f32-raddextexp/gen/avx2-p5-x72-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003181 "src/f32-raddextexp/gen/avx2-p5-x72.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003182 "src/f32-raddextexp/gen/avx2-p5-x80-acc2.c",
3183 "src/f32-raddextexp/gen/avx2-p5-x80-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003184 "src/f32-raddextexp/gen/avx2-p5-x80.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003185 "src/f32-raddextexp/gen/avx2-p5-x96-acc2.c",
3186 "src/f32-raddextexp/gen/avx2-p5-x96-acc3.c",
3187 "src/f32-raddextexp/gen/avx2-p5-x96-acc6.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003188 "src/f32-raddextexp/gen/avx2-p5-x96.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003189 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64-acc2.c",
3190 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003191 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003192 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x72-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003193 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x72.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003194 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80-acc2.c",
3195 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003196 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003197 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc2.c",
3198 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc3.c",
3199 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc6.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003200 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96.c",
Marat Dukhaned6baaf2020-12-01 15:07:08 -08003201 "src/f32-velu/gen/velu-avx2-rr1-lut4-p4-perm-x8.c",
3202 "src/f32-velu/gen/velu-avx2-rr1-lut4-p4-perm-x16.c",
3203 "src/f32-velu/gen/velu-avx2-rr1-lut4-p4-perm-x24.c",
3204 "src/f32-velu/gen/velu-avx2-rr1-lut4-p4-perm-x32.c",
3205 "src/f32-velu/gen/velu-avx2-rr1-lut4-p4-perm-x40.c",
3206 "src/f32-velu/gen/velu-avx2-rr1-lut4-p4-perm-x48.c",
3207 "src/f32-velu/gen/velu-avx2-rr1-lut4-p4-perm-x56.c",
3208 "src/f32-velu/gen/velu-avx2-rr1-lut4-p4-perm-x64.c",
3209 "src/f32-velu/gen/velu-avx2-rr1-lut4-p4-perm-x72.c",
3210 "src/f32-velu/gen/velu-avx2-rr1-lut4-p4-perm-x80.c",
3211 "src/f32-velu/gen/velu-avx2-rr1-lut8-p4-perm-x8.c",
3212 "src/f32-velu/gen/velu-avx2-rr1-lut8-p4-perm-x16.c",
3213 "src/f32-velu/gen/velu-avx2-rr1-lut8-p4-perm-x24.c",
3214 "src/f32-velu/gen/velu-avx2-rr1-lut8-p4-perm-x32.c",
3215 "src/f32-velu/gen/velu-avx2-rr1-lut8-p4-perm-x40.c",
3216 "src/f32-velu/gen/velu-avx2-rr1-lut8-p4-perm-x48.c",
3217 "src/f32-velu/gen/velu-avx2-rr1-lut8-p4-perm-x56.c",
3218 "src/f32-velu/gen/velu-avx2-rr1-lut8-p4-perm-x64.c",
3219 "src/f32-velu/gen/velu-avx2-rr1-lut8-p4-perm-x72.c",
3220 "src/f32-velu/gen/velu-avx2-rr1-lut8-p4-perm-x80.c",
3221 "src/f32-velu/gen/velu-avx2-rr1-lut16-p3-gather-x8.c",
3222 "src/f32-velu/gen/velu-avx2-rr1-lut16-p3-gather-x16.c",
3223 "src/f32-velu/gen/velu-avx2-rr1-lut16-p3-gather-x24.c",
3224 "src/f32-velu/gen/velu-avx2-rr1-lut16-p3-gather-x32.c",
3225 "src/f32-velu/gen/velu-avx2-rr1-lut16-p3-gather-x40.c",
3226 "src/f32-velu/gen/velu-avx2-rr1-lut16-p3-gather-x48.c",
3227 "src/f32-velu/gen/velu-avx2-rr1-lut16-p3-gather-x56.c",
3228 "src/f32-velu/gen/velu-avx2-rr1-lut16-p3-gather-x64.c",
3229 "src/f32-velu/gen/velu-avx2-rr1-lut16-p3-gather-x72.c",
3230 "src/f32-velu/gen/velu-avx2-rr1-lut16-p3-gather-x80.c",
3231 "src/f32-velu/gen/velu-avx2-rr1-p6-x8.c",
3232 "src/f32-velu/gen/velu-avx2-rr1-p6-x16.c",
3233 "src/f32-velu/gen/velu-avx2-rr1-p6-x24.c",
3234 "src/f32-velu/gen/velu-avx2-rr1-p6-x32.c",
3235 "src/f32-velu/gen/velu-avx2-rr1-p6-x40.c",
3236 "src/f32-velu/gen/velu-avx2-rr1-p6-x48.c",
3237 "src/f32-velu/gen/velu-avx2-rr1-p6-x56.c",
3238 "src/f32-velu/gen/velu-avx2-rr1-p6-x64.c",
3239 "src/f32-velu/gen/velu-avx2-rr1-p6-x72.c",
3240 "src/f32-velu/gen/velu-avx2-rr1-p6-x80.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003241 "src/f32-vscaleexpminusmax/gen/avx2-p5-x8.c",
3242 "src/f32-vscaleexpminusmax/gen/avx2-p5-x16.c",
3243 "src/f32-vscaleexpminusmax/gen/avx2-p5-x24.c",
3244 "src/f32-vscaleexpminusmax/gen/avx2-p5-x32.c",
3245 "src/f32-vscaleexpminusmax/gen/avx2-p5-x40.c",
3246 "src/f32-vscaleexpminusmax/gen/avx2-p5-x48.c",
3247 "src/f32-vscaleexpminusmax/gen/avx2-p5-x56.c",
3248 "src/f32-vscaleexpminusmax/gen/avx2-p5-x64.c",
3249 "src/f32-vscaleexpminusmax/gen/avx2-p5-x72.c",
3250 "src/f32-vscaleexpminusmax/gen/avx2-p5-x80.c",
3251 "src/f32-vscaleexpminusmax/gen/avx2-p5-x88.c",
3252 "src/f32-vscaleexpminusmax/gen/avx2-p5-x96.c",
3253 "src/f32-vscaleextexp/gen/avx2-p5-x8.c",
3254 "src/f32-vscaleextexp/gen/avx2-p5-x16.c",
3255 "src/f32-vscaleextexp/gen/avx2-p5-x24.c",
3256 "src/f32-vscaleextexp/gen/avx2-p5-x32.c",
3257 "src/f32-vscaleextexp/gen/avx2-p5-x40.c",
3258 "src/f32-vscaleextexp/gen/avx2-p5-x48.c",
3259 "src/f32-vscaleextexp/gen/avx2-p5-x56.c",
3260 "src/f32-vscaleextexp/gen/avx2-p5-x64.c",
3261 "src/f32-vscaleextexp/gen/avx2-p5-x72.c",
3262 "src/f32-vscaleextexp/gen/avx2-p5-x80.c",
3263 "src/f32-vscaleextexp/gen/avx2-p5-x88.c",
3264 "src/f32-vscaleextexp/gen/avx2-p5-x96.c",
Marat Dukhanb1eec082021-05-05 23:24:55 -07003265 "src/f32-vsigmoid/gen/vsigmoid-avx2-rr1-p5-div-x8.c",
3266 "src/f32-vsigmoid/gen/vsigmoid-avx2-rr1-p5-div-x16.c",
3267 "src/f32-vsigmoid/gen/vsigmoid-avx2-rr1-p5-div-x24.c",
3268 "src/f32-vsigmoid/gen/vsigmoid-avx2-rr1-p5-div-x32.c",
3269 "src/f32-vsigmoid/gen/vsigmoid-avx2-rr1-p5-div-x40.c",
3270 "src/f32-vsigmoid/gen/vsigmoid-avx2-rr1-p5-div-x48.c",
3271 "src/f32-vsigmoid/gen/vsigmoid-avx2-rr1-p5-div-x56.c",
3272 "src/f32-vsigmoid/gen/vsigmoid-avx2-rr1-p5-div-x64.c",
3273 "src/f32-vsigmoid/gen/vsigmoid-avx2-rr1-p5-div-x72.c",
3274 "src/f32-vsigmoid/gen/vsigmoid-avx2-rr1-p5-div-x80.c",
3275 "src/f32-vsigmoid/gen/vsigmoid-avx2-rr1-p5-nr1fma-x8.c",
3276 "src/f32-vsigmoid/gen/vsigmoid-avx2-rr1-p5-nr1fma-x16.c",
3277 "src/f32-vsigmoid/gen/vsigmoid-avx2-rr1-p5-nr1fma-x24.c",
3278 "src/f32-vsigmoid/gen/vsigmoid-avx2-rr1-p5-nr1fma-x32.c",
3279 "src/f32-vsigmoid/gen/vsigmoid-avx2-rr1-p5-nr1fma-x40.c",
3280 "src/f32-vsigmoid/gen/vsigmoid-avx2-rr1-p5-nr1fma-x48.c",
3281 "src/f32-vsigmoid/gen/vsigmoid-avx2-rr1-p5-nr1fma-x56.c",
3282 "src/f32-vsigmoid/gen/vsigmoid-avx2-rr1-p5-nr1fma-x64.c",
3283 "src/f32-vsigmoid/gen/vsigmoid-avx2-rr1-p5-nr1fma-x72.c",
3284 "src/f32-vsigmoid/gen/vsigmoid-avx2-rr1-p5-nr1fma-x80.c",
3285 "src/f32-vsigmoid/gen/vsigmoid-avx2-rr1-p5-nr2fma-x8.c",
3286 "src/f32-vsigmoid/gen/vsigmoid-avx2-rr1-p5-nr2fma-x16.c",
3287 "src/f32-vsigmoid/gen/vsigmoid-avx2-rr1-p5-nr2fma-x24.c",
3288 "src/f32-vsigmoid/gen/vsigmoid-avx2-rr1-p5-nr2fma-x32.c",
3289 "src/f32-vsigmoid/gen/vsigmoid-avx2-rr1-p5-nr2fma-x40.c",
3290 "src/f32-vsigmoid/gen/vsigmoid-avx2-rr1-p5-nr2fma-x48.c",
3291 "src/f32-vsigmoid/gen/vsigmoid-avx2-rr1-p5-nr2fma-x56.c",
3292 "src/f32-vsigmoid/gen/vsigmoid-avx2-rr1-p5-nr2fma-x64.c",
3293 "src/f32-vsigmoid/gen/vsigmoid-avx2-rr1-p5-nr2fma-x72.c",
3294 "src/f32-vsigmoid/gen/vsigmoid-avx2-rr1-p5-nr2fma-x80.c",
Marat Dukhanb7633f22020-11-20 16:34:56 -08003295 "src/math/exp-avx2-rr2-lut8-p3-perm.c",
3296 "src/math/exp-avx2-rr2-lut8-p4-perm.c",
3297 "src/math/exp-avx2-rr2-p5.c",
Marat Dukhande390d42020-11-29 19:32:18 -08003298 "src/math/expm1minus-avx2-rr1-lut4-p4-perm.c",
3299 "src/math/expm1minus-avx2-rr1-lut8-p4-perm.c",
3300 "src/math/expm1minus-avx2-rr1-lut16-p3-gather.c",
3301 "src/math/expm1minus-avx2-rr1-p6.c",
Frank Barcharde7223ee2020-12-04 19:04:01 -08003302 "src/math/expminus-avx2-rr2-p5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003303 "src/math/extexp-avx2-p5.c",
3304 "src/math/sigmoid-avx2-rr1-lut64-p2-gather-div.c",
3305 "src/math/sigmoid-avx2-rr1-lut64-p2-gather-nr1fma.c",
3306 "src/math/sigmoid-avx2-rr1-lut64-p2-gather-nr2fma.c",
3307 "src/math/sigmoid-avx2-rr1-lut64-p2-gather-nr2fma1adj.c",
3308 "src/math/sigmoid-avx2-rr1-p5-div.c",
3309 "src/math/sigmoid-avx2-rr1-p5-nr1fma.c",
3310 "src/math/sigmoid-avx2-rr1-p5-nr2fma.c",
3311 "src/math/sigmoid-avx2-rr2-lut64-p2-gather-div.c",
3312 "src/math/sigmoid-avx2-rr2-lut64-p2-gather-nr1fma.c",
3313 "src/math/sigmoid-avx2-rr2-lut64-p2-gather-nr2fma.c",
3314 "src/math/sigmoid-avx2-rr2-lut64-p2-gather-nr2fma1adj.c",
3315 "src/math/sigmoid-avx2-rr2-p5-div.c",
3316 "src/math/sigmoid-avx2-rr2-p5-nr1fma.c",
3317 "src/math/sigmoid-avx2-rr2-p5-nr2fma.c",
Marat Dukhand65a1522020-08-04 19:28:18 -07003318 "src/qs8-dwconv/gen/up8x9-minmax-avx2-mul32.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003319 "src/qs8-dwconv/gen/up16x9-minmax-avx2-mul16.c",
Marat Dukhand65a1522020-08-04 19:28:18 -07003320 "src/qs8-dwconv/gen/up16x9-minmax-avx2-mul32.c",
3321 "src/qs8-dwconv/gen/up24x9-minmax-avx2-mul32.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003322 "src/qs8-dwconv/gen/up32x9-minmax-avx2-mul16.c",
Marat Dukhand65a1522020-08-04 19:28:18 -07003323 "src/qs8-dwconv/gen/up32x9-minmax-avx2-mul32.c",
Marat Dukhan12809522020-08-02 22:23:51 -07003324 "src/qs8-gemm/gen/1x8c8-minmax-avx2.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07003325 "src/qs8-gemm/gen/1x8c8-xw-minmax-avx2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003326 "src/qs8-gemm/gen/2x8c8-minmax-avx2.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07003327 "src/qs8-gemm/gen/2x8c8-xw-minmax-avx2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003328 "src/qs8-gemm/gen/3x8c8-minmax-avx2.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07003329 "src/qs8-gemm/gen/3x8c8-xw-minmax-avx2.c",
Marat Dukhan12809522020-08-02 22:23:51 -07003330 "src/qs8-igemm/gen/1x8c8-minmax-avx2.c",
3331 "src/qs8-igemm/gen/2x8c8-minmax-avx2.c",
3332 "src/qs8-igemm/gen/3x8c8-minmax-avx2.c",
Marat Dukhane6dc0b62020-09-08 23:57:14 -07003333 "src/qs8-vadd/gen/minmax-avx2-mul32-ld64-x8.c",
3334 "src/qs8-vadd/gen/minmax-avx2-mul32-ld64-x16.c",
3335 "src/qs8-vadd/gen/minmax-avx2-mul32-ld64-x24.c",
3336 "src/qs8-vadd/gen/minmax-avx2-mul32-ld64-x32.c",
3337 "src/qs8-vaddc/gen/minmax-avx2-mul32-ld64-x8.c",
3338 "src/qs8-vaddc/gen/minmax-avx2-mul32-ld64-x16.c",
3339 "src/qs8-vaddc/gen/minmax-avx2-mul32-ld64-x24.c",
3340 "src/qs8-vaddc/gen/minmax-avx2-mul32-ld64-x32.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003341]
3342
Marat Dukhan08c4a432019-10-03 09:29:21 -07003343AVX512F_UKERNELS = [
Marat Dukhan1c587112020-04-08 20:04:28 -07003344 "src/f32-dwconv/gen/up16x4-minmax-avx512f-acc2.c",
3345 "src/f32-dwconv/gen/up16x4-minmax-avx512f.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07003346 "src/f32-dwconv/gen/up16x9-minmax-avx512f-acc2.c",
3347 "src/f32-dwconv/gen/up16x9-minmax-avx512f.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07003348 "src/f32-dwconv/gen/up16x25-minmax-avx512f-acc2.c",
3349 "src/f32-dwconv/gen/up16x25-minmax-avx512f.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003350 "src/f32-dwconv/gen/up32x4-minmax-avx512f-acc2.c",
3351 "src/f32-dwconv/gen/up32x4-minmax-avx512f.c",
3352 "src/f32-dwconv/gen/up32x9-minmax-avx512f-acc2.c",
3353 "src/f32-dwconv/gen/up32x9-minmax-avx512f.c",
3354 "src/f32-dwconv/gen/up32x25-minmax-avx512f-acc2.c",
3355 "src/f32-dwconv/gen/up32x25-minmax-avx512f.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07003356 "src/f32-gemm/gen-inc/1x16inc-minmax-avx512f-broadcast.c",
3357 "src/f32-gemm/gen-inc/4x16inc-minmax-avx512f-broadcast.c",
3358 "src/f32-gemm/gen-inc/5x16inc-minmax-avx512f-broadcast.c",
3359 "src/f32-gemm/gen-inc/6x16inc-minmax-avx512f-broadcast.c",
3360 "src/f32-gemm/gen-inc/7x16inc-minmax-avx512f-broadcast.c",
3361 "src/f32-gemm/gen-inc/8x16inc-minmax-avx512f-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003362 "src/f32-gemm/gen/1x16-minmax-avx512f-broadcast.c",
3363 "src/f32-gemm/gen/4x16-minmax-avx512f-broadcast.c",
3364 "src/f32-gemm/gen/5x16-minmax-avx512f-broadcast.c",
3365 "src/f32-gemm/gen/6x16-minmax-avx512f-broadcast.c",
3366 "src/f32-gemm/gen/7x16-minmax-avx512f-broadcast.c",
3367 "src/f32-gemm/gen/8x16-minmax-avx512f-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07003368 "src/f32-igemm/gen/1x16-minmax-avx512f-broadcast.c",
3369 "src/f32-igemm/gen/4x16-minmax-avx512f-broadcast.c",
3370 "src/f32-igemm/gen/5x16-minmax-avx512f-broadcast.c",
3371 "src/f32-igemm/gen/6x16-minmax-avx512f-broadcast.c",
3372 "src/f32-igemm/gen/7x16-minmax-avx512f-broadcast.c",
3373 "src/f32-igemm/gen/8x16-minmax-avx512f-broadcast.c",
Marat Dukhan90eca0a2020-03-11 00:52:23 -07003374 "src/f32-prelu/gen/avx512f-2x16.c",
3375 "src/f32-prelu/gen/avx512f-2x32.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003376 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128-acc2.c",
3377 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003378 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003379 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x144-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003380 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x144.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003381 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160-acc2.c",
3382 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003383 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003384 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc2.c",
3385 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc3.c",
3386 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc6.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003387 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003388 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128-acc2.c",
3389 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003390 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003391 "src/f32-raddextexp/gen/avx512f-p5-scalef-x144-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003392 "src/f32-raddextexp/gen/avx512f-p5-scalef-x144.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003393 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160-acc2.c",
3394 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003395 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003396 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc2.c",
3397 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc3.c",
3398 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc6.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003399 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003400 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128-acc2.c",
3401 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003402 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003403 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x144-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003404 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x144.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003405 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160-acc2.c",
3406 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003407 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003408 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc2.c",
3409 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc3.c",
3410 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc6.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003411 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003412 "src/f32-rmax/avx512f.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07003413 "src/f32-vbinary/gen/vadd-minmax-avx512f-x16.c",
3414 "src/f32-vbinary/gen/vadd-minmax-avx512f-x32.c",
3415 "src/f32-vbinary/gen/vaddc-minmax-avx512f-x16.c",
3416 "src/f32-vbinary/gen/vaddc-minmax-avx512f-x32.c",
3417 "src/f32-vbinary/gen/vdiv-minmax-avx512f-x16.c",
3418 "src/f32-vbinary/gen/vdiv-minmax-avx512f-x32.c",
3419 "src/f32-vbinary/gen/vdivc-minmax-avx512f-x16.c",
3420 "src/f32-vbinary/gen/vdivc-minmax-avx512f-x32.c",
Marat Dukhan9a88efe2019-12-10 15:54:24 -08003421 "src/f32-vbinary/gen/vmax-avx512f-x16.c",
3422 "src/f32-vbinary/gen/vmax-avx512f-x32.c",
3423 "src/f32-vbinary/gen/vmaxc-avx512f-x16.c",
3424 "src/f32-vbinary/gen/vmaxc-avx512f-x32.c",
3425 "src/f32-vbinary/gen/vmin-avx512f-x16.c",
3426 "src/f32-vbinary/gen/vmin-avx512f-x32.c",
3427 "src/f32-vbinary/gen/vminc-avx512f-x16.c",
3428 "src/f32-vbinary/gen/vminc-avx512f-x32.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07003429 "src/f32-vbinary/gen/vmul-minmax-avx512f-x16.c",
3430 "src/f32-vbinary/gen/vmul-minmax-avx512f-x32.c",
3431 "src/f32-vbinary/gen/vmulc-minmax-avx512f-x16.c",
3432 "src/f32-vbinary/gen/vmulc-minmax-avx512f-x32.c",
3433 "src/f32-vbinary/gen/vrdivc-minmax-avx512f-x16.c",
3434 "src/f32-vbinary/gen/vrdivc-minmax-avx512f-x32.c",
3435 "src/f32-vbinary/gen/vrsubc-minmax-avx512f-x16.c",
3436 "src/f32-vbinary/gen/vrsubc-minmax-avx512f-x32.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07003437 "src/f32-vbinary/gen/vsqrdiff-avx512f-x16.c",
3438 "src/f32-vbinary/gen/vsqrdiff-avx512f-x32.c",
3439 "src/f32-vbinary/gen/vsqrdiffc-avx512f-x16.c",
3440 "src/f32-vbinary/gen/vsqrdiffc-avx512f-x32.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07003441 "src/f32-vbinary/gen/vsub-minmax-avx512f-x16.c",
3442 "src/f32-vbinary/gen/vsub-minmax-avx512f-x32.c",
3443 "src/f32-vbinary/gen/vsubc-minmax-avx512f-x16.c",
3444 "src/f32-vbinary/gen/vsubc-minmax-avx512f-x32.c",
Marat Dukhanb1eec082021-05-05 23:24:55 -07003445 "src/f32-vclamp/gen/vclamp-avx512f-x16.c",
3446 "src/f32-vclamp/gen/vclamp-avx512f-x32.c",
Marat Dukhaned6baaf2020-12-01 15:07:08 -08003447 "src/f32-velu/gen/velu-avx512f-rr1-lut16-p3-perm-x16.c",
3448 "src/f32-velu/gen/velu-avx512f-rr1-lut16-p3-perm-x32.c",
3449 "src/f32-velu/gen/velu-avx512f-rr1-lut16-p3-perm-x48.c",
3450 "src/f32-velu/gen/velu-avx512f-rr1-lut16-p3-perm-x64.c",
3451 "src/f32-velu/gen/velu-avx512f-rr1-lut16-p3-perm-x80.c",
3452 "src/f32-velu/gen/velu-avx512f-rr1-lut16-p3-perm-x96.c",
3453 "src/f32-velu/gen/velu-avx512f-rr1-lut16-p3-perm-x112.c",
3454 "src/f32-velu/gen/velu-avx512f-rr1-lut16-p3-perm-x128.c",
3455 "src/f32-velu/gen/velu-avx512f-rr1-p6-x16.c",
3456 "src/f32-velu/gen/velu-avx512f-rr1-p6-x32.c",
3457 "src/f32-velu/gen/velu-avx512f-rr1-p6-x48.c",
3458 "src/f32-velu/gen/velu-avx512f-rr1-p6-x64.c",
3459 "src/f32-velu/gen/velu-avx512f-rr1-p6-x80.c",
3460 "src/f32-velu/gen/velu-avx512f-rr1-p6-x96.c",
3461 "src/f32-velu/gen/velu-avx512f-rr1-p6-x112.c",
3462 "src/f32-velu/gen/velu-avx512f-rr1-p6-x128.c",
Marat Dukhan6674d692021-05-05 22:27:00 -07003463 "src/f32-vhswish/gen/vhswish-avx512f-x16.c",
3464 "src/f32-vhswish/gen/vhswish-avx512f-x32.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07003465 "src/f32-vlrelu/gen/vlrelu-avx512f-x16.c",
3466 "src/f32-vlrelu/gen/vlrelu-avx512f-x32.c",
Marat Dukhanb1eec082021-05-05 23:24:55 -07003467 "src/f32-vrelu/gen/vrelu-avx512f-x16.c",
3468 "src/f32-vrelu/gen/vrelu-avx512f-x32.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003469 "src/f32-vrnd/gen/vrndd-avx512f-x16.c",
3470 "src/f32-vrnd/gen/vrndd-avx512f-x32.c",
3471 "src/f32-vrnd/gen/vrndne-avx512f-x16.c",
3472 "src/f32-vrnd/gen/vrndne-avx512f-x32.c",
3473 "src/f32-vrnd/gen/vrndu-avx512f-x16.c",
3474 "src/f32-vrnd/gen/vrndu-avx512f-x32.c",
3475 "src/f32-vrnd/gen/vrndz-avx512f-x16.c",
3476 "src/f32-vrnd/gen/vrndz-avx512f-x32.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07003477 "src/f32-vscale/avx512f-x64.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003478 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x16.c",
3479 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x32.c",
3480 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x48.c",
3481 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x64.c",
3482 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x80.c",
3483 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x96.c",
3484 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x112.c",
3485 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x128.c",
3486 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x144.c",
3487 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x160.c",
3488 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x176.c",
3489 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x192.c",
3490 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x16.c",
3491 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x32.c",
3492 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x48.c",
3493 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x64.c",
3494 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x80.c",
3495 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x96.c",
3496 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x112.c",
3497 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x128.c",
3498 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x144.c",
3499 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x160.c",
3500 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x176.c",
3501 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x192.c",
Marat Dukhanb1eec082021-05-05 23:24:55 -07003502 "src/f32-vsigmoid/gen/vsigmoid-avx512f-rr1-lut16-p3-perm-scalef-div-x16.c",
3503 "src/f32-vsigmoid/gen/vsigmoid-avx512f-rr1-lut16-p3-perm-scalef-div-x32.c",
3504 "src/f32-vsigmoid/gen/vsigmoid-avx512f-rr1-lut16-p3-perm-scalef-div-x48.c",
3505 "src/f32-vsigmoid/gen/vsigmoid-avx512f-rr1-lut16-p3-perm-scalef-div-x64.c",
3506 "src/f32-vsigmoid/gen/vsigmoid-avx512f-rr1-lut16-p3-perm-scalef-div-x80.c",
3507 "src/f32-vsigmoid/gen/vsigmoid-avx512f-rr1-lut16-p3-perm-scalef-div-x96.c",
3508 "src/f32-vsigmoid/gen/vsigmoid-avx512f-rr1-lut16-p3-perm-scalef-div-x112.c",
3509 "src/f32-vsigmoid/gen/vsigmoid-avx512f-rr1-lut16-p3-perm-scalef-div-x128.c",
3510 "src/f32-vsigmoid/gen/vsigmoid-avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x16.c",
3511 "src/f32-vsigmoid/gen/vsigmoid-avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x32.c",
3512 "src/f32-vsigmoid/gen/vsigmoid-avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x48.c",
3513 "src/f32-vsigmoid/gen/vsigmoid-avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x64.c",
3514 "src/f32-vsigmoid/gen/vsigmoid-avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x80.c",
3515 "src/f32-vsigmoid/gen/vsigmoid-avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x96.c",
3516 "src/f32-vsigmoid/gen/vsigmoid-avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x112.c",
3517 "src/f32-vsigmoid/gen/vsigmoid-avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x128.c",
3518 "src/f32-vsigmoid/gen/vsigmoid-avx512f-rr1-p5-scalef-div-x16.c",
3519 "src/f32-vsigmoid/gen/vsigmoid-avx512f-rr1-p5-scalef-div-x32.c",
3520 "src/f32-vsigmoid/gen/vsigmoid-avx512f-rr1-p5-scalef-div-x48.c",
3521 "src/f32-vsigmoid/gen/vsigmoid-avx512f-rr1-p5-scalef-div-x64.c",
3522 "src/f32-vsigmoid/gen/vsigmoid-avx512f-rr1-p5-scalef-div-x80.c",
3523 "src/f32-vsigmoid/gen/vsigmoid-avx512f-rr1-p5-scalef-div-x96.c",
3524 "src/f32-vsigmoid/gen/vsigmoid-avx512f-rr1-p5-scalef-div-x112.c",
3525 "src/f32-vsigmoid/gen/vsigmoid-avx512f-rr1-p5-scalef-div-x128.c",
3526 "src/f32-vsigmoid/gen/vsigmoid-avx512f-rr1-p5-scalef-nr1fma-x16.c",
3527 "src/f32-vsigmoid/gen/vsigmoid-avx512f-rr1-p5-scalef-nr1fma-x32.c",
3528 "src/f32-vsigmoid/gen/vsigmoid-avx512f-rr1-p5-scalef-nr1fma-x48.c",
3529 "src/f32-vsigmoid/gen/vsigmoid-avx512f-rr1-p5-scalef-nr1fma-x64.c",
3530 "src/f32-vsigmoid/gen/vsigmoid-avx512f-rr1-p5-scalef-nr1fma-x80.c",
3531 "src/f32-vsigmoid/gen/vsigmoid-avx512f-rr1-p5-scalef-nr1fma-x96.c",
3532 "src/f32-vsigmoid/gen/vsigmoid-avx512f-rr1-p5-scalef-nr1fma-x112.c",
3533 "src/f32-vsigmoid/gen/vsigmoid-avx512f-rr1-p5-scalef-nr1fma-x128.c",
3534 "src/f32-vsigmoid/gen/vsigmoid-avx512f-rr2-lut32-p2-perm2-scalef-div-x16.c",
3535 "src/f32-vsigmoid/gen/vsigmoid-avx512f-rr2-lut32-p2-perm2-scalef-div-x32.c",
3536 "src/f32-vsigmoid/gen/vsigmoid-avx512f-rr2-lut32-p2-perm2-scalef-div-x48.c",
3537 "src/f32-vsigmoid/gen/vsigmoid-avx512f-rr2-lut32-p2-perm2-scalef-div-x64.c",
3538 "src/f32-vsigmoid/gen/vsigmoid-avx512f-rr2-lut32-p2-perm2-scalef-div-x80.c",
3539 "src/f32-vsigmoid/gen/vsigmoid-avx512f-rr2-lut32-p2-perm2-scalef-div-x96.c",
3540 "src/f32-vsigmoid/gen/vsigmoid-avx512f-rr2-lut32-p2-perm2-scalef-div-x112.c",
3541 "src/f32-vsigmoid/gen/vsigmoid-avx512f-rr2-lut32-p2-perm2-scalef-div-x128.c",
3542 "src/f32-vsigmoid/gen/vsigmoid-avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x16.c",
3543 "src/f32-vsigmoid/gen/vsigmoid-avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x32.c",
3544 "src/f32-vsigmoid/gen/vsigmoid-avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x48.c",
3545 "src/f32-vsigmoid/gen/vsigmoid-avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x64.c",
3546 "src/f32-vsigmoid/gen/vsigmoid-avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x80.c",
3547 "src/f32-vsigmoid/gen/vsigmoid-avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x96.c",
3548 "src/f32-vsigmoid/gen/vsigmoid-avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x112.c",
3549 "src/f32-vsigmoid/gen/vsigmoid-avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x128.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07003550 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x16.c",
3551 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x32.c",
3552 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x48.c",
3553 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x64.c",
3554 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x80.c",
3555 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x96.c",
3556 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x112.c",
3557 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x128.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07003558 "src/f32-vunary/gen/vabs-avx512f-x16.c",
3559 "src/f32-vunary/gen/vabs-avx512f-x32.c",
3560 "src/f32-vunary/gen/vneg-avx512f-x16.c",
3561 "src/f32-vunary/gen/vneg-avx512f-x32.c",
3562 "src/f32-vunary/gen/vsqr-avx512f-x16.c",
3563 "src/f32-vunary/gen/vsqr-avx512f-x32.c",
Marat Dukhanb7633f22020-11-20 16:34:56 -08003564 "src/math/exp-avx512f-rr2-lut16-p3-perm-scalef.c",
3565 "src/math/exp-avx512f-rr2-lut16-p3-perm.c",
3566 "src/math/exp-avx512f-rr2-lut32-p2-perm2-scalef.c",
3567 "src/math/exp-avx512f-rr2-lut32-p2-perm2.c",
3568 "src/math/exp-avx512f-rr2-p5-scalef.c",
3569 "src/math/exp-avx512f-rr2-p5.c",
Marat Dukhande390d42020-11-29 19:32:18 -08003570 "src/math/expm1minus-avx512f-rr1-lut16-p3-perm.c",
3571 "src/math/expm1minus-avx512f-rr1-p6.c",
Marat Dukhan98ba4412019-10-23 02:14:28 -07003572 "src/math/extexp-avx512f-p5.c",
Marat Dukhan6a34c5f2020-09-22 21:44:15 -07003573 "src/math/sigmoid-avx512f-rr1-lut16-p3-perm-scalef-div.c",
Marat Dukhan6a34c5f2020-09-22 21:44:15 -07003574 "src/math/sigmoid-avx512f-rr1-lut16-p3-perm-scalef-nr1fma.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003575 "src/math/sigmoid-avx512f-rr1-lut16-p3-perm-scalef-nr1fma1adj.c",
Marat Dukhan6a34c5f2020-09-22 21:44:15 -07003576 "src/math/sigmoid-avx512f-rr1-lut32-p2-perm2-scalef-div.c",
Marat Dukhan6a34c5f2020-09-22 21:44:15 -07003577 "src/math/sigmoid-avx512f-rr1-lut32-p2-perm2-scalef-nr1fma.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003578 "src/math/sigmoid-avx512f-rr1-lut32-p2-perm2-scalef-nr1fma1adj.c",
Marat Dukhan36173d22020-10-15 17:14:26 -07003579 "src/math/sigmoid-avx512f-rr1-lut64-p2-gather-scalef-div.c",
Marat Dukhan36173d22020-10-15 17:14:26 -07003580 "src/math/sigmoid-avx512f-rr1-lut64-p2-gather-scalef-nr1fma.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003581 "src/math/sigmoid-avx512f-rr1-lut64-p2-gather-scalef-nr1fma1adj.c",
3582 "src/math/sigmoid-avx512f-rr1-p5-scalef-div.c",
3583 "src/math/sigmoid-avx512f-rr1-p5-scalef-nr1fma.c",
3584 "src/math/sigmoid-avx512f-rr1-p5-scalef-nr1fma1adj.c",
3585 "src/math/sigmoid-avx512f-rr2-lut16-p3-perm-scalef-div.c",
3586 "src/math/sigmoid-avx512f-rr2-lut16-p3-perm-scalef-nr1fma.c",
3587 "src/math/sigmoid-avx512f-rr2-lut16-p3-perm-scalef-nr1fma1adj.c",
3588 "src/math/sigmoid-avx512f-rr2-lut32-p2-perm2-scalef-div.c",
3589 "src/math/sigmoid-avx512f-rr2-lut32-p2-perm2-scalef-nr1fma.c",
3590 "src/math/sigmoid-avx512f-rr2-lut32-p2-perm2-scalef-nr1fma1adj.c",
Marat Dukhan36173d22020-10-15 17:14:26 -07003591 "src/math/sigmoid-avx512f-rr2-lut64-p2-gather-scalef-div.c",
Marat Dukhan36173d22020-10-15 17:14:26 -07003592 "src/math/sigmoid-avx512f-rr2-lut64-p2-gather-scalef-nr1fma.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003593 "src/math/sigmoid-avx512f-rr2-lut64-p2-gather-scalef-nr1fma1adj.c",
3594 "src/math/sigmoid-avx512f-rr2-p5-scalef-div.c",
3595 "src/math/sigmoid-avx512f-rr2-p5-scalef-nr1fma.c",
3596 "src/math/sigmoid-avx512f-rr2-p5-scalef-nr1fma1adj.c",
Marat Dukhan84000762020-06-29 18:38:43 -07003597 "src/math/sqrt-avx512f-nr1fma.c",
Marat Dukhan84000762020-06-29 18:38:43 -07003598 "src/math/sqrt-avx512f-nr1fma1adj.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003599 "src/math/sqrt-avx512f-nr2fma.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003600]
3601
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07003602AVX512SKX_UKERNELS = [
Marat Dukhan2ffc5e62020-09-06 22:33:38 -07003603 "src/qs8-dwconv/gen/up16x9-minmax-avx512skx-mul32.c",
3604 "src/qs8-dwconv/gen/up32x9-minmax-avx512skx-mul32.c",
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07003605 "src/qs8-gemm/gen/1x16c8-minmax-avx512skx.c",
3606 "src/qs8-gemm/gen/2x16c8-minmax-avx512skx.c",
3607 "src/qs8-gemm/gen/3x16c8-minmax-avx512skx.c",
3608 "src/qs8-gemm/gen/4x16c8-minmax-avx512skx.c",
3609 "src/qs8-igemm/gen/1x16c8-minmax-avx512skx.c",
3610 "src/qs8-igemm/gen/2x16c8-minmax-avx512skx.c",
3611 "src/qs8-igemm/gen/3x16c8-minmax-avx512skx.c",
3612 "src/qs8-igemm/gen/4x16c8-minmax-avx512skx.c",
3613]
3614
Frank Barchardbcedc082020-08-17 18:00:51 -07003615WASM32_ASM_UKERNELS = [
Marat Dukhan6674d692021-05-05 22:27:00 -07003616 "src/f32-vrelu/wasm_shr_x1.S",
3617 "src/f32-vrelu/wasm_shr_x2.S",
3618 "src/f32-vrelu/wasm_shr_x4.S",
Frank Barchardbcedc082020-08-17 18:00:51 -07003619]
3620
Marat Dukhan08c4a432019-10-03 09:29:21 -07003621AARCH32_ASM_UKERNELS = [
Marat Dukhan32f93812020-05-17 20:31:21 -07003622 "src/f32-gemm/4x4-aarch32-vfp-ld64.S",
Marat Dukhan3b98f6b2020-05-17 10:09:22 -07003623 "src/f32-gemm/4x4-minmax-aarch32-vfp-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003624 "src/f32-gemm/4x8-minmax-aarch32-neon-cortex-a53.S",
3625 "src/f32-gemm/4x8-minmax-aarch32-neon-cortex-a55.S",
Frank Barchard490febe2020-07-16 18:42:17 -07003626 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-cortex-a7.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003627 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-cortex-a75.S",
Frank Barchard569561d2020-06-17 13:11:12 -07003628 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-ld64.S",
Frank Barchard490febe2020-07-16 18:42:17 -07003629 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-pld-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003630 "src/f32-igemm/4x8-minmax-aarch32-neon-cortex-a53.S",
3631 "src/f32-igemm/4x8-minmax-aarch32-neon-cortex-a55.S",
Frank Barchard490febe2020-07-16 18:42:17 -07003632 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-cortex-a7.S",
3633 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-cortex-a75.S",
3634 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-ld64.S",
3635 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-pld-cortex-a75.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003636]
3637
3638AARCH64_ASM_UKERNELS = [
Frank Barchardbddfbcd2020-04-15 12:32:41 -07003639 "src/f16-gemm/gen-inc/1x8inc-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchard04336c12020-10-22 16:48:55 -07003640 "src/f16-gemm/gen-inc/1x16inc-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchardbddfbcd2020-04-15 12:32:41 -07003641 "src/f16-gemm/gen-inc/4x8inc-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchard04336c12020-10-22 16:48:55 -07003642 "src/f16-gemm/gen-inc/4x16inc-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchardbddfbcd2020-04-15 12:32:41 -07003643 "src/f16-gemm/gen-inc/6x8inc-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchard04336c12020-10-22 16:48:55 -07003644 "src/f16-gemm/gen-inc/6x16inc-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchard3b8e5662020-04-20 12:12:53 -07003645 "src/f16-gemm/gen-inc/8x8inc-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchard04336c12020-10-22 16:48:55 -07003646 "src/f16-gemm/gen/1x8-minmax-aarch64-neonfp16arith-ld64.S",
3647 "src/f16-gemm/gen/1x16-minmax-aarch64-neonfp16arith-ld32.S",
3648 "src/f16-gemm/gen/4x8-minmax-aarch64-neonfp16arith-ld64.S",
3649 "src/f16-gemm/gen/4x16-minmax-aarch64-neonfp16arith-ld32.S",
3650 "src/f16-gemm/gen/6x8-minmax-aarch64-neonfp16arith-ld64.S",
3651 "src/f16-gemm/gen/6x16-minmax-aarch64-neonfp16arith-ld32.S",
3652 "src/f16-gemm/gen/8x8-minmax-aarch64-neonfp16arith-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003653 "src/f32-dwconv/up4x9-minmax-aarch64-neonfma-cortex-a55.S",
3654 "src/f32-dwconv/up4x9-minmax-aarch64-neonfma.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003655 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a53.S",
3656 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a57.S",
3657 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a75.S",
Frank Barchard04336c12020-10-22 16:48:55 -07003658 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-ld64.S",
3659 "src/f32-gemm/gen-inc/1x12inc-minmax-aarch64-neonfma-cortex-a53.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003660 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a53.S",
3661 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a55.S",
3662 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a57.S",
3663 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003664 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-ld64.S",
Frank Barchard04336c12020-10-22 16:48:55 -07003665 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-ld128.S",
3666 "src/f32-gemm/gen-inc/4x12inc-minmax-aarch64-neonfma-cortex-a53.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003667 "src/f32-gemm/gen-inc/5x8inc-minmax-aarch64-neonfma-cortex-a57.S",
3668 "src/f32-gemm/gen-inc/5x8inc-minmax-aarch64-neonfma-cortex-a75.S",
3669 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a53.S",
3670 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a55.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003671 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a57.S",
Frank Barchard04336c12020-10-22 16:48:55 -07003672 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a73.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003673 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003674 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-ld64.S",
Frank Barchard04336c12020-10-22 16:48:55 -07003675 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-ld128.S",
3676 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a53.S",
3677 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a57.S",
3678 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a75.S",
3679 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-ld64.S",
3680 "src/f32-gemm/gen/1x12-minmax-aarch64-neonfma-cortex-a53.S",
3681 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a53.S",
3682 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a55.S",
3683 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a57.S",
3684 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a75.S",
3685 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-ld64.S",
3686 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-ld128.S",
3687 "src/f32-gemm/gen/4x12-minmax-aarch64-neonfma-cortex-a53.S",
3688 "src/f32-gemm/gen/5x8-minmax-aarch64-neonfma-cortex-a57.S",
3689 "src/f32-gemm/gen/5x8-minmax-aarch64-neonfma-cortex-a75.S",
3690 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a53.S",
3691 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a55.S",
3692 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a57.S",
3693 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a73.S",
3694 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a75.S",
3695 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-ld64.S",
3696 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-ld128.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003697 "src/f32-igemm/1x8-minmax-aarch64-neonfma-cortex-a53.S",
Frank Barchard04336c12020-10-22 16:48:55 -07003698 "src/f32-igemm/1x12-minmax-aarch64-neonfma-cortex-a53.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003699 "src/f32-igemm/4x8-minmax-aarch64-neonfma-cortex-a53.S",
3700 "src/f32-igemm/4x8-minmax-aarch64-neonfma-cortex-a55.S",
Frank Barchard04336c12020-10-22 16:48:55 -07003701 "src/f32-igemm/4x12-minmax-aarch64-neonfma-cortex-a53.S",
3702 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a53.S",
3703 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a55.S",
3704 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a73.S",
3705 "src/f32-igemm/gen/1x8-minmax-aarch64-neonfma-cortex-a57.S",
3706 "src/f32-igemm/gen/1x8-minmax-aarch64-neonfma-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003707 "src/f32-igemm/gen/4x8-minmax-aarch64-neonfma-cortex-a57.S",
3708 "src/f32-igemm/gen/4x8-minmax-aarch64-neonfma-cortex-a75.S",
3709 "src/f32-igemm/gen/5x8-minmax-aarch64-neonfma-cortex-a57.S",
3710 "src/f32-igemm/gen/5x8-minmax-aarch64-neonfma-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003711 "src/f32-igemm/gen/6x8-minmax-aarch64-neonfma-cortex-a57.S",
3712 "src/f32-igemm/gen/6x8-minmax-aarch64-neonfma-cortex-a75.S",
Frank Barchard4a4be4e2021-03-01 15:00:28 -08003713 "src/qs8-gemm/1x16c4-aarch64-neondot-ld32.S",
Frank Barchardf1fd89e2020-10-06 23:27:11 -07003714 "src/qs8-gemm/1x16c4-aarch64-neondot-ld64.S",
Frank Barchard1dc9fef2021-03-10 23:50:19 -08003715 "src/qs8-gemm/2x8c8-aarch64-neon-mull-padal.S",
Frank Barchard5655cb72021-03-09 20:57:14 -08003716 "src/qs8-gemm/2x8c16-aarch64-neon-mlal-padal.S",
Frank Barchard671d1b02021-03-10 15:42:08 -08003717 "src/qs8-gemm/4x16c4-aarch64-neondot-cortex-a55.S",
Frank Barchardda78da12021-03-02 14:28:00 -08003718 "src/qs8-gemm/4x16c4-aarch64-neondot-ld32.S",
Frank Barchard671d1b02021-03-10 15:42:08 -08003719 "src/qs8-gemm/4x16c4-aarch64-neondot-ld64.S",
Frank Barchard938ea812021-05-07 13:41:35 -07003720 "src/qs8-gemm/gen/1x8c8-aarch64-neon-mlal-padal.S",
3721 "src/qs8-gemm/gen/1x8c8-aarch64-neon-mlal-padal-prfm.S",
Frank Barcharddddb38f2021-05-03 16:51:48 -07003722 "src/qs8-gemm/gen/1x8c8-aarch64-neon-mlal-padal-cortex-a53.S",
3723 "src/qs8-gemm/gen/1x8c8-aarch64-neon-mlal-padal-prfm-cortex-a53.S",
3724 "src/qs8-gemm/gen/2x8c8-aarch64-neon-mlal-padal-cortex-a53.S",
3725 "src/qs8-gemm/gen/2x8c8-aarch64-neon-mlal-padal-prfm-cortex-a53.S",
Frank Barchard4a352042021-04-13 15:52:08 -07003726 "src/qs8-gemm/gen/2x8c8-aarch64-neon-mlal-padal-prfm.S",
3727 "src/qs8-gemm/gen/2x8c8-aarch64-neon-mlal-padal.S",
Frank Barchard7ca54df2021-03-12 00:09:02 -08003728 "src/qs8-igemm/2x8c16-aarch64-neon-mlal-padal.S",
Frank Barchard89e12f82021-03-08 11:42:44 -08003729 "src/qs8-igemm/4x16c4-aarch64-neondot-cortex-a55.S",
Frank Barchard671d1b02021-03-10 15:42:08 -08003730 "src/qs8-igemm/4x16c4-aarch64-neondot-ld64.S",
Frank Barchard938ea812021-05-07 13:41:35 -07003731 "src/qs8-igemm/gen/1x8c8-aarch64-neon-mlal-padal.S",
3732 "src/qs8-igemm/gen/1x8c8-aarch64-neon-mlal-padal-prfm.S",
Frank Barcharddddb38f2021-05-03 16:51:48 -07003733 "src/qs8-igemm/gen/1x8c8-aarch64-neon-mlal-padal-cortex-a53.S",
3734 "src/qs8-igemm/gen/1x8c8-aarch64-neon-mlal-padal-prfm-cortex-a53.S",
3735 "src/qs8-igemm/gen/2x8c8-aarch64-neon-mlal-padal-cortex-a53.S",
3736 "src/qs8-igemm/gen/2x8c8-aarch64-neon-mlal-padal-prfm-cortex-a53.S",
Frank Barchard4a352042021-04-13 15:52:08 -07003737 "src/qs8-igemm/gen/2x8c8-aarch64-neon-mlal-padal-prfm.S",
3738 "src/qs8-igemm/gen/2x8c8-aarch64-neon-mlal-padal.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003739]
3740
Marat Dukhan1b354632020-03-23 12:50:22 -07003741INTERNAL_MICROKERNEL_HDRS = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003742 "src/xnnpack/argmaxpool.h",
3743 "src/xnnpack/avgpool.h",
3744 "src/xnnpack/clamp.h",
3745 "src/xnnpack/common.h",
3746 "src/xnnpack/conv.h",
Yury Kartynnike7841862020-11-04 18:22:18 -08003747 "src/xnnpack/depthtospace.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003748 "src/xnnpack/dwconv.h",
Frank Barchard04336c12020-10-22 16:48:55 -07003749 "src/xnnpack/fill.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003750 "src/xnnpack/gavgpool.h",
3751 "src/xnnpack/gemm.h",
Marat Dukhan660fd192020-03-10 04:55:30 -07003752 "src/xnnpack/ibilinear.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003753 "src/xnnpack/igemm.h",
Marat Dukhancfb31342019-12-05 10:42:57 -08003754 "src/xnnpack/intrinsics-polyfill.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003755 "src/xnnpack/lut.h",
3756 "src/xnnpack/math.h",
3757 "src/xnnpack/maxpool.h",
3758 "src/xnnpack/packx.h",
3759 "src/xnnpack/pad.h",
3760 "src/xnnpack/params.h",
3761 "src/xnnpack/pavgpool.h",
3762 "src/xnnpack/ppmm.h",
3763 "src/xnnpack/prelu.h",
Marat Dukhan97579532019-10-18 16:40:39 -07003764 "src/xnnpack/raddexpminusmax.h",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07003765 "src/xnnpack/raddextexp.h",
Marat Dukhan97579532019-10-18 16:40:39 -07003766 "src/xnnpack/raddstoreexpminusmax.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003767 "src/xnnpack/rmax.h",
3768 "src/xnnpack/scalar-utils.h",
3769 "src/xnnpack/spmm.h",
3770 "src/xnnpack/unpool.h",
3771 "src/xnnpack/vadd.h",
Marat Dukhan1e782c42019-11-21 17:02:40 -08003772 "src/xnnpack/vbinary.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003773 "src/xnnpack/vmulcaddc.h",
Marat Dukhan05ac8e32019-10-21 15:39:33 -07003774 "src/xnnpack/vscale.h",
Marat Dukhan97579532019-10-18 16:40:39 -07003775 "src/xnnpack/vscaleexpminusmax.h",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07003776 "src/xnnpack/vscaleextexp.h",
Marat Dukhan1e782c42019-11-21 17:02:40 -08003777 "src/xnnpack/vunary.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003778 "src/xnnpack/zip.h",
Marat Dukhan1b354632020-03-23 12:50:22 -07003779]
3780
3781INTERNAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003782 "include/xnnpack.h",
3783 "src/xnnpack/allocator.h",
3784 "src/xnnpack/compute.h",
3785 "src/xnnpack/im2col.h",
3786 "src/xnnpack/indirection.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003787 "src/xnnpack/math-stubs.h",
Chao Mei6ddfc602020-05-13 22:29:36 -07003788 "src/xnnpack/memory-planner.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003789 "src/xnnpack/operator.h",
3790 "src/xnnpack/pack.h",
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07003791 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003792 "src/xnnpack/requantization-stubs.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003793 "src/xnnpack/requantization.h",
Marat Dukhan1d75a542020-02-03 12:23:01 -08003794 "src/xnnpack/subgraph.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003795]
3796
Marat Dukhan1b354632020-03-23 12:50:22 -07003797ACCURACY_EVAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003798 "src/xnnpack/math-stubs.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003799]
3800
Marat Dukhan1b354632020-03-23 12:50:22 -07003801MICROKERNEL_BENCHMARK_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003802 "include/xnnpack.h",
Frank Barchard35db7d02020-10-26 13:37:34 -07003803 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003804]
3805
Marat Dukhan1b354632020-03-23 12:50:22 -07003806MICROKERNEL_TEST_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Frank Barchard35db7d02020-10-26 13:37:34 -07003807 "include/xnnpack.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003808 "src/xnnpack/isa-checks.h",
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07003809 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003810 "src/xnnpack/requantization.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003811]
3812
3813OPERATOR_TEST_PARAMS_HDRS = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003814 "src/xnnpack/common.h",
Frank Barchard04336c12020-10-22 16:48:55 -07003815 "src/xnnpack/params.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003816]
3817
3818WEIGHTS_PACK_HDRS = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003819 "src/xnnpack/compute.h",
Frank Barchard04336c12020-10-22 16:48:55 -07003820 "src/xnnpack/operator.h",
3821 "src/xnnpack/pack.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003822]
3823
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07003824LOGGING_COPTS = select({
3825 # No logging in optimized mode
3826 ":optimized_build": ["-DXNN_LOG_LEVEL=0"],
3827 # Full logging in debug mode
3828 ":debug_build": ["-DXNN_LOG_LEVEL=5"],
3829 # Error-only logging in default (fastbuild) mode
3830 "//conditions:default": ["-DXNN_LOG_LEVEL=2"],
3831})
3832
Marat Dukhan3b59de22020-06-03 20:15:19 -07003833LOGGING_SRCS = select({
3834 # No logging in optimized mode
3835 ":optimized_build": [],
3836 "//conditions:default": [
Marat Dukhanccd3a1d2021-03-29 16:03:12 -07003837 "src/datatype-strings.c",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003838 "src/operator-strings.c",
3839 "src/subgraph-strings.c",
3840 ],
3841})
3842
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07003843LOGGING_HDRS = [
3844 "src/xnnpack/log.h",
3845]
3846
Marat Dukhan08c4a432019-10-03 09:29:21 -07003847xnnpack_cc_library(
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003848 name = "tables",
3849 srcs = TABLE_SRCS,
3850 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003851 gcc_copts = xnnpack_gcc_std_copts(),
3852 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003853)
3854
3855xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003856 name = "scalar_ukernels",
3857 srcs = SCALAR_UKERNELS,
3858 hdrs = INTERNAL_HDRS,
3859 aarch32_copts = ["-marm"],
Marat Dukhan10a38082020-04-17 03:58:35 -07003860 gcc_copts = xnnpack_gcc_std_copts(),
3861 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003862 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003863 ":tables",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003864 "@FP16",
3865 "@FXdiv",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003866 "@pthreadpool",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003867 ],
3868)
3869
3870xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003871 name = "scalar_ukernels_test_mode",
3872 srcs = SCALAR_UKERNELS,
3873 hdrs = INTERNAL_HDRS,
3874 aarch32_copts = ["-marm"],
3875 copts = [
3876 "-UNDEBUG",
3877 "-DXNN_TEST_MODE=1",
3878 ],
3879 gcc_copts = xnnpack_gcc_std_copts(),
3880 msvc_copts = xnnpack_msvc_std_copts(),
3881 deps = [
3882 ":tables",
3883 "@FP16",
3884 "@FXdiv",
3885 "@pthreadpool",
3886 ],
3887)
3888
3889xnnpack_cc_library(
Marat Dukhan436ebe62019-12-04 15:10:12 -08003890 name = "wasm_ukernels",
3891 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003892 gcc_copts = xnnpack_gcc_std_copts(),
3893 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan436ebe62019-12-04 15:10:12 -08003894 wasm_srcs = WASM_UKERNELS,
Marat Dukhan1c6e3892020-06-25 23:56:10 -07003895 wasmsimd_srcs = WASM_UKERNELS + WASMSIMD_UKERNELS,
Marat Dukhan436ebe62019-12-04 15:10:12 -08003896 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003897 ":tables",
Marat Dukhan436ebe62019-12-04 15:10:12 -08003898 "@FP16",
3899 "@FXdiv",
3900 "@pthreadpool",
3901 ],
3902)
3903
3904xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003905 name = "wasm_ukernels_test_mode",
3906 hdrs = INTERNAL_HDRS,
3907 copts = [
3908 "-UNDEBUG",
3909 "-DXNN_TEST_MODE=1",
3910 ],
3911 gcc_copts = xnnpack_gcc_std_copts(),
3912 msvc_copts = xnnpack_msvc_std_copts(),
3913 wasm_srcs = WASM_UKERNELS,
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07003914 wasmsimd_srcs = WASM_UKERNELS + WASMSIMD_UKERNELS,
Marat Dukhan33fcf782020-05-24 14:27:15 -07003915 deps = [
3916 ":tables",
3917 "@FP16",
3918 "@FXdiv",
3919 "@pthreadpool",
3920 ],
3921)
3922
3923xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003924 name = "neon_ukernels",
3925 hdrs = INTERNAL_HDRS,
3926 aarch32_copts = [
3927 "-marm",
Marat Dukhan8853b822020-05-07 12:19:01 -07003928 "-march=armv7-a",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003929 "-mfpu=neon",
3930 ],
3931 aarch32_srcs = NEON_UKERNELS,
3932 aarch64_srcs = NEON_UKERNELS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003933 gcc_copts = xnnpack_gcc_std_copts(),
3934 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08003935 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003936 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003937 "@FP16",
3938 "@pthreadpool",
3939 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003940)
3941
3942xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003943 name = "neon_ukernels_test_mode",
3944 hdrs = INTERNAL_HDRS,
3945 aarch32_copts = [
3946 "-marm",
3947 "-march=armv7-a",
3948 "-mfpu=neon",
3949 ],
3950 aarch32_srcs = NEON_UKERNELS,
3951 aarch64_srcs = NEON_UKERNELS,
3952 copts = [
3953 "-UNDEBUG",
3954 "-DXNN_TEST_MODE=1",
3955 ],
3956 gcc_copts = xnnpack_gcc_std_copts(),
3957 msvc_copts = xnnpack_msvc_std_copts(),
3958 deps = [
3959 ":tables",
3960 "@FP16",
3961 "@pthreadpool",
3962 ],
3963)
3964
3965xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003966 name = "neonfma_ukernels",
3967 hdrs = INTERNAL_HDRS,
3968 aarch32_copts = [
3969 "-marm",
Marat Dukhan8853b822020-05-07 12:19:01 -07003970 "-march=armv7-a",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003971 "-mfpu=neon-vfpv4",
3972 ],
3973 aarch32_srcs = NEONFMA_UKERNELS,
3974 aarch64_srcs = NEONFMA_UKERNELS + AARCH64_NEONFMA_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07003975 apple_aarch32_copts = [
3976 "-mcpu=swift",
3977 "-mtune=generic",
3978 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07003979 gcc_copts = xnnpack_gcc_std_copts(),
3980 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08003981 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003982 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003983 "@FP16",
3984 "@pthreadpool",
3985 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003986)
3987
3988xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003989 name = "neonfma_ukernels_test_mode",
3990 hdrs = INTERNAL_HDRS,
3991 aarch32_copts = [
3992 "-marm",
3993 "-march=armv7-a",
3994 "-mfpu=neon-vfpv4",
3995 ],
3996 aarch32_srcs = NEONFMA_UKERNELS,
3997 aarch64_srcs = NEONFMA_UKERNELS + AARCH64_NEONFMA_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07003998 apple_aarch32_copts = [
3999 "-mcpu=swift",
4000 "-mtune=generic",
4001 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07004002 copts = [
4003 "-UNDEBUG",
4004 "-DXNN_TEST_MODE=1",
4005 ],
4006 gcc_copts = xnnpack_gcc_std_copts(),
4007 msvc_copts = xnnpack_msvc_std_copts(),
4008 deps = [
4009 ":tables",
4010 "@FP16",
4011 "@pthreadpool",
4012 ],
4013)
4014
4015xnnpack_cc_library(
Marat Dukhan8853b822020-05-07 12:19:01 -07004016 name = "neonv8_ukernels",
4017 hdrs = INTERNAL_HDRS,
4018 aarch32_copts = [
4019 "-marm",
4020 "-march=armv8-a",
4021 "-mfpu=neon-fp-armv8",
4022 ],
4023 aarch32_srcs = NEONV8_UKERNELS,
4024 aarch64_srcs = NEONV8_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07004025 apple_aarch32_copts = [
4026 "-mcpu=cyclone",
4027 "-mtune=generic",
4028 ],
Marat Dukhan8853b822020-05-07 12:19:01 -07004029 gcc_copts = xnnpack_gcc_std_copts(),
4030 msvc_copts = xnnpack_msvc_std_copts(),
4031 deps = [
4032 ":tables",
4033 "@FP16",
4034 "@pthreadpool",
4035 ],
4036)
4037
4038xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07004039 name = "neonv8_ukernels_test_mode",
4040 hdrs = INTERNAL_HDRS,
4041 aarch32_copts = [
4042 "-marm",
4043 "-march=armv8-a",
4044 "-mfpu=neon-fp-armv8",
4045 ],
4046 aarch32_srcs = NEONV8_UKERNELS,
4047 aarch64_srcs = NEONV8_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07004048 apple_aarch32_copts = [
4049 "-mcpu=cyclone",
4050 "-mtune=generic",
4051 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07004052 copts = [
4053 "-UNDEBUG",
4054 "-DXNN_TEST_MODE=1",
4055 ],
4056 gcc_copts = xnnpack_gcc_std_copts(),
4057 msvc_copts = xnnpack_msvc_std_copts(),
4058 deps = [
4059 ":tables",
4060 "@FP16",
4061 "@pthreadpool",
4062 ],
4063)
4064
4065xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004066 name = "neonfp16arith_ukernels",
4067 hdrs = INTERNAL_HDRS,
4068 aarch64_copts = ["-march=armv8.2-a+fp16"],
4069 aarch64_srcs = AARCH64_NEONFP16ARITH_UKERNELS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004070 gcc_copts = xnnpack_gcc_std_copts(),
4071 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08004072 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08004073 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08004074 "@FP16",
4075 "@pthreadpool",
4076 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004077)
4078
4079xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07004080 name = "neonfp16arith_ukernels_test_mode",
4081 hdrs = INTERNAL_HDRS,
4082 aarch64_copts = ["-march=armv8.2-a+fp16"],
4083 aarch64_srcs = AARCH64_NEONFP16ARITH_UKERNELS,
4084 copts = [
4085 "-UNDEBUG",
4086 "-DXNN_TEST_MODE=1",
4087 ],
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(
Benoit Jacoba9644732020-08-13 12:48:55 -07004098 name = "neondot_ukernels",
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 gcc_copts = xnnpack_gcc_std_copts(),
4109 msvc_copts = xnnpack_msvc_std_copts(),
4110 deps = [
4111 ":tables",
4112 "@FP16",
4113 "@pthreadpool",
4114 ],
4115)
4116
4117xnnpack_cc_library(
4118 name = "neondot_ukernels_test_mode",
4119 hdrs = INTERNAL_HDRS,
Marat Dukhan799ac752020-08-13 16:08:03 -07004120 aarch32_copts = [
4121 "-marm",
4122 "-march=armv8.2-a+dotprod",
4123 "-mfpu=neon-fp-armv8",
4124 ],
Benoit Jacoba9644732020-08-13 12:48:55 -07004125 aarch32_srcs = NEONDOT_UKERNELS,
4126 aarch64_copts = ["-march=armv8.2-a+dotprod"],
4127 aarch64_srcs = NEONDOT_UKERNELS,
4128 copts = [
4129 "-UNDEBUG",
4130 "-DXNN_TEST_MODE=1",
4131 ],
4132 gcc_copts = xnnpack_gcc_std_copts(),
4133 msvc_copts = xnnpack_msvc_std_copts(),
4134 deps = [
4135 ":tables",
4136 "@FP16",
4137 "@pthreadpool",
4138 ],
4139)
4140
4141xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004142 name = "sse2_ukernels",
4143 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004144 gcc_copts = xnnpack_gcc_std_copts(),
4145 gcc_x86_copts = ["-msse2"],
4146 msvc_copts = xnnpack_msvc_std_copts(),
4147 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004148 x86_srcs = SSE_UKERNELS + SSE2_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08004149 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08004150 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08004151 "@FP16",
4152 "@pthreadpool",
4153 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004154)
4155
4156xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07004157 name = "sse2_ukernels_test_mode",
4158 hdrs = INTERNAL_HDRS,
4159 copts = [
4160 "-UNDEBUG",
4161 "-DXNN_TEST_MODE=1",
4162 ],
4163 gcc_copts = xnnpack_gcc_std_copts(),
4164 gcc_x86_copts = ["-msse2"],
4165 msvc_copts = xnnpack_msvc_std_copts(),
4166 msvc_x86_32_copts = ["/arch:SSE2"],
4167 x86_srcs = SSE_UKERNELS + SSE2_UKERNELS,
4168 deps = [
4169 ":tables",
4170 "@FP16",
4171 "@pthreadpool",
4172 ],
4173)
4174
4175xnnpack_cc_library(
Marat Dukhanfe7acb62020-03-09 19:30:05 -07004176 name = "ssse3_ukernels",
4177 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004178 gcc_copts = xnnpack_gcc_std_copts(),
4179 gcc_x86_copts = ["-mssse3"],
4180 msvc_copts = xnnpack_msvc_std_copts(),
4181 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhanfe7acb62020-03-09 19:30:05 -07004182 x86_srcs = SSSE3_UKERNELS,
4183 deps = [
4184 ":tables",
4185 "@FP16",
4186 "@pthreadpool",
4187 ],
4188)
4189
4190xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07004191 name = "ssse3_ukernels_test_mode",
4192 hdrs = INTERNAL_HDRS,
4193 copts = [
4194 "-UNDEBUG",
4195 "-DXNN_TEST_MODE=1",
4196 ],
4197 gcc_copts = xnnpack_gcc_std_copts(),
4198 gcc_x86_copts = ["-mssse3"],
4199 msvc_copts = xnnpack_msvc_std_copts(),
4200 msvc_x86_32_copts = ["/arch:SSE2"],
4201 x86_srcs = SSSE3_UKERNELS,
4202 deps = [
4203 ":tables",
4204 "@FP16",
4205 "@pthreadpool",
4206 ],
4207)
4208
4209xnnpack_cc_library(
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08004210 name = "sse41_ukernels",
4211 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004212 gcc_copts = xnnpack_gcc_std_copts(),
4213 gcc_x86_copts = ["-msse4.1"],
4214 msvc_copts = xnnpack_msvc_std_copts(),
4215 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08004216 x86_srcs = SSE41_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08004217 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08004218 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08004219 "@FP16",
4220 "@pthreadpool",
4221 ],
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08004222)
4223
4224xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07004225 name = "sse41_ukernels_test_mode",
4226 hdrs = INTERNAL_HDRS,
4227 copts = [
4228 "-UNDEBUG",
4229 "-DXNN_TEST_MODE=1",
4230 ],
4231 gcc_copts = xnnpack_gcc_std_copts(),
4232 gcc_x86_copts = ["-msse4.1"],
4233 msvc_copts = xnnpack_msvc_std_copts(),
4234 msvc_x86_32_copts = ["/arch:SSE2"],
4235 x86_srcs = SSE41_UKERNELS,
4236 deps = [
4237 ":tables",
4238 "@FP16",
4239 "@pthreadpool",
4240 ],
4241)
4242
4243xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004244 name = "avx_ukernels",
4245 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004246 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"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004251 x86_srcs = AVX_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08004252 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08004253 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08004254 "@FP16",
4255 "@pthreadpool",
4256 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004257)
4258
4259xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07004260 name = "avx_ukernels_test_mode",
4261 hdrs = INTERNAL_HDRS,
4262 copts = [
4263 "-UNDEBUG",
4264 "-DXNN_TEST_MODE=1",
4265 ],
4266 gcc_copts = xnnpack_gcc_std_copts(),
4267 gcc_x86_copts = ["-mavx"],
4268 msvc_copts = xnnpack_msvc_std_copts(),
4269 msvc_x86_32_copts = ["/arch:AVX"],
4270 msvc_x86_64_copts = ["/arch:AVX"],
4271 x86_srcs = AVX_UKERNELS,
4272 deps = [
4273 ":tables",
4274 "@FP16",
4275 "@pthreadpool",
4276 ],
4277)
4278
4279xnnpack_cc_library(
Marat Dukhan1566fee2020-08-02 21:55:41 -07004280 name = "xop_ukernels",
4281 hdrs = INTERNAL_HDRS,
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(
4296 name = "xop_ukernels_test_mode",
4297 hdrs = INTERNAL_HDRS,
4298 copts = [
4299 "-UNDEBUG",
4300 "-DXNN_TEST_MODE=1",
4301 ],
4302 gcc_copts = xnnpack_gcc_std_copts(),
4303 gcc_x86_copts = ["-mxop"],
4304 msvc_copts = xnnpack_msvc_std_copts(),
4305 msvc_x86_32_copts = ["/arch:AVX"],
4306 msvc_x86_64_copts = ["/arch:AVX"],
4307 x86_srcs = XOP_UKERNELS,
4308 deps = [
4309 ":tables",
4310 "@FP16",
4311 "@pthreadpool",
4312 ],
4313)
4314
4315xnnpack_cc_library(
Marat Dukhanfda12b82019-11-21 12:27:59 -08004316 name = "fma3_ukernels",
4317 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004318 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"],
Marat Dukhanfda12b82019-11-21 12:27:59 -08004323 x86_srcs = FMA3_UKERNELS,
4324 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08004325 ":tables",
Marat Dukhanfda12b82019-11-21 12:27:59 -08004326 "@FP16",
4327 "@pthreadpool",
4328 ],
4329)
4330
4331xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07004332 name = "fma3_ukernels_test_mode",
4333 hdrs = INTERNAL_HDRS,
4334 copts = [
4335 "-UNDEBUG",
4336 "-DXNN_TEST_MODE=1",
4337 ],
4338 gcc_copts = xnnpack_gcc_std_copts(),
4339 gcc_x86_copts = ["-mfma"],
4340 msvc_copts = xnnpack_msvc_std_copts(),
4341 msvc_x86_32_copts = ["/arch:AVX"],
4342 msvc_x86_64_copts = ["/arch:AVX"],
4343 x86_srcs = FMA3_UKERNELS,
4344 deps = [
4345 ":tables",
4346 "@FP16",
4347 "@pthreadpool",
4348 ],
4349)
4350
4351xnnpack_cc_library(
Marat Dukhan6adff4e2019-10-14 18:32:07 -07004352 name = "avx2_ukernels",
4353 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004354 gcc_copts = xnnpack_gcc_std_copts(),
4355 gcc_x86_copts = [
Marat Dukhan6adff4e2019-10-14 18:32:07 -07004356 "-mfma",
4357 "-mavx2",
4358 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07004359 msvc_copts = xnnpack_msvc_std_copts(),
4360 msvc_x86_32_copts = ["/arch:AVX2"],
4361 msvc_x86_64_copts = ["/arch:AVX2"],
Marat Dukhan6adff4e2019-10-14 18:32:07 -07004362 x86_srcs = AVX2_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08004363 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08004364 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08004365 "@FP16",
4366 "@pthreadpool",
4367 ],
Marat Dukhan6adff4e2019-10-14 18:32:07 -07004368)
4369
4370xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07004371 name = "avx2_ukernels_test_mode",
4372 hdrs = INTERNAL_HDRS,
4373 copts = [
4374 "-UNDEBUG",
4375 "-DXNN_TEST_MODE=1",
4376 ],
4377 gcc_copts = xnnpack_gcc_std_copts(),
4378 gcc_x86_copts = [
4379 "-mfma",
4380 "-mavx2",
4381 ],
4382 msvc_copts = xnnpack_msvc_std_copts(),
4383 msvc_x86_32_copts = ["/arch:AVX2"],
4384 msvc_x86_64_copts = ["/arch:AVX2"],
4385 x86_srcs = AVX2_UKERNELS,
4386 deps = [
4387 ":tables",
4388 "@FP16",
4389 "@pthreadpool",
4390 ],
4391)
4392
4393xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004394 name = "avx512f_ukernels",
4395 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004396 gcc_copts = xnnpack_gcc_std_copts(),
4397 gcc_x86_copts = ["-mavx512f"],
4398 mingw_copts = ["-fno-asynchronous-unwind-tables"],
4399 msvc_copts = xnnpack_msvc_std_copts(),
4400 msvc_x86_32_copts = ["/arch:AVX512"],
4401 msvc_x86_64_copts = ["/arch:AVX512"],
4402 msys_copts = ["-fno-asynchronous-unwind-tables"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004403 x86_srcs = AVX512F_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08004404 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08004405 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08004406 "@FP16",
4407 "@pthreadpool",
4408 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004409)
4410
4411xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07004412 name = "avx512f_ukernels_test_mode",
4413 hdrs = INTERNAL_HDRS,
4414 copts = [
4415 "-UNDEBUG",
4416 "-DXNN_TEST_MODE=1",
4417 ],
4418 gcc_copts = xnnpack_gcc_std_copts(),
4419 gcc_x86_copts = ["-mavx512f"],
4420 mingw_copts = ["-fno-asynchronous-unwind-tables"],
4421 msvc_copts = xnnpack_msvc_std_copts(),
4422 msvc_x86_32_copts = ["/arch:AVX512"],
4423 msvc_x86_64_copts = ["/arch:AVX512"],
4424 msys_copts = ["-fno-asynchronous-unwind-tables"],
4425 x86_srcs = AVX512F_UKERNELS,
4426 deps = [
4427 ":tables",
4428 "@FP16",
4429 "@pthreadpool",
4430 ],
4431)
4432
4433xnnpack_cc_library(
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07004434 name = "avx512skx_ukernels",
4435 hdrs = INTERNAL_HDRS,
4436 gcc_copts = xnnpack_gcc_std_copts(),
4437 gcc_x86_copts = [
4438 "-mavx512f",
4439 "-mavx512cd",
4440 "-mavx512bw",
4441 "-mavx512dq",
4442 "-mavx512vl",
4443 ],
4444 mingw_copts = ["-fno-asynchronous-unwind-tables"],
4445 msvc_copts = xnnpack_msvc_std_copts(),
4446 msvc_x86_32_copts = ["/arch:AVX512"],
4447 msvc_x86_64_copts = ["/arch:AVX512"],
4448 msys_copts = ["-fno-asynchronous-unwind-tables"],
4449 x86_srcs = AVX512SKX_UKERNELS,
4450 deps = [
4451 ":tables",
4452 "@FP16",
4453 "@pthreadpool",
4454 ],
4455)
4456
4457xnnpack_cc_library(
4458 name = "avx512skx_ukernels_test_mode",
4459 hdrs = INTERNAL_HDRS,
4460 copts = [
4461 "-UNDEBUG",
4462 "-DXNN_TEST_MODE=1",
4463 ],
4464 gcc_copts = xnnpack_gcc_std_copts(),
4465 gcc_x86_copts = [
4466 "-mavx512f",
4467 "-mavx512cd",
4468 "-mavx512bw",
4469 "-mavx512dq",
4470 "-mavx512vl",
4471 ],
4472 mingw_copts = ["-fno-asynchronous-unwind-tables"],
4473 msvc_copts = xnnpack_msvc_std_copts(),
4474 msvc_x86_32_copts = ["/arch:AVX512"],
4475 msvc_x86_64_copts = ["/arch:AVX512"],
4476 msys_copts = ["-fno-asynchronous-unwind-tables"],
4477 x86_srcs = AVX512SKX_UKERNELS,
4478 deps = [
4479 ":tables",
4480 "@FP16",
4481 "@pthreadpool",
4482 ],
4483)
4484
4485xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004486 name = "asm_ukernels",
4487 hdrs = ["src/xnnpack/assembly.h"],
4488 aarch32_srcs = AARCH32_ASM_UKERNELS,
Frank Barchard31bb45b2020-10-06 00:26:33 -07004489 aarch64_copts = ["-march=armv8.2-a+fp16+dotprod"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004490 aarch64_srcs = AARCH64_ASM_UKERNELS,
Frank Barchardbcedc082020-08-17 18:00:51 -07004491 wasm_srcs = WASM32_ASM_UKERNELS,
4492 wasmsimd_srcs = WASM32_ASM_UKERNELS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07004493)
4494
Marat Dukhan3b59de22020-06-03 20:15:19 -07004495xnnpack_cc_library(
4496 name = "logging_utils",
4497 srcs = LOGGING_SRCS,
4498 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
4499 copts = LOGGING_COPTS + [
4500 "-Isrc",
4501 "-Iinclude",
4502 ] + select({
4503 ":debug_build": [],
4504 "//conditions:default": xnnpack_min_size_copts(),
4505 }),
4506 gcc_copts = xnnpack_gcc_std_copts(),
4507 msvc_copts = xnnpack_msvc_std_copts(),
4508 visibility = xnnpack_visibility(),
4509 deps = [
4510 "@FP16",
4511 "@clog",
4512 "@pthreadpool",
4513 ],
4514)
4515
Marat Dukhan08c4a432019-10-03 09:29:21 -07004516xnnpack_aggregate_library(
4517 name = "ukernels",
Marat Dukhan6e8c0ce2021-04-13 14:35:08 -07004518 aarch32_ios_deps = [
4519 ":neon_ukernels",
4520 ":neonfma_ukernels",
4521 ":neonv8_ukernels",
4522 ":asm_ukernels",
4523 ],
4524 aarch32_nonios_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07004525 ":neon_ukernels",
4526 ":neonfma_ukernels",
Marat Dukhan8853b822020-05-07 12:19:01 -07004527 ":neonv8_ukernels",
Benoit Jacoba9644732020-08-13 12:48:55 -07004528 ":neondot_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004529 ":asm_ukernels",
4530 ],
4531 aarch64_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07004532 ":neon_ukernels",
4533 ":neonfma_ukernels",
Marat Dukhan8853b822020-05-07 12:19:01 -07004534 ":neonv8_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004535 ":neonfp16arith_ukernels",
Benoit Jacoba9644732020-08-13 12:48:55 -07004536 ":neondot_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004537 ":asm_ukernels",
4538 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07004539 generic_deps = [
4540 ":scalar_ukernels",
4541 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07004542 wasm_deps = [
4543 ":wasm_ukernels",
Frank Barchardbcedc082020-08-17 18:00:51 -07004544 ":asm_ukernels",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004545 ],
4546 wasmsimd_deps = [
4547 ":wasm_ukernels",
Frank Barchardbcedc082020-08-17 18:00:51 -07004548 ":asm_ukernels",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004549 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004550 x86_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07004551 ":sse2_ukernels",
Marat Dukhanfe7acb62020-03-09 19:30:05 -07004552 ":ssse3_ukernels",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08004553 ":sse41_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004554 ":avx_ukernels",
Marat Dukhan1566fee2020-08-02 21:55:41 -07004555 ":xop_ukernels",
Marat Dukhanfda12b82019-11-21 12:27:59 -08004556 ":fma3_ukernels",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07004557 ":avx2_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004558 ":avx512f_ukernels",
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07004559 ":avx512skx_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004560 ],
4561)
4562
Marat Dukhan33fcf782020-05-24 14:27:15 -07004563xnnpack_aggregate_library(
4564 name = "ukernels_test_mode",
Marat Dukhan6e8c0ce2021-04-13 14:35:08 -07004565 aarch32_ios_deps = [
4566 ":neon_ukernels_test_mode",
4567 ":neonfma_ukernels_test_mode",
4568 ":neonv8_ukernels_test_mode",
4569 ":asm_ukernels",
4570 ],
4571 aarch32_nonios_deps = [
Marat Dukhan33fcf782020-05-24 14:27:15 -07004572 ":neon_ukernels_test_mode",
4573 ":neonfma_ukernels_test_mode",
4574 ":neonv8_ukernels_test_mode",
Benoit Jacoba9644732020-08-13 12:48:55 -07004575 ":neondot_ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004576 ":asm_ukernels",
4577 ],
4578 aarch64_deps = [
4579 ":neon_ukernels_test_mode",
4580 ":neonfma_ukernels_test_mode",
4581 ":neonv8_ukernels_test_mode",
4582 ":neonfp16arith_ukernels_test_mode",
Benoit Jacoba9644732020-08-13 12:48:55 -07004583 ":neondot_ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004584 ":asm_ukernels",
4585 ],
4586 generic_deps = [
4587 ":scalar_ukernels_test_mode",
4588 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07004589 wasm_deps = [
4590 ":wasm_ukernels_test_mode",
Frank Barchardbcedc082020-08-17 18:00:51 -07004591 ":asm_ukernels",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004592 ],
4593 wasmsimd_deps = [
4594 ":wasm_ukernels_test_mode",
Frank Barchardbcedc082020-08-17 18:00:51 -07004595 ":asm_ukernels",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004596 ],
4597 x86_deps = [
4598 ":sse2_ukernels_test_mode",
4599 ":ssse3_ukernels_test_mode",
4600 ":sse41_ukernels_test_mode",
4601 ":avx_ukernels_test_mode",
Marat Dukhan1566fee2020-08-02 21:55:41 -07004602 ":xop_ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004603 ":fma3_ukernels_test_mode",
4604 ":avx2_ukernels_test_mode",
4605 ":avx512f_ukernels_test_mode",
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07004606 ":avx512skx_ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004607 ],
4608)
4609
Marat Dukhan08c4a432019-10-03 09:29:21 -07004610xnnpack_cc_library(
4611 name = "im2col",
4612 srcs = ["src/im2col.c"],
4613 hdrs = [
4614 "src/xnnpack/common.h",
4615 "src/xnnpack/im2col.h",
4616 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07004617 gcc_copts = xnnpack_gcc_std_copts(),
4618 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004619)
4620
4621xnnpack_cc_library(
4622 name = "indirection",
4623 srcs = ["src/indirection.c"],
4624 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004625 gcc_copts = xnnpack_gcc_std_copts(),
4626 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004627 deps = [
4628 "@FP16",
4629 "@FXdiv",
4630 "@pthreadpool",
4631 ],
4632)
4633
4634xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07004635 name = "indirection_test_mode",
4636 srcs = ["src/indirection.c"],
4637 hdrs = INTERNAL_HDRS,
4638 copts = [
4639 "-UNDEBUG",
4640 "-DXNN_TEST_MODE=1",
4641 ],
4642 gcc_copts = xnnpack_gcc_std_copts(),
4643 msvc_copts = xnnpack_msvc_std_copts(),
4644 deps = [
4645 "@FP16",
4646 "@FXdiv",
4647 "@pthreadpool",
4648 ],
4649)
4650
4651xnnpack_cc_library(
Marat Dukhanab582382020-07-06 13:32:08 -07004652 name = "packing",
4653 srcs = ["src/packing.c"],
4654 hdrs = INTERNAL_HDRS,
4655 gcc_copts = xnnpack_gcc_std_copts(),
4656 msvc_copts = xnnpack_msvc_std_copts(),
4657 deps = [
4658 "@FP16",
4659 "@FXdiv",
4660 "@pthreadpool",
4661 ],
4662)
4663
4664xnnpack_cc_library(
4665 name = "packing_test_mode",
4666 srcs = ["src/packing.c"],
4667 hdrs = INTERNAL_HDRS,
4668 copts = [
4669 "-UNDEBUG",
4670 "-DXNN_TEST_MODE=1",
4671 ],
4672 gcc_copts = xnnpack_gcc_std_copts(),
4673 msvc_copts = xnnpack_msvc_std_copts(),
4674 deps = [
4675 "@FP16",
4676 "@FXdiv",
4677 "@pthreadpool",
4678 ],
4679)
4680
4681xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004682 name = "operator_run",
4683 srcs = ["src/operator-run.c"],
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07004684 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004685 copts = LOGGING_COPTS + select({
Marat Dukhan05702cf2020-03-26 15:41:33 -07004686 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4687 "//conditions:default": [],
4688 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07004689 gcc_copts = xnnpack_gcc_std_copts(),
4690 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004691 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07004692 ":logging_utils",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004693 "@FP16",
4694 "@FXdiv",
4695 "@clog",
4696 "@pthreadpool",
4697 ],
4698)
4699
Chao Mei6ddfc602020-05-13 22:29:36 -07004700xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07004701 name = "operator_run_test_mode",
4702 srcs = ["src/operator-run.c"],
4703 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
4704 copts = LOGGING_COPTS + [
4705 "-UNDEBUG",
4706 "-DXNN_TEST_MODE=1",
4707 ] + select({
4708 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4709 "//conditions:default": [],
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",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004715 "@FP16",
4716 "@FXdiv",
4717 "@clog",
4718 "@pthreadpool",
4719 ],
4720)
4721
4722xnnpack_cc_library(
Chao Mei6ddfc602020-05-13 22:29:36 -07004723 name = "memory_planner",
4724 srcs = ["src/memory-planner.c"],
4725 hdrs = INTERNAL_HDRS,
4726 defines = select({
4727 ":xnn_enable_memopt_explicit_true": ["XNN_ENABLE_MEMOPT=1"],
4728 ":xnn_enable_memopt_explicit_false": ["XNN_ENABLE_MEMOPT=0"],
4729 "//conditions:default": ["XNN_ENABLE_MEMOPT=1"],
4730 }),
4731 gcc_copts = xnnpack_gcc_std_copts(),
4732 msvc_copts = xnnpack_msvc_std_copts(),
4733 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07004734 ":logging_utils",
Chao Mei6ddfc602020-05-13 22:29:36 -07004735 "@pthreadpool",
4736 ],
4737)
4738
Marat Dukhan33fcf782020-05-24 14:27:15 -07004739xnnpack_cc_library(
4740 name = "memory_planner_test_mode",
4741 srcs = ["src/memory-planner.c"],
4742 hdrs = INTERNAL_HDRS,
4743 copts = [
4744 "-UNDEBUG",
4745 "-DXNN_TEST_MODE=1",
4746 ],
4747 defines = select({
4748 ":xnn_enable_memopt_explicit_true": ["XNN_ENABLE_MEMOPT=1"],
4749 ":xnn_enable_memopt_explicit_false": ["XNN_ENABLE_MEMOPT=0"],
4750 "//conditions:default": ["XNN_ENABLE_MEMOPT=1"],
4751 }),
4752 gcc_copts = xnnpack_gcc_std_copts(),
4753 msvc_copts = xnnpack_msvc_std_copts(),
4754 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07004755 ":logging_utils",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004756 "@pthreadpool",
4757 ],
4758)
4759
Marat Dukhan08c4a432019-10-03 09:29:21 -07004760cc_library(
4761 name = "enable_assembly",
4762 defines = select({
4763 ":xnn_enable_assembly_explicit_true": ["XNN_ENABLE_ASSEMBLY=1"],
4764 ":xnn_enable_assembly_explicit_false": ["XNN_ENABLE_ASSEMBLY=0"],
Frank Barchard810171d2019-10-10 10:34:51 -07004765 "//conditions:default": ["XNN_ENABLE_ASSEMBLY=1"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004766 }),
4767)
4768
Marat Dukhan9de90e02020-06-18 16:04:12 -07004769cc_library(
4770 name = "enable_sparse",
4771 defines = select({
4772 ":xnn_enable_sparse_explicit_true": ["XNN_ENABLE_SPARSE=1"],
4773 ":xnn_enable_sparse_explicit_false": ["XNN_ENABLE_SPARSE=0"],
Marat Dukhanb36582b2020-12-08 11:16:28 -08004774 "//conditions:default": ["XNN_ENABLE_SPARSE=1"],
Marat Dukhan9de90e02020-06-18 16:04:12 -07004775 }),
4776)
4777
Marat Dukhancf056b22019-10-07 10:26:29 -07004778xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004779 name = "operators",
4780 srcs = OPERATOR_SRCS + [
Marat Dukhan496389f2021-04-07 15:47:12 -07004781 "src/allocator.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004782 "src/operator-delete.c",
Marat Dukhancf056b22019-10-07 10:26:29 -07004783 ],
4784 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004785 copts = LOGGING_COPTS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07004786 "-Isrc",
4787 "-Iinclude",
4788 ] + select({
4789 ":debug_build": [],
4790 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07004791 }) + select({
4792 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4793 "//conditions:default": [],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004794 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07004795 gcc_copts = xnnpack_gcc_std_copts(),
4796 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004797 deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07004798 ":indirection",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004799 ":logging_utils",
Marat Dukhanab582382020-07-06 13:32:08 -07004800 ":packing",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004801 "@FP16",
4802 "@FXdiv",
4803 "@clog",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004804 "@pthreadpool",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004805 ],
4806)
4807
Marat Dukhan10a38082020-04-17 03:58:35 -07004808xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07004809 name = "operators_test_mode",
4810 srcs = OPERATOR_SRCS + [
Marat Dukhan496389f2021-04-07 15:47:12 -07004811 "src/allocator.c",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004812 "src/operator-delete.c",
4813 ],
4814 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
4815 copts = LOGGING_COPTS + [
4816 "-Isrc",
4817 "-Iinclude",
4818 "-UNDEBUG",
4819 "-DXNN_TEST_MODE=1",
4820 ] + select({
4821 ":debug_build": [],
4822 "//conditions:default": xnnpack_min_size_copts(),
4823 }) + select({
4824 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4825 "//conditions:default": [],
4826 }),
4827 gcc_copts = xnnpack_gcc_std_copts(),
4828 msvc_copts = xnnpack_msvc_std_copts(),
4829 deps = [
4830 ":indirection_test_mode",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004831 ":logging_utils",
Marat Dukhanab582382020-07-06 13:32:08 -07004832 ":packing_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004833 "@FP16",
4834 "@FXdiv",
4835 "@clog",
4836 "@pthreadpool",
4837 ],
4838)
4839
4840xnnpack_cc_library(
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004841 name = "XNNPACK",
4842 srcs = [
4843 "src/init.c",
Marat Dukhanccfdbd12020-02-03 14:27:45 -08004844 "src/runtime.c",
4845 "src/subgraph.c",
4846 "src/tensor.c",
Marat Dukhanf03da0d2020-06-10 16:00:20 -07004847 ] + SUBGRAPH_SRCS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004848 hdrs = ["include/xnnpack.h"],
4849 copts = LOGGING_COPTS + [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004850 "-Isrc",
4851 "-Iinclude",
4852 ] + select({
4853 ":debug_build": [],
4854 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07004855 }) + select({
4856 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4857 "//conditions:default": [],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004858 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07004859 gcc_copts = xnnpack_gcc_std_copts(),
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004860 includes = ["include"],
Marat Dukhan10a38082020-04-17 03:58:35 -07004861 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004862 visibility = xnnpack_visibility(),
4863 deps = [
4864 ":enable_assembly",
Marat Dukhan9de90e02020-06-18 16:04:12 -07004865 ":enable_sparse",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004866 ":logging_utils",
Chao Mei6ddfc602020-05-13 22:29:36 -07004867 ":memory_planner",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004868 ":operator_run",
4869 ":operators",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004870 ":ukernels",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004871 "@clog",
Marat Dukhanab2946c2020-05-21 20:04:13 -07004872 "@FP16",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004873 "@pthreadpool",
Marat Dukhand343c222019-10-07 09:22:14 -07004874 ] + select({
4875 ":emscripten": [],
4876 "//conditions:default": ["@cpuinfo"],
4877 }),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004878)
4879
Marat Dukhan10a38082020-04-17 03:58:35 -07004880xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07004881 name = "XNNPACK_test_mode",
4882 srcs = [
4883 "src/init.c",
4884 "src/runtime.c",
4885 "src/subgraph.c",
4886 "src/tensor.c",
Marat Dukhanf03da0d2020-06-10 16:00:20 -07004887 ] + SUBGRAPH_SRCS,
Marat Dukhan33fcf782020-05-24 14:27:15 -07004888 hdrs = ["include/xnnpack.h"],
4889 copts = LOGGING_COPTS + [
4890 "-Isrc",
4891 "-Iinclude",
4892 "-UNDEBUG",
4893 "-DXNN_TEST_MODE=1",
4894 ] + select({
4895 ":debug_build": [],
4896 "//conditions:default": xnnpack_min_size_copts(),
4897 }) + select({
4898 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4899 "//conditions:default": [],
4900 }),
4901 gcc_copts = xnnpack_gcc_std_copts(),
4902 includes = ["include"],
4903 msvc_copts = xnnpack_msvc_std_copts(),
4904 visibility = xnnpack_visibility(),
4905 deps = [
4906 ":enable_assembly",
Marat Dukhan9de90e02020-06-18 16:04:12 -07004907 ":enable_sparse",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004908 ":logging_utils",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004909 ":memory_planner_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004910 ":operator_run_test_mode",
4911 ":operators_test_mode",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004912 ":ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004913 "@clog",
4914 "@FP16",
4915 "@pthreadpool",
4916 ] + select({
4917 ":emscripten": [],
4918 "//conditions:default": ["@cpuinfo"],
4919 }),
4920)
4921
Marat Dukhanf0cb70a2021-03-30 15:45:15 -07004922# Specialized XNNPACK version for TensorFlow Lite. Excludes operators currently
4923# not used by the TensorFlow Lite XNNPACK delegate to minimize code size.
Marat Dukhanae046f52020-06-15 13:16:14 -07004924xnnpack_cc_library(
Marat Dukhanf0cb70a2021-03-30 15:45:15 -07004925 name = "xnnpack_for_tflite",
4926 srcs = [
4927 "src/init.c",
4928 "src/runtime.c",
4929 "src/subgraph.c",
4930 "src/tensor.c",
4931 ] + SUBGRAPH_SRCS,
4932 hdrs = ["include/xnnpack.h"],
4933 copts = LOGGING_COPTS + [
4934 "-Isrc",
4935 "-Iinclude",
4936 ] + select({
4937 ":debug_build": [],
4938 "//conditions:default": xnnpack_min_size_copts(),
4939 }) + select({
4940 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4941 "//conditions:default": [],
4942 }),
4943 defines = [
Marat Dukhanf0cb70a2021-03-30 15:45:15 -07004944 "XNN_NO_QU8_OPERATORS",
4945 "XNN_NO_U8_OPERATORS",
4946 "XNN_NO_X8_OPERATORS",
4947 "XNN_NO_F16_OPERATORS",
4948 "XNN_NO_X16_OPERATORS",
Marat Dukhanb939cdb2021-03-30 18:51:51 -07004949 ] + select({
4950 ":xnn_enable_qs8_explicit_true": [],
4951 ":xnn_enable_qs8_explicit_false": ["XNN_NO_QS8_OPERATORS"],
4952 "//conditions:default": ["XNN_NO_QS8_OPERATORS"],
4953 }),
Marat Dukhanf0cb70a2021-03-30 15:45:15 -07004954 gcc_copts = xnnpack_gcc_std_copts(),
4955 includes = ["include"],
4956 msvc_copts = xnnpack_msvc_std_copts(),
4957 visibility = xnnpack_visibility(),
4958 deps = [
4959 ":enable_assembly",
4960 ":enable_sparse",
4961 ":logging_utils",
4962 ":memory_planner",
4963 ":operator_run",
4964 ":operators",
Marat Dukhand09ca262021-03-30 16:17:12 -07004965 ":ukernels",
Marat Dukhanf0cb70a2021-03-30 15:45:15 -07004966 "@clog",
4967 "@FP16",
4968 "@pthreadpool",
4969 ] + select({
4970 ":emscripten": [],
4971 "//conditions:default": ["@cpuinfo"],
4972 }),
4973)
4974
4975# Specialized XNNPACK version for TensorFlow.js. Excludes operators currently
4976# not used by the TensorFlow.js WebAssembly backend to minimize code size.
4977xnnpack_cc_library(
4978 name = "xnnpack_for_tfjs",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004979 srcs = [
4980 "src/init.c",
4981 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07004982 hdrs = ["include/xnnpack.h"],
4983 copts = LOGGING_COPTS + [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004984 "-Isrc",
4985 "-Iinclude",
4986 ] + select({
4987 ":debug_build": [],
4988 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07004989 }) + select({
4990 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4991 "//conditions:default": [],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004992 }),
4993 defines = [
Marat Dukhan16f1e1a2020-08-04 16:38:22 -07004994 "XNN_NO_QS8_OPERATORS",
Marat Dukhan08b7a972020-07-14 18:17:29 -07004995 "XNN_NO_QU8_OPERATORS",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004996 "XNN_NO_U8_OPERATORS",
4997 "XNN_NO_X8_OPERATORS",
Marat Dukhanefc47b82019-11-18 09:25:38 -08004998 "XNN_NO_NCHW_OPERATORS",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004999 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07005000 gcc_copts = xnnpack_gcc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07005001 includes = ["include"],
Marat Dukhan10a38082020-04-17 03:58:35 -07005002 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07005003 visibility = xnnpack_visibility(),
5004 deps = [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07005005 ":enable_assembly",
Marat Dukhan3b59de22020-06-03 20:15:19 -07005006 ":logging_utils",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005007 ":operator_run",
5008 ":operators",
Marat Dukhan3b59de22020-06-03 20:15:19 -07005009 ":ukernels",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07005010 "@clog",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005011 "@pthreadpool",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07005012 ] + select({
5013 ":emscripten": [],
5014 "//conditions:default": ["@cpuinfo"],
5015 }),
Marat Dukhan08c4a432019-10-03 09:29:21 -07005016)
5017
Marat Dukhancf056b22019-10-07 10:26:29 -07005018xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005019 name = "bench_utils",
5020 srcs = ["bench/utils.cc"],
5021 hdrs = ["bench/utils.h"],
Marat Dukhanbad48fe2019-11-04 10:35:22 -08005022 deps = [
5023 "@com_google_benchmark//:benchmark",
5024 "@cpuinfo",
5025 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005026)
5027
Frank Barchard7e955972019-10-11 10:34:25 -07005028######################### Benchmarks for micro-kernels #########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07005029
5030xnnpack_benchmark(
Marat Dukhan595e1702020-07-31 10:12:52 -07005031 name = "qs8_gemm_bench",
5032 srcs = [
5033 "bench/gemm.h",
5034 "bench/qs8-gemm.cc",
5035 "src/xnnpack/AlignedAllocator.h",
5036 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Frank Barchard31328cb2020-10-12 11:55:18 -07005037 copts = xnnpack_optional_ruy_copts() + xnnpack_optional_gemmlowp_copts(),
5038 deps = MICROKERNEL_BENCHMARK_DEPS + [":packing"] + xnnpack_optional_ruy_deps() + xnnpack_optional_gemmlowp_deps(),
Marat Dukhan595e1702020-07-31 10:12:52 -07005039)
5040
5041xnnpack_benchmark(
Marat Dukhan56bdd4a2020-08-03 19:47:04 -07005042 name = "qs8_requantization_bench",
5043 srcs = [
5044 "bench/qs8-requantization.cc",
Marat Dukhan56bdd4a2020-08-03 19:47:04 -07005045 "src/xnnpack/AlignedAllocator.h",
Frank Barchard04336c12020-10-22 16:48:55 -07005046 "src/xnnpack/requantization-stubs.h",
Marat Dukhan56bdd4a2020-08-03 19:47:04 -07005047 ] + MICROKERNEL_BENCHMARK_HDRS,
5048 deps = MICROKERNEL_BENCHMARK_DEPS,
5049)
5050
5051xnnpack_benchmark(
Marat Dukhan08b7a972020-07-14 18:17:29 -07005052 name = "qu8_gemm_bench",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005053 srcs = [
5054 "bench/gemm.h",
Marat Dukhan08b7a972020-07-14 18:17:29 -07005055 "bench/qu8-gemm.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005056 "src/xnnpack/AlignedAllocator.h",
5057 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07005058 copts = xnnpack_optional_ruy_copts() + xnnpack_optional_gemmlowp_copts(),
Marat Dukhanab582382020-07-06 13:32:08 -07005059 deps = MICROKERNEL_BENCHMARK_DEPS + [":packing"] + xnnpack_optional_ruy_deps() + xnnpack_optional_gemmlowp_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07005060)
5061
5062xnnpack_benchmark(
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07005063 name = "qu8_requantization_bench",
5064 srcs = [
5065 "bench/qu8-requantization.cc",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07005066 "src/xnnpack/AlignedAllocator.h",
Frank Barchard04336c12020-10-22 16:48:55 -07005067 "src/xnnpack/requantization-stubs.h",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07005068 ] + MICROKERNEL_BENCHMARK_HDRS,
5069 deps = MICROKERNEL_BENCHMARK_DEPS,
5070)
5071
5072xnnpack_benchmark(
Frank Barchard40d20fe2020-05-05 00:37:45 -07005073 name = "f16_igemm_bench",
5074 srcs = [
5075 "bench/f16-igemm.cc",
5076 "bench/conv.h",
5077 "bench/google/conv.h",
5078 "src/xnnpack/AlignedAllocator.h",
5079 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005080 deps = MICROKERNEL_BENCHMARK_DEPS + [
5081 ":indirection",
5082 ":packing",
5083 ],
Frank Barchard40d20fe2020-05-05 00:37:45 -07005084)
5085
5086xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005087 name = "f16_gemm_bench",
5088 srcs = [
5089 "bench/f16-gemm.cc",
5090 "bench/gemm.h",
5091 "src/xnnpack/AlignedAllocator.h",
5092 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005093 deps = MICROKERNEL_BENCHMARK_DEPS + [
5094 ":packing",
5095 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005096)
5097
5098xnnpack_benchmark(
Marat Dukhanbdb56f52020-02-05 21:42:49 -08005099 name = "f16_spmm_bench",
5100 srcs = [
5101 "bench/f16-spmm.cc",
Marat Dukhan1631e3e2020-12-06 19:29:31 -08005102 "bench/spmm.h",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08005103 "src/xnnpack/AlignedAllocator.h",
5104 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanbdb56f52020-02-05 21:42:49 -08005105 deps = MICROKERNEL_BENCHMARK_DEPS,
5106)
5107
5108xnnpack_benchmark(
Marat Dukhan6674d692021-05-05 22:27:00 -07005109 name = "f16_vrelu_bench",
5110 srcs = [
5111 "bench/f16-vrelu.cc",
5112 "src/xnnpack/AlignedAllocator.h",
5113 ] + MICROKERNEL_BENCHMARK_HDRS,
5114 deps = MICROKERNEL_BENCHMARK_DEPS,
5115)
5116
5117xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005118 name = "f32_igemm_bench",
5119 srcs = [
5120 "bench/f32-igemm.cc",
5121 "bench/conv.h",
5122 "src/xnnpack/AlignedAllocator.h",
5123 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005124 deps = MICROKERNEL_BENCHMARK_DEPS + [
5125 ":indirection",
5126 ":packing",
5127 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005128)
5129
5130xnnpack_benchmark(
5131 name = "f32_conv_hwc_bench",
5132 srcs = [
5133 "bench/f32-conv-hwc.cc",
5134 "bench/dconv.h",
5135 "src/xnnpack/AlignedAllocator.h",
5136 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005137 deps = MICROKERNEL_BENCHMARK_DEPS + [
5138 ":packing",
5139 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005140)
5141
5142xnnpack_benchmark(
Marat Dukhan1f29b802020-05-15 23:46:39 -07005143 name = "f32_conv_hwc2chw_bench",
Erich Elsen563df5f2019-10-23 08:02:21 -07005144 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07005145 "bench/f32-conv-hwc2chw.cc",
Erich Elsen563df5f2019-10-23 08:02:21 -07005146 "bench/dconv.h",
5147 "src/xnnpack/AlignedAllocator.h",
5148 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005149 deps = MICROKERNEL_BENCHMARK_DEPS + [
5150 ":packing",
5151 ],
Erich Elsen563df5f2019-10-23 08:02:21 -07005152)
5153
5154xnnpack_benchmark(
Frank Barchard5a599a62020-06-04 20:12:44 -07005155 name = "f16_dwconv_bench",
5156 srcs = [
5157 "bench/f16-dwconv.cc",
5158 "bench/dwconv.h",
5159 "bench/google/dwconv.h",
5160 "src/xnnpack/AlignedAllocator.h",
5161 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005162 deps = MICROKERNEL_BENCHMARK_DEPS + [
5163 ":indirection",
5164 ":packing",
5165 ],
Frank Barchard5a599a62020-06-04 20:12:44 -07005166)
5167
5168xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005169 name = "f32_dwconv_bench",
5170 srcs = [
5171 "bench/f32-dwconv.cc",
5172 "bench/dwconv.h",
5173 "src/xnnpack/AlignedAllocator.h",
5174 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005175 deps = MICROKERNEL_BENCHMARK_DEPS + [
5176 ":indirection",
5177 ":packing",
5178 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005179)
5180
5181xnnpack_benchmark(
Marat Dukhanbf715f92020-10-23 20:17:00 -07005182 name = "f32_dwconv2d_chw_bench",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005183 srcs = [
Marat Dukhanbf715f92020-10-23 20:17:00 -07005184 "bench/f32-dwconv2d-chw.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005185 "bench/dwconv.h",
5186 "src/xnnpack/AlignedAllocator.h",
5187 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005188 deps = MICROKERNEL_BENCHMARK_DEPS + [
5189 ":indirection",
5190 ":packing",
5191 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005192)
5193
5194xnnpack_benchmark(
5195 name = "f32_gemm_bench",
5196 srcs = [
5197 "bench/f32-gemm.cc",
5198 "bench/gemm.h",
5199 "src/xnnpack/AlignedAllocator.h",
5200 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07005201 copts = xnnpack_optional_ruy_copts(),
Marat Dukhanab582382020-07-06 13:32:08 -07005202 deps = MICROKERNEL_BENCHMARK_DEPS + [":packing"] + xnnpack_optional_ruy_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07005203)
5204
5205xnnpack_benchmark(
Marat Dukhan4c4eb002019-12-08 21:27:49 -08005206 name = "f32_raddexpminusmax_bench",
5207 srcs = [
5208 "bench/f32-raddexpminusmax.cc",
5209 "src/xnnpack/AlignedAllocator.h",
5210 ] + MICROKERNEL_BENCHMARK_HDRS,
5211 deps = MICROKERNEL_BENCHMARK_DEPS,
5212)
5213
5214xnnpack_benchmark(
5215 name = "f32_raddextexp_bench",
5216 srcs = [
5217 "bench/f32-raddextexp.cc",
5218 "src/xnnpack/AlignedAllocator.h",
5219 ] + MICROKERNEL_BENCHMARK_HDRS,
5220 deps = MICROKERNEL_BENCHMARK_DEPS,
5221)
5222
5223xnnpack_benchmark(
5224 name = "f32_raddstoreexpminusmax_bench",
5225 srcs = [
5226 "bench/f32-raddstoreexpminusmax.cc",
5227 "src/xnnpack/AlignedAllocator.h",
5228 ] + MICROKERNEL_BENCHMARK_HDRS,
5229 deps = MICROKERNEL_BENCHMARK_DEPS,
5230)
5231
5232xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005233 name = "f32_rmax_bench",
5234 srcs = [
5235 "bench/f32-rmax.cc",
5236 "src/xnnpack/AlignedAllocator.h",
5237 ] + MICROKERNEL_BENCHMARK_HDRS,
5238 deps = MICROKERNEL_BENCHMARK_DEPS,
5239)
5240
5241xnnpack_benchmark(
5242 name = "f32_spmm_bench",
5243 srcs = [
5244 "bench/f32-spmm.cc",
Marat Dukhan1631e3e2020-12-06 19:29:31 -08005245 "bench/spmm.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005246 "src/xnnpack/AlignedAllocator.h",
5247 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005248 deps = MICROKERNEL_BENCHMARK_DEPS,
5249)
5250
5251xnnpack_benchmark(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08005252 name = "f32_softmax_bench",
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07005253 srcs = [
Marat Dukhanfd8e6892020-01-27 15:25:25 -08005254 "bench/f32-softmax.cc",
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07005255 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07005256 copts = xnnpack_optional_dnnl_copts(),
Marat Dukhan8d3c6932020-03-06 20:27:27 -08005257 deps = MICROKERNEL_BENCHMARK_DEPS + xnnpack_optional_dnnl_deps(),
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07005258)
5259
5260xnnpack_benchmark(
Marat Dukhaned6baaf2020-12-01 15:07:08 -08005261 name = "f32_velu_bench",
5262 srcs = [
5263 "bench/f32-velu.cc",
5264 "src/xnnpack/AlignedAllocator.h",
5265 ] + MICROKERNEL_BENCHMARK_HDRS,
5266 deps = MICROKERNEL_BENCHMARK_DEPS,
5267)
5268
5269xnnpack_benchmark(
Marat Dukhan6674d692021-05-05 22:27:00 -07005270 name = "f32_vhswish_bench",
5271 srcs = [
5272 "bench/f32-vhswish.cc",
5273 "src/xnnpack/AlignedAllocator.h",
5274 ] + MICROKERNEL_BENCHMARK_HDRS,
5275 deps = MICROKERNEL_BENCHMARK_DEPS,
5276)
5277
5278xnnpack_benchmark(
5279 name = "f32_vrelu_bench",
5280 srcs = [
5281 "bench/f32-vrelu.cc",
5282 "src/xnnpack/AlignedAllocator.h",
5283 ] + MICROKERNEL_BENCHMARK_HDRS,
5284 deps = MICROKERNEL_BENCHMARK_DEPS,
5285)
5286
5287xnnpack_benchmark(
Marat Dukhan4c4eb002019-12-08 21:27:49 -08005288 name = "f32_vscaleexpminusmax_bench",
5289 srcs = [
5290 "bench/f32-vscaleexpminusmax.cc",
5291 "src/xnnpack/AlignedAllocator.h",
5292 ] + MICROKERNEL_BENCHMARK_HDRS,
5293 deps = MICROKERNEL_BENCHMARK_DEPS,
5294)
5295
5296xnnpack_benchmark(
5297 name = "f32_vscaleextexp_bench",
5298 srcs = [
5299 "bench/f32-vscaleextexp.cc",
5300 "src/xnnpack/AlignedAllocator.h",
5301 ] + MICROKERNEL_BENCHMARK_HDRS,
5302 deps = MICROKERNEL_BENCHMARK_DEPS,
5303)
5304
5305xnnpack_benchmark(
Marat Dukhan6674d692021-05-05 22:27:00 -07005306 name = "f32_vsigmoid_bench",
5307 srcs = [
5308 "bench/f32-vsigmoid.cc",
5309 "src/xnnpack/AlignedAllocator.h",
5310 ] + MICROKERNEL_BENCHMARK_HDRS,
5311 deps = MICROKERNEL_BENCHMARK_DEPS,
5312)
5313
5314xnnpack_benchmark(
Marat Dukhanf4db2f32020-06-30 10:55:30 -07005315 name = "f32_vsqrt_bench",
5316 srcs = [
5317 "bench/f32-vsqrt.cc",
5318 "src/xnnpack/AlignedAllocator.h",
5319 ] + MICROKERNEL_BENCHMARK_HDRS,
5320 deps = MICROKERNEL_BENCHMARK_DEPS,
5321)
5322
5323xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005324 name = "f32_im2col_gemm_bench",
5325 srcs = [
5326 "bench/f32-im2col-gemm.cc",
5327 "bench/conv.h",
5328 "src/xnnpack/AlignedAllocator.h",
5329 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005330 deps = MICROKERNEL_BENCHMARK_DEPS + [
5331 ":im2col",
5332 ":packing",
5333 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005334)
5335
Marat Dukhanfe7acb62020-03-09 19:30:05 -07005336xnnpack_benchmark(
Marat Dukhanffbf96a2020-05-14 02:59:08 -07005337 name = "rounding_bench",
5338 srcs = [
5339 "bench/rounding.cc",
Marat Dukhanffbf96a2020-05-14 02:59:08 -07005340 "src/xnnpack/AlignedAllocator.h",
Frank Barchard04336c12020-10-22 16:48:55 -07005341 "src/xnnpack/math-stubs.h",
Marat Dukhanffbf96a2020-05-14 02:59:08 -07005342 ] + MICROKERNEL_BENCHMARK_HDRS,
5343 deps = MICROKERNEL_BENCHMARK_DEPS,
5344)
5345
Marat Dukhan08c4a432019-10-03 09:29:21 -07005346########################### Benchmarks for operators ###########################
5347
5348xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005349 name = "average_pooling_bench",
5350 srcs = ["bench/average-pooling.cc"],
Marat Dukhan7a16d8b2020-03-11 04:22:44 -07005351 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07005352 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07005353 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07005354)
5355
5356xnnpack_benchmark(
Marat Dukhan87727142020-06-24 15:24:10 -07005357 name = "bankers_rounding_bench",
5358 srcs = ["bench/bankers-rounding.cc"],
5359 copts = xnnpack_optional_tflite_copts(),
5360 tags = ["nowin32"],
5361 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
5362)
5363
5364xnnpack_benchmark(
5365 name = "ceiling_bench",
5366 srcs = ["bench/ceiling.cc"],
5367 copts = xnnpack_optional_tflite_copts(),
5368 tags = ["nowin32"],
5369 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
5370)
5371
5372xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005373 name = "channel_shuffle_bench",
5374 srcs = ["bench/channel-shuffle.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07005375 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005376)
5377
5378xnnpack_benchmark(
5379 name = "convolution_bench",
5380 srcs = ["bench/convolution.cc"],
5381 copts = xnnpack_optional_tflite_copts() + xnnpack_optional_armcl_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07005382 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07005383 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps() + xnnpack_optional_armcl_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07005384)
5385
5386xnnpack_benchmark(
5387 name = "deconvolution_bench",
5388 srcs = ["bench/deconvolution.cc"],
5389 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07005390 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07005391 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07005392)
5393
5394xnnpack_benchmark(
Marat Dukhanb6bd4bc2020-12-01 17:01:40 -08005395 name = "elu_bench",
5396 srcs = ["bench/elu.cc"],
5397 copts = xnnpack_optional_tflite_copts(),
5398 tags = ["nowin32"],
5399 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
5400)
5401
5402xnnpack_benchmark(
Marat Dukhan87727142020-06-24 15:24:10 -07005403 name = "floor_bench",
5404 srcs = ["bench/floor.cc"],
5405 copts = xnnpack_optional_tflite_copts(),
5406 tags = ["nowin32"],
5407 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
5408)
5409
5410xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005411 name = "global_average_pooling_bench",
5412 srcs = ["bench/global-average-pooling.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07005413 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005414)
5415
5416xnnpack_benchmark(
Marat Dukhanad352602020-06-25 21:50:54 -07005417 name = "hardswish_bench",
5418 srcs = ["bench/hardswish.cc"],
5419 copts = xnnpack_optional_tflite_copts(),
5420 tags = ["nowin32"],
5421 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
5422)
5423
5424xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005425 name = "max_pooling_bench",
5426 srcs = ["bench/max-pooling.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07005427 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005428)
5429
5430xnnpack_benchmark(
5431 name = "sigmoid_bench",
5432 srcs = ["bench/sigmoid.cc"],
Marat Dukhanc3b9e862019-11-17 13:18:54 -08005433 copts = xnnpack_optional_tflite_copts(),
Marat Dukhanca2ba702020-04-24 01:31:47 -07005434 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07005435 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07005436)
5437
5438xnnpack_benchmark(
Marat Dukhan95b22432019-10-30 16:30:14 -07005439 name = "prelu_bench",
5440 srcs = ["bench/prelu.cc"],
5441 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07005442 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07005443 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan95b22432019-10-30 16:30:14 -07005444)
5445
5446xnnpack_benchmark(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08005447 name = "softmax_bench",
5448 srcs = ["bench/softmax.cc"],
Marat Dukhan9c0db962020-01-28 12:30:14 -08005449 copts = xnnpack_optional_tflite_copts(),
Marat Dukhanca2ba702020-04-24 01:31:47 -07005450 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07005451 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07005452)
5453
Marat Dukhan87727142020-06-24 15:24:10 -07005454xnnpack_benchmark(
Marat Dukhan6804bbd2020-06-30 19:26:11 -07005455 name = "square_root_bench",
5456 srcs = ["bench/square-root.cc"],
5457 copts = xnnpack_optional_tflite_copts(),
5458 tags = ["nowin32"],
5459 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
5460)
5461
5462xnnpack_benchmark(
Marat Dukhan87727142020-06-24 15:24:10 -07005463 name = "truncation_bench",
5464 srcs = ["bench/truncation.cc"],
5465 deps = OPERATOR_BENCHMARK_DEPS,
5466)
5467
Marat Dukhanc068bb62019-10-04 13:24:39 -07005468############################# End-to-end benchmarks ############################
5469
5470cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07005471 name = "fp32_mobilenet_v1",
5472 srcs = ["models/fp32-mobilenet-v1.cc"],
Marat Dukhanc068bb62019-10-04 13:24:39 -07005473 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08005474 copts = xnnpack_std_cxxopts(),
Marat Dukhanc068bb62019-10-04 13:24:39 -07005475 linkstatic = True,
5476 deps = [
5477 ":XNNPACK",
5478 "@pthreadpool",
5479 ],
5480)
5481
5482cc_library(
Marat Dukhan4cea2322021-03-09 09:35:36 -08005483 name = "fp32_sparse_mobilenet_v1",
5484 srcs = ["models/fp32-sparse-mobilenet-v1.cc"],
5485 hdrs = ["models/models.h"],
5486 copts = xnnpack_std_cxxopts(),
5487 linkstatic = True,
5488 deps = [
5489 ":XNNPACK",
5490 "@pthreadpool",
5491 ],
5492)
5493
5494cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07005495 name = "fp16_mobilenet_v1",
5496 srcs = ["models/fp16-mobilenet-v1.cc"],
5497 hdrs = ["models/models.h"],
5498 copts = xnnpack_std_cxxopts(),
5499 linkstatic = True,
5500 deps = [
5501 ":XNNPACK",
5502 "@FP16",
5503 "@pthreadpool",
5504 ],
5505)
5506
5507cc_library(
Marat Dukhan0743cdf2020-08-04 18:52:07 -07005508 name = "qs8_mobilenet_v1",
5509 srcs = ["models/qs8-mobilenet-v1.cc"],
5510 hdrs = ["models/models.h"],
5511 copts = xnnpack_std_cxxopts(),
5512 linkstatic = True,
5513 deps = [
5514 ":XNNPACK",
5515 "@pthreadpool",
5516 ],
5517)
5518
5519cc_library(
Marat Dukhan70a96182020-09-03 17:13:58 -07005520 name = "qs8_mobilenet_v2",
5521 srcs = ["models/qs8-mobilenet-v2.cc"],
5522 hdrs = ["models/models.h"],
5523 copts = xnnpack_std_cxxopts(),
5524 linkstatic = True,
5525 deps = [
5526 ":XNNPACK",
5527 "@pthreadpool",
5528 ],
5529)
5530
5531cc_library(
Marat Dukhan12a23bb2021-03-08 08:13:21 -08005532 name = "qu8_mobilenet_v1",
5533 srcs = ["models/qu8-mobilenet-v1.cc"],
5534 hdrs = ["models/models.h"],
5535 copts = xnnpack_std_cxxopts(),
5536 linkstatic = True,
5537 deps = [
5538 ":XNNPACK",
5539 "@pthreadpool",
5540 ],
5541)
5542
5543cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07005544 name = "fp32_mobilenet_v2",
5545 srcs = ["models/fp32-mobilenet-v2.cc"],
Marat Dukhanc068bb62019-10-04 13:24:39 -07005546 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08005547 copts = xnnpack_std_cxxopts(),
Marat Dukhanc068bb62019-10-04 13:24:39 -07005548 linkstatic = True,
5549 deps = [
5550 ":XNNPACK",
5551 "@pthreadpool",
5552 ],
5553)
5554
Marat Dukhanc08cdf52019-12-09 09:17:51 -08005555cc_library(
Marat Dukhan4cea2322021-03-09 09:35:36 -08005556 name = "fp32_sparse_mobilenet_v2",
5557 srcs = ["models/fp32-sparse-mobilenet-v2.cc"],
5558 hdrs = ["models/models.h"],
5559 copts = xnnpack_std_cxxopts(),
5560 linkstatic = True,
5561 deps = [
5562 ":XNNPACK",
5563 "@pthreadpool",
5564 ],
5565)
5566
5567cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07005568 name = "fp16_mobilenet_v2",
5569 srcs = ["models/fp16-mobilenet-v2.cc"],
5570 hdrs = ["models/models.h"],
5571 copts = xnnpack_std_cxxopts(),
5572 linkstatic = True,
5573 deps = [
5574 ":XNNPACK",
5575 "@FP16",
5576 "@pthreadpool",
5577 ],
5578)
5579
5580cc_library(
5581 name = "fp32_mobilenet_v3_large",
5582 srcs = ["models/fp32-mobilenet-v3-large.cc"],
Marat Dukhanc08cdf52019-12-09 09:17:51 -08005583 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08005584 copts = xnnpack_std_cxxopts(),
Marat Dukhanc08cdf52019-12-09 09:17:51 -08005585 linkstatic = True,
5586 deps = [
5587 ":XNNPACK",
5588 "@pthreadpool",
5589 ],
5590)
5591
5592cc_library(
Marat Dukhan4cea2322021-03-09 09:35:36 -08005593 name = "fp32_sparse_mobilenet_v3_large",
5594 srcs = ["models/fp32-sparse-mobilenet-v3-large.cc"],
5595 hdrs = ["models/models.h"],
5596 copts = xnnpack_std_cxxopts(),
5597 linkstatic = True,
5598 deps = [
5599 ":XNNPACK",
5600 "@pthreadpool",
5601 ],
5602)
5603
5604cc_library(
Marat Dukhan66f3ccd2020-09-14 16:09:38 -07005605 name = "fp16_mobilenet_v3_large",
5606 srcs = ["models/fp16-mobilenet-v3-large.cc"],
5607 hdrs = ["models/models.h"],
5608 copts = xnnpack_std_cxxopts(),
5609 linkstatic = True,
5610 deps = [
5611 ":XNNPACK",
5612 "@FP16",
5613 "@pthreadpool",
5614 ],
5615)
5616
5617cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07005618 name = "fp32_mobilenet_v3_small",
5619 srcs = ["models/fp32-mobilenet-v3-small.cc"],
Marat Dukhanc08cdf52019-12-09 09:17:51 -08005620 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08005621 copts = xnnpack_std_cxxopts(),
Marat Dukhanc08cdf52019-12-09 09:17:51 -08005622 linkstatic = True,
5623 deps = [
5624 ":XNNPACK",
5625 "@pthreadpool",
5626 ],
5627)
5628
Marat Dukhan66f3ccd2020-09-14 16:09:38 -07005629cc_library(
Marat Dukhan4cea2322021-03-09 09:35:36 -08005630 name = "fp32_sparse_mobilenet_v3_small",
5631 srcs = ["models/fp32-sparse-mobilenet-v3-small.cc"],
5632 hdrs = ["models/models.h"],
5633 copts = xnnpack_std_cxxopts(),
5634 linkstatic = True,
5635 deps = [
5636 ":XNNPACK",
5637 "@pthreadpool",
5638 ],
5639)
5640
5641cc_library(
Marat Dukhan66f3ccd2020-09-14 16:09:38 -07005642 name = "fp16_mobilenet_v3_small",
5643 srcs = ["models/fp16-mobilenet-v3-small.cc"],
5644 hdrs = ["models/models.h"],
5645 copts = xnnpack_std_cxxopts(),
5646 linkstatic = True,
5647 deps = [
5648 ":XNNPACK",
5649 "@FP16",
5650 "@pthreadpool",
5651 ],
5652)
5653
Marat Dukhanc068bb62019-10-04 13:24:39 -07005654xnnpack_benchmark(
Marat Dukhanef4416e2019-10-31 13:44:40 -07005655 name = "f32_dwconv_e2e_bench",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08005656 srcs = [
5657 "bench/f32-dwconv-e2e.cc",
5658 "bench/end2end.h",
5659 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanef4416e2019-10-31 13:44:40 -07005660 deps = MICROKERNEL_BENCHMARK_DEPS + [
5661 ":XNNPACK",
Marat Dukhan270a2c42020-06-26 16:45:52 -07005662 ":fp32_mobilenet_v1",
5663 ":fp32_mobilenet_v2",
5664 ":fp32_mobilenet_v3_large",
5665 ":fp32_mobilenet_v3_small",
Marat Dukhanef4416e2019-10-31 13:44:40 -07005666 ],
5667)
5668
5669xnnpack_benchmark(
Marat Dukhan5f18d262019-10-31 10:24:14 -07005670 name = "f32_gemm_e2e_bench",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08005671 srcs = [
5672 "bench/f32-gemm-e2e.cc",
5673 "bench/end2end.h",
5674 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan5f18d262019-10-31 10:24:14 -07005675 deps = MICROKERNEL_BENCHMARK_DEPS + [
5676 ":XNNPACK",
Marat Dukhan270a2c42020-06-26 16:45:52 -07005677 ":fp32_mobilenet_v1",
5678 ":fp32_mobilenet_v2",
5679 ":fp32_mobilenet_v3_large",
5680 ":fp32_mobilenet_v3_small",
Marat Dukhan5f18d262019-10-31 10:24:14 -07005681 ],
5682)
5683
5684xnnpack_benchmark(
Frank Barcharddc909cb2021-02-08 13:59:31 -08005685 name = "qs8_gemm_e2e_bench",
5686 srcs = [
5687 "bench/qs8-gemm-e2e.cc",
5688 "bench/end2end.h",
5689 ] + MICROKERNEL_BENCHMARK_HDRS,
5690 deps = MICROKERNEL_BENCHMARK_DEPS + [
5691 ":XNNPACK",
5692 ":qs8_mobilenet_v1",
5693 ":qs8_mobilenet_v2",
5694 ],
5695)
5696
5697xnnpack_benchmark(
Marat Dukhanc068bb62019-10-04 13:24:39 -07005698 name = "end2end_bench",
5699 srcs = ["bench/end2end.cc"],
5700 deps = [
5701 ":XNNPACK",
Frank Barchardc712fa42019-10-31 14:00:21 -07005702 ":bench_utils",
Marat Dukhan270a2c42020-06-26 16:45:52 -07005703 ":fp16_mobilenet_v1",
5704 ":fp16_mobilenet_v2",
Marat Dukhan66f3ccd2020-09-14 16:09:38 -07005705 ":fp16_mobilenet_v3_large",
5706 ":fp16_mobilenet_v3_small",
Marat Dukhan270a2c42020-06-26 16:45:52 -07005707 ":fp32_mobilenet_v1",
5708 ":fp32_mobilenet_v2",
5709 ":fp32_mobilenet_v3_large",
5710 ":fp32_mobilenet_v3_small",
Marat Dukhan4cea2322021-03-09 09:35:36 -08005711 ":fp32_sparse_mobilenet_v1",
5712 ":fp32_sparse_mobilenet_v2",
5713 ":fp32_sparse_mobilenet_v3_large",
5714 ":fp32_sparse_mobilenet_v3_small",
Marat Dukhan0743cdf2020-08-04 18:52:07 -07005715 ":qs8_mobilenet_v1",
Marat Dukhan70a96182020-09-03 17:13:58 -07005716 ":qs8_mobilenet_v2",
Marat Dukhan12a23bb2021-03-08 08:13:21 -08005717 ":qu8_mobilenet_v1",
Marat Dukhanc068bb62019-10-04 13:24:39 -07005718 "@pthreadpool",
5719 ],
5720)
5721
Marat Dukhan6adff4e2019-10-14 18:32:07 -07005722#################### Accuracy evaluation for math functions ####################
5723
5724xnnpack_benchmark(
Marat Dukhancf67ec62020-12-14 09:56:37 -08005725 name = "f32_exp_ulp_eval",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07005726 srcs = [
Marat Dukhancf67ec62020-12-14 09:56:37 -08005727 "eval/f32-exp-ulp.cc",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07005728 "src/xnnpack/AlignedAllocator.h",
5729 ] + ACCURACY_EVAL_HDRS,
Marat Dukhan3a305212020-12-06 19:24:27 -08005730 deps = ACCURACY_EVAL_DEPS + [
5731 ":bench_utils",
5732 "@cpuinfo",
5733 ],
Marat Dukhan6adff4e2019-10-14 18:32:07 -07005734)
5735
Marat Dukhan515c9772019-10-17 18:07:57 -07005736xnnpack_benchmark(
Marat Dukhancf67ec62020-12-14 09:56:37 -08005737 name = "f32_expminus_ulp_eval",
Marat Dukhan515c9772019-10-17 18:07:57 -07005738 srcs = [
Marat Dukhancf67ec62020-12-14 09:56:37 -08005739 "eval/f32-expminus-ulp.cc",
Marat Dukhan515c9772019-10-17 18:07:57 -07005740 "src/xnnpack/AlignedAllocator.h",
5741 ] + ACCURACY_EVAL_HDRS,
Marat Dukhan3a305212020-12-06 19:24:27 -08005742 deps = ACCURACY_EVAL_DEPS + [
5743 ":bench_utils",
5744 "@cpuinfo",
5745 ],
Marat Dukhan515c9772019-10-17 18:07:57 -07005746)
5747
Marat Dukhan98ba4412019-10-23 02:14:28 -07005748xnnpack_benchmark(
Marat Dukhancf67ec62020-12-14 09:56:37 -08005749 name = "f32_expm1minus_ulp_eval",
Marat Dukhana438aca2020-11-20 15:45:01 -08005750 srcs = [
Marat Dukhancf67ec62020-12-14 09:56:37 -08005751 "eval/f32-expm1minus-ulp.cc",
Marat Dukhana438aca2020-11-20 15:45:01 -08005752 "src/xnnpack/AlignedAllocator.h",
5753 ] + ACCURACY_EVAL_HDRS,
Marat Dukhande390d42020-11-29 19:32:18 -08005754 deps = ACCURACY_EVAL_DEPS + [
5755 ":bench_utils",
Marat Dukhan3a305212020-12-06 19:24:27 -08005756 "@cpuinfo",
Marat Dukhande390d42020-11-29 19:32:18 -08005757 ],
Marat Dukhana438aca2020-11-20 15:45:01 -08005758)
5759
5760xnnpack_benchmark(
Marat Dukhancf67ec62020-12-14 09:56:37 -08005761 name = "f32_extexp_ulp_eval",
Marat Dukhan98ba4412019-10-23 02:14:28 -07005762 srcs = [
Marat Dukhancf67ec62020-12-14 09:56:37 -08005763 "eval/f32-extexp-ulp.cc",
Marat Dukhan98ba4412019-10-23 02:14:28 -07005764 "src/xnnpack/AlignedAllocator.h",
5765 ] + ACCURACY_EVAL_HDRS,
Marat Dukhan3a305212020-12-06 19:24:27 -08005766 deps = ACCURACY_EVAL_DEPS + [
5767 ":bench_utils",
5768 "@cpuinfo",
5769 ],
Marat Dukhan98ba4412019-10-23 02:14:28 -07005770)
5771
Marat Dukhanf44f0222020-12-14 11:53:27 -08005772xnnpack_benchmark(
5773 name = "f32_sigmoid_ulp_eval",
5774 srcs = [
5775 "eval/f32-sigmoid-ulp.cc",
5776 "src/xnnpack/AlignedAllocator.h",
5777 ] + ACCURACY_EVAL_HDRS,
5778 deps = ACCURACY_EVAL_DEPS + [
5779 ":bench_utils",
5780 "@cpuinfo",
5781 ],
5782)
5783
5784xnnpack_benchmark(
5785 name = "f32_sqrt_ulp_eval",
5786 srcs = [
5787 "eval/f32-sqrt-ulp.cc",
5788 "src/xnnpack/AlignedAllocator.h",
5789 ] + ACCURACY_EVAL_HDRS,
5790 deps = ACCURACY_EVAL_DEPS + [
5791 ":bench_utils",
5792 "@cpuinfo",
5793 ],
5794)
5795
5796################### Accuracy verification for math functions ##################
5797
5798xnnpack_unit_test(
Marat Dukhanf7291fc2020-12-15 11:02:50 -08005799 name = "f32_exp_eval",
5800 srcs = [
5801 "eval/f32-exp.cc",
5802 "src/xnnpack/AlignedAllocator.h",
5803 "src/xnnpack/math-stubs.h",
5804 ] + MICROKERNEL_TEST_HDRS,
5805 automatic = False,
5806 deps = MICROKERNEL_TEST_DEPS,
5807)
5808
5809xnnpack_unit_test(
Marat Dukhanf44f0222020-12-14 11:53:27 -08005810 name = "f32_expm1minus_eval",
5811 srcs = [
5812 "eval/f32-expm1minus.cc",
5813 "src/xnnpack/AlignedAllocator.h",
5814 "src/xnnpack/math-stubs.h",
5815 ] + MICROKERNEL_TEST_HDRS,
5816 automatic = False,
5817 deps = MICROKERNEL_TEST_DEPS,
5818)
5819
Marat Dukhan8853b822020-05-07 12:19:01 -07005820xnnpack_unit_test(
Marat Dukhand28a5a22020-12-14 15:27:22 -08005821 name = "f32_expminus_eval",
5822 srcs = [
5823 "eval/f32-expminus.cc",
5824 "src/xnnpack/AlignedAllocator.h",
5825 "src/xnnpack/math-stubs.h",
5826 ] + MICROKERNEL_TEST_HDRS,
5827 automatic = False,
5828 deps = MICROKERNEL_TEST_DEPS,
5829)
5830
5831xnnpack_unit_test(
Marat Dukhan8853b822020-05-07 12:19:01 -07005832 name = "f32_roundne_eval",
5833 srcs = [
5834 "eval/f32-roundne.cc",
5835 "src/xnnpack/AlignedAllocator.h",
5836 "src/xnnpack/math-stubs.h",
5837 ] + MICROKERNEL_TEST_HDRS,
Marat Dukhan22eed3d2020-05-11 20:13:37 -07005838 automatic = False,
Marat Dukhan8853b822020-05-07 12:19:01 -07005839 deps = MICROKERNEL_TEST_DEPS,
5840)
5841
Marat Dukhan2dbb9442020-05-12 20:43:43 -07005842xnnpack_unit_test(
Marat Dukhanc9852ba2020-05-13 17:21:29 -07005843 name = "f32_roundd_eval",
5844 srcs = [
5845 "eval/f32-roundd.cc",
5846 "src/xnnpack/AlignedAllocator.h",
5847 "src/xnnpack/math-stubs.h",
5848 ] + MICROKERNEL_TEST_HDRS,
5849 automatic = False,
5850 deps = MICROKERNEL_TEST_DEPS,
5851)
5852
5853xnnpack_unit_test(
5854 name = "f32_roundu_eval",
5855 srcs = [
5856 "eval/f32-roundu.cc",
5857 "src/xnnpack/AlignedAllocator.h",
5858 "src/xnnpack/math-stubs.h",
5859 ] + MICROKERNEL_TEST_HDRS,
5860 automatic = False,
5861 deps = MICROKERNEL_TEST_DEPS,
5862)
5863
5864xnnpack_unit_test(
Marat Dukhan2dbb9442020-05-12 20:43:43 -07005865 name = "f32_roundz_eval",
5866 srcs = [
5867 "eval/f32-roundz.cc",
5868 "src/xnnpack/AlignedAllocator.h",
5869 "src/xnnpack/math-stubs.h",
5870 ] + MICROKERNEL_TEST_HDRS,
Marat Dukhanc9852ba2020-05-13 17:21:29 -07005871 automatic = False,
Marat Dukhan2dbb9442020-05-12 20:43:43 -07005872 deps = MICROKERNEL_TEST_DEPS,
5873)
5874
Marat Dukhan08c4a432019-10-03 09:29:21 -07005875######################### Unit tests for micro-kernels #########################
5876
5877xnnpack_unit_test(
Marat Dukhan6674d692021-05-05 22:27:00 -07005878 name = "f16_dwconv_minmax_test",
5879 srcs = [
5880 "test/f16-dwconv-minmax.cc",
5881 "test/dwconv-microkernel-tester.h",
5882 "src/xnnpack/AlignedAllocator.h",
5883 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
5884 deps = MICROKERNEL_TEST_DEPS + [":packing"],
5885)
5886
5887xnnpack_unit_test(
5888 name = "f16_gavgpool_minmax_test",
5889 srcs = [
5890 "test/f16-gavgpool-minmax.cc",
5891 "test/gavgpool-microkernel-tester.h",
5892 "src/xnnpack/AlignedAllocator.h",
5893 ] + MICROKERNEL_TEST_HDRS,
5894 deps = MICROKERNEL_TEST_DEPS,
5895)
5896
5897xnnpack_unit_test(
Marat Dukhande06f492020-04-09 00:19:31 -07005898 name = "f16_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005899 srcs = [
Marat Dukhande06f492020-04-09 00:19:31 -07005900 "test/f16-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005901 "test/gemm-microkernel-tester.h",
5902 "src/xnnpack/AlignedAllocator.h",
5903 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005904 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005905)
5906
5907xnnpack_unit_test(
Marat Dukhan6674d692021-05-05 22:27:00 -07005908 name = "f16_igemm_minmax_test",
5909 srcs = [
5910 "test/f16-igemm-minmax.cc",
5911 "test/gemm-microkernel-tester.h",
5912 "src/xnnpack/AlignedAllocator.h",
5913 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
5914 deps = MICROKERNEL_TEST_DEPS + [":packing"],
5915)
5916
5917xnnpack_unit_test(
Marat Dukhan355ab432020-04-09 19:01:52 -07005918 name = "f16_spmm_minmax_test",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08005919 srcs = [
Marat Dukhan355ab432020-04-09 19:01:52 -07005920 "test/f16-spmm-minmax.cc",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08005921 "test/spmm-microkernel-tester.h",
5922 "src/xnnpack/AlignedAllocator.h",
5923 ] + MICROKERNEL_TEST_HDRS,
5924 deps = MICROKERNEL_TEST_DEPS,
5925)
5926
5927xnnpack_unit_test(
Marat Dukhan6674d692021-05-05 22:27:00 -07005928 name = "f16_vadd_minmax_test",
5929 srcs = [
5930 "test/f16-vadd-minmax.cc",
5931 "test/vbinary-microkernel-tester.h",
5932 ] + MICROKERNEL_TEST_HDRS,
5933 deps = MICROKERNEL_TEST_DEPS,
5934)
5935
5936xnnpack_unit_test(
5937 name = "f16_vaddc_minmax_test",
5938 srcs = [
5939 "test/f16-vaddc-minmax.cc",
5940 "test/vbinaryc-microkernel-tester.h",
5941 ] + MICROKERNEL_TEST_HDRS,
5942 deps = MICROKERNEL_TEST_DEPS,
5943)
5944
5945xnnpack_unit_test(
5946 name = "f16_vclamp_test",
5947 srcs = [
5948 "test/f16-vclamp.cc",
5949 "test/vclamp-microkernel-tester.h",
5950 ] + MICROKERNEL_TEST_HDRS,
5951 deps = MICROKERNEL_TEST_DEPS,
5952)
5953
5954xnnpack_unit_test(
5955 name = "f16_vdiv_minmax_test",
5956 srcs = [
5957 "test/f16-vdiv-minmax.cc",
5958 "test/vbinary-microkernel-tester.h",
5959 ] + MICROKERNEL_TEST_HDRS,
5960 deps = MICROKERNEL_TEST_DEPS,
5961)
5962
5963xnnpack_unit_test(
5964 name = "f16_vdivc_minmax_test",
5965 srcs = [
5966 "test/f16-vdivc-minmax.cc",
5967 "test/vbinaryc-microkernel-tester.h",
5968 ] + MICROKERNEL_TEST_HDRS,
5969 deps = MICROKERNEL_TEST_DEPS,
5970)
5971
5972xnnpack_unit_test(
5973 name = "f16_vrdivc_minmax_test",
5974 srcs = [
5975 "test/f16-vrdivc-minmax.cc",
5976 "test/vbinaryc-microkernel-tester.h",
5977 ] + MICROKERNEL_TEST_HDRS,
5978 deps = MICROKERNEL_TEST_DEPS,
5979)
5980
5981xnnpack_unit_test(
5982 name = "f16_vhswish_test",
5983 srcs = [
5984 "test/f16-vhswish.cc",
5985 "test/vhswish-microkernel-tester.h",
5986 ] + MICROKERNEL_TEST_HDRS,
5987 deps = MICROKERNEL_TEST_DEPS,
5988)
5989
5990xnnpack_unit_test(
5991 name = "f16_vmax_test",
5992 srcs = [
5993 "test/f16-vmax.cc",
5994 "test/vbinary-microkernel-tester.h",
5995 ] + MICROKERNEL_TEST_HDRS,
5996 deps = MICROKERNEL_TEST_DEPS,
5997)
5998
5999xnnpack_unit_test(
6000 name = "f16_vmaxc_test",
6001 srcs = [
6002 "test/f16-vmaxc.cc",
6003 "test/vbinaryc-microkernel-tester.h",
6004 ] + MICROKERNEL_TEST_HDRS,
6005 deps = MICROKERNEL_TEST_DEPS,
6006)
6007
6008xnnpack_unit_test(
6009 name = "f16_vmin_test",
6010 srcs = [
6011 "test/f16-vmin.cc",
6012 "test/vbinary-microkernel-tester.h",
6013 ] + MICROKERNEL_TEST_HDRS,
6014 deps = MICROKERNEL_TEST_DEPS,
6015)
6016
6017xnnpack_unit_test(
6018 name = "f16_vminc_test",
6019 srcs = [
6020 "test/f16-vminc.cc",
6021 "test/vbinaryc-microkernel-tester.h",
6022 ] + MICROKERNEL_TEST_HDRS,
6023 deps = MICROKERNEL_TEST_DEPS,
6024)
6025
6026xnnpack_unit_test(
6027 name = "f16_vmul_minmax_test",
6028 srcs = [
6029 "test/f16-vmul-minmax.cc",
6030 "test/vbinary-microkernel-tester.h",
6031 ] + MICROKERNEL_TEST_HDRS,
6032 deps = MICROKERNEL_TEST_DEPS,
6033)
6034
6035xnnpack_unit_test(
6036 name = "f16_vmulc_minmax_test",
6037 srcs = [
6038 "test/f16-vmulc-minmax.cc",
6039 "test/vbinaryc-microkernel-tester.h",
6040 ] + MICROKERNEL_TEST_HDRS,
6041 deps = MICROKERNEL_TEST_DEPS,
6042)
6043
6044xnnpack_unit_test(
6045 name = "f16_vmulcaddc_minmax_test",
6046 srcs = [
6047 "test/f16-vmulcaddc-minmax.cc",
6048 "test/vmulcaddc-microkernel-tester.h",
6049 "src/xnnpack/AlignedAllocator.h",
6050 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
6051 deps = MICROKERNEL_TEST_DEPS + [":packing"],
6052)
6053
6054xnnpack_unit_test(
6055 name = "f16_vsub_minmax_test",
6056 srcs = [
6057 "test/f16-vsub-minmax.cc",
6058 "test/vbinary-microkernel-tester.h",
6059 ] + MICROKERNEL_TEST_HDRS,
6060 deps = MICROKERNEL_TEST_DEPS,
6061)
6062
6063xnnpack_unit_test(
6064 name = "f16_vsubc_minmax_test",
6065 srcs = [
6066 "test/f16-vsubc-minmax.cc",
6067 "test/vbinaryc-microkernel-tester.h",
6068 ] + MICROKERNEL_TEST_HDRS,
6069 deps = MICROKERNEL_TEST_DEPS,
6070)
6071
6072xnnpack_unit_test(
6073 name = "f16_vrsubc_minmax_test",
6074 srcs = [
6075 "test/f16-vrsubc-minmax.cc",
6076 "test/vbinaryc-microkernel-tester.h",
6077 ] + MICROKERNEL_TEST_HDRS,
6078 deps = MICROKERNEL_TEST_DEPS,
6079)
6080
6081xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07006082 name = "f32_argmaxpool_test",
6083 srcs = [
6084 "test/f32-argmaxpool.cc",
6085 "test/argmaxpool-microkernel-tester.h",
6086 "src/xnnpack/AlignedAllocator.h",
6087 ] + MICROKERNEL_TEST_HDRS,
6088 deps = MICROKERNEL_TEST_DEPS,
6089)
6090
6091xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07006092 name = "f32_avgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006093 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07006094 "test/f32-avgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006095 "test/avgpool-microkernel-tester.h",
6096 "src/xnnpack/AlignedAllocator.h",
6097 ] + MICROKERNEL_TEST_HDRS,
6098 deps = MICROKERNEL_TEST_DEPS,
6099)
6100
6101xnnpack_unit_test(
Marat Dukhan660fd192020-03-10 04:55:30 -07006102 name = "f32_ibilinear_test",
Marat Dukhan35dacfb2019-11-07 19:18:16 -08006103 srcs = [
Marat Dukhan660fd192020-03-10 04:55:30 -07006104 "test/f32-ibilinear.cc",
6105 "test/ibilinear-microkernel-tester.h",
Marat Dukhan35dacfb2019-11-07 19:18:16 -08006106 "src/xnnpack/AlignedAllocator.h",
6107 ] + MICROKERNEL_TEST_HDRS,
6108 deps = MICROKERNEL_TEST_DEPS,
6109)
6110
6111xnnpack_unit_test(
XNNPACK Team21432672020-10-19 19:58:48 -07006112 name = "f32_ibilinear_chw_test",
6113 srcs = [
6114 "test/f32-ibilinear-chw.cc",
XNNPACK Team6be46b22020-10-22 23:34:54 -07006115 "test/ibilinear-microkernel-tester.h",
XNNPACK Team21432672020-10-19 19:58:48 -07006116 "src/xnnpack/AlignedAllocator.h",
6117 ] + MICROKERNEL_TEST_HDRS,
6118 deps = MICROKERNEL_TEST_DEPS,
6119)
6120
6121xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07006122 name = "f32_igemm_test",
6123 srcs = [
6124 "test/f32-igemm.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 Dukhan163a7e62020-04-09 04:19:26 -07006129)
6130
6131xnnpack_unit_test(
Marat Dukhan467f6362020-05-22 23:21:55 -07006132 name = "f32_igemm_relu_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006133 srcs = [
Marat Dukhan467f6362020-05-22 23:21:55 -07006134 "test/f32-igemm-relu.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006135 "test/gemm-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 Dukhane207b7b2020-05-28 16:27:42 -07006142 name = "f32_igemm_minmax_test",
6143 srcs = [
6144 "test/f32-igemm-minmax.cc",
6145 "test/gemm-microkernel-tester.h",
6146 "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 Dukhane207b7b2020-05-28 16:27:42 -07006149)
6150
6151xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07006152 name = "f32_conv_hwc_test",
6153 srcs = [
6154 "test/f32-conv-hwc.cc",
6155 "test/conv-hwc-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 Dukhan08c4a432019-10-03 09:29:21 -07006159)
6160
6161xnnpack_unit_test(
Marat Dukhan1f29b802020-05-15 23:46:39 -07006162 name = "f32_conv_hwc2chw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006163 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07006164 "test/f32-conv-hwc2chw.cc",
6165 "test/conv-hwc2chw-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006166 "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 Dukhan163a7e62020-04-09 04:19:26 -07006172 name = "f32_dwconv_test",
6173 srcs = [
6174 "test/f32-dwconv.cc",
6175 "test/dwconv-microkernel-tester.h",
6176 "src/xnnpack/AlignedAllocator.h",
6177 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07006178 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan163a7e62020-04-09 04:19:26 -07006179)
6180
6181xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07006182 name = "f32_dwconv_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006183 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07006184 "test/f32-dwconv-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006185 "test/dwconv-microkernel-tester.h",
6186 "src/xnnpack/AlignedAllocator.h",
6187 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07006188 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07006189)
6190
6191xnnpack_unit_test(
Marat Dukhanbf715f92020-10-23 20:17:00 -07006192 name = "f32_dwconv2d_chw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006193 srcs = [
Marat Dukhanbf715f92020-10-23 20:17:00 -07006194 "test/f32-dwconv2d-chw.cc",
6195 "test/dwconv2d-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006196 "src/xnnpack/AlignedAllocator.h",
6197 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07006198 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07006199)
6200
6201xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07006202 name = "f32_gavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006203 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07006204 "test/f32-gavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006205 "test/gavgpool-microkernel-tester.h",
6206 "src/xnnpack/AlignedAllocator.h",
6207 ] + MICROKERNEL_TEST_HDRS,
6208 deps = MICROKERNEL_TEST_DEPS,
6209)
6210
6211xnnpack_unit_test(
Marat Dukhan1f29b802020-05-15 23:46:39 -07006212 name = "f32_gavgpool_cw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006213 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07006214 "test/f32-gavgpool-cw.cc",
6215 "test/gavgpool-cw-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006216 "src/xnnpack/AlignedAllocator.h",
6217 ] + MICROKERNEL_TEST_HDRS,
6218 deps = MICROKERNEL_TEST_DEPS,
6219)
6220
6221xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07006222 name = "f32_gemm_test",
6223 srcs = [
6224 "test/f32-gemm.cc",
6225 "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 Dukhan163a7e62020-04-09 04:19:26 -07006229)
6230
6231xnnpack_unit_test(
Marat Dukhan467f6362020-05-22 23:21:55 -07006232 name = "f32_gemm_relu_test",
6233 srcs = [
6234 "test/f32-gemm-relu.cc",
6235 "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 Dukhan467f6362020-05-22 23:21:55 -07006239)
6240
6241xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07006242 name = "f32_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006243 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07006244 "test/f32-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006245 "test/gemm-microkernel-tester.h",
6246 "src/xnnpack/AlignedAllocator.h",
6247 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07006248 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07006249)
6250
6251xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07006252 name = "f32_gemminc_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006253 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07006254 "test/f32-gemminc-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006255 "test/gemm-microkernel-tester.h",
6256 "src/xnnpack/AlignedAllocator.h",
6257 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07006258 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07006259)
6260
6261xnnpack_unit_test(
Marat Dukhan6674d692021-05-05 22:27:00 -07006262 name = "f32_vhswish_test",
Frank Barchardb1966592020-05-12 13:47:06 -07006263 srcs = [
Marat Dukhan6674d692021-05-05 22:27:00 -07006264 "test/f32-vhswish.cc",
6265 "test/vhswish-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006266 ] + MICROKERNEL_TEST_HDRS,
6267 deps = MICROKERNEL_TEST_DEPS,
6268)
6269
6270xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07006271 name = "f32_maxpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006272 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07006273 "test/f32-maxpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006274 "test/maxpool-microkernel-tester.h",
6275 ] + MICROKERNEL_TEST_HDRS,
6276 deps = MICROKERNEL_TEST_DEPS,
6277)
6278
6279xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07006280 name = "f32_pavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006281 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07006282 "test/f32-pavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006283 "test/avgpool-microkernel-tester.h",
6284 "src/xnnpack/AlignedAllocator.h",
6285 ] + MICROKERNEL_TEST_HDRS,
6286 deps = MICROKERNEL_TEST_DEPS,
6287)
6288
6289xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07006290 name = "f32_ppmm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006291 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07006292 "test/f32-ppmm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006293 "test/gemm-microkernel-tester.h",
6294 "src/xnnpack/AlignedAllocator.h",
6295 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07006296 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07006297)
6298
6299xnnpack_unit_test(
Frank Barchardb1966592020-05-12 13:47:06 -07006300 name = "f16_prelu_test",
6301 srcs = [
6302 "test/f16-prelu.cc",
6303 "test/prelu-microkernel-tester.h",
6304 "src/xnnpack/AlignedAllocator.h",
6305 ] + MICROKERNEL_TEST_HDRS,
6306 deps = MICROKERNEL_TEST_DEPS,
6307)
6308
6309xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07006310 name = "f32_prelu_test",
6311 srcs = [
6312 "test/f32-prelu.cc",
6313 "test/prelu-microkernel-tester.h",
6314 "src/xnnpack/AlignedAllocator.h",
6315 ] + MICROKERNEL_TEST_HDRS,
6316 deps = MICROKERNEL_TEST_DEPS,
6317)
6318
6319xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07006320 name = "f32_raddexpminusmax_test",
6321 srcs = [
6322 "test/f32-raddexpminusmax.cc",
6323 "test/raddexpminusmax-microkernel-tester.h",
6324 ] + MICROKERNEL_TEST_HDRS,
6325 deps = MICROKERNEL_TEST_DEPS,
6326)
6327
6328xnnpack_unit_test(
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07006329 name = "f32_raddextexp_test",
6330 srcs = [
6331 "test/f32-raddextexp.cc",
6332 "test/raddextexp-microkernel-tester.h",
6333 ] + MICROKERNEL_TEST_HDRS,
6334 deps = MICROKERNEL_TEST_DEPS,
6335)
6336
6337xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07006338 name = "f32_raddstoreexpminusmax_test",
6339 srcs = [
6340 "test/f32-raddstoreexpminusmax.cc",
6341 "test/raddstoreexpminusmax-microkernel-tester.h",
6342 ] + MICROKERNEL_TEST_HDRS,
6343 deps = MICROKERNEL_TEST_DEPS,
6344)
6345
6346xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07006347 name = "f32_rmax_test",
6348 srcs = [
6349 "test/f32-rmax.cc",
6350 "test/rmax-microkernel-tester.h",
6351 ] + MICROKERNEL_TEST_HDRS,
6352 deps = MICROKERNEL_TEST_DEPS,
6353)
6354
6355xnnpack_unit_test(
Marat Dukhan355ab432020-04-09 19:01:52 -07006356 name = "f32_spmm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006357 srcs = [
Marat Dukhan355ab432020-04-09 19:01:52 -07006358 "test/f32-spmm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006359 "test/spmm-microkernel-tester.h",
6360 "src/xnnpack/AlignedAllocator.h",
6361 ] + MICROKERNEL_TEST_HDRS,
6362 deps = MICROKERNEL_TEST_DEPS,
6363)
6364
6365xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07006366 name = "f32_vabs_test",
6367 srcs = [
6368 "test/f32-vabs.cc",
6369 "test/vunary-microkernel-tester.h",
6370 ] + MICROKERNEL_TEST_HDRS,
6371 deps = MICROKERNEL_TEST_DEPS,
6372)
6373
6374xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07006375 name = "f32_vadd_test",
6376 srcs = [
6377 "test/f32-vadd.cc",
6378 "test/vbinary-microkernel-tester.h",
6379 ] + MICROKERNEL_TEST_HDRS,
6380 deps = MICROKERNEL_TEST_DEPS,
6381)
6382
6383xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006384 name = "f32_vadd_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006385 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006386 "test/f32-vadd-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08006387 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08006388 ] + MICROKERNEL_TEST_HDRS,
6389 deps = MICROKERNEL_TEST_DEPS,
6390)
6391
6392xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07006393 name = "f32_vadd_relu_test",
6394 srcs = [
6395 "test/f32-vadd-relu.cc",
6396 "test/vbinary-microkernel-tester.h",
6397 ] + MICROKERNEL_TEST_HDRS,
6398 deps = MICROKERNEL_TEST_DEPS,
6399)
6400
6401xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07006402 name = "f32_vaddc_test",
6403 srcs = [
6404 "test/f32-vaddc.cc",
6405 "test/vbinaryc-microkernel-tester.h",
6406 ] + MICROKERNEL_TEST_HDRS,
6407 deps = MICROKERNEL_TEST_DEPS,
6408)
6409
6410xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006411 name = "f32_vaddc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08006412 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006413 "test/f32-vaddc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08006414 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006415 ] + MICROKERNEL_TEST_HDRS,
6416 deps = MICROKERNEL_TEST_DEPS,
6417)
6418
6419xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07006420 name = "f32_vaddc_relu_test",
6421 srcs = [
6422 "test/f32-vaddc-relu.cc",
6423 "test/vbinaryc-microkernel-tester.h",
6424 ] + MICROKERNEL_TEST_HDRS,
6425 deps = MICROKERNEL_TEST_DEPS,
6426)
6427
6428xnnpack_unit_test(
Marat Dukhan6674d692021-05-05 22:27:00 -07006429 name = "f32_vclamp_test",
6430 srcs = [
6431 "test/f32-vclamp.cc",
6432 "test/vclamp-microkernel-tester.h",
6433 ] + MICROKERNEL_TEST_HDRS,
6434 deps = MICROKERNEL_TEST_DEPS,
6435)
6436
6437xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07006438 name = "f32_vdiv_test",
6439 srcs = [
6440 "test/f32-vdiv.cc",
6441 "test/vbinary-microkernel-tester.h",
6442 ] + MICROKERNEL_TEST_HDRS,
6443 deps = MICROKERNEL_TEST_DEPS,
6444)
6445
6446xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006447 name = "f32_vdiv_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08006448 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006449 "test/f32-vdiv-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08006450 "test/vbinary-microkernel-tester.h",
6451 ] + MICROKERNEL_TEST_HDRS,
6452 deps = MICROKERNEL_TEST_DEPS,
6453)
6454
6455xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07006456 name = "f32_vdiv_relu_test",
6457 srcs = [
6458 "test/f32-vdiv-relu.cc",
6459 "test/vbinary-microkernel-tester.h",
6460 ] + MICROKERNEL_TEST_HDRS,
6461 deps = MICROKERNEL_TEST_DEPS,
6462)
6463
6464xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07006465 name = "f32_vdivc_test",
6466 srcs = [
6467 "test/f32-vdivc.cc",
6468 "test/vbinaryc-microkernel-tester.h",
6469 ] + MICROKERNEL_TEST_HDRS,
6470 deps = MICROKERNEL_TEST_DEPS,
6471)
6472
6473xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006474 name = "f32_vdivc_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08006475 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006476 "test/f32-vdivc-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08006477 "test/vbinaryc-microkernel-tester.h",
6478 ] + MICROKERNEL_TEST_HDRS,
6479 deps = MICROKERNEL_TEST_DEPS,
6480)
6481
6482xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07006483 name = "f32_vdivc_relu_test",
6484 srcs = [
6485 "test/f32-vdivc-relu.cc",
6486 "test/vbinaryc-microkernel-tester.h",
6487 ] + MICROKERNEL_TEST_HDRS,
6488 deps = MICROKERNEL_TEST_DEPS,
6489)
6490
6491xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07006492 name = "f32_vrdivc_test",
6493 srcs = [
6494 "test/f32-vrdivc.cc",
6495 "test/vbinaryc-microkernel-tester.h",
6496 ] + MICROKERNEL_TEST_HDRS,
6497 deps = MICROKERNEL_TEST_DEPS,
6498)
6499
6500xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006501 name = "f32_vrdivc_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08006502 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006503 "test/f32-vrdivc-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08006504 "test/vbinaryc-microkernel-tester.h",
6505 ] + MICROKERNEL_TEST_HDRS,
6506 deps = MICROKERNEL_TEST_DEPS,
6507)
6508
6509xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07006510 name = "f32_vrdivc_relu_test",
6511 srcs = [
6512 "test/f32-vrdivc-relu.cc",
6513 "test/vbinaryc-microkernel-tester.h",
6514 ] + MICROKERNEL_TEST_HDRS,
6515 deps = MICROKERNEL_TEST_DEPS,
6516)
6517
6518xnnpack_unit_test(
Marat Dukhaned6baaf2020-12-01 15:07:08 -08006519 name = "f32_velu_test",
6520 srcs = [
6521 "test/f32-velu.cc",
6522 "test/vunary-microkernel-tester.h",
6523 ] + MICROKERNEL_TEST_HDRS,
6524 deps = MICROKERNEL_TEST_DEPS,
6525)
6526
6527xnnpack_unit_test(
Marat Dukhan403b7d42019-12-05 12:49:11 -08006528 name = "f32_vmax_test",
6529 srcs = [
6530 "test/f32-vmax.cc",
6531 "test/vbinary-microkernel-tester.h",
6532 ] + MICROKERNEL_TEST_HDRS,
6533 deps = MICROKERNEL_TEST_DEPS,
6534)
6535
6536xnnpack_unit_test(
6537 name = "f32_vmaxc_test",
6538 srcs = [
6539 "test/f32-vmaxc.cc",
6540 "test/vbinaryc-microkernel-tester.h",
6541 ] + MICROKERNEL_TEST_HDRS,
6542 deps = MICROKERNEL_TEST_DEPS,
6543)
6544
6545xnnpack_unit_test(
6546 name = "f32_vmin_test",
6547 srcs = [
6548 "test/f32-vmin.cc",
6549 "test/vbinary-microkernel-tester.h",
6550 ] + MICROKERNEL_TEST_HDRS,
6551 deps = MICROKERNEL_TEST_DEPS,
6552)
6553
6554xnnpack_unit_test(
6555 name = "f32_vminc_test",
6556 srcs = [
6557 "test/f32-vminc.cc",
6558 "test/vbinaryc-microkernel-tester.h",
6559 ] + MICROKERNEL_TEST_HDRS,
6560 deps = MICROKERNEL_TEST_DEPS,
6561)
6562
6563xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07006564 name = "f32_vmul_test",
6565 srcs = [
6566 "test/f32-vmul.cc",
6567 "test/vbinary-microkernel-tester.h",
6568 ] + MICROKERNEL_TEST_HDRS,
6569 deps = MICROKERNEL_TEST_DEPS,
6570)
6571
6572xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006573 name = "f32_vmul_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006574 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006575 "test/f32-vmul-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08006576 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08006577 ] + MICROKERNEL_TEST_HDRS,
6578 deps = MICROKERNEL_TEST_DEPS,
6579)
6580
6581xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07006582 name = "f32_vmul_relu_test",
6583 srcs = [
6584 "test/f32-vmul-relu.cc",
6585 "test/vbinary-microkernel-tester.h",
6586 ] + MICROKERNEL_TEST_HDRS,
6587 deps = MICROKERNEL_TEST_DEPS,
6588)
6589
6590xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07006591 name = "f32_vmulc_test",
6592 srcs = [
6593 "test/f32-vmulc.cc",
6594 "test/vbinaryc-microkernel-tester.h",
6595 ] + MICROKERNEL_TEST_HDRS,
6596 deps = MICROKERNEL_TEST_DEPS,
6597)
6598
6599xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006600 name = "f32_vmulc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08006601 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006602 "test/f32-vmulc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08006603 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006604 ] + MICROKERNEL_TEST_HDRS,
6605 deps = MICROKERNEL_TEST_DEPS,
6606)
6607
6608xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07006609 name = "f32_vmulc_relu_test",
6610 srcs = [
6611 "test/f32-vmulc-relu.cc",
6612 "test/vbinaryc-microkernel-tester.h",
6613 ] + MICROKERNEL_TEST_HDRS,
6614 deps = MICROKERNEL_TEST_DEPS,
6615)
6616
6617xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07006618 name = "f32_vmulcaddc_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006619 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07006620 "test/f32-vmulcaddc-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006621 "test/vmulcaddc-microkernel-tester.h",
6622 "src/xnnpack/AlignedAllocator.h",
6623 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07006624 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07006625)
6626
6627xnnpack_unit_test(
Marat Dukhan8cc7efe2020-06-10 16:24:27 -07006628 name = "f32_vlrelu_test",
6629 srcs = [
6630 "test/f32-vlrelu.cc",
6631 "test/vunary-microkernel-tester.h",
6632 ] + MICROKERNEL_TEST_HDRS,
6633 deps = MICROKERNEL_TEST_DEPS,
6634)
6635
6636xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07006637 name = "f32_vneg_test",
6638 srcs = [
6639 "test/f32-vneg.cc",
6640 "test/vunary-microkernel-tester.h",
6641 ] + MICROKERNEL_TEST_HDRS,
6642 deps = MICROKERNEL_TEST_DEPS,
6643)
6644
6645xnnpack_unit_test(
Marat Dukhan6674d692021-05-05 22:27:00 -07006646 name = "f32_vrelu_test",
6647 srcs = [
6648 "test/f32-vrelu.cc",
6649 "test/vunary-microkernel-tester.h",
6650 ] + MICROKERNEL_TEST_HDRS,
6651 deps = MICROKERNEL_TEST_DEPS,
6652)
6653
6654xnnpack_unit_test(
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07006655 name = "f32_vrndne_test",
6656 srcs = [
6657 "test/f32-vrndne.cc",
6658 "test/vunary-microkernel-tester.h",
6659 ] + MICROKERNEL_TEST_HDRS,
6660 deps = MICROKERNEL_TEST_DEPS,
6661)
6662
6663xnnpack_unit_test(
6664 name = "f32_vrndz_test",
6665 srcs = [
6666 "test/f32-vrndz.cc",
6667 "test/vunary-microkernel-tester.h",
6668 ] + MICROKERNEL_TEST_HDRS,
6669 deps = MICROKERNEL_TEST_DEPS,
6670)
6671
6672xnnpack_unit_test(
6673 name = "f32_vrndu_test",
6674 srcs = [
6675 "test/f32-vrndu.cc",
6676 "test/vunary-microkernel-tester.h",
6677 ] + MICROKERNEL_TEST_HDRS,
6678 deps = MICROKERNEL_TEST_DEPS,
6679)
6680
6681xnnpack_unit_test(
6682 name = "f32_vrndd_test",
6683 srcs = [
6684 "test/f32-vrndd.cc",
6685 "test/vunary-microkernel-tester.h",
6686 ] + MICROKERNEL_TEST_HDRS,
6687 deps = MICROKERNEL_TEST_DEPS,
6688)
6689
6690xnnpack_unit_test(
Marat Dukhan05ac8e32019-10-21 15:39:33 -07006691 name = "f32_vscale_test",
6692 srcs = [
6693 "test/f32-vscale.cc",
6694 "test/vscale-microkernel-tester.h",
6695 ] + MICROKERNEL_TEST_HDRS,
6696 deps = MICROKERNEL_TEST_DEPS,
6697)
6698
6699xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07006700 name = "f32_vscaleexpminusmax_test",
6701 srcs = [
6702 "test/f32-vscaleexpminusmax.cc",
6703 "test/vscaleexpminusmax-microkernel-tester.h",
6704 ] + MICROKERNEL_TEST_HDRS,
6705 deps = MICROKERNEL_TEST_DEPS,
6706)
6707
6708xnnpack_unit_test(
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07006709 name = "f32_vscaleextexp_test",
6710 srcs = [
6711 "test/f32-vscaleextexp.cc",
6712 "test/vscaleextexp-microkernel-tester.h",
6713 ] + MICROKERNEL_TEST_HDRS,
6714 deps = MICROKERNEL_TEST_DEPS,
6715)
6716
6717xnnpack_unit_test(
Marat Dukhan6674d692021-05-05 22:27:00 -07006718 name = "f32_vsigmoid_test",
6719 srcs = [
6720 "test/f32-vsigmoid.cc",
6721 "test/vunary-microkernel-tester.h",
6722 ] + MICROKERNEL_TEST_HDRS,
6723 deps = MICROKERNEL_TEST_DEPS,
6724)
6725
6726xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07006727 name = "f32_vsqr_test",
6728 srcs = [
6729 "test/f32-vsqr.cc",
6730 "test/vunary-microkernel-tester.h",
6731 ] + MICROKERNEL_TEST_HDRS,
6732 deps = MICROKERNEL_TEST_DEPS,
6733)
6734
6735xnnpack_unit_test(
Marat Dukhan13bafb02020-06-05 00:43:11 -07006736 name = "f32_vsqrdiff_test",
6737 srcs = [
6738 "test/f32-vsqrdiff.cc",
6739 "test/vbinary-microkernel-tester.h",
6740 ] + MICROKERNEL_TEST_HDRS,
6741 deps = MICROKERNEL_TEST_DEPS,
6742)
6743
6744xnnpack_unit_test(
6745 name = "f32_vsqrdiffc_test",
6746 srcs = [
6747 "test/f32-vsqrdiffc.cc",
6748 "test/vbinaryc-microkernel-tester.h",
6749 ] + MICROKERNEL_TEST_HDRS,
6750 deps = MICROKERNEL_TEST_DEPS,
6751)
6752
6753xnnpack_unit_test(
Marat Dukhanf4db2f32020-06-30 10:55:30 -07006754 name = "f32_vsqrt_test",
6755 srcs = [
6756 "test/f32-vsqrt.cc",
6757 "test/vunary-microkernel-tester.h",
6758 ] + MICROKERNEL_TEST_HDRS,
6759 deps = MICROKERNEL_TEST_DEPS,
6760)
6761
6762xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07006763 name = "f32_vsub_test",
6764 srcs = [
6765 "test/f32-vsub.cc",
6766 "test/vbinary-microkernel-tester.h",
6767 ] + MICROKERNEL_TEST_HDRS,
6768 deps = MICROKERNEL_TEST_DEPS,
6769)
6770
6771xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006772 name = "f32_vsub_minmax_test",
Marat Dukhan97579532019-10-18 16:40:39 -07006773 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006774 "test/f32-vsub-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08006775 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08006776 ] + MICROKERNEL_TEST_HDRS,
6777 deps = MICROKERNEL_TEST_DEPS,
6778)
6779
6780xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07006781 name = "f32_vsub_relu_test",
6782 srcs = [
6783 "test/f32-vsub-relu.cc",
6784 "test/vbinary-microkernel-tester.h",
6785 ] + MICROKERNEL_TEST_HDRS,
6786 deps = MICROKERNEL_TEST_DEPS,
6787)
6788
6789xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07006790 name = "f32_vsubc_test",
6791 srcs = [
6792 "test/f32-vsubc.cc",
6793 "test/vbinaryc-microkernel-tester.h",
6794 ] + MICROKERNEL_TEST_HDRS,
6795 deps = MICROKERNEL_TEST_DEPS,
6796)
6797
6798xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006799 name = "f32_vsubc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08006800 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006801 "test/f32-vsubc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08006802 "test/vbinaryc-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08006803 ] + MICROKERNEL_TEST_HDRS,
6804 deps = MICROKERNEL_TEST_DEPS,
6805)
6806
6807xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07006808 name = "f32_vsubc_relu_test",
6809 srcs = [
6810 "test/f32-vsubc-relu.cc",
6811 "test/vbinaryc-microkernel-tester.h",
6812 ] + MICROKERNEL_TEST_HDRS,
6813 deps = MICROKERNEL_TEST_DEPS,
6814)
6815
6816xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07006817 name = "f32_vrsubc_test",
6818 srcs = [
6819 "test/f32-vrsubc.cc",
6820 "test/vbinaryc-microkernel-tester.h",
6821 ] + MICROKERNEL_TEST_HDRS,
6822 deps = MICROKERNEL_TEST_DEPS,
6823)
6824
6825xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006826 name = "f32_vrsubc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08006827 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006828 "test/f32-vrsubc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08006829 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan97579532019-10-18 16:40:39 -07006830 ] + MICROKERNEL_TEST_HDRS,
6831 deps = MICROKERNEL_TEST_DEPS,
6832)
6833
6834xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07006835 name = "f32_vrsubc_relu_test",
6836 srcs = [
6837 "test/f32-vrsubc-relu.cc",
6838 "test/vbinaryc-microkernel-tester.h",
6839 ] + MICROKERNEL_TEST_HDRS,
6840 deps = MICROKERNEL_TEST_DEPS,
6841)
6842
6843xnnpack_unit_test(
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07006844 name = "qs8_dwconv_minmax_test",
6845 srcs = [
6846 "test/qs8-dwconv-minmax.cc",
6847 "test/dwconv-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 Dukhan4ed53f42020-08-06 01:12:55 -07006854 name = "qs8_gavgpool_minmax_test",
6855 srcs = [
6856 "test/qs8-gavgpool-minmax.cc",
6857 "test/gavgpool-microkernel-tester.h",
6858 "src/xnnpack/AlignedAllocator.h",
6859 ] + MICROKERNEL_TEST_HDRS,
6860 deps = MICROKERNEL_TEST_DEPS,
6861)
6862
6863xnnpack_unit_test(
Marat Dukhan595e1702020-07-31 10:12:52 -07006864 name = "qs8_gemm_minmax_test",
6865 srcs = [
6866 "test/qs8-gemm-minmax.cc",
6867 "test/gemm-microkernel-tester.h",
6868 "src/xnnpack/AlignedAllocator.h",
6869 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
6870 deps = MICROKERNEL_TEST_DEPS + [":packing"],
6871)
6872
6873xnnpack_unit_test(
Marat Dukhanf9480682020-07-31 14:50:24 -07006874 name = "qs8_igemm_minmax_test",
6875 srcs = [
6876 "test/qs8-igemm-minmax.cc",
6877 "test/gemm-microkernel-tester.h",
6878 "src/xnnpack/AlignedAllocator.h",
6879 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
6880 deps = MICROKERNEL_TEST_DEPS + [":packing"],
6881)
6882
6883xnnpack_unit_test(
6884 name = "qs8_requantization_test",
6885 srcs = [
6886 "src/xnnpack/requantization-stubs.h",
6887 "test/qs8-requantization.cc",
6888 "test/requantization-tester.h",
6889 ] + MICROKERNEL_TEST_HDRS,
6890 deps = MICROKERNEL_TEST_DEPS,
6891)
6892
6893xnnpack_unit_test(
Marat Dukhand9f3ad42020-08-10 12:30:58 -07006894 name = "qs8_vadd_minmax_test",
6895 srcs = [
6896 "test/qs8-vadd-minmax.cc",
6897 "test/vadd-microkernel-tester.h",
6898 ] + MICROKERNEL_TEST_HDRS,
6899 deps = MICROKERNEL_TEST_DEPS,
6900)
6901
6902xnnpack_unit_test(
Marat Dukhan0270d9f2020-08-11 00:56:46 -07006903 name = "qs8_vaddc_minmax_test",
6904 srcs = [
6905 "test/qs8-vaddc-minmax.cc",
6906 "test/vaddc-microkernel-tester.h",
6907 ] + MICROKERNEL_TEST_HDRS,
6908 deps = MICROKERNEL_TEST_DEPS,
6909)
6910
6911xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07006912 name = "qu8_avgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006913 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07006914 "test/qu8-avgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006915 "test/avgpool-microkernel-tester.h",
6916 "src/xnnpack/AlignedAllocator.h",
6917 ] + MICROKERNEL_TEST_HDRS,
6918 deps = MICROKERNEL_TEST_DEPS,
6919)
6920
6921xnnpack_unit_test(
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07006922 name = "qu8_dwconv_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006923 srcs = [
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07006924 "test/qu8-dwconv-minmax.cc",
6925 "test/dwconv-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006926 "src/xnnpack/AlignedAllocator.h",
6927 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07006928 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07006929)
6930
6931xnnpack_unit_test(
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07006932 name = "qu8_igemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006933 srcs = [
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07006934 "test/qu8-igemm-minmax.cc",
6935 "test/gemm-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006936 "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 Dukhan08b7a972020-07-14 18:17:29 -07006942 name = "qu8_gavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006943 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07006944 "test/qu8-gavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006945 "test/gavgpool-microkernel-tester.h",
6946 "src/xnnpack/AlignedAllocator.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_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006953 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07006954 "test/qu8-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006955 "test/gemm-microkernel-tester.h",
6956 "src/xnnpack/AlignedAllocator.h",
6957 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07006958 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07006959)
6960
6961xnnpack_unit_test(
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07006962 name = "qu8_requantization_test",
6963 srcs = [
6964 "src/xnnpack/requantization-stubs.h",
6965 "test/qu8-requantization.cc",
6966 "test/requantization-tester.h",
6967 ] + MICROKERNEL_TEST_HDRS,
6968 deps = MICROKERNEL_TEST_DEPS,
6969)
6970
6971xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07006972 name = "qu8_vadd_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006973 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07006974 "test/qu8-vadd-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006975 "test/vadd-microkernel-tester.h",
6976 ] + MICROKERNEL_TEST_HDRS,
6977 deps = MICROKERNEL_TEST_DEPS,
6978)
6979
6980xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07006981 name = "u8_lut32norm_test",
6982 srcs = [
6983 "test/u8-lut32norm.cc",
6984 "test/lut-norm-microkernel-tester.h",
6985 ] + MICROKERNEL_TEST_HDRS,
6986 deps = MICROKERNEL_TEST_DEPS,
6987)
6988
6989xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07006990 name = "u8_maxpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006991 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07006992 "test/u8-maxpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006993 "test/maxpool-microkernel-tester.h",
6994 ] + MICROKERNEL_TEST_HDRS,
6995 deps = MICROKERNEL_TEST_DEPS,
6996)
6997
6998xnnpack_unit_test(
6999 name = "u8_rmax_test",
7000 srcs = [
7001 "test/u8-rmax.cc",
7002 "test/rmax-microkernel-tester.h",
7003 ] + MICROKERNEL_TEST_HDRS,
7004 deps = MICROKERNEL_TEST_DEPS,
7005)
7006
7007xnnpack_unit_test(
Marat Dukhan6674d692021-05-05 22:27:00 -07007008 name = "u8_vclamp_test",
7009 srcs = [
7010 "test/u8-vclamp.cc",
7011 "test/vclamp-microkernel-tester.h",
7012 ] + MICROKERNEL_TEST_HDRS,
7013 deps = MICROKERNEL_TEST_DEPS,
7014)
7015
7016xnnpack_unit_test(
Marat Dukhanad71b9a2020-11-20 00:01:51 -08007017 name = "x32_depthtospace2d_chw2hwc_test",
Yury Kartynnike7841862020-11-04 18:22:18 -08007018 srcs = [
Marat Dukhanad71b9a2020-11-20 00:01:51 -08007019 "test/x32-depthtospace2d-chw2hwc.cc",
7020 "test/depthtospace-microkernel-tester.h",
Yury Kartynnike7841862020-11-04 18:22:18 -08007021 ] + MICROKERNEL_TEST_HDRS,
7022 deps = MICROKERNEL_TEST_DEPS,
7023)
7024
7025xnnpack_unit_test(
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -07007026 name = "x32_fill_test",
7027 srcs = [
7028 "test/x32-fill.cc",
7029 "test/fill-microkernel-tester.h",
7030 ] + MICROKERNEL_TEST_HDRS,
7031 deps = MICROKERNEL_TEST_DEPS,
7032)
7033
7034xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07007035 name = "x32_packx_test",
7036 srcs = [
7037 "test/x32-packx.cc",
7038 "test/pack-microkernel-tester.h",
7039 "src/xnnpack/AlignedAllocator.h",
7040 ] + MICROKERNEL_TEST_HDRS,
7041 deps = MICROKERNEL_TEST_DEPS,
7042)
7043
7044xnnpack_unit_test(
7045 name = "x32_pad_test",
7046 srcs = [
7047 "test/x32-pad.cc",
7048 "test/pad-microkernel-tester.h",
7049 ] + MICROKERNEL_TEST_HDRS,
7050 deps = MICROKERNEL_TEST_DEPS,
7051)
7052
7053xnnpack_unit_test(
7054 name = "x32_unpool_test",
7055 srcs = [
7056 "test/x32-unpool.cc",
7057 "test/unpool-microkernel-tester.h",
7058 ] + MICROKERNEL_TEST_HDRS,
7059 deps = MICROKERNEL_TEST_DEPS,
7060)
7061
7062xnnpack_unit_test(
7063 name = "x32_zip_test",
7064 srcs = [
7065 "test/x32-zip.cc",
7066 "test/zip-microkernel-tester.h",
7067 ] + MICROKERNEL_TEST_HDRS,
7068 deps = MICROKERNEL_TEST_DEPS,
7069)
7070
7071xnnpack_unit_test(
7072 name = "x8_lut_test",
7073 srcs = [
7074 "test/x8-lut.cc",
7075 "test/lut-microkernel-tester.h",
7076 ] + MICROKERNEL_TEST_HDRS,
7077 deps = MICROKERNEL_TEST_DEPS,
7078)
7079
7080xnnpack_unit_test(
7081 name = "x8_zip_test",
7082 srcs = [
7083 "test/x8-zip.cc",
7084 "test/zip-microkernel-tester.h",
7085 ] + MICROKERNEL_TEST_HDRS,
7086 deps = MICROKERNEL_TEST_DEPS,
7087)
7088
Marat Dukhan20c3b922020-03-10 03:45:06 -07007089########################## Size tests for the library #########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07007090
7091xnnpack_binary(
Marat Dukhan20c3b922020-03-10 03:45:06 -07007092 name = "operator_size_test",
7093 srcs = ["test/operator-size.c"],
Marat Dukhanf0cb70a2021-03-30 15:45:15 -07007094 deps = [":xnnpack_for_tfjs"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07007095)
7096
Marat Dukhan20c3b922020-03-10 03:45:06 -07007097xnnpack_binary(
7098 name = "subgraph_size_test",
7099 srcs = ["test/subgraph-size.c"],
7100 deps = [":XNNPACK"],
7101)
7102
7103########################### Unit tests for operators ##########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07007104
7105xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07007106 name = "abs_nc_test",
7107 srcs = [
7108 "test/abs-nc.cc",
7109 "test/abs-operator-tester.h",
7110 ],
7111 deps = OPERATOR_TEST_DEPS,
7112)
7113
7114xnnpack_unit_test(
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08007115 name = "add_nd_test",
Artsiom Ablavatski2202c812021-01-22 14:16:43 -08007116 timeout = "moderate",
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08007117 srcs = [
7118 "test/add-nd.cc",
7119 "test/binary-elementwise-operator-tester.h",
7120 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07007121 deps = OPERATOR_TEST_DEPS,
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08007122)
7123
7124xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08007125 name = "argmax_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007126 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08007127 "test/argmax-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007128 "test/argmax-pooling-operator-tester.h",
7129 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07007130 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07007131)
7132
7133xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08007134 name = "average_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007135 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08007136 "test/average-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007137 "test/average-pooling-operator-tester.h",
7138 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07007139 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07007140)
7141
7142xnnpack_unit_test(
Marat Dukhan64e52512020-06-09 13:41:16 -07007143 name = "bankers_rounding_nc_test",
7144 srcs = [
7145 "test/bankers-rounding-nc.cc",
7146 "test/bankers-rounding-operator-tester.h",
7147 ],
7148 deps = OPERATOR_TEST_DEPS,
7149)
7150
7151xnnpack_unit_test(
7152 name = "ceiling_nc_test",
7153 srcs = [
7154 "test/ceiling-nc.cc",
7155 "test/ceiling-operator-tester.h",
7156 ],
7157 deps = OPERATOR_TEST_DEPS,
7158)
7159
7160xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08007161 name = "channel_shuffle_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007162 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08007163 "test/channel-shuffle-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007164 "test/channel-shuffle-operator-tester.h",
7165 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07007166 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07007167)
7168
7169xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08007170 name = "clamp_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007171 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08007172 "test/clamp-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007173 "test/clamp-operator-tester.h",
7174 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07007175 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07007176)
7177
7178xnnpack_unit_test(
Marat Dukhan065b11e2020-05-22 09:49:41 -07007179 name = "constant_pad_nd_test",
7180 srcs = [
7181 "test/constant-pad-nd.cc",
7182 "test/constant-pad-operator-tester.h",
7183 ],
7184 deps = OPERATOR_TEST_DEPS,
7185)
7186
7187xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08007188 name = "convolution_nhwc_test",
Artsiom Ablavatski2202c812021-01-22 14:16:43 -08007189 timeout = "moderate",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007190 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08007191 "test/convolution-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007192 "test/convolution-operator-tester.h",
7193 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07007194 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07007195)
7196
7197xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08007198 name = "convolution_nchw_test",
Artsiom Ablavatski2202c812021-01-22 14:16:43 -08007199 timeout = "moderate",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007200 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08007201 "test/convolution-nchw.cc",
7202 "test/convolution-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007203 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07007204 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07007205)
7206
7207xnnpack_unit_test(
Marat Dukhan4e21b272020-06-04 18:45:01 -07007208 name = "copy_nc_test",
7209 srcs = [
7210 "test/copy-nc.cc",
7211 "test/copy-operator-tester.h",
7212 ],
7213 deps = OPERATOR_TEST_DEPS,
7214)
7215
7216xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08007217 name = "deconvolution_nhwc_test",
Artsiom Ablavatskic1aa2972020-12-08 11:23:34 -08007218 timeout = "moderate",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007219 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08007220 "test/deconvolution-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007221 "test/deconvolution-operator-tester.h",
7222 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07007223 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07007224)
7225
7226xnnpack_unit_test(
Marat Dukhan188d1042020-11-24 23:39:40 -08007227 name = "depth_to_space_nchw2nhwc_test",
Artsiom Ablavatski0f1dc182020-11-05 19:21:50 -08007228 srcs = [
Marat Dukhan188d1042020-11-24 23:39:40 -08007229 "test/depth-to-space-nchw2nhwc.cc",
Artsiom Ablavatski0f1dc182020-11-05 19:21:50 -08007230 "test/depth-to-space-operator-tester.h",
7231 ] + OPERATOR_TEST_PARAMS_HDRS,
7232 deps = OPERATOR_TEST_DEPS,
7233)
7234
7235xnnpack_unit_test(
Marat Dukhan0e521172020-11-25 13:10:04 -08007236 name = "depth_to_space_nhwc_test",
7237 srcs = [
7238 "test/depth-to-space-nhwc.cc",
7239 "test/depth-to-space-operator-tester.h",
7240 ] + OPERATOR_TEST_PARAMS_HDRS,
7241 deps = OPERATOR_TEST_DEPS,
7242)
7243
7244xnnpack_unit_test(
Marat Dukhan69180502019-12-06 15:00:31 -08007245 name = "divide_nd_test",
7246 srcs = [
7247 "test/binary-elementwise-operator-tester.h",
7248 "test/divide-nd.cc",
7249 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07007250 deps = OPERATOR_TEST_DEPS,
Marat Dukhan69180502019-12-06 15:00:31 -08007251)
7252
7253xnnpack_unit_test(
Marat Dukhanb6bd4bc2020-12-01 17:01:40 -08007254 name = "elu_nc_test",
7255 srcs = [
7256 "test/elu-nc.cc",
7257 "test/elu-operator-tester.h",
7258 ],
7259 deps = OPERATOR_TEST_DEPS,
7260)
7261
7262xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08007263 name = "fully_connected_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007264 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08007265 "test/fully-connected-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007266 "test/fully-connected-operator-tester.h",
7267 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07007268 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07007269)
7270
7271xnnpack_unit_test(
Marat Dukhan64e52512020-06-09 13:41:16 -07007272 name = "floor_nc_test",
7273 srcs = [
7274 "test/floor-nc.cc",
7275 "test/floor-operator-tester.h",
7276 ],
7277 deps = OPERATOR_TEST_DEPS,
7278)
7279
7280xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08007281 name = "global_average_pooling_nwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007282 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08007283 "test/global-average-pooling-nwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007284 "test/global-average-pooling-operator-tester.h",
Marat Dukhanef61d022020-06-19 13:54:49 -07007285 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07007286 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07007287)
7288
7289xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08007290 name = "global_average_pooling_ncw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007291 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08007292 "test/global-average-pooling-ncw.cc",
7293 "test/global-average-pooling-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007294 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07007295 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07007296)
7297
7298xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08007299 name = "hardswish_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007300 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08007301 "test/hardswish-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007302 "test/hardswish-operator-tester.h",
7303 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07007304 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07007305)
7306
7307xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08007308 name = "leaky_relu_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007309 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08007310 "test/leaky-relu-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007311 "test/leaky-relu-operator-tester.h",
7312 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07007313 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07007314)
7315
7316xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08007317 name = "max_pooling_nhwc_test",
Artsiom Ablavatski2202c812021-01-22 14:16:43 -08007318 timeout = "moderate",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007319 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08007320 "test/max-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007321 "test/max-pooling-operator-tester.h",
7322 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07007323 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07007324)
7325
7326xnnpack_unit_test(
Marat Dukhan79e7f842019-12-05 14:35:50 -08007327 name = "maximum_nd_test",
7328 srcs = [
7329 "test/binary-elementwise-operator-tester.h",
7330 "test/maximum-nd.cc",
7331 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07007332 deps = OPERATOR_TEST_DEPS,
Marat Dukhan79e7f842019-12-05 14:35:50 -08007333)
7334
7335xnnpack_unit_test(
7336 name = "minimum_nd_test",
7337 srcs = [
7338 "test/binary-elementwise-operator-tester.h",
7339 "test/minimum-nd.cc",
7340 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07007341 deps = OPERATOR_TEST_DEPS,
Marat Dukhan79e7f842019-12-05 14:35:50 -08007342)
7343
7344xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08007345 name = "multiply_nd_test",
Marat Dukhanca2733c2019-11-15 23:21:17 -08007346 srcs = [
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08007347 "test/binary-elementwise-operator-tester.h",
Marat Dukhanefc47b82019-11-18 09:25:38 -08007348 "test/multiply-nd.cc",
Marat Dukhanca2733c2019-11-15 23:21:17 -08007349 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07007350 deps = OPERATOR_TEST_DEPS,
Marat Dukhanca2733c2019-11-15 23:21:17 -08007351)
7352
7353xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07007354 name = "negate_nc_test",
7355 srcs = [
7356 "test/negate-nc.cc",
7357 "test/negate-operator-tester.h",
7358 ],
7359 deps = OPERATOR_TEST_DEPS,
7360)
7361
7362xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08007363 name = "prelu_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007364 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08007365 "test/prelu-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007366 "test/prelu-operator-tester.h",
7367 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07007368 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07007369)
7370
7371xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08007372 name = "resize_bilinear_nhwc_test",
Marat Dukhan69722492019-11-11 19:55:50 -08007373 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08007374 "test/resize-bilinear-nhwc.cc",
Marat Dukhan69722492019-11-11 19:55:50 -08007375 "test/resize-bilinear-operator-tester.h",
7376 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07007377 deps = OPERATOR_TEST_DEPS,
Marat Dukhan69722492019-11-11 19:55:50 -08007378)
7379
7380xnnpack_unit_test(
Artsiom Ablavatski97918102020-10-27 15:52:59 -07007381 name = "resize_bilinear_nchw_test",
7382 srcs = [
7383 "test/resize-bilinear-nchw.cc",
7384 "test/resize-bilinear-operator-tester.h",
7385 ] + OPERATOR_TEST_PARAMS_HDRS,
7386 deps = OPERATOR_TEST_DEPS,
7387)
7388
7389xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08007390 name = "sigmoid_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007391 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08007392 "test/sigmoid-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007393 "test/sigmoid-operator-tester.h",
7394 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07007395 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07007396)
7397
7398xnnpack_unit_test(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08007399 name = "softmax_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007400 srcs = [
Marat Dukhanfd8e6892020-01-27 15:25:25 -08007401 "test/softmax-nc.cc",
7402 "test/softmax-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007403 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07007404 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07007405)
7406
7407xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07007408 name = "square_nc_test",
7409 srcs = [
7410 "test/square-nc.cc",
7411 "test/square-operator-tester.h",
7412 ],
7413 deps = OPERATOR_TEST_DEPS,
7414)
7415
7416xnnpack_unit_test(
Marat Dukhan6804bbd2020-06-30 19:26:11 -07007417 name = "square_root_nc_test",
7418 srcs = [
7419 "test/square-root-nc.cc",
7420 "test/square-root-operator-tester.h",
7421 ],
7422 deps = OPERATOR_TEST_DEPS,
7423)
7424
7425xnnpack_unit_test(
Marat Dukhanf7399262020-06-05 10:58:44 -07007426 name = "squared_difference_nd_test",
7427 srcs = [
7428 "test/binary-elementwise-operator-tester.h",
7429 "test/squared-difference-nd.cc",
7430 ],
7431 deps = OPERATOR_TEST_DEPS,
7432)
7433
7434xnnpack_unit_test(
Marat Dukhan05f3f6d2019-12-03 15:13:53 -08007435 name = "subtract_nd_test",
7436 srcs = [
7437 "test/binary-elementwise-operator-tester.h",
7438 "test/subtract-nd.cc",
7439 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07007440 deps = OPERATOR_TEST_DEPS,
Marat Dukhan05f3f6d2019-12-03 15:13:53 -08007441)
7442
7443xnnpack_unit_test(
Marat Dukhan64e52512020-06-09 13:41:16 -07007444 name = "truncation_nc_test",
7445 srcs = [
7446 "test/truncation-nc.cc",
7447 "test/truncation-operator-tester.h",
7448 ],
7449 deps = OPERATOR_TEST_DEPS,
7450)
7451
7452xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08007453 name = "unpooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007454 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08007455 "test/unpooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007456 "test/unpooling-operator-tester.h",
7457 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07007458 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07007459)
7460
Chao Mei6ddfc602020-05-13 22:29:36 -07007461############################### Misc unit tests ###############################
7462
7463xnnpack_unit_test(
7464 name = "memory_planner_test",
7465 srcs = [
7466 "test/memory-planner-test.cc",
7467 ],
7468 deps = [
7469 ":XNNPACK",
7470 ":memory_planner",
7471 ],
7472)
7473
XNNPACK Teamab8c4c82020-10-09 08:05:51 -07007474xnnpack_unit_test(
7475 name = "subgraph_nchw_test",
7476 srcs = [
7477 "src/xnnpack/subgraph.h",
7478 "test/subgraph-nchw.cc",
7479 "test/subgraph-tester.h",
7480 ],
7481 deps = [
7482 ":XNNPACK",
7483 ],
7484)
7485
Marat Dukhan08c4a432019-10-03 09:29:21 -07007486############################# Build configurations #############################
7487
Marat Dukhanb8642352019-10-30 15:43:02 -07007488# Enables usage of assembly kernels.
Marat Dukhan08c4a432019-10-03 09:29:21 -07007489config_setting(
Marat Dukhanb8642352019-10-30 15:43:02 -07007490 name = "xnn_enable_assembly_explicit_true",
7491 define_values = {"xnn_enable_assembly": "true"},
7492)
7493
7494# Disables usage of assembly kernels.
7495config_setting(
7496 name = "xnn_enable_assembly_explicit_false",
7497 define_values = {"xnn_enable_assembly": "false"},
7498)
7499
Marat Dukhan9de90e02020-06-18 16:04:12 -07007500# Enables usage of sparse inference.
7501config_setting(
7502 name = "xnn_enable_sparse_explicit_true",
7503 define_values = {"xnn_enable_sparse": "true"},
7504)
7505
7506# Disables usage of sparse inference.
7507config_setting(
7508 name = "xnn_enable_sparse_explicit_false",
7509 define_values = {"xnn_enable_sparse": "false"},
7510)
7511
Marat Dukhan05702cf2020-03-26 15:41:33 -07007512# Disables usage of HMP-aware optimizations.
7513config_setting(
7514 name = "xnn_enable_hmp_explicit_false",
7515 define_values = {"xnn_enable_hmp": "false"},
7516)
7517
Chao Mei6ddfc602020-05-13 22:29:36 -07007518# Enable usage of optimized memory allocation
7519config_setting(
7520 name = "xnn_enable_memopt_explicit_true",
Marat Dukhan03f46212021-03-30 21:29:49 -07007521 define_values = {"xnn_enable_memopt": "true"},
Chao Mei6ddfc602020-05-13 22:29:36 -07007522)
7523
7524# Disable usage of optimized memory allocation
7525config_setting(
7526 name = "xnn_enable_memopt_explicit_false",
Marat Dukhan03f46212021-03-30 21:29:49 -07007527 define_values = {"xnn_enable_memopt": "false"},
Chao Mei6ddfc602020-05-13 22:29:36 -07007528)
7529
Marat Dukhanb939cdb2021-03-30 18:51:51 -07007530# Enable QS8 inference in TFLite-specific version
7531config_setting(
7532 name = "xnn_enable_qs8_explicit_true",
7533 define_values = {"xnn_enable_qs8": "true"},
7534)
7535
7536# Disable QS8 inference in TFLite-specific version
7537config_setting(
7538 name = "xnn_enable_qs8_explicit_false",
7539 define_values = {"xnn_enable_qs8": "false"},
7540)
7541
Marat Dukhanb8642352019-10-30 15:43:02 -07007542# Builds with -c dbg
7543config_setting(
7544 name = "debug_build",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007545 values = {
Marat Dukhanb8642352019-10-30 15:43:02 -07007546 "compilation_mode": "dbg",
7547 },
7548)
7549
7550# Builds with -c opt
7551config_setting(
7552 name = "optimized_build",
7553 values = {
7554 "compilation_mode": "opt",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007555 },
7556)
7557
7558config_setting(
Marat Dukhanb8642352019-10-30 15:43:02 -07007559 name = "linux_k8",
7560 values = {"cpu": "k8"},
7561)
7562
7563config_setting(
Marat Dukhan582094e2020-04-30 17:21:25 -07007564 name = "linux_arm",
7565 values = {"cpu": "arm"},
Marat Dukhan4e45e662019-10-03 15:40:24 -07007566)
7567
7568config_setting(
Marat Dukhanf0bd4de2020-06-15 15:53:19 -07007569 name = "linux_armeabi",
7570 values = {"cpu": "armeabi"},
7571)
7572
7573config_setting(
Terry Heo68eef3f2020-04-13 22:53:52 -07007574 name = "linux_armhf",
7575 values = {"cpu": "armhf"},
7576)
7577
7578config_setting(
Marat Dukhana720e932020-06-10 13:01:11 -07007579 name = "linux_armv7a",
7580 values = {"cpu": "armv7a"},
7581)
7582
7583config_setting(
Marat Dukhan582094e2020-04-30 17:21:25 -07007584 name = "linux_aarch64",
7585 values = {"cpu": "aarch64"},
7586)
7587
7588config_setting(
Marat Dukhan08c4a432019-10-03 09:29:21 -07007589 name = "android",
7590 values = {"crosstool_top": "//external:android/crosstool"},
7591)
7592
7593config_setting(
7594 name = "android_armv7",
7595 values = {
7596 "crosstool_top": "//external:android/crosstool",
7597 "cpu": "armeabi-v7a",
7598 },
7599)
7600
7601config_setting(
7602 name = "android_arm64",
7603 values = {
7604 "crosstool_top": "//external:android/crosstool",
7605 "cpu": "arm64-v8a",
7606 },
7607)
7608
7609config_setting(
7610 name = "android_x86",
7611 values = {
7612 "crosstool_top": "//external:android/crosstool",
7613 "cpu": "x86",
7614 },
7615)
7616
7617config_setting(
7618 name = "android_x86_64",
7619 values = {
7620 "crosstool_top": "//external:android/crosstool",
7621 "cpu": "x86_64",
7622 },
7623)
7624
7625config_setting(
Marat Dukhan10a38082020-04-17 03:58:35 -07007626 name = "windows_x86_64",
7627 values = {"cpu": "x64_windows"},
Marat Dukhan9fe932e2020-04-11 17:14:15 -07007628)
7629
7630config_setting(
Marat Dukhan10a38082020-04-17 03:58:35 -07007631 name = "windows_x86_64_clang",
7632 values = {
7633 "compiler": "clang-cl",
7634 "cpu": "x64_windows",
7635 },
7636)
7637
7638config_setting(
7639 name = "windows_x86_64_mingw",
7640 values = {
7641 "compiler": "mingw-gcc",
7642 "cpu": "x64_windows",
7643 },
7644)
7645
7646config_setting(
7647 name = "windows_x86_64_msys",
7648 values = {
7649 "compiler": "msys-gcc",
7650 "cpu": "x64_windows",
7651 },
Marat Dukhan9fe932e2020-04-11 17:14:15 -07007652)
7653
7654config_setting(
Marat Dukhan885ca242019-10-07 09:17:32 -07007655 name = "macos_x86_64",
7656 values = {
7657 "apple_platform_type": "macos",
7658 "cpu": "darwin",
7659 },
7660)
7661
7662config_setting(
Simon Maurerae33ab82021-03-03 23:38:22 +01007663 name = "macos_arm64",
7664 values = {
7665 "apple_platform_type": "macos",
7666 "cpu": "darwin_arm64",
7667 },
7668)
7669
7670config_setting(
Marat Dukhan08c4a432019-10-03 09:29:21 -07007671 name = "emscripten",
XNNPACK Team3bfbdaf2021-03-29 15:26:23 -07007672 values = {"crosstool_top": "@emsdk//emscripten_toolchain:everything"},
Marat Dukhan08c4a432019-10-03 09:29:21 -07007673)
7674
7675config_setting(
7676 name = "emscripten_wasm",
7677 values = {
XNNPACK Team3bfbdaf2021-03-29 15:26:23 -07007678 "crosstool_top": "@emsdk//emscripten_toolchain:everything",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007679 "cpu": "wasm",
7680 },
7681)
7682
7683config_setting(
7684 name = "emscripten_wasmsimd",
7685 values = {
XNNPACK Team3bfbdaf2021-03-29 15:26:23 -07007686 "crosstool_top": "@emsdk//emscripten_toolchain:everything",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007687 "cpu": "wasm",
Marat Dukhan81c62602020-05-29 13:22:49 -07007688 "copt": "-msimd128",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007689 },
7690)
7691
7692config_setting(
Marat Dukhan1498d1d2020-02-11 20:00:05 -08007693 name = "ios_armv7",
7694 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08007695 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08007696 "cpu": "ios_armv7",
7697 },
7698)
7699
7700config_setting(
7701 name = "ios_arm64",
7702 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08007703 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08007704 "cpu": "ios_arm64",
7705 },
7706)
7707
7708config_setting(
7709 name = "ios_arm64e",
7710 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08007711 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08007712 "cpu": "ios_arm64e",
7713 },
7714)
7715
7716config_setting(
7717 name = "ios_x86",
7718 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08007719 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08007720 "cpu": "ios_i386",
7721 },
7722)
7723
7724config_setting(
7725 name = "ios_x86_64",
7726 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08007727 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08007728 "cpu": "ios_x86_64",
7729 },
7730)
7731
7732config_setting(
7733 name = "watchos_armv7k",
7734 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08007735 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08007736 "cpu": "watchos_armv7k",
7737 },
7738)
7739
7740config_setting(
7741 name = "watchos_arm64_32",
7742 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08007743 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08007744 "cpu": "watchos_arm64_32",
7745 },
7746)
7747
7748config_setting(
7749 name = "watchos_x86",
7750 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08007751 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08007752 "cpu": "watchos_i386",
7753 },
7754)
7755
7756config_setting(
7757 name = "watchos_x86_64",
7758 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08007759 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08007760 "cpu": "watchos_x86_64",
7761 },
7762)
7763
7764config_setting(
7765 name = "tvos_arm64",
7766 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08007767 "apple_platform_type": "tvos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08007768 "cpu": "tvos_arm64",
7769 },
7770)
7771
7772config_setting(
7773 name = "tvos_x86_64",
7774 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08007775 "apple_platform_type": "tvos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08007776 "cpu": "tvos_x86_64",
7777 },
7778)