blob: 04b766685a52b075270aff53af22bf32e450f78c [file] [log] [blame]
Frank Barchard1a953052020-11-16 18:44:58 -08001# Copyright 2020 Google LLC
Marat Dukhan08c4a432019-10-03 09:29:21 -07002#
3# This source code is licensed under the BSD-style license found in the
4# LICENSE file in the root directory of this source tree.
5#
6# Description:
7# XNNPACK - optimized floating-point neural network operators library
8
Marat Dukhan10a38082020-04-17 03:58:35 -07009load(":build_defs.bzl", "xnnpack_aggregate_library", "xnnpack_benchmark", "xnnpack_binary", "xnnpack_cc_library", "xnnpack_gcc_std_copts", "xnnpack_min_size_copts", "xnnpack_msvc_std_copts", "xnnpack_optional_armcl_copts", "xnnpack_optional_armcl_deps", "xnnpack_optional_dnnl_copts", "xnnpack_optional_dnnl_deps", "xnnpack_optional_gemmlowp_copts", "xnnpack_optional_gemmlowp_deps", "xnnpack_optional_ruy_copts", "xnnpack_optional_ruy_deps", "xnnpack_optional_tflite_copts", "xnnpack_optional_tflite_deps", "xnnpack_std_cxxopts", "xnnpack_unit_test", "xnnpack_visibility")
Marat Dukhan69c3f2c2019-11-06 12:30:01 -080010
Marat Dukhan08c4a432019-10-03 09:29:21 -070011licenses(["notice"])
12
13exports_files(["LICENSE"])
14
Marat Dukhan1b354632020-03-23 12:50:22 -070015OPERATOR_BENCHMARK_DEPS = [
16 ":XNNPACK",
17 ":bench_utils",
18 "@cpuinfo",
Frank Barchard0c849732020-06-12 13:31:32 -070019 "@FP16",
Marat Dukhan1b354632020-03-23 12:50:22 -070020 "@pthreadpool",
21]
22
Marat Dukhan08c4a432019-10-03 09:29:21 -070023MICROKERNEL_BENCHMARK_DEPS = [
24 ":ukernels",
25 ":bench_utils",
Frank Barchard7e955972019-10-11 10:34:25 -070026 ":enable_assembly",
Marat Dukhan08c4a432019-10-03 09:29:21 -070027 "@cpuinfo",
28 "@FP16",
29 "@pthreadpool",
30]
31
Marat Dukhan6adff4e2019-10-14 18:32:07 -070032ACCURACY_EVAL_DEPS = [
33 ":XNNPACK",
34 ":ukernels",
35 "@FP16",
36 "@pthreadpool",
37]
38
Marat Dukhan08c4a432019-10-03 09:29:21 -070039MICROKERNEL_TEST_DEPS = [
Marat Dukhan33fcf782020-05-24 14:27:15 -070040 ":ukernels_test_mode",
Frank Barchard7e955972019-10-11 10:34:25 -070041 ":enable_assembly",
Marat Dukhan08c4a432019-10-03 09:29:21 -070042 "@cpuinfo",
43 "@FP16",
44 "@pthreadpool",
45]
46
Marat Dukhan1b354632020-03-23 12:50:22 -070047OPERATOR_TEST_DEPS = [
Marat Dukhan33fcf782020-05-24 14:27:15 -070048 ":XNNPACK_test_mode",
Marat Dukhan1b354632020-03-23 12:50:22 -070049 "@pthreadpool",
50 "@FP16",
51]
52
Marat Dukhan08c4a432019-10-03 09:29:21 -070053OPERATOR_SRCS = [
Marat Dukhane8265432020-04-28 18:42:59 -070054 "src/operators/argmax-pooling-nhwc.c",
55 "src/operators/average-pooling-nhwc.c",
56 "src/operators/binary-elementwise-nd.c",
Marat Dukhane8265432020-04-28 18:42:59 -070057 "src/operators/channel-shuffle-nc.c",
Marat Dukhan065b11e2020-05-22 09:49:41 -070058 "src/operators/constant-pad-nd.c",
Marat Dukhane8265432020-04-28 18:42:59 -070059 "src/operators/convolution-nchw.c",
60 "src/operators/convolution-nhwc.c",
61 "src/operators/deconvolution-nhwc.c",
Marat Dukhan13b68f22020-11-12 11:55:19 -080062 "src/operators/depth-to-space-nchw2nhwc.c",
Marat Dukhan0e521172020-11-25 13:10:04 -080063 "src/operators/depth-to-space-nhwc.c",
Marat Dukhane8265432020-04-28 18:42:59 -070064 "src/operators/fully-connected-nc.c",
65 "src/operators/global-average-pooling-ncw.c",
66 "src/operators/global-average-pooling-nwc.c",
Marat Dukhane8265432020-04-28 18:42:59 -070067 "src/operators/leaky-relu-nc.c",
68 "src/operators/max-pooling-nhwc.c",
69 "src/operators/prelu-nc.c",
Artsiom Ablavatski97918102020-10-27 15:52:59 -070070 "src/operators/resize-bilinear-nchw.c",
Marat Dukhane8265432020-04-28 18:42:59 -070071 "src/operators/resize-bilinear-nhwc.c",
72 "src/operators/sigmoid-nc.c",
73 "src/operators/softmax-nc.c",
Marat Dukhanc3065f52020-06-04 13:33:32 -070074 "src/operators/unary-elementwise-nc.c",
Marat Dukhane8265432020-04-28 18:42:59 -070075 "src/operators/unpooling-nhwc.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -070076]
77
Marat Dukhan95e8b7a2020-06-03 12:46:26 -070078SUBGRAPH_SRCS = [
Marat Dukhan5fab4092020-06-10 01:28:28 -070079 "src/subgraph/abs.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -070080 "src/subgraph/add2.c",
81 "src/subgraph/argmax-pooling-2d.c",
82 "src/subgraph/average-pooling-2d.c",
Marat Dukhan5fab4092020-06-10 01:28:28 -070083 "src/subgraph/bankers-rounding.c",
84 "src/subgraph/ceiling.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -070085 "src/subgraph/clamp.c",
86 "src/subgraph/convolution-2d.c",
87 "src/subgraph/deconvolution-2d.c",
Artsiom Ablavatskibbe85062020-11-05 14:07:37 -080088 "src/subgraph/depth-to-space.c",
Frank Barchard9cef5ea2020-11-18 14:52:08 -080089 "src/subgraph/depthwise-convolution-2d.c",
Marat Dukhan9d3a4592020-06-05 16:52:42 -070090 "src/subgraph/divide.c",
Marat Dukhana1600202020-12-01 22:17:16 -080091 "src/subgraph/elu.c",
Marat Dukhan5fab4092020-06-10 01:28:28 -070092 "src/subgraph/floor.c",
Frank Barchard04336c12020-10-22 16:48:55 -070093 "src/subgraph/fully-connected.c",
Marat Dukhana059b7d2020-06-11 11:41:27 -070094 "src/subgraph/global-average-pooling-2d.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -070095 "src/subgraph/hardswish.c",
Marat Dukhan5bbebac2020-06-10 19:42:15 -070096 "src/subgraph/leaky-relu.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -070097 "src/subgraph/max-pooling-2d.c",
Marat Dukhan9d3a4592020-06-05 16:52:42 -070098 "src/subgraph/maximum2.c",
99 "src/subgraph/minimum2.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -0700100 "src/subgraph/multiply2.c",
Marat Dukhan5fab4092020-06-10 01:28:28 -0700101 "src/subgraph/negate.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -0700102 "src/subgraph/prelu.c",
103 "src/subgraph/sigmoid.c",
104 "src/subgraph/softmax.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700105 "src/subgraph/square-root.c",
106 "src/subgraph/square.c",
107 "src/subgraph/squared-difference.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -0700108 "src/subgraph/static-constant-pad.c",
Marat Dukhand27202d2020-07-09 23:43:40 -0700109 "src/subgraph/static-reshape.c",
Marat Dukhanaff24e22020-07-23 01:43:58 -0700110 "src/subgraph/static-resize-bilinear-2d.c",
Marat Dukhan9d3a4592020-06-05 16:52:42 -0700111 "src/subgraph/subtract.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -0700112 "src/subgraph/unpooling-2d.c",
113]
114
Marat Dukhan3a77ea72019-12-23 12:10:24 -0800115TABLE_SRCS = [
116 "src/tables/exp2-k-over-64.c",
117 "src/tables/exp2-k-over-2048.c",
Marat Dukhande390d42020-11-29 19:32:18 -0800118 "src/tables/exp2minus-k-over-4.c",
119 "src/tables/exp2minus-k-over-8.c",
Marat Dukhanc60742b2020-11-23 12:33:27 -0800120 "src/tables/exp2minus-k-over-16.c",
Marat Dukhan1f256fc2020-09-24 21:27:14 -0700121 "src/tables/exp2minus-k-over-64.c",
122 "src/tables/exp2minus-k-over-2048.c",
Marat Dukhan3a77ea72019-12-23 12:10:24 -0800123]
124
Marat Dukhan08c4a432019-10-03 09:29:21 -0700125SCALAR_UKERNELS = [
Marat Dukhan329da642019-11-19 21:44:39 -0800126 "src/f32-argmaxpool/4x-scalar-c1.c",
Marat Dukhan1e782c42019-11-21 17:02:40 -0800127 "src/f32-argmaxpool/9p8x-scalar-c1.c",
Marat Dukhan329da642019-11-19 21:44:39 -0800128 "src/f32-argmaxpool/9x-scalar-c1.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700129 "src/f32-avgpool/9p8x-minmax-scalar-c1.c",
130 "src/f32-avgpool/9x-minmax-scalar-c1.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -0700131 "src/f32-clamp/gen/scalar-x1.c",
132 "src/f32-clamp/gen/scalar-x2.c",
133 "src/f32-clamp/gen/scalar-x4.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700134 "src/f32-conv-hwc/3x3s2p0p1c3x4-scalar-1x1.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700135 "src/f32-conv-hwc/3x3s2p1c3x4-scalar-1x1.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -0700136 "src/f32-conv-hwc2chw/3x3s2p1c3x4-scalar-1x1.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700137 "src/f32-dwconv/gen/up1x4-minmax-scalar-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700138 "src/f32-dwconv/gen/up1x4-minmax-scalar.c",
139 "src/f32-dwconv/gen/up1x4-scalar-acc2.c",
140 "src/f32-dwconv/gen/up1x4-scalar.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700141 "src/f32-dwconv/gen/up1x9-minmax-scalar-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700142 "src/f32-dwconv/gen/up1x9-minmax-scalar.c",
143 "src/f32-dwconv/gen/up1x9-scalar-acc2.c",
144 "src/f32-dwconv/gen/up1x9-scalar.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700145 "src/f32-dwconv/gen/up1x25-minmax-scalar-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700146 "src/f32-dwconv/gen/up1x25-minmax-scalar.c",
147 "src/f32-dwconv/gen/up1x25-scalar-acc2.c",
148 "src/f32-dwconv/gen/up1x25-scalar.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700149 "src/f32-dwconv/gen/up2x4-minmax-scalar-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700150 "src/f32-dwconv/gen/up2x4-minmax-scalar.c",
151 "src/f32-dwconv/gen/up2x4-scalar-acc2.c",
152 "src/f32-dwconv/gen/up2x4-scalar.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700153 "src/f32-dwconv/gen/up2x9-minmax-scalar-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700154 "src/f32-dwconv/gen/up2x9-minmax-scalar.c",
155 "src/f32-dwconv/gen/up2x9-scalar-acc2.c",
156 "src/f32-dwconv/gen/up2x9-scalar.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700157 "src/f32-dwconv/gen/up2x25-minmax-scalar-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700158 "src/f32-dwconv/gen/up2x25-minmax-scalar.c",
159 "src/f32-dwconv/gen/up2x25-scalar-acc2.c",
160 "src/f32-dwconv/gen/up2x25-scalar.c",
Frank Barchard35db7d02020-10-26 13:37:34 -0700161 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-scalar-1x1-acc2.c",
162 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-scalar-1x1-acc3.c",
163 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-scalar-1x1-acc4.c",
Marat Dukhan91249d22020-10-24 12:02:51 -0700164 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-scalar-1x1.c",
Frank Barchard35db7d02020-10-26 13:37:34 -0700165 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-scalar-2x1-acc2.c",
Marat Dukhan91249d22020-10-24 12:02:51 -0700166 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-scalar-2x1.c",
167 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-scalar-3x1.c",
168 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-scalar-4x1.c",
169 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-scalar-5x1.c",
170 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-scalar-6x1.c",
Marat Dukhancf5b3c32020-10-25 19:21:10 -0700171 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-scalar-1x1-acc2.c",
172 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-scalar-1x1-acc3.c",
173 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-scalar-1x1-acc4.c",
Frank Barchard35db7d02020-10-26 13:37:34 -0700174 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-scalar-1x1.c",
Marat Dukhancf5b3c32020-10-25 19:21:10 -0700175 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-scalar-2x1-acc2.c",
Frank Barchard35db7d02020-10-26 13:37:34 -0700176 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-scalar-2x1.c",
177 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-scalar-3x1.c",
178 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-scalar-4x1.c",
Marat Dukhanc4efb002020-10-25 23:14:47 -0700179 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-scalar-1x1-acc2.c",
180 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-scalar-1x1-acc3.c",
181 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-scalar-1x1-acc4.c",
182 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-scalar-1x1-acc5.c",
Frank Barchard35db7d02020-10-26 13:37:34 -0700183 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-scalar-1x1.c",
Marat Dukhanc4efb002020-10-25 23:14:47 -0700184 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-scalar-2x1-acc2.c",
185 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-scalar-2x1-acc3.c",
Frank Barchard35db7d02020-10-26 13:37:34 -0700186 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-scalar-2x1.c",
Marat Dukhanc4efb002020-10-25 23:14:47 -0700187 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-scalar-3x1-acc2.c",
Frank Barchard35db7d02020-10-26 13:37:34 -0700188 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-scalar-3x1.c",
Marat Dukhan29c0c332020-10-28 22:11:00 -0700189 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-scalar-1x1-acc2.c",
190 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-scalar-1x1-acc3.c",
191 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-scalar-1x1-acc4.c",
192 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-scalar-1x1-acc5.c",
193 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-scalar-1x1.c",
194 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-scalar-2x1-acc2.c",
195 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-scalar-2x1-acc3.c",
196 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-scalar-2x1.c",
197 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-scalar-3x1-acc2.c",
198 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-scalar-3x1.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -0700199 "src/f32-gavgpool-cw/scalar-x1.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700200 "src/f32-gavgpool/7p7x-minmax-scalar-c1.c",
201 "src/f32-gavgpool/7x-minmax-scalar-c1.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700202 "src/f32-gemm/gen-inc/1x4inc-minmax-scalar.c",
203 "src/f32-gemm/gen-inc/2x4inc-minmax-scalar.c",
204 "src/f32-gemm/gen-inc/4x4inc-minmax-scalar.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700205 "src/f32-gemm/gen/1x4-minmax-scalar.c",
206 "src/f32-gemm/gen/1x4-relu-scalar.c",
207 "src/f32-gemm/gen/1x4-scalar.c",
208 "src/f32-gemm/gen/2x4-minmax-scalar.c",
209 "src/f32-gemm/gen/2x4-relu-scalar.c",
210 "src/f32-gemm/gen/2x4-scalar.c",
211 "src/f32-gemm/gen/4x2-minmax-scalar.c",
212 "src/f32-gemm/gen/4x2-relu-scalar.c",
213 "src/f32-gemm/gen/4x2-scalar.c",
214 "src/f32-gemm/gen/4x4-minmax-scalar.c",
215 "src/f32-gemm/gen/4x4-relu-scalar.c",
216 "src/f32-gemm/gen/4x4-scalar.c",
Marat Dukhanefaac272020-07-16 16:58:07 -0700217 "src/f32-hswish/gen/hswish-scalar-x1.c",
218 "src/f32-hswish/gen/hswish-scalar-x2.c",
219 "src/f32-hswish/gen/hswish-scalar-x4.c",
XNNPACK Team21432672020-10-19 19:58:48 -0700220 "src/f32-ibilinear-chw/gen/scalar-p1.c",
221 "src/f32-ibilinear-chw/gen/scalar-p2.c",
222 "src/f32-ibilinear-chw/gen/scalar-p4.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700223 "src/f32-ibilinear/gen/scalar-c1.c",
224 "src/f32-ibilinear/gen/scalar-c2.c",
225 "src/f32-ibilinear/gen/scalar-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700226 "src/f32-igemm/gen/1x4-minmax-scalar.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700227 "src/f32-igemm/gen/1x4-relu-scalar.c",
228 "src/f32-igemm/gen/1x4-scalar.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700229 "src/f32-igemm/gen/2x4-minmax-scalar.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700230 "src/f32-igemm/gen/2x4-relu-scalar.c",
231 "src/f32-igemm/gen/2x4-scalar.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700232 "src/f32-igemm/gen/4x2-minmax-scalar.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700233 "src/f32-igemm/gen/4x2-relu-scalar.c",
234 "src/f32-igemm/gen/4x2-scalar.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700235 "src/f32-igemm/gen/4x4-minmax-scalar.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700236 "src/f32-igemm/gen/4x4-relu-scalar.c",
237 "src/f32-igemm/gen/4x4-scalar.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700238 "src/f32-maxpool/9p8x-minmax-scalar-c1.c",
239 "src/f32-pavgpool/9p8x-minmax-scalar-c1.c",
240 "src/f32-pavgpool/9x-minmax-scalar-c1.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700241 "src/f32-ppmm/gen/2x4-minmax-scalar.c",
242 "src/f32-ppmm/gen/3x3-minmax-scalar.c",
243 "src/f32-ppmm/gen/4x2-minmax-scalar.c",
244 "src/f32-ppmm/gen/4x4-minmax-scalar.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800245 "src/f32-prelu/gen/scalar-2x1.c",
246 "src/f32-prelu/gen/scalar-2x4.c",
Marat Dukhanf46f6752020-01-21 11:03:49 -0800247 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x1.c",
Marat Dukhanf46f6752020-01-21 11:03:49 -0800248 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x2-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700249 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x2.c",
Marat Dukhanf46f6752020-01-21 11:03:49 -0800250 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x4-acc2.c",
251 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x4-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700252 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x4.c",
Marat Dukhanf46f6752020-01-21 11:03:49 -0800253 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x1.c",
Marat Dukhanf46f6752020-01-21 11:03:49 -0800254 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x2-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700255 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x2.c",
Marat Dukhanf46f6752020-01-21 11:03:49 -0800256 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x4-acc2.c",
257 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x4-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700258 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x4.c",
Frank Barchardfb158e22020-07-15 16:10:10 -0700259 "src/f32-relu/gen/scalar-x1.c",
260 "src/f32-relu/gen/scalar-x2.c",
261 "src/f32-relu/gen/scalar-x4.c",
Frank Barchard37297a62020-08-27 12:15:10 -0700262 "src/f32-relu/gen/scalar-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700263 "src/f32-rmax/scalar.c",
Marat Dukhan3a77ea72019-12-23 12:10:24 -0800264 "src/f32-sigmoid/gen/scalar-lut64-p2-div-x1.c",
265 "src/f32-sigmoid/gen/scalar-lut64-p2-div-x2.c",
266 "src/f32-sigmoid/gen/scalar-lut64-p2-div-x4.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700267 "src/f32-sigmoid/gen/scalar-lut2048-p1-div-x1.c",
268 "src/f32-sigmoid/gen/scalar-lut2048-p1-div-x2.c",
269 "src/f32-sigmoid/gen/scalar-lut2048-p1-div-x4.c",
Marat Dukhan3a77ea72019-12-23 12:10:24 -0800270 "src/f32-sigmoid/gen/scalar-p5-div-x1.c",
271 "src/f32-sigmoid/gen/scalar-p5-div-x2.c",
272 "src/f32-sigmoid/gen/scalar-p5-div-x4.c",
Marat Dukhan355ab432020-04-09 19:01:52 -0700273 "src/f32-spmm/gen/1x1-minmax-scalar-pipelined.c",
274 "src/f32-spmm/gen/1x1-minmax-scalar.c",
275 "src/f32-spmm/gen/2x1-minmax-scalar-pipelined.c",
276 "src/f32-spmm/gen/2x1-minmax-scalar.c",
277 "src/f32-spmm/gen/4x1-minmax-scalar-pipelined.c",
278 "src/f32-spmm/gen/4x1-minmax-scalar.c",
279 "src/f32-spmm/gen/8x1-minmax-scalar-pipelined.c",
280 "src/f32-spmm/gen/8x1-minmax-scalar.c",
281 "src/f32-spmm/gen/8x2-minmax-scalar.c",
282 "src/f32-spmm/gen/8x4-minmax-scalar.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700283 "src/f32-vbinary/gen/vadd-minmax-scalar-x1.c",
284 "src/f32-vbinary/gen/vadd-minmax-scalar-x2.c",
285 "src/f32-vbinary/gen/vadd-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700286 "src/f32-vbinary/gen/vadd-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700287 "src/f32-vbinary/gen/vadd-relu-scalar-x1.c",
288 "src/f32-vbinary/gen/vadd-relu-scalar-x2.c",
289 "src/f32-vbinary/gen/vadd-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700290 "src/f32-vbinary/gen/vadd-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700291 "src/f32-vbinary/gen/vadd-scalar-x1.c",
292 "src/f32-vbinary/gen/vadd-scalar-x2.c",
293 "src/f32-vbinary/gen/vadd-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700294 "src/f32-vbinary/gen/vadd-scalar-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700295 "src/f32-vbinary/gen/vaddc-minmax-scalar-x1.c",
296 "src/f32-vbinary/gen/vaddc-minmax-scalar-x2.c",
297 "src/f32-vbinary/gen/vaddc-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700298 "src/f32-vbinary/gen/vaddc-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700299 "src/f32-vbinary/gen/vaddc-relu-scalar-x1.c",
300 "src/f32-vbinary/gen/vaddc-relu-scalar-x2.c",
301 "src/f32-vbinary/gen/vaddc-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700302 "src/f32-vbinary/gen/vaddc-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700303 "src/f32-vbinary/gen/vaddc-scalar-x1.c",
304 "src/f32-vbinary/gen/vaddc-scalar-x2.c",
305 "src/f32-vbinary/gen/vaddc-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700306 "src/f32-vbinary/gen/vaddc-scalar-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700307 "src/f32-vbinary/gen/vdiv-minmax-scalar-x1.c",
308 "src/f32-vbinary/gen/vdiv-minmax-scalar-x2.c",
309 "src/f32-vbinary/gen/vdiv-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700310 "src/f32-vbinary/gen/vdiv-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700311 "src/f32-vbinary/gen/vdiv-relu-scalar-x1.c",
312 "src/f32-vbinary/gen/vdiv-relu-scalar-x2.c",
313 "src/f32-vbinary/gen/vdiv-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700314 "src/f32-vbinary/gen/vdiv-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700315 "src/f32-vbinary/gen/vdiv-scalar-x1.c",
316 "src/f32-vbinary/gen/vdiv-scalar-x2.c",
317 "src/f32-vbinary/gen/vdiv-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700318 "src/f32-vbinary/gen/vdiv-scalar-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700319 "src/f32-vbinary/gen/vdivc-minmax-scalar-x1.c",
320 "src/f32-vbinary/gen/vdivc-minmax-scalar-x2.c",
321 "src/f32-vbinary/gen/vdivc-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700322 "src/f32-vbinary/gen/vdivc-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700323 "src/f32-vbinary/gen/vdivc-relu-scalar-x1.c",
324 "src/f32-vbinary/gen/vdivc-relu-scalar-x2.c",
325 "src/f32-vbinary/gen/vdivc-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700326 "src/f32-vbinary/gen/vdivc-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700327 "src/f32-vbinary/gen/vdivc-scalar-x1.c",
328 "src/f32-vbinary/gen/vdivc-scalar-x2.c",
329 "src/f32-vbinary/gen/vdivc-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700330 "src/f32-vbinary/gen/vdivc-scalar-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800331 "src/f32-vbinary/gen/vmax-scalar-x1.c",
332 "src/f32-vbinary/gen/vmax-scalar-x2.c",
333 "src/f32-vbinary/gen/vmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700334 "src/f32-vbinary/gen/vmax-scalar-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800335 "src/f32-vbinary/gen/vmaxc-scalar-x1.c",
336 "src/f32-vbinary/gen/vmaxc-scalar-x2.c",
337 "src/f32-vbinary/gen/vmaxc-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700338 "src/f32-vbinary/gen/vmaxc-scalar-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800339 "src/f32-vbinary/gen/vmin-scalar-x1.c",
340 "src/f32-vbinary/gen/vmin-scalar-x2.c",
341 "src/f32-vbinary/gen/vmin-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700342 "src/f32-vbinary/gen/vmin-scalar-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800343 "src/f32-vbinary/gen/vminc-scalar-x1.c",
344 "src/f32-vbinary/gen/vminc-scalar-x2.c",
345 "src/f32-vbinary/gen/vminc-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700346 "src/f32-vbinary/gen/vminc-scalar-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700347 "src/f32-vbinary/gen/vmul-minmax-scalar-x1.c",
348 "src/f32-vbinary/gen/vmul-minmax-scalar-x2.c",
349 "src/f32-vbinary/gen/vmul-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700350 "src/f32-vbinary/gen/vmul-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700351 "src/f32-vbinary/gen/vmul-relu-scalar-x1.c",
352 "src/f32-vbinary/gen/vmul-relu-scalar-x2.c",
353 "src/f32-vbinary/gen/vmul-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700354 "src/f32-vbinary/gen/vmul-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700355 "src/f32-vbinary/gen/vmul-scalar-x1.c",
356 "src/f32-vbinary/gen/vmul-scalar-x2.c",
357 "src/f32-vbinary/gen/vmul-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700358 "src/f32-vbinary/gen/vmul-scalar-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700359 "src/f32-vbinary/gen/vmulc-minmax-scalar-x1.c",
360 "src/f32-vbinary/gen/vmulc-minmax-scalar-x2.c",
361 "src/f32-vbinary/gen/vmulc-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700362 "src/f32-vbinary/gen/vmulc-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700363 "src/f32-vbinary/gen/vmulc-relu-scalar-x1.c",
364 "src/f32-vbinary/gen/vmulc-relu-scalar-x2.c",
365 "src/f32-vbinary/gen/vmulc-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700366 "src/f32-vbinary/gen/vmulc-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700367 "src/f32-vbinary/gen/vmulc-scalar-x1.c",
368 "src/f32-vbinary/gen/vmulc-scalar-x2.c",
369 "src/f32-vbinary/gen/vmulc-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700370 "src/f32-vbinary/gen/vmulc-scalar-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700371 "src/f32-vbinary/gen/vrdivc-minmax-scalar-x1.c",
372 "src/f32-vbinary/gen/vrdivc-minmax-scalar-x2.c",
373 "src/f32-vbinary/gen/vrdivc-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700374 "src/f32-vbinary/gen/vrdivc-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700375 "src/f32-vbinary/gen/vrdivc-relu-scalar-x1.c",
376 "src/f32-vbinary/gen/vrdivc-relu-scalar-x2.c",
377 "src/f32-vbinary/gen/vrdivc-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700378 "src/f32-vbinary/gen/vrdivc-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700379 "src/f32-vbinary/gen/vrdivc-scalar-x1.c",
380 "src/f32-vbinary/gen/vrdivc-scalar-x2.c",
381 "src/f32-vbinary/gen/vrdivc-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700382 "src/f32-vbinary/gen/vrdivc-scalar-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700383 "src/f32-vbinary/gen/vrsubc-minmax-scalar-x1.c",
384 "src/f32-vbinary/gen/vrsubc-minmax-scalar-x2.c",
385 "src/f32-vbinary/gen/vrsubc-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700386 "src/f32-vbinary/gen/vrsubc-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700387 "src/f32-vbinary/gen/vrsubc-relu-scalar-x1.c",
388 "src/f32-vbinary/gen/vrsubc-relu-scalar-x2.c",
389 "src/f32-vbinary/gen/vrsubc-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700390 "src/f32-vbinary/gen/vrsubc-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700391 "src/f32-vbinary/gen/vrsubc-scalar-x1.c",
392 "src/f32-vbinary/gen/vrsubc-scalar-x2.c",
393 "src/f32-vbinary/gen/vrsubc-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700394 "src/f32-vbinary/gen/vrsubc-scalar-x8.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -0700395 "src/f32-vbinary/gen/vsqrdiff-scalar-x1.c",
396 "src/f32-vbinary/gen/vsqrdiff-scalar-x2.c",
397 "src/f32-vbinary/gen/vsqrdiff-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700398 "src/f32-vbinary/gen/vsqrdiff-scalar-x8.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -0700399 "src/f32-vbinary/gen/vsqrdiffc-scalar-x1.c",
400 "src/f32-vbinary/gen/vsqrdiffc-scalar-x2.c",
401 "src/f32-vbinary/gen/vsqrdiffc-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700402 "src/f32-vbinary/gen/vsqrdiffc-scalar-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700403 "src/f32-vbinary/gen/vsub-minmax-scalar-x1.c",
404 "src/f32-vbinary/gen/vsub-minmax-scalar-x2.c",
405 "src/f32-vbinary/gen/vsub-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700406 "src/f32-vbinary/gen/vsub-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700407 "src/f32-vbinary/gen/vsub-relu-scalar-x1.c",
408 "src/f32-vbinary/gen/vsub-relu-scalar-x2.c",
409 "src/f32-vbinary/gen/vsub-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700410 "src/f32-vbinary/gen/vsub-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700411 "src/f32-vbinary/gen/vsub-scalar-x1.c",
412 "src/f32-vbinary/gen/vsub-scalar-x2.c",
413 "src/f32-vbinary/gen/vsub-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700414 "src/f32-vbinary/gen/vsub-scalar-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700415 "src/f32-vbinary/gen/vsubc-minmax-scalar-x1.c",
416 "src/f32-vbinary/gen/vsubc-minmax-scalar-x2.c",
417 "src/f32-vbinary/gen/vsubc-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700418 "src/f32-vbinary/gen/vsubc-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700419 "src/f32-vbinary/gen/vsubc-relu-scalar-x1.c",
420 "src/f32-vbinary/gen/vsubc-relu-scalar-x2.c",
421 "src/f32-vbinary/gen/vsubc-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700422 "src/f32-vbinary/gen/vsubc-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700423 "src/f32-vbinary/gen/vsubc-scalar-x1.c",
424 "src/f32-vbinary/gen/vsubc-scalar-x2.c",
425 "src/f32-vbinary/gen/vsubc-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700426 "src/f32-vbinary/gen/vsubc-scalar-x8.c",
Marat Dukhaned6baaf2020-12-01 15:07:08 -0800427 "src/f32-velu/gen/velu-scalar-rr2-lut16-p3-x1.c",
428 "src/f32-velu/gen/velu-scalar-rr2-lut16-p3-x2.c",
429 "src/f32-velu/gen/velu-scalar-rr2-lut16-p3-x3.c",
430 "src/f32-velu/gen/velu-scalar-rr2-lut16-p3-x4.c",
431 "src/f32-velu/gen/velu-scalar-rr2-lut16-p3-x5.c",
432 "src/f32-velu/gen/velu-scalar-rr2-lut16-p3-x6.c",
433 "src/f32-velu/gen/velu-scalar-rr2-p6-x1.c",
434 "src/f32-velu/gen/velu-scalar-rr2-p6-x2.c",
435 "src/f32-velu/gen/velu-scalar-rr2-p6-x3.c",
436 "src/f32-velu/gen/velu-scalar-rr2-p6-x4.c",
437 "src/f32-velu/gen/velu-scalar-rr2-p6-x5.c",
438 "src/f32-velu/gen/velu-scalar-rr2-p6-x6.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700439 "src/f32-vlrelu/gen/vlrelu-scalar-x1.c",
440 "src/f32-vlrelu/gen/vlrelu-scalar-x2.c",
441 "src/f32-vlrelu/gen/vlrelu-scalar-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700442 "src/f32-vmulcaddc/gen/c1-minmax-scalar-2x.c",
443 "src/f32-vmulcaddc/gen/c2-minmax-scalar-2x.c",
444 "src/f32-vmulcaddc/gen/c4-minmax-scalar-2x.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -0700445 "src/f32-vrnd/gen/vrndd-scalar-libm-x1.c",
446 "src/f32-vrnd/gen/vrndd-scalar-libm-x2.c",
447 "src/f32-vrnd/gen/vrndd-scalar-libm-x4.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700448 "src/f32-vrnd/gen/vrndne-scalar-libm-x1.c",
449 "src/f32-vrnd/gen/vrndne-scalar-libm-x2.c",
450 "src/f32-vrnd/gen/vrndne-scalar-libm-x4.c",
451 "src/f32-vrnd/gen/vrndu-scalar-libm-x1.c",
452 "src/f32-vrnd/gen/vrndu-scalar-libm-x2.c",
453 "src/f32-vrnd/gen/vrndu-scalar-libm-x4.c",
454 "src/f32-vrnd/gen/vrndz-scalar-libm-x1.c",
455 "src/f32-vrnd/gen/vrndz-scalar-libm-x2.c",
456 "src/f32-vrnd/gen/vrndz-scalar-libm-x4.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -0700457 "src/f32-vsqrt/gen/scalar-sqrt-x1.c",
458 "src/f32-vsqrt/gen/scalar-sqrt-x2.c",
459 "src/f32-vsqrt/gen/scalar-sqrt-x4.c",
Marat Dukhan5020b962020-06-08 13:30:10 -0700460 "src/f32-vunary/gen/vabs-scalar-x1.c",
461 "src/f32-vunary/gen/vabs-scalar-x2.c",
462 "src/f32-vunary/gen/vabs-scalar-x4.c",
463 "src/f32-vunary/gen/vneg-scalar-x1.c",
464 "src/f32-vunary/gen/vneg-scalar-x2.c",
465 "src/f32-vunary/gen/vneg-scalar-x4.c",
466 "src/f32-vunary/gen/vsqr-scalar-x1.c",
467 "src/f32-vunary/gen/vsqr-scalar-x2.c",
468 "src/f32-vunary/gen/vsqr-scalar-x4.c",
Marat Dukhande390d42020-11-29 19:32:18 -0800469 "src/math/expm1minus-scalar-rr2-lut4-p4.c",
470 "src/math/expm1minus-scalar-rr2-lut8-p3.c",
471 "src/math/expm1minus-scalar-rr2-lut8-p4.c",
Marat Dukhanc60742b2020-11-23 12:33:27 -0800472 "src/math/expm1minus-scalar-rr2-lut16-p3.c",
473 "src/math/expm1minus-scalar-rr2-lut16-p4.c",
474 "src/math/expm1minus-scalar-rr2-p5.c",
475 "src/math/expm1minus-scalar-rr2-p6.c",
Frank Barchard22136062020-11-24 18:44:46 -0800476 "src/math/expminus-scalar-rr2-lut64-p2.c",
477 "src/math/expminus-scalar-rr2-lut2048-p1.c",
478 "src/math/expminus-scalar-rr2-p5.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -0700479 "src/math/roundd-scalar-addsub.c",
480 "src/math/roundd-scalar-cvt.c",
Marat Dukhanffbf96a2020-05-14 02:59:08 -0700481 "src/math/roundd-scalar-floor.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700482 "src/math/roundne-scalar-addsub.c",
483 "src/math/roundne-scalar-nearbyint.c",
484 "src/math/roundne-scalar-rint.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -0700485 "src/math/roundu-scalar-addsub.c",
Marat Dukhanffbf96a2020-05-14 02:59:08 -0700486 "src/math/roundu-scalar-ceil.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -0700487 "src/math/roundu-scalar-cvt.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -0700488 "src/math/roundz-scalar-addsub.c",
489 "src/math/roundz-scalar-cvt.c",
Marat Dukhanffbf96a2020-05-14 02:59:08 -0700490 "src/math/roundz-scalar-trunc.c",
Marat Dukhanf8475d62020-09-17 15:01:43 -0700491 "src/math/sigmoid-scalar-rr2-lut64-p2-div.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700492 "src/math/sigmoid-scalar-rr2-lut2048-p1-div.c",
Marat Dukhanf8475d62020-09-17 15:01:43 -0700493 "src/math/sigmoid-scalar-rr2-p5-div.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -0700494 "src/qs8-requantization/fp32-scalar-lrintf.c",
495 "src/qs8-requantization/fp32-scalar-magic.c",
496 "src/qs8-requantization/precise-scalar-signed64.c",
497 "src/qs8-requantization/precise-scalar-unsigned32.c",
498 "src/qs8-requantization/precise-scalar-unsigned64.c",
499 "src/qs8-requantization/q31-scalar.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -0700500 "src/qu8-avgpool/9p8x-minmax-scalar-c1.c",
501 "src/qu8-avgpool/9x-minmax-scalar-c1.c",
502 "src/qu8-dwconv/up1x9-minmax-scalar.c",
503 "src/qu8-gavgpool/7p7x-minmax-scalar-c1.c",
504 "src/qu8-gavgpool/7x-minmax-scalar-c1.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700505 "src/qu8-gemm/2x2-minmax-scalar.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -0700506 "src/qu8-igemm/2x2-minmax-scalar.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -0700507 "src/qu8-requantization/fp32-scalar-lrintf.c",
508 "src/qu8-requantization/fp32-scalar-magic.c",
509 "src/qu8-requantization/precise-scalar-signed64.c",
510 "src/qu8-requantization/precise-scalar-unsigned32.c",
511 "src/qu8-requantization/precise-scalar-unsigned64.c",
512 "src/qu8-requantization/q31-scalar.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -0700513 "src/qu8-vadd/minmax-scalar.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -0700514 "src/u8-clamp/scalar-x4.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700515 "src/u8-lut32norm/scalar.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700516 "src/u8-maxpool/9p8x-minmax-scalar-c1.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700517 "src/u8-rmax/scalar.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700518 "src/x8-lut/scalar.c",
519 "src/x8-zip/x2-scalar.c",
520 "src/x8-zip/x3-scalar.c",
521 "src/x8-zip/x4-scalar.c",
522 "src/x8-zip/xm-scalar.c",
Marat Dukhanad71b9a2020-11-20 00:01:51 -0800523 "src/x32-depthtospace2d-chw2hwc/scalar.c",
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -0700524 "src/x32-fill/scalar-float.c",
525 "src/x32-fill/scalar-int.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700526 "src/x32-packx/x2-scalar.c",
527 "src/x32-packx/x3-scalar.c",
528 "src/x32-packx/x4-scalar.c",
Marat Dukhan63523d42020-05-22 17:07:33 -0700529 "src/x32-pad/scalar-float.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700530 "src/x32-pad/scalar-int.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700531 "src/x32-unpool/scalar.c",
532 "src/x32-zip/x2-scalar.c",
533 "src/x32-zip/x3-scalar.c",
534 "src/x32-zip/x4-scalar.c",
535 "src/x32-zip/xm-scalar.c",
Marat Dukhan048931b2020-11-24 20:53:54 -0800536 "src/xx-copy/memcpy.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700537]
538
Marat Dukhan436ebe62019-12-04 15:10:12 -0800539WASM_UKERNELS = [
Marat Dukhan99936602020-04-11 16:47:01 -0700540 "src/f32-avgpool/9p8x-minmax-wasm-c1.c",
541 "src/f32-avgpool/9x-minmax-wasm-c1.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -0700542 "src/f32-clamp/gen/wasm-x1.c",
543 "src/f32-clamp/gen/wasm-x2.c",
544 "src/f32-clamp/gen/wasm-x4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700545 "src/f32-dwconv/gen/up1x4-minmax-wasm-acc2.c",
546 "src/f32-dwconv/gen/up1x4-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700547 "src/f32-dwconv/gen/up1x4-wasm-acc2.c",
548 "src/f32-dwconv/gen/up1x4-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700549 "src/f32-dwconv/gen/up1x9-minmax-wasm-acc2.c",
550 "src/f32-dwconv/gen/up1x9-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700551 "src/f32-dwconv/gen/up1x9-wasm-acc2.c",
552 "src/f32-dwconv/gen/up1x9-wasm.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700553 "src/f32-dwconv/gen/up1x25-minmax-wasm-acc2.c",
554 "src/f32-dwconv/gen/up1x25-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700555 "src/f32-dwconv/gen/up1x25-wasm-acc2.c",
556 "src/f32-dwconv/gen/up1x25-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700557 "src/f32-dwconv/gen/up2x4-minmax-wasm-acc2.c",
558 "src/f32-dwconv/gen/up2x4-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700559 "src/f32-dwconv/gen/up2x4-wasm-acc2.c",
560 "src/f32-dwconv/gen/up2x4-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700561 "src/f32-dwconv/gen/up2x9-minmax-wasm-acc2.c",
562 "src/f32-dwconv/gen/up2x9-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700563 "src/f32-dwconv/gen/up2x9-wasm-acc2.c",
564 "src/f32-dwconv/gen/up2x9-wasm.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700565 "src/f32-dwconv/gen/up2x25-minmax-wasm-acc2.c",
566 "src/f32-dwconv/gen/up2x25-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700567 "src/f32-dwconv/gen/up2x25-wasm-acc2.c",
568 "src/f32-dwconv/gen/up2x25-wasm.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700569 "src/f32-gavgpool/7p7x-minmax-wasm-c1.c",
570 "src/f32-gavgpool/7x-minmax-wasm-c1.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700571 "src/f32-gemm/gen-inc/1x4inc-minmax-wasm.c",
572 "src/f32-gemm/gen-inc/2x4inc-minmax-wasm.c",
573 "src/f32-gemm/gen-inc/4x4inc-minmax-wasm.c",
574 "src/f32-gemm/gen/1x4-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700575 "src/f32-gemm/gen/1x4-relu-wasm.c",
576 "src/f32-gemm/gen/1x4-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700577 "src/f32-gemm/gen/2x4-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700578 "src/f32-gemm/gen/2x4-relu-wasm.c",
579 "src/f32-gemm/gen/2x4-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700580 "src/f32-gemm/gen/4x2-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700581 "src/f32-gemm/gen/4x2-relu-wasm.c",
582 "src/f32-gemm/gen/4x2-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700583 "src/f32-gemm/gen/4x4-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700584 "src/f32-gemm/gen/4x4-relu-wasm.c",
585 "src/f32-gemm/gen/4x4-wasm.c",
Marat Dukhanefaac272020-07-16 16:58:07 -0700586 "src/f32-hswish/gen/hswish-wasm-x1.c",
587 "src/f32-hswish/gen/hswish-wasm-x2.c",
588 "src/f32-hswish/gen/hswish-wasm-x4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700589 "src/f32-igemm/gen/1x4-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700590 "src/f32-igemm/gen/1x4-relu-wasm.c",
591 "src/f32-igemm/gen/1x4-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700592 "src/f32-igemm/gen/2x4-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700593 "src/f32-igemm/gen/2x4-relu-wasm.c",
594 "src/f32-igemm/gen/2x4-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700595 "src/f32-igemm/gen/4x2-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700596 "src/f32-igemm/gen/4x2-relu-wasm.c",
597 "src/f32-igemm/gen/4x2-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700598 "src/f32-igemm/gen/4x4-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700599 "src/f32-igemm/gen/4x4-relu-wasm.c",
600 "src/f32-igemm/gen/4x4-wasm.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700601 "src/f32-maxpool/9p8x-minmax-wasm-c1.c",
602 "src/f32-pavgpool/9p8x-minmax-wasm-c1.c",
603 "src/f32-pavgpool/9x-minmax-wasm-c1.c",
Marat Dukhan7c1f8082020-06-25 13:26:20 -0700604 "src/f32-prelu/gen/wasm-2x1.c",
605 "src/f32-prelu/gen/wasm-2x4.c",
Frank Barchardfb158e22020-07-15 16:10:10 -0700606 "src/f32-relu/gen/wasm-x1.c",
607 "src/f32-relu/gen/wasm-x2.c",
608 "src/f32-relu/gen/wasm-x4.c",
Frank Barchard37297a62020-08-27 12:15:10 -0700609 "src/f32-relu/gen/wasm-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700610 "src/f32-vbinary/gen/vadd-minmax-wasm-x1.c",
611 "src/f32-vbinary/gen/vadd-minmax-wasm-x2.c",
612 "src/f32-vbinary/gen/vadd-minmax-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700613 "src/f32-vbinary/gen/vadd-minmax-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700614 "src/f32-vbinary/gen/vadd-relu-wasm-x1.c",
615 "src/f32-vbinary/gen/vadd-relu-wasm-x2.c",
616 "src/f32-vbinary/gen/vadd-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700617 "src/f32-vbinary/gen/vadd-relu-wasm-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700618 "src/f32-vbinary/gen/vaddc-minmax-wasm-x1.c",
619 "src/f32-vbinary/gen/vaddc-minmax-wasm-x2.c",
620 "src/f32-vbinary/gen/vaddc-minmax-wasm-x4.c",
621 "src/f32-vbinary/gen/vaddc-minmax-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700622 "src/f32-vbinary/gen/vaddc-relu-wasm-x1.c",
623 "src/f32-vbinary/gen/vaddc-relu-wasm-x2.c",
624 "src/f32-vbinary/gen/vaddc-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700625 "src/f32-vbinary/gen/vaddc-relu-wasm-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700626 "src/f32-vbinary/gen/vdiv-minmax-wasm-x1.c",
627 "src/f32-vbinary/gen/vdiv-minmax-wasm-x2.c",
628 "src/f32-vbinary/gen/vdiv-minmax-wasm-x4.c",
629 "src/f32-vbinary/gen/vdiv-minmax-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700630 "src/f32-vbinary/gen/vdiv-relu-wasm-x1.c",
631 "src/f32-vbinary/gen/vdiv-relu-wasm-x2.c",
632 "src/f32-vbinary/gen/vdiv-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700633 "src/f32-vbinary/gen/vdiv-relu-wasm-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700634 "src/f32-vbinary/gen/vdivc-minmax-wasm-x1.c",
635 "src/f32-vbinary/gen/vdivc-minmax-wasm-x2.c",
636 "src/f32-vbinary/gen/vdivc-minmax-wasm-x4.c",
637 "src/f32-vbinary/gen/vdivc-minmax-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700638 "src/f32-vbinary/gen/vdivc-relu-wasm-x1.c",
639 "src/f32-vbinary/gen/vdivc-relu-wasm-x2.c",
640 "src/f32-vbinary/gen/vdivc-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700641 "src/f32-vbinary/gen/vdivc-relu-wasm-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800642 "src/f32-vbinary/gen/vmax-wasm-x1.c",
643 "src/f32-vbinary/gen/vmax-wasm-x2.c",
644 "src/f32-vbinary/gen/vmax-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700645 "src/f32-vbinary/gen/vmax-wasm-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800646 "src/f32-vbinary/gen/vmaxc-wasm-x1.c",
647 "src/f32-vbinary/gen/vmaxc-wasm-x2.c",
648 "src/f32-vbinary/gen/vmaxc-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700649 "src/f32-vbinary/gen/vmaxc-wasm-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800650 "src/f32-vbinary/gen/vmin-wasm-x1.c",
651 "src/f32-vbinary/gen/vmin-wasm-x2.c",
652 "src/f32-vbinary/gen/vmin-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700653 "src/f32-vbinary/gen/vmin-wasm-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800654 "src/f32-vbinary/gen/vminc-wasm-x1.c",
655 "src/f32-vbinary/gen/vminc-wasm-x2.c",
656 "src/f32-vbinary/gen/vminc-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700657 "src/f32-vbinary/gen/vminc-wasm-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700658 "src/f32-vbinary/gen/vmul-minmax-wasm-x1.c",
659 "src/f32-vbinary/gen/vmul-minmax-wasm-x2.c",
660 "src/f32-vbinary/gen/vmul-minmax-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700661 "src/f32-vbinary/gen/vmul-minmax-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700662 "src/f32-vbinary/gen/vmul-relu-wasm-x1.c",
663 "src/f32-vbinary/gen/vmul-relu-wasm-x2.c",
664 "src/f32-vbinary/gen/vmul-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700665 "src/f32-vbinary/gen/vmul-relu-wasm-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700666 "src/f32-vbinary/gen/vmulc-minmax-wasm-x1.c",
667 "src/f32-vbinary/gen/vmulc-minmax-wasm-x2.c",
668 "src/f32-vbinary/gen/vmulc-minmax-wasm-x4.c",
669 "src/f32-vbinary/gen/vmulc-minmax-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700670 "src/f32-vbinary/gen/vmulc-relu-wasm-x1.c",
671 "src/f32-vbinary/gen/vmulc-relu-wasm-x2.c",
672 "src/f32-vbinary/gen/vmulc-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700673 "src/f32-vbinary/gen/vmulc-relu-wasm-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700674 "src/f32-vbinary/gen/vrdivc-minmax-wasm-x1.c",
675 "src/f32-vbinary/gen/vrdivc-minmax-wasm-x2.c",
676 "src/f32-vbinary/gen/vrdivc-minmax-wasm-x4.c",
677 "src/f32-vbinary/gen/vrdivc-minmax-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700678 "src/f32-vbinary/gen/vrdivc-relu-wasm-x1.c",
679 "src/f32-vbinary/gen/vrdivc-relu-wasm-x2.c",
680 "src/f32-vbinary/gen/vrdivc-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700681 "src/f32-vbinary/gen/vrdivc-relu-wasm-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700682 "src/f32-vbinary/gen/vrsubc-minmax-wasm-x1.c",
683 "src/f32-vbinary/gen/vrsubc-minmax-wasm-x2.c",
684 "src/f32-vbinary/gen/vrsubc-minmax-wasm-x4.c",
685 "src/f32-vbinary/gen/vrsubc-minmax-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700686 "src/f32-vbinary/gen/vrsubc-relu-wasm-x1.c",
687 "src/f32-vbinary/gen/vrsubc-relu-wasm-x2.c",
688 "src/f32-vbinary/gen/vrsubc-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700689 "src/f32-vbinary/gen/vrsubc-relu-wasm-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700690 "src/f32-vbinary/gen/vsub-minmax-wasm-x1.c",
691 "src/f32-vbinary/gen/vsub-minmax-wasm-x2.c",
692 "src/f32-vbinary/gen/vsub-minmax-wasm-x4.c",
693 "src/f32-vbinary/gen/vsub-minmax-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700694 "src/f32-vbinary/gen/vsub-relu-wasm-x1.c",
695 "src/f32-vbinary/gen/vsub-relu-wasm-x2.c",
696 "src/f32-vbinary/gen/vsub-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700697 "src/f32-vbinary/gen/vsub-relu-wasm-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700698 "src/f32-vbinary/gen/vsubc-minmax-wasm-x1.c",
699 "src/f32-vbinary/gen/vsubc-minmax-wasm-x2.c",
700 "src/f32-vbinary/gen/vsubc-minmax-wasm-x4.c",
701 "src/f32-vbinary/gen/vsubc-minmax-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700702 "src/f32-vbinary/gen/vsubc-relu-wasm-x1.c",
703 "src/f32-vbinary/gen/vsubc-relu-wasm-x2.c",
704 "src/f32-vbinary/gen/vsubc-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700705 "src/f32-vbinary/gen/vsubc-relu-wasm-x8.c",
Marat Dukhaned6baaf2020-12-01 15:07:08 -0800706 "src/f32-velu/gen/velu-wasm-rr2-lut16-p3-x1.c",
707 "src/f32-velu/gen/velu-wasm-rr2-lut16-p3-x2.c",
708 "src/f32-velu/gen/velu-wasm-rr2-lut16-p3-x3.c",
709 "src/f32-velu/gen/velu-wasm-rr2-lut16-p3-x4.c",
710 "src/f32-velu/gen/velu-wasm-rr2-lut16-p3-x5.c",
711 "src/f32-velu/gen/velu-wasm-rr2-lut16-p3-x6.c",
712 "src/f32-velu/gen/velu-wasm-rr2-p6-x1.c",
713 "src/f32-velu/gen/velu-wasm-rr2-p6-x2.c",
714 "src/f32-velu/gen/velu-wasm-rr2-p6-x3.c",
715 "src/f32-velu/gen/velu-wasm-rr2-p6-x4.c",
716 "src/f32-velu/gen/velu-wasm-rr2-p6-x5.c",
717 "src/f32-velu/gen/velu-wasm-rr2-p6-x6.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -0700718 "src/f32-vlrelu/gen/vlrelu-wasm-x1.c",
719 "src/f32-vlrelu/gen/vlrelu-wasm-x2.c",
720 "src/f32-vlrelu/gen/vlrelu-wasm-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700721 "src/f32-vmulcaddc/gen/c1-minmax-wasm-2x.c",
722 "src/f32-vmulcaddc/gen/c2-minmax-wasm-2x.c",
723 "src/f32-vmulcaddc/gen/c4-minmax-wasm-2x.c",
Marat Dukhan436ebe62019-12-04 15:10:12 -0800724]
725
Marat Dukhan290055c2020-06-09 12:24:29 -0700726WASMSIMD_UKERNELS = [
Marat Dukhan40f05522020-07-16 22:33:12 -0700727 "src/f32-argmaxpool/4x-wasmsimd-c4.c",
728 "src/f32-argmaxpool/9p8x-wasmsimd-c4.c",
729 "src/f32-argmaxpool/9x-wasmsimd-c4.c",
Marat Dukhan3b7432d2020-07-16 17:46:32 -0700730 "src/f32-avgpool/9p8x-minmax-wasmsimd-arm-c4.c",
731 "src/f32-avgpool/9p8x-minmax-wasmsimd-x86-c4.c",
732 "src/f32-avgpool/9x-minmax-wasmsimd-arm-c4.c",
733 "src/f32-avgpool/9x-minmax-wasmsimd-x86-c4.c",
Marat Dukhan3fa52c82020-07-08 12:54:33 -0700734 "src/f32-clamp/gen/wasmsimd-arm-x4.c",
735 "src/f32-clamp/gen/wasmsimd-arm-x8.c",
736 "src/f32-clamp/gen/wasmsimd-x86-x4.c",
737 "src/f32-clamp/gen/wasmsimd-x86-x8.c",
Frank Barchard22136062020-11-24 18:44:46 -0800738 "src/f32-conv-hwc2chw/3x3s2p1c3x4-wasmsimd-2x2.c",
Frank Barchard0725b8d2020-12-07 11:07:35 -0800739 "src/f32-dwconv/gen/up4x4-minmax-wasmsimd-arm-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700740 "src/f32-dwconv/gen/up4x4-minmax-wasmsimd-arm.c",
Frank Barchard0725b8d2020-12-07 11:07:35 -0800741 "src/f32-dwconv/gen/up4x4-minmax-wasmsimd-x86-acc2.c",
Marat Dukhanac014d72020-06-16 08:36:47 -0700742 "src/f32-dwconv/gen/up4x4-minmax-wasmsimd-x86.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700743 "src/f32-dwconv/gen/up4x4-wasmsimd.c",
Frank Barchard0725b8d2020-12-07 11:07:35 -0800744 "src/f32-dwconv/gen/up4x9-minmax-wasmsimd-arm-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700745 "src/f32-dwconv/gen/up4x9-minmax-wasmsimd-arm.c",
Frank Barchard0725b8d2020-12-07 11:07:35 -0800746 "src/f32-dwconv/gen/up4x9-minmax-wasmsimd-x86-acc2.c",
Marat Dukhanac014d72020-06-16 08:36:47 -0700747 "src/f32-dwconv/gen/up4x9-minmax-wasmsimd-x86.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700748 "src/f32-dwconv/gen/up4x9-wasmsimd.c",
Frank Barchard0725b8d2020-12-07 11:07:35 -0800749 "src/f32-dwconv/gen/up4x25-minmax-wasmsimd-arm-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700750 "src/f32-dwconv/gen/up4x25-minmax-wasmsimd-arm.c",
Frank Barchard0725b8d2020-12-07 11:07:35 -0800751 "src/f32-dwconv/gen/up4x25-minmax-wasmsimd-x86-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700752 "src/f32-dwconv/gen/up4x25-minmax-wasmsimd-x86.c",
753 "src/f32-dwconv/gen/up4x25-wasmsimd.c",
Frank Barchard0725b8d2020-12-07 11:07:35 -0800754 "src/f32-dwconv/gen/up8x4-minmax-wasmsimd-arm-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700755 "src/f32-dwconv/gen/up8x4-minmax-wasmsimd-arm.c",
Frank Barchard0725b8d2020-12-07 11:07:35 -0800756 "src/f32-dwconv/gen/up8x4-minmax-wasmsimd-x86-acc2.c",
Marat Dukhanac014d72020-06-16 08:36:47 -0700757 "src/f32-dwconv/gen/up8x4-minmax-wasmsimd-x86.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700758 "src/f32-dwconv/gen/up8x4-wasmsimd.c",
Frank Barchard0725b8d2020-12-07 11:07:35 -0800759 "src/f32-dwconv/gen/up8x9-minmax-wasmsimd-arm-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700760 "src/f32-dwconv/gen/up8x9-minmax-wasmsimd-arm.c",
Frank Barchard0725b8d2020-12-07 11:07:35 -0800761 "src/f32-dwconv/gen/up8x9-minmax-wasmsimd-x86-acc2.c",
Marat Dukhanac014d72020-06-16 08:36:47 -0700762 "src/f32-dwconv/gen/up8x9-minmax-wasmsimd-x86.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700763 "src/f32-dwconv/gen/up8x9-wasmsimd.c",
Frank Barchard0725b8d2020-12-07 11:07:35 -0800764 "src/f32-dwconv/gen/up8x25-minmax-wasmsimd-arm-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700765 "src/f32-dwconv/gen/up8x25-minmax-wasmsimd-arm.c",
Frank Barchard0725b8d2020-12-07 11:07:35 -0800766 "src/f32-dwconv/gen/up8x25-minmax-wasmsimd-x86-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700767 "src/f32-dwconv/gen/up8x25-minmax-wasmsimd-x86.c",
768 "src/f32-dwconv/gen/up8x25-wasmsimd.c",
Frank Barchard412e2f42020-12-11 11:40:50 -0800769 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-loadsplat-1x4-acc2.c",
770 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-loadsplat-1x4-acc3.c",
771 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-loadsplat-1x4-acc4.c",
772 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-loadsplat-1x4.c",
773 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-loadsplat-2x4-acc2.c",
774 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-loadsplat-2x4.c",
775 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-loadsplat-3x4.c",
776 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-loadsplat-4x4.c",
777 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-loadsplat-5x4.c",
778 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-loadsplat-6x4.c",
Frank Barchard02bb4292020-12-15 18:25:32 -0800779 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-splat-1x4-acc2.c",
780 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-splat-1x4-acc3.c",
781 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-splat-1x4-acc4.c",
782 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-splat-1x4.c",
783 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-splat-2x4-acc2.c",
784 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-splat-2x4.c",
785 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-splat-3x4.c",
786 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-splat-4x4.c",
787 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-splat-5x4.c",
788 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-splat-6x4.c",
Frank Barchard412e2f42020-12-11 11:40:50 -0800789 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-loadsplat-1x4-acc2.c",
790 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-loadsplat-1x4-acc3.c",
791 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-loadsplat-1x4-acc4.c",
792 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-loadsplat-1x4.c",
793 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-loadsplat-2x4-acc2.c",
794 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-loadsplat-2x4.c",
795 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-loadsplat-3x4.c",
796 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-loadsplat-4x4.c",
797 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-loadsplat-5x4.c",
798 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-loadsplat-6x4.c",
Frank Barchard02bb4292020-12-15 18:25:32 -0800799 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-splat-1x4-acc2.c",
800 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-splat-1x4-acc3.c",
801 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-splat-1x4-acc4.c",
802 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-splat-1x4.c",
803 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-splat-2x4-acc2.c",
804 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-splat-2x4.c",
805 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-splat-3x4.c",
806 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-splat-4x4.c",
807 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-splat-5x4.c",
808 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-splat-6x4.c",
Frank Barchardc5704bf2020-12-21 23:09:00 -0800809 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-arm-loadsplat-1x4-acc2.c",
810 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-arm-loadsplat-1x4-acc3.c",
811 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-arm-loadsplat-1x4-acc4.c",
812 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-arm-loadsplat-1x4.c",
813 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-arm-loadsplat-2x4-acc2.c",
814 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-arm-loadsplat-2x4.c",
815 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-arm-loadsplat-3x4.c",
816 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-arm-loadsplat-4x4.c",
Frank Barchard412e2f42020-12-11 11:40:50 -0800817 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-arm-splat-1x4-acc2.c",
818 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-arm-splat-1x4-acc3.c",
819 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-arm-splat-1x4-acc4.c",
820 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-arm-splat-1x4.c",
821 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-arm-splat-2x4-acc2.c",
822 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-arm-splat-2x4.c",
823 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-arm-splat-3x4.c",
824 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-arm-splat-4x4.c",
Frank Barchardcadd4222021-01-20 16:27:25 -0800825 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-x86-loadsplat-1x4-acc2.c",
826 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-x86-loadsplat-1x4-acc3.c",
827 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-x86-loadsplat-1x4-acc4.c",
828 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-x86-loadsplat-1x4.c",
829 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-x86-loadsplat-2x4-acc2.c",
830 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-x86-loadsplat-2x4.c",
831 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-x86-loadsplat-3x4.c",
832 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-x86-loadsplat-4x4.c",
Frank Barchard412e2f42020-12-11 11:40:50 -0800833 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-x86-splat-1x4-acc2.c",
834 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-x86-splat-1x4-acc3.c",
835 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-x86-splat-1x4-acc4.c",
836 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-x86-splat-1x4.c",
837 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-x86-splat-2x4-acc2.c",
838 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-x86-splat-2x4.c",
839 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-x86-splat-3x4.c",
840 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-x86-splat-4x4.c",
Frank Barchardb20dcd62020-12-15 16:46:14 -0800841 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-loadsplat-1x4-acc2.c",
842 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-loadsplat-1x4-acc3.c",
843 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-loadsplat-1x4-acc4.c",
844 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-loadsplat-1x4-acc5.c",
845 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-loadsplat-1x4.c",
846 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-loadsplat-2x4-acc2.c",
847 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-loadsplat-2x4-acc3.c",
848 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-loadsplat-2x4.c",
849 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-loadsplat-3x4-acc2.c",
850 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-loadsplat-3x4.c",
851 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-loadsplat-4x4-acc2.c",
852 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-loadsplat-4x4.c",
853 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-loadsplat-5x4.c",
Frank Barchard412e2f42020-12-11 11:40:50 -0800854 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-splat-1x4-acc2.c",
855 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-splat-1x4-acc3.c",
856 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-splat-1x4-acc4.c",
857 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-splat-1x4-acc5.c",
858 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-splat-1x4.c",
859 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-splat-2x4-acc2.c",
860 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-splat-2x4-acc3.c",
861 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-splat-2x4.c",
862 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-splat-3x4-acc2.c",
863 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-splat-3x4.c",
864 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-splat-4x4-acc2.c",
865 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-splat-4x4.c",
866 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-splat-5x4.c",
Frank Barchardb20dcd62020-12-15 16:46:14 -0800867 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-loadsplat-1x4-acc2.c",
868 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-loadsplat-1x4-acc3.c",
869 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-loadsplat-1x4-acc4.c",
870 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-loadsplat-1x4-acc5.c",
871 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-loadsplat-1x4.c",
872 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-loadsplat-2x4-acc2.c",
873 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-loadsplat-2x4-acc3.c",
874 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-loadsplat-2x4.c",
875 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-loadsplat-3x4-acc2.c",
876 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-loadsplat-3x4.c",
877 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-loadsplat-4x4-acc2.c",
878 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-loadsplat-4x4.c",
879 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-loadsplat-5x4.c",
Frank Barchard412e2f42020-12-11 11:40:50 -0800880 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-splat-1x4-acc2.c",
881 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-splat-1x4-acc3.c",
882 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-splat-1x4-acc4.c",
883 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-splat-1x4-acc5.c",
884 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-splat-1x4.c",
885 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-splat-2x4-acc2.c",
886 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-splat-2x4-acc3.c",
887 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-splat-2x4.c",
888 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-splat-3x4-acc2.c",
889 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-splat-3x4.c",
890 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-splat-4x4-acc2.c",
891 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-splat-4x4.c",
892 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-splat-5x4.c",
Frank Barchardc6889b32020-12-21 11:27:22 -0800893 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-arm-loadsplat-1x4-acc2.c",
894 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-arm-loadsplat-1x4-acc3.c",
895 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-arm-loadsplat-1x4-acc4.c",
896 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-arm-loadsplat-1x4-acc5.c",
897 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-arm-loadsplat-1x4.c",
898 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-arm-loadsplat-2x4-acc2.c",
899 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-arm-loadsplat-2x4-acc3.c",
900 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-arm-loadsplat-2x4.c",
901 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-arm-loadsplat-3x4-acc2.c",
902 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-arm-loadsplat-3x4.c",
Frank Barchard412e2f42020-12-11 11:40:50 -0800903 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-arm-splat-1x4-acc2.c",
904 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-arm-splat-1x4-acc3.c",
905 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-arm-splat-1x4-acc4.c",
906 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-arm-splat-1x4-acc5.c",
907 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-arm-splat-1x4.c",
908 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-arm-splat-2x4-acc2.c",
909 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-arm-splat-2x4-acc3.c",
910 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-arm-splat-2x4.c",
911 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-arm-splat-3x4-acc2.c",
912 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-arm-splat-3x4.c",
Frank Barchardc6889b32020-12-21 11:27:22 -0800913 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-x86-loadsplat-1x4-acc2.c",
914 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-x86-loadsplat-1x4-acc3.c",
915 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-x86-loadsplat-1x4-acc4.c",
916 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-x86-loadsplat-1x4-acc5.c",
917 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-x86-loadsplat-1x4.c",
918 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-x86-loadsplat-2x4-acc2.c",
919 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-x86-loadsplat-2x4-acc3.c",
920 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-x86-loadsplat-2x4.c",
921 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-x86-loadsplat-3x4-acc2.c",
922 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-x86-loadsplat-3x4.c",
Frank Barchard412e2f42020-12-11 11:40:50 -0800923 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-x86-splat-1x4-acc2.c",
924 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-x86-splat-1x4-acc3.c",
925 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-x86-splat-1x4-acc4.c",
926 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-x86-splat-1x4-acc5.c",
927 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-x86-splat-1x4.c",
928 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-x86-splat-2x4-acc2.c",
929 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-x86-splat-2x4-acc3.c",
930 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-x86-splat-2x4.c",
931 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-x86-splat-3x4-acc2.c",
932 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-x86-splat-3x4.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700933 "src/f32-gavgpool-cw/wasmsimd-arm-x4.c",
934 "src/f32-gavgpool-cw/wasmsimd-x86-x4.c",
Marat Dukhanc6016802020-07-16 18:51:28 -0700935 "src/f32-gavgpool/7p7x-minmax-wasmsimd-arm-c4.c",
936 "src/f32-gavgpool/7p7x-minmax-wasmsimd-x86-c4.c",
937 "src/f32-gavgpool/7x-minmax-wasmsimd-arm-c4.c",
938 "src/f32-gavgpool/7x-minmax-wasmsimd-x86-c4.c",
Frank Barchard0725b8d2020-12-07 11:07:35 -0800939 "src/f32-gemm/gen-inc/1x8inc-minmax-wasmsimd-arm-loadsplat.c",
940 "src/f32-gemm/gen-inc/1x8inc-minmax-wasmsimd-arm-splat.c",
941 "src/f32-gemm/gen-inc/1x8inc-minmax-wasmsimd-x86-loadsplat.c",
942 "src/f32-gemm/gen-inc/1x8inc-minmax-wasmsimd-x86-splat.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700943 "src/f32-gemm/gen-inc/1x8s4inc-minmax-wasmsimd-arm.c",
944 "src/f32-gemm/gen-inc/1x8s4inc-minmax-wasmsimd-x86.c",
Frank Barchard0725b8d2020-12-07 11:07:35 -0800945 "src/f32-gemm/gen-inc/3x8inc-minmax-wasmsimd-arm-loadsplat.c",
946 "src/f32-gemm/gen-inc/3x8inc-minmax-wasmsimd-arm-splat.c",
947 "src/f32-gemm/gen-inc/3x8inc-minmax-wasmsimd-x86-loadsplat.c",
948 "src/f32-gemm/gen-inc/3x8inc-minmax-wasmsimd-x86-splat.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700949 "src/f32-gemm/gen-inc/3x8s4inc-minmax-wasmsimd-arm.c",
950 "src/f32-gemm/gen-inc/3x8s4inc-minmax-wasmsimd-x86.c",
Frank Barchard0725b8d2020-12-07 11:07:35 -0800951 "src/f32-gemm/gen-inc/4x8inc-minmax-wasmsimd-arm-loadsplat.c",
952 "src/f32-gemm/gen-inc/4x8inc-minmax-wasmsimd-arm-splat.c",
953 "src/f32-gemm/gen-inc/4x8inc-minmax-wasmsimd-x86-loadsplat.c",
954 "src/f32-gemm/gen-inc/4x8inc-minmax-wasmsimd-x86-splat.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700955 "src/f32-gemm/gen-inc/4x8s4inc-minmax-wasmsimd-arm.c",
956 "src/f32-gemm/gen-inc/4x8s4inc-minmax-wasmsimd-x86.c",
Frank Barchard0725b8d2020-12-07 11:07:35 -0800957 "src/f32-gemm/gen-inc/5x8inc-minmax-wasmsimd-arm-loadsplat.c",
958 "src/f32-gemm/gen-inc/5x8inc-minmax-wasmsimd-arm-splat.c",
959 "src/f32-gemm/gen-inc/5x8inc-minmax-wasmsimd-x86-loadsplat.c",
960 "src/f32-gemm/gen-inc/5x8inc-minmax-wasmsimd-x86-splat.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700961 "src/f32-gemm/gen-inc/5x8s4inc-minmax-wasmsimd-arm.c",
962 "src/f32-gemm/gen-inc/5x8s4inc-minmax-wasmsimd-x86.c",
Frank Barchard0725b8d2020-12-07 11:07:35 -0800963 "src/f32-gemm/gen-inc/6x8inc-minmax-wasmsimd-arm-loadsplat.c",
964 "src/f32-gemm/gen-inc/6x8inc-minmax-wasmsimd-arm-splat.c",
965 "src/f32-gemm/gen-inc/6x8inc-minmax-wasmsimd-x86-loadsplat.c",
966 "src/f32-gemm/gen-inc/6x8inc-minmax-wasmsimd-x86-splat.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700967 "src/f32-gemm/gen-inc/6x8s4inc-minmax-wasmsimd-arm.c",
968 "src/f32-gemm/gen-inc/6x8s4inc-minmax-wasmsimd-x86.c",
Frank Barchard0725b8d2020-12-07 11:07:35 -0800969 "src/f32-gemm/gen/1x8-minmax-wasmsimd-arm-loadsplat.c",
970 "src/f32-gemm/gen/1x8-minmax-wasmsimd-arm-splat.c",
971 "src/f32-gemm/gen/1x8-minmax-wasmsimd-x86-loadsplat.c",
972 "src/f32-gemm/gen/1x8-minmax-wasmsimd-x86-splat.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700973 "src/f32-gemm/gen/1x8-relu-wasmsimd-splat.c",
974 "src/f32-gemm/gen/1x8-wasmsimd-splat.c",
975 "src/f32-gemm/gen/1x8s4-minmax-wasmsimd-arm.c",
976 "src/f32-gemm/gen/1x8s4-minmax-wasmsimd-x86.c",
Frank Barchard0725b8d2020-12-07 11:07:35 -0800977 "src/f32-gemm/gen/3x8-minmax-wasmsimd-arm-loadsplat.c",
978 "src/f32-gemm/gen/3x8-minmax-wasmsimd-arm-splat.c",
979 "src/f32-gemm/gen/3x8-minmax-wasmsimd-x86-loadsplat.c",
980 "src/f32-gemm/gen/3x8-minmax-wasmsimd-x86-splat.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700981 "src/f32-gemm/gen/3x8s4-minmax-wasmsimd-arm.c",
982 "src/f32-gemm/gen/3x8s4-minmax-wasmsimd-x86.c",
983 "src/f32-gemm/gen/4x2c4-minmax-wasmsimd-arm.c",
984 "src/f32-gemm/gen/4x2c4-minmax-wasmsimd-x86.c",
985 "src/f32-gemm/gen/4x2c4-relu-wasmsimd.c",
986 "src/f32-gemm/gen/4x2c4-wasmsimd.c",
Frank Barchard0725b8d2020-12-07 11:07:35 -0800987 "src/f32-gemm/gen/4x8-minmax-wasmsimd-arm-loadsplat.c",
988 "src/f32-gemm/gen/4x8-minmax-wasmsimd-arm-splat.c",
989 "src/f32-gemm/gen/4x8-minmax-wasmsimd-x86-loadsplat.c",
990 "src/f32-gemm/gen/4x8-minmax-wasmsimd-x86-splat.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700991 "src/f32-gemm/gen/4x8-relu-wasmsimd-splat.c",
992 "src/f32-gemm/gen/4x8-wasmsimd-splat.c",
993 "src/f32-gemm/gen/4x8s4-minmax-wasmsimd-arm.c",
994 "src/f32-gemm/gen/4x8s4-minmax-wasmsimd-x86.c",
Frank Barchard0725b8d2020-12-07 11:07:35 -0800995 "src/f32-gemm/gen/5x8-minmax-wasmsimd-arm-loadsplat.c",
996 "src/f32-gemm/gen/5x8-minmax-wasmsimd-arm-splat.c",
997 "src/f32-gemm/gen/5x8-minmax-wasmsimd-x86-loadsplat.c",
998 "src/f32-gemm/gen/5x8-minmax-wasmsimd-x86-splat.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700999 "src/f32-gemm/gen/5x8-relu-wasmsimd-splat.c",
1000 "src/f32-gemm/gen/5x8-wasmsimd-splat.c",
1001 "src/f32-gemm/gen/5x8s4-minmax-wasmsimd-arm.c",
1002 "src/f32-gemm/gen/5x8s4-minmax-wasmsimd-x86.c",
Frank Barchard0725b8d2020-12-07 11:07:35 -08001003 "src/f32-gemm/gen/6x8-minmax-wasmsimd-arm-loadsplat.c",
1004 "src/f32-gemm/gen/6x8-minmax-wasmsimd-arm-splat.c",
1005 "src/f32-gemm/gen/6x8-minmax-wasmsimd-x86-loadsplat.c",
1006 "src/f32-gemm/gen/6x8-minmax-wasmsimd-x86-splat.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001007 "src/f32-gemm/gen/6x8s4-minmax-wasmsimd-arm.c",
1008 "src/f32-gemm/gen/6x8s4-minmax-wasmsimd-x86.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07001009 "src/f32-hswish/gen/hswish-wasmsimd-x4.c",
1010 "src/f32-hswish/gen/hswish-wasmsimd-x8.c",
1011 "src/f32-hswish/gen/hswish-wasmsimd-x16.c",
XNNPACK Team965272b2020-10-23 21:10:15 -07001012 "src/f32-ibilinear-chw/gen/wasmsimd-p4.c",
1013 "src/f32-ibilinear-chw/gen/wasmsimd-p8.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001014 "src/f32-ibilinear/gen/wasmsimd-c4.c",
1015 "src/f32-ibilinear/gen/wasmsimd-c8.c",
Frank Barchard0725b8d2020-12-07 11:07:35 -08001016 "src/f32-igemm/gen/1x8-minmax-wasmsimd-arm-loadsplat.c",
1017 "src/f32-igemm/gen/1x8-minmax-wasmsimd-arm-splat.c",
1018 "src/f32-igemm/gen/1x8-minmax-wasmsimd-x86-loadsplat.c",
1019 "src/f32-igemm/gen/1x8-minmax-wasmsimd-x86-splat.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001020 "src/f32-igemm/gen/1x8-relu-wasmsimd-splat.c",
1021 "src/f32-igemm/gen/1x8-wasmsimd-splat.c",
1022 "src/f32-igemm/gen/1x8s4-minmax-wasmsimd-arm.c",
1023 "src/f32-igemm/gen/1x8s4-minmax-wasmsimd-x86.c",
Frank Barchard0725b8d2020-12-07 11:07:35 -08001024 "src/f32-igemm/gen/3x8-minmax-wasmsimd-arm-loadsplat.c",
1025 "src/f32-igemm/gen/3x8-minmax-wasmsimd-arm-splat.c",
1026 "src/f32-igemm/gen/3x8-minmax-wasmsimd-x86-loadsplat.c",
1027 "src/f32-igemm/gen/3x8-minmax-wasmsimd-x86-splat.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001028 "src/f32-igemm/gen/3x8s4-minmax-wasmsimd-arm.c",
1029 "src/f32-igemm/gen/3x8s4-minmax-wasmsimd-x86.c",
1030 "src/f32-igemm/gen/4x2c4-minmax-wasmsimd-arm.c",
1031 "src/f32-igemm/gen/4x2c4-minmax-wasmsimd-x86.c",
1032 "src/f32-igemm/gen/4x2c4-relu-wasmsimd.c",
1033 "src/f32-igemm/gen/4x2c4-wasmsimd.c",
Frank Barchard0725b8d2020-12-07 11:07:35 -08001034 "src/f32-igemm/gen/4x8-minmax-wasmsimd-arm-loadsplat.c",
1035 "src/f32-igemm/gen/4x8-minmax-wasmsimd-arm-splat.c",
1036 "src/f32-igemm/gen/4x8-minmax-wasmsimd-x86-loadsplat.c",
1037 "src/f32-igemm/gen/4x8-minmax-wasmsimd-x86-splat.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001038 "src/f32-igemm/gen/4x8-relu-wasmsimd-splat.c",
1039 "src/f32-igemm/gen/4x8-wasmsimd-splat.c",
1040 "src/f32-igemm/gen/4x8s4-minmax-wasmsimd-arm.c",
1041 "src/f32-igemm/gen/4x8s4-minmax-wasmsimd-x86.c",
Frank Barchard0725b8d2020-12-07 11:07:35 -08001042 "src/f32-igemm/gen/5x8-minmax-wasmsimd-arm-loadsplat.c",
1043 "src/f32-igemm/gen/5x8-minmax-wasmsimd-arm-splat.c",
1044 "src/f32-igemm/gen/5x8-minmax-wasmsimd-x86-loadsplat.c",
1045 "src/f32-igemm/gen/5x8-minmax-wasmsimd-x86-splat.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001046 "src/f32-igemm/gen/5x8-relu-wasmsimd-splat.c",
1047 "src/f32-igemm/gen/5x8-wasmsimd-splat.c",
1048 "src/f32-igemm/gen/5x8s4-minmax-wasmsimd-arm.c",
1049 "src/f32-igemm/gen/5x8s4-minmax-wasmsimd-x86.c",
Frank Barchard0725b8d2020-12-07 11:07:35 -08001050 "src/f32-igemm/gen/6x8-minmax-wasmsimd-arm-loadsplat.c",
1051 "src/f32-igemm/gen/6x8-minmax-wasmsimd-arm-splat.c",
1052 "src/f32-igemm/gen/6x8-minmax-wasmsimd-x86-loadsplat.c",
1053 "src/f32-igemm/gen/6x8-minmax-wasmsimd-x86-splat.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001054 "src/f32-igemm/gen/6x8s4-minmax-wasmsimd-arm.c",
1055 "src/f32-igemm/gen/6x8s4-minmax-wasmsimd-x86.c",
Marat Dukhanf6e24802020-07-08 22:20:40 -07001056 "src/f32-maxpool/9p8x-minmax-wasmsimd-arm-c4.c",
1057 "src/f32-maxpool/9p8x-minmax-wasmsimd-x86-c4.c",
Marat Dukhan1483c532020-07-16 18:08:19 -07001058 "src/f32-pavgpool/9p8x-minmax-wasmsimd-arm-c4.c",
1059 "src/f32-pavgpool/9p8x-minmax-wasmsimd-x86-c4.c",
1060 "src/f32-pavgpool/9x-minmax-wasmsimd-arm-c4.c",
1061 "src/f32-pavgpool/9x-minmax-wasmsimd-x86-c4.c",
Frank Barchard0725b8d2020-12-07 11:07:35 -08001062 "src/f32-ppmm/gen/4x8-minmax-wasmsimd-arm-splat.c",
1063 "src/f32-ppmm/gen/4x8-minmax-wasmsimd-x86-splat.c",
Frank Barcharda5316982020-07-23 13:19:28 -07001064 "src/f32-prelu/gen/wasmsimd-bitselect-1x4.c",
1065 "src/f32-prelu/gen/wasmsimd-bitselect-1x8.c",
1066 "src/f32-prelu/gen/wasmsimd-bitselect-1x16.c",
Marat Dukhan195f8eb2020-06-25 12:50:57 -07001067 "src/f32-prelu/gen/wasmsimd-bitselect-2x4.c",
1068 "src/f32-prelu/gen/wasmsimd-bitselect-2x8.c",
Frank Barcharda5316982020-07-23 13:19:28 -07001069 "src/f32-prelu/gen/wasmsimd-bitselect-2x16.c",
1070 "src/f32-prelu/gen/wasmsimd-bitselect-4x4.c",
1071 "src/f32-prelu/gen/wasmsimd-bitselect-4x8.c",
1072 "src/f32-prelu/gen/wasmsimd-bitselect-4x16.c",
1073 "src/f32-prelu/gen/wasmsimd-minmax-1x4.c",
1074 "src/f32-prelu/gen/wasmsimd-minmax-1x8.c",
1075 "src/f32-prelu/gen/wasmsimd-minmax-1x16.c",
Marat Dukhan195f8eb2020-06-25 12:50:57 -07001076 "src/f32-prelu/gen/wasmsimd-minmax-2x4.c",
1077 "src/f32-prelu/gen/wasmsimd-minmax-2x8.c",
Frank Barcharda5316982020-07-23 13:19:28 -07001078 "src/f32-prelu/gen/wasmsimd-minmax-2x16.c",
1079 "src/f32-prelu/gen/wasmsimd-minmax-4x4.c",
1080 "src/f32-prelu/gen/wasmsimd-minmax-4x8.c",
1081 "src/f32-prelu/gen/wasmsimd-minmax-4x16.c",
Marat Dukhan52238f02020-07-16 15:30:28 -07001082 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x4.c",
Marat Dukhan52238f02020-07-16 15:30:28 -07001083 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x8-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001084 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x8.c",
Marat Dukhan52238f02020-07-16 15:30:28 -07001085 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x12-acc2.c",
1086 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x12-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001087 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x12.c",
Marat Dukhan52238f02020-07-16 15:30:28 -07001088 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x16-acc2.c",
1089 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x16-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001090 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x16.c",
Marat Dukhan52238f02020-07-16 15:30:28 -07001091 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x20-acc2.c",
1092 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x20-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001093 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x20.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07001094 "src/f32-relu/gen/wasmsimd-x4.c",
1095 "src/f32-relu/gen/wasmsimd-x8.c",
Frank Barchard37297a62020-08-27 12:15:10 -07001096 "src/f32-relu/gen/wasmsimd-x16.c",
Marat Dukhan8c417962020-07-08 12:27:50 -07001097 "src/f32-rmax/wasmsimd-arm.c",
1098 "src/f32-rmax/wasmsimd-x86.c",
Marat Dukhand187a5b2020-07-20 01:07:17 -07001099 "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x4.c",
1100 "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x8.c",
1101 "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x12.c",
1102 "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x16.c",
1103 "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x20.c",
1104 "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x24.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001105 "src/f32-sigmoid/gen/wasmsimd-p5-div-x4.c",
1106 "src/f32-sigmoid/gen/wasmsimd-p5-div-x8.c",
1107 "src/f32-sigmoid/gen/wasmsimd-p5-div-x12.c",
1108 "src/f32-sigmoid/gen/wasmsimd-p5-div-x16.c",
1109 "src/f32-sigmoid/gen/wasmsimd-p5-div-x20.c",
1110 "src/f32-sigmoid/gen/wasmsimd-p5-div-x24.c",
Frank Barchard8ef44cd2020-11-03 12:30:23 -08001111 "src/f32-spmm/gen/4x1-minmax-wasmsimd-arm-pipelined-x2.c",
1112 "src/f32-spmm/gen/4x1-minmax-wasmsimd-arm-pipelined.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07001113 "src/f32-spmm/gen/4x1-minmax-wasmsimd-arm-x2.c",
1114 "src/f32-spmm/gen/4x1-minmax-wasmsimd-arm-x4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001115 "src/f32-spmm/gen/4x1-minmax-wasmsimd-arm.c",
Frank Barchard8ef44cd2020-11-03 12:30:23 -08001116 "src/f32-spmm/gen/4x1-minmax-wasmsimd-x86-pipelined-x2.c",
1117 "src/f32-spmm/gen/4x1-minmax-wasmsimd-x86-pipelined.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07001118 "src/f32-spmm/gen/4x1-minmax-wasmsimd-x86-x2.c",
1119 "src/f32-spmm/gen/4x1-minmax-wasmsimd-x86-x4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001120 "src/f32-spmm/gen/4x1-minmax-wasmsimd-x86.c",
Frank Barchard8ef44cd2020-11-03 12:30:23 -08001121 "src/f32-spmm/gen/8x1-minmax-wasmsimd-arm-pipelined-x2.c",
1122 "src/f32-spmm/gen/8x1-minmax-wasmsimd-arm-pipelined.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07001123 "src/f32-spmm/gen/8x1-minmax-wasmsimd-arm-x2.c",
1124 "src/f32-spmm/gen/8x1-minmax-wasmsimd-arm-x4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001125 "src/f32-spmm/gen/8x1-minmax-wasmsimd-arm.c",
Frank Barchard8ef44cd2020-11-03 12:30:23 -08001126 "src/f32-spmm/gen/8x1-minmax-wasmsimd-x86-pipelined-x2.c",
1127 "src/f32-spmm/gen/8x1-minmax-wasmsimd-x86-pipelined.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07001128 "src/f32-spmm/gen/8x1-minmax-wasmsimd-x86-x2.c",
1129 "src/f32-spmm/gen/8x1-minmax-wasmsimd-x86-x4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001130 "src/f32-spmm/gen/8x1-minmax-wasmsimd-x86.c",
Frank Barchard8ef44cd2020-11-03 12:30:23 -08001131 "src/f32-spmm/gen/16x1-minmax-wasmsimd-arm-pipelined-x2.c",
1132 "src/f32-spmm/gen/16x1-minmax-wasmsimd-arm-pipelined.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07001133 "src/f32-spmm/gen/16x1-minmax-wasmsimd-arm-x2.c",
1134 "src/f32-spmm/gen/16x1-minmax-wasmsimd-arm-x4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001135 "src/f32-spmm/gen/16x1-minmax-wasmsimd-arm.c",
Frank Barchard8ef44cd2020-11-03 12:30:23 -08001136 "src/f32-spmm/gen/16x1-minmax-wasmsimd-x86-pipelined-x2.c",
1137 "src/f32-spmm/gen/16x1-minmax-wasmsimd-x86-pipelined.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07001138 "src/f32-spmm/gen/16x1-minmax-wasmsimd-x86-x2.c",
1139 "src/f32-spmm/gen/16x1-minmax-wasmsimd-x86-x4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001140 "src/f32-spmm/gen/16x1-minmax-wasmsimd-x86.c",
Frank Barchard8ef44cd2020-11-03 12:30:23 -08001141 "src/f32-spmm/gen/32x1-minmax-wasmsimd-arm-pipelined-x2.c",
1142 "src/f32-spmm/gen/32x1-minmax-wasmsimd-arm-pipelined.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07001143 "src/f32-spmm/gen/32x1-minmax-wasmsimd-arm-x2.c",
1144 "src/f32-spmm/gen/32x1-minmax-wasmsimd-arm-x4.c",
Frank Barchard846c0c62020-10-26 15:01:39 -07001145 "src/f32-spmm/gen/32x1-minmax-wasmsimd-arm.c",
Frank Barchard8ef44cd2020-11-03 12:30:23 -08001146 "src/f32-spmm/gen/32x1-minmax-wasmsimd-x86-pipelined-x2.c",
1147 "src/f32-spmm/gen/32x1-minmax-wasmsimd-x86-pipelined.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07001148 "src/f32-spmm/gen/32x1-minmax-wasmsimd-x86-x2.c",
1149 "src/f32-spmm/gen/32x1-minmax-wasmsimd-x86-x4.c",
Frank Barchard846c0c62020-10-26 15:01:39 -07001150 "src/f32-spmm/gen/32x1-minmax-wasmsimd-x86.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -07001151 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-arm-x4.c",
1152 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001153 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-arm-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -07001154 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-x86-x4.c",
1155 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001156 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -07001157 "src/f32-vbinary/gen/vadd-relu-wasmsimd-x4.c",
1158 "src/f32-vbinary/gen/vadd-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001159 "src/f32-vbinary/gen/vadd-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07001160 "src/f32-vbinary/gen/vadd-wasmsimd-x4.c",
1161 "src/f32-vbinary/gen/vadd-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001162 "src/f32-vbinary/gen/vadd-wasmsimd-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001163 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-arm-x4.c",
1164 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001165 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-arm-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001166 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-x86-x4.c",
1167 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001168 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -07001169 "src/f32-vbinary/gen/vaddc-relu-wasmsimd-x4.c",
1170 "src/f32-vbinary/gen/vaddc-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001171 "src/f32-vbinary/gen/vaddc-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07001172 "src/f32-vbinary/gen/vaddc-wasmsimd-x4.c",
1173 "src/f32-vbinary/gen/vaddc-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001174 "src/f32-vbinary/gen/vaddc-wasmsimd-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -07001175 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-arm-x4.c",
1176 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001177 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-arm-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -07001178 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-x86-x4.c",
1179 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001180 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -07001181 "src/f32-vbinary/gen/vdiv-relu-wasmsimd-x4.c",
1182 "src/f32-vbinary/gen/vdiv-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001183 "src/f32-vbinary/gen/vdiv-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07001184 "src/f32-vbinary/gen/vdiv-wasmsimd-x4.c",
1185 "src/f32-vbinary/gen/vdiv-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001186 "src/f32-vbinary/gen/vdiv-wasmsimd-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001187 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-arm-x4.c",
1188 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001189 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-arm-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001190 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-x86-x4.c",
1191 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001192 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -07001193 "src/f32-vbinary/gen/vdivc-relu-wasmsimd-x4.c",
1194 "src/f32-vbinary/gen/vdivc-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001195 "src/f32-vbinary/gen/vdivc-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07001196 "src/f32-vbinary/gen/vdivc-wasmsimd-x4.c",
1197 "src/f32-vbinary/gen/vdivc-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001198 "src/f32-vbinary/gen/vdivc-wasmsimd-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -07001199 "src/f32-vbinary/gen/vmax-wasmsimd-arm-x4.c",
1200 "src/f32-vbinary/gen/vmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001201 "src/f32-vbinary/gen/vmax-wasmsimd-arm-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -07001202 "src/f32-vbinary/gen/vmax-wasmsimd-x86-x4.c",
1203 "src/f32-vbinary/gen/vmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001204 "src/f32-vbinary/gen/vmax-wasmsimd-x86-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001205 "src/f32-vbinary/gen/vmaxc-wasmsimd-arm-x4.c",
1206 "src/f32-vbinary/gen/vmaxc-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001207 "src/f32-vbinary/gen/vmaxc-wasmsimd-arm-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001208 "src/f32-vbinary/gen/vmaxc-wasmsimd-x86-x4.c",
1209 "src/f32-vbinary/gen/vmaxc-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001210 "src/f32-vbinary/gen/vmaxc-wasmsimd-x86-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -07001211 "src/f32-vbinary/gen/vmin-wasmsimd-arm-x4.c",
1212 "src/f32-vbinary/gen/vmin-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001213 "src/f32-vbinary/gen/vmin-wasmsimd-arm-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -07001214 "src/f32-vbinary/gen/vmin-wasmsimd-x86-x4.c",
1215 "src/f32-vbinary/gen/vmin-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001216 "src/f32-vbinary/gen/vmin-wasmsimd-x86-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001217 "src/f32-vbinary/gen/vminc-wasmsimd-arm-x4.c",
1218 "src/f32-vbinary/gen/vminc-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001219 "src/f32-vbinary/gen/vminc-wasmsimd-arm-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001220 "src/f32-vbinary/gen/vminc-wasmsimd-x86-x4.c",
1221 "src/f32-vbinary/gen/vminc-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001222 "src/f32-vbinary/gen/vminc-wasmsimd-x86-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -07001223 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-arm-x4.c",
1224 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001225 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-arm-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -07001226 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-x86-x4.c",
1227 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001228 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -07001229 "src/f32-vbinary/gen/vmul-relu-wasmsimd-x4.c",
1230 "src/f32-vbinary/gen/vmul-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001231 "src/f32-vbinary/gen/vmul-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07001232 "src/f32-vbinary/gen/vmul-wasmsimd-x4.c",
1233 "src/f32-vbinary/gen/vmul-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001234 "src/f32-vbinary/gen/vmul-wasmsimd-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001235 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-arm-x4.c",
1236 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001237 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-arm-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001238 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-x86-x4.c",
1239 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001240 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -07001241 "src/f32-vbinary/gen/vmulc-relu-wasmsimd-x4.c",
1242 "src/f32-vbinary/gen/vmulc-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001243 "src/f32-vbinary/gen/vmulc-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07001244 "src/f32-vbinary/gen/vmulc-wasmsimd-x4.c",
1245 "src/f32-vbinary/gen/vmulc-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001246 "src/f32-vbinary/gen/vmulc-wasmsimd-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001247 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-arm-x4.c",
1248 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001249 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-arm-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001250 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-x86-x4.c",
1251 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001252 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -07001253 "src/f32-vbinary/gen/vrdivc-relu-wasmsimd-x4.c",
1254 "src/f32-vbinary/gen/vrdivc-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001255 "src/f32-vbinary/gen/vrdivc-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07001256 "src/f32-vbinary/gen/vrdivc-wasmsimd-x4.c",
1257 "src/f32-vbinary/gen/vrdivc-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001258 "src/f32-vbinary/gen/vrdivc-wasmsimd-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001259 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-arm-x4.c",
1260 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001261 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-arm-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001262 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-x86-x4.c",
1263 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001264 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -07001265 "src/f32-vbinary/gen/vrsubc-relu-wasmsimd-x4.c",
1266 "src/f32-vbinary/gen/vrsubc-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001267 "src/f32-vbinary/gen/vrsubc-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07001268 "src/f32-vbinary/gen/vrsubc-wasmsimd-x4.c",
1269 "src/f32-vbinary/gen/vrsubc-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001270 "src/f32-vbinary/gen/vrsubc-wasmsimd-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -07001271 "src/f32-vbinary/gen/vsqrdiff-wasmsimd-x4.c",
1272 "src/f32-vbinary/gen/vsqrdiff-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001273 "src/f32-vbinary/gen/vsqrdiff-wasmsimd-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001274 "src/f32-vbinary/gen/vsqrdiffc-wasmsimd-x4.c",
1275 "src/f32-vbinary/gen/vsqrdiffc-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001276 "src/f32-vbinary/gen/vsqrdiffc-wasmsimd-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -07001277 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-arm-x4.c",
1278 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001279 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-arm-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -07001280 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-x86-x4.c",
1281 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001282 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -07001283 "src/f32-vbinary/gen/vsub-relu-wasmsimd-x4.c",
1284 "src/f32-vbinary/gen/vsub-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001285 "src/f32-vbinary/gen/vsub-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07001286 "src/f32-vbinary/gen/vsub-wasmsimd-x4.c",
1287 "src/f32-vbinary/gen/vsub-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001288 "src/f32-vbinary/gen/vsub-wasmsimd-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001289 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-arm-x4.c",
1290 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001291 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-arm-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001292 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-x86-x4.c",
1293 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001294 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -07001295 "src/f32-vbinary/gen/vsubc-relu-wasmsimd-x4.c",
1296 "src/f32-vbinary/gen/vsubc-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001297 "src/f32-vbinary/gen/vsubc-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07001298 "src/f32-vbinary/gen/vsubc-wasmsimd-x4.c",
1299 "src/f32-vbinary/gen/vsubc-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001300 "src/f32-vbinary/gen/vsubc-wasmsimd-x16.c",
Marat Dukhaned6baaf2020-12-01 15:07:08 -08001301 "src/f32-velu/gen/velu-wasmsimd-arm-rr2-lut16-p3-x4.c",
1302 "src/f32-velu/gen/velu-wasmsimd-arm-rr2-lut16-p3-x8.c",
1303 "src/f32-velu/gen/velu-wasmsimd-arm-rr2-lut16-p3-x12.c",
1304 "src/f32-velu/gen/velu-wasmsimd-arm-rr2-lut16-p3-x16.c",
1305 "src/f32-velu/gen/velu-wasmsimd-arm-rr2-lut16-p3-x20.c",
1306 "src/f32-velu/gen/velu-wasmsimd-arm-rr2-lut16-p3-x24.c",
Marat Dukhaned6baaf2020-12-01 15:07:08 -08001307 "src/f32-velu/gen/velu-wasmsimd-arm-rr2-p6-x4.c",
1308 "src/f32-velu/gen/velu-wasmsimd-arm-rr2-p6-x8.c",
1309 "src/f32-velu/gen/velu-wasmsimd-arm-rr2-p6-x12.c",
1310 "src/f32-velu/gen/velu-wasmsimd-arm-rr2-p6-x16.c",
1311 "src/f32-velu/gen/velu-wasmsimd-arm-rr2-p6-x20.c",
1312 "src/f32-velu/gen/velu-wasmsimd-arm-rr2-p6-x24.c",
Frank Barcharde7223ee2020-12-04 19:04:01 -08001313 "src/f32-velu/gen/velu-wasmsimd-x86-rr2-lut16-p3-x4.c",
1314 "src/f32-velu/gen/velu-wasmsimd-x86-rr2-lut16-p3-x8.c",
1315 "src/f32-velu/gen/velu-wasmsimd-x86-rr2-lut16-p3-x12.c",
1316 "src/f32-velu/gen/velu-wasmsimd-x86-rr2-lut16-p3-x16.c",
1317 "src/f32-velu/gen/velu-wasmsimd-x86-rr2-lut16-p3-x20.c",
1318 "src/f32-velu/gen/velu-wasmsimd-x86-rr2-lut16-p3-x24.c",
Marat Dukhaned6baaf2020-12-01 15:07:08 -08001319 "src/f32-velu/gen/velu-wasmsimd-x86-rr2-p6-x4.c",
1320 "src/f32-velu/gen/velu-wasmsimd-x86-rr2-p6-x8.c",
1321 "src/f32-velu/gen/velu-wasmsimd-x86-rr2-p6-x12.c",
1322 "src/f32-velu/gen/velu-wasmsimd-x86-rr2-p6-x16.c",
1323 "src/f32-velu/gen/velu-wasmsimd-x86-rr2-p6-x20.c",
1324 "src/f32-velu/gen/velu-wasmsimd-x86-rr2-p6-x24.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07001325 "src/f32-vlrelu/gen/vlrelu-wasmsimd-bitselect-x4.c",
1326 "src/f32-vlrelu/gen/vlrelu-wasmsimd-bitselect-x8.c",
1327 "src/f32-vlrelu/gen/vlrelu-wasmsimd-minmax-x4.c",
1328 "src/f32-vlrelu/gen/vlrelu-wasmsimd-minmax-x8.c",
Marat Dukhand816f622020-07-15 10:14:39 -07001329 "src/f32-vmulcaddc/gen/c4-minmax-wasmsimd-arm-2x.c",
Marat Dukhand816f622020-07-15 10:14:39 -07001330 "src/f32-vmulcaddc/gen/c4-minmax-wasmsimd-x86-2x.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001331 "src/f32-vmulcaddc/gen/c8-minmax-wasmsimd-arm-2x.c",
Marat Dukhand816f622020-07-15 10:14:39 -07001332 "src/f32-vmulcaddc/gen/c8-minmax-wasmsimd-x86-2x.c",
Marat Dukhanb82b2cd2020-07-16 02:23:42 -07001333 "src/f32-vrnd/gen/vrndd-wasmsimd-addsub-x4.c",
1334 "src/f32-vrnd/gen/vrndd-wasmsimd-addsub-x8.c",
1335 "src/f32-vrnd/gen/vrndd-wasmsimd-cvt-x4.c",
1336 "src/f32-vrnd/gen/vrndd-wasmsimd-cvt-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001337 "src/f32-vrnd/gen/vrndne-wasmsimd-addsub-x4.c",
1338 "src/f32-vrnd/gen/vrndne-wasmsimd-addsub-x8.c",
1339 "src/f32-vrnd/gen/vrndu-wasmsimd-addsub-x4.c",
1340 "src/f32-vrnd/gen/vrndu-wasmsimd-addsub-x8.c",
1341 "src/f32-vrnd/gen/vrndu-wasmsimd-cvt-x4.c",
1342 "src/f32-vrnd/gen/vrndu-wasmsimd-cvt-x8.c",
1343 "src/f32-vrnd/gen/vrndz-wasmsimd-addsub-x4.c",
1344 "src/f32-vrnd/gen/vrndz-wasmsimd-addsub-x8.c",
1345 "src/f32-vrnd/gen/vrndz-wasmsimd-cvt-x4.c",
1346 "src/f32-vrnd/gen/vrndz-wasmsimd-cvt-x8.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07001347 "src/f32-vsqrt/gen/wasmsimd-sqrt-x4.c",
1348 "src/f32-vsqrt/gen/wasmsimd-sqrt-x8.c",
Marat Dukhan37c83512020-06-29 13:25:53 -07001349 "src/f32-vunary/gen/vabs-wasmsimd-x4.c",
1350 "src/f32-vunary/gen/vabs-wasmsimd-x8.c",
1351 "src/f32-vunary/gen/vneg-wasmsimd-x4.c",
1352 "src/f32-vunary/gen/vneg-wasmsimd-x8.c",
1353 "src/f32-vunary/gen/vsqr-wasmsimd-x4.c",
1354 "src/f32-vunary/gen/vsqr-wasmsimd-x8.c",
Marat Dukhande390d42020-11-29 19:32:18 -08001355 "src/math/expm1minus-wasmsimd-rr2-lut16-p3-andnot.c",
1356 "src/math/expm1minus-wasmsimd-rr2-lut16-p3-max.c",
1357 "src/math/expm1minus-wasmsimd-rr2-p6-andnot.c",
1358 "src/math/expm1minus-wasmsimd-rr2-p6-max.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001359 "src/math/roundd-wasmsimd-addsub.c",
1360 "src/math/roundd-wasmsimd-cvt.c",
1361 "src/math/roundne-wasmsimd-addsub.c",
1362 "src/math/roundu-wasmsimd-addsub.c",
1363 "src/math/roundu-wasmsimd-cvt.c",
1364 "src/math/roundz-wasmsimd-addsub.c",
1365 "src/math/roundz-wasmsimd-cvt.c",
1366 "src/math/sigmoid-wasmsimd-rr2-lut64-p2-div.c",
1367 "src/math/sigmoid-wasmsimd-rr2-p5-div.c",
Marat Dukhancc8f34c2020-08-05 16:36:38 -07001368 "src/qs8-dwconv/gen/up8x9-minmax-wasmsimd-mul16.c",
1369 "src/qs8-dwconv/gen/up16x9-minmax-wasmsimd-mul16.c",
1370 "src/qs8-dwconv/gen/up24x9-minmax-wasmsimd-mul16.c",
Marat Dukhanb5e3d172020-08-06 13:29:53 -07001371 "src/qs8-gavgpool/gen/7p7x-minmax-wasmsimd-c8-acc2.c",
1372 "src/qs8-gavgpool/gen/7p7x-minmax-wasmsimd-c16-acc2.c",
1373 "src/qs8-gavgpool/gen/7p7x-minmax-wasmsimd-c24-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001374 "src/qs8-gavgpool/gen/7x-minmax-wasmsimd-c8-acc2.c",
1375 "src/qs8-gavgpool/gen/7x-minmax-wasmsimd-c16-acc2.c",
1376 "src/qs8-gavgpool/gen/7x-minmax-wasmsimd-c24-acc2.c",
Marat Dukhan27203da2020-08-05 15:19:03 -07001377 "src/qs8-gemm/gen/1x4c8-minmax-wasmsimd-ld64.c",
Marat Dukhan27203da2020-08-05 15:19:03 -07001378 "src/qs8-gemm/gen/1x4c8-minmax-wasmsimd-ld128.c",
Marat Dukhan27203da2020-08-05 15:19:03 -07001379 "src/qs8-gemm/gen/1x4c8-xw-minmax-wasmsimd.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001380 "src/qs8-gemm/gen/2x4c8-minmax-wasmsimd-ld64.c",
1381 "src/qs8-gemm/gen/2x4c8-minmax-wasmsimd-ld128.c",
Marat Dukhan27203da2020-08-05 15:19:03 -07001382 "src/qs8-gemm/gen/2x4c8-xw-minmax-wasmsimd.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001383 "src/qs8-gemm/gen/3x4c8-minmax-wasmsimd-ld64.c",
1384 "src/qs8-gemm/gen/3x4c8-minmax-wasmsimd-ld128.c",
Marat Dukhan27203da2020-08-05 15:19:03 -07001385 "src/qs8-gemm/gen/3x4c8-xw-minmax-wasmsimd.c",
1386 "src/qs8-igemm/gen/1x4c8-minmax-wasmsimd-ld64.c",
Marat Dukhan27203da2020-08-05 15:19:03 -07001387 "src/qs8-igemm/gen/1x4c8-minmax-wasmsimd-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001388 "src/qs8-igemm/gen/2x4c8-minmax-wasmsimd-ld64.c",
Marat Dukhan27203da2020-08-05 15:19:03 -07001389 "src/qs8-igemm/gen/2x4c8-minmax-wasmsimd-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001390 "src/qs8-igemm/gen/3x4c8-minmax-wasmsimd-ld64.c",
Marat Dukhan27203da2020-08-05 15:19:03 -07001391 "src/qs8-igemm/gen/3x4c8-minmax-wasmsimd-ld128.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -07001392 "src/qs8-requantization/fp32-wasmsimd.c",
Marat Dukhan138560c2020-08-03 18:57:34 -07001393 "src/qs8-requantization/q31-wasmsimd.c",
Marat Dukhan5df27f82020-09-02 23:59:21 -07001394 "src/qs8-vadd/gen/minmax-wasmsimd-x8.c",
1395 "src/qs8-vadd/gen/minmax-wasmsimd-x16.c",
1396 "src/qs8-vadd/gen/minmax-wasmsimd-x24.c",
1397 "src/qs8-vadd/gen/minmax-wasmsimd-x32.c",
1398 "src/qs8-vaddc/gen/minmax-wasmsimd-x8.c",
1399 "src/qs8-vaddc/gen/minmax-wasmsimd-x16.c",
1400 "src/qs8-vaddc/gen/minmax-wasmsimd-x24.c",
1401 "src/qs8-vaddc/gen/minmax-wasmsimd-x32.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07001402 "src/qu8-requantization/fp32-wasmsimd.c",
Marat Dukhan22de5e72020-08-03 21:53:18 -07001403 "src/qu8-requantization/q31-wasmsimd.c",
Marat Dukhan8ee37012020-07-16 13:17:13 -07001404 "src/x32-fill/wasmsimd.c",
Marat Dukhan66d99e92020-07-16 12:56:21 -07001405 "src/x32-packx/x4-wasmsimd.c",
Marat Dukhan9306ae02020-07-16 15:51:13 -07001406 "src/x32-pad/wasmsimd.c",
Marat Dukhan9d4bfa22020-07-16 19:07:04 -07001407 "src/x32-unpool/wasmsimd.c",
Marat Dukhane3b78762020-07-16 20:02:58 -07001408 "src/x32-zip/x2-wasmsimd.c",
1409 "src/x32-zip/x3-wasmsimd.c",
1410 "src/x32-zip/x4-wasmsimd.c",
1411 "src/x32-zip/xm-wasmsimd.c",
Marat Dukhan290055c2020-06-09 12:24:29 -07001412]
1413
Marat Dukhan08c4a432019-10-03 09:29:21 -07001414# ISA-specific micro-kernels
1415NEON_UKERNELS = [
Marat Dukhanef25c6d2020-07-24 00:59:40 -07001416 "src/f32-argmaxpool/4x-neon-c4.c",
1417 "src/f32-argmaxpool/9p8x-neon-c4.c",
1418 "src/f32-argmaxpool/9x-neon-c4.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001419 "src/f32-avgpool/9p8x-minmax-neon-c4.c",
1420 "src/f32-avgpool/9x-minmax-neon-c4.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001421 "src/f32-clamp/gen/neon-x4.c",
1422 "src/f32-clamp/gen/neon-x8.c",
Marat Dukhan56b10cd2020-05-18 09:35:49 -07001423 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neon-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001424 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neon-2x2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001425 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neon-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001426 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neon-2x2.c",
Marat Dukhan56b10cd2020-05-18 09:35:49 -07001427 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neon-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001428 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neon-2x2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001429 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neon-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001430 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neon-2x2.c",
Marat Dukhanc7634882020-12-07 15:11:12 -08001431 "src/f32-conv-hwc2chw/3x3s2p1c3x4-neon-2x2.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -07001432 "src/f32-dwconv/gen/up4x4-minmax-neon-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001433 "src/f32-dwconv/gen/up4x4-minmax-neon.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001434 "src/f32-dwconv/gen/up4x9-minmax-neon-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001435 "src/f32-dwconv/gen/up4x9-minmax-neon.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -07001436 "src/f32-dwconv/gen/up4x25-minmax-neon-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001437 "src/f32-dwconv/gen/up4x25-minmax-neon.c",
1438 "src/f32-dwconv/gen/up8x4-minmax-neon-acc2.c",
1439 "src/f32-dwconv/gen/up8x4-minmax-neon.c",
1440 "src/f32-dwconv/gen/up8x9-minmax-neon-acc2.c",
1441 "src/f32-dwconv/gen/up8x9-minmax-neon.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -07001442 "src/f32-dwconv/gen/up8x25-minmax-neon-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001443 "src/f32-dwconv/gen/up8x25-minmax-neon.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001444 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-1x4-acc2.c",
1445 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-1x4-acc3.c",
1446 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-1x4-acc4.c",
Marat Dukhanc581e482020-10-24 01:28:11 -07001447 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-1x4.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001448 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-2x4-acc2.c",
Marat Dukhanc581e482020-10-24 01:28:11 -07001449 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-2x4.c",
1450 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-3x4.c",
1451 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-4x4.c",
1452 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-5x4.c",
1453 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-6x4.c",
Marat Dukhan82f0c322020-10-25 19:17:35 -07001454 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neon-1x4-acc2.c",
1455 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neon-1x4-acc3.c",
1456 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neon-1x4-acc4.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001457 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neon-1x4.c",
Marat Dukhan82f0c322020-10-25 19:17:35 -07001458 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neon-2x4-acc2.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001459 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neon-2x4.c",
1460 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neon-3x4.c",
1461 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neon-4x4.c",
Marat Dukhan149f0ea2020-10-26 12:50:33 -07001462 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-1x4-acc2.c",
1463 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-1x4-acc3.c",
1464 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-1x4-acc4.c",
1465 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-1x4-acc5.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001466 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-1x4.c",
Marat Dukhan149f0ea2020-10-26 12:50:33 -07001467 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-2x4-acc2.c",
1468 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-2x4-acc3.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001469 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-2x4.c",
Marat Dukhan149f0ea2020-10-26 12:50:33 -07001470 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-3x4-acc2.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001471 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-3x4.c",
Marat Dukhan149f0ea2020-10-26 12:50:33 -07001472 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-4x4-acc2.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001473 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-4x4.c",
1474 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-5x4.c",
Marat Dukhan30d4b252020-10-29 16:33:22 -07001475 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neon-1x4-acc2.c",
1476 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neon-1x4-acc3.c",
1477 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neon-1x4-acc4.c",
1478 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neon-1x4-acc5.c",
1479 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neon-1x4.c",
1480 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neon-2x4-acc2.c",
1481 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neon-2x4-acc3.c",
1482 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neon-2x4.c",
Marat Dukhan30d4b252020-10-29 16:33:22 -07001483 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neon-3x4-acc2.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07001484 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neon-3x4.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -07001485 "src/f32-gavgpool-cw/neon-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001486 "src/f32-gavgpool/7p7x-minmax-neon-c4.c",
1487 "src/f32-gavgpool/7x-minmax-neon-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001488 "src/f32-gemm/gen-inc/1x8inc-minmax-neon-dup-ld64.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001489 "src/f32-gemm/gen-inc/1x8inc-minmax-neon-lane-ld64.c",
1490 "src/f32-gemm/gen-inc/1x8s4inc-minmax-neon.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001491 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-dup-ld64.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001492 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-dup-ld128.c",
1493 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-lane-ld64.c",
1494 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-lane-ld128.c",
1495 "src/f32-gemm/gen-inc/4x8s4inc-minmax-neon.c",
1496 "src/f32-gemm/gen-inc/5x8inc-minmax-neon-lane-ld64.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001497 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-dup-ld64.c",
1498 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-dup-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001499 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-lane-ld64.c",
1500 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-lane-ld128.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001501 "src/f32-gemm/gen-inc/6x8s4inc-minmax-neon.c",
1502 "src/f32-gemm/gen-inc/8x8s4inc-minmax-neon.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001503 "src/f32-gemm/gen/1x8-minmax-neon-dup-ld64.c",
1504 "src/f32-gemm/gen/1x8-minmax-neon-lane-ld64.c",
1505 "src/f32-gemm/gen/1x8s4-minmax-neon.c",
1506 "src/f32-gemm/gen/4x2-minmax-neon-lane-ld64.c",
1507 "src/f32-gemm/gen/4x8-minmax-neon-dup-ld64.c",
1508 "src/f32-gemm/gen/4x8-minmax-neon-dup-ld128.c",
1509 "src/f32-gemm/gen/4x8-minmax-neon-lane-ld64.c",
1510 "src/f32-gemm/gen/4x8-minmax-neon-lane-ld128.c",
1511 "src/f32-gemm/gen/4x8s4-minmax-neon.c",
1512 "src/f32-gemm/gen/5x8-minmax-neon-lane-ld64.c",
1513 "src/f32-gemm/gen/6x8-minmax-neon-dup-ld64.c",
1514 "src/f32-gemm/gen/6x8-minmax-neon-dup-ld128.c",
1515 "src/f32-gemm/gen/6x8-minmax-neon-lane-ld64.c",
1516 "src/f32-gemm/gen/6x8-minmax-neon-lane-ld128.c",
1517 "src/f32-gemm/gen/6x8s4-minmax-neon.c",
1518 "src/f32-gemm/gen/8x8s4-minmax-neon.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07001519 "src/f32-hswish/gen/hswish-neon-x4.c",
1520 "src/f32-hswish/gen/hswish-neon-x8.c",
1521 "src/f32-hswish/gen/hswish-neon-x16.c",
Artsiom Ablavatski2202c812021-01-22 14:16:43 -08001522 "src/f32-ibilinear-chw/gen/neon-p4.c",
1523 "src/f32-ibilinear-chw/gen/neon-p8.c",
Frank Barchard8247e212021-02-03 18:12:33 -08001524 "src/f32-ibilinear/gen/neon-c4.c",
1525 "src/f32-ibilinear/gen/neon-c8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001526 "src/f32-igemm/gen/1x8-minmax-neon-dup-ld64.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001527 "src/f32-igemm/gen/1x8-minmax-neon-lane-ld64.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001528 "src/f32-igemm/gen/1x8s4-minmax-neon.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001529 "src/f32-igemm/gen/4x2-minmax-neon-lane-ld64.c",
1530 "src/f32-igemm/gen/4x4-minmax-neon-lane-ld64.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001531 "src/f32-igemm/gen/4x8-minmax-neon-dup-ld64.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001532 "src/f32-igemm/gen/4x8-minmax-neon-dup-ld128.c",
1533 "src/f32-igemm/gen/4x8-minmax-neon-lane-ld64.c",
1534 "src/f32-igemm/gen/4x8-minmax-neon-lane-ld128.c",
1535 "src/f32-igemm/gen/4x8s4-minmax-neon.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001536 "src/f32-igemm/gen/6x8-minmax-neon-dup-ld64.c",
1537 "src/f32-igemm/gen/6x8-minmax-neon-dup-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001538 "src/f32-igemm/gen/6x8-minmax-neon-lane-ld64.c",
1539 "src/f32-igemm/gen/6x8-minmax-neon-lane-ld128.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001540 "src/f32-igemm/gen/6x8s4-minmax-neon.c",
1541 "src/f32-igemm/gen/8x8s4-minmax-neon.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001542 "src/f32-maxpool/9p8x-minmax-neon-c4.c",
1543 "src/f32-pavgpool/9p8x-minmax-neon-c4.c",
1544 "src/f32-pavgpool/9x-minmax-neon-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001545 "src/f32-ppmm/gen/4x8-minmax-neon.c",
1546 "src/f32-ppmm/gen/8x8-minmax-neon.c",
Frank Barcharda5316982020-07-23 13:19:28 -07001547 "src/f32-prelu/gen/neon-1x4.c",
1548 "src/f32-prelu/gen/neon-1x8.c",
1549 "src/f32-prelu/gen/neon-1x16.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001550 "src/f32-prelu/gen/neon-2x4.c",
1551 "src/f32-prelu/gen/neon-2x8.c",
Frank Barcharda5316982020-07-23 13:19:28 -07001552 "src/f32-prelu/gen/neon-2x16.c",
1553 "src/f32-prelu/gen/neon-4x4.c",
1554 "src/f32-prelu/gen/neon-4x8.c",
1555 "src/f32-prelu/gen/neon-4x16.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001556 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x4.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001557 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x8-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001558 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x8.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001559 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12-acc2.c",
1560 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001561 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001562 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16-acc2.c",
1563 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001564 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001565 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20-acc2.c",
1566 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001567 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20.c",
1568 "src/f32-raddstoreexpminusmax/gen/neon-p5-x4.c",
1569 "src/f32-raddstoreexpminusmax/gen/neon-p5-x8-acc2.c",
1570 "src/f32-raddstoreexpminusmax/gen/neon-p5-x8.c",
1571 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12-acc2.c",
1572 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12-acc3.c",
1573 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12.c",
1574 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16-acc2.c",
1575 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16-acc4.c",
1576 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16.c",
1577 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20-acc2.c",
1578 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20-acc5.c",
1579 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07001580 "src/f32-relu/gen/neon-x4.c",
1581 "src/f32-relu/gen/neon-x8.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001582 "src/f32-rmax/neon.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -08001583 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x4.c",
1584 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x8.c",
1585 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x12.c",
1586 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x16.c",
1587 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x20.c",
1588 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x24.c",
1589 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x4.c",
1590 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x8.c",
1591 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x12.c",
1592 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x16.c",
1593 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x20.c",
1594 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x24.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001595 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x4.c",
1596 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x8.c",
1597 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x12.c",
1598 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x16.c",
1599 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x20.c",
1600 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x24.c",
Marat Dukhan5b86c432020-12-06 19:15:03 -08001601 "src/f32-spmm/gen/4x1-minmax-neon-pipelined.c",
1602 "src/f32-spmm/gen/4x1-minmax-neon-x2.c",
1603 "src/f32-spmm/gen/4x1-minmax-neon.c",
1604 "src/f32-spmm/gen/8x1-minmax-neon-pipelined.c",
1605 "src/f32-spmm/gen/8x1-minmax-neon-x2.c",
1606 "src/f32-spmm/gen/8x1-minmax-neon.c",
1607 "src/f32-spmm/gen/12x1-minmax-neon.c",
1608 "src/f32-spmm/gen/16x1-minmax-neon-pipelined.c",
1609 "src/f32-spmm/gen/16x1-minmax-neon-x2.c",
1610 "src/f32-spmm/gen/16x1-minmax-neon.c",
1611 "src/f32-spmm/gen/32x1-minmax-neon-pipelined.c",
1612 "src/f32-spmm/gen/32x1-minmax-neon-x2.c",
1613 "src/f32-spmm/gen/32x1-minmax-neon.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001614 "src/f32-vbinary/gen/vadd-minmax-neon-x4.c",
1615 "src/f32-vbinary/gen/vadd-minmax-neon-x8.c",
1616 "src/f32-vbinary/gen/vaddc-minmax-neon-x4.c",
1617 "src/f32-vbinary/gen/vaddc-minmax-neon-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -08001618 "src/f32-vbinary/gen/vmax-neon-x4.c",
1619 "src/f32-vbinary/gen/vmax-neon-x8.c",
1620 "src/f32-vbinary/gen/vmaxc-neon-x4.c",
1621 "src/f32-vbinary/gen/vmaxc-neon-x8.c",
1622 "src/f32-vbinary/gen/vmin-neon-x4.c",
1623 "src/f32-vbinary/gen/vmin-neon-x8.c",
1624 "src/f32-vbinary/gen/vminc-neon-x4.c",
1625 "src/f32-vbinary/gen/vminc-neon-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001626 "src/f32-vbinary/gen/vmul-minmax-neon-x4.c",
1627 "src/f32-vbinary/gen/vmul-minmax-neon-x8.c",
1628 "src/f32-vbinary/gen/vmulc-minmax-neon-x4.c",
1629 "src/f32-vbinary/gen/vmulc-minmax-neon-x8.c",
1630 "src/f32-vbinary/gen/vrsubc-minmax-neon-x4.c",
1631 "src/f32-vbinary/gen/vrsubc-minmax-neon-x8.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07001632 "src/f32-vbinary/gen/vsqrdiff-neon-x4.c",
1633 "src/f32-vbinary/gen/vsqrdiff-neon-x8.c",
1634 "src/f32-vbinary/gen/vsqrdiffc-neon-x4.c",
1635 "src/f32-vbinary/gen/vsqrdiffc-neon-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001636 "src/f32-vbinary/gen/vsub-minmax-neon-x4.c",
1637 "src/f32-vbinary/gen/vsub-minmax-neon-x8.c",
1638 "src/f32-vbinary/gen/vsubc-minmax-neon-x4.c",
1639 "src/f32-vbinary/gen/vsubc-minmax-neon-x8.c",
Marat Dukhaned6baaf2020-12-01 15:07:08 -08001640 "src/f32-velu/gen/velu-neon-rr2-lut16-p3-x4.c",
1641 "src/f32-velu/gen/velu-neon-rr2-lut16-p3-x8.c",
1642 "src/f32-velu/gen/velu-neon-rr2-lut16-p3-x12.c",
1643 "src/f32-velu/gen/velu-neon-rr2-lut16-p3-x16.c",
1644 "src/f32-velu/gen/velu-neon-rr2-lut16-p3-x20.c",
1645 "src/f32-velu/gen/velu-neon-rr2-lut16-p3-x24.c",
1646 "src/f32-velu/gen/velu-neon-rr2-p6-x4.c",
1647 "src/f32-velu/gen/velu-neon-rr2-p6-x8.c",
1648 "src/f32-velu/gen/velu-neon-rr2-p6-x12.c",
1649 "src/f32-velu/gen/velu-neon-rr2-p6-x16.c",
1650 "src/f32-velu/gen/velu-neon-rr2-p6-x20.c",
1651 "src/f32-velu/gen/velu-neon-rr2-p6-x24.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07001652 "src/f32-vlrelu/gen/vlrelu-neon-x4.c",
1653 "src/f32-vlrelu/gen/vlrelu-neon-x8.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001654 "src/f32-vmulcaddc/gen/c4-minmax-neon-2x.c",
1655 "src/f32-vmulcaddc/gen/c8-minmax-neon-2x.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07001656 "src/f32-vrnd/gen/vrndd-neon-x4.c",
1657 "src/f32-vrnd/gen/vrndd-neon-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001658 "src/f32-vrnd/gen/vrndne-neon-x4.c",
1659 "src/f32-vrnd/gen/vrndne-neon-x8.c",
1660 "src/f32-vrnd/gen/vrndu-neon-x4.c",
1661 "src/f32-vrnd/gen/vrndu-neon-x8.c",
1662 "src/f32-vrnd/gen/vrndz-neon-x4.c",
1663 "src/f32-vrnd/gen/vrndz-neon-x8.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07001664 "src/f32-vunary/gen/vabs-neon-x4.c",
1665 "src/f32-vunary/gen/vabs-neon-x8.c",
1666 "src/f32-vunary/gen/vneg-neon-x4.c",
1667 "src/f32-vunary/gen/vneg-neon-x8.c",
1668 "src/f32-vunary/gen/vsqr-neon-x4.c",
1669 "src/f32-vunary/gen/vsqr-neon-x8.c",
Marat Dukhande390d42020-11-29 19:32:18 -08001670 "src/math/expm1minus-neon-rr2-lut16-p3.c",
1671 "src/math/expm1minus-neon-rr2-p6.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001672 "src/math/roundd-neon-addsub.c",
1673 "src/math/roundd-neon-cvt.c",
1674 "src/math/roundne-neon-addsub.c",
1675 "src/math/roundu-neon-addsub.c",
1676 "src/math/roundu-neon-cvt.c",
1677 "src/math/roundz-neon-addsub.c",
1678 "src/math/roundz-neon-cvt.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001679 "src/math/sigmoid-neon-rr2-lut64-p2-nr2recps.c",
1680 "src/math/sigmoid-neon-rr2-lut2048-p1-nr2recps.c",
1681 "src/math/sigmoid-neon-rr2-p5-nr2recps.c",
1682 "src/math/sqrt-neon-nr1rsqrts.c",
1683 "src/math/sqrt-neon-nr2rsqrts.c",
1684 "src/math/sqrt-neon-nr3rsqrts.c",
Marat Dukhan023bcf92020-08-10 12:40:50 -07001685 "src/qs8-dwconv/gen/up8x9-minmax-neon-mul16.c",
1686 "src/qs8-dwconv/gen/up16x9-minmax-neon-mul16.c",
1687 "src/qs8-dwconv/gen/up24x9-minmax-neon-mul16.c",
1688 "src/qs8-dwconv/gen/up32x9-minmax-neon-mul16.c",
Marat Dukhan281262d2020-08-10 13:23:21 -07001689 "src/qs8-gavgpool/gen/7p7x-minmax-neon-c8-acc2.c",
1690 "src/qs8-gavgpool/gen/7p7x-minmax-neon-c16-acc2.c",
1691 "src/qs8-gavgpool/gen/7p7x-minmax-neon-c24-acc2.c",
1692 "src/qs8-gavgpool/gen/7p7x-minmax-neon-c32-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001693 "src/qs8-gavgpool/gen/7x-minmax-neon-c8-acc2.c",
1694 "src/qs8-gavgpool/gen/7x-minmax-neon-c16-acc2.c",
1695 "src/qs8-gavgpool/gen/7x-minmax-neon-c24-acc2.c",
1696 "src/qs8-gavgpool/gen/7x-minmax-neon-c32-acc2.c",
Frank Barchardec0bf142021-01-15 10:14:05 -08001697 "src/qs8-gemm/gen/1x8-minmax-neon-mlal-lane.c",
Frank Barchardec0bf142021-01-15 10:14:05 -08001698 "src/qs8-gemm/gen/1x8-minmax-neon-mull-addw-dup.c",
Frank Barchard8247e212021-02-03 18:12:33 -08001699 "src/qs8-gemm/gen/1x8c2-minmax-neon-mlal-padal-dup.c",
Frank Barchard2302ffd2021-01-22 14:08:02 -08001700 "src/qs8-gemm/gen/1x8c2-minmax-neon-mull-padal-dup.c",
Frank Barchard8247e212021-02-03 18:12:33 -08001701 "src/qs8-gemm/gen/1x8c8-minmax-neon-mull-padal.c",
Frank Barchardcadd4222021-01-20 16:27:25 -08001702 "src/qs8-gemm/gen/1x16-minmax-neon-mlal-lane.c",
1703 "src/qs8-gemm/gen/1x16-minmax-neon-mull-addw-dup.c",
Frank Barchard8247e212021-02-03 18:12:33 -08001704 "src/qs8-gemm/gen/1x16c2-minmax-neon-mlal-padal-dup.c",
Frank Barchard2302ffd2021-01-22 14:08:02 -08001705 "src/qs8-gemm/gen/1x16c2-minmax-neon-mull-padal-dup.c",
Frank Barchard8247e212021-02-03 18:12:33 -08001706 "src/qs8-gemm/gen/1x16c8-minmax-neon-mull-padal.c",
Frank Barchardcadd4222021-01-20 16:27:25 -08001707 "src/qs8-gemm/gen/2x8-minmax-neon-mlal-lane.c",
Frank Barchardec0bf142021-01-15 10:14:05 -08001708 "src/qs8-gemm/gen/2x8-minmax-neon-mull-addw-dup.c",
Frank Barchard8247e212021-02-03 18:12:33 -08001709 "src/qs8-gemm/gen/2x8c2-minmax-neon-mlal-padal-dup.c",
Frank Barchard2302ffd2021-01-22 14:08:02 -08001710 "src/qs8-gemm/gen/2x8c2-minmax-neon-mull-padal-dup.c",
Frank Barchard8247e212021-02-03 18:12:33 -08001711 "src/qs8-gemm/gen/2x8c8-minmax-neon-mull-padal.c",
Frank Barchardcadd4222021-01-20 16:27:25 -08001712 "src/qs8-gemm/gen/2x16-minmax-neon-mlal-lane.c",
1713 "src/qs8-gemm/gen/2x16-minmax-neon-mull-addw-dup.c",
Frank Barchard8247e212021-02-03 18:12:33 -08001714 "src/qs8-gemm/gen/2x16c2-minmax-neon-mlal-padal-dup.c",
Frank Barchard2302ffd2021-01-22 14:08:02 -08001715 "src/qs8-gemm/gen/2x16c2-minmax-neon-mull-padal-dup.c",
Frank Barchard8247e212021-02-03 18:12:33 -08001716 "src/qs8-gemm/gen/2x16c8-minmax-neon-mull-padal.c",
Frank Barchard2302ffd2021-01-22 14:08:02 -08001717 "src/qs8-gemm/gen/3x8-minmax-neon-mlal-lane.c",
1718 "src/qs8-gemm/gen/3x8-minmax-neon-mull-addw-dup.c",
Frank Barchard8247e212021-02-03 18:12:33 -08001719 "src/qs8-gemm/gen/3x8c2-minmax-neon-mlal-padal-dup.c",
Frank Barchard2302ffd2021-01-22 14:08:02 -08001720 "src/qs8-gemm/gen/3x8c2-minmax-neon-mull-padal-dup.c",
Frank Barchard8247e212021-02-03 18:12:33 -08001721 "src/qs8-gemm/gen/3x8c8-minmax-neon-mull-padal.c",
1722 "src/qs8-gemm/gen/3x16-minmax-neon-mlal-lane.c",
1723 "src/qs8-gemm/gen/3x16-minmax-neon-mull-addw-dup.c",
1724 "src/qs8-gemm/gen/3x16c2-minmax-neon-mlal-padal-dup.c",
1725 "src/qs8-gemm/gen/3x16c2-minmax-neon-mull-padal-dup.c",
1726 "src/qs8-gemm/gen/3x16c8-minmax-neon-mull-padal.c",
Frank Barchardcadd4222021-01-20 16:27:25 -08001727 "src/qs8-gemm/gen/4x8-minmax-neon-mlal-lane.c",
Frank Barchardec0bf142021-01-15 10:14:05 -08001728 "src/qs8-gemm/gen/4x8-minmax-neon-mull-addw-dup.c",
Frank Barchard8247e212021-02-03 18:12:33 -08001729 "src/qs8-gemm/gen/4x8c2-minmax-neon-mlal-padal-dup.c",
Frank Barchard2302ffd2021-01-22 14:08:02 -08001730 "src/qs8-gemm/gen/4x8c2-minmax-neon-mull-padal-dup.c",
Frank Barchard8247e212021-02-03 18:12:33 -08001731 "src/qs8-gemm/gen/4x8c8-minmax-neon-mull-padal.c",
Frank Barchardcadd4222021-01-20 16:27:25 -08001732 "src/qs8-gemm/gen/4x16-minmax-neon-mlal-lane.c",
1733 "src/qs8-gemm/gen/4x16-minmax-neon-mull-addw-dup.c",
Frank Barchard8247e212021-02-03 18:12:33 -08001734 "src/qs8-gemm/gen/4x16c2-minmax-neon-mlal-padal-dup.c",
Frank Barchard2302ffd2021-01-22 14:08:02 -08001735 "src/qs8-gemm/gen/4x16c2-minmax-neon-mull-padal-dup.c",
Frank Barchard58990122021-01-29 16:09:56 -08001736 "src/qs8-gemm/gen/4x16c8-minmax-neon-mull-padal.c",
Frank Barchard71c4d1a2021-02-19 10:07:23 -08001737 "src/qs8-gemm/gen/1x8c16-minmax-neon-mlal-padal.c",
1738 "src/qs8-gemm/gen/2x8c16-minmax-neon-mlal-padal.c",
1739 "src/qs8-gemm/gen/3x8c16-minmax-neon-mlal-padal.c",
1740 "src/qs8-gemm/gen/4x8c16-minmax-neon-mlal-padal.c",
1741 "src/qs8-gemm/gen/1x16c16-minmax-neon-mlal-padal.c",
1742 "src/qs8-gemm/gen/2x16c16-minmax-neon-mlal-padal.c",
1743 "src/qs8-gemm/gen/3x16c16-minmax-neon-mlal-padal.c",
1744 "src/qs8-gemm/gen/4x16c16-minmax-neon-mlal-padal.c",
Frank Barchardec0bf142021-01-15 10:14:05 -08001745 "src/qs8-igemm/gen/1x8-minmax-neon-mlal-lane.c",
Frank Barchardcadd4222021-01-20 16:27:25 -08001746 "src/qs8-igemm/gen/1x16-minmax-neon-mlal-lane.c",
Frank Barchardec0bf142021-01-15 10:14:05 -08001747 "src/qs8-igemm/gen/2x8-minmax-neon-mlal-lane.c",
Frank Barchardcadd4222021-01-20 16:27:25 -08001748 "src/qs8-igemm/gen/2x16-minmax-neon-mlal-lane.c",
Frank Barchardec0bf142021-01-15 10:14:05 -08001749 "src/qs8-igemm/gen/3x8-minmax-neon-mlal-lane.c",
Frank Barchardcadd4222021-01-20 16:27:25 -08001750 "src/qs8-igemm/gen/3x16-minmax-neon-mlal-lane.c",
Frank Barchardec0bf142021-01-15 10:14:05 -08001751 "src/qs8-igemm/gen/4x8-minmax-neon-mlal-lane.c",
Frank Barchardcadd4222021-01-20 16:27:25 -08001752 "src/qs8-igemm/gen/4x16-minmax-neon-mlal-lane.c",
Frank Barchardd79391d2021-02-15 14:22:32 -08001753 "src/qs8-igemm/gen/1x16c8-minmax-neon-mull-padal.c",
1754 "src/qs8-igemm/gen/1x8c8-minmax-neon-mull-padal.c",
1755 "src/qs8-igemm/gen/2x16c8-minmax-neon-mull-padal.c",
1756 "src/qs8-igemm/gen/2x8c8-minmax-neon-mull-padal.c",
1757 "src/qs8-igemm/gen/3x16c8-minmax-neon-mull-padal.c",
1758 "src/qs8-igemm/gen/3x8c8-minmax-neon-mull-padal.c",
1759 "src/qs8-igemm/gen/4x16c8-minmax-neon-mull-padal.c",
1760 "src/qs8-igemm/gen/4x8c8-minmax-neon-mull-padal.c",
Frank Barchard6fa80782021-02-15 16:18:26 -08001761 "src/qs8-igemm/gen/1x16c2-minmax-neon-mlal-padal-dup.c",
1762 "src/qs8-igemm/gen/1x8c2-minmax-neon-mlal-padal-dup.c",
1763 "src/qs8-igemm/gen/2x16c2-minmax-neon-mlal-padal-dup.c",
1764 "src/qs8-igemm/gen/2x8c2-minmax-neon-mlal-padal-dup.c",
1765 "src/qs8-igemm/gen/3x16c2-minmax-neon-mlal-padal-dup.c",
1766 "src/qs8-igemm/gen/3x8c2-minmax-neon-mlal-padal-dup.c",
1767 "src/qs8-igemm/gen/4x16c2-minmax-neon-mlal-padal-dup.c",
1768 "src/qs8-igemm/gen/4x8c2-minmax-neon-mlal-padal-dup.c",
1769 "src/qs8-igemm/gen/1x16c2-minmax-neon-mull-padal-dup.c",
1770 "src/qs8-igemm/gen/1x8c2-minmax-neon-mull-padal-dup.c",
1771 "src/qs8-igemm/gen/2x16c2-minmax-neon-mull-padal-dup.c",
1772 "src/qs8-igemm/gen/2x8c2-minmax-neon-mull-padal-dup.c",
1773 "src/qs8-igemm/gen/3x16c2-minmax-neon-mull-padal-dup.c",
1774 "src/qs8-igemm/gen/3x8c2-minmax-neon-mull-padal-dup.c",
1775 "src/qs8-igemm/gen/4x16c2-minmax-neon-mull-padal-dup.c",
1776 "src/qs8-igemm/gen/4x8c2-minmax-neon-mull-padal-dup.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -07001777 "src/qs8-requantization/fp32-neon.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001778 "src/qs8-requantization/precise-neon.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -07001779 "src/qs8-requantization/q31-neon.c",
Marat Dukhanba7b2792020-09-02 14:26:45 -07001780 "src/qs8-vadd/gen/minmax-neon-ld64-x8.c",
1781 "src/qs8-vadd/gen/minmax-neon-ld64-x16.c",
1782 "src/qs8-vadd/gen/minmax-neon-ld64-x24.c",
1783 "src/qs8-vadd/gen/minmax-neon-ld64-x32.c",
1784 "src/qs8-vaddc/gen/minmax-neon-ld64-x8.c",
1785 "src/qs8-vaddc/gen/minmax-neon-ld64-x16.c",
1786 "src/qs8-vaddc/gen/minmax-neon-ld64-x24.c",
1787 "src/qs8-vaddc/gen/minmax-neon-ld64-x32.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -07001788 "src/qu8-avgpool/9p8x-minmax-neon-c8.c",
1789 "src/qu8-avgpool/9x-minmax-neon-c8.c",
1790 "src/qu8-dwconv/up8x9-minmax-neon.c",
1791 "src/qu8-gavgpool/7p7x-minmax-neon-c8.c",
1792 "src/qu8-gavgpool/7x-minmax-neon-c8.c",
1793 "src/qu8-gemm/4x8-minmax-neon.c",
1794 "src/qu8-gemm/8x8-minmax-neon.c",
1795 "src/qu8-igemm/4x8-minmax-neon.c",
1796 "src/qu8-igemm/8x8-minmax-neon.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07001797 "src/qu8-requantization/fp32-neon.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001798 "src/qu8-requantization/precise-neon.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07001799 "src/qu8-requantization/q31-neon.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -07001800 "src/qu8-vadd/minmax-neon.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001801 "src/u8-clamp/neon-x64.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001802 "src/u8-maxpool/9p8x-minmax-neon-c16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001803 "src/u8-rmax/neon.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001804 "src/x8-zip/x2-neon.c",
1805 "src/x8-zip/x3-neon.c",
1806 "src/x8-zip/x4-neon.c",
1807 "src/x8-zip/xm-neon.c",
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -07001808 "src/x32-fill/neon.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001809 "src/x32-packx/x4-neon-st4.c",
Marat Dukhan63523d42020-05-22 17:07:33 -07001810 "src/x32-pad/neon.c",
Marat Dukhan57dccd82020-04-14 00:53:10 -07001811 "src/x32-unpool/neon.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001812 "src/x32-zip/x2-neon.c",
1813 "src/x32-zip/x3-neon.c",
1814 "src/x32-zip/x4-neon.c",
1815 "src/x32-zip/xm-neon.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001816]
1817
1818NEONFMA_UKERNELS = [
Frank Barchard04336c12020-10-22 16:48:55 -07001819 "src/f32-dwconv/gen/up4x4-minmax-neonfma-acc2.c",
1820 "src/f32-dwconv/gen/up4x4-minmax-neonfma.c",
1821 "src/f32-dwconv/gen/up4x9-minmax-neonfma-acc2.c",
1822 "src/f32-dwconv/gen/up4x9-minmax-neonfma.c",
1823 "src/f32-dwconv/gen/up4x25-minmax-neonfma-acc2.c",
1824 "src/f32-dwconv/gen/up4x25-minmax-neonfma.c",
1825 "src/f32-dwconv/gen/up8x4-minmax-neonfma-acc2.c",
1826 "src/f32-dwconv/gen/up8x4-minmax-neonfma.c",
1827 "src/f32-dwconv/gen/up8x9-minmax-neonfma-acc2.c",
1828 "src/f32-dwconv/gen/up8x9-minmax-neonfma.c",
1829 "src/f32-dwconv/gen/up8x25-minmax-neonfma-acc2.c",
1830 "src/f32-dwconv/gen/up8x25-minmax-neonfma.c",
1831 "src/f32-gemm/gen-inc/1x8inc-minmax-neonfma-dup-ld64.c",
1832 "src/f32-gemm/gen-inc/1x8s4inc-minmax-neonfma.c",
1833 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-dup-ld64.c",
1834 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-dup-ld128.c",
1835 "src/f32-gemm/gen-inc/4x8s4inc-minmax-neonfma.c",
1836 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-dup-ld64.c",
1837 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-dup-ld128.c",
1838 "src/f32-gemm/gen-inc/6x8s4inc-minmax-neonfma.c",
1839 "src/f32-gemm/gen-inc/8x8s4inc-minmax-neonfma.c",
1840 "src/f32-gemm/gen/1x8-minmax-neonfma-dup-ld64.c",
1841 "src/f32-gemm/gen/1x8s4-minmax-neonfma.c",
1842 "src/f32-gemm/gen/4x8-minmax-neonfma-dup-ld64.c",
1843 "src/f32-gemm/gen/4x8-minmax-neonfma-dup-ld128.c",
1844 "src/f32-gemm/gen/4x8s4-minmax-neonfma.c",
1845 "src/f32-gemm/gen/6x8-minmax-neonfma-dup-ld64.c",
1846 "src/f32-gemm/gen/6x8-minmax-neonfma-dup-ld128.c",
1847 "src/f32-gemm/gen/6x8s4-minmax-neonfma.c",
1848 "src/f32-gemm/gen/8x8s4-minmax-neonfma.c",
Artsiom Ablavatski2202c812021-01-22 14:16:43 -08001849 "src/f32-ibilinear-chw/gen/neonfma-p4.c",
1850 "src/f32-ibilinear-chw/gen/neonfma-p8.c",
Frank Barchard8247e212021-02-03 18:12:33 -08001851 "src/f32-ibilinear/gen/neonfma-c4.c",
1852 "src/f32-ibilinear/gen/neonfma-c8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001853 "src/f32-igemm/gen/1x8-minmax-neonfma-dup-ld64.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001854 "src/f32-igemm/gen/1x8s4-minmax-neonfma.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001855 "src/f32-igemm/gen/4x8-minmax-neonfma-dup-ld64.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001856 "src/f32-igemm/gen/4x8-minmax-neonfma-dup-ld128.c",
1857 "src/f32-igemm/gen/4x8s4-minmax-neonfma.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001858 "src/f32-igemm/gen/6x8-minmax-neonfma-dup-ld64.c",
1859 "src/f32-igemm/gen/6x8-minmax-neonfma-dup-ld128.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001860 "src/f32-igemm/gen/6x8s4-minmax-neonfma.c",
1861 "src/f32-igemm/gen/8x8s4-minmax-neonfma.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001862 "src/f32-ppmm/gen/4x8-minmax-neonfma.c",
1863 "src/f32-ppmm/gen/8x8-minmax-neonfma.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001864 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x4.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001865 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x8-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001866 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x8.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001867 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12-acc2.c",
1868 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001869 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001870 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16-acc2.c",
1871 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001872 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001873 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20-acc2.c",
1874 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001875 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20.c",
1876 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x4.c",
1877 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x8-acc2.c",
1878 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x8.c",
1879 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12-acc2.c",
1880 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12-acc3.c",
1881 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12.c",
1882 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16-acc2.c",
1883 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16-acc4.c",
1884 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16.c",
1885 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20-acc2.c",
1886 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20-acc5.c",
1887 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -08001888 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x4.c",
1889 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x8.c",
1890 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x12.c",
1891 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x16.c",
1892 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x20.c",
1893 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x24.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001894 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x4.c",
1895 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x8.c",
1896 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x12.c",
1897 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x16.c",
1898 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x20.c",
1899 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x24.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -08001900 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x4.c",
1901 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x8.c",
1902 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x12.c",
1903 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x16.c",
1904 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x20.c",
1905 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x24.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -08001906 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x4.c",
1907 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x8.c",
1908 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x12.c",
1909 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x16.c",
1910 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x20.c",
1911 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x24.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001912 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x4.c",
1913 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x8.c",
1914 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x12.c",
1915 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x16.c",
1916 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x20.c",
1917 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x24.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -08001918 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x4.c",
1919 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x8.c",
1920 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x12.c",
1921 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x16.c",
1922 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x20.c",
1923 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x24.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001924 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x4.c",
1925 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x8.c",
1926 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x12.c",
1927 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x16.c",
1928 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x20.c",
1929 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x24.c",
1930 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x4.c",
1931 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x8.c",
1932 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x12.c",
1933 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x16.c",
1934 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x20.c",
1935 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x24.c",
1936 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x4.c",
1937 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x8.c",
1938 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x12.c",
1939 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x16.c",
1940 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x20.c",
1941 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x24.c",
Marat Dukhan2fa7a0c2020-12-06 19:09:02 -08001942 "src/f32-spmm/gen/4x1-minmax-neonfma-pipelined.c",
1943 "src/f32-spmm/gen/4x1-minmax-neonfma-x2.c",
1944 "src/f32-spmm/gen/4x1-minmax-neonfma.c",
1945 "src/f32-spmm/gen/8x1-minmax-neonfma-pipelined.c",
1946 "src/f32-spmm/gen/8x1-minmax-neonfma-x2.c",
1947 "src/f32-spmm/gen/8x1-minmax-neonfma.c",
1948 "src/f32-spmm/gen/12x1-minmax-neonfma.c",
1949 "src/f32-spmm/gen/16x1-minmax-neonfma-pipelined.c",
1950 "src/f32-spmm/gen/16x1-minmax-neonfma-x2.c",
1951 "src/f32-spmm/gen/16x1-minmax-neonfma.c",
1952 "src/f32-spmm/gen/32x1-minmax-neonfma-pipelined.c",
1953 "src/f32-spmm/gen/32x1-minmax-neonfma-x2.c",
1954 "src/f32-spmm/gen/32x1-minmax-neonfma.c",
Marat Dukhaned6baaf2020-12-01 15:07:08 -08001955 "src/f32-velu/gen/velu-neonfma-rr1-lut16-p3-x4.c",
1956 "src/f32-velu/gen/velu-neonfma-rr1-lut16-p3-x8.c",
1957 "src/f32-velu/gen/velu-neonfma-rr1-lut16-p3-x12.c",
1958 "src/f32-velu/gen/velu-neonfma-rr1-lut16-p3-x16.c",
1959 "src/f32-velu/gen/velu-neonfma-rr1-lut16-p3-x20.c",
1960 "src/f32-velu/gen/velu-neonfma-rr1-lut16-p3-x24.c",
1961 "src/f32-velu/gen/velu-neonfma-rr1-p6-x4.c",
1962 "src/f32-velu/gen/velu-neonfma-rr1-p6-x8.c",
1963 "src/f32-velu/gen/velu-neonfma-rr1-p6-x12.c",
1964 "src/f32-velu/gen/velu-neonfma-rr1-p6-x16.c",
1965 "src/f32-velu/gen/velu-neonfma-rr1-p6-x20.c",
1966 "src/f32-velu/gen/velu-neonfma-rr1-p6-x24.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001967 "src/f32-vmulcaddc/gen/c4-minmax-neonfma-2x.c",
1968 "src/f32-vmulcaddc/gen/c8-minmax-neonfma-2x.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07001969 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x4.c",
1970 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x8.c",
1971 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x12.c",
1972 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x16.c",
1973 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x20.c",
1974 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x24.c",
1975 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x28.c",
1976 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x32.c",
1977 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x36.c",
1978 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x40.c",
1979 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x4.c",
1980 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x8.c",
1981 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x12.c",
1982 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x16.c",
1983 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x20.c",
1984 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x24.c",
1985 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x28.c",
1986 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x32.c",
1987 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x36.c",
1988 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x40.c",
Marat Dukhanb7633f22020-11-20 16:34:56 -08001989 "src/math/exp-neonfma-rr2-lut64-p2.c",
1990 "src/math/exp-neonfma-rr2-p5.c",
Frank Barcharde7223ee2020-12-04 19:04:01 -08001991 "src/math/expm1minus-neonfma-rr1-lut16-p3.c",
1992 "src/math/expm1minus-neonfma-rr1-p6.c",
Marat Dukhan9dd119a2020-11-20 18:20:04 -08001993 "src/math/expminus-neonfma-rr2-lut64-p2.c",
1994 "src/math/expminus-neonfma-rr2-lut2048-p1.c",
1995 "src/math/expminus-neonfma-rr2-p5.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001996 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr1recps1fma.c",
1997 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr2fma.c",
1998 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr2recps.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001999 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr1recps1fma.c",
2000 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr2fma.c",
2001 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr2recps.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08002002 "src/math/sigmoid-neonfma-rr1-p5-nr1recps1fma.c",
2003 "src/math/sigmoid-neonfma-rr1-p5-nr2fma.c",
2004 "src/math/sigmoid-neonfma-rr1-p5-nr2recps.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08002005 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr1recps1fma.c",
2006 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr2fma.c",
2007 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr2recps.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002008 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr1recps1fma.c",
2009 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr2fma.c",
2010 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr2recps.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08002011 "src/math/sigmoid-neonfma-rr2-p5-nr1recps1fma.c",
2012 "src/math/sigmoid-neonfma-rr2-p5-nr2fma.c",
2013 "src/math/sigmoid-neonfma-rr2-p5-nr2recps.c",
Marat Dukhan84000762020-06-29 18:38:43 -07002014 "src/math/sqrt-neonfma-nr1fma.c",
Marat Dukhan84000762020-06-29 18:38:43 -07002015 "src/math/sqrt-neonfma-nr1rsqrts1fma1adj.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002016 "src/math/sqrt-neonfma-nr2fma.c",
2017 "src/math/sqrt-neonfma-nr2fma1adj.c",
2018 "src/math/sqrt-neonfma-nr3fma.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002019]
2020
2021AARCH64_NEONFMA_UKERNELS = [
Marat Dukhan56b10cd2020-05-18 09:35:49 -07002022 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neonfma-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07002023 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neonfma-2x2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002024 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neonfma-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07002025 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neonfma-2x2.c",
Marat Dukhan56b10cd2020-05-18 09:35:49 -07002026 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neonfma-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07002027 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neonfma-2x2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002028 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neonfma-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07002029 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neonfma-2x2.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -07002030 "src/f32-conv-hwc2chw/3x3s2p1c3x4-neonfma-2x2.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07002031 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-1x4-acc2.c",
2032 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-1x4-acc3.c",
2033 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-1x4-acc4.c",
Marat Dukhan1268a242020-10-24 00:36:32 -07002034 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-1x4.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07002035 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-2x4-acc2.c",
Marat Dukhan1268a242020-10-24 00:36:32 -07002036 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-2x4.c",
2037 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-3x4.c",
2038 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-4x4.c",
2039 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-5x4.c",
2040 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-6x4.c",
Marat Dukhan82f0c322020-10-25 19:17:35 -07002041 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neonfma-1x4-acc2.c",
2042 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neonfma-1x4-acc3.c",
2043 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neonfma-1x4-acc4.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07002044 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neonfma-1x4.c",
Marat Dukhan82f0c322020-10-25 19:17:35 -07002045 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neonfma-2x4-acc2.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07002046 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neonfma-2x4.c",
2047 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neonfma-3x4.c",
2048 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neonfma-4x4.c",
Marat Dukhan149f0ea2020-10-26 12:50:33 -07002049 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-1x4-acc2.c",
2050 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-1x4-acc3.c",
2051 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-1x4-acc4.c",
2052 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-1x4-acc5.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07002053 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-1x4.c",
Marat Dukhan149f0ea2020-10-26 12:50:33 -07002054 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-2x4-acc2.c",
2055 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-2x4-acc3.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07002056 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-2x4.c",
Marat Dukhan149f0ea2020-10-26 12:50:33 -07002057 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-3x4-acc2.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07002058 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-3x4.c",
Marat Dukhan149f0ea2020-10-26 12:50:33 -07002059 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-4x4-acc2.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07002060 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-4x4.c",
2061 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-5x4.c",
Marat Dukhan30d4b252020-10-29 16:33:22 -07002062 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neonfma-1x4-acc2.c",
2063 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neonfma-1x4-acc3.c",
2064 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neonfma-1x4-acc4.c",
2065 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neonfma-1x4-acc5.c",
2066 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neonfma-1x4.c",
2067 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neonfma-2x4-acc2.c",
2068 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neonfma-2x4-acc3.c",
2069 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neonfma-2x4.c",
Marat Dukhan30d4b252020-10-29 16:33:22 -07002070 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neonfma-3x4-acc2.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07002071 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neonfma-3x4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002072 "src/f32-gemm/gen-inc/1x8inc-minmax-neonfma-lane-ld64.c",
2073 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-lane-ld64.c",
2074 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-lane-ld128.c",
2075 "src/f32-gemm/gen-inc/5x8inc-minmax-neonfma-lane-ld64.c",
2076 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-lane-ld64.c",
2077 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-lane-ld128.c",
2078 "src/f32-gemm/gen/1x8-minmax-neonfma-lane-ld64.c",
2079 "src/f32-gemm/gen/4x2-minmax-neonfma-lane-ld64.c",
2080 "src/f32-gemm/gen/4x8-minmax-neonfma-lane-ld64.c",
2081 "src/f32-gemm/gen/4x8-minmax-neonfma-lane-ld128.c",
2082 "src/f32-gemm/gen/5x8-minmax-neonfma-lane-ld64.c",
2083 "src/f32-gemm/gen/6x8-minmax-neonfma-lane-ld64.c",
2084 "src/f32-gemm/gen/6x8-minmax-neonfma-lane-ld128.c",
2085 "src/f32-igemm/gen/1x8-minmax-neonfma-lane-ld64.c",
2086 "src/f32-igemm/gen/4x2-minmax-neonfma-lane-ld64.c",
2087 "src/f32-igemm/gen/4x4-minmax-neonfma-lane-ld64.c",
2088 "src/f32-igemm/gen/4x8-minmax-neonfma-lane-ld64.c",
2089 "src/f32-igemm/gen/4x8-minmax-neonfma-lane-ld128.c",
2090 "src/f32-igemm/gen/6x8-minmax-neonfma-lane-ld64.c",
2091 "src/f32-igemm/gen/6x8-minmax-neonfma-lane-ld128.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -08002092 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x4.c",
2093 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x8.c",
2094 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x12.c",
2095 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x16.c",
2096 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x20.c",
2097 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x24.c",
2098 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x4.c",
2099 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x8.c",
2100 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x12.c",
2101 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x16.c",
2102 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x20.c",
2103 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x24.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002104 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x4.c",
2105 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x8.c",
2106 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x12.c",
2107 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x16.c",
2108 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x20.c",
2109 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x24.c",
Marat Dukhan355ab432020-04-09 19:01:52 -07002110 "src/f32-spmm/gen/4x2-minmax-neonfma.c",
2111 "src/f32-spmm/gen/4x4-minmax-neonfma.c",
Marat Dukhan355ab432020-04-09 19:01:52 -07002112 "src/f32-spmm/gen/8x2-minmax-neonfma.c",
2113 "src/f32-spmm/gen/8x4-minmax-neonfma.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002114 "src/f32-spmm/gen/12x2-minmax-neonfma.c",
2115 "src/f32-spmm/gen/12x4-minmax-neonfma.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002116 "src/f32-spmm/gen/16x2-minmax-neonfma.c",
2117 "src/f32-spmm/gen/16x4-minmax-neonfma.c",
Frank Barchard846c0c62020-10-26 15:01:39 -07002118 "src/f32-spmm/gen/32x2-minmax-neonfma.c",
2119 "src/f32-spmm/gen/32x4-minmax-neonfma.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002120 "src/f32-vbinary/gen/vdiv-minmax-neon-x4.c",
2121 "src/f32-vbinary/gen/vdiv-minmax-neon-x8.c",
2122 "src/f32-vbinary/gen/vdivc-minmax-neon-x4.c",
2123 "src/f32-vbinary/gen/vdivc-minmax-neon-x8.c",
2124 "src/f32-vbinary/gen/vrdivc-minmax-neon-x4.c",
2125 "src/f32-vbinary/gen/vrdivc-minmax-neon-x8.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07002126 "src/f32-vsqrt/gen/neon-sqrt-x4.c",
2127 "src/f32-vsqrt/gen/neon-sqrt-x8.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08002128 "src/math/sigmoid-neonfma-rr1-lut64-p2-div.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002129 "src/math/sigmoid-neonfma-rr1-lut2048-p1-div.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08002130 "src/math/sigmoid-neonfma-rr1-p5-div.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08002131 "src/math/sigmoid-neonfma-rr2-lut64-p2-div.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002132 "src/math/sigmoid-neonfma-rr2-lut2048-p1-div.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08002133 "src/math/sigmoid-neonfma-rr2-p5-div.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002134]
2135
Marat Dukhan8853b822020-05-07 12:19:01 -07002136NEONV8_UKERNELS = [
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07002137 "src/f32-vrnd/gen/vrndd-neonv8-x4.c",
2138 "src/f32-vrnd/gen/vrndd-neonv8-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002139 "src/f32-vrnd/gen/vrndne-neonv8-x4.c",
2140 "src/f32-vrnd/gen/vrndne-neonv8-x8.c",
2141 "src/f32-vrnd/gen/vrndu-neonv8-x4.c",
2142 "src/f32-vrnd/gen/vrndu-neonv8-x8.c",
2143 "src/f32-vrnd/gen/vrndz-neonv8-x4.c",
2144 "src/f32-vrnd/gen/vrndz-neonv8-x8.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07002145 "src/math/roundd-neonv8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002146 "src/math/roundne-neonv8.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07002147 "src/math/roundu-neonv8.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07002148 "src/math/roundz-neonv8.c",
Marat Dukhan8853b822020-05-07 12:19:01 -07002149]
2150
Marat Dukhan08c4a432019-10-03 09:29:21 -07002151AARCH64_NEONFP16ARITH_UKERNELS = [
Frank Barchardb1966592020-05-12 13:47:06 -07002152 "src/f16-clamp/gen/neonfp16arith-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002153 "src/f16-clamp/gen/neonfp16arith-x16.c",
Frank Barchard5a599a62020-06-04 20:12:44 -07002154 "src/f16-dwconv/gen/up8x4-minmax-neonfp16arith-acc2.c",
2155 "src/f16-dwconv/gen/up8x4-minmax-neonfp16arith.c",
2156 "src/f16-dwconv/gen/up8x9-minmax-neonfp16arith-acc2.c",
2157 "src/f16-dwconv/gen/up8x9-minmax-neonfp16arith.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002158 "src/f16-dwconv/gen/up8x25-minmax-neonfp16arith-acc2.c",
2159 "src/f16-dwconv/gen/up8x25-minmax-neonfp16arith.c",
2160 "src/f16-dwconv/gen/up16x4-minmax-neonfp16arith-acc2.c",
2161 "src/f16-dwconv/gen/up16x4-minmax-neonfp16arith.c",
2162 "src/f16-dwconv/gen/up16x9-minmax-neonfp16arith-acc2.c",
2163 "src/f16-dwconv/gen/up16x9-minmax-neonfp16arith.c",
2164 "src/f16-dwconv/gen/up16x25-minmax-neonfp16arith-acc2.c",
2165 "src/f16-dwconv/gen/up16x25-minmax-neonfp16arith.c",
Frank Barchard0bb49a72020-06-04 11:35:11 -07002166 "src/f16-gavgpool/7p7x-minmax-neonfp16arith-c8.c",
2167 "src/f16-gavgpool/7x-minmax-neonfp16arith-c8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002168 "src/f16-gemm/gen-inc/1x8inc-minmax-neonfp16arith-ld64.c",
2169 "src/f16-gemm/gen-inc/1x16inc-minmax-neonfp16arith-ld64.c",
2170 "src/f16-gemm/gen-inc/4x8inc-minmax-neonfp16arith-ld64.c",
2171 "src/f16-gemm/gen-inc/4x16inc-minmax-neonfp16arith-ld64.c",
2172 "src/f16-gemm/gen-inc/6x8inc-minmax-neonfp16arith-ld64.c",
2173 "src/f16-gemm/gen-inc/6x16inc-minmax-neonfp16arith-ld64.c",
2174 "src/f16-gemm/gen-inc/8x8inc-minmax-neonfp16arith-ld64.c",
2175 "src/f16-gemm/gen-inc/8x16inc-minmax-neonfp16arith-ld64.c",
2176 "src/f16-gemm/gen/1x8-minmax-neonfp16arith-ld64.c",
2177 "src/f16-gemm/gen/1x16-minmax-neonfp16arith-ld64.c",
2178 "src/f16-gemm/gen/4x8-minmax-neonfp16arith-ld64.c",
2179 "src/f16-gemm/gen/4x16-minmax-neonfp16arith-ld64.c",
2180 "src/f16-gemm/gen/6x8-minmax-neonfp16arith-ld64.c",
2181 "src/f16-gemm/gen/6x16-minmax-neonfp16arith-ld64.c",
2182 "src/f16-gemm/gen/8x8-minmax-neonfp16arith-ld64.c",
2183 "src/f16-gemm/gen/8x16-minmax-neonfp16arith-ld64.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07002184 "src/f16-hswish/gen/hswish-neonfp16arith-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002185 "src/f16-hswish/gen/hswish-neonfp16arith-x16.c",
2186 "src/f16-igemm/gen/1x8-minmax-neonfp16arith-ld64.c",
2187 "src/f16-igemm/gen/1x16-minmax-neonfp16arith-ld64.c",
2188 "src/f16-igemm/gen/4x8-minmax-neonfp16arith-ld64.c",
2189 "src/f16-igemm/gen/4x16-minmax-neonfp16arith-ld64.c",
2190 "src/f16-igemm/gen/6x8-minmax-neonfp16arith-ld64.c",
2191 "src/f16-igemm/gen/6x16-minmax-neonfp16arith-ld64.c",
2192 "src/f16-igemm/gen/8x8-minmax-neonfp16arith-ld64.c",
2193 "src/f16-igemm/gen/8x16-minmax-neonfp16arith-ld64.c",
Frank Barchardb1966592020-05-12 13:47:06 -07002194 "src/f16-prelu/gen/neonfp16arith-2x8.c",
Frank Barcharda5316982020-07-23 13:19:28 -07002195 "src/f16-prelu/gen/neonfp16arith-2x16.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07002196 "src/f16-relu/gen/neonfp16arith-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002197 "src/f16-relu/gen/neonfp16arith-x16.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07002198 "src/f16-spmm/gen/8x1-minmax-neonfp16arith-x2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002199 "src/f16-spmm/gen/8x1-minmax-neonfp16arith.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07002200 "src/f16-spmm/gen/16x1-minmax-neonfp16arith-x2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002201 "src/f16-spmm/gen/16x1-minmax-neonfp16arith.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07002202 "src/f16-spmm/gen/24x1-minmax-neonfp16arith-x2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002203 "src/f16-spmm/gen/24x1-minmax-neonfp16arith.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07002204 "src/f16-spmm/gen/32x1-minmax-neonfp16arith-x2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002205 "src/f16-spmm/gen/32x1-minmax-neonfp16arith.c",
2206 "src/f16-vbinary/gen/vadd-minmax-neonfp16arith-x8.c",
2207 "src/f16-vbinary/gen/vadd-minmax-neonfp16arith-x16.c",
2208 "src/f16-vbinary/gen/vaddc-minmax-neonfp16arith-x8.c",
2209 "src/f16-vbinary/gen/vaddc-minmax-neonfp16arith-x16.c",
2210 "src/f16-vbinary/gen/vdiv-minmax-neonfp16arith-x8.c",
2211 "src/f16-vbinary/gen/vdiv-minmax-neonfp16arith-x16.c",
2212 "src/f16-vbinary/gen/vdivc-minmax-neonfp16arith-x8.c",
2213 "src/f16-vbinary/gen/vdivc-minmax-neonfp16arith-x16.c",
2214 "src/f16-vbinary/gen/vmax-neonfp16arith-x8.c",
2215 "src/f16-vbinary/gen/vmax-neonfp16arith-x16.c",
2216 "src/f16-vbinary/gen/vmaxc-neonfp16arith-x8.c",
2217 "src/f16-vbinary/gen/vmaxc-neonfp16arith-x16.c",
2218 "src/f16-vbinary/gen/vmin-neonfp16arith-x8.c",
2219 "src/f16-vbinary/gen/vmin-neonfp16arith-x16.c",
2220 "src/f16-vbinary/gen/vminc-neonfp16arith-x8.c",
2221 "src/f16-vbinary/gen/vminc-neonfp16arith-x16.c",
2222 "src/f16-vbinary/gen/vmul-minmax-neonfp16arith-x8.c",
2223 "src/f16-vbinary/gen/vmul-minmax-neonfp16arith-x16.c",
2224 "src/f16-vbinary/gen/vmulc-minmax-neonfp16arith-x8.c",
2225 "src/f16-vbinary/gen/vmulc-minmax-neonfp16arith-x16.c",
2226 "src/f16-vbinary/gen/vrdivc-minmax-neonfp16arith-x8.c",
2227 "src/f16-vbinary/gen/vrdivc-minmax-neonfp16arith-x16.c",
2228 "src/f16-vbinary/gen/vrsubc-minmax-neonfp16arith-x8.c",
2229 "src/f16-vbinary/gen/vrsubc-minmax-neonfp16arith-x16.c",
2230 "src/f16-vbinary/gen/vsub-minmax-neonfp16arith-x8.c",
2231 "src/f16-vbinary/gen/vsub-minmax-neonfp16arith-x16.c",
2232 "src/f16-vbinary/gen/vsubc-minmax-neonfp16arith-x8.c",
2233 "src/f16-vbinary/gen/vsubc-minmax-neonfp16arith-x16.c",
2234 "src/f16-vmulcaddc/gen/c8-minmax-neonfp16arith-2x.c",
2235 "src/f16-vmulcaddc/gen/c16-minmax-neonfp16arith-2x.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002236]
2237
Benoit Jacoba9644732020-08-13 12:48:55 -07002238NEONDOT_UKERNELS = [
Marat Dukhand4c83032020-09-09 23:20:12 -07002239 "src/qs8-gemm/gen/1x8c4-minmax-neondot.c",
Marat Dukhand4c83032020-09-09 23:20:12 -07002240 "src/qs8-gemm/gen/1x16c4-minmax-neondot.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002241 "src/qs8-gemm/gen/4x8c4-minmax-neondot.c",
Benoit Jacoba9644732020-08-13 12:48:55 -07002242 "src/qs8-gemm/gen/4x16c4-minmax-neondot.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002243 "src/qs8-gemm/gen/6x8c4-minmax-neondot.c",
Frank Barchard2fa17452020-09-22 09:51:46 -07002244 "src/qs8-gemm/gen/6x16c4-minmax-neondot.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002245 "src/qs8-gemm/gen/8x8c4-minmax-neondot.c",
Frank Barcharda48848f2020-09-24 00:59:24 -07002246 "src/qs8-gemm/gen/8x16c4-minmax-neondot.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002247 "src/qs8-gemm/gen/12x8c4-minmax-neondot.c",
Marat Dukhand4c83032020-09-09 23:20:12 -07002248 "src/qs8-igemm/gen/1x8c4-minmax-neondot.c",
Marat Dukhand4c83032020-09-09 23:20:12 -07002249 "src/qs8-igemm/gen/1x16c4-minmax-neondot.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002250 "src/qs8-igemm/gen/4x8c4-minmax-neondot.c",
Benoit Jacoba05487f2020-08-14 13:37:21 -07002251 "src/qs8-igemm/gen/4x16c4-minmax-neondot.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002252 "src/qs8-igemm/gen/6x8c4-minmax-neondot.c",
Frank Barchard2fa17452020-09-22 09:51:46 -07002253 "src/qs8-igemm/gen/6x16c4-minmax-neondot.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002254 "src/qs8-igemm/gen/8x8c4-minmax-neondot.c",
Frank Barcharda48848f2020-09-24 00:59:24 -07002255 "src/qs8-igemm/gen/8x16c4-minmax-neondot.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002256 "src/qs8-igemm/gen/12x8c4-minmax-neondot.c",
Benoit Jacoba9644732020-08-13 12:48:55 -07002257]
2258
Marat Dukhan08c4a432019-10-03 09:29:21 -07002259SSE_UKERNELS = [
Marat Dukhan99936602020-04-11 16:47:01 -07002260 "src/f32-avgpool/9p8x-minmax-sse-c4.c",
2261 "src/f32-avgpool/9x-minmax-sse-c4.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -07002262 "src/f32-clamp/gen/sse-x4.c",
2263 "src/f32-clamp/gen/sse-x8.c",
Erich Elsenb1233402020-06-08 15:53:15 -07002264 "src/f32-conv-hwc2chw/3x3s2p1c3x4-sse-1x1.c",
2265 "src/f32-conv-hwc2chw/3x3s2p1c3x4-sse-2x2.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002266 "src/f32-dwconv/gen/up4x4-minmax-sse-acc2.c",
2267 "src/f32-dwconv/gen/up4x4-minmax-sse.c",
2268 "src/f32-dwconv/gen/up4x9-minmax-sse-acc2.c",
2269 "src/f32-dwconv/gen/up4x9-minmax-sse.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002270 "src/f32-dwconv/gen/up4x25-minmax-sse-acc2.c",
2271 "src/f32-dwconv/gen/up4x25-minmax-sse.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002272 "src/f32-dwconv/gen/up8x4-minmax-sse-acc2.c",
2273 "src/f32-dwconv/gen/up8x4-minmax-sse.c",
2274 "src/f32-dwconv/gen/up8x9-minmax-sse-acc2.c",
2275 "src/f32-dwconv/gen/up8x9-minmax-sse.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002276 "src/f32-dwconv/gen/up8x25-minmax-sse-acc2.c",
2277 "src/f32-dwconv/gen/up8x25-minmax-sse.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07002278 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-1x4-acc2.c",
2279 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-1x4-acc3.c",
2280 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-1x4-acc4.c",
Marat Dukhan470078a2020-10-23 22:36:52 -07002281 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-1x4.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07002282 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-2x4-acc2.c",
Marat Dukhan470078a2020-10-23 22:36:52 -07002283 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-2x4.c",
2284 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-3x4.c",
2285 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-4x4.c",
2286 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-5x4.c",
2287 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-6x4.c",
Marat Dukhan0ff97182020-10-25 19:14:03 -07002288 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-sse-1x4-acc2.c",
2289 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-sse-1x4-acc3.c",
2290 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-sse-1x4-acc4.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07002291 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-sse-1x4.c",
Marat Dukhan0ff97182020-10-25 19:14:03 -07002292 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-sse-2x4-acc2.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07002293 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-sse-2x4.c",
2294 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-sse-3x4.c",
2295 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-sse-4x4.c",
Marat Dukhand0503892020-10-30 08:22:04 -07002296 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-1x4-acc2.c",
2297 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-1x4-acc3.c",
2298 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-1x4-acc4.c",
2299 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-1x4-acc5.c",
2300 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-1x4.c",
2301 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-2x4-acc2.c",
2302 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-2x4-acc3.c",
2303 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-2x4.c",
2304 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-3x4-acc2.c",
2305 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-3x4.c",
2306 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-4x4-acc2.c",
2307 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-4x4.c",
2308 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-5x4.c",
Marat Dukhanccca2142020-10-30 17:32:45 -07002309 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-sse-1x4-acc2.c",
2310 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-sse-1x4-acc3.c",
2311 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-sse-1x4-acc4.c",
2312 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-sse-1x4-acc5.c",
2313 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-sse-1x4.c",
2314 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-sse-2x4-acc2.c",
2315 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-sse-2x4-acc3.c",
2316 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-sse-2x4.c",
Marat Dukhanccca2142020-10-30 17:32:45 -07002317 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-sse-3x4-acc2.c",
Frank Barchard8ef44cd2020-11-03 12:30:23 -08002318 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-sse-3x4.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -07002319 "src/f32-gavgpool-cw/sse-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -07002320 "src/f32-gavgpool/7p7x-minmax-sse-c4.c",
2321 "src/f32-gavgpool/7x-minmax-sse-c4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002322 "src/f32-gemm/gen-inc/1x8inc-minmax-sse-dup.c",
2323 "src/f32-gemm/gen-inc/1x8inc-minmax-sse-load1.c",
2324 "src/f32-gemm/gen-inc/1x8s4inc-minmax-sse.c",
Marat Dukhan802fcae2020-12-11 14:37:25 -08002325 "src/f32-gemm/gen-inc/3x8inc-minmax-sse-dup.c",
2326 "src/f32-gemm/gen-inc/3x8inc-minmax-sse-load1.c",
2327 "src/f32-gemm/gen-inc/3x8s4inc-minmax-sse.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002328 "src/f32-gemm/gen-inc/4x8inc-minmax-sse-dup.c",
2329 "src/f32-gemm/gen-inc/4x8inc-minmax-sse-load1.c",
2330 "src/f32-gemm/gen-inc/4x8s4inc-minmax-sse.c",
Marat Dukhan802fcae2020-12-11 14:37:25 -08002331 "src/f32-gemm/gen-inc/5x8inc-minmax-sse-dup.c",
2332 "src/f32-gemm/gen-inc/5x8inc-minmax-sse-load1.c",
2333 "src/f32-gemm/gen-inc/5x8s4inc-minmax-sse.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002334 "src/f32-gemm/gen/1x8-minmax-sse-dup.c",
2335 "src/f32-gemm/gen/1x8-minmax-sse-load1.c",
2336 "src/f32-gemm/gen/1x8s4-minmax-sse.c",
Marat Dukhan802fcae2020-12-11 14:37:25 -08002337 "src/f32-gemm/gen/3x8-minmax-sse-dup.c",
2338 "src/f32-gemm/gen/3x8-minmax-sse-load1.c",
2339 "src/f32-gemm/gen/3x8s4-minmax-sse.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002340 "src/f32-gemm/gen/4x2c4-minmax-sse.c",
2341 "src/f32-gemm/gen/4x8-minmax-sse-dup.c",
2342 "src/f32-gemm/gen/4x8-minmax-sse-load1.c",
2343 "src/f32-gemm/gen/4x8s4-minmax-sse.c",
Marat Dukhan802fcae2020-12-11 14:37:25 -08002344 "src/f32-gemm/gen/5x8-minmax-sse-dup.c",
2345 "src/f32-gemm/gen/5x8-minmax-sse-load1.c",
2346 "src/f32-gemm/gen/5x8s4-minmax-sse.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07002347 "src/f32-hswish/gen/hswish-sse-x4.c",
2348 "src/f32-hswish/gen/hswish-sse-x8.c",
Marat Dukhan660fd192020-03-10 04:55:30 -07002349 "src/f32-ibilinear/gen/sse-c4.c",
2350 "src/f32-ibilinear/gen/sse-c8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002351 "src/f32-igemm/gen/1x8-minmax-sse-dup.c",
2352 "src/f32-igemm/gen/1x8-minmax-sse-load1.c",
2353 "src/f32-igemm/gen/1x8s4-minmax-sse.c",
Marat Dukhan802fcae2020-12-11 14:37:25 -08002354 "src/f32-igemm/gen/3x8-minmax-sse-dup.c",
2355 "src/f32-igemm/gen/3x8-minmax-sse-load1.c",
2356 "src/f32-igemm/gen/3x8s4-minmax-sse.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002357 "src/f32-igemm/gen/4x2c4-minmax-sse.c",
2358 "src/f32-igemm/gen/4x8-minmax-sse-dup.c",
2359 "src/f32-igemm/gen/4x8-minmax-sse-load1.c",
2360 "src/f32-igemm/gen/4x8s4-minmax-sse.c",
Marat Dukhan802fcae2020-12-11 14:37:25 -08002361 "src/f32-igemm/gen/5x8-minmax-sse-dup.c",
2362 "src/f32-igemm/gen/5x8-minmax-sse-load1.c",
2363 "src/f32-igemm/gen/5x8s4-minmax-sse.c",
Marat Dukhan99936602020-04-11 16:47:01 -07002364 "src/f32-maxpool/9p8x-minmax-sse-c4.c",
2365 "src/f32-pavgpool/9p8x-minmax-sse-c4.c",
2366 "src/f32-pavgpool/9x-minmax-sse-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002367 "src/f32-ppmm/gen/4x8-minmax-sse.c",
Marat Dukhan39b5e942020-06-24 15:03:48 -07002368 "src/f32-prelu/gen/sse-2x4.c",
2369 "src/f32-prelu/gen/sse-2x8.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07002370 "src/f32-relu/gen/sse-x4.c",
2371 "src/f32-relu/gen/sse-x8.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002372 "src/f32-rmax/sse.c",
Marat Dukhan355ab432020-04-09 19:01:52 -07002373 "src/f32-spmm/gen/4x1-minmax-sse.c",
2374 "src/f32-spmm/gen/8x1-minmax-sse.c",
Erich Elsen6e80fdc2020-06-09 15:35:37 -07002375 "src/f32-spmm/gen/16x1-minmax-sse.c",
Frank Barchard846c0c62020-10-26 15:01:39 -07002376 "src/f32-spmm/gen/32x1-minmax-sse.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002377 "src/f32-vbinary/gen/vadd-minmax-sse-x4.c",
2378 "src/f32-vbinary/gen/vadd-minmax-sse-x8.c",
2379 "src/f32-vbinary/gen/vaddc-minmax-sse-x4.c",
2380 "src/f32-vbinary/gen/vaddc-minmax-sse-x8.c",
2381 "src/f32-vbinary/gen/vdiv-minmax-sse-x4.c",
2382 "src/f32-vbinary/gen/vdiv-minmax-sse-x8.c",
2383 "src/f32-vbinary/gen/vdivc-minmax-sse-x4.c",
2384 "src/f32-vbinary/gen/vdivc-minmax-sse-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -08002385 "src/f32-vbinary/gen/vmax-sse-x4.c",
2386 "src/f32-vbinary/gen/vmax-sse-x8.c",
2387 "src/f32-vbinary/gen/vmaxc-sse-x4.c",
2388 "src/f32-vbinary/gen/vmaxc-sse-x8.c",
2389 "src/f32-vbinary/gen/vmin-sse-x4.c",
2390 "src/f32-vbinary/gen/vmin-sse-x8.c",
2391 "src/f32-vbinary/gen/vminc-sse-x4.c",
2392 "src/f32-vbinary/gen/vminc-sse-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002393 "src/f32-vbinary/gen/vmul-minmax-sse-x4.c",
2394 "src/f32-vbinary/gen/vmul-minmax-sse-x8.c",
2395 "src/f32-vbinary/gen/vmulc-minmax-sse-x4.c",
2396 "src/f32-vbinary/gen/vmulc-minmax-sse-x8.c",
2397 "src/f32-vbinary/gen/vrdivc-minmax-sse-x4.c",
2398 "src/f32-vbinary/gen/vrdivc-minmax-sse-x8.c",
2399 "src/f32-vbinary/gen/vrsubc-minmax-sse-x4.c",
2400 "src/f32-vbinary/gen/vrsubc-minmax-sse-x8.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07002401 "src/f32-vbinary/gen/vsqrdiff-sse-x4.c",
2402 "src/f32-vbinary/gen/vsqrdiff-sse-x8.c",
2403 "src/f32-vbinary/gen/vsqrdiffc-sse-x4.c",
2404 "src/f32-vbinary/gen/vsqrdiffc-sse-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002405 "src/f32-vbinary/gen/vsub-minmax-sse-x4.c",
2406 "src/f32-vbinary/gen/vsub-minmax-sse-x8.c",
2407 "src/f32-vbinary/gen/vsubc-minmax-sse-x4.c",
2408 "src/f32-vbinary/gen/vsubc-minmax-sse-x8.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07002409 "src/f32-vlrelu/gen/vlrelu-sse-x4.c",
2410 "src/f32-vlrelu/gen/vlrelu-sse-x8.c",
Marat Dukhan99936602020-04-11 16:47:01 -07002411 "src/f32-vmulcaddc/gen/c4-minmax-sse-2x.c",
2412 "src/f32-vmulcaddc/gen/c8-minmax-sse-2x.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07002413 "src/f32-vsqrt/gen/sse-sqrt-x4.c",
2414 "src/f32-vsqrt/gen/sse-sqrt-x8.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07002415 "src/f32-vunary/gen/vabs-sse-x4.c",
2416 "src/f32-vunary/gen/vabs-sse-x8.c",
2417 "src/f32-vunary/gen/vneg-sse-x4.c",
2418 "src/f32-vunary/gen/vneg-sse-x8.c",
2419 "src/f32-vunary/gen/vsqr-sse-x4.c",
2420 "src/f32-vunary/gen/vsqr-sse-x8.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07002421 "src/math/roundd-sse-addsub.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002422 "src/math/roundne-sse-addsub.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07002423 "src/math/roundu-sse-addsub.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07002424 "src/math/roundz-sse-addsub.c",
Marat Dukhan84000762020-06-29 18:38:43 -07002425 "src/math/sqrt-sse-hh1mac.c",
2426 "src/math/sqrt-sse-nr1mac.c",
2427 "src/math/sqrt-sse-nr2mac.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002428 "src/x32-fill/sse.c",
2429 "src/x32-packx/x4-sse.c",
2430 "src/x32-pad/sse.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002431]
2432
2433SSE2_UKERNELS = [
Marat Dukhan329da642019-11-19 21:44:39 -08002434 "src/f32-argmaxpool/4x-sse2-c4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002435 "src/f32-argmaxpool/9p8x-sse2-c4.c",
Marat Dukhan329da642019-11-19 21:44:39 -08002436 "src/f32-argmaxpool/9x-sse2-c4.c",
Marat Dukhan802fcae2020-12-11 14:37:25 -08002437 "src/f32-gemm/gen-inc/1x8inc-minmax-sse2-dup.c",
2438 "src/f32-gemm/gen-inc/3x8inc-minmax-sse2-dup.c",
2439 "src/f32-gemm/gen-inc/4x8inc-minmax-sse2-dup.c",
2440 "src/f32-gemm/gen-inc/5x8inc-minmax-sse2-dup.c",
2441 "src/f32-gemm/gen/1x8-minmax-sse2-dup.c",
2442 "src/f32-gemm/gen/3x8-minmax-sse2-dup.c",
2443 "src/f32-gemm/gen/4x8-minmax-sse2-dup.c",
2444 "src/f32-gemm/gen/5x8-minmax-sse2-dup.c",
2445 "src/f32-igemm/gen/1x8-minmax-sse2-dup.c",
2446 "src/f32-igemm/gen/3x8-minmax-sse2-dup.c",
2447 "src/f32-igemm/gen/4x8-minmax-sse2-dup.c",
2448 "src/f32-igemm/gen/5x8-minmax-sse2-dup.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08002449 "src/f32-prelu/gen/sse2-2x4.c",
2450 "src/f32-prelu/gen/sse2-2x8.c",
Marat Dukhanb39689d2020-01-24 13:32:20 -08002451 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x4.c",
Marat Dukhanb39689d2020-01-24 13:32:20 -08002452 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x8-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002453 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x8.c",
Marat Dukhanb39689d2020-01-24 13:32:20 -08002454 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12-acc2.c",
2455 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002456 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12.c",
Marat Dukhanb39689d2020-01-24 13:32:20 -08002457 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16-acc2.c",
2458 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002459 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16.c",
Marat Dukhanb39689d2020-01-24 13:32:20 -08002460 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20-acc2.c",
2461 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002462 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20.c",
Marat Dukhand243c1a2020-09-17 18:48:10 -07002463 "src/f32-sigmoid/gen/sse2-lut64-p2-div-x4.c",
2464 "src/f32-sigmoid/gen/sse2-lut64-p2-div-x8.c",
2465 "src/f32-sigmoid/gen/sse2-lut64-p2-div-x12.c",
2466 "src/f32-sigmoid/gen/sse2-lut64-p2-div-x16.c",
2467 "src/f32-sigmoid/gen/sse2-lut64-p2-div-x20.c",
2468 "src/f32-sigmoid/gen/sse2-lut64-p2-div-x24.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002469 "src/f32-sigmoid/gen/sse2-p5-div-x4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08002470 "src/f32-sigmoid/gen/sse2-p5-div-x8.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002471 "src/f32-sigmoid/gen/sse2-p5-div-x12.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08002472 "src/f32-sigmoid/gen/sse2-p5-div-x16.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002473 "src/f32-sigmoid/gen/sse2-p5-div-x20.c",
2474 "src/f32-sigmoid/gen/sse2-p5-div-x24.c",
Marat Dukhaned6baaf2020-12-01 15:07:08 -08002475 "src/f32-velu/gen/velu-sse2-rr2-lut16-p3-x4.c",
2476 "src/f32-velu/gen/velu-sse2-rr2-lut16-p3-x8.c",
2477 "src/f32-velu/gen/velu-sse2-rr2-lut16-p3-x12.c",
2478 "src/f32-velu/gen/velu-sse2-rr2-lut16-p3-x16.c",
2479 "src/f32-velu/gen/velu-sse2-rr2-lut16-p3-x20.c",
2480 "src/f32-velu/gen/velu-sse2-rr2-lut16-p3-x24.c",
2481 "src/f32-velu/gen/velu-sse2-rr2-p6-x4.c",
2482 "src/f32-velu/gen/velu-sse2-rr2-p6-x8.c",
2483 "src/f32-velu/gen/velu-sse2-rr2-p6-x12.c",
2484 "src/f32-velu/gen/velu-sse2-rr2-p6-x16.c",
2485 "src/f32-velu/gen/velu-sse2-rr2-p6-x20.c",
2486 "src/f32-velu/gen/velu-sse2-rr2-p6-x24.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07002487 "src/f32-vlrelu/gen/vlrelu-sse2-x4.c",
2488 "src/f32-vlrelu/gen/vlrelu-sse2-x8.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07002489 "src/f32-vrnd/gen/vrndd-sse2-x4.c",
2490 "src/f32-vrnd/gen/vrndd-sse2-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002491 "src/f32-vrnd/gen/vrndne-sse2-x4.c",
2492 "src/f32-vrnd/gen/vrndne-sse2-x8.c",
2493 "src/f32-vrnd/gen/vrndu-sse2-x4.c",
2494 "src/f32-vrnd/gen/vrndu-sse2-x8.c",
2495 "src/f32-vrnd/gen/vrndz-sse2-x4.c",
2496 "src/f32-vrnd/gen/vrndz-sse2-x8.c",
Marat Dukhanb7633f22020-11-20 16:34:56 -08002497 "src/math/exp-sse2-rr2-lut64-p2.c",
2498 "src/math/exp-sse2-rr2-p5.c",
Marat Dukhande390d42020-11-29 19:32:18 -08002499 "src/math/expm1minus-sse2-rr2-lut16-p3.c",
Marat Dukhana438aca2020-11-20 15:45:01 -08002500 "src/math/expm1minus-sse2-rr2-p6.c",
Frank Barchard3b800452020-11-22 12:12:35 -08002501 "src/math/expminus-sse2-rr2-p5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002502 "src/math/roundd-sse2-cvt.c",
2503 "src/math/roundne-sse2-cvt.c",
2504 "src/math/roundu-sse2-cvt.c",
2505 "src/math/roundz-sse2-cvt.c",
2506 "src/math/sigmoid-sse2-rr2-lut64-p2-div.c",
2507 "src/math/sigmoid-sse2-rr2-lut64-p2-nr1.c",
2508 "src/math/sigmoid-sse2-rr2-lut64-p2-nr2.c",
2509 "src/math/sigmoid-sse2-rr2-p5-div.c",
2510 "src/math/sigmoid-sse2-rr2-p5-nr1.c",
2511 "src/math/sigmoid-sse2-rr2-p5-nr2.c",
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07002512 "src/qs8-dwconv/gen/up8x9-minmax-sse2-mul16.c",
2513 "src/qs8-dwconv/gen/up16x9-minmax-sse2-mul16.c",
2514 "src/qs8-dwconv/gen/up24x9-minmax-sse2-mul16.c",
Marat Dukhan159688f2020-08-06 10:34:29 -07002515 "src/qs8-gavgpool/gen/7p7x-minmax-sse2-c8-acc2.c",
2516 "src/qs8-gavgpool/gen/7p7x-minmax-sse2-c16-acc2.c",
2517 "src/qs8-gavgpool/gen/7p7x-minmax-sse2-c24-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002518 "src/qs8-gavgpool/gen/7x-minmax-sse2-c8-acc2.c",
2519 "src/qs8-gavgpool/gen/7x-minmax-sse2-c16-acc2.c",
2520 "src/qs8-gavgpool/gen/7x-minmax-sse2-c24-acc2.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002521 "src/qs8-gemm/gen/1x4c2-minmax-sse2-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002522 "src/qs8-gemm/gen/1x4c2-minmax-sse2-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002523 "src/qs8-gemm/gen/1x4c2-xw-minmax-sse2.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002524 "src/qs8-gemm/gen/1x4c8-minmax-sse2-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002525 "src/qs8-gemm/gen/1x4c8-minmax-sse2-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002526 "src/qs8-gemm/gen/1x4c8-xw-minmax-sse2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002527 "src/qs8-gemm/gen/2x4c8-minmax-sse2-ld64.c",
2528 "src/qs8-gemm/gen/2x4c8-minmax-sse2-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002529 "src/qs8-gemm/gen/2x4c8-xw-minmax-sse2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002530 "src/qs8-gemm/gen/3x4c8-minmax-sse2-ld64.c",
2531 "src/qs8-gemm/gen/3x4c8-minmax-sse2-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002532 "src/qs8-gemm/gen/3x4c8-xw-minmax-sse2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002533 "src/qs8-gemm/gen/4x4c2-minmax-sse2-ld64.c",
2534 "src/qs8-gemm/gen/4x4c2-minmax-sse2-ld128.c",
2535 "src/qs8-gemm/gen/4x4c2-xw-minmax-sse2.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002536 "src/qs8-igemm/gen/1x4c2-minmax-sse2-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002537 "src/qs8-igemm/gen/1x4c2-minmax-sse2-ld128.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002538 "src/qs8-igemm/gen/1x4c8-minmax-sse2-ld64.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002539 "src/qs8-igemm/gen/1x4c8-minmax-sse2-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002540 "src/qs8-igemm/gen/2x4c8-minmax-sse2-ld64.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002541 "src/qs8-igemm/gen/2x4c8-minmax-sse2-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002542 "src/qs8-igemm/gen/3x4c8-minmax-sse2-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002543 "src/qs8-igemm/gen/3x4c8-minmax-sse2-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002544 "src/qs8-igemm/gen/4x4c2-minmax-sse2-ld64.c",
2545 "src/qs8-igemm/gen/4x4c2-minmax-sse2-ld128.c",
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07002546 "src/qs8-requantization/fp32-sse2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002547 "src/qs8-requantization/precise-sse2.c",
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07002548 "src/qs8-requantization/q31-sse2.c",
Marat Dukhand9f3ad42020-08-10 12:30:58 -07002549 "src/qs8-vadd/gen/minmax-sse2-mul16-ld64-x8.c",
2550 "src/qs8-vadd/gen/minmax-sse2-mul16-ld64-x16.c",
2551 "src/qs8-vadd/gen/minmax-sse2-mul16-ld64-x24.c",
2552 "src/qs8-vadd/gen/minmax-sse2-mul16-ld64-x32.c",
Marat Dukhan0270d9f2020-08-11 00:56:46 -07002553 "src/qs8-vaddc/gen/minmax-sse2-mul16-ld64-x8.c",
2554 "src/qs8-vaddc/gen/minmax-sse2-mul16-ld64-x16.c",
2555 "src/qs8-vaddc/gen/minmax-sse2-mul16-ld64-x24.c",
2556 "src/qs8-vaddc/gen/minmax-sse2-mul16-ld64-x32.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -07002557 "src/qu8-avgpool/9p8x-minmax-sse2-c8.c",
2558 "src/qu8-avgpool/9x-minmax-sse2-c8.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -07002559 "src/qu8-dwconv/up8x9-minmax-sse2.c",
2560 "src/qu8-gavgpool/7p7x-minmax-sse2-c8.c",
2561 "src/qu8-gavgpool/7x-minmax-sse2-c8.c",
2562 "src/qu8-gemm/2x4c8-minmax-sse2.c",
2563 "src/qu8-gemm/4x4c2-minmax-sse2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002564 "src/qu8-igemm/4x4c2-minmax-sse2.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07002565 "src/qu8-requantization/fp32-sse2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002566 "src/qu8-requantization/precise-sse2.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07002567 "src/qu8-requantization/q31-sse2.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -07002568 "src/qu8-vadd/minmax-sse2.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -07002569 "src/u8-clamp/sse2-x64.c",
Marat Dukhan99936602020-04-11 16:47:01 -07002570 "src/u8-maxpool/9p8x-minmax-sse2-c16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002571 "src/u8-rmax/sse2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002572 "src/x8-zip/x2-sse2.c",
2573 "src/x8-zip/x3-sse2.c",
2574 "src/x8-zip/x4-sse2.c",
2575 "src/x8-zip/xm-sse2.c",
Marat Dukhan57dccd82020-04-14 00:53:10 -07002576 "src/x32-unpool/sse2.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002577 "src/x32-zip/x2-sse2.c",
2578 "src/x32-zip/x3-sse2.c",
2579 "src/x32-zip/x4-sse2.c",
2580 "src/x32-zip/xm-sse2.c",
Marat Dukhanfe7acb62020-03-09 19:30:05 -07002581]
2582
2583SSSE3_UKERNELS = [
Frank Barchard35db7d02020-10-26 13:37:34 -07002584 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-1x4-acc2.c",
2585 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-1x4-acc3.c",
2586 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-1x4-acc4.c",
Marat Dukhan98f2eeb2020-10-23 23:13:41 -07002587 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-1x4.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07002588 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-2x4-acc2.c",
Marat Dukhan98f2eeb2020-10-23 23:13:41 -07002589 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-2x4.c",
2590 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-3x4.c",
2591 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-4x4.c",
2592 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-5x4.c",
2593 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-6x4.c",
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07002594 "src/qs8-dwconv/gen/up8x9-minmax-ssse3-mul16.c",
2595 "src/qs8-dwconv/gen/up16x9-minmax-ssse3-mul16.c",
2596 "src/qs8-dwconv/gen/up24x9-minmax-ssse3-mul16.c",
Marat Dukhan159688f2020-08-06 10:34:29 -07002597 "src/qs8-gavgpool/gen/7p7x-minmax-ssse3-c8-acc2.c",
2598 "src/qs8-gavgpool/gen/7p7x-minmax-ssse3-c16-acc2.c",
2599 "src/qs8-gavgpool/gen/7p7x-minmax-ssse3-c24-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002600 "src/qs8-gavgpool/gen/7x-minmax-ssse3-c8-acc2.c",
2601 "src/qs8-gavgpool/gen/7x-minmax-ssse3-c16-acc2.c",
2602 "src/qs8-gavgpool/gen/7x-minmax-ssse3-c24-acc2.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002603 "src/qs8-gemm/gen/1x4c2-minmax-ssse3-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002604 "src/qs8-gemm/gen/1x4c2-minmax-ssse3-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002605 "src/qs8-gemm/gen/1x4c2-xw-minmax-ssse3.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002606 "src/qs8-gemm/gen/1x4c8-minmax-ssse3-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002607 "src/qs8-gemm/gen/1x4c8-minmax-ssse3-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002608 "src/qs8-gemm/gen/1x4c8-xw-minmax-ssse3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002609 "src/qs8-gemm/gen/2x4c8-minmax-ssse3-ld64.c",
2610 "src/qs8-gemm/gen/2x4c8-minmax-ssse3-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002611 "src/qs8-gemm/gen/2x4c8-xw-minmax-ssse3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002612 "src/qs8-gemm/gen/3x4c8-minmax-ssse3-ld64.c",
2613 "src/qs8-gemm/gen/3x4c8-minmax-ssse3-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002614 "src/qs8-gemm/gen/3x4c8-xw-minmax-ssse3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002615 "src/qs8-gemm/gen/4x4c2-minmax-ssse3-ld64.c",
2616 "src/qs8-gemm/gen/4x4c2-minmax-ssse3-ld128.c",
2617 "src/qs8-gemm/gen/4x4c2-xw-minmax-ssse3.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002618 "src/qs8-igemm/gen/1x4c2-minmax-ssse3-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002619 "src/qs8-igemm/gen/1x4c2-minmax-ssse3-ld128.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002620 "src/qs8-igemm/gen/1x4c8-minmax-ssse3-ld64.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002621 "src/qs8-igemm/gen/1x4c8-minmax-ssse3-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002622 "src/qs8-igemm/gen/2x4c8-minmax-ssse3-ld64.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002623 "src/qs8-igemm/gen/2x4c8-minmax-ssse3-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002624 "src/qs8-igemm/gen/3x4c8-minmax-ssse3-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002625 "src/qs8-igemm/gen/3x4c8-minmax-ssse3-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002626 "src/qs8-igemm/gen/4x4c2-minmax-ssse3-ld64.c",
2627 "src/qs8-igemm/gen/4x4c2-minmax-ssse3-ld128.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -07002628 "src/qs8-requantization/precise-ssse3.c",
2629 "src/qs8-requantization/q31-ssse3.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07002630 "src/qu8-requantization/precise-ssse3.c",
2631 "src/qu8-requantization/q31-ssse3.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002632]
2633
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08002634SSE41_UKERNELS = [
Marat Dukhan40a672f2019-11-25 03:08:22 -08002635 "src/f32-prelu/gen/sse41-2x4.c",
2636 "src/f32-prelu/gen/sse41-2x8.c",
Marat Dukhand243c1a2020-09-17 18:48:10 -07002637 "src/f32-sigmoid/gen/sse41-lut64-p2-div-x4.c",
2638 "src/f32-sigmoid/gen/sse41-lut64-p2-div-x8.c",
2639 "src/f32-sigmoid/gen/sse41-lut64-p2-div-x12.c",
2640 "src/f32-sigmoid/gen/sse41-lut64-p2-div-x16.c",
2641 "src/f32-sigmoid/gen/sse41-lut64-p2-div-x20.c",
2642 "src/f32-sigmoid/gen/sse41-lut64-p2-div-x24.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002643 "src/f32-sigmoid/gen/sse41-p5-div-x4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08002644 "src/f32-sigmoid/gen/sse41-p5-div-x8.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002645 "src/f32-sigmoid/gen/sse41-p5-div-x12.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08002646 "src/f32-sigmoid/gen/sse41-p5-div-x16.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002647 "src/f32-sigmoid/gen/sse41-p5-div-x20.c",
2648 "src/f32-sigmoid/gen/sse41-p5-div-x24.c",
Marat Dukhaned6baaf2020-12-01 15:07:08 -08002649 "src/f32-velu/gen/velu-sse41-rr2-lut16-p3-x4.c",
2650 "src/f32-velu/gen/velu-sse41-rr2-lut16-p3-x8.c",
2651 "src/f32-velu/gen/velu-sse41-rr2-lut16-p3-x12.c",
2652 "src/f32-velu/gen/velu-sse41-rr2-lut16-p3-x16.c",
2653 "src/f32-velu/gen/velu-sse41-rr2-lut16-p3-x20.c",
2654 "src/f32-velu/gen/velu-sse41-rr2-lut16-p3-x24.c",
2655 "src/f32-velu/gen/velu-sse41-rr2-p6-x4.c",
2656 "src/f32-velu/gen/velu-sse41-rr2-p6-x8.c",
2657 "src/f32-velu/gen/velu-sse41-rr2-p6-x12.c",
2658 "src/f32-velu/gen/velu-sse41-rr2-p6-x16.c",
2659 "src/f32-velu/gen/velu-sse41-rr2-p6-x20.c",
2660 "src/f32-velu/gen/velu-sse41-rr2-p6-x24.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07002661 "src/f32-vlrelu/gen/vlrelu-sse41-x4.c",
2662 "src/f32-vlrelu/gen/vlrelu-sse41-x8.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07002663 "src/f32-vrnd/gen/vrndd-sse41-x4.c",
2664 "src/f32-vrnd/gen/vrndd-sse41-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002665 "src/f32-vrnd/gen/vrndne-sse41-x4.c",
2666 "src/f32-vrnd/gen/vrndne-sse41-x8.c",
2667 "src/f32-vrnd/gen/vrndu-sse41-x4.c",
2668 "src/f32-vrnd/gen/vrndu-sse41-x8.c",
2669 "src/f32-vrnd/gen/vrndz-sse41-x4.c",
2670 "src/f32-vrnd/gen/vrndz-sse41-x8.c",
2671 "src/math/roundd-sse41.c",
2672 "src/math/roundne-sse41.c",
2673 "src/math/roundu-sse41.c",
2674 "src/math/roundz-sse41.c",
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07002675 "src/qs8-dwconv/gen/up8x9-minmax-sse41-mul16.c",
2676 "src/qs8-dwconv/gen/up16x9-minmax-sse41-mul16.c",
2677 "src/qs8-dwconv/gen/up24x9-minmax-sse41-mul16.c",
Marat Dukhan159688f2020-08-06 10:34:29 -07002678 "src/qs8-gavgpool/gen/7p7x-minmax-sse41-c8-acc2.c",
2679 "src/qs8-gavgpool/gen/7p7x-minmax-sse41-c16-acc2.c",
2680 "src/qs8-gavgpool/gen/7p7x-minmax-sse41-c24-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002681 "src/qs8-gavgpool/gen/7x-minmax-sse41-c8-acc2.c",
2682 "src/qs8-gavgpool/gen/7x-minmax-sse41-c16-acc2.c",
2683 "src/qs8-gavgpool/gen/7x-minmax-sse41-c24-acc2.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002684 "src/qs8-gemm/gen/1x4c2-minmax-sse41-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002685 "src/qs8-gemm/gen/1x4c2-minmax-sse41-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002686 "src/qs8-gemm/gen/1x4c2-xw-minmax-sse41.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002687 "src/qs8-gemm/gen/1x4c8-minmax-sse41-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002688 "src/qs8-gemm/gen/1x4c8-minmax-sse41-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002689 "src/qs8-gemm/gen/1x4c8-xw-minmax-sse41.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002690 "src/qs8-gemm/gen/2x4c8-minmax-sse41-ld64.c",
2691 "src/qs8-gemm/gen/2x4c8-minmax-sse41-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002692 "src/qs8-gemm/gen/2x4c8-xw-minmax-sse41.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002693 "src/qs8-gemm/gen/3x4c8-minmax-sse41-ld64.c",
2694 "src/qs8-gemm/gen/3x4c8-minmax-sse41-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002695 "src/qs8-gemm/gen/3x4c8-xw-minmax-sse41.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002696 "src/qs8-gemm/gen/4x4c2-minmax-sse41-ld64.c",
2697 "src/qs8-gemm/gen/4x4c2-minmax-sse41-ld128.c",
2698 "src/qs8-gemm/gen/4x4c2-xw-minmax-sse41.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002699 "src/qs8-igemm/gen/1x4c2-minmax-sse41-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002700 "src/qs8-igemm/gen/1x4c2-minmax-sse41-ld128.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002701 "src/qs8-igemm/gen/1x4c8-minmax-sse41-ld64.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002702 "src/qs8-igemm/gen/1x4c8-minmax-sse41-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002703 "src/qs8-igemm/gen/2x4c8-minmax-sse41-ld64.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002704 "src/qs8-igemm/gen/2x4c8-minmax-sse41-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002705 "src/qs8-igemm/gen/3x4c8-minmax-sse41-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002706 "src/qs8-igemm/gen/3x4c8-minmax-sse41-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002707 "src/qs8-igemm/gen/4x4c2-minmax-sse41-ld64.c",
2708 "src/qs8-igemm/gen/4x4c2-minmax-sse41-ld128.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -07002709 "src/qs8-requantization/fp32-sse4.c",
2710 "src/qs8-requantization/precise-sse4.c",
2711 "src/qs8-requantization/q31-sse4.c",
Marat Dukhand9f3ad42020-08-10 12:30:58 -07002712 "src/qs8-vadd/gen/minmax-sse41-mul16-ld64-x8.c",
2713 "src/qs8-vadd/gen/minmax-sse41-mul16-ld64-x16.c",
2714 "src/qs8-vadd/gen/minmax-sse41-mul16-ld64-x24.c",
2715 "src/qs8-vadd/gen/minmax-sse41-mul16-ld64-x32.c",
Marat Dukhanbb9225e2020-09-06 22:40:56 -07002716 "src/qs8-vadd/gen/minmax-sse41-mul32-ld32-x8.c",
2717 "src/qs8-vadd/gen/minmax-sse41-mul32-ld32-x16.c",
2718 "src/qs8-vadd/gen/minmax-sse41-mul32-ld32-x24.c",
2719 "src/qs8-vadd/gen/minmax-sse41-mul32-ld32-x32.c",
Marat Dukhan0270d9f2020-08-11 00:56:46 -07002720 "src/qs8-vaddc/gen/minmax-sse41-mul16-ld64-x8.c",
2721 "src/qs8-vaddc/gen/minmax-sse41-mul16-ld64-x16.c",
2722 "src/qs8-vaddc/gen/minmax-sse41-mul16-ld64-x24.c",
2723 "src/qs8-vaddc/gen/minmax-sse41-mul16-ld64-x32.c",
Marat Dukhanbb9225e2020-09-06 22:40:56 -07002724 "src/qs8-vaddc/gen/minmax-sse41-mul32-ld32-x8.c",
2725 "src/qs8-vaddc/gen/minmax-sse41-mul32-ld32-x16.c",
2726 "src/qs8-vaddc/gen/minmax-sse41-mul32-ld32-x24.c",
2727 "src/qs8-vaddc/gen/minmax-sse41-mul32-ld32-x32.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07002728 "src/qu8-requantization/precise-sse4.c",
2729 "src/qu8-requantization/q31-sse4.c",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08002730]
2731
Marat Dukhan08c4a432019-10-03 09:29:21 -07002732AVX_UKERNELS = [
Marat Dukhan5c5fa962020-03-10 18:38:33 -07002733 "src/f32-clamp/gen/avx-x8.c",
2734 "src/f32-clamp/gen/avx-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002735 "src/f32-dwconv/gen/up8x4-minmax-avx-acc2.c",
2736 "src/f32-dwconv/gen/up8x4-minmax-avx.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002737 "src/f32-dwconv/gen/up8x9-minmax-avx-acc2.c",
2738 "src/f32-dwconv/gen/up8x9-minmax-avx.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002739 "src/f32-dwconv/gen/up8x25-minmax-avx-acc2.c",
2740 "src/f32-dwconv/gen/up8x25-minmax-avx.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002741 "src/f32-dwconv/gen/up16x4-minmax-avx-acc2.c",
2742 "src/f32-dwconv/gen/up16x4-minmax-avx.c",
2743 "src/f32-dwconv/gen/up16x9-minmax-avx-acc2.c",
2744 "src/f32-dwconv/gen/up16x9-minmax-avx.c",
2745 "src/f32-dwconv/gen/up16x25-minmax-avx-acc2.c",
2746 "src/f32-dwconv/gen/up16x25-minmax-avx.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002747 "src/f32-gemm/gen-inc/1x8inc-minmax-avx-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002748 "src/f32-gemm/gen-inc/1x16inc-minmax-avx-broadcast.c",
2749 "src/f32-gemm/gen-inc/3x16inc-minmax-avx-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002750 "src/f32-gemm/gen-inc/4x8inc-minmax-avx-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002751 "src/f32-gemm/gen-inc/4x16inc-minmax-avx-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002752 "src/f32-gemm/gen-inc/5x8inc-minmax-avx-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002753 "src/f32-gemm/gen-inc/5x16inc-minmax-avx-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002754 "src/f32-gemm/gen-inc/6x8inc-minmax-avx-broadcast.c",
2755 "src/f32-gemm/gen-inc/7x8inc-minmax-avx-broadcast.c",
2756 "src/f32-gemm/gen/1x8-minmax-avx-broadcast.c",
2757 "src/f32-gemm/gen/1x16-minmax-avx-broadcast.c",
2758 "src/f32-gemm/gen/3x16-minmax-avx-broadcast.c",
2759 "src/f32-gemm/gen/4x8-minmax-avx-broadcast.c",
2760 "src/f32-gemm/gen/4x16-minmax-avx-broadcast.c",
2761 "src/f32-gemm/gen/5x8-minmax-avx-broadcast.c",
2762 "src/f32-gemm/gen/5x16-minmax-avx-broadcast.c",
2763 "src/f32-gemm/gen/6x8-minmax-avx-broadcast.c",
2764 "src/f32-gemm/gen/7x8-minmax-avx-broadcast.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07002765 "src/f32-hswish/gen/hswish-avx-x8.c",
2766 "src/f32-hswish/gen/hswish-avx-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002767 "src/f32-igemm/gen/1x8-minmax-avx-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002768 "src/f32-igemm/gen/1x16-minmax-avx-broadcast.c",
2769 "src/f32-igemm/gen/3x16-minmax-avx-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002770 "src/f32-igemm/gen/4x8-minmax-avx-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002771 "src/f32-igemm/gen/4x16-minmax-avx-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002772 "src/f32-igemm/gen/5x8-minmax-avx-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002773 "src/f32-igemm/gen/5x16-minmax-avx-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002774 "src/f32-igemm/gen/6x8-minmax-avx-broadcast.c",
2775 "src/f32-igemm/gen/7x8-minmax-avx-broadcast.c",
Marat Dukhan90eca0a2020-03-11 00:52:23 -07002776 "src/f32-prelu/gen/avx-2x8.c",
2777 "src/f32-prelu/gen/avx-2x16.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07002778 "src/f32-relu/gen/avx-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002779 "src/f32-relu/gen/avx-x16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002780 "src/f32-rmax/avx.c",
T.J. Alumbaughdc2b29c2020-10-14 13:56:08 -07002781 "src/f32-sigmoid/gen/avx-rr2-p5-div-x8.c",
2782 "src/f32-sigmoid/gen/avx-rr2-p5-div-x16.c",
2783 "src/f32-sigmoid/gen/avx-rr2-p5-div-x24.c",
2784 "src/f32-sigmoid/gen/avx-rr2-p5-div-x32.c",
2785 "src/f32-sigmoid/gen/avx-rr2-p5-div-x40.c",
2786 "src/f32-sigmoid/gen/avx-rr2-p5-div-x48.c",
2787 "src/f32-sigmoid/gen/avx-rr2-p5-div-x56.c",
2788 "src/f32-sigmoid/gen/avx-rr2-p5-div-x64.c",
2789 "src/f32-sigmoid/gen/avx-rr2-p5-div-x72.c",
2790 "src/f32-sigmoid/gen/avx-rr2-p5-div-x80.c",
2791 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x8.c",
2792 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x16.c",
2793 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x24.c",
2794 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x32.c",
2795 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x40.c",
2796 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x48.c",
2797 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x56.c",
2798 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x64.c",
2799 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x72.c",
2800 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x80.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002801 "src/f32-vbinary/gen/vadd-minmax-avx-x8.c",
2802 "src/f32-vbinary/gen/vadd-minmax-avx-x16.c",
2803 "src/f32-vbinary/gen/vaddc-minmax-avx-x8.c",
2804 "src/f32-vbinary/gen/vaddc-minmax-avx-x16.c",
2805 "src/f32-vbinary/gen/vdiv-minmax-avx-x8.c",
2806 "src/f32-vbinary/gen/vdiv-minmax-avx-x16.c",
2807 "src/f32-vbinary/gen/vdivc-minmax-avx-x8.c",
2808 "src/f32-vbinary/gen/vdivc-minmax-avx-x16.c",
Marat Dukhan9a88efe2019-12-10 15:54:24 -08002809 "src/f32-vbinary/gen/vmax-avx-x8.c",
2810 "src/f32-vbinary/gen/vmax-avx-x16.c",
2811 "src/f32-vbinary/gen/vmaxc-avx-x8.c",
2812 "src/f32-vbinary/gen/vmaxc-avx-x16.c",
2813 "src/f32-vbinary/gen/vmin-avx-x8.c",
2814 "src/f32-vbinary/gen/vmin-avx-x16.c",
2815 "src/f32-vbinary/gen/vminc-avx-x8.c",
2816 "src/f32-vbinary/gen/vminc-avx-x16.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002817 "src/f32-vbinary/gen/vmul-minmax-avx-x8.c",
2818 "src/f32-vbinary/gen/vmul-minmax-avx-x16.c",
2819 "src/f32-vbinary/gen/vmulc-minmax-avx-x8.c",
2820 "src/f32-vbinary/gen/vmulc-minmax-avx-x16.c",
2821 "src/f32-vbinary/gen/vrdivc-minmax-avx-x8.c",
2822 "src/f32-vbinary/gen/vrdivc-minmax-avx-x16.c",
2823 "src/f32-vbinary/gen/vrsubc-minmax-avx-x8.c",
2824 "src/f32-vbinary/gen/vrsubc-minmax-avx-x16.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07002825 "src/f32-vbinary/gen/vsqrdiff-avx-x8.c",
2826 "src/f32-vbinary/gen/vsqrdiff-avx-x16.c",
2827 "src/f32-vbinary/gen/vsqrdiffc-avx-x8.c",
2828 "src/f32-vbinary/gen/vsqrdiffc-avx-x16.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002829 "src/f32-vbinary/gen/vsub-minmax-avx-x8.c",
2830 "src/f32-vbinary/gen/vsub-minmax-avx-x16.c",
2831 "src/f32-vbinary/gen/vsubc-minmax-avx-x8.c",
2832 "src/f32-vbinary/gen/vsubc-minmax-avx-x16.c",
Marat Dukhaned6baaf2020-12-01 15:07:08 -08002833 "src/f32-velu/gen/velu-avx-rr2-lut4-p4-perm-x8.c",
2834 "src/f32-velu/gen/velu-avx-rr2-lut4-p4-perm-x16.c",
2835 "src/f32-velu/gen/velu-avx-rr2-lut4-p4-perm-x24.c",
2836 "src/f32-velu/gen/velu-avx-rr2-lut4-p4-perm-x32.c",
2837 "src/f32-velu/gen/velu-avx-rr2-lut4-p4-perm-x40.c",
2838 "src/f32-velu/gen/velu-avx-rr2-lut4-p4-perm-x48.c",
2839 "src/f32-velu/gen/velu-avx-rr2-lut16-p3-x8.c",
2840 "src/f32-velu/gen/velu-avx-rr2-lut16-p3-x16.c",
2841 "src/f32-velu/gen/velu-avx-rr2-lut16-p3-x24.c",
2842 "src/f32-velu/gen/velu-avx-rr2-lut16-p3-x32.c",
2843 "src/f32-velu/gen/velu-avx-rr2-lut16-p3-x40.c",
2844 "src/f32-velu/gen/velu-avx-rr2-lut16-p3-x48.c",
2845 "src/f32-velu/gen/velu-avx-rr2-p6-x8.c",
2846 "src/f32-velu/gen/velu-avx-rr2-p6-x16.c",
2847 "src/f32-velu/gen/velu-avx-rr2-p6-x24.c",
2848 "src/f32-velu/gen/velu-avx-rr2-p6-x32.c",
2849 "src/f32-velu/gen/velu-avx-rr2-p6-x40.c",
2850 "src/f32-velu/gen/velu-avx-rr2-p6-x48.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07002851 "src/f32-vlrelu/gen/vlrelu-avx-x8.c",
2852 "src/f32-vlrelu/gen/vlrelu-avx-x16.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07002853 "src/f32-vrnd/gen/vrndd-avx-x8.c",
2854 "src/f32-vrnd/gen/vrndd-avx-x16.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002855 "src/f32-vrnd/gen/vrndne-avx-x8.c",
2856 "src/f32-vrnd/gen/vrndne-avx-x16.c",
2857 "src/f32-vrnd/gen/vrndu-avx-x8.c",
2858 "src/f32-vrnd/gen/vrndu-avx-x16.c",
2859 "src/f32-vrnd/gen/vrndz-avx-x8.c",
2860 "src/f32-vrnd/gen/vrndz-avx-x16.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07002861 "src/f32-vscale/avx-x32.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07002862 "src/f32-vsqrt/gen/avx-sqrt-x8.c",
2863 "src/f32-vsqrt/gen/avx-sqrt-x16.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07002864 "src/f32-vunary/gen/vabs-avx-x8.c",
2865 "src/f32-vunary/gen/vabs-avx-x16.c",
2866 "src/f32-vunary/gen/vneg-avx-x8.c",
2867 "src/f32-vunary/gen/vneg-avx-x16.c",
2868 "src/f32-vunary/gen/vsqr-avx-x8.c",
2869 "src/f32-vunary/gen/vsqr-avx-x16.c",
Marat Dukhanb7633f22020-11-20 16:34:56 -08002870 "src/math/exp-avx-rr2-p5.c",
Marat Dukhande390d42020-11-29 19:32:18 -08002871 "src/math/expm1minus-avx-rr2-lut4-p4-perm.c",
2872 "src/math/expm1minus-avx-rr2-lut16-p3.c",
2873 "src/math/expm1minus-avx-rr2-p6.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002874 "src/math/sigmoid-avx-rr2-lut64-p2-div.c",
Marat Dukhan2e4727b2020-09-24 00:48:39 -07002875 "src/math/sigmoid-avx-rr2-p5-div.c",
Marat Dukhan39ab16f2020-10-16 14:30:35 -07002876 "src/math/sigmoid-avx-rr2-p5-nr1.c",
2877 "src/math/sigmoid-avx-rr2-p5-nr2.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002878]
2879
Marat Dukhan1566fee2020-08-02 21:55:41 -07002880XOP_UKERNELS = [
2881 "src/qs8-gemm/gen/1x4c2-minmax-xop-ld64.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002882 "src/qs8-gemm/gen/1x4c2-minmax-xop-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002883 "src/qs8-gemm/gen/1x4c2-xw-minmax-xop.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002884 "src/qs8-gemm/gen/1x4c8-minmax-xop-ld64.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002885 "src/qs8-gemm/gen/1x4c8-minmax-xop-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002886 "src/qs8-gemm/gen/1x4c8-xw-minmax-xop.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002887 "src/qs8-gemm/gen/2x4c8-minmax-xop-ld64.c",
2888 "src/qs8-gemm/gen/2x4c8-minmax-xop-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002889 "src/qs8-gemm/gen/2x4c8-xw-minmax-xop.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002890 "src/qs8-gemm/gen/3x4c8-minmax-xop-ld64.c",
2891 "src/qs8-gemm/gen/3x4c8-minmax-xop-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002892 "src/qs8-gemm/gen/3x4c8-xw-minmax-xop.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002893 "src/qs8-gemm/gen/4x4c2-minmax-xop-ld64.c",
2894 "src/qs8-gemm/gen/4x4c2-minmax-xop-ld128.c",
2895 "src/qs8-gemm/gen/4x4c2-xw-minmax-xop.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002896 "src/qs8-igemm/gen/1x4c2-minmax-xop-ld64.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002897 "src/qs8-igemm/gen/1x4c2-minmax-xop-ld128.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002898 "src/qs8-igemm/gen/1x4c8-minmax-xop-ld64.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002899 "src/qs8-igemm/gen/1x4c8-minmax-xop-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002900 "src/qs8-igemm/gen/2x4c8-minmax-xop-ld64.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002901 "src/qs8-igemm/gen/2x4c8-minmax-xop-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002902 "src/qs8-igemm/gen/3x4c8-minmax-xop-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002903 "src/qs8-igemm/gen/3x4c8-minmax-xop-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002904 "src/qs8-igemm/gen/4x4c2-minmax-xop-ld64.c",
2905 "src/qs8-igemm/gen/4x4c2-minmax-xop-ld128.c",
Marat Dukhanbb9225e2020-09-06 22:40:56 -07002906 "src/qs8-vadd/gen/minmax-xop-mul32-ld32-x8.c",
2907 "src/qs8-vadd/gen/minmax-xop-mul32-ld32-x16.c",
2908 "src/qs8-vadd/gen/minmax-xop-mul32-ld32-x24.c",
2909 "src/qs8-vadd/gen/minmax-xop-mul32-ld32-x32.c",
2910 "src/qs8-vaddc/gen/minmax-xop-mul32-ld32-x8.c",
2911 "src/qs8-vaddc/gen/minmax-xop-mul32-ld32-x16.c",
2912 "src/qs8-vaddc/gen/minmax-xop-mul32-ld32-x24.c",
2913 "src/qs8-vaddc/gen/minmax-xop-mul32-ld32-x32.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002914]
2915
Marat Dukhanfda12b82019-11-21 12:27:59 -08002916FMA3_UKERNELS = [
Marat Dukhan1c587112020-04-08 20:04:28 -07002917 "src/f32-dwconv/gen/up8x4-minmax-fma3-acc2.c",
2918 "src/f32-dwconv/gen/up8x4-minmax-fma3.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002919 "src/f32-dwconv/gen/up8x9-minmax-fma3-acc2.c",
2920 "src/f32-dwconv/gen/up8x9-minmax-fma3.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002921 "src/f32-dwconv/gen/up8x25-minmax-fma3-acc2.c",
2922 "src/f32-dwconv/gen/up8x25-minmax-fma3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002923 "src/f32-dwconv/gen/up16x4-minmax-fma3-acc2.c",
2924 "src/f32-dwconv/gen/up16x4-minmax-fma3.c",
2925 "src/f32-dwconv/gen/up16x9-minmax-fma3-acc2.c",
2926 "src/f32-dwconv/gen/up16x9-minmax-fma3.c",
2927 "src/f32-dwconv/gen/up16x25-minmax-fma3-acc2.c",
2928 "src/f32-dwconv/gen/up16x25-minmax-fma3.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002929 "src/f32-gemm/gen-inc/1x8inc-minmax-fma3-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002930 "src/f32-gemm/gen-inc/1x16inc-minmax-fma3-broadcast.c",
2931 "src/f32-gemm/gen-inc/1x16s4inc-minmax-fma3-broadcast.c",
2932 "src/f32-gemm/gen-inc/3x16inc-minmax-fma3-broadcast.c",
2933 "src/f32-gemm/gen-inc/3x16s4inc-minmax-fma3-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002934 "src/f32-gemm/gen-inc/4x8inc-minmax-fma3-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002935 "src/f32-gemm/gen-inc/4x16inc-minmax-fma3-broadcast.c",
2936 "src/f32-gemm/gen-inc/4x16s4inc-minmax-fma3-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002937 "src/f32-gemm/gen-inc/5x8inc-minmax-fma3-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002938 "src/f32-gemm/gen-inc/5x16inc-minmax-fma3-broadcast.c",
2939 "src/f32-gemm/gen-inc/5x16s4inc-minmax-fma3-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002940 "src/f32-gemm/gen-inc/6x8inc-minmax-fma3-broadcast.c",
2941 "src/f32-gemm/gen-inc/7x8inc-minmax-fma3-broadcast.c",
2942 "src/f32-gemm/gen-inc/8x8inc-minmax-fma3-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002943 "src/f32-gemm/gen/1x8-minmax-fma3-broadcast.c",
2944 "src/f32-gemm/gen/1x16-minmax-fma3-broadcast.c",
2945 "src/f32-gemm/gen/1x16s4-minmax-fma3-broadcast.c",
2946 "src/f32-gemm/gen/3x16-minmax-fma3-broadcast.c",
2947 "src/f32-gemm/gen/3x16s4-minmax-fma3-broadcast.c",
2948 "src/f32-gemm/gen/4x8-minmax-fma3-broadcast.c",
2949 "src/f32-gemm/gen/4x16-minmax-fma3-broadcast.c",
2950 "src/f32-gemm/gen/4x16s4-minmax-fma3-broadcast.c",
2951 "src/f32-gemm/gen/5x8-minmax-fma3-broadcast.c",
2952 "src/f32-gemm/gen/5x16-minmax-fma3-broadcast.c",
2953 "src/f32-gemm/gen/5x16s4-minmax-fma3-broadcast.c",
2954 "src/f32-gemm/gen/6x8-minmax-fma3-broadcast.c",
2955 "src/f32-gemm/gen/7x8-minmax-fma3-broadcast.c",
2956 "src/f32-gemm/gen/8x8-minmax-fma3-broadcast.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07002957 "src/f32-hswish/gen/hswish-fma3-x8.c",
2958 "src/f32-hswish/gen/hswish-fma3-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002959 "src/f32-igemm/gen/1x8-minmax-fma3-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002960 "src/f32-igemm/gen/1x16-minmax-fma3-broadcast.c",
2961 "src/f32-igemm/gen/1x16s4-minmax-fma3-broadcast.c",
2962 "src/f32-igemm/gen/3x16-minmax-fma3-broadcast.c",
2963 "src/f32-igemm/gen/3x16s4-minmax-fma3-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002964 "src/f32-igemm/gen/4x8-minmax-fma3-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002965 "src/f32-igemm/gen/4x16-minmax-fma3-broadcast.c",
2966 "src/f32-igemm/gen/4x16s4-minmax-fma3-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002967 "src/f32-igemm/gen/5x8-minmax-fma3-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002968 "src/f32-igemm/gen/5x16-minmax-fma3-broadcast.c",
2969 "src/f32-igemm/gen/5x16s4-minmax-fma3-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002970 "src/f32-igemm/gen/6x8-minmax-fma3-broadcast.c",
2971 "src/f32-igemm/gen/7x8-minmax-fma3-broadcast.c",
2972 "src/f32-igemm/gen/8x8-minmax-fma3-broadcast.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07002973 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x8.c",
2974 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x16.c",
2975 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x24.c",
2976 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x32.c",
2977 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x40.c",
2978 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x48.c",
2979 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x56.c",
2980 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x64.c",
Marat Dukhan84000762020-06-29 18:38:43 -07002981 "src/math/sqrt-fma3-nr1fma.c",
Marat Dukhan84000762020-06-29 18:38:43 -07002982 "src/math/sqrt-fma3-nr1fma1adj.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002983 "src/math/sqrt-fma3-nr2fma.c",
Marat Dukhanfda12b82019-11-21 12:27:59 -08002984]
2985
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002986AVX2_UKERNELS = [
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002987 "src/f32-raddexpminusmax/gen/avx2-p5-x64-acc2.c",
2988 "src/f32-raddexpminusmax/gen/avx2-p5-x64-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002989 "src/f32-raddexpminusmax/gen/avx2-p5-x64.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002990 "src/f32-raddexpminusmax/gen/avx2-p5-x72-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002991 "src/f32-raddexpminusmax/gen/avx2-p5-x72.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002992 "src/f32-raddexpminusmax/gen/avx2-p5-x80-acc2.c",
2993 "src/f32-raddexpminusmax/gen/avx2-p5-x80-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002994 "src/f32-raddexpminusmax/gen/avx2-p5-x80.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002995 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc2.c",
2996 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc3.c",
2997 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc6.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002998 "src/f32-raddexpminusmax/gen/avx2-p5-x96.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002999 "src/f32-raddextexp/gen/avx2-p5-x64-acc2.c",
3000 "src/f32-raddextexp/gen/avx2-p5-x64-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003001 "src/f32-raddextexp/gen/avx2-p5-x64.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003002 "src/f32-raddextexp/gen/avx2-p5-x72-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003003 "src/f32-raddextexp/gen/avx2-p5-x72.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003004 "src/f32-raddextexp/gen/avx2-p5-x80-acc2.c",
3005 "src/f32-raddextexp/gen/avx2-p5-x80-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003006 "src/f32-raddextexp/gen/avx2-p5-x80.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003007 "src/f32-raddextexp/gen/avx2-p5-x96-acc2.c",
3008 "src/f32-raddextexp/gen/avx2-p5-x96-acc3.c",
3009 "src/f32-raddextexp/gen/avx2-p5-x96-acc6.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003010 "src/f32-raddextexp/gen/avx2-p5-x96.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003011 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64-acc2.c",
3012 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003013 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003014 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x72-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003015 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x72.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003016 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80-acc2.c",
3017 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003018 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003019 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc2.c",
3020 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc3.c",
3021 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc6.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003022 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96.c",
Marat Dukhanfa0a4322020-01-06 16:14:29 -08003023 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x8.c",
3024 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x16.c",
3025 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x24.c",
3026 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x32.c",
3027 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x40.c",
3028 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x48.c",
3029 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x56.c",
3030 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x64.c",
3031 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x72.c",
3032 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x80.c",
3033 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x8.c",
3034 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x16.c",
3035 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x24.c",
3036 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x32.c",
3037 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x40.c",
3038 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x48.c",
3039 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x56.c",
3040 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x64.c",
3041 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x72.c",
3042 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x80.c",
3043 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x8.c",
3044 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x16.c",
3045 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x24.c",
3046 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x32.c",
3047 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x40.c",
3048 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x48.c",
3049 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x56.c",
3050 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x64.c",
3051 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x72.c",
3052 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x80.c",
Marat Dukhaned6baaf2020-12-01 15:07:08 -08003053 "src/f32-velu/gen/velu-avx2-rr1-lut4-p4-perm-x8.c",
3054 "src/f32-velu/gen/velu-avx2-rr1-lut4-p4-perm-x16.c",
3055 "src/f32-velu/gen/velu-avx2-rr1-lut4-p4-perm-x24.c",
3056 "src/f32-velu/gen/velu-avx2-rr1-lut4-p4-perm-x32.c",
3057 "src/f32-velu/gen/velu-avx2-rr1-lut4-p4-perm-x40.c",
3058 "src/f32-velu/gen/velu-avx2-rr1-lut4-p4-perm-x48.c",
3059 "src/f32-velu/gen/velu-avx2-rr1-lut4-p4-perm-x56.c",
3060 "src/f32-velu/gen/velu-avx2-rr1-lut4-p4-perm-x64.c",
3061 "src/f32-velu/gen/velu-avx2-rr1-lut4-p4-perm-x72.c",
3062 "src/f32-velu/gen/velu-avx2-rr1-lut4-p4-perm-x80.c",
3063 "src/f32-velu/gen/velu-avx2-rr1-lut8-p4-perm-x8.c",
3064 "src/f32-velu/gen/velu-avx2-rr1-lut8-p4-perm-x16.c",
3065 "src/f32-velu/gen/velu-avx2-rr1-lut8-p4-perm-x24.c",
3066 "src/f32-velu/gen/velu-avx2-rr1-lut8-p4-perm-x32.c",
3067 "src/f32-velu/gen/velu-avx2-rr1-lut8-p4-perm-x40.c",
3068 "src/f32-velu/gen/velu-avx2-rr1-lut8-p4-perm-x48.c",
3069 "src/f32-velu/gen/velu-avx2-rr1-lut8-p4-perm-x56.c",
3070 "src/f32-velu/gen/velu-avx2-rr1-lut8-p4-perm-x64.c",
3071 "src/f32-velu/gen/velu-avx2-rr1-lut8-p4-perm-x72.c",
3072 "src/f32-velu/gen/velu-avx2-rr1-lut8-p4-perm-x80.c",
3073 "src/f32-velu/gen/velu-avx2-rr1-lut16-p3-gather-x8.c",
3074 "src/f32-velu/gen/velu-avx2-rr1-lut16-p3-gather-x16.c",
3075 "src/f32-velu/gen/velu-avx2-rr1-lut16-p3-gather-x24.c",
3076 "src/f32-velu/gen/velu-avx2-rr1-lut16-p3-gather-x32.c",
3077 "src/f32-velu/gen/velu-avx2-rr1-lut16-p3-gather-x40.c",
3078 "src/f32-velu/gen/velu-avx2-rr1-lut16-p3-gather-x48.c",
3079 "src/f32-velu/gen/velu-avx2-rr1-lut16-p3-gather-x56.c",
3080 "src/f32-velu/gen/velu-avx2-rr1-lut16-p3-gather-x64.c",
3081 "src/f32-velu/gen/velu-avx2-rr1-lut16-p3-gather-x72.c",
3082 "src/f32-velu/gen/velu-avx2-rr1-lut16-p3-gather-x80.c",
3083 "src/f32-velu/gen/velu-avx2-rr1-p6-x8.c",
3084 "src/f32-velu/gen/velu-avx2-rr1-p6-x16.c",
3085 "src/f32-velu/gen/velu-avx2-rr1-p6-x24.c",
3086 "src/f32-velu/gen/velu-avx2-rr1-p6-x32.c",
3087 "src/f32-velu/gen/velu-avx2-rr1-p6-x40.c",
3088 "src/f32-velu/gen/velu-avx2-rr1-p6-x48.c",
3089 "src/f32-velu/gen/velu-avx2-rr1-p6-x56.c",
3090 "src/f32-velu/gen/velu-avx2-rr1-p6-x64.c",
3091 "src/f32-velu/gen/velu-avx2-rr1-p6-x72.c",
3092 "src/f32-velu/gen/velu-avx2-rr1-p6-x80.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003093 "src/f32-vscaleexpminusmax/gen/avx2-p5-x8.c",
3094 "src/f32-vscaleexpminusmax/gen/avx2-p5-x16.c",
3095 "src/f32-vscaleexpminusmax/gen/avx2-p5-x24.c",
3096 "src/f32-vscaleexpminusmax/gen/avx2-p5-x32.c",
3097 "src/f32-vscaleexpminusmax/gen/avx2-p5-x40.c",
3098 "src/f32-vscaleexpminusmax/gen/avx2-p5-x48.c",
3099 "src/f32-vscaleexpminusmax/gen/avx2-p5-x56.c",
3100 "src/f32-vscaleexpminusmax/gen/avx2-p5-x64.c",
3101 "src/f32-vscaleexpminusmax/gen/avx2-p5-x72.c",
3102 "src/f32-vscaleexpminusmax/gen/avx2-p5-x80.c",
3103 "src/f32-vscaleexpminusmax/gen/avx2-p5-x88.c",
3104 "src/f32-vscaleexpminusmax/gen/avx2-p5-x96.c",
3105 "src/f32-vscaleextexp/gen/avx2-p5-x8.c",
3106 "src/f32-vscaleextexp/gen/avx2-p5-x16.c",
3107 "src/f32-vscaleextexp/gen/avx2-p5-x24.c",
3108 "src/f32-vscaleextexp/gen/avx2-p5-x32.c",
3109 "src/f32-vscaleextexp/gen/avx2-p5-x40.c",
3110 "src/f32-vscaleextexp/gen/avx2-p5-x48.c",
3111 "src/f32-vscaleextexp/gen/avx2-p5-x56.c",
3112 "src/f32-vscaleextexp/gen/avx2-p5-x64.c",
3113 "src/f32-vscaleextexp/gen/avx2-p5-x72.c",
3114 "src/f32-vscaleextexp/gen/avx2-p5-x80.c",
3115 "src/f32-vscaleextexp/gen/avx2-p5-x88.c",
3116 "src/f32-vscaleextexp/gen/avx2-p5-x96.c",
Marat Dukhanb7633f22020-11-20 16:34:56 -08003117 "src/math/exp-avx2-rr2-lut8-p3-perm.c",
3118 "src/math/exp-avx2-rr2-lut8-p4-perm.c",
3119 "src/math/exp-avx2-rr2-p5.c",
Marat Dukhande390d42020-11-29 19:32:18 -08003120 "src/math/expm1minus-avx2-rr1-lut4-p4-perm.c",
3121 "src/math/expm1minus-avx2-rr1-lut8-p4-perm.c",
3122 "src/math/expm1minus-avx2-rr1-lut16-p3-gather.c",
3123 "src/math/expm1minus-avx2-rr1-p6.c",
Frank Barcharde7223ee2020-12-04 19:04:01 -08003124 "src/math/expminus-avx2-rr2-p5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003125 "src/math/extexp-avx2-p5.c",
3126 "src/math/sigmoid-avx2-rr1-lut64-p2-gather-div.c",
3127 "src/math/sigmoid-avx2-rr1-lut64-p2-gather-nr1fma.c",
3128 "src/math/sigmoid-avx2-rr1-lut64-p2-gather-nr2fma.c",
3129 "src/math/sigmoid-avx2-rr1-lut64-p2-gather-nr2fma1adj.c",
3130 "src/math/sigmoid-avx2-rr1-p5-div.c",
3131 "src/math/sigmoid-avx2-rr1-p5-nr1fma.c",
3132 "src/math/sigmoid-avx2-rr1-p5-nr2fma.c",
3133 "src/math/sigmoid-avx2-rr2-lut64-p2-gather-div.c",
3134 "src/math/sigmoid-avx2-rr2-lut64-p2-gather-nr1fma.c",
3135 "src/math/sigmoid-avx2-rr2-lut64-p2-gather-nr2fma.c",
3136 "src/math/sigmoid-avx2-rr2-lut64-p2-gather-nr2fma1adj.c",
3137 "src/math/sigmoid-avx2-rr2-p5-div.c",
3138 "src/math/sigmoid-avx2-rr2-p5-nr1fma.c",
3139 "src/math/sigmoid-avx2-rr2-p5-nr2fma.c",
Marat Dukhand65a1522020-08-04 19:28:18 -07003140 "src/qs8-dwconv/gen/up8x9-minmax-avx2-mul32.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003141 "src/qs8-dwconv/gen/up16x9-minmax-avx2-mul16.c",
Marat Dukhand65a1522020-08-04 19:28:18 -07003142 "src/qs8-dwconv/gen/up16x9-minmax-avx2-mul32.c",
3143 "src/qs8-dwconv/gen/up24x9-minmax-avx2-mul32.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003144 "src/qs8-dwconv/gen/up32x9-minmax-avx2-mul16.c",
Marat Dukhand65a1522020-08-04 19:28:18 -07003145 "src/qs8-dwconv/gen/up32x9-minmax-avx2-mul32.c",
Marat Dukhan12809522020-08-02 22:23:51 -07003146 "src/qs8-gemm/gen/1x8c8-minmax-avx2.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07003147 "src/qs8-gemm/gen/1x8c8-xw-minmax-avx2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003148 "src/qs8-gemm/gen/2x8c8-minmax-avx2.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07003149 "src/qs8-gemm/gen/2x8c8-xw-minmax-avx2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003150 "src/qs8-gemm/gen/3x8c8-minmax-avx2.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07003151 "src/qs8-gemm/gen/3x8c8-xw-minmax-avx2.c",
Marat Dukhan12809522020-08-02 22:23:51 -07003152 "src/qs8-igemm/gen/1x8c8-minmax-avx2.c",
3153 "src/qs8-igemm/gen/2x8c8-minmax-avx2.c",
3154 "src/qs8-igemm/gen/3x8c8-minmax-avx2.c",
Marat Dukhane6dc0b62020-09-08 23:57:14 -07003155 "src/qs8-vadd/gen/minmax-avx2-mul32-ld64-x8.c",
3156 "src/qs8-vadd/gen/minmax-avx2-mul32-ld64-x16.c",
3157 "src/qs8-vadd/gen/minmax-avx2-mul32-ld64-x24.c",
3158 "src/qs8-vadd/gen/minmax-avx2-mul32-ld64-x32.c",
3159 "src/qs8-vaddc/gen/minmax-avx2-mul32-ld64-x8.c",
3160 "src/qs8-vaddc/gen/minmax-avx2-mul32-ld64-x16.c",
3161 "src/qs8-vaddc/gen/minmax-avx2-mul32-ld64-x24.c",
3162 "src/qs8-vaddc/gen/minmax-avx2-mul32-ld64-x32.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003163]
3164
Marat Dukhan08c4a432019-10-03 09:29:21 -07003165AVX512F_UKERNELS = [
Marat Dukhan5c5fa962020-03-10 18:38:33 -07003166 "src/f32-clamp/gen/avx512f-x16.c",
3167 "src/f32-clamp/gen/avx512f-x32.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07003168 "src/f32-dwconv/gen/up16x4-minmax-avx512f-acc2.c",
3169 "src/f32-dwconv/gen/up16x4-minmax-avx512f.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07003170 "src/f32-dwconv/gen/up16x9-minmax-avx512f-acc2.c",
3171 "src/f32-dwconv/gen/up16x9-minmax-avx512f.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07003172 "src/f32-dwconv/gen/up16x25-minmax-avx512f-acc2.c",
3173 "src/f32-dwconv/gen/up16x25-minmax-avx512f.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003174 "src/f32-dwconv/gen/up32x4-minmax-avx512f-acc2.c",
3175 "src/f32-dwconv/gen/up32x4-minmax-avx512f.c",
3176 "src/f32-dwconv/gen/up32x9-minmax-avx512f-acc2.c",
3177 "src/f32-dwconv/gen/up32x9-minmax-avx512f.c",
3178 "src/f32-dwconv/gen/up32x25-minmax-avx512f-acc2.c",
3179 "src/f32-dwconv/gen/up32x25-minmax-avx512f.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07003180 "src/f32-gemm/gen-inc/1x16inc-minmax-avx512f-broadcast.c",
3181 "src/f32-gemm/gen-inc/4x16inc-minmax-avx512f-broadcast.c",
3182 "src/f32-gemm/gen-inc/5x16inc-minmax-avx512f-broadcast.c",
3183 "src/f32-gemm/gen-inc/6x16inc-minmax-avx512f-broadcast.c",
3184 "src/f32-gemm/gen-inc/7x16inc-minmax-avx512f-broadcast.c",
3185 "src/f32-gemm/gen-inc/8x16inc-minmax-avx512f-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003186 "src/f32-gemm/gen/1x16-minmax-avx512f-broadcast.c",
3187 "src/f32-gemm/gen/4x16-minmax-avx512f-broadcast.c",
3188 "src/f32-gemm/gen/5x16-minmax-avx512f-broadcast.c",
3189 "src/f32-gemm/gen/6x16-minmax-avx512f-broadcast.c",
3190 "src/f32-gemm/gen/7x16-minmax-avx512f-broadcast.c",
3191 "src/f32-gemm/gen/8x16-minmax-avx512f-broadcast.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07003192 "src/f32-hswish/gen/hswish-avx512f-x16.c",
3193 "src/f32-hswish/gen/hswish-avx512f-x32.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07003194 "src/f32-igemm/gen/1x16-minmax-avx512f-broadcast.c",
3195 "src/f32-igemm/gen/4x16-minmax-avx512f-broadcast.c",
3196 "src/f32-igemm/gen/5x16-minmax-avx512f-broadcast.c",
3197 "src/f32-igemm/gen/6x16-minmax-avx512f-broadcast.c",
3198 "src/f32-igemm/gen/7x16-minmax-avx512f-broadcast.c",
3199 "src/f32-igemm/gen/8x16-minmax-avx512f-broadcast.c",
Marat Dukhan90eca0a2020-03-11 00:52:23 -07003200 "src/f32-prelu/gen/avx512f-2x16.c",
3201 "src/f32-prelu/gen/avx512f-2x32.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003202 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128-acc2.c",
3203 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003204 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003205 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x144-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003206 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x144.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003207 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160-acc2.c",
3208 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003209 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003210 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc2.c",
3211 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc3.c",
3212 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc6.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003213 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003214 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128-acc2.c",
3215 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003216 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003217 "src/f32-raddextexp/gen/avx512f-p5-scalef-x144-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003218 "src/f32-raddextexp/gen/avx512f-p5-scalef-x144.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003219 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160-acc2.c",
3220 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003221 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003222 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc2.c",
3223 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc3.c",
3224 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc6.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003225 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003226 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128-acc2.c",
3227 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003228 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003229 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x144-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003230 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x144.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003231 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160-acc2.c",
3232 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003233 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003234 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc2.c",
3235 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc3.c",
3236 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc6.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003237 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07003238 "src/f32-relu/gen/avx512f-x16.c",
3239 "src/f32-relu/gen/avx512f-x32.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003240 "src/f32-rmax/avx512f.c",
Marat Dukhand9ca7e62020-09-23 23:45:29 -07003241 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x16.c",
3242 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x32.c",
3243 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x48.c",
3244 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x64.c",
3245 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x80.c",
3246 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x96.c",
3247 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x112.c",
3248 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x128.c",
3249 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x16.c",
3250 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x32.c",
3251 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x48.c",
3252 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x64.c",
3253 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x80.c",
3254 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x96.c",
3255 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x112.c",
3256 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003257 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x16.c",
3258 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x32.c",
3259 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x48.c",
3260 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x64.c",
3261 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x80.c",
3262 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x96.c",
3263 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x112.c",
3264 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x128.c",
3265 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x16.c",
3266 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x32.c",
3267 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x48.c",
3268 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x64.c",
3269 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x80.c",
3270 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x96.c",
3271 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x112.c",
3272 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x128.c",
Marat Dukhand9ca7e62020-09-23 23:45:29 -07003273 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x16.c",
3274 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x32.c",
3275 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x48.c",
3276 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x64.c",
3277 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x80.c",
3278 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x96.c",
3279 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x112.c",
3280 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x128.c",
3281 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x16.c",
3282 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x32.c",
3283 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x48.c",
3284 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x64.c",
3285 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x80.c",
3286 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x96.c",
3287 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x112.c",
3288 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x128.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07003289 "src/f32-vbinary/gen/vadd-minmax-avx512f-x16.c",
3290 "src/f32-vbinary/gen/vadd-minmax-avx512f-x32.c",
3291 "src/f32-vbinary/gen/vaddc-minmax-avx512f-x16.c",
3292 "src/f32-vbinary/gen/vaddc-minmax-avx512f-x32.c",
3293 "src/f32-vbinary/gen/vdiv-minmax-avx512f-x16.c",
3294 "src/f32-vbinary/gen/vdiv-minmax-avx512f-x32.c",
3295 "src/f32-vbinary/gen/vdivc-minmax-avx512f-x16.c",
3296 "src/f32-vbinary/gen/vdivc-minmax-avx512f-x32.c",
Marat Dukhan9a88efe2019-12-10 15:54:24 -08003297 "src/f32-vbinary/gen/vmax-avx512f-x16.c",
3298 "src/f32-vbinary/gen/vmax-avx512f-x32.c",
3299 "src/f32-vbinary/gen/vmaxc-avx512f-x16.c",
3300 "src/f32-vbinary/gen/vmaxc-avx512f-x32.c",
3301 "src/f32-vbinary/gen/vmin-avx512f-x16.c",
3302 "src/f32-vbinary/gen/vmin-avx512f-x32.c",
3303 "src/f32-vbinary/gen/vminc-avx512f-x16.c",
3304 "src/f32-vbinary/gen/vminc-avx512f-x32.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07003305 "src/f32-vbinary/gen/vmul-minmax-avx512f-x16.c",
3306 "src/f32-vbinary/gen/vmul-minmax-avx512f-x32.c",
3307 "src/f32-vbinary/gen/vmulc-minmax-avx512f-x16.c",
3308 "src/f32-vbinary/gen/vmulc-minmax-avx512f-x32.c",
3309 "src/f32-vbinary/gen/vrdivc-minmax-avx512f-x16.c",
3310 "src/f32-vbinary/gen/vrdivc-minmax-avx512f-x32.c",
3311 "src/f32-vbinary/gen/vrsubc-minmax-avx512f-x16.c",
3312 "src/f32-vbinary/gen/vrsubc-minmax-avx512f-x32.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07003313 "src/f32-vbinary/gen/vsqrdiff-avx512f-x16.c",
3314 "src/f32-vbinary/gen/vsqrdiff-avx512f-x32.c",
3315 "src/f32-vbinary/gen/vsqrdiffc-avx512f-x16.c",
3316 "src/f32-vbinary/gen/vsqrdiffc-avx512f-x32.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07003317 "src/f32-vbinary/gen/vsub-minmax-avx512f-x16.c",
3318 "src/f32-vbinary/gen/vsub-minmax-avx512f-x32.c",
3319 "src/f32-vbinary/gen/vsubc-minmax-avx512f-x16.c",
3320 "src/f32-vbinary/gen/vsubc-minmax-avx512f-x32.c",
Marat Dukhaned6baaf2020-12-01 15:07:08 -08003321 "src/f32-velu/gen/velu-avx512f-rr1-lut16-p3-perm-x16.c",
3322 "src/f32-velu/gen/velu-avx512f-rr1-lut16-p3-perm-x32.c",
3323 "src/f32-velu/gen/velu-avx512f-rr1-lut16-p3-perm-x48.c",
3324 "src/f32-velu/gen/velu-avx512f-rr1-lut16-p3-perm-x64.c",
3325 "src/f32-velu/gen/velu-avx512f-rr1-lut16-p3-perm-x80.c",
3326 "src/f32-velu/gen/velu-avx512f-rr1-lut16-p3-perm-x96.c",
3327 "src/f32-velu/gen/velu-avx512f-rr1-lut16-p3-perm-x112.c",
3328 "src/f32-velu/gen/velu-avx512f-rr1-lut16-p3-perm-x128.c",
3329 "src/f32-velu/gen/velu-avx512f-rr1-p6-x16.c",
3330 "src/f32-velu/gen/velu-avx512f-rr1-p6-x32.c",
3331 "src/f32-velu/gen/velu-avx512f-rr1-p6-x48.c",
3332 "src/f32-velu/gen/velu-avx512f-rr1-p6-x64.c",
3333 "src/f32-velu/gen/velu-avx512f-rr1-p6-x80.c",
3334 "src/f32-velu/gen/velu-avx512f-rr1-p6-x96.c",
3335 "src/f32-velu/gen/velu-avx512f-rr1-p6-x112.c",
3336 "src/f32-velu/gen/velu-avx512f-rr1-p6-x128.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07003337 "src/f32-vlrelu/gen/vlrelu-avx512f-x16.c",
3338 "src/f32-vlrelu/gen/vlrelu-avx512f-x32.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003339 "src/f32-vrnd/gen/vrndd-avx512f-x16.c",
3340 "src/f32-vrnd/gen/vrndd-avx512f-x32.c",
3341 "src/f32-vrnd/gen/vrndne-avx512f-x16.c",
3342 "src/f32-vrnd/gen/vrndne-avx512f-x32.c",
3343 "src/f32-vrnd/gen/vrndu-avx512f-x16.c",
3344 "src/f32-vrnd/gen/vrndu-avx512f-x32.c",
3345 "src/f32-vrnd/gen/vrndz-avx512f-x16.c",
3346 "src/f32-vrnd/gen/vrndz-avx512f-x32.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07003347 "src/f32-vscale/avx512f-x64.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003348 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x16.c",
3349 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x32.c",
3350 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x48.c",
3351 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x64.c",
3352 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x80.c",
3353 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x96.c",
3354 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x112.c",
3355 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x128.c",
3356 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x144.c",
3357 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x160.c",
3358 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x176.c",
3359 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x192.c",
3360 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x16.c",
3361 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x32.c",
3362 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x48.c",
3363 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x64.c",
3364 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x80.c",
3365 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x96.c",
3366 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x112.c",
3367 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x128.c",
3368 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x144.c",
3369 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x160.c",
3370 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x176.c",
3371 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x192.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07003372 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x16.c",
3373 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x32.c",
3374 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x48.c",
3375 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x64.c",
3376 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x80.c",
3377 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x96.c",
3378 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x112.c",
3379 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x128.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07003380 "src/f32-vunary/gen/vabs-avx512f-x16.c",
3381 "src/f32-vunary/gen/vabs-avx512f-x32.c",
3382 "src/f32-vunary/gen/vneg-avx512f-x16.c",
3383 "src/f32-vunary/gen/vneg-avx512f-x32.c",
3384 "src/f32-vunary/gen/vsqr-avx512f-x16.c",
3385 "src/f32-vunary/gen/vsqr-avx512f-x32.c",
Marat Dukhanb7633f22020-11-20 16:34:56 -08003386 "src/math/exp-avx512f-rr2-lut16-p3-perm-scalef.c",
3387 "src/math/exp-avx512f-rr2-lut16-p3-perm.c",
3388 "src/math/exp-avx512f-rr2-lut32-p2-perm2-scalef.c",
3389 "src/math/exp-avx512f-rr2-lut32-p2-perm2.c",
3390 "src/math/exp-avx512f-rr2-p5-scalef.c",
3391 "src/math/exp-avx512f-rr2-p5.c",
Marat Dukhande390d42020-11-29 19:32:18 -08003392 "src/math/expm1minus-avx512f-rr1-lut16-p3-perm.c",
3393 "src/math/expm1minus-avx512f-rr1-p6.c",
Marat Dukhan98ba4412019-10-23 02:14:28 -07003394 "src/math/extexp-avx512f-p5.c",
Marat Dukhan6a34c5f2020-09-22 21:44:15 -07003395 "src/math/sigmoid-avx512f-rr1-lut16-p3-perm-scalef-div.c",
Marat Dukhan6a34c5f2020-09-22 21:44:15 -07003396 "src/math/sigmoid-avx512f-rr1-lut16-p3-perm-scalef-nr1fma.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003397 "src/math/sigmoid-avx512f-rr1-lut16-p3-perm-scalef-nr1fma1adj.c",
Marat Dukhan6a34c5f2020-09-22 21:44:15 -07003398 "src/math/sigmoid-avx512f-rr1-lut32-p2-perm2-scalef-div.c",
Marat Dukhan6a34c5f2020-09-22 21:44:15 -07003399 "src/math/sigmoid-avx512f-rr1-lut32-p2-perm2-scalef-nr1fma.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003400 "src/math/sigmoid-avx512f-rr1-lut32-p2-perm2-scalef-nr1fma1adj.c",
Marat Dukhan36173d22020-10-15 17:14:26 -07003401 "src/math/sigmoid-avx512f-rr1-lut64-p2-gather-scalef-div.c",
Marat Dukhan36173d22020-10-15 17:14:26 -07003402 "src/math/sigmoid-avx512f-rr1-lut64-p2-gather-scalef-nr1fma.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003403 "src/math/sigmoid-avx512f-rr1-lut64-p2-gather-scalef-nr1fma1adj.c",
3404 "src/math/sigmoid-avx512f-rr1-p5-scalef-div.c",
3405 "src/math/sigmoid-avx512f-rr1-p5-scalef-nr1fma.c",
3406 "src/math/sigmoid-avx512f-rr1-p5-scalef-nr1fma1adj.c",
3407 "src/math/sigmoid-avx512f-rr2-lut16-p3-perm-scalef-div.c",
3408 "src/math/sigmoid-avx512f-rr2-lut16-p3-perm-scalef-nr1fma.c",
3409 "src/math/sigmoid-avx512f-rr2-lut16-p3-perm-scalef-nr1fma1adj.c",
3410 "src/math/sigmoid-avx512f-rr2-lut32-p2-perm2-scalef-div.c",
3411 "src/math/sigmoid-avx512f-rr2-lut32-p2-perm2-scalef-nr1fma.c",
3412 "src/math/sigmoid-avx512f-rr2-lut32-p2-perm2-scalef-nr1fma1adj.c",
Marat Dukhan36173d22020-10-15 17:14:26 -07003413 "src/math/sigmoid-avx512f-rr2-lut64-p2-gather-scalef-div.c",
Marat Dukhan36173d22020-10-15 17:14:26 -07003414 "src/math/sigmoid-avx512f-rr2-lut64-p2-gather-scalef-nr1fma.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003415 "src/math/sigmoid-avx512f-rr2-lut64-p2-gather-scalef-nr1fma1adj.c",
3416 "src/math/sigmoid-avx512f-rr2-p5-scalef-div.c",
3417 "src/math/sigmoid-avx512f-rr2-p5-scalef-nr1fma.c",
3418 "src/math/sigmoid-avx512f-rr2-p5-scalef-nr1fma1adj.c",
Marat Dukhan84000762020-06-29 18:38:43 -07003419 "src/math/sqrt-avx512f-nr1fma.c",
Marat Dukhan84000762020-06-29 18:38:43 -07003420 "src/math/sqrt-avx512f-nr1fma1adj.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003421 "src/math/sqrt-avx512f-nr2fma.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003422]
3423
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07003424AVX512SKX_UKERNELS = [
Marat Dukhan2ffc5e62020-09-06 22:33:38 -07003425 "src/qs8-dwconv/gen/up16x9-minmax-avx512skx-mul32.c",
3426 "src/qs8-dwconv/gen/up32x9-minmax-avx512skx-mul32.c",
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07003427 "src/qs8-gemm/gen/1x16c8-minmax-avx512skx.c",
3428 "src/qs8-gemm/gen/2x16c8-minmax-avx512skx.c",
3429 "src/qs8-gemm/gen/3x16c8-minmax-avx512skx.c",
3430 "src/qs8-gemm/gen/4x16c8-minmax-avx512skx.c",
3431 "src/qs8-igemm/gen/1x16c8-minmax-avx512skx.c",
3432 "src/qs8-igemm/gen/2x16c8-minmax-avx512skx.c",
3433 "src/qs8-igemm/gen/3x16c8-minmax-avx512skx.c",
3434 "src/qs8-igemm/gen/4x16c8-minmax-avx512skx.c",
3435]
3436
Frank Barchardbcedc082020-08-17 18:00:51 -07003437WASM32_ASM_UKERNELS = [
3438 "src/f32-relu/wasm_shr_x1.S",
3439 "src/f32-relu/wasm_shr_x2.S",
3440 "src/f32-relu/wasm_shr_x4.S",
3441]
3442
Marat Dukhan08c4a432019-10-03 09:29:21 -07003443AARCH32_ASM_UKERNELS = [
Marat Dukhan32f93812020-05-17 20:31:21 -07003444 "src/f32-gemm/4x4-aarch32-vfp-ld64.S",
Marat Dukhan3b98f6b2020-05-17 10:09:22 -07003445 "src/f32-gemm/4x4-minmax-aarch32-vfp-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003446 "src/f32-gemm/4x8-minmax-aarch32-neon-cortex-a53.S",
3447 "src/f32-gemm/4x8-minmax-aarch32-neon-cortex-a55.S",
Frank Barchard490febe2020-07-16 18:42:17 -07003448 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-cortex-a7.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003449 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-cortex-a75.S",
Frank Barchard569561d2020-06-17 13:11:12 -07003450 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-ld64.S",
Frank Barchard490febe2020-07-16 18:42:17 -07003451 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-pld-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003452 "src/f32-igemm/4x8-minmax-aarch32-neon-cortex-a53.S",
3453 "src/f32-igemm/4x8-minmax-aarch32-neon-cortex-a55.S",
Frank Barchard490febe2020-07-16 18:42:17 -07003454 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-cortex-a7.S",
3455 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-cortex-a75.S",
3456 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-ld64.S",
3457 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-pld-cortex-a75.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003458]
3459
3460AARCH64_ASM_UKERNELS = [
Frank Barchardbddfbcd2020-04-15 12:32:41 -07003461 "src/f16-gemm/gen-inc/1x8inc-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchard04336c12020-10-22 16:48:55 -07003462 "src/f16-gemm/gen-inc/1x16inc-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchardbddfbcd2020-04-15 12:32:41 -07003463 "src/f16-gemm/gen-inc/4x8inc-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchard04336c12020-10-22 16:48:55 -07003464 "src/f16-gemm/gen-inc/4x16inc-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchardbddfbcd2020-04-15 12:32:41 -07003465 "src/f16-gemm/gen-inc/6x8inc-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchard04336c12020-10-22 16:48:55 -07003466 "src/f16-gemm/gen-inc/6x16inc-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchard3b8e5662020-04-20 12:12:53 -07003467 "src/f16-gemm/gen-inc/8x8inc-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchard04336c12020-10-22 16:48:55 -07003468 "src/f16-gemm/gen/1x8-minmax-aarch64-neonfp16arith-ld64.S",
3469 "src/f16-gemm/gen/1x16-minmax-aarch64-neonfp16arith-ld32.S",
3470 "src/f16-gemm/gen/4x8-minmax-aarch64-neonfp16arith-ld64.S",
3471 "src/f16-gemm/gen/4x16-minmax-aarch64-neonfp16arith-ld32.S",
3472 "src/f16-gemm/gen/6x8-minmax-aarch64-neonfp16arith-ld64.S",
3473 "src/f16-gemm/gen/6x16-minmax-aarch64-neonfp16arith-ld32.S",
3474 "src/f16-gemm/gen/8x8-minmax-aarch64-neonfp16arith-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003475 "src/f32-dwconv/up4x9-minmax-aarch64-neonfma-cortex-a55.S",
3476 "src/f32-dwconv/up4x9-minmax-aarch64-neonfma.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003477 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a53.S",
3478 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a57.S",
3479 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a75.S",
Frank Barchard04336c12020-10-22 16:48:55 -07003480 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-ld64.S",
3481 "src/f32-gemm/gen-inc/1x12inc-minmax-aarch64-neonfma-cortex-a53.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003482 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a53.S",
3483 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a55.S",
3484 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a57.S",
3485 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003486 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-ld64.S",
Frank Barchard04336c12020-10-22 16:48:55 -07003487 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-ld128.S",
3488 "src/f32-gemm/gen-inc/4x12inc-minmax-aarch64-neonfma-cortex-a53.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003489 "src/f32-gemm/gen-inc/5x8inc-minmax-aarch64-neonfma-cortex-a57.S",
3490 "src/f32-gemm/gen-inc/5x8inc-minmax-aarch64-neonfma-cortex-a75.S",
3491 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a53.S",
3492 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a55.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003493 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a57.S",
Frank Barchard04336c12020-10-22 16:48:55 -07003494 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a73.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003495 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003496 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-ld64.S",
Frank Barchard04336c12020-10-22 16:48:55 -07003497 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-ld128.S",
3498 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a53.S",
3499 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a57.S",
3500 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a75.S",
3501 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-ld64.S",
3502 "src/f32-gemm/gen/1x12-minmax-aarch64-neonfma-cortex-a53.S",
3503 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a53.S",
3504 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a55.S",
3505 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a57.S",
3506 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a75.S",
3507 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-ld64.S",
3508 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-ld128.S",
3509 "src/f32-gemm/gen/4x12-minmax-aarch64-neonfma-cortex-a53.S",
3510 "src/f32-gemm/gen/5x8-minmax-aarch64-neonfma-cortex-a57.S",
3511 "src/f32-gemm/gen/5x8-minmax-aarch64-neonfma-cortex-a75.S",
3512 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a53.S",
3513 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a55.S",
3514 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a57.S",
3515 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a73.S",
3516 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a75.S",
3517 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-ld64.S",
3518 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-ld128.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003519 "src/f32-igemm/1x8-minmax-aarch64-neonfma-cortex-a53.S",
Frank Barchard04336c12020-10-22 16:48:55 -07003520 "src/f32-igemm/1x12-minmax-aarch64-neonfma-cortex-a53.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003521 "src/f32-igemm/4x8-minmax-aarch64-neonfma-cortex-a53.S",
3522 "src/f32-igemm/4x8-minmax-aarch64-neonfma-cortex-a55.S",
Frank Barchard04336c12020-10-22 16:48:55 -07003523 "src/f32-igemm/4x12-minmax-aarch64-neonfma-cortex-a53.S",
3524 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a53.S",
3525 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a55.S",
3526 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a73.S",
3527 "src/f32-igemm/gen/1x8-minmax-aarch64-neonfma-cortex-a57.S",
3528 "src/f32-igemm/gen/1x8-minmax-aarch64-neonfma-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003529 "src/f32-igemm/gen/4x8-minmax-aarch64-neonfma-cortex-a57.S",
3530 "src/f32-igemm/gen/4x8-minmax-aarch64-neonfma-cortex-a75.S",
3531 "src/f32-igemm/gen/5x8-minmax-aarch64-neonfma-cortex-a57.S",
3532 "src/f32-igemm/gen/5x8-minmax-aarch64-neonfma-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003533 "src/f32-igemm/gen/6x8-minmax-aarch64-neonfma-cortex-a57.S",
3534 "src/f32-igemm/gen/6x8-minmax-aarch64-neonfma-cortex-a75.S",
Frank Barchardf1fd89e2020-10-06 23:27:11 -07003535 "src/qs8-gemm/1x16c4-aarch64-neondot-ld64.S",
Frank Barchardf2742c42020-10-12 21:16:46 -07003536 "src/qs8-gemm/4x16c4-aarch64-neondot-cortex-a55.S",
Frank Barcharda5237a52020-10-06 16:18:46 -07003537 "src/qs8-gemm/4x16c4-aarch64-neondot-ld64.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003538]
3539
Marat Dukhan1b354632020-03-23 12:50:22 -07003540INTERNAL_MICROKERNEL_HDRS = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003541 "src/xnnpack/argmaxpool.h",
3542 "src/xnnpack/avgpool.h",
3543 "src/xnnpack/clamp.h",
3544 "src/xnnpack/common.h",
3545 "src/xnnpack/conv.h",
Yury Kartynnike7841862020-11-04 18:22:18 -08003546 "src/xnnpack/depthtospace.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003547 "src/xnnpack/dwconv.h",
Frank Barchard04336c12020-10-22 16:48:55 -07003548 "src/xnnpack/fill.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003549 "src/xnnpack/gavgpool.h",
3550 "src/xnnpack/gemm.h",
3551 "src/xnnpack/hswish.h",
Marat Dukhan660fd192020-03-10 04:55:30 -07003552 "src/xnnpack/ibilinear.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003553 "src/xnnpack/igemm.h",
Marat Dukhancfb31342019-12-05 10:42:57 -08003554 "src/xnnpack/intrinsics-polyfill.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003555 "src/xnnpack/lut.h",
3556 "src/xnnpack/math.h",
3557 "src/xnnpack/maxpool.h",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003558 "src/xnnpack/memory.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003559 "src/xnnpack/packx.h",
3560 "src/xnnpack/pad.h",
3561 "src/xnnpack/params.h",
3562 "src/xnnpack/pavgpool.h",
3563 "src/xnnpack/ppmm.h",
3564 "src/xnnpack/prelu.h",
Marat Dukhan97579532019-10-18 16:40:39 -07003565 "src/xnnpack/raddexpminusmax.h",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07003566 "src/xnnpack/raddextexp.h",
Marat Dukhan97579532019-10-18 16:40:39 -07003567 "src/xnnpack/raddstoreexpminusmax.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003568 "src/xnnpack/rmax.h",
3569 "src/xnnpack/scalar-utils.h",
3570 "src/xnnpack/spmm.h",
3571 "src/xnnpack/unpool.h",
3572 "src/xnnpack/vadd.h",
Marat Dukhan1e782c42019-11-21 17:02:40 -08003573 "src/xnnpack/vbinary.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003574 "src/xnnpack/vmulcaddc.h",
Marat Dukhan05ac8e32019-10-21 15:39:33 -07003575 "src/xnnpack/vscale.h",
Marat Dukhan97579532019-10-18 16:40:39 -07003576 "src/xnnpack/vscaleexpminusmax.h",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07003577 "src/xnnpack/vscaleextexp.h",
Marat Dukhan1e782c42019-11-21 17:02:40 -08003578 "src/xnnpack/vunary.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003579 "src/xnnpack/zip.h",
Marat Dukhan1b354632020-03-23 12:50:22 -07003580]
3581
3582INTERNAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003583 "include/xnnpack.h",
3584 "src/xnnpack/allocator.h",
3585 "src/xnnpack/compute.h",
3586 "src/xnnpack/im2col.h",
3587 "src/xnnpack/indirection.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003588 "src/xnnpack/math-stubs.h",
Chao Mei6ddfc602020-05-13 22:29:36 -07003589 "src/xnnpack/memory-planner.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003590 "src/xnnpack/operator.h",
3591 "src/xnnpack/pack.h",
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07003592 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003593 "src/xnnpack/requantization-stubs.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003594 "src/xnnpack/requantization.h",
Marat Dukhan1d75a542020-02-03 12:23:01 -08003595 "src/xnnpack/subgraph.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003596]
3597
Marat Dukhan1b354632020-03-23 12:50:22 -07003598ACCURACY_EVAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003599 "src/xnnpack/math-stubs.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003600]
3601
Marat Dukhan1b354632020-03-23 12:50:22 -07003602MICROKERNEL_BENCHMARK_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003603 "include/xnnpack.h",
Frank Barchard35db7d02020-10-26 13:37:34 -07003604 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003605]
3606
Marat Dukhan1b354632020-03-23 12:50:22 -07003607MICROKERNEL_TEST_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Frank Barchard35db7d02020-10-26 13:37:34 -07003608 "include/xnnpack.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003609 "src/xnnpack/isa-checks.h",
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07003610 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003611 "src/xnnpack/requantization.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003612]
3613
3614OPERATOR_TEST_PARAMS_HDRS = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003615 "src/xnnpack/common.h",
Frank Barchard04336c12020-10-22 16:48:55 -07003616 "src/xnnpack/params.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003617]
3618
3619WEIGHTS_PACK_HDRS = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003620 "src/xnnpack/compute.h",
Frank Barchard04336c12020-10-22 16:48:55 -07003621 "src/xnnpack/operator.h",
3622 "src/xnnpack/pack.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003623]
3624
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07003625LOGGING_COPTS = select({
3626 # No logging in optimized mode
3627 ":optimized_build": ["-DXNN_LOG_LEVEL=0"],
3628 # Full logging in debug mode
3629 ":debug_build": ["-DXNN_LOG_LEVEL=5"],
3630 # Error-only logging in default (fastbuild) mode
3631 "//conditions:default": ["-DXNN_LOG_LEVEL=2"],
3632})
3633
Marat Dukhan3b59de22020-06-03 20:15:19 -07003634LOGGING_SRCS = select({
3635 # No logging in optimized mode
3636 ":optimized_build": [],
3637 "//conditions:default": [
3638 "src/operator-strings.c",
3639 "src/subgraph-strings.c",
3640 ],
3641})
3642
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07003643LOGGING_HDRS = [
3644 "src/xnnpack/log.h",
3645]
3646
Marat Dukhan08c4a432019-10-03 09:29:21 -07003647xnnpack_cc_library(
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003648 name = "tables",
3649 srcs = TABLE_SRCS,
3650 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003651 gcc_copts = xnnpack_gcc_std_copts(),
3652 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003653)
3654
3655xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003656 name = "scalar_ukernels",
3657 srcs = SCALAR_UKERNELS,
3658 hdrs = INTERNAL_HDRS,
3659 aarch32_copts = ["-marm"],
Marat Dukhan10a38082020-04-17 03:58:35 -07003660 gcc_copts = xnnpack_gcc_std_copts(),
3661 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003662 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003663 ":tables",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003664 "@FP16",
3665 "@FXdiv",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003666 "@pthreadpool",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003667 ],
3668)
3669
3670xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003671 name = "scalar_ukernels_test_mode",
3672 srcs = SCALAR_UKERNELS,
3673 hdrs = INTERNAL_HDRS,
3674 aarch32_copts = ["-marm"],
3675 copts = [
3676 "-UNDEBUG",
3677 "-DXNN_TEST_MODE=1",
3678 ],
3679 gcc_copts = xnnpack_gcc_std_copts(),
3680 msvc_copts = xnnpack_msvc_std_copts(),
3681 deps = [
3682 ":tables",
3683 "@FP16",
3684 "@FXdiv",
3685 "@pthreadpool",
3686 ],
3687)
3688
3689xnnpack_cc_library(
Marat Dukhan436ebe62019-12-04 15:10:12 -08003690 name = "wasm_ukernels",
3691 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003692 gcc_copts = xnnpack_gcc_std_copts(),
3693 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan436ebe62019-12-04 15:10:12 -08003694 wasm_srcs = WASM_UKERNELS,
Marat Dukhan1c6e3892020-06-25 23:56:10 -07003695 wasmsimd_srcs = WASM_UKERNELS + WASMSIMD_UKERNELS,
Marat Dukhan436ebe62019-12-04 15:10:12 -08003696 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003697 ":tables",
Marat Dukhan436ebe62019-12-04 15:10:12 -08003698 "@FP16",
3699 "@FXdiv",
3700 "@pthreadpool",
3701 ],
3702)
3703
3704xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003705 name = "wasm_ukernels_test_mode",
3706 hdrs = INTERNAL_HDRS,
3707 copts = [
3708 "-UNDEBUG",
3709 "-DXNN_TEST_MODE=1",
3710 ],
3711 gcc_copts = xnnpack_gcc_std_copts(),
3712 msvc_copts = xnnpack_msvc_std_copts(),
3713 wasm_srcs = WASM_UKERNELS,
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07003714 wasmsimd_srcs = WASM_UKERNELS + WASMSIMD_UKERNELS,
Marat Dukhan33fcf782020-05-24 14:27:15 -07003715 deps = [
3716 ":tables",
3717 "@FP16",
3718 "@FXdiv",
3719 "@pthreadpool",
3720 ],
3721)
3722
3723xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003724 name = "neon_ukernels",
3725 hdrs = INTERNAL_HDRS,
3726 aarch32_copts = [
3727 "-marm",
Marat Dukhan8853b822020-05-07 12:19:01 -07003728 "-march=armv7-a",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003729 "-mfpu=neon",
3730 ],
3731 aarch32_srcs = NEON_UKERNELS,
3732 aarch64_srcs = NEON_UKERNELS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003733 gcc_copts = xnnpack_gcc_std_copts(),
3734 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08003735 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003736 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003737 "@FP16",
3738 "@pthreadpool",
3739 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003740)
3741
3742xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003743 name = "neon_ukernels_test_mode",
3744 hdrs = INTERNAL_HDRS,
3745 aarch32_copts = [
3746 "-marm",
3747 "-march=armv7-a",
3748 "-mfpu=neon",
3749 ],
3750 aarch32_srcs = NEON_UKERNELS,
3751 aarch64_srcs = NEON_UKERNELS,
3752 copts = [
3753 "-UNDEBUG",
3754 "-DXNN_TEST_MODE=1",
3755 ],
3756 gcc_copts = xnnpack_gcc_std_copts(),
3757 msvc_copts = xnnpack_msvc_std_copts(),
3758 deps = [
3759 ":tables",
3760 "@FP16",
3761 "@pthreadpool",
3762 ],
3763)
3764
3765xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003766 name = "neonfma_ukernels",
3767 hdrs = INTERNAL_HDRS,
3768 aarch32_copts = [
3769 "-marm",
Marat Dukhan8853b822020-05-07 12:19:01 -07003770 "-march=armv7-a",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003771 "-mfpu=neon-vfpv4",
3772 ],
3773 aarch32_srcs = NEONFMA_UKERNELS,
3774 aarch64_srcs = NEONFMA_UKERNELS + AARCH64_NEONFMA_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07003775 apple_aarch32_copts = [
3776 "-mcpu=swift",
3777 "-mtune=generic",
3778 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07003779 gcc_copts = xnnpack_gcc_std_copts(),
3780 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08003781 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003782 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003783 "@FP16",
3784 "@pthreadpool",
3785 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003786)
3787
3788xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003789 name = "neonfma_ukernels_test_mode",
3790 hdrs = INTERNAL_HDRS,
3791 aarch32_copts = [
3792 "-marm",
3793 "-march=armv7-a",
3794 "-mfpu=neon-vfpv4",
3795 ],
3796 aarch32_srcs = NEONFMA_UKERNELS,
3797 aarch64_srcs = NEONFMA_UKERNELS + AARCH64_NEONFMA_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07003798 apple_aarch32_copts = [
3799 "-mcpu=swift",
3800 "-mtune=generic",
3801 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07003802 copts = [
3803 "-UNDEBUG",
3804 "-DXNN_TEST_MODE=1",
3805 ],
3806 gcc_copts = xnnpack_gcc_std_copts(),
3807 msvc_copts = xnnpack_msvc_std_copts(),
3808 deps = [
3809 ":tables",
3810 "@FP16",
3811 "@pthreadpool",
3812 ],
3813)
3814
3815xnnpack_cc_library(
Marat Dukhan8853b822020-05-07 12:19:01 -07003816 name = "neonv8_ukernels",
3817 hdrs = INTERNAL_HDRS,
3818 aarch32_copts = [
3819 "-marm",
3820 "-march=armv8-a",
3821 "-mfpu=neon-fp-armv8",
3822 ],
3823 aarch32_srcs = NEONV8_UKERNELS,
3824 aarch64_srcs = NEONV8_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07003825 apple_aarch32_copts = [
3826 "-mcpu=cyclone",
3827 "-mtune=generic",
3828 ],
Marat Dukhan8853b822020-05-07 12:19:01 -07003829 gcc_copts = xnnpack_gcc_std_copts(),
3830 msvc_copts = xnnpack_msvc_std_copts(),
3831 deps = [
3832 ":tables",
3833 "@FP16",
3834 "@pthreadpool",
3835 ],
3836)
3837
3838xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003839 name = "neonv8_ukernels_test_mode",
3840 hdrs = INTERNAL_HDRS,
3841 aarch32_copts = [
3842 "-marm",
3843 "-march=armv8-a",
3844 "-mfpu=neon-fp-armv8",
3845 ],
3846 aarch32_srcs = NEONV8_UKERNELS,
3847 aarch64_srcs = NEONV8_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07003848 apple_aarch32_copts = [
3849 "-mcpu=cyclone",
3850 "-mtune=generic",
3851 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07003852 copts = [
3853 "-UNDEBUG",
3854 "-DXNN_TEST_MODE=1",
3855 ],
3856 gcc_copts = xnnpack_gcc_std_copts(),
3857 msvc_copts = xnnpack_msvc_std_copts(),
3858 deps = [
3859 ":tables",
3860 "@FP16",
3861 "@pthreadpool",
3862 ],
3863)
3864
3865xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003866 name = "neonfp16arith_ukernels",
3867 hdrs = INTERNAL_HDRS,
3868 aarch64_copts = ["-march=armv8.2-a+fp16"],
3869 aarch64_srcs = AARCH64_NEONFP16ARITH_UKERNELS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003870 gcc_copts = xnnpack_gcc_std_copts(),
3871 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08003872 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003873 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003874 "@FP16",
3875 "@pthreadpool",
3876 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003877)
3878
3879xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003880 name = "neonfp16arith_ukernels_test_mode",
3881 hdrs = INTERNAL_HDRS,
3882 aarch64_copts = ["-march=armv8.2-a+fp16"],
3883 aarch64_srcs = AARCH64_NEONFP16ARITH_UKERNELS,
3884 copts = [
3885 "-UNDEBUG",
3886 "-DXNN_TEST_MODE=1",
3887 ],
3888 gcc_copts = xnnpack_gcc_std_copts(),
3889 msvc_copts = xnnpack_msvc_std_copts(),
3890 deps = [
3891 ":tables",
3892 "@FP16",
3893 "@pthreadpool",
3894 ],
3895)
3896
3897xnnpack_cc_library(
Benoit Jacoba9644732020-08-13 12:48:55 -07003898 name = "neondot_ukernels",
3899 hdrs = INTERNAL_HDRS,
Marat Dukhan799ac752020-08-13 16:08:03 -07003900 aarch32_copts = [
3901 "-marm",
3902 "-march=armv8.2-a+dotprod",
3903 "-mfpu=neon-fp-armv8",
3904 ],
Benoit Jacoba9644732020-08-13 12:48:55 -07003905 aarch32_srcs = NEONDOT_UKERNELS,
3906 aarch64_copts = ["-march=armv8.2-a+dotprod"],
3907 aarch64_srcs = NEONDOT_UKERNELS,
3908 gcc_copts = xnnpack_gcc_std_copts(),
3909 msvc_copts = xnnpack_msvc_std_copts(),
3910 deps = [
3911 ":tables",
3912 "@FP16",
3913 "@pthreadpool",
3914 ],
3915)
3916
3917xnnpack_cc_library(
3918 name = "neondot_ukernels_test_mode",
3919 hdrs = INTERNAL_HDRS,
Marat Dukhan799ac752020-08-13 16:08:03 -07003920 aarch32_copts = [
3921 "-marm",
3922 "-march=armv8.2-a+dotprod",
3923 "-mfpu=neon-fp-armv8",
3924 ],
Benoit Jacoba9644732020-08-13 12:48:55 -07003925 aarch32_srcs = NEONDOT_UKERNELS,
3926 aarch64_copts = ["-march=armv8.2-a+dotprod"],
3927 aarch64_srcs = NEONDOT_UKERNELS,
3928 copts = [
3929 "-UNDEBUG",
3930 "-DXNN_TEST_MODE=1",
3931 ],
3932 gcc_copts = xnnpack_gcc_std_copts(),
3933 msvc_copts = xnnpack_msvc_std_copts(),
3934 deps = [
3935 ":tables",
3936 "@FP16",
3937 "@pthreadpool",
3938 ],
3939)
3940
3941xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003942 name = "sse2_ukernels",
3943 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003944 gcc_copts = xnnpack_gcc_std_copts(),
3945 gcc_x86_copts = ["-msse2"],
3946 msvc_copts = xnnpack_msvc_std_copts(),
3947 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003948 x86_srcs = SSE_UKERNELS + SSE2_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08003949 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003950 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003951 "@FP16",
3952 "@pthreadpool",
3953 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003954)
3955
3956xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003957 name = "sse2_ukernels_test_mode",
3958 hdrs = INTERNAL_HDRS,
3959 copts = [
3960 "-UNDEBUG",
3961 "-DXNN_TEST_MODE=1",
3962 ],
3963 gcc_copts = xnnpack_gcc_std_copts(),
3964 gcc_x86_copts = ["-msse2"],
3965 msvc_copts = xnnpack_msvc_std_copts(),
3966 msvc_x86_32_copts = ["/arch:SSE2"],
3967 x86_srcs = SSE_UKERNELS + SSE2_UKERNELS,
3968 deps = [
3969 ":tables",
3970 "@FP16",
3971 "@pthreadpool",
3972 ],
3973)
3974
3975xnnpack_cc_library(
Marat Dukhanfe7acb62020-03-09 19:30:05 -07003976 name = "ssse3_ukernels",
3977 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003978 gcc_copts = xnnpack_gcc_std_copts(),
3979 gcc_x86_copts = ["-mssse3"],
3980 msvc_copts = xnnpack_msvc_std_copts(),
3981 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhanfe7acb62020-03-09 19:30:05 -07003982 x86_srcs = SSSE3_UKERNELS,
3983 deps = [
3984 ":tables",
3985 "@FP16",
3986 "@pthreadpool",
3987 ],
3988)
3989
3990xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003991 name = "ssse3_ukernels_test_mode",
3992 hdrs = INTERNAL_HDRS,
3993 copts = [
3994 "-UNDEBUG",
3995 "-DXNN_TEST_MODE=1",
3996 ],
3997 gcc_copts = xnnpack_gcc_std_copts(),
3998 gcc_x86_copts = ["-mssse3"],
3999 msvc_copts = xnnpack_msvc_std_copts(),
4000 msvc_x86_32_copts = ["/arch:SSE2"],
4001 x86_srcs = SSSE3_UKERNELS,
4002 deps = [
4003 ":tables",
4004 "@FP16",
4005 "@pthreadpool",
4006 ],
4007)
4008
4009xnnpack_cc_library(
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08004010 name = "sse41_ukernels",
4011 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004012 gcc_copts = xnnpack_gcc_std_copts(),
4013 gcc_x86_copts = ["-msse4.1"],
4014 msvc_copts = xnnpack_msvc_std_copts(),
4015 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08004016 x86_srcs = SSE41_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08004017 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08004018 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08004019 "@FP16",
4020 "@pthreadpool",
4021 ],
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08004022)
4023
4024xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07004025 name = "sse41_ukernels_test_mode",
4026 hdrs = INTERNAL_HDRS,
4027 copts = [
4028 "-UNDEBUG",
4029 "-DXNN_TEST_MODE=1",
4030 ],
4031 gcc_copts = xnnpack_gcc_std_copts(),
4032 gcc_x86_copts = ["-msse4.1"],
4033 msvc_copts = xnnpack_msvc_std_copts(),
4034 msvc_x86_32_copts = ["/arch:SSE2"],
4035 x86_srcs = SSE41_UKERNELS,
4036 deps = [
4037 ":tables",
4038 "@FP16",
4039 "@pthreadpool",
4040 ],
4041)
4042
4043xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004044 name = "avx_ukernels",
4045 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004046 gcc_copts = xnnpack_gcc_std_copts(),
4047 gcc_x86_copts = ["-mavx"],
4048 msvc_copts = xnnpack_msvc_std_copts(),
4049 msvc_x86_32_copts = ["/arch:AVX"],
4050 msvc_x86_64_copts = ["/arch:AVX"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004051 x86_srcs = AVX_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08004052 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08004053 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08004054 "@FP16",
4055 "@pthreadpool",
4056 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004057)
4058
4059xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07004060 name = "avx_ukernels_test_mode",
4061 hdrs = INTERNAL_HDRS,
4062 copts = [
4063 "-UNDEBUG",
4064 "-DXNN_TEST_MODE=1",
4065 ],
4066 gcc_copts = xnnpack_gcc_std_copts(),
4067 gcc_x86_copts = ["-mavx"],
4068 msvc_copts = xnnpack_msvc_std_copts(),
4069 msvc_x86_32_copts = ["/arch:AVX"],
4070 msvc_x86_64_copts = ["/arch:AVX"],
4071 x86_srcs = AVX_UKERNELS,
4072 deps = [
4073 ":tables",
4074 "@FP16",
4075 "@pthreadpool",
4076 ],
4077)
4078
4079xnnpack_cc_library(
Marat Dukhan1566fee2020-08-02 21:55:41 -07004080 name = "xop_ukernels",
4081 hdrs = INTERNAL_HDRS,
4082 gcc_copts = xnnpack_gcc_std_copts(),
4083 gcc_x86_copts = ["-mxop"],
4084 msvc_copts = xnnpack_msvc_std_copts(),
4085 msvc_x86_32_copts = ["/arch:AVX"],
4086 msvc_x86_64_copts = ["/arch:AVX"],
4087 x86_srcs = XOP_UKERNELS,
4088 deps = [
4089 ":tables",
4090 "@FP16",
4091 "@pthreadpool",
4092 ],
4093)
4094
4095xnnpack_cc_library(
4096 name = "xop_ukernels_test_mode",
4097 hdrs = INTERNAL_HDRS,
4098 copts = [
4099 "-UNDEBUG",
4100 "-DXNN_TEST_MODE=1",
4101 ],
4102 gcc_copts = xnnpack_gcc_std_copts(),
4103 gcc_x86_copts = ["-mxop"],
4104 msvc_copts = xnnpack_msvc_std_copts(),
4105 msvc_x86_32_copts = ["/arch:AVX"],
4106 msvc_x86_64_copts = ["/arch:AVX"],
4107 x86_srcs = XOP_UKERNELS,
4108 deps = [
4109 ":tables",
4110 "@FP16",
4111 "@pthreadpool",
4112 ],
4113)
4114
4115xnnpack_cc_library(
Marat Dukhanfda12b82019-11-21 12:27:59 -08004116 name = "fma3_ukernels",
4117 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004118 gcc_copts = xnnpack_gcc_std_copts(),
4119 gcc_x86_copts = ["-mfma"],
4120 msvc_copts = xnnpack_msvc_std_copts(),
4121 msvc_x86_32_copts = ["/arch:AVX"],
4122 msvc_x86_64_copts = ["/arch:AVX"],
Marat Dukhanfda12b82019-11-21 12:27:59 -08004123 x86_srcs = FMA3_UKERNELS,
4124 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08004125 ":tables",
Marat Dukhanfda12b82019-11-21 12:27:59 -08004126 "@FP16",
4127 "@pthreadpool",
4128 ],
4129)
4130
4131xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07004132 name = "fma3_ukernels_test_mode",
4133 hdrs = INTERNAL_HDRS,
4134 copts = [
4135 "-UNDEBUG",
4136 "-DXNN_TEST_MODE=1",
4137 ],
4138 gcc_copts = xnnpack_gcc_std_copts(),
4139 gcc_x86_copts = ["-mfma"],
4140 msvc_copts = xnnpack_msvc_std_copts(),
4141 msvc_x86_32_copts = ["/arch:AVX"],
4142 msvc_x86_64_copts = ["/arch:AVX"],
4143 x86_srcs = FMA3_UKERNELS,
4144 deps = [
4145 ":tables",
4146 "@FP16",
4147 "@pthreadpool",
4148 ],
4149)
4150
4151xnnpack_cc_library(
Marat Dukhan6adff4e2019-10-14 18:32:07 -07004152 name = "avx2_ukernels",
4153 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004154 gcc_copts = xnnpack_gcc_std_copts(),
4155 gcc_x86_copts = [
Marat Dukhan6adff4e2019-10-14 18:32:07 -07004156 "-mfma",
4157 "-mavx2",
4158 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07004159 msvc_copts = xnnpack_msvc_std_copts(),
4160 msvc_x86_32_copts = ["/arch:AVX2"],
4161 msvc_x86_64_copts = ["/arch:AVX2"],
Marat Dukhan6adff4e2019-10-14 18:32:07 -07004162 x86_srcs = AVX2_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08004163 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08004164 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08004165 "@FP16",
4166 "@pthreadpool",
4167 ],
Marat Dukhan6adff4e2019-10-14 18:32:07 -07004168)
4169
4170xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07004171 name = "avx2_ukernels_test_mode",
4172 hdrs = INTERNAL_HDRS,
4173 copts = [
4174 "-UNDEBUG",
4175 "-DXNN_TEST_MODE=1",
4176 ],
4177 gcc_copts = xnnpack_gcc_std_copts(),
4178 gcc_x86_copts = [
4179 "-mfma",
4180 "-mavx2",
4181 ],
4182 msvc_copts = xnnpack_msvc_std_copts(),
4183 msvc_x86_32_copts = ["/arch:AVX2"],
4184 msvc_x86_64_copts = ["/arch:AVX2"],
4185 x86_srcs = AVX2_UKERNELS,
4186 deps = [
4187 ":tables",
4188 "@FP16",
4189 "@pthreadpool",
4190 ],
4191)
4192
4193xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004194 name = "avx512f_ukernels",
4195 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004196 gcc_copts = xnnpack_gcc_std_copts(),
4197 gcc_x86_copts = ["-mavx512f"],
4198 mingw_copts = ["-fno-asynchronous-unwind-tables"],
4199 msvc_copts = xnnpack_msvc_std_copts(),
4200 msvc_x86_32_copts = ["/arch:AVX512"],
4201 msvc_x86_64_copts = ["/arch:AVX512"],
4202 msys_copts = ["-fno-asynchronous-unwind-tables"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004203 x86_srcs = AVX512F_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08004204 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08004205 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08004206 "@FP16",
4207 "@pthreadpool",
4208 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004209)
4210
4211xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07004212 name = "avx512f_ukernels_test_mode",
4213 hdrs = INTERNAL_HDRS,
4214 copts = [
4215 "-UNDEBUG",
4216 "-DXNN_TEST_MODE=1",
4217 ],
4218 gcc_copts = xnnpack_gcc_std_copts(),
4219 gcc_x86_copts = ["-mavx512f"],
4220 mingw_copts = ["-fno-asynchronous-unwind-tables"],
4221 msvc_copts = xnnpack_msvc_std_copts(),
4222 msvc_x86_32_copts = ["/arch:AVX512"],
4223 msvc_x86_64_copts = ["/arch:AVX512"],
4224 msys_copts = ["-fno-asynchronous-unwind-tables"],
4225 x86_srcs = AVX512F_UKERNELS,
4226 deps = [
4227 ":tables",
4228 "@FP16",
4229 "@pthreadpool",
4230 ],
4231)
4232
4233xnnpack_cc_library(
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07004234 name = "avx512skx_ukernels",
4235 hdrs = INTERNAL_HDRS,
4236 gcc_copts = xnnpack_gcc_std_copts(),
4237 gcc_x86_copts = [
4238 "-mavx512f",
4239 "-mavx512cd",
4240 "-mavx512bw",
4241 "-mavx512dq",
4242 "-mavx512vl",
4243 ],
4244 mingw_copts = ["-fno-asynchronous-unwind-tables"],
4245 msvc_copts = xnnpack_msvc_std_copts(),
4246 msvc_x86_32_copts = ["/arch:AVX512"],
4247 msvc_x86_64_copts = ["/arch:AVX512"],
4248 msys_copts = ["-fno-asynchronous-unwind-tables"],
4249 x86_srcs = AVX512SKX_UKERNELS,
4250 deps = [
4251 ":tables",
4252 "@FP16",
4253 "@pthreadpool",
4254 ],
4255)
4256
4257xnnpack_cc_library(
4258 name = "avx512skx_ukernels_test_mode",
4259 hdrs = INTERNAL_HDRS,
4260 copts = [
4261 "-UNDEBUG",
4262 "-DXNN_TEST_MODE=1",
4263 ],
4264 gcc_copts = xnnpack_gcc_std_copts(),
4265 gcc_x86_copts = [
4266 "-mavx512f",
4267 "-mavx512cd",
4268 "-mavx512bw",
4269 "-mavx512dq",
4270 "-mavx512vl",
4271 ],
4272 mingw_copts = ["-fno-asynchronous-unwind-tables"],
4273 msvc_copts = xnnpack_msvc_std_copts(),
4274 msvc_x86_32_copts = ["/arch:AVX512"],
4275 msvc_x86_64_copts = ["/arch:AVX512"],
4276 msys_copts = ["-fno-asynchronous-unwind-tables"],
4277 x86_srcs = AVX512SKX_UKERNELS,
4278 deps = [
4279 ":tables",
4280 "@FP16",
4281 "@pthreadpool",
4282 ],
4283)
4284
4285xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004286 name = "asm_ukernels",
4287 hdrs = ["src/xnnpack/assembly.h"],
4288 aarch32_srcs = AARCH32_ASM_UKERNELS,
Frank Barchard31bb45b2020-10-06 00:26:33 -07004289 aarch64_copts = ["-march=armv8.2-a+fp16+dotprod"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004290 aarch64_srcs = AARCH64_ASM_UKERNELS,
Frank Barchardbcedc082020-08-17 18:00:51 -07004291 wasm_srcs = WASM32_ASM_UKERNELS,
4292 wasmsimd_srcs = WASM32_ASM_UKERNELS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07004293)
4294
Marat Dukhan3b59de22020-06-03 20:15:19 -07004295xnnpack_cc_library(
4296 name = "logging_utils",
4297 srcs = LOGGING_SRCS,
4298 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
4299 copts = LOGGING_COPTS + [
4300 "-Isrc",
4301 "-Iinclude",
4302 ] + select({
4303 ":debug_build": [],
4304 "//conditions:default": xnnpack_min_size_copts(),
4305 }),
4306 gcc_copts = xnnpack_gcc_std_copts(),
4307 msvc_copts = xnnpack_msvc_std_copts(),
4308 visibility = xnnpack_visibility(),
4309 deps = [
4310 "@FP16",
4311 "@clog",
4312 "@pthreadpool",
4313 ],
4314)
4315
Marat Dukhan08c4a432019-10-03 09:29:21 -07004316xnnpack_aggregate_library(
4317 name = "ukernels",
4318 aarch32_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07004319 ":neon_ukernels",
4320 ":neonfma_ukernels",
Marat Dukhan8853b822020-05-07 12:19:01 -07004321 ":neonv8_ukernels",
Benoit Jacoba9644732020-08-13 12:48:55 -07004322 ":neondot_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004323 ":asm_ukernels",
4324 ],
4325 aarch64_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07004326 ":neon_ukernels",
4327 ":neonfma_ukernels",
Marat Dukhan8853b822020-05-07 12:19:01 -07004328 ":neonv8_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004329 ":neonfp16arith_ukernels",
Benoit Jacoba9644732020-08-13 12:48:55 -07004330 ":neondot_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004331 ":asm_ukernels",
4332 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07004333 generic_deps = [
4334 ":scalar_ukernels",
4335 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07004336 wasm_deps = [
4337 ":wasm_ukernels",
Frank Barchardbcedc082020-08-17 18:00:51 -07004338 ":asm_ukernels",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004339 ],
4340 wasmsimd_deps = [
4341 ":wasm_ukernels",
Frank Barchardbcedc082020-08-17 18:00:51 -07004342 ":asm_ukernels",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004343 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004344 x86_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07004345 ":sse2_ukernels",
Marat Dukhanfe7acb62020-03-09 19:30:05 -07004346 ":ssse3_ukernels",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08004347 ":sse41_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004348 ":avx_ukernels",
Marat Dukhan1566fee2020-08-02 21:55:41 -07004349 ":xop_ukernels",
Marat Dukhanfda12b82019-11-21 12:27:59 -08004350 ":fma3_ukernels",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07004351 ":avx2_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004352 ":avx512f_ukernels",
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07004353 ":avx512skx_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004354 ],
4355)
4356
Marat Dukhan33fcf782020-05-24 14:27:15 -07004357xnnpack_aggregate_library(
4358 name = "ukernels_test_mode",
4359 aarch32_deps = [
4360 ":neon_ukernels_test_mode",
4361 ":neonfma_ukernels_test_mode",
4362 ":neonv8_ukernels_test_mode",
Benoit Jacoba9644732020-08-13 12:48:55 -07004363 ":neondot_ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004364 ":asm_ukernels",
4365 ],
4366 aarch64_deps = [
4367 ":neon_ukernels_test_mode",
4368 ":neonfma_ukernels_test_mode",
4369 ":neonv8_ukernels_test_mode",
4370 ":neonfp16arith_ukernels_test_mode",
Benoit Jacoba9644732020-08-13 12:48:55 -07004371 ":neondot_ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004372 ":asm_ukernels",
4373 ],
4374 generic_deps = [
4375 ":scalar_ukernels_test_mode",
4376 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07004377 wasm_deps = [
4378 ":wasm_ukernels_test_mode",
Frank Barchardbcedc082020-08-17 18:00:51 -07004379 ":asm_ukernels",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004380 ],
4381 wasmsimd_deps = [
4382 ":wasm_ukernels_test_mode",
Frank Barchardbcedc082020-08-17 18:00:51 -07004383 ":asm_ukernels",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004384 ],
4385 x86_deps = [
4386 ":sse2_ukernels_test_mode",
4387 ":ssse3_ukernels_test_mode",
4388 ":sse41_ukernels_test_mode",
4389 ":avx_ukernels_test_mode",
Marat Dukhan1566fee2020-08-02 21:55:41 -07004390 ":xop_ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004391 ":fma3_ukernels_test_mode",
4392 ":avx2_ukernels_test_mode",
4393 ":avx512f_ukernels_test_mode",
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07004394 ":avx512skx_ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004395 ],
4396)
4397
Marat Dukhan86df3842020-06-17 14:03:31 -07004398# This target must be compatible with Android NDK r17 and compile with Clang 6.
4399# It is similar to :ukernels target, but excludes microkernels which require
4400# NEON FP16 arithmetics, NEON Dot Product, and newer intrinsics.
Marat Dukhanae046f52020-06-15 13:16:14 -07004401xnnpack_aggregate_library(
Marat Dukhan86df3842020-06-17 14:03:31 -07004402 name = "ukernels_max_compatibility",
Marat Dukhanae046f52020-06-15 13:16:14 -07004403 aarch32_deps = [
4404 ":neon_ukernels",
4405 ":neonfma_ukernels",
4406 ":neonv8_ukernels",
4407 ":asm_ukernels",
4408 ],
4409 aarch64_deps = [
4410 ":neon_ukernels",
4411 ":neonfma_ukernels",
4412 ":neonv8_ukernels",
4413 ":asm_ukernels",
4414 ],
4415 generic_deps = [
4416 ":scalar_ukernels",
4417 ],
Marat Dukhanae046f52020-06-15 13:16:14 -07004418 wasm_deps = [
4419 ":wasm_ukernels",
4420 ],
4421 wasmsimd_deps = [
4422 ":wasm_ukernels",
4423 ],
4424 x86_deps = [
4425 ":sse2_ukernels",
4426 ":ssse3_ukernels",
4427 ":sse41_ukernels",
4428 ":avx_ukernels",
Marat Dukhan1566fee2020-08-02 21:55:41 -07004429 ":xop_ukernels",
Marat Dukhanae046f52020-06-15 13:16:14 -07004430 ":fma3_ukernels",
4431 ":avx2_ukernels",
4432 ":avx512f_ukernels",
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07004433 ":avx512skx_ukernels",
Marat Dukhanae046f52020-06-15 13:16:14 -07004434 ],
4435)
4436
Marat Dukhan08c4a432019-10-03 09:29:21 -07004437xnnpack_cc_library(
4438 name = "im2col",
4439 srcs = ["src/im2col.c"],
4440 hdrs = [
4441 "src/xnnpack/common.h",
4442 "src/xnnpack/im2col.h",
4443 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07004444 gcc_copts = xnnpack_gcc_std_copts(),
4445 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004446)
4447
4448xnnpack_cc_library(
4449 name = "indirection",
4450 srcs = ["src/indirection.c"],
4451 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004452 gcc_copts = xnnpack_gcc_std_copts(),
4453 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004454 deps = [
4455 "@FP16",
4456 "@FXdiv",
4457 "@pthreadpool",
4458 ],
4459)
4460
4461xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07004462 name = "indirection_test_mode",
4463 srcs = ["src/indirection.c"],
4464 hdrs = INTERNAL_HDRS,
4465 copts = [
4466 "-UNDEBUG",
4467 "-DXNN_TEST_MODE=1",
4468 ],
4469 gcc_copts = xnnpack_gcc_std_copts(),
4470 msvc_copts = xnnpack_msvc_std_copts(),
4471 deps = [
4472 "@FP16",
4473 "@FXdiv",
4474 "@pthreadpool",
4475 ],
4476)
4477
4478xnnpack_cc_library(
Marat Dukhanab582382020-07-06 13:32:08 -07004479 name = "packing",
4480 srcs = ["src/packing.c"],
4481 hdrs = INTERNAL_HDRS,
4482 gcc_copts = xnnpack_gcc_std_copts(),
4483 msvc_copts = xnnpack_msvc_std_copts(),
4484 deps = [
4485 "@FP16",
4486 "@FXdiv",
4487 "@pthreadpool",
4488 ],
4489)
4490
4491xnnpack_cc_library(
4492 name = "packing_test_mode",
4493 srcs = ["src/packing.c"],
4494 hdrs = INTERNAL_HDRS,
4495 copts = [
4496 "-UNDEBUG",
4497 "-DXNN_TEST_MODE=1",
4498 ],
4499 gcc_copts = xnnpack_gcc_std_copts(),
4500 msvc_copts = xnnpack_msvc_std_copts(),
4501 deps = [
4502 "@FP16",
4503 "@FXdiv",
4504 "@pthreadpool",
4505 ],
4506)
4507
4508xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004509 name = "operator_run",
4510 srcs = ["src/operator-run.c"],
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07004511 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004512 copts = LOGGING_COPTS + select({
Marat Dukhan05702cf2020-03-26 15:41:33 -07004513 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4514 "//conditions:default": [],
4515 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07004516 gcc_copts = xnnpack_gcc_std_copts(),
4517 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004518 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07004519 ":logging_utils",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004520 "@FP16",
4521 "@FXdiv",
4522 "@clog",
4523 "@pthreadpool",
4524 ],
4525)
4526
Chao Mei6ddfc602020-05-13 22:29:36 -07004527xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07004528 name = "operator_run_test_mode",
4529 srcs = ["src/operator-run.c"],
4530 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
4531 copts = LOGGING_COPTS + [
4532 "-UNDEBUG",
4533 "-DXNN_TEST_MODE=1",
4534 ] + select({
4535 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4536 "//conditions:default": [],
4537 }),
4538 gcc_copts = xnnpack_gcc_std_copts(),
4539 msvc_copts = xnnpack_msvc_std_copts(),
4540 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07004541 ":logging_utils",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004542 "@FP16",
4543 "@FXdiv",
4544 "@clog",
4545 "@pthreadpool",
4546 ],
4547)
4548
4549xnnpack_cc_library(
Chao Mei6ddfc602020-05-13 22:29:36 -07004550 name = "memory_planner",
4551 srcs = ["src/memory-planner.c"],
4552 hdrs = INTERNAL_HDRS,
4553 defines = select({
4554 ":xnn_enable_memopt_explicit_true": ["XNN_ENABLE_MEMOPT=1"],
4555 ":xnn_enable_memopt_explicit_false": ["XNN_ENABLE_MEMOPT=0"],
4556 "//conditions:default": ["XNN_ENABLE_MEMOPT=1"],
4557 }),
4558 gcc_copts = xnnpack_gcc_std_copts(),
4559 msvc_copts = xnnpack_msvc_std_copts(),
4560 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07004561 ":logging_utils",
Chao Mei6ddfc602020-05-13 22:29:36 -07004562 "@pthreadpool",
4563 ],
4564)
4565
Marat Dukhan33fcf782020-05-24 14:27:15 -07004566xnnpack_cc_library(
4567 name = "memory_planner_test_mode",
4568 srcs = ["src/memory-planner.c"],
4569 hdrs = INTERNAL_HDRS,
4570 copts = [
4571 "-UNDEBUG",
4572 "-DXNN_TEST_MODE=1",
4573 ],
4574 defines = select({
4575 ":xnn_enable_memopt_explicit_true": ["XNN_ENABLE_MEMOPT=1"],
4576 ":xnn_enable_memopt_explicit_false": ["XNN_ENABLE_MEMOPT=0"],
4577 "//conditions:default": ["XNN_ENABLE_MEMOPT=1"],
4578 }),
4579 gcc_copts = xnnpack_gcc_std_copts(),
4580 msvc_copts = xnnpack_msvc_std_copts(),
4581 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07004582 ":logging_utils",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004583 "@pthreadpool",
4584 ],
4585)
4586
Marat Dukhan08c4a432019-10-03 09:29:21 -07004587cc_library(
4588 name = "enable_assembly",
4589 defines = select({
4590 ":xnn_enable_assembly_explicit_true": ["XNN_ENABLE_ASSEMBLY=1"],
4591 ":xnn_enable_assembly_explicit_false": ["XNN_ENABLE_ASSEMBLY=0"],
Frank Barchard810171d2019-10-10 10:34:51 -07004592 "//conditions:default": ["XNN_ENABLE_ASSEMBLY=1"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004593 }),
4594)
4595
Marat Dukhan9de90e02020-06-18 16:04:12 -07004596cc_library(
4597 name = "enable_sparse",
4598 defines = select({
4599 ":xnn_enable_sparse_explicit_true": ["XNN_ENABLE_SPARSE=1"],
4600 ":xnn_enable_sparse_explicit_false": ["XNN_ENABLE_SPARSE=0"],
Marat Dukhanb36582b2020-12-08 11:16:28 -08004601 "//conditions:default": ["XNN_ENABLE_SPARSE=1"],
Marat Dukhan9de90e02020-06-18 16:04:12 -07004602 }),
4603)
4604
Marat Dukhancf056b22019-10-07 10:26:29 -07004605xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004606 name = "operators",
4607 srcs = OPERATOR_SRCS + [
Marat Dukhan04f03be2019-11-19 12:36:47 -08004608 "src/memory.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004609 "src/operator-delete.c",
Marat Dukhancf056b22019-10-07 10:26:29 -07004610 ],
4611 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004612 copts = LOGGING_COPTS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07004613 "-Isrc",
4614 "-Iinclude",
4615 ] + select({
4616 ":debug_build": [],
4617 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07004618 }) + select({
4619 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4620 "//conditions:default": [],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004621 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07004622 gcc_copts = xnnpack_gcc_std_copts(),
4623 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004624 deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07004625 ":indirection",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004626 ":logging_utils",
Marat Dukhanab582382020-07-06 13:32:08 -07004627 ":packing",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004628 "@FP16",
4629 "@FXdiv",
4630 "@clog",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004631 "@pthreadpool",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004632 ],
4633)
4634
Marat Dukhan10a38082020-04-17 03:58:35 -07004635xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07004636 name = "operators_test_mode",
4637 srcs = OPERATOR_SRCS + [
4638 "src/memory.c",
4639 "src/operator-delete.c",
4640 ],
4641 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
4642 copts = LOGGING_COPTS + [
4643 "-Isrc",
4644 "-Iinclude",
4645 "-UNDEBUG",
4646 "-DXNN_TEST_MODE=1",
4647 ] + select({
4648 ":debug_build": [],
4649 "//conditions:default": xnnpack_min_size_copts(),
4650 }) + select({
4651 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4652 "//conditions:default": [],
4653 }),
4654 gcc_copts = xnnpack_gcc_std_copts(),
4655 msvc_copts = xnnpack_msvc_std_copts(),
4656 deps = [
4657 ":indirection_test_mode",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004658 ":logging_utils",
Marat Dukhanab582382020-07-06 13:32:08 -07004659 ":packing_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004660 "@FP16",
4661 "@FXdiv",
4662 "@clog",
4663 "@pthreadpool",
4664 ],
4665)
4666
4667xnnpack_cc_library(
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004668 name = "XNNPACK",
4669 srcs = [
4670 "src/init.c",
Marat Dukhanccfdbd12020-02-03 14:27:45 -08004671 "src/runtime.c",
4672 "src/subgraph.c",
4673 "src/tensor.c",
Marat Dukhanf03da0d2020-06-10 16:00:20 -07004674 ] + SUBGRAPH_SRCS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004675 hdrs = ["include/xnnpack.h"],
4676 copts = LOGGING_COPTS + [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004677 "-Isrc",
4678 "-Iinclude",
4679 ] + select({
4680 ":debug_build": [],
4681 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07004682 }) + select({
4683 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4684 "//conditions:default": [],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004685 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07004686 gcc_copts = xnnpack_gcc_std_copts(),
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004687 includes = ["include"],
Marat Dukhan10a38082020-04-17 03:58:35 -07004688 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004689 visibility = xnnpack_visibility(),
4690 deps = [
4691 ":enable_assembly",
Marat Dukhan9de90e02020-06-18 16:04:12 -07004692 ":enable_sparse",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004693 ":logging_utils",
Chao Mei6ddfc602020-05-13 22:29:36 -07004694 ":memory_planner",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004695 ":operator_run",
4696 ":operators",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004697 ":ukernels",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004698 "@clog",
Marat Dukhanab2946c2020-05-21 20:04:13 -07004699 "@FP16",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004700 "@pthreadpool",
Marat Dukhand343c222019-10-07 09:22:14 -07004701 ] + select({
4702 ":emscripten": [],
4703 "//conditions:default": ["@cpuinfo"],
4704 }),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004705)
4706
Marat Dukhan10a38082020-04-17 03:58:35 -07004707xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07004708 name = "XNNPACK_test_mode",
4709 srcs = [
4710 "src/init.c",
4711 "src/runtime.c",
4712 "src/subgraph.c",
4713 "src/tensor.c",
Marat Dukhanf03da0d2020-06-10 16:00:20 -07004714 ] + SUBGRAPH_SRCS,
Marat Dukhan33fcf782020-05-24 14:27:15 -07004715 hdrs = ["include/xnnpack.h"],
4716 copts = LOGGING_COPTS + [
4717 "-Isrc",
4718 "-Iinclude",
4719 "-UNDEBUG",
4720 "-DXNN_TEST_MODE=1",
4721 ] + select({
4722 ":debug_build": [],
4723 "//conditions:default": xnnpack_min_size_copts(),
4724 }) + select({
4725 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4726 "//conditions:default": [],
4727 }),
4728 gcc_copts = xnnpack_gcc_std_copts(),
4729 includes = ["include"],
4730 msvc_copts = xnnpack_msvc_std_copts(),
4731 visibility = xnnpack_visibility(),
4732 deps = [
4733 ":enable_assembly",
Marat Dukhan9de90e02020-06-18 16:04:12 -07004734 ":enable_sparse",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004735 ":logging_utils",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004736 ":memory_planner_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004737 ":operator_run_test_mode",
4738 ":operators_test_mode",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004739 ":ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004740 "@clog",
4741 "@FP16",
4742 "@pthreadpool",
4743 ] + select({
4744 ":emscripten": [],
4745 "//conditions:default": ["@cpuinfo"],
4746 }),
4747)
4748
4749xnnpack_cc_library(
Marat Dukhanae046f52020-06-15 13:16:14 -07004750 name = "xnnpack_f32",
4751 srcs = [
4752 "src/init.c",
4753 "src/runtime.c",
4754 "src/subgraph.c",
4755 "src/tensor.c",
4756 ] + SUBGRAPH_SRCS,
4757 hdrs = ["include/xnnpack.h"],
4758 copts = LOGGING_COPTS + [
4759 "-Isrc",
4760 "-Iinclude",
4761 ] + select({
4762 ":debug_build": [],
4763 "//conditions:default": xnnpack_min_size_copts(),
4764 }) + select({
4765 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4766 "//conditions:default": [],
4767 }),
4768 defines = [
Marat Dukhan16f1e1a2020-08-04 16:38:22 -07004769 "XNN_NO_QS8_OPERATORS",
Marat Dukhan08b7a972020-07-14 18:17:29 -07004770 "XNN_NO_QU8_OPERATORS",
Marat Dukhanae046f52020-06-15 13:16:14 -07004771 "XNN_NO_U8_OPERATORS",
4772 "XNN_NO_X8_OPERATORS",
4773 "XNN_NO_F16_OPERATORS",
4774 "XNN_NO_X16_OPERATORS",
4775 ],
4776 gcc_copts = xnnpack_gcc_std_copts(),
4777 includes = ["include"],
4778 msvc_copts = xnnpack_msvc_std_copts(),
4779 visibility = xnnpack_visibility(),
4780 deps = [
4781 ":enable_assembly",
Marat Dukhan9de90e02020-06-18 16:04:12 -07004782 ":enable_sparse",
Marat Dukhanae046f52020-06-15 13:16:14 -07004783 ":logging_utils",
4784 ":memory_planner",
4785 ":operator_run",
4786 ":operators",
Marat Dukhan86df3842020-06-17 14:03:31 -07004787 ":ukernels_max_compatibility",
Marat Dukhanae046f52020-06-15 13:16:14 -07004788 "@clog",
4789 "@FP16",
4790 "@pthreadpool",
4791 ] + select({
4792 ":emscripten": [],
4793 "//conditions:default": ["@cpuinfo"],
4794 }),
4795)
4796
4797xnnpack_cc_library(
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004798 name = "xnnpack_operators_nhwc_f32",
4799 srcs = [
4800 "src/init.c",
4801 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07004802 hdrs = ["include/xnnpack.h"],
4803 copts = LOGGING_COPTS + [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004804 "-Isrc",
4805 "-Iinclude",
4806 ] + select({
4807 ":debug_build": [],
4808 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07004809 }) + select({
4810 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4811 "//conditions:default": [],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004812 }),
4813 defines = [
Marat Dukhan16f1e1a2020-08-04 16:38:22 -07004814 "XNN_NO_QS8_OPERATORS",
Marat Dukhan08b7a972020-07-14 18:17:29 -07004815 "XNN_NO_QU8_OPERATORS",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004816 "XNN_NO_U8_OPERATORS",
4817 "XNN_NO_X8_OPERATORS",
Marat Dukhanefc47b82019-11-18 09:25:38 -08004818 "XNN_NO_NCHW_OPERATORS",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004819 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07004820 gcc_copts = xnnpack_gcc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004821 includes = ["include"],
Marat Dukhan10a38082020-04-17 03:58:35 -07004822 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004823 visibility = xnnpack_visibility(),
4824 deps = [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004825 ":enable_assembly",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004826 ":logging_utils",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004827 ":operator_run",
4828 ":operators",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004829 ":ukernels",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004830 "@clog",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004831 "@pthreadpool",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004832 ] + select({
4833 ":emscripten": [],
4834 "//conditions:default": ["@cpuinfo"],
4835 }),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004836)
4837
Marat Dukhancf056b22019-10-07 10:26:29 -07004838xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004839 name = "bench_utils",
4840 srcs = ["bench/utils.cc"],
4841 hdrs = ["bench/utils.h"],
Marat Dukhanbad48fe2019-11-04 10:35:22 -08004842 deps = [
4843 "@com_google_benchmark//:benchmark",
4844 "@cpuinfo",
4845 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004846)
4847
Frank Barchard7e955972019-10-11 10:34:25 -07004848######################### Benchmarks for micro-kernels #########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07004849
4850xnnpack_benchmark(
Marat Dukhan595e1702020-07-31 10:12:52 -07004851 name = "qs8_gemm_bench",
4852 srcs = [
4853 "bench/gemm.h",
4854 "bench/qs8-gemm.cc",
4855 "src/xnnpack/AlignedAllocator.h",
4856 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Frank Barchard31328cb2020-10-12 11:55:18 -07004857 copts = xnnpack_optional_ruy_copts() + xnnpack_optional_gemmlowp_copts(),
4858 deps = MICROKERNEL_BENCHMARK_DEPS + [":packing"] + xnnpack_optional_ruy_deps() + xnnpack_optional_gemmlowp_deps(),
Marat Dukhan595e1702020-07-31 10:12:52 -07004859)
4860
4861xnnpack_benchmark(
Marat Dukhan56bdd4a2020-08-03 19:47:04 -07004862 name = "qs8_requantization_bench",
4863 srcs = [
4864 "bench/qs8-requantization.cc",
Marat Dukhan56bdd4a2020-08-03 19:47:04 -07004865 "src/xnnpack/AlignedAllocator.h",
Frank Barchard04336c12020-10-22 16:48:55 -07004866 "src/xnnpack/requantization-stubs.h",
Marat Dukhan56bdd4a2020-08-03 19:47:04 -07004867 ] + MICROKERNEL_BENCHMARK_HDRS,
4868 deps = MICROKERNEL_BENCHMARK_DEPS,
4869)
4870
4871xnnpack_benchmark(
Marat Dukhan08b7a972020-07-14 18:17:29 -07004872 name = "qu8_gemm_bench",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004873 srcs = [
4874 "bench/gemm.h",
Marat Dukhan08b7a972020-07-14 18:17:29 -07004875 "bench/qu8-gemm.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004876 "src/xnnpack/AlignedAllocator.h",
4877 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004878 copts = xnnpack_optional_ruy_copts() + xnnpack_optional_gemmlowp_copts(),
Marat Dukhanab582382020-07-06 13:32:08 -07004879 deps = MICROKERNEL_BENCHMARK_DEPS + [":packing"] + xnnpack_optional_ruy_deps() + xnnpack_optional_gemmlowp_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004880)
4881
4882xnnpack_benchmark(
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07004883 name = "qu8_requantization_bench",
4884 srcs = [
4885 "bench/qu8-requantization.cc",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07004886 "src/xnnpack/AlignedAllocator.h",
Frank Barchard04336c12020-10-22 16:48:55 -07004887 "src/xnnpack/requantization-stubs.h",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07004888 ] + MICROKERNEL_BENCHMARK_HDRS,
4889 deps = MICROKERNEL_BENCHMARK_DEPS,
4890)
4891
4892xnnpack_benchmark(
Frank Barchard40d20fe2020-05-05 00:37:45 -07004893 name = "f16_igemm_bench",
4894 srcs = [
4895 "bench/f16-igemm.cc",
4896 "bench/conv.h",
4897 "bench/google/conv.h",
4898 "src/xnnpack/AlignedAllocator.h",
4899 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004900 deps = MICROKERNEL_BENCHMARK_DEPS + [
4901 ":indirection",
4902 ":packing",
4903 ],
Frank Barchard40d20fe2020-05-05 00:37:45 -07004904)
4905
4906xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004907 name = "f16_gemm_bench",
4908 srcs = [
4909 "bench/f16-gemm.cc",
4910 "bench/gemm.h",
4911 "src/xnnpack/AlignedAllocator.h",
4912 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004913 deps = MICROKERNEL_BENCHMARK_DEPS + [
4914 ":packing",
4915 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004916)
4917
4918xnnpack_benchmark(
Marat Dukhanbdb56f52020-02-05 21:42:49 -08004919 name = "f16_spmm_bench",
4920 srcs = [
4921 "bench/f16-spmm.cc",
Marat Dukhan1631e3e2020-12-06 19:29:31 -08004922 "bench/spmm.h",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08004923 "src/xnnpack/AlignedAllocator.h",
4924 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanbdb56f52020-02-05 21:42:49 -08004925 deps = MICROKERNEL_BENCHMARK_DEPS,
4926)
4927
4928xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004929 name = "f32_igemm_bench",
4930 srcs = [
4931 "bench/f32-igemm.cc",
4932 "bench/conv.h",
4933 "src/xnnpack/AlignedAllocator.h",
4934 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004935 deps = MICROKERNEL_BENCHMARK_DEPS + [
4936 ":indirection",
4937 ":packing",
4938 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004939)
4940
4941xnnpack_benchmark(
Frank Barchardc2010062020-07-19 10:04:12 -07004942 name = "f16_relu_bench",
4943 srcs = [
4944 "bench/f16-relu.cc",
4945 "src/xnnpack/AlignedAllocator.h",
4946 ] + MICROKERNEL_BENCHMARK_HDRS,
4947 deps = MICROKERNEL_BENCHMARK_DEPS,
4948)
4949
4950xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004951 name = "f32_conv_hwc_bench",
4952 srcs = [
4953 "bench/f32-conv-hwc.cc",
4954 "bench/dconv.h",
4955 "src/xnnpack/AlignedAllocator.h",
4956 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004957 deps = MICROKERNEL_BENCHMARK_DEPS + [
4958 ":packing",
4959 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004960)
4961
4962xnnpack_benchmark(
Marat Dukhan1f29b802020-05-15 23:46:39 -07004963 name = "f32_conv_hwc2chw_bench",
Erich Elsen563df5f2019-10-23 08:02:21 -07004964 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07004965 "bench/f32-conv-hwc2chw.cc",
Erich Elsen563df5f2019-10-23 08:02:21 -07004966 "bench/dconv.h",
4967 "src/xnnpack/AlignedAllocator.h",
4968 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004969 deps = MICROKERNEL_BENCHMARK_DEPS + [
4970 ":packing",
4971 ],
Erich Elsen563df5f2019-10-23 08:02:21 -07004972)
4973
4974xnnpack_benchmark(
Frank Barchard5a599a62020-06-04 20:12:44 -07004975 name = "f16_dwconv_bench",
4976 srcs = [
4977 "bench/f16-dwconv.cc",
4978 "bench/dwconv.h",
4979 "bench/google/dwconv.h",
4980 "src/xnnpack/AlignedAllocator.h",
4981 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004982 deps = MICROKERNEL_BENCHMARK_DEPS + [
4983 ":indirection",
4984 ":packing",
4985 ],
Frank Barchard5a599a62020-06-04 20:12:44 -07004986)
4987
4988xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004989 name = "f32_dwconv_bench",
4990 srcs = [
4991 "bench/f32-dwconv.cc",
4992 "bench/dwconv.h",
4993 "src/xnnpack/AlignedAllocator.h",
4994 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004995 deps = MICROKERNEL_BENCHMARK_DEPS + [
4996 ":indirection",
4997 ":packing",
4998 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004999)
5000
5001xnnpack_benchmark(
Marat Dukhanbf715f92020-10-23 20:17:00 -07005002 name = "f32_dwconv2d_chw_bench",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005003 srcs = [
Marat Dukhanbf715f92020-10-23 20:17:00 -07005004 "bench/f32-dwconv2d-chw.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005005 "bench/dwconv.h",
5006 "src/xnnpack/AlignedAllocator.h",
5007 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005008 deps = MICROKERNEL_BENCHMARK_DEPS + [
5009 ":indirection",
5010 ":packing",
5011 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005012)
5013
5014xnnpack_benchmark(
5015 name = "f32_gemm_bench",
5016 srcs = [
5017 "bench/f32-gemm.cc",
5018 "bench/gemm.h",
5019 "src/xnnpack/AlignedAllocator.h",
5020 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07005021 copts = xnnpack_optional_ruy_copts(),
Marat Dukhanab582382020-07-06 13:32:08 -07005022 deps = MICROKERNEL_BENCHMARK_DEPS + [":packing"] + xnnpack_optional_ruy_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07005023)
5024
5025xnnpack_benchmark(
Marat Dukhana11ca342020-06-25 23:45:07 -07005026 name = "f32_hswish_bench",
5027 srcs = [
5028 "bench/f32-hswish.cc",
5029 "src/xnnpack/AlignedAllocator.h",
5030 ] + MICROKERNEL_BENCHMARK_HDRS,
5031 deps = MICROKERNEL_BENCHMARK_DEPS,
5032)
5033
5034xnnpack_benchmark(
Marat Dukhan4c4eb002019-12-08 21:27:49 -08005035 name = "f32_raddexpminusmax_bench",
5036 srcs = [
5037 "bench/f32-raddexpminusmax.cc",
5038 "src/xnnpack/AlignedAllocator.h",
5039 ] + MICROKERNEL_BENCHMARK_HDRS,
5040 deps = MICROKERNEL_BENCHMARK_DEPS,
5041)
5042
5043xnnpack_benchmark(
5044 name = "f32_raddextexp_bench",
5045 srcs = [
5046 "bench/f32-raddextexp.cc",
5047 "src/xnnpack/AlignedAllocator.h",
5048 ] + MICROKERNEL_BENCHMARK_HDRS,
5049 deps = MICROKERNEL_BENCHMARK_DEPS,
5050)
5051
5052xnnpack_benchmark(
5053 name = "f32_raddstoreexpminusmax_bench",
5054 srcs = [
5055 "bench/f32-raddstoreexpminusmax.cc",
5056 "src/xnnpack/AlignedAllocator.h",
5057 ] + MICROKERNEL_BENCHMARK_HDRS,
5058 deps = MICROKERNEL_BENCHMARK_DEPS,
5059)
5060
5061xnnpack_benchmark(
Frank Barchard3c4a9522020-07-18 16:11:22 -07005062 name = "f32_relu_bench",
5063 srcs = [
5064 "bench/f32-relu.cc",
5065 "src/xnnpack/AlignedAllocator.h",
5066 ] + MICROKERNEL_BENCHMARK_HDRS,
5067 deps = MICROKERNEL_BENCHMARK_DEPS,
5068)
5069
5070xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005071 name = "f32_rmax_bench",
5072 srcs = [
5073 "bench/f32-rmax.cc",
5074 "src/xnnpack/AlignedAllocator.h",
5075 ] + MICROKERNEL_BENCHMARK_HDRS,
5076 deps = MICROKERNEL_BENCHMARK_DEPS,
5077)
5078
5079xnnpack_benchmark(
Marat Dukhan14bec502019-11-18 11:35:31 -08005080 name = "f32_sigmoid_bench",
5081 srcs = [
5082 "bench/f32-sigmoid.cc",
5083 "src/xnnpack/AlignedAllocator.h",
5084 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan14bec502019-11-18 11:35:31 -08005085 deps = MICROKERNEL_BENCHMARK_DEPS,
5086)
5087
5088xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005089 name = "f32_spmm_bench",
5090 srcs = [
5091 "bench/f32-spmm.cc",
Marat Dukhan1631e3e2020-12-06 19:29:31 -08005092 "bench/spmm.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005093 "src/xnnpack/AlignedAllocator.h",
5094 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005095 deps = MICROKERNEL_BENCHMARK_DEPS,
5096)
5097
5098xnnpack_benchmark(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08005099 name = "f32_softmax_bench",
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07005100 srcs = [
Marat Dukhanfd8e6892020-01-27 15:25:25 -08005101 "bench/f32-softmax.cc",
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07005102 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07005103 copts = xnnpack_optional_dnnl_copts(),
Marat Dukhan8d3c6932020-03-06 20:27:27 -08005104 deps = MICROKERNEL_BENCHMARK_DEPS + xnnpack_optional_dnnl_deps(),
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07005105)
5106
5107xnnpack_benchmark(
Marat Dukhaned6baaf2020-12-01 15:07:08 -08005108 name = "f32_velu_bench",
5109 srcs = [
5110 "bench/f32-velu.cc",
5111 "src/xnnpack/AlignedAllocator.h",
5112 ] + MICROKERNEL_BENCHMARK_HDRS,
5113 deps = MICROKERNEL_BENCHMARK_DEPS,
5114)
5115
5116xnnpack_benchmark(
Marat Dukhan4c4eb002019-12-08 21:27:49 -08005117 name = "f32_vscaleexpminusmax_bench",
5118 srcs = [
5119 "bench/f32-vscaleexpminusmax.cc",
5120 "src/xnnpack/AlignedAllocator.h",
5121 ] + MICROKERNEL_BENCHMARK_HDRS,
5122 deps = MICROKERNEL_BENCHMARK_DEPS,
5123)
5124
5125xnnpack_benchmark(
5126 name = "f32_vscaleextexp_bench",
5127 srcs = [
5128 "bench/f32-vscaleextexp.cc",
5129 "src/xnnpack/AlignedAllocator.h",
5130 ] + MICROKERNEL_BENCHMARK_HDRS,
5131 deps = MICROKERNEL_BENCHMARK_DEPS,
5132)
5133
5134xnnpack_benchmark(
Marat Dukhanf4db2f32020-06-30 10:55:30 -07005135 name = "f32_vsqrt_bench",
5136 srcs = [
5137 "bench/f32-vsqrt.cc",
5138 "src/xnnpack/AlignedAllocator.h",
5139 ] + MICROKERNEL_BENCHMARK_HDRS,
5140 deps = MICROKERNEL_BENCHMARK_DEPS,
5141)
5142
5143xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005144 name = "f32_im2col_gemm_bench",
5145 srcs = [
5146 "bench/f32-im2col-gemm.cc",
5147 "bench/conv.h",
5148 "src/xnnpack/AlignedAllocator.h",
5149 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005150 deps = MICROKERNEL_BENCHMARK_DEPS + [
5151 ":im2col",
5152 ":packing",
5153 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005154)
5155
Marat Dukhanfe7acb62020-03-09 19:30:05 -07005156xnnpack_benchmark(
Marat Dukhanffbf96a2020-05-14 02:59:08 -07005157 name = "rounding_bench",
5158 srcs = [
5159 "bench/rounding.cc",
Marat Dukhanffbf96a2020-05-14 02:59:08 -07005160 "src/xnnpack/AlignedAllocator.h",
Frank Barchard04336c12020-10-22 16:48:55 -07005161 "src/xnnpack/math-stubs.h",
Marat Dukhanffbf96a2020-05-14 02:59:08 -07005162 ] + MICROKERNEL_BENCHMARK_HDRS,
5163 deps = MICROKERNEL_BENCHMARK_DEPS,
5164)
5165
Marat Dukhan08c4a432019-10-03 09:29:21 -07005166########################### Benchmarks for operators ###########################
5167
5168xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005169 name = "average_pooling_bench",
5170 srcs = ["bench/average-pooling.cc"],
Marat Dukhan7a16d8b2020-03-11 04:22:44 -07005171 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07005172 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07005173 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07005174)
5175
5176xnnpack_benchmark(
Marat Dukhan87727142020-06-24 15:24:10 -07005177 name = "bankers_rounding_bench",
5178 srcs = ["bench/bankers-rounding.cc"],
5179 copts = xnnpack_optional_tflite_copts(),
5180 tags = ["nowin32"],
5181 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
5182)
5183
5184xnnpack_benchmark(
5185 name = "ceiling_bench",
5186 srcs = ["bench/ceiling.cc"],
5187 copts = xnnpack_optional_tflite_copts(),
5188 tags = ["nowin32"],
5189 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
5190)
5191
5192xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005193 name = "channel_shuffle_bench",
5194 srcs = ["bench/channel-shuffle.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07005195 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005196)
5197
5198xnnpack_benchmark(
5199 name = "convolution_bench",
5200 srcs = ["bench/convolution.cc"],
5201 copts = xnnpack_optional_tflite_copts() + xnnpack_optional_armcl_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07005202 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07005203 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps() + xnnpack_optional_armcl_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07005204)
5205
5206xnnpack_benchmark(
5207 name = "deconvolution_bench",
5208 srcs = ["bench/deconvolution.cc"],
5209 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07005210 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07005211 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07005212)
5213
5214xnnpack_benchmark(
Marat Dukhanb6bd4bc2020-12-01 17:01:40 -08005215 name = "elu_bench",
5216 srcs = ["bench/elu.cc"],
5217 copts = xnnpack_optional_tflite_copts(),
5218 tags = ["nowin32"],
5219 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
5220)
5221
5222xnnpack_benchmark(
Marat Dukhan87727142020-06-24 15:24:10 -07005223 name = "floor_bench",
5224 srcs = ["bench/floor.cc"],
5225 copts = xnnpack_optional_tflite_copts(),
5226 tags = ["nowin32"],
5227 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
5228)
5229
5230xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005231 name = "global_average_pooling_bench",
5232 srcs = ["bench/global-average-pooling.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07005233 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005234)
5235
5236xnnpack_benchmark(
Marat Dukhanad352602020-06-25 21:50:54 -07005237 name = "hardswish_bench",
5238 srcs = ["bench/hardswish.cc"],
5239 copts = xnnpack_optional_tflite_copts(),
5240 tags = ["nowin32"],
5241 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
5242)
5243
5244xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005245 name = "max_pooling_bench",
5246 srcs = ["bench/max-pooling.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07005247 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005248)
5249
5250xnnpack_benchmark(
5251 name = "sigmoid_bench",
5252 srcs = ["bench/sigmoid.cc"],
Marat Dukhanc3b9e862019-11-17 13:18:54 -08005253 copts = xnnpack_optional_tflite_copts(),
Marat Dukhanca2ba702020-04-24 01:31:47 -07005254 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07005255 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07005256)
5257
5258xnnpack_benchmark(
Marat Dukhan95b22432019-10-30 16:30:14 -07005259 name = "prelu_bench",
5260 srcs = ["bench/prelu.cc"],
5261 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07005262 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07005263 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan95b22432019-10-30 16:30:14 -07005264)
5265
5266xnnpack_benchmark(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08005267 name = "softmax_bench",
5268 srcs = ["bench/softmax.cc"],
Marat Dukhan9c0db962020-01-28 12:30:14 -08005269 copts = xnnpack_optional_tflite_copts(),
Marat Dukhanca2ba702020-04-24 01:31:47 -07005270 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07005271 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07005272)
5273
Marat Dukhan87727142020-06-24 15:24:10 -07005274xnnpack_benchmark(
Marat Dukhan6804bbd2020-06-30 19:26:11 -07005275 name = "square_root_bench",
5276 srcs = ["bench/square-root.cc"],
5277 copts = xnnpack_optional_tflite_copts(),
5278 tags = ["nowin32"],
5279 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
5280)
5281
5282xnnpack_benchmark(
Marat Dukhan87727142020-06-24 15:24:10 -07005283 name = "truncation_bench",
5284 srcs = ["bench/truncation.cc"],
5285 deps = OPERATOR_BENCHMARK_DEPS,
5286)
5287
Marat Dukhanc068bb62019-10-04 13:24:39 -07005288############################# End-to-end benchmarks ############################
5289
5290cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07005291 name = "fp32_mobilenet_v1",
5292 srcs = ["models/fp32-mobilenet-v1.cc"],
Marat Dukhanc068bb62019-10-04 13:24:39 -07005293 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08005294 copts = xnnpack_std_cxxopts(),
Marat Dukhanc068bb62019-10-04 13:24:39 -07005295 linkstatic = True,
5296 deps = [
5297 ":XNNPACK",
5298 "@pthreadpool",
5299 ],
5300)
5301
5302cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07005303 name = "fp16_mobilenet_v1",
5304 srcs = ["models/fp16-mobilenet-v1.cc"],
5305 hdrs = ["models/models.h"],
5306 copts = xnnpack_std_cxxopts(),
5307 linkstatic = True,
5308 deps = [
5309 ":XNNPACK",
5310 "@FP16",
5311 "@pthreadpool",
5312 ],
5313)
5314
5315cc_library(
Marat Dukhan0743cdf2020-08-04 18:52:07 -07005316 name = "qs8_mobilenet_v1",
5317 srcs = ["models/qs8-mobilenet-v1.cc"],
5318 hdrs = ["models/models.h"],
5319 copts = xnnpack_std_cxxopts(),
5320 linkstatic = True,
5321 deps = [
5322 ":XNNPACK",
5323 "@pthreadpool",
5324 ],
5325)
5326
5327cc_library(
Marat Dukhan70a96182020-09-03 17:13:58 -07005328 name = "qs8_mobilenet_v2",
5329 srcs = ["models/qs8-mobilenet-v2.cc"],
5330 hdrs = ["models/models.h"],
5331 copts = xnnpack_std_cxxopts(),
5332 linkstatic = True,
5333 deps = [
5334 ":XNNPACK",
5335 "@pthreadpool",
5336 ],
5337)
5338
5339cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07005340 name = "fp32_mobilenet_v2",
5341 srcs = ["models/fp32-mobilenet-v2.cc"],
Marat Dukhanc068bb62019-10-04 13:24:39 -07005342 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08005343 copts = xnnpack_std_cxxopts(),
Marat Dukhanc068bb62019-10-04 13:24:39 -07005344 linkstatic = True,
5345 deps = [
5346 ":XNNPACK",
5347 "@pthreadpool",
5348 ],
5349)
5350
Marat Dukhanc08cdf52019-12-09 09:17:51 -08005351cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07005352 name = "fp16_mobilenet_v2",
5353 srcs = ["models/fp16-mobilenet-v2.cc"],
5354 hdrs = ["models/models.h"],
5355 copts = xnnpack_std_cxxopts(),
5356 linkstatic = True,
5357 deps = [
5358 ":XNNPACK",
5359 "@FP16",
5360 "@pthreadpool",
5361 ],
5362)
5363
5364cc_library(
5365 name = "fp32_mobilenet_v3_large",
5366 srcs = ["models/fp32-mobilenet-v3-large.cc"],
Marat Dukhanc08cdf52019-12-09 09:17:51 -08005367 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08005368 copts = xnnpack_std_cxxopts(),
Marat Dukhanc08cdf52019-12-09 09:17:51 -08005369 linkstatic = True,
5370 deps = [
5371 ":XNNPACK",
5372 "@pthreadpool",
5373 ],
5374)
5375
5376cc_library(
Marat Dukhan66f3ccd2020-09-14 16:09:38 -07005377 name = "fp16_mobilenet_v3_large",
5378 srcs = ["models/fp16-mobilenet-v3-large.cc"],
5379 hdrs = ["models/models.h"],
5380 copts = xnnpack_std_cxxopts(),
5381 linkstatic = True,
5382 deps = [
5383 ":XNNPACK",
5384 "@FP16",
5385 "@pthreadpool",
5386 ],
5387)
5388
5389cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07005390 name = "fp32_mobilenet_v3_small",
5391 srcs = ["models/fp32-mobilenet-v3-small.cc"],
Marat Dukhanc08cdf52019-12-09 09:17:51 -08005392 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08005393 copts = xnnpack_std_cxxopts(),
Marat Dukhanc08cdf52019-12-09 09:17:51 -08005394 linkstatic = True,
5395 deps = [
5396 ":XNNPACK",
5397 "@pthreadpool",
5398 ],
5399)
5400
Marat Dukhan66f3ccd2020-09-14 16:09:38 -07005401cc_library(
5402 name = "fp16_mobilenet_v3_small",
5403 srcs = ["models/fp16-mobilenet-v3-small.cc"],
5404 hdrs = ["models/models.h"],
5405 copts = xnnpack_std_cxxopts(),
5406 linkstatic = True,
5407 deps = [
5408 ":XNNPACK",
5409 "@FP16",
5410 "@pthreadpool",
5411 ],
5412)
5413
Marat Dukhanc068bb62019-10-04 13:24:39 -07005414xnnpack_benchmark(
Marat Dukhanef4416e2019-10-31 13:44:40 -07005415 name = "f32_dwconv_e2e_bench",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08005416 srcs = [
5417 "bench/f32-dwconv-e2e.cc",
5418 "bench/end2end.h",
5419 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanef4416e2019-10-31 13:44:40 -07005420 deps = MICROKERNEL_BENCHMARK_DEPS + [
5421 ":XNNPACK",
Marat Dukhan270a2c42020-06-26 16:45:52 -07005422 ":fp32_mobilenet_v1",
5423 ":fp32_mobilenet_v2",
5424 ":fp32_mobilenet_v3_large",
5425 ":fp32_mobilenet_v3_small",
Marat Dukhanef4416e2019-10-31 13:44:40 -07005426 ],
5427)
5428
5429xnnpack_benchmark(
Marat Dukhan5f18d262019-10-31 10:24:14 -07005430 name = "f32_gemm_e2e_bench",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08005431 srcs = [
5432 "bench/f32-gemm-e2e.cc",
5433 "bench/end2end.h",
5434 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan5f18d262019-10-31 10:24:14 -07005435 deps = MICROKERNEL_BENCHMARK_DEPS + [
5436 ":XNNPACK",
Marat Dukhan270a2c42020-06-26 16:45:52 -07005437 ":fp32_mobilenet_v1",
5438 ":fp32_mobilenet_v2",
5439 ":fp32_mobilenet_v3_large",
5440 ":fp32_mobilenet_v3_small",
Marat Dukhan5f18d262019-10-31 10:24:14 -07005441 ],
5442)
5443
5444xnnpack_benchmark(
Frank Barcharddc909cb2021-02-08 13:59:31 -08005445 name = "qs8_gemm_e2e_bench",
5446 srcs = [
5447 "bench/qs8-gemm-e2e.cc",
5448 "bench/end2end.h",
5449 ] + MICROKERNEL_BENCHMARK_HDRS,
5450 deps = MICROKERNEL_BENCHMARK_DEPS + [
5451 ":XNNPACK",
5452 ":qs8_mobilenet_v1",
5453 ":qs8_mobilenet_v2",
5454 ],
5455)
5456
5457xnnpack_benchmark(
Marat Dukhanc068bb62019-10-04 13:24:39 -07005458 name = "end2end_bench",
5459 srcs = ["bench/end2end.cc"],
5460 deps = [
5461 ":XNNPACK",
Frank Barchardc712fa42019-10-31 14:00:21 -07005462 ":bench_utils",
Marat Dukhan270a2c42020-06-26 16:45:52 -07005463 ":fp16_mobilenet_v1",
5464 ":fp16_mobilenet_v2",
Marat Dukhan66f3ccd2020-09-14 16:09:38 -07005465 ":fp16_mobilenet_v3_large",
5466 ":fp16_mobilenet_v3_small",
Marat Dukhan270a2c42020-06-26 16:45:52 -07005467 ":fp32_mobilenet_v1",
5468 ":fp32_mobilenet_v2",
5469 ":fp32_mobilenet_v3_large",
5470 ":fp32_mobilenet_v3_small",
Marat Dukhan0743cdf2020-08-04 18:52:07 -07005471 ":qs8_mobilenet_v1",
Marat Dukhan70a96182020-09-03 17:13:58 -07005472 ":qs8_mobilenet_v2",
Marat Dukhanc068bb62019-10-04 13:24:39 -07005473 "@pthreadpool",
5474 ],
5475)
5476
Marat Dukhan6adff4e2019-10-14 18:32:07 -07005477#################### Accuracy evaluation for math functions ####################
5478
5479xnnpack_benchmark(
Marat Dukhancf67ec62020-12-14 09:56:37 -08005480 name = "f32_exp_ulp_eval",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07005481 srcs = [
Marat Dukhancf67ec62020-12-14 09:56:37 -08005482 "eval/f32-exp-ulp.cc",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07005483 "src/xnnpack/AlignedAllocator.h",
5484 ] + ACCURACY_EVAL_HDRS,
Marat Dukhan3a305212020-12-06 19:24:27 -08005485 deps = ACCURACY_EVAL_DEPS + [
5486 ":bench_utils",
5487 "@cpuinfo",
5488 ],
Marat Dukhan6adff4e2019-10-14 18:32:07 -07005489)
5490
Marat Dukhan515c9772019-10-17 18:07:57 -07005491xnnpack_benchmark(
Marat Dukhancf67ec62020-12-14 09:56:37 -08005492 name = "f32_expminus_ulp_eval",
Marat Dukhan515c9772019-10-17 18:07:57 -07005493 srcs = [
Marat Dukhancf67ec62020-12-14 09:56:37 -08005494 "eval/f32-expminus-ulp.cc",
Marat Dukhan515c9772019-10-17 18:07:57 -07005495 "src/xnnpack/AlignedAllocator.h",
5496 ] + ACCURACY_EVAL_HDRS,
Marat Dukhan3a305212020-12-06 19:24:27 -08005497 deps = ACCURACY_EVAL_DEPS + [
5498 ":bench_utils",
5499 "@cpuinfo",
5500 ],
Marat Dukhan515c9772019-10-17 18:07:57 -07005501)
5502
Marat Dukhan98ba4412019-10-23 02:14:28 -07005503xnnpack_benchmark(
Marat Dukhancf67ec62020-12-14 09:56:37 -08005504 name = "f32_expm1minus_ulp_eval",
Marat Dukhana438aca2020-11-20 15:45:01 -08005505 srcs = [
Marat Dukhancf67ec62020-12-14 09:56:37 -08005506 "eval/f32-expm1minus-ulp.cc",
Marat Dukhana438aca2020-11-20 15:45:01 -08005507 "src/xnnpack/AlignedAllocator.h",
5508 ] + ACCURACY_EVAL_HDRS,
Marat Dukhande390d42020-11-29 19:32:18 -08005509 deps = ACCURACY_EVAL_DEPS + [
5510 ":bench_utils",
Marat Dukhan3a305212020-12-06 19:24:27 -08005511 "@cpuinfo",
Marat Dukhande390d42020-11-29 19:32:18 -08005512 ],
Marat Dukhana438aca2020-11-20 15:45:01 -08005513)
5514
5515xnnpack_benchmark(
Marat Dukhancf67ec62020-12-14 09:56:37 -08005516 name = "f32_extexp_ulp_eval",
Marat Dukhan98ba4412019-10-23 02:14:28 -07005517 srcs = [
Marat Dukhancf67ec62020-12-14 09:56:37 -08005518 "eval/f32-extexp-ulp.cc",
Marat Dukhan98ba4412019-10-23 02:14:28 -07005519 "src/xnnpack/AlignedAllocator.h",
5520 ] + ACCURACY_EVAL_HDRS,
Marat Dukhan3a305212020-12-06 19:24:27 -08005521 deps = ACCURACY_EVAL_DEPS + [
5522 ":bench_utils",
5523 "@cpuinfo",
5524 ],
Marat Dukhan98ba4412019-10-23 02:14:28 -07005525)
5526
Marat Dukhanf44f0222020-12-14 11:53:27 -08005527xnnpack_benchmark(
5528 name = "f32_sigmoid_ulp_eval",
5529 srcs = [
5530 "eval/f32-sigmoid-ulp.cc",
5531 "src/xnnpack/AlignedAllocator.h",
5532 ] + ACCURACY_EVAL_HDRS,
5533 deps = ACCURACY_EVAL_DEPS + [
5534 ":bench_utils",
5535 "@cpuinfo",
5536 ],
5537)
5538
5539xnnpack_benchmark(
5540 name = "f32_sqrt_ulp_eval",
5541 srcs = [
5542 "eval/f32-sqrt-ulp.cc",
5543 "src/xnnpack/AlignedAllocator.h",
5544 ] + ACCURACY_EVAL_HDRS,
5545 deps = ACCURACY_EVAL_DEPS + [
5546 ":bench_utils",
5547 "@cpuinfo",
5548 ],
5549)
5550
5551################### Accuracy verification for math functions ##################
5552
5553xnnpack_unit_test(
Marat Dukhanf7291fc2020-12-15 11:02:50 -08005554 name = "f32_exp_eval",
5555 srcs = [
5556 "eval/f32-exp.cc",
5557 "src/xnnpack/AlignedAllocator.h",
5558 "src/xnnpack/math-stubs.h",
5559 ] + MICROKERNEL_TEST_HDRS,
5560 automatic = False,
5561 deps = MICROKERNEL_TEST_DEPS,
5562)
5563
5564xnnpack_unit_test(
Marat Dukhanf44f0222020-12-14 11:53:27 -08005565 name = "f32_expm1minus_eval",
5566 srcs = [
5567 "eval/f32-expm1minus.cc",
5568 "src/xnnpack/AlignedAllocator.h",
5569 "src/xnnpack/math-stubs.h",
5570 ] + MICROKERNEL_TEST_HDRS,
5571 automatic = False,
5572 deps = MICROKERNEL_TEST_DEPS,
5573)
5574
Marat Dukhan8853b822020-05-07 12:19:01 -07005575xnnpack_unit_test(
Marat Dukhand28a5a22020-12-14 15:27:22 -08005576 name = "f32_expminus_eval",
5577 srcs = [
5578 "eval/f32-expminus.cc",
5579 "src/xnnpack/AlignedAllocator.h",
5580 "src/xnnpack/math-stubs.h",
5581 ] + MICROKERNEL_TEST_HDRS,
5582 automatic = False,
5583 deps = MICROKERNEL_TEST_DEPS,
5584)
5585
5586xnnpack_unit_test(
Marat Dukhan8853b822020-05-07 12:19:01 -07005587 name = "f32_roundne_eval",
5588 srcs = [
5589 "eval/f32-roundne.cc",
5590 "src/xnnpack/AlignedAllocator.h",
5591 "src/xnnpack/math-stubs.h",
5592 ] + MICROKERNEL_TEST_HDRS,
Marat Dukhan22eed3d2020-05-11 20:13:37 -07005593 automatic = False,
Marat Dukhan8853b822020-05-07 12:19:01 -07005594 deps = MICROKERNEL_TEST_DEPS,
5595)
5596
Marat Dukhan2dbb9442020-05-12 20:43:43 -07005597xnnpack_unit_test(
Marat Dukhanc9852ba2020-05-13 17:21:29 -07005598 name = "f32_roundd_eval",
5599 srcs = [
5600 "eval/f32-roundd.cc",
5601 "src/xnnpack/AlignedAllocator.h",
5602 "src/xnnpack/math-stubs.h",
5603 ] + MICROKERNEL_TEST_HDRS,
5604 automatic = False,
5605 deps = MICROKERNEL_TEST_DEPS,
5606)
5607
5608xnnpack_unit_test(
5609 name = "f32_roundu_eval",
5610 srcs = [
5611 "eval/f32-roundu.cc",
5612 "src/xnnpack/AlignedAllocator.h",
5613 "src/xnnpack/math-stubs.h",
5614 ] + MICROKERNEL_TEST_HDRS,
5615 automatic = False,
5616 deps = MICROKERNEL_TEST_DEPS,
5617)
5618
5619xnnpack_unit_test(
Marat Dukhan2dbb9442020-05-12 20:43:43 -07005620 name = "f32_roundz_eval",
5621 srcs = [
5622 "eval/f32-roundz.cc",
5623 "src/xnnpack/AlignedAllocator.h",
5624 "src/xnnpack/math-stubs.h",
5625 ] + MICROKERNEL_TEST_HDRS,
Marat Dukhanc9852ba2020-05-13 17:21:29 -07005626 automatic = False,
Marat Dukhan2dbb9442020-05-12 20:43:43 -07005627 deps = MICROKERNEL_TEST_DEPS,
5628)
5629
Marat Dukhan08c4a432019-10-03 09:29:21 -07005630######################### Unit tests for micro-kernels #########################
5631
5632xnnpack_unit_test(
Marat Dukhande06f492020-04-09 00:19:31 -07005633 name = "f16_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005634 srcs = [
Marat Dukhande06f492020-04-09 00:19:31 -07005635 "test/f16-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005636 "test/gemm-microkernel-tester.h",
5637 "src/xnnpack/AlignedAllocator.h",
5638 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005639 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005640)
5641
5642xnnpack_unit_test(
Marat Dukhan355ab432020-04-09 19:01:52 -07005643 name = "f16_spmm_minmax_test",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08005644 srcs = [
Marat Dukhan355ab432020-04-09 19:01:52 -07005645 "test/f16-spmm-minmax.cc",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08005646 "test/spmm-microkernel-tester.h",
5647 "src/xnnpack/AlignedAllocator.h",
5648 ] + MICROKERNEL_TEST_HDRS,
5649 deps = MICROKERNEL_TEST_DEPS,
5650)
5651
5652xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005653 name = "f32_argmaxpool_test",
5654 srcs = [
5655 "test/f32-argmaxpool.cc",
5656 "test/argmaxpool-microkernel-tester.h",
5657 "src/xnnpack/AlignedAllocator.h",
5658 ] + MICROKERNEL_TEST_HDRS,
5659 deps = MICROKERNEL_TEST_DEPS,
5660)
5661
5662xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07005663 name = "f32_avgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005664 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07005665 "test/f32-avgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005666 "test/avgpool-microkernel-tester.h",
5667 "src/xnnpack/AlignedAllocator.h",
5668 ] + MICROKERNEL_TEST_HDRS,
5669 deps = MICROKERNEL_TEST_DEPS,
5670)
5671
5672xnnpack_unit_test(
Marat Dukhan660fd192020-03-10 04:55:30 -07005673 name = "f32_ibilinear_test",
Marat Dukhan35dacfb2019-11-07 19:18:16 -08005674 srcs = [
Marat Dukhan660fd192020-03-10 04:55:30 -07005675 "test/f32-ibilinear.cc",
5676 "test/ibilinear-microkernel-tester.h",
Marat Dukhan35dacfb2019-11-07 19:18:16 -08005677 "src/xnnpack/AlignedAllocator.h",
5678 ] + MICROKERNEL_TEST_HDRS,
5679 deps = MICROKERNEL_TEST_DEPS,
5680)
5681
5682xnnpack_unit_test(
XNNPACK Team21432672020-10-19 19:58:48 -07005683 name = "f32_ibilinear_chw_test",
5684 srcs = [
5685 "test/f32-ibilinear-chw.cc",
XNNPACK Team6be46b22020-10-22 23:34:54 -07005686 "test/ibilinear-microkernel-tester.h",
XNNPACK Team21432672020-10-19 19:58:48 -07005687 "src/xnnpack/AlignedAllocator.h",
5688 ] + MICROKERNEL_TEST_HDRS,
5689 deps = MICROKERNEL_TEST_DEPS,
5690)
5691
5692xnnpack_unit_test(
Frank Barchardb1966592020-05-12 13:47:06 -07005693 name = "f16_clamp_test",
5694 srcs = [
5695 "test/f16-clamp.cc",
5696 "test/clamp-microkernel-tester.h",
5697 ] + MICROKERNEL_TEST_HDRS,
5698 deps = MICROKERNEL_TEST_DEPS,
5699)
5700
5701xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005702 name = "f32_clamp_test",
5703 srcs = [
5704 "test/f32-clamp.cc",
5705 "test/clamp-microkernel-tester.h",
5706 ] + MICROKERNEL_TEST_HDRS,
5707 deps = MICROKERNEL_TEST_DEPS,
5708)
5709
5710xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07005711 name = "f32_igemm_test",
5712 srcs = [
5713 "test/f32-igemm.cc",
5714 "test/gemm-microkernel-tester.h",
5715 "src/xnnpack/AlignedAllocator.h",
5716 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005717 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan163a7e62020-04-09 04:19:26 -07005718)
5719
5720xnnpack_unit_test(
Marat Dukhan467f6362020-05-22 23:21:55 -07005721 name = "f32_igemm_relu_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005722 srcs = [
Marat Dukhan467f6362020-05-22 23:21:55 -07005723 "test/f32-igemm-relu.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005724 "test/gemm-microkernel-tester.h",
5725 "src/xnnpack/AlignedAllocator.h",
5726 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005727 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005728)
5729
5730xnnpack_unit_test(
Marat Dukhane207b7b2020-05-28 16:27:42 -07005731 name = "f32_igemm_minmax_test",
5732 srcs = [
5733 "test/f32-igemm-minmax.cc",
5734 "test/gemm-microkernel-tester.h",
5735 "src/xnnpack/AlignedAllocator.h",
5736 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005737 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhane207b7b2020-05-28 16:27:42 -07005738)
5739
5740xnnpack_unit_test(
Frank Barchardb0e4fae2020-05-04 15:27:51 -07005741 name = "f16_igemm_minmax_test",
5742 srcs = [
5743 "test/f16-igemm-minmax.cc",
5744 "test/gemm-microkernel-tester.h",
5745 "src/xnnpack/AlignedAllocator.h",
5746 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005747 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Frank Barchardb0e4fae2020-05-04 15:27:51 -07005748)
5749
5750xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005751 name = "f32_conv_hwc_test",
5752 srcs = [
5753 "test/f32-conv-hwc.cc",
5754 "test/conv-hwc-microkernel-tester.h",
5755 "src/xnnpack/AlignedAllocator.h",
5756 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005757 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005758)
5759
5760xnnpack_unit_test(
Marat Dukhan1f29b802020-05-15 23:46:39 -07005761 name = "f32_conv_hwc2chw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005762 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07005763 "test/f32-conv-hwc2chw.cc",
5764 "test/conv-hwc2chw-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005765 "src/xnnpack/AlignedAllocator.h",
5766 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005767 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005768)
5769
5770xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07005771 name = "f32_dwconv_test",
5772 srcs = [
5773 "test/f32-dwconv.cc",
5774 "test/dwconv-microkernel-tester.h",
5775 "src/xnnpack/AlignedAllocator.h",
5776 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005777 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan163a7e62020-04-09 04:19:26 -07005778)
5779
5780xnnpack_unit_test(
Frank Barchard5a599a62020-06-04 20:12:44 -07005781 name = "f16_dwconv_minmax_test",
5782 srcs = [
5783 "test/f16-dwconv-minmax.cc",
5784 "test/dwconv-microkernel-tester.h",
5785 "src/xnnpack/AlignedAllocator.h",
5786 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005787 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Frank Barchard5a599a62020-06-04 20:12:44 -07005788)
5789
5790xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07005791 name = "f32_dwconv_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005792 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07005793 "test/f32-dwconv-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005794 "test/dwconv-microkernel-tester.h",
5795 "src/xnnpack/AlignedAllocator.h",
5796 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005797 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005798)
5799
5800xnnpack_unit_test(
Marat Dukhanbf715f92020-10-23 20:17:00 -07005801 name = "f32_dwconv2d_chw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005802 srcs = [
Marat Dukhanbf715f92020-10-23 20:17:00 -07005803 "test/f32-dwconv2d-chw.cc",
5804 "test/dwconv2d-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005805 "src/xnnpack/AlignedAllocator.h",
5806 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005807 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005808)
5809
5810xnnpack_unit_test(
Frank Barchard0bb49a72020-06-04 11:35:11 -07005811 name = "f16_gavgpool_minmax_test",
5812 srcs = [
5813 "test/f16-gavgpool-minmax.cc",
5814 "test/gavgpool-microkernel-tester.h",
5815 "src/xnnpack/AlignedAllocator.h",
5816 ] + MICROKERNEL_TEST_HDRS,
5817 deps = MICROKERNEL_TEST_DEPS,
5818)
5819
5820xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07005821 name = "f32_gavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005822 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07005823 "test/f32-gavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005824 "test/gavgpool-microkernel-tester.h",
5825 "src/xnnpack/AlignedAllocator.h",
5826 ] + MICROKERNEL_TEST_HDRS,
5827 deps = MICROKERNEL_TEST_DEPS,
5828)
5829
5830xnnpack_unit_test(
Marat Dukhan1f29b802020-05-15 23:46:39 -07005831 name = "f32_gavgpool_cw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005832 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07005833 "test/f32-gavgpool-cw.cc",
5834 "test/gavgpool-cw-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005835 "src/xnnpack/AlignedAllocator.h",
5836 ] + MICROKERNEL_TEST_HDRS,
5837 deps = MICROKERNEL_TEST_DEPS,
5838)
5839
5840xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07005841 name = "f32_gemm_test",
5842 srcs = [
5843 "test/f32-gemm.cc",
5844 "test/gemm-microkernel-tester.h",
5845 "src/xnnpack/AlignedAllocator.h",
5846 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005847 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan163a7e62020-04-09 04:19:26 -07005848)
5849
5850xnnpack_unit_test(
Marat Dukhan467f6362020-05-22 23:21:55 -07005851 name = "f32_gemm_relu_test",
5852 srcs = [
5853 "test/f32-gemm-relu.cc",
5854 "test/gemm-microkernel-tester.h",
5855 "src/xnnpack/AlignedAllocator.h",
5856 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005857 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan467f6362020-05-22 23:21:55 -07005858)
5859
5860xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07005861 name = "f32_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005862 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07005863 "test/f32-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005864 "test/gemm-microkernel-tester.h",
5865 "src/xnnpack/AlignedAllocator.h",
5866 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005867 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005868)
5869
5870xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07005871 name = "f32_gemminc_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005872 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07005873 "test/f32-gemminc-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005874 "test/gemm-microkernel-tester.h",
5875 "src/xnnpack/AlignedAllocator.h",
5876 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005877 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005878)
5879
5880xnnpack_unit_test(
Frank Barchardb1966592020-05-12 13:47:06 -07005881 name = "f16_hswish_test",
5882 srcs = [
5883 "test/f16-hswish.cc",
5884 "test/hswish-microkernel-tester.h",
5885 ] + MICROKERNEL_TEST_HDRS,
5886 deps = MICROKERNEL_TEST_DEPS,
5887)
5888
5889xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005890 name = "f32_hswish_test",
5891 srcs = [
5892 "test/f32-hswish.cc",
5893 "test/hswish-microkernel-tester.h",
5894 ] + MICROKERNEL_TEST_HDRS,
5895 deps = MICROKERNEL_TEST_DEPS,
5896)
5897
5898xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07005899 name = "f32_maxpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005900 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07005901 "test/f32-maxpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005902 "test/maxpool-microkernel-tester.h",
5903 ] + MICROKERNEL_TEST_HDRS,
5904 deps = MICROKERNEL_TEST_DEPS,
5905)
5906
5907xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07005908 name = "f32_pavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005909 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07005910 "test/f32-pavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005911 "test/avgpool-microkernel-tester.h",
5912 "src/xnnpack/AlignedAllocator.h",
5913 ] + MICROKERNEL_TEST_HDRS,
5914 deps = MICROKERNEL_TEST_DEPS,
5915)
5916
5917xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07005918 name = "f32_ppmm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005919 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07005920 "test/f32-ppmm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005921 "test/gemm-microkernel-tester.h",
5922 "src/xnnpack/AlignedAllocator.h",
5923 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005924 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005925)
5926
5927xnnpack_unit_test(
Frank Barchardb1966592020-05-12 13:47:06 -07005928 name = "f16_prelu_test",
5929 srcs = [
5930 "test/f16-prelu.cc",
5931 "test/prelu-microkernel-tester.h",
5932 "src/xnnpack/AlignedAllocator.h",
5933 ] + MICROKERNEL_TEST_HDRS,
5934 deps = MICROKERNEL_TEST_DEPS,
5935)
5936
5937xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005938 name = "f32_prelu_test",
5939 srcs = [
5940 "test/f32-prelu.cc",
5941 "test/prelu-microkernel-tester.h",
5942 "src/xnnpack/AlignedAllocator.h",
5943 ] + MICROKERNEL_TEST_HDRS,
5944 deps = MICROKERNEL_TEST_DEPS,
5945)
5946
5947xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07005948 name = "f32_raddexpminusmax_test",
5949 srcs = [
5950 "test/f32-raddexpminusmax.cc",
5951 "test/raddexpminusmax-microkernel-tester.h",
5952 ] + MICROKERNEL_TEST_HDRS,
5953 deps = MICROKERNEL_TEST_DEPS,
5954)
5955
5956xnnpack_unit_test(
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07005957 name = "f32_raddextexp_test",
5958 srcs = [
5959 "test/f32-raddextexp.cc",
5960 "test/raddextexp-microkernel-tester.h",
5961 ] + MICROKERNEL_TEST_HDRS,
5962 deps = MICROKERNEL_TEST_DEPS,
5963)
5964
5965xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07005966 name = "f32_raddstoreexpminusmax_test",
5967 srcs = [
5968 "test/f32-raddstoreexpminusmax.cc",
5969 "test/raddstoreexpminusmax-microkernel-tester.h",
5970 ] + MICROKERNEL_TEST_HDRS,
5971 deps = MICROKERNEL_TEST_DEPS,
5972)
5973
5974xnnpack_unit_test(
Frank Barchardfb158e22020-07-15 16:10:10 -07005975 name = "f32_relu_test",
5976 srcs = [
5977 "test/f32-relu.cc",
5978 "test/vunary-microkernel-tester.h",
5979 ] + MICROKERNEL_TEST_HDRS,
5980 deps = MICROKERNEL_TEST_DEPS,
5981)
5982
5983xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005984 name = "f32_rmax_test",
5985 srcs = [
5986 "test/f32-rmax.cc",
5987 "test/rmax-microkernel-tester.h",
5988 ] + MICROKERNEL_TEST_HDRS,
5989 deps = MICROKERNEL_TEST_DEPS,
5990)
5991
5992xnnpack_unit_test(
Marat Dukhan346a9e52019-11-15 09:06:30 -08005993 name = "f32_sigmoid_test",
5994 srcs = [
5995 "test/f32-sigmoid.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005996 "test/vunary-microkernel-tester.h",
Marat Dukhan346a9e52019-11-15 09:06:30 -08005997 ] + MICROKERNEL_TEST_HDRS,
5998 deps = MICROKERNEL_TEST_DEPS,
5999)
6000
6001xnnpack_unit_test(
Marat Dukhan355ab432020-04-09 19:01:52 -07006002 name = "f32_spmm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006003 srcs = [
Marat Dukhan355ab432020-04-09 19:01:52 -07006004 "test/f32-spmm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006005 "test/spmm-microkernel-tester.h",
6006 "src/xnnpack/AlignedAllocator.h",
6007 ] + MICROKERNEL_TEST_HDRS,
6008 deps = MICROKERNEL_TEST_DEPS,
6009)
6010
6011xnnpack_unit_test(
Frank Barchardd793f6c2020-05-08 13:37:43 -07006012 name = "f16_vadd_minmax_test",
6013 srcs = [
6014 "test/f16-vadd-minmax.cc",
6015 "test/vbinary-microkernel-tester.h",
6016 ] + MICROKERNEL_TEST_HDRS,
6017 deps = MICROKERNEL_TEST_DEPS,
6018)
6019
6020xnnpack_unit_test(
6021 name = "f16_vaddc_minmax_test",
6022 srcs = [
6023 "test/f16-vaddc-minmax.cc",
6024 "test/vbinaryc-microkernel-tester.h",
6025 ] + MICROKERNEL_TEST_HDRS,
6026 deps = MICROKERNEL_TEST_DEPS,
6027)
6028
6029xnnpack_unit_test(
6030 name = "f16_vdiv_minmax_test",
6031 srcs = [
6032 "test/f16-vdiv-minmax.cc",
6033 "test/vbinary-microkernel-tester.h",
6034 ] + MICROKERNEL_TEST_HDRS,
6035 deps = MICROKERNEL_TEST_DEPS,
6036)
6037
6038xnnpack_unit_test(
6039 name = "f16_vdivc_minmax_test",
6040 srcs = [
6041 "test/f16-vdivc-minmax.cc",
6042 "test/vbinaryc-microkernel-tester.h",
6043 ] + MICROKERNEL_TEST_HDRS,
6044 deps = MICROKERNEL_TEST_DEPS,
6045)
6046
6047xnnpack_unit_test(
6048 name = "f16_vrdivc_minmax_test",
6049 srcs = [
6050 "test/f16-vrdivc-minmax.cc",
6051 "test/vbinaryc-microkernel-tester.h",
6052 ] + MICROKERNEL_TEST_HDRS,
6053 deps = MICROKERNEL_TEST_DEPS,
6054)
6055
6056xnnpack_unit_test(
6057 name = "f16_vmax_test",
6058 srcs = [
6059 "test/f16-vmax.cc",
6060 "test/vbinary-microkernel-tester.h",
6061 ] + MICROKERNEL_TEST_HDRS,
6062 deps = MICROKERNEL_TEST_DEPS,
6063)
6064
6065xnnpack_unit_test(
6066 name = "f16_vmaxc_test",
6067 srcs = [
6068 "test/f16-vmaxc.cc",
6069 "test/vbinaryc-microkernel-tester.h",
6070 ] + MICROKERNEL_TEST_HDRS,
6071 deps = MICROKERNEL_TEST_DEPS,
6072)
6073
6074xnnpack_unit_test(
6075 name = "f16_vmin_test",
6076 srcs = [
6077 "test/f16-vmin.cc",
6078 "test/vbinary-microkernel-tester.h",
6079 ] + MICROKERNEL_TEST_HDRS,
6080 deps = MICROKERNEL_TEST_DEPS,
6081)
6082
6083xnnpack_unit_test(
6084 name = "f16_vminc_test",
6085 srcs = [
6086 "test/f16-vminc.cc",
6087 "test/vbinaryc-microkernel-tester.h",
6088 ] + MICROKERNEL_TEST_HDRS,
6089 deps = MICROKERNEL_TEST_DEPS,
6090)
6091
6092xnnpack_unit_test(
6093 name = "f16_vmul_minmax_test",
6094 srcs = [
6095 "test/f16-vmul-minmax.cc",
6096 "test/vbinary-microkernel-tester.h",
6097 ] + MICROKERNEL_TEST_HDRS,
6098 deps = MICROKERNEL_TEST_DEPS,
6099)
6100
6101xnnpack_unit_test(
6102 name = "f16_vmulc_minmax_test",
6103 srcs = [
6104 "test/f16-vmulc-minmax.cc",
6105 "test/vbinaryc-microkernel-tester.h",
6106 ] + MICROKERNEL_TEST_HDRS,
6107 deps = MICROKERNEL_TEST_DEPS,
6108)
6109
6110xnnpack_unit_test(
6111 name = "f16_vsub_minmax_test",
6112 srcs = [
6113 "test/f16-vsub-minmax.cc",
6114 "test/vbinary-microkernel-tester.h",
6115 ] + MICROKERNEL_TEST_HDRS,
6116 deps = MICROKERNEL_TEST_DEPS,
6117)
6118
6119xnnpack_unit_test(
6120 name = "f16_vsubc_minmax_test",
6121 srcs = [
6122 "test/f16-vsubc-minmax.cc",
6123 "test/vbinaryc-microkernel-tester.h",
6124 ] + MICROKERNEL_TEST_HDRS,
6125 deps = MICROKERNEL_TEST_DEPS,
6126)
6127
6128xnnpack_unit_test(
6129 name = "f16_vrsubc_minmax_test",
6130 srcs = [
6131 "test/f16-vrsubc-minmax.cc",
6132 "test/vbinaryc-microkernel-tester.h",
6133 ] + MICROKERNEL_TEST_HDRS,
6134 deps = MICROKERNEL_TEST_DEPS,
6135)
6136
6137xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07006138 name = "f32_vabs_test",
6139 srcs = [
6140 "test/f32-vabs.cc",
6141 "test/vunary-microkernel-tester.h",
6142 ] + MICROKERNEL_TEST_HDRS,
6143 deps = MICROKERNEL_TEST_DEPS,
6144)
6145
6146xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07006147 name = "f32_vadd_test",
6148 srcs = [
6149 "test/f32-vadd.cc",
6150 "test/vbinary-microkernel-tester.h",
6151 ] + MICROKERNEL_TEST_HDRS,
6152 deps = MICROKERNEL_TEST_DEPS,
6153)
6154
6155xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006156 name = "f32_vadd_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006157 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006158 "test/f32-vadd-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08006159 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08006160 ] + MICROKERNEL_TEST_HDRS,
6161 deps = MICROKERNEL_TEST_DEPS,
6162)
6163
6164xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07006165 name = "f32_vadd_relu_test",
6166 srcs = [
6167 "test/f32-vadd-relu.cc",
6168 "test/vbinary-microkernel-tester.h",
6169 ] + MICROKERNEL_TEST_HDRS,
6170 deps = MICROKERNEL_TEST_DEPS,
6171)
6172
6173xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07006174 name = "f32_vaddc_test",
6175 srcs = [
6176 "test/f32-vaddc.cc",
6177 "test/vbinaryc-microkernel-tester.h",
6178 ] + MICROKERNEL_TEST_HDRS,
6179 deps = MICROKERNEL_TEST_DEPS,
6180)
6181
6182xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006183 name = "f32_vaddc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08006184 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006185 "test/f32-vaddc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08006186 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006187 ] + MICROKERNEL_TEST_HDRS,
6188 deps = MICROKERNEL_TEST_DEPS,
6189)
6190
6191xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07006192 name = "f32_vaddc_relu_test",
6193 srcs = [
6194 "test/f32-vaddc-relu.cc",
6195 "test/vbinaryc-microkernel-tester.h",
6196 ] + MICROKERNEL_TEST_HDRS,
6197 deps = MICROKERNEL_TEST_DEPS,
6198)
6199
6200xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07006201 name = "f32_vdiv_test",
6202 srcs = [
6203 "test/f32-vdiv.cc",
6204 "test/vbinary-microkernel-tester.h",
6205 ] + MICROKERNEL_TEST_HDRS,
6206 deps = MICROKERNEL_TEST_DEPS,
6207)
6208
6209xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006210 name = "f32_vdiv_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08006211 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006212 "test/f32-vdiv-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08006213 "test/vbinary-microkernel-tester.h",
6214 ] + MICROKERNEL_TEST_HDRS,
6215 deps = MICROKERNEL_TEST_DEPS,
6216)
6217
6218xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07006219 name = "f32_vdiv_relu_test",
6220 srcs = [
6221 "test/f32-vdiv-relu.cc",
6222 "test/vbinary-microkernel-tester.h",
6223 ] + MICROKERNEL_TEST_HDRS,
6224 deps = MICROKERNEL_TEST_DEPS,
6225)
6226
6227xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07006228 name = "f32_vdivc_test",
6229 srcs = [
6230 "test/f32-vdivc.cc",
6231 "test/vbinaryc-microkernel-tester.h",
6232 ] + MICROKERNEL_TEST_HDRS,
6233 deps = MICROKERNEL_TEST_DEPS,
6234)
6235
6236xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006237 name = "f32_vdivc_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08006238 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006239 "test/f32-vdivc-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08006240 "test/vbinaryc-microkernel-tester.h",
6241 ] + MICROKERNEL_TEST_HDRS,
6242 deps = MICROKERNEL_TEST_DEPS,
6243)
6244
6245xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07006246 name = "f32_vdivc_relu_test",
6247 srcs = [
6248 "test/f32-vdivc-relu.cc",
6249 "test/vbinaryc-microkernel-tester.h",
6250 ] + MICROKERNEL_TEST_HDRS,
6251 deps = MICROKERNEL_TEST_DEPS,
6252)
6253
6254xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07006255 name = "f32_vrdivc_test",
6256 srcs = [
6257 "test/f32-vrdivc.cc",
6258 "test/vbinaryc-microkernel-tester.h",
6259 ] + MICROKERNEL_TEST_HDRS,
6260 deps = MICROKERNEL_TEST_DEPS,
6261)
6262
6263xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006264 name = "f32_vrdivc_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08006265 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006266 "test/f32-vrdivc-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08006267 "test/vbinaryc-microkernel-tester.h",
6268 ] + MICROKERNEL_TEST_HDRS,
6269 deps = MICROKERNEL_TEST_DEPS,
6270)
6271
6272xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07006273 name = "f32_vrdivc_relu_test",
6274 srcs = [
6275 "test/f32-vrdivc-relu.cc",
6276 "test/vbinaryc-microkernel-tester.h",
6277 ] + MICROKERNEL_TEST_HDRS,
6278 deps = MICROKERNEL_TEST_DEPS,
6279)
6280
6281xnnpack_unit_test(
Marat Dukhaned6baaf2020-12-01 15:07:08 -08006282 name = "f32_velu_test",
6283 srcs = [
6284 "test/f32-velu.cc",
6285 "test/vunary-microkernel-tester.h",
6286 ] + MICROKERNEL_TEST_HDRS,
6287 deps = MICROKERNEL_TEST_DEPS,
6288)
6289
6290xnnpack_unit_test(
Marat Dukhan403b7d42019-12-05 12:49:11 -08006291 name = "f32_vmax_test",
6292 srcs = [
6293 "test/f32-vmax.cc",
6294 "test/vbinary-microkernel-tester.h",
6295 ] + MICROKERNEL_TEST_HDRS,
6296 deps = MICROKERNEL_TEST_DEPS,
6297)
6298
6299xnnpack_unit_test(
6300 name = "f32_vmaxc_test",
6301 srcs = [
6302 "test/f32-vmaxc.cc",
6303 "test/vbinaryc-microkernel-tester.h",
6304 ] + MICROKERNEL_TEST_HDRS,
6305 deps = MICROKERNEL_TEST_DEPS,
6306)
6307
6308xnnpack_unit_test(
6309 name = "f32_vmin_test",
6310 srcs = [
6311 "test/f32-vmin.cc",
6312 "test/vbinary-microkernel-tester.h",
6313 ] + MICROKERNEL_TEST_HDRS,
6314 deps = MICROKERNEL_TEST_DEPS,
6315)
6316
6317xnnpack_unit_test(
6318 name = "f32_vminc_test",
6319 srcs = [
6320 "test/f32-vminc.cc",
6321 "test/vbinaryc-microkernel-tester.h",
6322 ] + MICROKERNEL_TEST_HDRS,
6323 deps = MICROKERNEL_TEST_DEPS,
6324)
6325
6326xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07006327 name = "f32_vmul_test",
6328 srcs = [
6329 "test/f32-vmul.cc",
6330 "test/vbinary-microkernel-tester.h",
6331 ] + MICROKERNEL_TEST_HDRS,
6332 deps = MICROKERNEL_TEST_DEPS,
6333)
6334
6335xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006336 name = "f32_vmul_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006337 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006338 "test/f32-vmul-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08006339 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08006340 ] + MICROKERNEL_TEST_HDRS,
6341 deps = MICROKERNEL_TEST_DEPS,
6342)
6343
6344xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07006345 name = "f32_vmul_relu_test",
6346 srcs = [
6347 "test/f32-vmul-relu.cc",
6348 "test/vbinary-microkernel-tester.h",
6349 ] + MICROKERNEL_TEST_HDRS,
6350 deps = MICROKERNEL_TEST_DEPS,
6351)
6352
6353xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07006354 name = "f32_vmulc_test",
6355 srcs = [
6356 "test/f32-vmulc.cc",
6357 "test/vbinaryc-microkernel-tester.h",
6358 ] + MICROKERNEL_TEST_HDRS,
6359 deps = MICROKERNEL_TEST_DEPS,
6360)
6361
6362xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006363 name = "f32_vmulc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08006364 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006365 "test/f32-vmulc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08006366 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006367 ] + MICROKERNEL_TEST_HDRS,
6368 deps = MICROKERNEL_TEST_DEPS,
6369)
6370
6371xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07006372 name = "f32_vmulc_relu_test",
6373 srcs = [
6374 "test/f32-vmulc-relu.cc",
6375 "test/vbinaryc-microkernel-tester.h",
6376 ] + MICROKERNEL_TEST_HDRS,
6377 deps = MICROKERNEL_TEST_DEPS,
6378)
6379
6380xnnpack_unit_test(
Frank Barchard2a1049c2020-06-03 02:31:27 -07006381 name = "f16_vmulcaddc_minmax_test",
6382 srcs = [
6383 "test/f16-vmulcaddc-minmax.cc",
6384 "test/vmulcaddc-microkernel-tester.h",
6385 "src/xnnpack/AlignedAllocator.h",
6386 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07006387 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Frank Barchard2a1049c2020-06-03 02:31:27 -07006388)
6389
6390xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07006391 name = "f32_vmulcaddc_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006392 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07006393 "test/f32-vmulcaddc-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006394 "test/vmulcaddc-microkernel-tester.h",
6395 "src/xnnpack/AlignedAllocator.h",
6396 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07006397 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07006398)
6399
6400xnnpack_unit_test(
Marat Dukhan8cc7efe2020-06-10 16:24:27 -07006401 name = "f32_vlrelu_test",
6402 srcs = [
6403 "test/f32-vlrelu.cc",
6404 "test/vunary-microkernel-tester.h",
6405 ] + MICROKERNEL_TEST_HDRS,
6406 deps = MICROKERNEL_TEST_DEPS,
6407)
6408
6409xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07006410 name = "f32_vneg_test",
6411 srcs = [
6412 "test/f32-vneg.cc",
6413 "test/vunary-microkernel-tester.h",
6414 ] + MICROKERNEL_TEST_HDRS,
6415 deps = MICROKERNEL_TEST_DEPS,
6416)
6417
6418xnnpack_unit_test(
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07006419 name = "f32_vrndne_test",
6420 srcs = [
6421 "test/f32-vrndne.cc",
6422 "test/vunary-microkernel-tester.h",
6423 ] + MICROKERNEL_TEST_HDRS,
6424 deps = MICROKERNEL_TEST_DEPS,
6425)
6426
6427xnnpack_unit_test(
6428 name = "f32_vrndz_test",
6429 srcs = [
6430 "test/f32-vrndz.cc",
6431 "test/vunary-microkernel-tester.h",
6432 ] + MICROKERNEL_TEST_HDRS,
6433 deps = MICROKERNEL_TEST_DEPS,
6434)
6435
6436xnnpack_unit_test(
6437 name = "f32_vrndu_test",
6438 srcs = [
6439 "test/f32-vrndu.cc",
6440 "test/vunary-microkernel-tester.h",
6441 ] + MICROKERNEL_TEST_HDRS,
6442 deps = MICROKERNEL_TEST_DEPS,
6443)
6444
6445xnnpack_unit_test(
6446 name = "f32_vrndd_test",
6447 srcs = [
6448 "test/f32-vrndd.cc",
6449 "test/vunary-microkernel-tester.h",
6450 ] + MICROKERNEL_TEST_HDRS,
6451 deps = MICROKERNEL_TEST_DEPS,
6452)
6453
6454xnnpack_unit_test(
Marat Dukhan05ac8e32019-10-21 15:39:33 -07006455 name = "f32_vscale_test",
6456 srcs = [
6457 "test/f32-vscale.cc",
6458 "test/vscale-microkernel-tester.h",
6459 ] + MICROKERNEL_TEST_HDRS,
6460 deps = MICROKERNEL_TEST_DEPS,
6461)
6462
6463xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07006464 name = "f32_vscaleexpminusmax_test",
6465 srcs = [
6466 "test/f32-vscaleexpminusmax.cc",
6467 "test/vscaleexpminusmax-microkernel-tester.h",
6468 ] + MICROKERNEL_TEST_HDRS,
6469 deps = MICROKERNEL_TEST_DEPS,
6470)
6471
6472xnnpack_unit_test(
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07006473 name = "f32_vscaleextexp_test",
6474 srcs = [
6475 "test/f32-vscaleextexp.cc",
6476 "test/vscaleextexp-microkernel-tester.h",
6477 ] + MICROKERNEL_TEST_HDRS,
6478 deps = MICROKERNEL_TEST_DEPS,
6479)
6480
6481xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07006482 name = "f32_vsqr_test",
6483 srcs = [
6484 "test/f32-vsqr.cc",
6485 "test/vunary-microkernel-tester.h",
6486 ] + MICROKERNEL_TEST_HDRS,
6487 deps = MICROKERNEL_TEST_DEPS,
6488)
6489
6490xnnpack_unit_test(
Marat Dukhan13bafb02020-06-05 00:43:11 -07006491 name = "f32_vsqrdiff_test",
6492 srcs = [
6493 "test/f32-vsqrdiff.cc",
6494 "test/vbinary-microkernel-tester.h",
6495 ] + MICROKERNEL_TEST_HDRS,
6496 deps = MICROKERNEL_TEST_DEPS,
6497)
6498
6499xnnpack_unit_test(
6500 name = "f32_vsqrdiffc_test",
6501 srcs = [
6502 "test/f32-vsqrdiffc.cc",
6503 "test/vbinaryc-microkernel-tester.h",
6504 ] + MICROKERNEL_TEST_HDRS,
6505 deps = MICROKERNEL_TEST_DEPS,
6506)
6507
6508xnnpack_unit_test(
Marat Dukhanf4db2f32020-06-30 10:55:30 -07006509 name = "f32_vsqrt_test",
6510 srcs = [
6511 "test/f32-vsqrt.cc",
6512 "test/vunary-microkernel-tester.h",
6513 ] + MICROKERNEL_TEST_HDRS,
6514 deps = MICROKERNEL_TEST_DEPS,
6515)
6516
6517xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07006518 name = "f32_vsub_test",
6519 srcs = [
6520 "test/f32-vsub.cc",
6521 "test/vbinary-microkernel-tester.h",
6522 ] + MICROKERNEL_TEST_HDRS,
6523 deps = MICROKERNEL_TEST_DEPS,
6524)
6525
6526xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006527 name = "f32_vsub_minmax_test",
Marat Dukhan97579532019-10-18 16:40:39 -07006528 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006529 "test/f32-vsub-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08006530 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08006531 ] + MICROKERNEL_TEST_HDRS,
6532 deps = MICROKERNEL_TEST_DEPS,
6533)
6534
6535xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07006536 name = "f32_vsub_relu_test",
6537 srcs = [
6538 "test/f32-vsub-relu.cc",
6539 "test/vbinary-microkernel-tester.h",
6540 ] + MICROKERNEL_TEST_HDRS,
6541 deps = MICROKERNEL_TEST_DEPS,
6542)
6543
6544xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07006545 name = "f32_vsubc_test",
6546 srcs = [
6547 "test/f32-vsubc.cc",
6548 "test/vbinaryc-microkernel-tester.h",
6549 ] + MICROKERNEL_TEST_HDRS,
6550 deps = MICROKERNEL_TEST_DEPS,
6551)
6552
6553xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006554 name = "f32_vsubc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08006555 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006556 "test/f32-vsubc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08006557 "test/vbinaryc-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08006558 ] + MICROKERNEL_TEST_HDRS,
6559 deps = MICROKERNEL_TEST_DEPS,
6560)
6561
6562xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07006563 name = "f32_vsubc_relu_test",
6564 srcs = [
6565 "test/f32-vsubc-relu.cc",
6566 "test/vbinaryc-microkernel-tester.h",
6567 ] + MICROKERNEL_TEST_HDRS,
6568 deps = MICROKERNEL_TEST_DEPS,
6569)
6570
6571xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07006572 name = "f32_vrsubc_test",
6573 srcs = [
6574 "test/f32-vrsubc.cc",
6575 "test/vbinaryc-microkernel-tester.h",
6576 ] + MICROKERNEL_TEST_HDRS,
6577 deps = MICROKERNEL_TEST_DEPS,
6578)
6579
6580xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006581 name = "f32_vrsubc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08006582 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006583 "test/f32-vrsubc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08006584 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan97579532019-10-18 16:40:39 -07006585 ] + MICROKERNEL_TEST_HDRS,
6586 deps = MICROKERNEL_TEST_DEPS,
6587)
6588
6589xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07006590 name = "f32_vrsubc_relu_test",
6591 srcs = [
6592 "test/f32-vrsubc-relu.cc",
6593 "test/vbinaryc-microkernel-tester.h",
6594 ] + MICROKERNEL_TEST_HDRS,
6595 deps = MICROKERNEL_TEST_DEPS,
6596)
6597
6598xnnpack_unit_test(
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07006599 name = "qs8_dwconv_minmax_test",
6600 srcs = [
6601 "test/qs8-dwconv-minmax.cc",
6602 "test/dwconv-microkernel-tester.h",
6603 "src/xnnpack/AlignedAllocator.h",
6604 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
6605 deps = MICROKERNEL_TEST_DEPS + [":packing"],
6606)
6607
6608xnnpack_unit_test(
Marat Dukhan4ed53f42020-08-06 01:12:55 -07006609 name = "qs8_gavgpool_minmax_test",
6610 srcs = [
6611 "test/qs8-gavgpool-minmax.cc",
6612 "test/gavgpool-microkernel-tester.h",
6613 "src/xnnpack/AlignedAllocator.h",
6614 ] + MICROKERNEL_TEST_HDRS,
6615 deps = MICROKERNEL_TEST_DEPS,
6616)
6617
6618xnnpack_unit_test(
Marat Dukhan595e1702020-07-31 10:12:52 -07006619 name = "qs8_gemm_minmax_test",
6620 srcs = [
6621 "test/qs8-gemm-minmax.cc",
6622 "test/gemm-microkernel-tester.h",
6623 "src/xnnpack/AlignedAllocator.h",
6624 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
6625 deps = MICROKERNEL_TEST_DEPS + [":packing"],
6626)
6627
6628xnnpack_unit_test(
Marat Dukhanf9480682020-07-31 14:50:24 -07006629 name = "qs8_igemm_minmax_test",
6630 srcs = [
6631 "test/qs8-igemm-minmax.cc",
6632 "test/gemm-microkernel-tester.h",
6633 "src/xnnpack/AlignedAllocator.h",
6634 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
6635 deps = MICROKERNEL_TEST_DEPS + [":packing"],
6636)
6637
6638xnnpack_unit_test(
6639 name = "qs8_requantization_test",
6640 srcs = [
6641 "src/xnnpack/requantization-stubs.h",
6642 "test/qs8-requantization.cc",
6643 "test/requantization-tester.h",
6644 ] + MICROKERNEL_TEST_HDRS,
6645 deps = MICROKERNEL_TEST_DEPS,
6646)
6647
6648xnnpack_unit_test(
Marat Dukhand9f3ad42020-08-10 12:30:58 -07006649 name = "qs8_vadd_minmax_test",
6650 srcs = [
6651 "test/qs8-vadd-minmax.cc",
6652 "test/vadd-microkernel-tester.h",
6653 ] + MICROKERNEL_TEST_HDRS,
6654 deps = MICROKERNEL_TEST_DEPS,
6655)
6656
6657xnnpack_unit_test(
Marat Dukhan0270d9f2020-08-11 00:56:46 -07006658 name = "qs8_vaddc_minmax_test",
6659 srcs = [
6660 "test/qs8-vaddc-minmax.cc",
6661 "test/vaddc-microkernel-tester.h",
6662 ] + MICROKERNEL_TEST_HDRS,
6663 deps = MICROKERNEL_TEST_DEPS,
6664)
6665
6666xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07006667 name = "qu8_avgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006668 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07006669 "test/qu8-avgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006670 "test/avgpool-microkernel-tester.h",
6671 "src/xnnpack/AlignedAllocator.h",
6672 ] + MICROKERNEL_TEST_HDRS,
6673 deps = MICROKERNEL_TEST_DEPS,
6674)
6675
6676xnnpack_unit_test(
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07006677 name = "qu8_dwconv_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006678 srcs = [
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07006679 "test/qu8-dwconv-minmax.cc",
6680 "test/dwconv-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006681 "src/xnnpack/AlignedAllocator.h",
6682 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07006683 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07006684)
6685
6686xnnpack_unit_test(
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07006687 name = "qu8_igemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006688 srcs = [
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07006689 "test/qu8-igemm-minmax.cc",
6690 "test/gemm-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006691 "src/xnnpack/AlignedAllocator.h",
6692 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07006693 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07006694)
6695
6696xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07006697 name = "qu8_gavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006698 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07006699 "test/qu8-gavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006700 "test/gavgpool-microkernel-tester.h",
6701 "src/xnnpack/AlignedAllocator.h",
6702 ] + MICROKERNEL_TEST_HDRS,
6703 deps = MICROKERNEL_TEST_DEPS,
6704)
6705
6706xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07006707 name = "qu8_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006708 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07006709 "test/qu8-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006710 "test/gemm-microkernel-tester.h",
6711 "src/xnnpack/AlignedAllocator.h",
6712 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07006713 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07006714)
6715
6716xnnpack_unit_test(
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07006717 name = "qu8_requantization_test",
6718 srcs = [
6719 "src/xnnpack/requantization-stubs.h",
6720 "test/qu8-requantization.cc",
6721 "test/requantization-tester.h",
6722 ] + MICROKERNEL_TEST_HDRS,
6723 deps = MICROKERNEL_TEST_DEPS,
6724)
6725
6726xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07006727 name = "qu8_vadd_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006728 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07006729 "test/qu8-vadd-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006730 "test/vadd-microkernel-tester.h",
6731 ] + MICROKERNEL_TEST_HDRS,
6732 deps = MICROKERNEL_TEST_DEPS,
6733)
6734
6735xnnpack_unit_test(
6736 name = "u8_clamp_test",
6737 srcs = [
6738 "test/u8-clamp.cc",
6739 "test/clamp-microkernel-tester.h",
6740 ] + MICROKERNEL_TEST_HDRS,
6741 deps = MICROKERNEL_TEST_DEPS,
6742)
6743
6744xnnpack_unit_test(
6745 name = "u8_lut32norm_test",
6746 srcs = [
6747 "test/u8-lut32norm.cc",
6748 "test/lut-norm-microkernel-tester.h",
6749 ] + MICROKERNEL_TEST_HDRS,
6750 deps = MICROKERNEL_TEST_DEPS,
6751)
6752
6753xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07006754 name = "u8_maxpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006755 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07006756 "test/u8-maxpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006757 "test/maxpool-microkernel-tester.h",
6758 ] + MICROKERNEL_TEST_HDRS,
6759 deps = MICROKERNEL_TEST_DEPS,
6760)
6761
6762xnnpack_unit_test(
6763 name = "u8_rmax_test",
6764 srcs = [
6765 "test/u8-rmax.cc",
6766 "test/rmax-microkernel-tester.h",
6767 ] + MICROKERNEL_TEST_HDRS,
6768 deps = MICROKERNEL_TEST_DEPS,
6769)
6770
6771xnnpack_unit_test(
Marat Dukhanad71b9a2020-11-20 00:01:51 -08006772 name = "x32_depthtospace2d_chw2hwc_test",
Yury Kartynnike7841862020-11-04 18:22:18 -08006773 srcs = [
Marat Dukhanad71b9a2020-11-20 00:01:51 -08006774 "test/x32-depthtospace2d-chw2hwc.cc",
6775 "test/depthtospace-microkernel-tester.h",
Yury Kartynnike7841862020-11-04 18:22:18 -08006776 ] + MICROKERNEL_TEST_HDRS,
6777 deps = MICROKERNEL_TEST_DEPS,
6778)
6779
6780xnnpack_unit_test(
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -07006781 name = "x32_fill_test",
6782 srcs = [
6783 "test/x32-fill.cc",
6784 "test/fill-microkernel-tester.h",
6785 ] + MICROKERNEL_TEST_HDRS,
6786 deps = MICROKERNEL_TEST_DEPS,
6787)
6788
6789xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07006790 name = "x32_packx_test",
6791 srcs = [
6792 "test/x32-packx.cc",
6793 "test/pack-microkernel-tester.h",
6794 "src/xnnpack/AlignedAllocator.h",
6795 ] + MICROKERNEL_TEST_HDRS,
6796 deps = MICROKERNEL_TEST_DEPS,
6797)
6798
6799xnnpack_unit_test(
6800 name = "x32_pad_test",
6801 srcs = [
6802 "test/x32-pad.cc",
6803 "test/pad-microkernel-tester.h",
6804 ] + MICROKERNEL_TEST_HDRS,
6805 deps = MICROKERNEL_TEST_DEPS,
6806)
6807
6808xnnpack_unit_test(
6809 name = "x32_unpool_test",
6810 srcs = [
6811 "test/x32-unpool.cc",
6812 "test/unpool-microkernel-tester.h",
6813 ] + MICROKERNEL_TEST_HDRS,
6814 deps = MICROKERNEL_TEST_DEPS,
6815)
6816
6817xnnpack_unit_test(
6818 name = "x32_zip_test",
6819 srcs = [
6820 "test/x32-zip.cc",
6821 "test/zip-microkernel-tester.h",
6822 ] + MICROKERNEL_TEST_HDRS,
6823 deps = MICROKERNEL_TEST_DEPS,
6824)
6825
6826xnnpack_unit_test(
6827 name = "x8_lut_test",
6828 srcs = [
6829 "test/x8-lut.cc",
6830 "test/lut-microkernel-tester.h",
6831 ] + MICROKERNEL_TEST_HDRS,
6832 deps = MICROKERNEL_TEST_DEPS,
6833)
6834
6835xnnpack_unit_test(
6836 name = "x8_zip_test",
6837 srcs = [
6838 "test/x8-zip.cc",
6839 "test/zip-microkernel-tester.h",
6840 ] + MICROKERNEL_TEST_HDRS,
6841 deps = MICROKERNEL_TEST_DEPS,
6842)
6843
Marat Dukhan20c3b922020-03-10 03:45:06 -07006844########################## Size tests for the library #########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07006845
6846xnnpack_binary(
Marat Dukhan20c3b922020-03-10 03:45:06 -07006847 name = "operator_size_test",
6848 srcs = ["test/operator-size.c"],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07006849 deps = [":xnnpack_operators_nhwc_f32"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07006850)
6851
Marat Dukhan20c3b922020-03-10 03:45:06 -07006852xnnpack_binary(
6853 name = "subgraph_size_test",
6854 srcs = ["test/subgraph-size.c"],
6855 deps = [":XNNPACK"],
6856)
6857
6858########################### Unit tests for operators ##########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07006859
6860xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07006861 name = "abs_nc_test",
6862 srcs = [
6863 "test/abs-nc.cc",
6864 "test/abs-operator-tester.h",
6865 ],
6866 deps = OPERATOR_TEST_DEPS,
6867)
6868
6869xnnpack_unit_test(
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08006870 name = "add_nd_test",
Artsiom Ablavatski2202c812021-01-22 14:16:43 -08006871 timeout = "moderate",
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08006872 srcs = [
6873 "test/add-nd.cc",
6874 "test/binary-elementwise-operator-tester.h",
6875 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006876 deps = OPERATOR_TEST_DEPS,
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08006877)
6878
6879xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006880 name = "argmax_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006881 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006882 "test/argmax-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006883 "test/argmax-pooling-operator-tester.h",
6884 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07006885 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006886)
6887
6888xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006889 name = "average_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006890 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006891 "test/average-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006892 "test/average-pooling-operator-tester.h",
6893 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07006894 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006895)
6896
6897xnnpack_unit_test(
Marat Dukhan64e52512020-06-09 13:41:16 -07006898 name = "bankers_rounding_nc_test",
6899 srcs = [
6900 "test/bankers-rounding-nc.cc",
6901 "test/bankers-rounding-operator-tester.h",
6902 ],
6903 deps = OPERATOR_TEST_DEPS,
6904)
6905
6906xnnpack_unit_test(
6907 name = "ceiling_nc_test",
6908 srcs = [
6909 "test/ceiling-nc.cc",
6910 "test/ceiling-operator-tester.h",
6911 ],
6912 deps = OPERATOR_TEST_DEPS,
6913)
6914
6915xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006916 name = "channel_shuffle_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006917 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006918 "test/channel-shuffle-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006919 "test/channel-shuffle-operator-tester.h",
6920 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006921 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006922)
6923
6924xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006925 name = "clamp_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006926 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006927 "test/clamp-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006928 "test/clamp-operator-tester.h",
6929 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006930 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006931)
6932
6933xnnpack_unit_test(
Marat Dukhan065b11e2020-05-22 09:49:41 -07006934 name = "constant_pad_nd_test",
6935 srcs = [
6936 "test/constant-pad-nd.cc",
6937 "test/constant-pad-operator-tester.h",
6938 ],
6939 deps = OPERATOR_TEST_DEPS,
6940)
6941
6942xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006943 name = "convolution_nhwc_test",
Artsiom Ablavatski2202c812021-01-22 14:16:43 -08006944 timeout = "moderate",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006945 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006946 "test/convolution-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006947 "test/convolution-operator-tester.h",
6948 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006949 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006950)
6951
6952xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006953 name = "convolution_nchw_test",
Artsiom Ablavatski2202c812021-01-22 14:16:43 -08006954 timeout = "moderate",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006955 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006956 "test/convolution-nchw.cc",
6957 "test/convolution-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006958 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006959 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006960)
6961
6962xnnpack_unit_test(
Marat Dukhan4e21b272020-06-04 18:45:01 -07006963 name = "copy_nc_test",
6964 srcs = [
6965 "test/copy-nc.cc",
6966 "test/copy-operator-tester.h",
6967 ],
6968 deps = OPERATOR_TEST_DEPS,
6969)
6970
6971xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006972 name = "deconvolution_nhwc_test",
Artsiom Ablavatskic1aa2972020-12-08 11:23:34 -08006973 timeout = "moderate",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006974 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006975 "test/deconvolution-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006976 "test/deconvolution-operator-tester.h",
6977 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07006978 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006979)
6980
6981xnnpack_unit_test(
Marat Dukhan188d1042020-11-24 23:39:40 -08006982 name = "depth_to_space_nchw2nhwc_test",
Artsiom Ablavatski0f1dc182020-11-05 19:21:50 -08006983 srcs = [
Marat Dukhan188d1042020-11-24 23:39:40 -08006984 "test/depth-to-space-nchw2nhwc.cc",
Artsiom Ablavatski0f1dc182020-11-05 19:21:50 -08006985 "test/depth-to-space-operator-tester.h",
6986 ] + OPERATOR_TEST_PARAMS_HDRS,
6987 deps = OPERATOR_TEST_DEPS,
6988)
6989
6990xnnpack_unit_test(
Marat Dukhan0e521172020-11-25 13:10:04 -08006991 name = "depth_to_space_nhwc_test",
6992 srcs = [
6993 "test/depth-to-space-nhwc.cc",
6994 "test/depth-to-space-operator-tester.h",
6995 ] + OPERATOR_TEST_PARAMS_HDRS,
6996 deps = OPERATOR_TEST_DEPS,
6997)
6998
6999xnnpack_unit_test(
Marat Dukhan69180502019-12-06 15:00:31 -08007000 name = "divide_nd_test",
7001 srcs = [
7002 "test/binary-elementwise-operator-tester.h",
7003 "test/divide-nd.cc",
7004 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07007005 deps = OPERATOR_TEST_DEPS,
Marat Dukhan69180502019-12-06 15:00:31 -08007006)
7007
7008xnnpack_unit_test(
Marat Dukhanb6bd4bc2020-12-01 17:01:40 -08007009 name = "elu_nc_test",
7010 srcs = [
7011 "test/elu-nc.cc",
7012 "test/elu-operator-tester.h",
7013 ],
7014 deps = OPERATOR_TEST_DEPS,
7015)
7016
7017xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08007018 name = "fully_connected_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007019 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08007020 "test/fully-connected-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007021 "test/fully-connected-operator-tester.h",
7022 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07007023 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07007024)
7025
7026xnnpack_unit_test(
Marat Dukhan64e52512020-06-09 13:41:16 -07007027 name = "floor_nc_test",
7028 srcs = [
7029 "test/floor-nc.cc",
7030 "test/floor-operator-tester.h",
7031 ],
7032 deps = OPERATOR_TEST_DEPS,
7033)
7034
7035xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08007036 name = "global_average_pooling_nwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007037 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08007038 "test/global-average-pooling-nwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007039 "test/global-average-pooling-operator-tester.h",
Marat Dukhanef61d022020-06-19 13:54:49 -07007040 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07007041 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07007042)
7043
7044xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08007045 name = "global_average_pooling_ncw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007046 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08007047 "test/global-average-pooling-ncw.cc",
7048 "test/global-average-pooling-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007049 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07007050 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07007051)
7052
7053xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08007054 name = "hardswish_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007055 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08007056 "test/hardswish-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007057 "test/hardswish-operator-tester.h",
7058 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07007059 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07007060)
7061
7062xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08007063 name = "leaky_relu_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007064 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08007065 "test/leaky-relu-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007066 "test/leaky-relu-operator-tester.h",
7067 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07007068 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07007069)
7070
7071xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08007072 name = "max_pooling_nhwc_test",
Artsiom Ablavatski2202c812021-01-22 14:16:43 -08007073 timeout = "moderate",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007074 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08007075 "test/max-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007076 "test/max-pooling-operator-tester.h",
7077 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07007078 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07007079)
7080
7081xnnpack_unit_test(
Marat Dukhan79e7f842019-12-05 14:35:50 -08007082 name = "maximum_nd_test",
7083 srcs = [
7084 "test/binary-elementwise-operator-tester.h",
7085 "test/maximum-nd.cc",
7086 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07007087 deps = OPERATOR_TEST_DEPS,
Marat Dukhan79e7f842019-12-05 14:35:50 -08007088)
7089
7090xnnpack_unit_test(
7091 name = "minimum_nd_test",
7092 srcs = [
7093 "test/binary-elementwise-operator-tester.h",
7094 "test/minimum-nd.cc",
7095 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07007096 deps = OPERATOR_TEST_DEPS,
Marat Dukhan79e7f842019-12-05 14:35:50 -08007097)
7098
7099xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08007100 name = "multiply_nd_test",
Marat Dukhanca2733c2019-11-15 23:21:17 -08007101 srcs = [
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08007102 "test/binary-elementwise-operator-tester.h",
Marat Dukhanefc47b82019-11-18 09:25:38 -08007103 "test/multiply-nd.cc",
Marat Dukhanca2733c2019-11-15 23:21:17 -08007104 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07007105 deps = OPERATOR_TEST_DEPS,
Marat Dukhanca2733c2019-11-15 23:21:17 -08007106)
7107
7108xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07007109 name = "negate_nc_test",
7110 srcs = [
7111 "test/negate-nc.cc",
7112 "test/negate-operator-tester.h",
7113 ],
7114 deps = OPERATOR_TEST_DEPS,
7115)
7116
7117xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08007118 name = "prelu_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007119 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08007120 "test/prelu-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007121 "test/prelu-operator-tester.h",
7122 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07007123 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07007124)
7125
7126xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08007127 name = "resize_bilinear_nhwc_test",
Marat Dukhan69722492019-11-11 19:55:50 -08007128 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08007129 "test/resize-bilinear-nhwc.cc",
Marat Dukhan69722492019-11-11 19:55:50 -08007130 "test/resize-bilinear-operator-tester.h",
7131 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07007132 deps = OPERATOR_TEST_DEPS,
Marat Dukhan69722492019-11-11 19:55:50 -08007133)
7134
7135xnnpack_unit_test(
Artsiom Ablavatski97918102020-10-27 15:52:59 -07007136 name = "resize_bilinear_nchw_test",
7137 srcs = [
7138 "test/resize-bilinear-nchw.cc",
7139 "test/resize-bilinear-operator-tester.h",
7140 ] + OPERATOR_TEST_PARAMS_HDRS,
7141 deps = OPERATOR_TEST_DEPS,
7142)
7143
7144xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08007145 name = "sigmoid_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007146 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08007147 "test/sigmoid-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007148 "test/sigmoid-operator-tester.h",
7149 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07007150 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07007151)
7152
7153xnnpack_unit_test(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08007154 name = "softmax_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007155 srcs = [
Marat Dukhanfd8e6892020-01-27 15:25:25 -08007156 "test/softmax-nc.cc",
7157 "test/softmax-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007158 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07007159 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07007160)
7161
7162xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07007163 name = "square_nc_test",
7164 srcs = [
7165 "test/square-nc.cc",
7166 "test/square-operator-tester.h",
7167 ],
7168 deps = OPERATOR_TEST_DEPS,
7169)
7170
7171xnnpack_unit_test(
Marat Dukhan6804bbd2020-06-30 19:26:11 -07007172 name = "square_root_nc_test",
7173 srcs = [
7174 "test/square-root-nc.cc",
7175 "test/square-root-operator-tester.h",
7176 ],
7177 deps = OPERATOR_TEST_DEPS,
7178)
7179
7180xnnpack_unit_test(
Marat Dukhanf7399262020-06-05 10:58:44 -07007181 name = "squared_difference_nd_test",
7182 srcs = [
7183 "test/binary-elementwise-operator-tester.h",
7184 "test/squared-difference-nd.cc",
7185 ],
7186 deps = OPERATOR_TEST_DEPS,
7187)
7188
7189xnnpack_unit_test(
Marat Dukhan05f3f6d2019-12-03 15:13:53 -08007190 name = "subtract_nd_test",
7191 srcs = [
7192 "test/binary-elementwise-operator-tester.h",
7193 "test/subtract-nd.cc",
7194 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07007195 deps = OPERATOR_TEST_DEPS,
Marat Dukhan05f3f6d2019-12-03 15:13:53 -08007196)
7197
7198xnnpack_unit_test(
Marat Dukhan64e52512020-06-09 13:41:16 -07007199 name = "truncation_nc_test",
7200 srcs = [
7201 "test/truncation-nc.cc",
7202 "test/truncation-operator-tester.h",
7203 ],
7204 deps = OPERATOR_TEST_DEPS,
7205)
7206
7207xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08007208 name = "unpooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007209 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08007210 "test/unpooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007211 "test/unpooling-operator-tester.h",
7212 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07007213 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07007214)
7215
Chao Mei6ddfc602020-05-13 22:29:36 -07007216############################### Misc unit tests ###############################
7217
7218xnnpack_unit_test(
7219 name = "memory_planner_test",
7220 srcs = [
7221 "test/memory-planner-test.cc",
7222 ],
7223 deps = [
7224 ":XNNPACK",
7225 ":memory_planner",
7226 ],
7227)
7228
XNNPACK Teamab8c4c82020-10-09 08:05:51 -07007229xnnpack_unit_test(
7230 name = "subgraph_nchw_test",
7231 srcs = [
7232 "src/xnnpack/subgraph.h",
7233 "test/subgraph-nchw.cc",
7234 "test/subgraph-tester.h",
7235 ],
7236 deps = [
7237 ":XNNPACK",
7238 ],
7239)
7240
Marat Dukhan08c4a432019-10-03 09:29:21 -07007241############################# Build configurations #############################
7242
Marat Dukhanb8642352019-10-30 15:43:02 -07007243# Enables usage of assembly kernels.
Marat Dukhan08c4a432019-10-03 09:29:21 -07007244config_setting(
Marat Dukhanb8642352019-10-30 15:43:02 -07007245 name = "xnn_enable_assembly_explicit_true",
7246 define_values = {"xnn_enable_assembly": "true"},
7247)
7248
7249# Disables usage of assembly kernels.
7250config_setting(
7251 name = "xnn_enable_assembly_explicit_false",
7252 define_values = {"xnn_enable_assembly": "false"},
7253)
7254
Marat Dukhan9de90e02020-06-18 16:04:12 -07007255# Enables usage of sparse inference.
7256config_setting(
7257 name = "xnn_enable_sparse_explicit_true",
7258 define_values = {"xnn_enable_sparse": "true"},
7259)
7260
7261# Disables usage of sparse inference.
7262config_setting(
7263 name = "xnn_enable_sparse_explicit_false",
7264 define_values = {"xnn_enable_sparse": "false"},
7265)
7266
Marat Dukhan05702cf2020-03-26 15:41:33 -07007267# Disables usage of HMP-aware optimizations.
7268config_setting(
7269 name = "xnn_enable_hmp_explicit_false",
7270 define_values = {"xnn_enable_hmp": "false"},
7271)
7272
Chao Mei6ddfc602020-05-13 22:29:36 -07007273# Enable usage of optimized memory allocation
7274config_setting(
7275 name = "xnn_enable_memopt_explicit_true",
7276 define_values = {"xnnpack_enable_memopt": "true"},
7277)
7278
7279# Disable usage of optimized memory allocation
7280config_setting(
7281 name = "xnn_enable_memopt_explicit_false",
7282 define_values = {"xnnpack_enable_memopt": "false"},
7283)
7284
Marat Dukhanb8642352019-10-30 15:43:02 -07007285# Builds with -c dbg
7286config_setting(
7287 name = "debug_build",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007288 values = {
Marat Dukhanb8642352019-10-30 15:43:02 -07007289 "compilation_mode": "dbg",
7290 },
7291)
7292
7293# Builds with -c opt
7294config_setting(
7295 name = "optimized_build",
7296 values = {
7297 "compilation_mode": "opt",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007298 },
7299)
7300
7301config_setting(
Marat Dukhanb8642352019-10-30 15:43:02 -07007302 name = "linux_k8",
7303 values = {"cpu": "k8"},
7304)
7305
7306config_setting(
Marat Dukhan582094e2020-04-30 17:21:25 -07007307 name = "linux_arm",
7308 values = {"cpu": "arm"},
Marat Dukhan4e45e662019-10-03 15:40:24 -07007309)
7310
7311config_setting(
Marat Dukhanf0bd4de2020-06-15 15:53:19 -07007312 name = "linux_armeabi",
7313 values = {"cpu": "armeabi"},
7314)
7315
7316config_setting(
Terry Heo68eef3f2020-04-13 22:53:52 -07007317 name = "linux_armhf",
7318 values = {"cpu": "armhf"},
7319)
7320
7321config_setting(
Marat Dukhana720e932020-06-10 13:01:11 -07007322 name = "linux_armv7a",
7323 values = {"cpu": "armv7a"},
7324)
7325
7326config_setting(
Marat Dukhan582094e2020-04-30 17:21:25 -07007327 name = "linux_aarch64",
7328 values = {"cpu": "aarch64"},
7329)
7330
7331config_setting(
Marat Dukhan08c4a432019-10-03 09:29:21 -07007332 name = "android",
7333 values = {"crosstool_top": "//external:android/crosstool"},
7334)
7335
7336config_setting(
7337 name = "android_armv7",
7338 values = {
7339 "crosstool_top": "//external:android/crosstool",
7340 "cpu": "armeabi-v7a",
7341 },
7342)
7343
7344config_setting(
7345 name = "android_arm64",
7346 values = {
7347 "crosstool_top": "//external:android/crosstool",
7348 "cpu": "arm64-v8a",
7349 },
7350)
7351
7352config_setting(
7353 name = "android_x86",
7354 values = {
7355 "crosstool_top": "//external:android/crosstool",
7356 "cpu": "x86",
7357 },
7358)
7359
7360config_setting(
7361 name = "android_x86_64",
7362 values = {
7363 "crosstool_top": "//external:android/crosstool",
7364 "cpu": "x86_64",
7365 },
7366)
7367
7368config_setting(
Marat Dukhan10a38082020-04-17 03:58:35 -07007369 name = "windows_x86_64",
7370 values = {"cpu": "x64_windows"},
Marat Dukhan9fe932e2020-04-11 17:14:15 -07007371)
7372
7373config_setting(
Marat Dukhan10a38082020-04-17 03:58:35 -07007374 name = "windows_x86_64_clang",
7375 values = {
7376 "compiler": "clang-cl",
7377 "cpu": "x64_windows",
7378 },
7379)
7380
7381config_setting(
7382 name = "windows_x86_64_mingw",
7383 values = {
7384 "compiler": "mingw-gcc",
7385 "cpu": "x64_windows",
7386 },
7387)
7388
7389config_setting(
7390 name = "windows_x86_64_msys",
7391 values = {
7392 "compiler": "msys-gcc",
7393 "cpu": "x64_windows",
7394 },
Marat Dukhan9fe932e2020-04-11 17:14:15 -07007395)
7396
7397config_setting(
Marat Dukhan885ca242019-10-07 09:17:32 -07007398 name = "macos_x86_64",
7399 values = {
7400 "apple_platform_type": "macos",
7401 "cpu": "darwin",
7402 },
7403)
7404
7405config_setting(
Marat Dukhan08c4a432019-10-03 09:29:21 -07007406 name = "emscripten",
Marat Dukhan1a729ec2019-10-07 09:31:44 -07007407 values = {"crosstool_top": "//toolchain:emscripten"},
Marat Dukhan08c4a432019-10-03 09:29:21 -07007408)
7409
7410config_setting(
7411 name = "emscripten_wasm",
7412 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07007413 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007414 "cpu": "wasm",
7415 },
7416)
7417
7418config_setting(
7419 name = "emscripten_wasmsimd",
7420 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07007421 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007422 "cpu": "wasm",
Marat Dukhan81c62602020-05-29 13:22:49 -07007423 "copt": "-msimd128",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007424 },
7425)
7426
7427config_setting(
Marat Dukhan1498d1d2020-02-11 20:00:05 -08007428 name = "ios_armv7",
7429 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08007430 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08007431 "cpu": "ios_armv7",
7432 },
7433)
7434
7435config_setting(
7436 name = "ios_arm64",
7437 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08007438 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08007439 "cpu": "ios_arm64",
7440 },
7441)
7442
7443config_setting(
7444 name = "ios_arm64e",
7445 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08007446 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08007447 "cpu": "ios_arm64e",
7448 },
7449)
7450
7451config_setting(
7452 name = "ios_x86",
7453 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08007454 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08007455 "cpu": "ios_i386",
7456 },
7457)
7458
7459config_setting(
7460 name = "ios_x86_64",
7461 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08007462 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08007463 "cpu": "ios_x86_64",
7464 },
7465)
7466
7467config_setting(
7468 name = "watchos_armv7k",
7469 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08007470 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08007471 "cpu": "watchos_armv7k",
7472 },
7473)
7474
7475config_setting(
7476 name = "watchos_arm64_32",
7477 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08007478 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08007479 "cpu": "watchos_arm64_32",
7480 },
7481)
7482
7483config_setting(
7484 name = "watchos_x86",
7485 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08007486 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08007487 "cpu": "watchos_i386",
7488 },
7489)
7490
7491config_setting(
7492 name = "watchos_x86_64",
7493 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08007494 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08007495 "cpu": "watchos_x86_64",
7496 },
7497)
7498
7499config_setting(
7500 name = "tvos_arm64",
7501 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08007502 "apple_platform_type": "tvos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08007503 "cpu": "tvos_arm64",
7504 },
7505)
7506
7507config_setting(
7508 name = "tvos_x86_64",
7509 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08007510 "apple_platform_type": "tvos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08007511 "cpu": "tvos_x86_64",
7512 },
7513)