blob: 1675b74377b63052bc97a3320d9cedd26541fe87 [file] [log] [blame]
Frank Barchard1a953052020-11-16 18:44:58 -08001# Copyright 2020 Google LLC
Marat Dukhan08c4a432019-10-03 09:29:21 -07002#
3# This source code is licensed under the BSD-style license found in the
4# LICENSE file in the root directory of this source tree.
5#
6# Description:
7# XNNPACK - optimized floating-point neural network operators library
8
Marat Dukhan10a38082020-04-17 03:58:35 -07009load(":build_defs.bzl", "xnnpack_aggregate_library", "xnnpack_benchmark", "xnnpack_binary", "xnnpack_cc_library", "xnnpack_gcc_std_copts", "xnnpack_min_size_copts", "xnnpack_msvc_std_copts", "xnnpack_optional_armcl_copts", "xnnpack_optional_armcl_deps", "xnnpack_optional_dnnl_copts", "xnnpack_optional_dnnl_deps", "xnnpack_optional_gemmlowp_copts", "xnnpack_optional_gemmlowp_deps", "xnnpack_optional_ruy_copts", "xnnpack_optional_ruy_deps", "xnnpack_optional_tflite_copts", "xnnpack_optional_tflite_deps", "xnnpack_std_cxxopts", "xnnpack_unit_test", "xnnpack_visibility")
Marat Dukhan69c3f2c2019-11-06 12:30:01 -080010
Marat Dukhan08c4a432019-10-03 09:29:21 -070011licenses(["notice"])
12
13exports_files(["LICENSE"])
14
Marat Dukhan1b354632020-03-23 12:50:22 -070015OPERATOR_BENCHMARK_DEPS = [
16 ":XNNPACK",
17 ":bench_utils",
18 "@cpuinfo",
Frank Barchard0c849732020-06-12 13:31:32 -070019 "@FP16",
Marat Dukhan1b354632020-03-23 12:50:22 -070020 "@pthreadpool",
21]
22
Marat Dukhan08c4a432019-10-03 09:29:21 -070023MICROKERNEL_BENCHMARK_DEPS = [
24 ":ukernels",
25 ":bench_utils",
Frank Barchard7e955972019-10-11 10:34:25 -070026 ":enable_assembly",
Marat Dukhan08c4a432019-10-03 09:29:21 -070027 "@cpuinfo",
28 "@FP16",
29 "@pthreadpool",
30]
31
Marat Dukhan6adff4e2019-10-14 18:32:07 -070032ACCURACY_EVAL_DEPS = [
33 ":XNNPACK",
34 ":ukernels",
35 "@FP16",
36 "@pthreadpool",
37]
38
Marat Dukhan08c4a432019-10-03 09:29:21 -070039MICROKERNEL_TEST_DEPS = [
Marat Dukhan33fcf782020-05-24 14:27:15 -070040 ":ukernels_test_mode",
Frank Barchard7e955972019-10-11 10:34:25 -070041 ":enable_assembly",
Marat Dukhan08c4a432019-10-03 09:29:21 -070042 "@cpuinfo",
43 "@FP16",
44 "@pthreadpool",
45]
46
Marat Dukhan1b354632020-03-23 12:50:22 -070047OPERATOR_TEST_DEPS = [
Marat Dukhan33fcf782020-05-24 14:27:15 -070048 ":XNNPACK_test_mode",
Marat Dukhan1b354632020-03-23 12:50:22 -070049 "@pthreadpool",
50 "@FP16",
51]
52
Marat Dukhan08c4a432019-10-03 09:29:21 -070053OPERATOR_SRCS = [
Marat Dukhane8265432020-04-28 18:42:59 -070054 "src/operators/argmax-pooling-nhwc.c",
55 "src/operators/average-pooling-nhwc.c",
56 "src/operators/binary-elementwise-nd.c",
Marat Dukhane8265432020-04-28 18:42:59 -070057 "src/operators/channel-shuffle-nc.c",
Marat Dukhan065b11e2020-05-22 09:49:41 -070058 "src/operators/constant-pad-nd.c",
Marat Dukhane8265432020-04-28 18:42:59 -070059 "src/operators/convolution-nchw.c",
60 "src/operators/convolution-nhwc.c",
61 "src/operators/deconvolution-nhwc.c",
Marat Dukhan13b68f22020-11-12 11:55:19 -080062 "src/operators/depth-to-space-nchw2nhwc.c",
Marat Dukhan0e521172020-11-25 13:10:04 -080063 "src/operators/depth-to-space-nhwc.c",
Marat Dukhane8265432020-04-28 18:42:59 -070064 "src/operators/fully-connected-nc.c",
65 "src/operators/global-average-pooling-ncw.c",
66 "src/operators/global-average-pooling-nwc.c",
Marat Dukhane8265432020-04-28 18:42:59 -070067 "src/operators/leaky-relu-nc.c",
68 "src/operators/max-pooling-nhwc.c",
69 "src/operators/prelu-nc.c",
Artsiom Ablavatski97918102020-10-27 15:52:59 -070070 "src/operators/resize-bilinear-nchw.c",
Marat Dukhane8265432020-04-28 18:42:59 -070071 "src/operators/resize-bilinear-nhwc.c",
72 "src/operators/sigmoid-nc.c",
73 "src/operators/softmax-nc.c",
Marat Dukhanc3065f52020-06-04 13:33:32 -070074 "src/operators/unary-elementwise-nc.c",
Marat Dukhane8265432020-04-28 18:42:59 -070075 "src/operators/unpooling-nhwc.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -070076]
77
Marat Dukhan95e8b7a2020-06-03 12:46:26 -070078SUBGRAPH_SRCS = [
Marat Dukhan5fab4092020-06-10 01:28:28 -070079 "src/subgraph/abs.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -070080 "src/subgraph/add2.c",
81 "src/subgraph/argmax-pooling-2d.c",
82 "src/subgraph/average-pooling-2d.c",
Marat Dukhan5fab4092020-06-10 01:28:28 -070083 "src/subgraph/bankers-rounding.c",
84 "src/subgraph/ceiling.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -070085 "src/subgraph/clamp.c",
86 "src/subgraph/convolution-2d.c",
87 "src/subgraph/deconvolution-2d.c",
Artsiom Ablavatskibbe85062020-11-05 14:07:37 -080088 "src/subgraph/depth-to-space.c",
Frank Barchard9cef5ea2020-11-18 14:52:08 -080089 "src/subgraph/depthwise-convolution-2d.c",
Marat Dukhan9d3a4592020-06-05 16:52:42 -070090 "src/subgraph/divide.c",
Marat Dukhana1600202020-12-01 22:17:16 -080091 "src/subgraph/elu.c",
Marat Dukhan5fab4092020-06-10 01:28:28 -070092 "src/subgraph/floor.c",
Frank Barchard04336c12020-10-22 16:48:55 -070093 "src/subgraph/fully-connected.c",
Marat Dukhana059b7d2020-06-11 11:41:27 -070094 "src/subgraph/global-average-pooling-2d.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -070095 "src/subgraph/hardswish.c",
Marat Dukhan5bbebac2020-06-10 19:42:15 -070096 "src/subgraph/leaky-relu.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -070097 "src/subgraph/max-pooling-2d.c",
Marat Dukhan9d3a4592020-06-05 16:52:42 -070098 "src/subgraph/maximum2.c",
99 "src/subgraph/minimum2.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -0700100 "src/subgraph/multiply2.c",
Marat Dukhan5fab4092020-06-10 01:28:28 -0700101 "src/subgraph/negate.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -0700102 "src/subgraph/prelu.c",
103 "src/subgraph/sigmoid.c",
104 "src/subgraph/softmax.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700105 "src/subgraph/square-root.c",
106 "src/subgraph/square.c",
107 "src/subgraph/squared-difference.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -0700108 "src/subgraph/static-constant-pad.c",
Marat Dukhand27202d2020-07-09 23:43:40 -0700109 "src/subgraph/static-reshape.c",
Marat Dukhanaff24e22020-07-23 01:43:58 -0700110 "src/subgraph/static-resize-bilinear-2d.c",
Marat Dukhan9d3a4592020-06-05 16:52:42 -0700111 "src/subgraph/subtract.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -0700112 "src/subgraph/unpooling-2d.c",
113]
114
Marat Dukhan3a77ea72019-12-23 12:10:24 -0800115TABLE_SRCS = [
116 "src/tables/exp2-k-over-64.c",
117 "src/tables/exp2-k-over-2048.c",
Marat Dukhande390d42020-11-29 19:32:18 -0800118 "src/tables/exp2minus-k-over-4.c",
119 "src/tables/exp2minus-k-over-8.c",
Marat Dukhanc60742b2020-11-23 12:33:27 -0800120 "src/tables/exp2minus-k-over-16.c",
Marat Dukhan1f256fc2020-09-24 21:27:14 -0700121 "src/tables/exp2minus-k-over-64.c",
122 "src/tables/exp2minus-k-over-2048.c",
Marat Dukhan3a77ea72019-12-23 12:10:24 -0800123]
124
Marat Dukhan08c4a432019-10-03 09:29:21 -0700125SCALAR_UKERNELS = [
Marat Dukhan329da642019-11-19 21:44:39 -0800126 "src/f32-argmaxpool/4x-scalar-c1.c",
Marat Dukhan1e782c42019-11-21 17:02:40 -0800127 "src/f32-argmaxpool/9p8x-scalar-c1.c",
Marat Dukhan329da642019-11-19 21:44:39 -0800128 "src/f32-argmaxpool/9x-scalar-c1.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700129 "src/f32-avgpool/9p8x-minmax-scalar-c1.c",
130 "src/f32-avgpool/9x-minmax-scalar-c1.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -0700131 "src/f32-clamp/gen/scalar-x1.c",
132 "src/f32-clamp/gen/scalar-x2.c",
133 "src/f32-clamp/gen/scalar-x4.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700134 "src/f32-conv-hwc/3x3s2p0p1c3x4-scalar-1x1.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700135 "src/f32-conv-hwc/3x3s2p1c3x4-scalar-1x1.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -0700136 "src/f32-conv-hwc2chw/3x3s2p1c3x4-scalar-1x1.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700137 "src/f32-dwconv/gen/up1x4-minmax-scalar-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700138 "src/f32-dwconv/gen/up1x4-minmax-scalar.c",
139 "src/f32-dwconv/gen/up1x4-scalar-acc2.c",
140 "src/f32-dwconv/gen/up1x4-scalar.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700141 "src/f32-dwconv/gen/up1x9-minmax-scalar-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700142 "src/f32-dwconv/gen/up1x9-minmax-scalar.c",
143 "src/f32-dwconv/gen/up1x9-scalar-acc2.c",
144 "src/f32-dwconv/gen/up1x9-scalar.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700145 "src/f32-dwconv/gen/up1x25-minmax-scalar-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700146 "src/f32-dwconv/gen/up1x25-minmax-scalar.c",
147 "src/f32-dwconv/gen/up1x25-scalar-acc2.c",
148 "src/f32-dwconv/gen/up1x25-scalar.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700149 "src/f32-dwconv/gen/up2x4-minmax-scalar-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700150 "src/f32-dwconv/gen/up2x4-minmax-scalar.c",
151 "src/f32-dwconv/gen/up2x4-scalar-acc2.c",
152 "src/f32-dwconv/gen/up2x4-scalar.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700153 "src/f32-dwconv/gen/up2x9-minmax-scalar-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700154 "src/f32-dwconv/gen/up2x9-minmax-scalar.c",
155 "src/f32-dwconv/gen/up2x9-scalar-acc2.c",
156 "src/f32-dwconv/gen/up2x9-scalar.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700157 "src/f32-dwconv/gen/up2x25-minmax-scalar-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700158 "src/f32-dwconv/gen/up2x25-minmax-scalar.c",
159 "src/f32-dwconv/gen/up2x25-scalar-acc2.c",
160 "src/f32-dwconv/gen/up2x25-scalar.c",
Frank Barchard35db7d02020-10-26 13:37:34 -0700161 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-scalar-1x1-acc2.c",
162 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-scalar-1x1-acc3.c",
163 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-scalar-1x1-acc4.c",
Marat Dukhan91249d22020-10-24 12:02:51 -0700164 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-scalar-1x1.c",
Frank Barchard35db7d02020-10-26 13:37:34 -0700165 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-scalar-2x1-acc2.c",
Marat Dukhan91249d22020-10-24 12:02:51 -0700166 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-scalar-2x1.c",
167 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-scalar-3x1.c",
168 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-scalar-4x1.c",
169 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-scalar-5x1.c",
170 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-scalar-6x1.c",
Marat Dukhancf5b3c32020-10-25 19:21:10 -0700171 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-scalar-1x1-acc2.c",
172 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-scalar-1x1-acc3.c",
173 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-scalar-1x1-acc4.c",
Frank Barchard35db7d02020-10-26 13:37:34 -0700174 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-scalar-1x1.c",
Marat Dukhancf5b3c32020-10-25 19:21:10 -0700175 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-scalar-2x1-acc2.c",
Frank Barchard35db7d02020-10-26 13:37:34 -0700176 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-scalar-2x1.c",
177 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-scalar-3x1.c",
178 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-scalar-4x1.c",
Marat Dukhanc4efb002020-10-25 23:14:47 -0700179 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-scalar-1x1-acc2.c",
180 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-scalar-1x1-acc3.c",
181 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-scalar-1x1-acc4.c",
182 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-scalar-1x1-acc5.c",
Frank Barchard35db7d02020-10-26 13:37:34 -0700183 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-scalar-1x1.c",
Marat Dukhanc4efb002020-10-25 23:14:47 -0700184 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-scalar-2x1-acc2.c",
185 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-scalar-2x1-acc3.c",
Frank Barchard35db7d02020-10-26 13:37:34 -0700186 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-scalar-2x1.c",
Marat Dukhanc4efb002020-10-25 23:14:47 -0700187 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-scalar-3x1-acc2.c",
Frank Barchard35db7d02020-10-26 13:37:34 -0700188 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-scalar-3x1.c",
Marat Dukhan29c0c332020-10-28 22:11:00 -0700189 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-scalar-1x1-acc2.c",
190 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-scalar-1x1-acc3.c",
191 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-scalar-1x1-acc4.c",
192 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-scalar-1x1-acc5.c",
193 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-scalar-1x1.c",
194 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-scalar-2x1-acc2.c",
195 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-scalar-2x1-acc3.c",
196 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-scalar-2x1.c",
197 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-scalar-3x1-acc2.c",
198 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-scalar-3x1.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -0700199 "src/f32-gavgpool-cw/scalar-x1.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700200 "src/f32-gavgpool/7p7x-minmax-scalar-c1.c",
201 "src/f32-gavgpool/7x-minmax-scalar-c1.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700202 "src/f32-gemm/gen-inc/1x4inc-minmax-scalar.c",
203 "src/f32-gemm/gen-inc/2x4inc-minmax-scalar.c",
204 "src/f32-gemm/gen-inc/4x4inc-minmax-scalar.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700205 "src/f32-gemm/gen/1x4-minmax-scalar.c",
206 "src/f32-gemm/gen/1x4-relu-scalar.c",
207 "src/f32-gemm/gen/1x4-scalar.c",
208 "src/f32-gemm/gen/2x4-minmax-scalar.c",
209 "src/f32-gemm/gen/2x4-relu-scalar.c",
210 "src/f32-gemm/gen/2x4-scalar.c",
211 "src/f32-gemm/gen/4x2-minmax-scalar.c",
212 "src/f32-gemm/gen/4x2-relu-scalar.c",
213 "src/f32-gemm/gen/4x2-scalar.c",
214 "src/f32-gemm/gen/4x4-minmax-scalar.c",
215 "src/f32-gemm/gen/4x4-relu-scalar.c",
216 "src/f32-gemm/gen/4x4-scalar.c",
Marat Dukhanefaac272020-07-16 16:58:07 -0700217 "src/f32-hswish/gen/hswish-scalar-x1.c",
218 "src/f32-hswish/gen/hswish-scalar-x2.c",
219 "src/f32-hswish/gen/hswish-scalar-x4.c",
XNNPACK Team21432672020-10-19 19:58:48 -0700220 "src/f32-ibilinear-chw/gen/scalar-p1.c",
221 "src/f32-ibilinear-chw/gen/scalar-p2.c",
222 "src/f32-ibilinear-chw/gen/scalar-p4.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700223 "src/f32-ibilinear/gen/scalar-c1.c",
224 "src/f32-ibilinear/gen/scalar-c2.c",
225 "src/f32-ibilinear/gen/scalar-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700226 "src/f32-igemm/gen/1x4-minmax-scalar.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700227 "src/f32-igemm/gen/1x4-relu-scalar.c",
228 "src/f32-igemm/gen/1x4-scalar.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700229 "src/f32-igemm/gen/2x4-minmax-scalar.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700230 "src/f32-igemm/gen/2x4-relu-scalar.c",
231 "src/f32-igemm/gen/2x4-scalar.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700232 "src/f32-igemm/gen/4x2-minmax-scalar.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700233 "src/f32-igemm/gen/4x2-relu-scalar.c",
234 "src/f32-igemm/gen/4x2-scalar.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700235 "src/f32-igemm/gen/4x4-minmax-scalar.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700236 "src/f32-igemm/gen/4x4-relu-scalar.c",
237 "src/f32-igemm/gen/4x4-scalar.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700238 "src/f32-maxpool/9p8x-minmax-scalar-c1.c",
239 "src/f32-pavgpool/9p8x-minmax-scalar-c1.c",
240 "src/f32-pavgpool/9x-minmax-scalar-c1.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700241 "src/f32-ppmm/gen/2x4-minmax-scalar.c",
242 "src/f32-ppmm/gen/3x3-minmax-scalar.c",
243 "src/f32-ppmm/gen/4x2-minmax-scalar.c",
244 "src/f32-ppmm/gen/4x4-minmax-scalar.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800245 "src/f32-prelu/gen/scalar-2x1.c",
246 "src/f32-prelu/gen/scalar-2x4.c",
Marat Dukhanf46f6752020-01-21 11:03:49 -0800247 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x1.c",
Marat Dukhanf46f6752020-01-21 11:03:49 -0800248 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x2-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700249 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x2.c",
Marat Dukhanf46f6752020-01-21 11:03:49 -0800250 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x4-acc2.c",
251 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x4-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700252 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x4.c",
Marat Dukhanf46f6752020-01-21 11:03:49 -0800253 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x1.c",
Marat Dukhanf46f6752020-01-21 11:03:49 -0800254 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x2-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700255 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x2.c",
Marat Dukhanf46f6752020-01-21 11:03:49 -0800256 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x4-acc2.c",
257 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x4-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700258 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x4.c",
Frank Barchardfb158e22020-07-15 16:10:10 -0700259 "src/f32-relu/gen/scalar-x1.c",
260 "src/f32-relu/gen/scalar-x2.c",
261 "src/f32-relu/gen/scalar-x4.c",
Frank Barchard37297a62020-08-27 12:15:10 -0700262 "src/f32-relu/gen/scalar-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700263 "src/f32-rmax/scalar.c",
Marat Dukhan3a77ea72019-12-23 12:10:24 -0800264 "src/f32-sigmoid/gen/scalar-lut64-p2-div-x1.c",
265 "src/f32-sigmoid/gen/scalar-lut64-p2-div-x2.c",
266 "src/f32-sigmoid/gen/scalar-lut64-p2-div-x4.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700267 "src/f32-sigmoid/gen/scalar-lut2048-p1-div-x1.c",
268 "src/f32-sigmoid/gen/scalar-lut2048-p1-div-x2.c",
269 "src/f32-sigmoid/gen/scalar-lut2048-p1-div-x4.c",
Marat Dukhan3a77ea72019-12-23 12:10:24 -0800270 "src/f32-sigmoid/gen/scalar-p5-div-x1.c",
271 "src/f32-sigmoid/gen/scalar-p5-div-x2.c",
272 "src/f32-sigmoid/gen/scalar-p5-div-x4.c",
Marat Dukhan355ab432020-04-09 19:01:52 -0700273 "src/f32-spmm/gen/1x1-minmax-scalar-pipelined.c",
274 "src/f32-spmm/gen/1x1-minmax-scalar.c",
275 "src/f32-spmm/gen/2x1-minmax-scalar-pipelined.c",
276 "src/f32-spmm/gen/2x1-minmax-scalar.c",
277 "src/f32-spmm/gen/4x1-minmax-scalar-pipelined.c",
278 "src/f32-spmm/gen/4x1-minmax-scalar.c",
279 "src/f32-spmm/gen/8x1-minmax-scalar-pipelined.c",
280 "src/f32-spmm/gen/8x1-minmax-scalar.c",
281 "src/f32-spmm/gen/8x2-minmax-scalar.c",
282 "src/f32-spmm/gen/8x4-minmax-scalar.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700283 "src/f32-vbinary/gen/vadd-minmax-scalar-x1.c",
284 "src/f32-vbinary/gen/vadd-minmax-scalar-x2.c",
285 "src/f32-vbinary/gen/vadd-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700286 "src/f32-vbinary/gen/vadd-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700287 "src/f32-vbinary/gen/vadd-relu-scalar-x1.c",
288 "src/f32-vbinary/gen/vadd-relu-scalar-x2.c",
289 "src/f32-vbinary/gen/vadd-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700290 "src/f32-vbinary/gen/vadd-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700291 "src/f32-vbinary/gen/vadd-scalar-x1.c",
292 "src/f32-vbinary/gen/vadd-scalar-x2.c",
293 "src/f32-vbinary/gen/vadd-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700294 "src/f32-vbinary/gen/vadd-scalar-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700295 "src/f32-vbinary/gen/vaddc-minmax-scalar-x1.c",
296 "src/f32-vbinary/gen/vaddc-minmax-scalar-x2.c",
297 "src/f32-vbinary/gen/vaddc-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700298 "src/f32-vbinary/gen/vaddc-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700299 "src/f32-vbinary/gen/vaddc-relu-scalar-x1.c",
300 "src/f32-vbinary/gen/vaddc-relu-scalar-x2.c",
301 "src/f32-vbinary/gen/vaddc-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700302 "src/f32-vbinary/gen/vaddc-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700303 "src/f32-vbinary/gen/vaddc-scalar-x1.c",
304 "src/f32-vbinary/gen/vaddc-scalar-x2.c",
305 "src/f32-vbinary/gen/vaddc-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700306 "src/f32-vbinary/gen/vaddc-scalar-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700307 "src/f32-vbinary/gen/vdiv-minmax-scalar-x1.c",
308 "src/f32-vbinary/gen/vdiv-minmax-scalar-x2.c",
309 "src/f32-vbinary/gen/vdiv-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700310 "src/f32-vbinary/gen/vdiv-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700311 "src/f32-vbinary/gen/vdiv-relu-scalar-x1.c",
312 "src/f32-vbinary/gen/vdiv-relu-scalar-x2.c",
313 "src/f32-vbinary/gen/vdiv-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700314 "src/f32-vbinary/gen/vdiv-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700315 "src/f32-vbinary/gen/vdiv-scalar-x1.c",
316 "src/f32-vbinary/gen/vdiv-scalar-x2.c",
317 "src/f32-vbinary/gen/vdiv-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700318 "src/f32-vbinary/gen/vdiv-scalar-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700319 "src/f32-vbinary/gen/vdivc-minmax-scalar-x1.c",
320 "src/f32-vbinary/gen/vdivc-minmax-scalar-x2.c",
321 "src/f32-vbinary/gen/vdivc-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700322 "src/f32-vbinary/gen/vdivc-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700323 "src/f32-vbinary/gen/vdivc-relu-scalar-x1.c",
324 "src/f32-vbinary/gen/vdivc-relu-scalar-x2.c",
325 "src/f32-vbinary/gen/vdivc-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700326 "src/f32-vbinary/gen/vdivc-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700327 "src/f32-vbinary/gen/vdivc-scalar-x1.c",
328 "src/f32-vbinary/gen/vdivc-scalar-x2.c",
329 "src/f32-vbinary/gen/vdivc-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700330 "src/f32-vbinary/gen/vdivc-scalar-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800331 "src/f32-vbinary/gen/vmax-scalar-x1.c",
332 "src/f32-vbinary/gen/vmax-scalar-x2.c",
333 "src/f32-vbinary/gen/vmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700334 "src/f32-vbinary/gen/vmax-scalar-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800335 "src/f32-vbinary/gen/vmaxc-scalar-x1.c",
336 "src/f32-vbinary/gen/vmaxc-scalar-x2.c",
337 "src/f32-vbinary/gen/vmaxc-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700338 "src/f32-vbinary/gen/vmaxc-scalar-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800339 "src/f32-vbinary/gen/vmin-scalar-x1.c",
340 "src/f32-vbinary/gen/vmin-scalar-x2.c",
341 "src/f32-vbinary/gen/vmin-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700342 "src/f32-vbinary/gen/vmin-scalar-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800343 "src/f32-vbinary/gen/vminc-scalar-x1.c",
344 "src/f32-vbinary/gen/vminc-scalar-x2.c",
345 "src/f32-vbinary/gen/vminc-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700346 "src/f32-vbinary/gen/vminc-scalar-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700347 "src/f32-vbinary/gen/vmul-minmax-scalar-x1.c",
348 "src/f32-vbinary/gen/vmul-minmax-scalar-x2.c",
349 "src/f32-vbinary/gen/vmul-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700350 "src/f32-vbinary/gen/vmul-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700351 "src/f32-vbinary/gen/vmul-relu-scalar-x1.c",
352 "src/f32-vbinary/gen/vmul-relu-scalar-x2.c",
353 "src/f32-vbinary/gen/vmul-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700354 "src/f32-vbinary/gen/vmul-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700355 "src/f32-vbinary/gen/vmul-scalar-x1.c",
356 "src/f32-vbinary/gen/vmul-scalar-x2.c",
357 "src/f32-vbinary/gen/vmul-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700358 "src/f32-vbinary/gen/vmul-scalar-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700359 "src/f32-vbinary/gen/vmulc-minmax-scalar-x1.c",
360 "src/f32-vbinary/gen/vmulc-minmax-scalar-x2.c",
361 "src/f32-vbinary/gen/vmulc-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700362 "src/f32-vbinary/gen/vmulc-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700363 "src/f32-vbinary/gen/vmulc-relu-scalar-x1.c",
364 "src/f32-vbinary/gen/vmulc-relu-scalar-x2.c",
365 "src/f32-vbinary/gen/vmulc-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700366 "src/f32-vbinary/gen/vmulc-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700367 "src/f32-vbinary/gen/vmulc-scalar-x1.c",
368 "src/f32-vbinary/gen/vmulc-scalar-x2.c",
369 "src/f32-vbinary/gen/vmulc-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700370 "src/f32-vbinary/gen/vmulc-scalar-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700371 "src/f32-vbinary/gen/vrdivc-minmax-scalar-x1.c",
372 "src/f32-vbinary/gen/vrdivc-minmax-scalar-x2.c",
373 "src/f32-vbinary/gen/vrdivc-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700374 "src/f32-vbinary/gen/vrdivc-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700375 "src/f32-vbinary/gen/vrdivc-relu-scalar-x1.c",
376 "src/f32-vbinary/gen/vrdivc-relu-scalar-x2.c",
377 "src/f32-vbinary/gen/vrdivc-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700378 "src/f32-vbinary/gen/vrdivc-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700379 "src/f32-vbinary/gen/vrdivc-scalar-x1.c",
380 "src/f32-vbinary/gen/vrdivc-scalar-x2.c",
381 "src/f32-vbinary/gen/vrdivc-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700382 "src/f32-vbinary/gen/vrdivc-scalar-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700383 "src/f32-vbinary/gen/vrsubc-minmax-scalar-x1.c",
384 "src/f32-vbinary/gen/vrsubc-minmax-scalar-x2.c",
385 "src/f32-vbinary/gen/vrsubc-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700386 "src/f32-vbinary/gen/vrsubc-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700387 "src/f32-vbinary/gen/vrsubc-relu-scalar-x1.c",
388 "src/f32-vbinary/gen/vrsubc-relu-scalar-x2.c",
389 "src/f32-vbinary/gen/vrsubc-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700390 "src/f32-vbinary/gen/vrsubc-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700391 "src/f32-vbinary/gen/vrsubc-scalar-x1.c",
392 "src/f32-vbinary/gen/vrsubc-scalar-x2.c",
393 "src/f32-vbinary/gen/vrsubc-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700394 "src/f32-vbinary/gen/vrsubc-scalar-x8.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -0700395 "src/f32-vbinary/gen/vsqrdiff-scalar-x1.c",
396 "src/f32-vbinary/gen/vsqrdiff-scalar-x2.c",
397 "src/f32-vbinary/gen/vsqrdiff-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700398 "src/f32-vbinary/gen/vsqrdiff-scalar-x8.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -0700399 "src/f32-vbinary/gen/vsqrdiffc-scalar-x1.c",
400 "src/f32-vbinary/gen/vsqrdiffc-scalar-x2.c",
401 "src/f32-vbinary/gen/vsqrdiffc-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700402 "src/f32-vbinary/gen/vsqrdiffc-scalar-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700403 "src/f32-vbinary/gen/vsub-minmax-scalar-x1.c",
404 "src/f32-vbinary/gen/vsub-minmax-scalar-x2.c",
405 "src/f32-vbinary/gen/vsub-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700406 "src/f32-vbinary/gen/vsub-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700407 "src/f32-vbinary/gen/vsub-relu-scalar-x1.c",
408 "src/f32-vbinary/gen/vsub-relu-scalar-x2.c",
409 "src/f32-vbinary/gen/vsub-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700410 "src/f32-vbinary/gen/vsub-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700411 "src/f32-vbinary/gen/vsub-scalar-x1.c",
412 "src/f32-vbinary/gen/vsub-scalar-x2.c",
413 "src/f32-vbinary/gen/vsub-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700414 "src/f32-vbinary/gen/vsub-scalar-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700415 "src/f32-vbinary/gen/vsubc-minmax-scalar-x1.c",
416 "src/f32-vbinary/gen/vsubc-minmax-scalar-x2.c",
417 "src/f32-vbinary/gen/vsubc-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700418 "src/f32-vbinary/gen/vsubc-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700419 "src/f32-vbinary/gen/vsubc-relu-scalar-x1.c",
420 "src/f32-vbinary/gen/vsubc-relu-scalar-x2.c",
421 "src/f32-vbinary/gen/vsubc-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700422 "src/f32-vbinary/gen/vsubc-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700423 "src/f32-vbinary/gen/vsubc-scalar-x1.c",
424 "src/f32-vbinary/gen/vsubc-scalar-x2.c",
425 "src/f32-vbinary/gen/vsubc-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700426 "src/f32-vbinary/gen/vsubc-scalar-x8.c",
Marat Dukhaned6baaf2020-12-01 15:07:08 -0800427 "src/f32-velu/gen/velu-scalar-rr2-lut16-p3-x1.c",
428 "src/f32-velu/gen/velu-scalar-rr2-lut16-p3-x2.c",
429 "src/f32-velu/gen/velu-scalar-rr2-lut16-p3-x3.c",
430 "src/f32-velu/gen/velu-scalar-rr2-lut16-p3-x4.c",
431 "src/f32-velu/gen/velu-scalar-rr2-lut16-p3-x5.c",
432 "src/f32-velu/gen/velu-scalar-rr2-lut16-p3-x6.c",
433 "src/f32-velu/gen/velu-scalar-rr2-p6-x1.c",
434 "src/f32-velu/gen/velu-scalar-rr2-p6-x2.c",
435 "src/f32-velu/gen/velu-scalar-rr2-p6-x3.c",
436 "src/f32-velu/gen/velu-scalar-rr2-p6-x4.c",
437 "src/f32-velu/gen/velu-scalar-rr2-p6-x5.c",
438 "src/f32-velu/gen/velu-scalar-rr2-p6-x6.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700439 "src/f32-vlrelu/gen/vlrelu-scalar-x1.c",
440 "src/f32-vlrelu/gen/vlrelu-scalar-x2.c",
441 "src/f32-vlrelu/gen/vlrelu-scalar-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700442 "src/f32-vmulcaddc/gen/c1-minmax-scalar-2x.c",
443 "src/f32-vmulcaddc/gen/c2-minmax-scalar-2x.c",
444 "src/f32-vmulcaddc/gen/c4-minmax-scalar-2x.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -0700445 "src/f32-vrnd/gen/vrndd-scalar-libm-x1.c",
446 "src/f32-vrnd/gen/vrndd-scalar-libm-x2.c",
447 "src/f32-vrnd/gen/vrndd-scalar-libm-x4.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700448 "src/f32-vrnd/gen/vrndne-scalar-libm-x1.c",
449 "src/f32-vrnd/gen/vrndne-scalar-libm-x2.c",
450 "src/f32-vrnd/gen/vrndne-scalar-libm-x4.c",
451 "src/f32-vrnd/gen/vrndu-scalar-libm-x1.c",
452 "src/f32-vrnd/gen/vrndu-scalar-libm-x2.c",
453 "src/f32-vrnd/gen/vrndu-scalar-libm-x4.c",
454 "src/f32-vrnd/gen/vrndz-scalar-libm-x1.c",
455 "src/f32-vrnd/gen/vrndz-scalar-libm-x2.c",
456 "src/f32-vrnd/gen/vrndz-scalar-libm-x4.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -0700457 "src/f32-vsqrt/gen/scalar-sqrt-x1.c",
458 "src/f32-vsqrt/gen/scalar-sqrt-x2.c",
459 "src/f32-vsqrt/gen/scalar-sqrt-x4.c",
Marat Dukhan5020b962020-06-08 13:30:10 -0700460 "src/f32-vunary/gen/vabs-scalar-x1.c",
461 "src/f32-vunary/gen/vabs-scalar-x2.c",
462 "src/f32-vunary/gen/vabs-scalar-x4.c",
463 "src/f32-vunary/gen/vneg-scalar-x1.c",
464 "src/f32-vunary/gen/vneg-scalar-x2.c",
465 "src/f32-vunary/gen/vneg-scalar-x4.c",
466 "src/f32-vunary/gen/vsqr-scalar-x1.c",
467 "src/f32-vunary/gen/vsqr-scalar-x2.c",
468 "src/f32-vunary/gen/vsqr-scalar-x4.c",
Marat Dukhande390d42020-11-29 19:32:18 -0800469 "src/math/expm1minus-scalar-rr2-lut4-p4.c",
470 "src/math/expm1minus-scalar-rr2-lut8-p3.c",
471 "src/math/expm1minus-scalar-rr2-lut8-p4.c",
Marat Dukhanc60742b2020-11-23 12:33:27 -0800472 "src/math/expm1minus-scalar-rr2-lut16-p3.c",
473 "src/math/expm1minus-scalar-rr2-lut16-p4.c",
474 "src/math/expm1minus-scalar-rr2-p5.c",
475 "src/math/expm1minus-scalar-rr2-p6.c",
Frank Barchard22136062020-11-24 18:44:46 -0800476 "src/math/expminus-scalar-rr2-lut64-p2.c",
477 "src/math/expminus-scalar-rr2-lut2048-p1.c",
478 "src/math/expminus-scalar-rr2-p5.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -0700479 "src/math/roundd-scalar-addsub.c",
480 "src/math/roundd-scalar-cvt.c",
Marat Dukhanffbf96a2020-05-14 02:59:08 -0700481 "src/math/roundd-scalar-floor.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700482 "src/math/roundne-scalar-addsub.c",
483 "src/math/roundne-scalar-nearbyint.c",
484 "src/math/roundne-scalar-rint.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -0700485 "src/math/roundu-scalar-addsub.c",
Marat Dukhanffbf96a2020-05-14 02:59:08 -0700486 "src/math/roundu-scalar-ceil.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -0700487 "src/math/roundu-scalar-cvt.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -0700488 "src/math/roundz-scalar-addsub.c",
489 "src/math/roundz-scalar-cvt.c",
Marat Dukhanffbf96a2020-05-14 02:59:08 -0700490 "src/math/roundz-scalar-trunc.c",
Marat Dukhanf8475d62020-09-17 15:01:43 -0700491 "src/math/sigmoid-scalar-rr2-lut64-p2-div.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700492 "src/math/sigmoid-scalar-rr2-lut2048-p1-div.c",
Marat Dukhanf8475d62020-09-17 15:01:43 -0700493 "src/math/sigmoid-scalar-rr2-p5-div.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -0700494 "src/qs8-requantization/fp32-scalar-lrintf.c",
495 "src/qs8-requantization/fp32-scalar-magic.c",
496 "src/qs8-requantization/precise-scalar-signed64.c",
497 "src/qs8-requantization/precise-scalar-unsigned32.c",
498 "src/qs8-requantization/precise-scalar-unsigned64.c",
499 "src/qs8-requantization/q31-scalar.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -0700500 "src/qu8-avgpool/9p8x-minmax-scalar-c1.c",
501 "src/qu8-avgpool/9x-minmax-scalar-c1.c",
502 "src/qu8-dwconv/up1x9-minmax-scalar.c",
503 "src/qu8-gavgpool/7p7x-minmax-scalar-c1.c",
504 "src/qu8-gavgpool/7x-minmax-scalar-c1.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700505 "src/qu8-gemm/2x2-minmax-scalar.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -0700506 "src/qu8-igemm/2x2-minmax-scalar.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -0700507 "src/qu8-requantization/fp32-scalar-lrintf.c",
508 "src/qu8-requantization/fp32-scalar-magic.c",
509 "src/qu8-requantization/precise-scalar-signed64.c",
510 "src/qu8-requantization/precise-scalar-unsigned32.c",
511 "src/qu8-requantization/precise-scalar-unsigned64.c",
512 "src/qu8-requantization/q31-scalar.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -0700513 "src/qu8-vadd/minmax-scalar.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -0700514 "src/u8-clamp/scalar-x4.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700515 "src/u8-lut32norm/scalar.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700516 "src/u8-maxpool/9p8x-minmax-scalar-c1.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700517 "src/u8-rmax/scalar.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700518 "src/x8-lut/scalar.c",
519 "src/x8-zip/x2-scalar.c",
520 "src/x8-zip/x3-scalar.c",
521 "src/x8-zip/x4-scalar.c",
522 "src/x8-zip/xm-scalar.c",
Marat Dukhanad71b9a2020-11-20 00:01:51 -0800523 "src/x32-depthtospace2d-chw2hwc/scalar.c",
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -0700524 "src/x32-fill/scalar-float.c",
525 "src/x32-fill/scalar-int.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700526 "src/x32-packx/x2-scalar.c",
527 "src/x32-packx/x3-scalar.c",
528 "src/x32-packx/x4-scalar.c",
Marat Dukhan63523d42020-05-22 17:07:33 -0700529 "src/x32-pad/scalar-float.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700530 "src/x32-pad/scalar-int.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700531 "src/x32-unpool/scalar.c",
532 "src/x32-zip/x2-scalar.c",
533 "src/x32-zip/x3-scalar.c",
534 "src/x32-zip/x4-scalar.c",
535 "src/x32-zip/xm-scalar.c",
Marat Dukhan048931b2020-11-24 20:53:54 -0800536 "src/xx-copy/memcpy.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700537]
538
Marat Dukhan436ebe62019-12-04 15:10:12 -0800539WASM_UKERNELS = [
Marat Dukhan99936602020-04-11 16:47:01 -0700540 "src/f32-avgpool/9p8x-minmax-wasm-c1.c",
541 "src/f32-avgpool/9x-minmax-wasm-c1.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -0700542 "src/f32-clamp/gen/wasm-x1.c",
543 "src/f32-clamp/gen/wasm-x2.c",
544 "src/f32-clamp/gen/wasm-x4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700545 "src/f32-dwconv/gen/up1x4-minmax-wasm-acc2.c",
546 "src/f32-dwconv/gen/up1x4-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700547 "src/f32-dwconv/gen/up1x4-wasm-acc2.c",
548 "src/f32-dwconv/gen/up1x4-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700549 "src/f32-dwconv/gen/up1x9-minmax-wasm-acc2.c",
550 "src/f32-dwconv/gen/up1x9-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700551 "src/f32-dwconv/gen/up1x9-wasm-acc2.c",
552 "src/f32-dwconv/gen/up1x9-wasm.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700553 "src/f32-dwconv/gen/up1x25-minmax-wasm-acc2.c",
554 "src/f32-dwconv/gen/up1x25-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700555 "src/f32-dwconv/gen/up1x25-wasm-acc2.c",
556 "src/f32-dwconv/gen/up1x25-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700557 "src/f32-dwconv/gen/up2x4-minmax-wasm-acc2.c",
558 "src/f32-dwconv/gen/up2x4-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700559 "src/f32-dwconv/gen/up2x4-wasm-acc2.c",
560 "src/f32-dwconv/gen/up2x4-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700561 "src/f32-dwconv/gen/up2x9-minmax-wasm-acc2.c",
562 "src/f32-dwconv/gen/up2x9-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700563 "src/f32-dwconv/gen/up2x9-wasm-acc2.c",
564 "src/f32-dwconv/gen/up2x9-wasm.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700565 "src/f32-dwconv/gen/up2x25-minmax-wasm-acc2.c",
566 "src/f32-dwconv/gen/up2x25-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700567 "src/f32-dwconv/gen/up2x25-wasm-acc2.c",
568 "src/f32-dwconv/gen/up2x25-wasm.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700569 "src/f32-gavgpool/7p7x-minmax-wasm-c1.c",
570 "src/f32-gavgpool/7x-minmax-wasm-c1.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700571 "src/f32-gemm/gen-inc/1x4inc-minmax-wasm.c",
572 "src/f32-gemm/gen-inc/2x4inc-minmax-wasm.c",
573 "src/f32-gemm/gen-inc/4x4inc-minmax-wasm.c",
574 "src/f32-gemm/gen/1x4-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700575 "src/f32-gemm/gen/1x4-relu-wasm.c",
576 "src/f32-gemm/gen/1x4-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700577 "src/f32-gemm/gen/2x4-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700578 "src/f32-gemm/gen/2x4-relu-wasm.c",
579 "src/f32-gemm/gen/2x4-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700580 "src/f32-gemm/gen/4x2-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700581 "src/f32-gemm/gen/4x2-relu-wasm.c",
582 "src/f32-gemm/gen/4x2-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700583 "src/f32-gemm/gen/4x4-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700584 "src/f32-gemm/gen/4x4-relu-wasm.c",
585 "src/f32-gemm/gen/4x4-wasm.c",
Marat Dukhanefaac272020-07-16 16:58:07 -0700586 "src/f32-hswish/gen/hswish-wasm-x1.c",
587 "src/f32-hswish/gen/hswish-wasm-x2.c",
588 "src/f32-hswish/gen/hswish-wasm-x4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700589 "src/f32-igemm/gen/1x4-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700590 "src/f32-igemm/gen/1x4-relu-wasm.c",
591 "src/f32-igemm/gen/1x4-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700592 "src/f32-igemm/gen/2x4-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700593 "src/f32-igemm/gen/2x4-relu-wasm.c",
594 "src/f32-igemm/gen/2x4-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700595 "src/f32-igemm/gen/4x2-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700596 "src/f32-igemm/gen/4x2-relu-wasm.c",
597 "src/f32-igemm/gen/4x2-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700598 "src/f32-igemm/gen/4x4-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700599 "src/f32-igemm/gen/4x4-relu-wasm.c",
600 "src/f32-igemm/gen/4x4-wasm.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700601 "src/f32-maxpool/9p8x-minmax-wasm-c1.c",
602 "src/f32-pavgpool/9p8x-minmax-wasm-c1.c",
603 "src/f32-pavgpool/9x-minmax-wasm-c1.c",
Marat Dukhan7c1f8082020-06-25 13:26:20 -0700604 "src/f32-prelu/gen/wasm-2x1.c",
605 "src/f32-prelu/gen/wasm-2x4.c",
Frank Barchardfb158e22020-07-15 16:10:10 -0700606 "src/f32-relu/gen/wasm-x1.c",
607 "src/f32-relu/gen/wasm-x2.c",
608 "src/f32-relu/gen/wasm-x4.c",
Frank Barchard37297a62020-08-27 12:15:10 -0700609 "src/f32-relu/gen/wasm-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700610 "src/f32-vbinary/gen/vadd-minmax-wasm-x1.c",
611 "src/f32-vbinary/gen/vadd-minmax-wasm-x2.c",
612 "src/f32-vbinary/gen/vadd-minmax-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700613 "src/f32-vbinary/gen/vadd-minmax-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700614 "src/f32-vbinary/gen/vadd-relu-wasm-x1.c",
615 "src/f32-vbinary/gen/vadd-relu-wasm-x2.c",
616 "src/f32-vbinary/gen/vadd-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700617 "src/f32-vbinary/gen/vadd-relu-wasm-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700618 "src/f32-vbinary/gen/vaddc-minmax-wasm-x1.c",
619 "src/f32-vbinary/gen/vaddc-minmax-wasm-x2.c",
620 "src/f32-vbinary/gen/vaddc-minmax-wasm-x4.c",
621 "src/f32-vbinary/gen/vaddc-minmax-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700622 "src/f32-vbinary/gen/vaddc-relu-wasm-x1.c",
623 "src/f32-vbinary/gen/vaddc-relu-wasm-x2.c",
624 "src/f32-vbinary/gen/vaddc-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700625 "src/f32-vbinary/gen/vaddc-relu-wasm-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700626 "src/f32-vbinary/gen/vdiv-minmax-wasm-x1.c",
627 "src/f32-vbinary/gen/vdiv-minmax-wasm-x2.c",
628 "src/f32-vbinary/gen/vdiv-minmax-wasm-x4.c",
629 "src/f32-vbinary/gen/vdiv-minmax-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700630 "src/f32-vbinary/gen/vdiv-relu-wasm-x1.c",
631 "src/f32-vbinary/gen/vdiv-relu-wasm-x2.c",
632 "src/f32-vbinary/gen/vdiv-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700633 "src/f32-vbinary/gen/vdiv-relu-wasm-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700634 "src/f32-vbinary/gen/vdivc-minmax-wasm-x1.c",
635 "src/f32-vbinary/gen/vdivc-minmax-wasm-x2.c",
636 "src/f32-vbinary/gen/vdivc-minmax-wasm-x4.c",
637 "src/f32-vbinary/gen/vdivc-minmax-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700638 "src/f32-vbinary/gen/vdivc-relu-wasm-x1.c",
639 "src/f32-vbinary/gen/vdivc-relu-wasm-x2.c",
640 "src/f32-vbinary/gen/vdivc-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700641 "src/f32-vbinary/gen/vdivc-relu-wasm-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800642 "src/f32-vbinary/gen/vmax-wasm-x1.c",
643 "src/f32-vbinary/gen/vmax-wasm-x2.c",
644 "src/f32-vbinary/gen/vmax-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700645 "src/f32-vbinary/gen/vmax-wasm-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800646 "src/f32-vbinary/gen/vmaxc-wasm-x1.c",
647 "src/f32-vbinary/gen/vmaxc-wasm-x2.c",
648 "src/f32-vbinary/gen/vmaxc-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700649 "src/f32-vbinary/gen/vmaxc-wasm-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800650 "src/f32-vbinary/gen/vmin-wasm-x1.c",
651 "src/f32-vbinary/gen/vmin-wasm-x2.c",
652 "src/f32-vbinary/gen/vmin-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700653 "src/f32-vbinary/gen/vmin-wasm-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800654 "src/f32-vbinary/gen/vminc-wasm-x1.c",
655 "src/f32-vbinary/gen/vminc-wasm-x2.c",
656 "src/f32-vbinary/gen/vminc-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700657 "src/f32-vbinary/gen/vminc-wasm-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700658 "src/f32-vbinary/gen/vmul-minmax-wasm-x1.c",
659 "src/f32-vbinary/gen/vmul-minmax-wasm-x2.c",
660 "src/f32-vbinary/gen/vmul-minmax-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700661 "src/f32-vbinary/gen/vmul-minmax-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700662 "src/f32-vbinary/gen/vmul-relu-wasm-x1.c",
663 "src/f32-vbinary/gen/vmul-relu-wasm-x2.c",
664 "src/f32-vbinary/gen/vmul-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700665 "src/f32-vbinary/gen/vmul-relu-wasm-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700666 "src/f32-vbinary/gen/vmulc-minmax-wasm-x1.c",
667 "src/f32-vbinary/gen/vmulc-minmax-wasm-x2.c",
668 "src/f32-vbinary/gen/vmulc-minmax-wasm-x4.c",
669 "src/f32-vbinary/gen/vmulc-minmax-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700670 "src/f32-vbinary/gen/vmulc-relu-wasm-x1.c",
671 "src/f32-vbinary/gen/vmulc-relu-wasm-x2.c",
672 "src/f32-vbinary/gen/vmulc-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700673 "src/f32-vbinary/gen/vmulc-relu-wasm-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700674 "src/f32-vbinary/gen/vrdivc-minmax-wasm-x1.c",
675 "src/f32-vbinary/gen/vrdivc-minmax-wasm-x2.c",
676 "src/f32-vbinary/gen/vrdivc-minmax-wasm-x4.c",
677 "src/f32-vbinary/gen/vrdivc-minmax-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700678 "src/f32-vbinary/gen/vrdivc-relu-wasm-x1.c",
679 "src/f32-vbinary/gen/vrdivc-relu-wasm-x2.c",
680 "src/f32-vbinary/gen/vrdivc-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700681 "src/f32-vbinary/gen/vrdivc-relu-wasm-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700682 "src/f32-vbinary/gen/vrsubc-minmax-wasm-x1.c",
683 "src/f32-vbinary/gen/vrsubc-minmax-wasm-x2.c",
684 "src/f32-vbinary/gen/vrsubc-minmax-wasm-x4.c",
685 "src/f32-vbinary/gen/vrsubc-minmax-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700686 "src/f32-vbinary/gen/vrsubc-relu-wasm-x1.c",
687 "src/f32-vbinary/gen/vrsubc-relu-wasm-x2.c",
688 "src/f32-vbinary/gen/vrsubc-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700689 "src/f32-vbinary/gen/vrsubc-relu-wasm-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700690 "src/f32-vbinary/gen/vsub-minmax-wasm-x1.c",
691 "src/f32-vbinary/gen/vsub-minmax-wasm-x2.c",
692 "src/f32-vbinary/gen/vsub-minmax-wasm-x4.c",
693 "src/f32-vbinary/gen/vsub-minmax-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700694 "src/f32-vbinary/gen/vsub-relu-wasm-x1.c",
695 "src/f32-vbinary/gen/vsub-relu-wasm-x2.c",
696 "src/f32-vbinary/gen/vsub-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700697 "src/f32-vbinary/gen/vsub-relu-wasm-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700698 "src/f32-vbinary/gen/vsubc-minmax-wasm-x1.c",
699 "src/f32-vbinary/gen/vsubc-minmax-wasm-x2.c",
700 "src/f32-vbinary/gen/vsubc-minmax-wasm-x4.c",
701 "src/f32-vbinary/gen/vsubc-minmax-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700702 "src/f32-vbinary/gen/vsubc-relu-wasm-x1.c",
703 "src/f32-vbinary/gen/vsubc-relu-wasm-x2.c",
704 "src/f32-vbinary/gen/vsubc-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700705 "src/f32-vbinary/gen/vsubc-relu-wasm-x8.c",
Marat Dukhaned6baaf2020-12-01 15:07:08 -0800706 "src/f32-velu/gen/velu-wasm-rr2-lut16-p3-x1.c",
707 "src/f32-velu/gen/velu-wasm-rr2-lut16-p3-x2.c",
708 "src/f32-velu/gen/velu-wasm-rr2-lut16-p3-x3.c",
709 "src/f32-velu/gen/velu-wasm-rr2-lut16-p3-x4.c",
710 "src/f32-velu/gen/velu-wasm-rr2-lut16-p3-x5.c",
711 "src/f32-velu/gen/velu-wasm-rr2-lut16-p3-x6.c",
712 "src/f32-velu/gen/velu-wasm-rr2-p6-x1.c",
713 "src/f32-velu/gen/velu-wasm-rr2-p6-x2.c",
714 "src/f32-velu/gen/velu-wasm-rr2-p6-x3.c",
715 "src/f32-velu/gen/velu-wasm-rr2-p6-x4.c",
716 "src/f32-velu/gen/velu-wasm-rr2-p6-x5.c",
717 "src/f32-velu/gen/velu-wasm-rr2-p6-x6.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -0700718 "src/f32-vlrelu/gen/vlrelu-wasm-x1.c",
719 "src/f32-vlrelu/gen/vlrelu-wasm-x2.c",
720 "src/f32-vlrelu/gen/vlrelu-wasm-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700721 "src/f32-vmulcaddc/gen/c1-minmax-wasm-2x.c",
722 "src/f32-vmulcaddc/gen/c2-minmax-wasm-2x.c",
723 "src/f32-vmulcaddc/gen/c4-minmax-wasm-2x.c",
Marat Dukhan436ebe62019-12-04 15:10:12 -0800724]
725
Marat Dukhan290055c2020-06-09 12:24:29 -0700726WASMSIMD_UKERNELS = [
Marat Dukhan40f05522020-07-16 22:33:12 -0700727 "src/f32-argmaxpool/4x-wasmsimd-c4.c",
728 "src/f32-argmaxpool/9p8x-wasmsimd-c4.c",
729 "src/f32-argmaxpool/9x-wasmsimd-c4.c",
Marat Dukhan3b7432d2020-07-16 17:46:32 -0700730 "src/f32-avgpool/9p8x-minmax-wasmsimd-arm-c4.c",
731 "src/f32-avgpool/9p8x-minmax-wasmsimd-x86-c4.c",
732 "src/f32-avgpool/9x-minmax-wasmsimd-arm-c4.c",
733 "src/f32-avgpool/9x-minmax-wasmsimd-x86-c4.c",
Marat Dukhan3fa52c82020-07-08 12:54:33 -0700734 "src/f32-clamp/gen/wasmsimd-arm-x4.c",
735 "src/f32-clamp/gen/wasmsimd-arm-x8.c",
736 "src/f32-clamp/gen/wasmsimd-x86-x4.c",
737 "src/f32-clamp/gen/wasmsimd-x86-x8.c",
Frank Barchard22136062020-11-24 18:44:46 -0800738 "src/f32-conv-hwc2chw/3x3s2p1c3x4-wasmsimd-2x2.c",
Frank Barchard0725b8d2020-12-07 11:07:35 -0800739 "src/f32-dwconv/gen/up4x4-minmax-wasmsimd-arm-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700740 "src/f32-dwconv/gen/up4x4-minmax-wasmsimd-arm.c",
Frank Barchard0725b8d2020-12-07 11:07:35 -0800741 "src/f32-dwconv/gen/up4x4-minmax-wasmsimd-x86-acc2.c",
Marat Dukhanac014d72020-06-16 08:36:47 -0700742 "src/f32-dwconv/gen/up4x4-minmax-wasmsimd-x86.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700743 "src/f32-dwconv/gen/up4x4-wasmsimd.c",
Frank Barchard0725b8d2020-12-07 11:07:35 -0800744 "src/f32-dwconv/gen/up4x9-minmax-wasmsimd-arm-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700745 "src/f32-dwconv/gen/up4x9-minmax-wasmsimd-arm.c",
Frank Barchard0725b8d2020-12-07 11:07:35 -0800746 "src/f32-dwconv/gen/up4x9-minmax-wasmsimd-x86-acc2.c",
Marat Dukhanac014d72020-06-16 08:36:47 -0700747 "src/f32-dwconv/gen/up4x9-minmax-wasmsimd-x86.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700748 "src/f32-dwconv/gen/up4x9-wasmsimd.c",
Frank Barchard0725b8d2020-12-07 11:07:35 -0800749 "src/f32-dwconv/gen/up4x25-minmax-wasmsimd-arm-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700750 "src/f32-dwconv/gen/up4x25-minmax-wasmsimd-arm.c",
Frank Barchard0725b8d2020-12-07 11:07:35 -0800751 "src/f32-dwconv/gen/up4x25-minmax-wasmsimd-x86-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700752 "src/f32-dwconv/gen/up4x25-minmax-wasmsimd-x86.c",
753 "src/f32-dwconv/gen/up4x25-wasmsimd.c",
Frank Barchard0725b8d2020-12-07 11:07:35 -0800754 "src/f32-dwconv/gen/up8x4-minmax-wasmsimd-arm-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700755 "src/f32-dwconv/gen/up8x4-minmax-wasmsimd-arm.c",
Frank Barchard0725b8d2020-12-07 11:07:35 -0800756 "src/f32-dwconv/gen/up8x4-minmax-wasmsimd-x86-acc2.c",
Marat Dukhanac014d72020-06-16 08:36:47 -0700757 "src/f32-dwconv/gen/up8x4-minmax-wasmsimd-x86.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700758 "src/f32-dwconv/gen/up8x4-wasmsimd.c",
Frank Barchard0725b8d2020-12-07 11:07:35 -0800759 "src/f32-dwconv/gen/up8x9-minmax-wasmsimd-arm-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700760 "src/f32-dwconv/gen/up8x9-minmax-wasmsimd-arm.c",
Frank Barchard0725b8d2020-12-07 11:07:35 -0800761 "src/f32-dwconv/gen/up8x9-minmax-wasmsimd-x86-acc2.c",
Marat Dukhanac014d72020-06-16 08:36:47 -0700762 "src/f32-dwconv/gen/up8x9-minmax-wasmsimd-x86.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700763 "src/f32-dwconv/gen/up8x9-wasmsimd.c",
Frank Barchard0725b8d2020-12-07 11:07:35 -0800764 "src/f32-dwconv/gen/up8x25-minmax-wasmsimd-arm-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700765 "src/f32-dwconv/gen/up8x25-minmax-wasmsimd-arm.c",
Frank Barchard0725b8d2020-12-07 11:07:35 -0800766 "src/f32-dwconv/gen/up8x25-minmax-wasmsimd-x86-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700767 "src/f32-dwconv/gen/up8x25-minmax-wasmsimd-x86.c",
768 "src/f32-dwconv/gen/up8x25-wasmsimd.c",
Frank Barchard412e2f42020-12-11 11:40:50 -0800769 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-loadsplat-1x4-acc2.c",
770 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-loadsplat-1x4-acc3.c",
771 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-loadsplat-1x4-acc4.c",
772 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-loadsplat-1x4.c",
773 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-loadsplat-2x4-acc2.c",
774 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-loadsplat-2x4.c",
775 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-loadsplat-3x4.c",
776 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-loadsplat-4x4.c",
777 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-loadsplat-5x4.c",
778 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-loadsplat-6x4.c",
Frank Barchard02bb4292020-12-15 18:25:32 -0800779 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-splat-1x4-acc2.c",
780 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-splat-1x4-acc3.c",
781 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-splat-1x4-acc4.c",
782 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-splat-1x4.c",
783 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-splat-2x4-acc2.c",
784 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-splat-2x4.c",
785 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-splat-3x4.c",
786 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-splat-4x4.c",
787 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-splat-5x4.c",
788 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-splat-6x4.c",
Frank Barchard412e2f42020-12-11 11:40:50 -0800789 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-loadsplat-1x4-acc2.c",
790 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-loadsplat-1x4-acc3.c",
791 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-loadsplat-1x4-acc4.c",
792 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-loadsplat-1x4.c",
793 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-loadsplat-2x4-acc2.c",
794 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-loadsplat-2x4.c",
795 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-loadsplat-3x4.c",
796 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-loadsplat-4x4.c",
797 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-loadsplat-5x4.c",
798 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-loadsplat-6x4.c",
Frank Barchard02bb4292020-12-15 18:25:32 -0800799 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-splat-1x4-acc2.c",
800 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-splat-1x4-acc3.c",
801 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-splat-1x4-acc4.c",
802 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-splat-1x4.c",
803 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-splat-2x4-acc2.c",
804 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-splat-2x4.c",
805 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-splat-3x4.c",
806 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-splat-4x4.c",
807 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-splat-5x4.c",
808 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-splat-6x4.c",
Frank Barchardc5704bf2020-12-21 23:09:00 -0800809 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-arm-loadsplat-1x4-acc2.c",
810 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-arm-loadsplat-1x4-acc3.c",
811 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-arm-loadsplat-1x4-acc4.c",
812 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-arm-loadsplat-1x4.c",
813 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-arm-loadsplat-2x4-acc2.c",
814 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-arm-loadsplat-2x4.c",
815 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-arm-loadsplat-3x4.c",
816 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-arm-loadsplat-4x4.c",
Frank Barchard412e2f42020-12-11 11:40:50 -0800817 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-arm-splat-1x4-acc2.c",
818 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-arm-splat-1x4-acc3.c",
819 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-arm-splat-1x4-acc4.c",
820 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-arm-splat-1x4.c",
821 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-arm-splat-2x4-acc2.c",
822 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-arm-splat-2x4.c",
823 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-arm-splat-3x4.c",
824 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-arm-splat-4x4.c",
Frank Barchardcadd4222021-01-20 16:27:25 -0800825 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-x86-loadsplat-1x4-acc2.c",
826 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-x86-loadsplat-1x4-acc3.c",
827 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-x86-loadsplat-1x4-acc4.c",
828 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-x86-loadsplat-1x4.c",
829 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-x86-loadsplat-2x4-acc2.c",
830 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-x86-loadsplat-2x4.c",
831 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-x86-loadsplat-3x4.c",
832 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-x86-loadsplat-4x4.c",
Frank Barchard412e2f42020-12-11 11:40:50 -0800833 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-x86-splat-1x4-acc2.c",
834 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-x86-splat-1x4-acc3.c",
835 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-x86-splat-1x4-acc4.c",
836 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-x86-splat-1x4.c",
837 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-x86-splat-2x4-acc2.c",
838 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-x86-splat-2x4.c",
839 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-x86-splat-3x4.c",
840 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-x86-splat-4x4.c",
Frank Barchardb20dcd62020-12-15 16:46:14 -0800841 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-loadsplat-1x4-acc2.c",
842 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-loadsplat-1x4-acc3.c",
843 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-loadsplat-1x4-acc4.c",
844 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-loadsplat-1x4-acc5.c",
845 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-loadsplat-1x4.c",
846 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-loadsplat-2x4-acc2.c",
847 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-loadsplat-2x4-acc3.c",
848 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-loadsplat-2x4.c",
849 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-loadsplat-3x4-acc2.c",
850 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-loadsplat-3x4.c",
851 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-loadsplat-4x4-acc2.c",
852 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-loadsplat-4x4.c",
853 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-loadsplat-5x4.c",
Frank Barchard412e2f42020-12-11 11:40:50 -0800854 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-splat-1x4-acc2.c",
855 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-splat-1x4-acc3.c",
856 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-splat-1x4-acc4.c",
857 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-splat-1x4-acc5.c",
858 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-splat-1x4.c",
859 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-splat-2x4-acc2.c",
860 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-splat-2x4-acc3.c",
861 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-splat-2x4.c",
862 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-splat-3x4-acc2.c",
863 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-splat-3x4.c",
864 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-splat-4x4-acc2.c",
865 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-splat-4x4.c",
866 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-splat-5x4.c",
Frank Barchardb20dcd62020-12-15 16:46:14 -0800867 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-loadsplat-1x4-acc2.c",
868 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-loadsplat-1x4-acc3.c",
869 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-loadsplat-1x4-acc4.c",
870 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-loadsplat-1x4-acc5.c",
871 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-loadsplat-1x4.c",
872 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-loadsplat-2x4-acc2.c",
873 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-loadsplat-2x4-acc3.c",
874 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-loadsplat-2x4.c",
875 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-loadsplat-3x4-acc2.c",
876 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-loadsplat-3x4.c",
877 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-loadsplat-4x4-acc2.c",
878 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-loadsplat-4x4.c",
879 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-loadsplat-5x4.c",
Frank Barchard412e2f42020-12-11 11:40:50 -0800880 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-splat-1x4-acc2.c",
881 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-splat-1x4-acc3.c",
882 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-splat-1x4-acc4.c",
883 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-splat-1x4-acc5.c",
884 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-splat-1x4.c",
885 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-splat-2x4-acc2.c",
886 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-splat-2x4-acc3.c",
887 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-splat-2x4.c",
888 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-splat-3x4-acc2.c",
889 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-splat-3x4.c",
890 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-splat-4x4-acc2.c",
891 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-splat-4x4.c",
892 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-splat-5x4.c",
Frank Barchardc6889b32020-12-21 11:27:22 -0800893 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-arm-loadsplat-1x4-acc2.c",
894 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-arm-loadsplat-1x4-acc3.c",
895 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-arm-loadsplat-1x4-acc4.c",
896 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-arm-loadsplat-1x4-acc5.c",
897 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-arm-loadsplat-1x4.c",
898 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-arm-loadsplat-2x4-acc2.c",
899 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-arm-loadsplat-2x4-acc3.c",
900 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-arm-loadsplat-2x4.c",
901 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-arm-loadsplat-3x4-acc2.c",
902 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-arm-loadsplat-3x4.c",
Frank Barchard412e2f42020-12-11 11:40:50 -0800903 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-arm-splat-1x4-acc2.c",
904 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-arm-splat-1x4-acc3.c",
905 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-arm-splat-1x4-acc4.c",
906 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-arm-splat-1x4-acc5.c",
907 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-arm-splat-1x4.c",
908 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-arm-splat-2x4-acc2.c",
909 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-arm-splat-2x4-acc3.c",
910 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-arm-splat-2x4.c",
911 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-arm-splat-3x4-acc2.c",
912 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-arm-splat-3x4.c",
Frank Barchardc6889b32020-12-21 11:27:22 -0800913 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-x86-loadsplat-1x4-acc2.c",
914 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-x86-loadsplat-1x4-acc3.c",
915 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-x86-loadsplat-1x4-acc4.c",
916 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-x86-loadsplat-1x4-acc5.c",
917 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-x86-loadsplat-1x4.c",
918 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-x86-loadsplat-2x4-acc2.c",
919 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-x86-loadsplat-2x4-acc3.c",
920 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-x86-loadsplat-2x4.c",
921 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-x86-loadsplat-3x4-acc2.c",
922 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-x86-loadsplat-3x4.c",
Frank Barchard412e2f42020-12-11 11:40:50 -0800923 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-x86-splat-1x4-acc2.c",
924 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-x86-splat-1x4-acc3.c",
925 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-x86-splat-1x4-acc4.c",
926 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-x86-splat-1x4-acc5.c",
927 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-x86-splat-1x4.c",
928 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-x86-splat-2x4-acc2.c",
929 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-x86-splat-2x4-acc3.c",
930 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-x86-splat-2x4.c",
931 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-x86-splat-3x4-acc2.c",
932 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-x86-splat-3x4.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700933 "src/f32-gavgpool-cw/wasmsimd-arm-x4.c",
934 "src/f32-gavgpool-cw/wasmsimd-x86-x4.c",
Marat Dukhanc6016802020-07-16 18:51:28 -0700935 "src/f32-gavgpool/7p7x-minmax-wasmsimd-arm-c4.c",
936 "src/f32-gavgpool/7p7x-minmax-wasmsimd-x86-c4.c",
937 "src/f32-gavgpool/7x-minmax-wasmsimd-arm-c4.c",
938 "src/f32-gavgpool/7x-minmax-wasmsimd-x86-c4.c",
Frank Barchard0725b8d2020-12-07 11:07:35 -0800939 "src/f32-gemm/gen-inc/1x8inc-minmax-wasmsimd-arm-loadsplat.c",
940 "src/f32-gemm/gen-inc/1x8inc-minmax-wasmsimd-arm-splat.c",
941 "src/f32-gemm/gen-inc/1x8inc-minmax-wasmsimd-x86-loadsplat.c",
942 "src/f32-gemm/gen-inc/1x8inc-minmax-wasmsimd-x86-splat.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700943 "src/f32-gemm/gen-inc/1x8s4inc-minmax-wasmsimd-arm.c",
944 "src/f32-gemm/gen-inc/1x8s4inc-minmax-wasmsimd-x86.c",
Frank Barchard0725b8d2020-12-07 11:07:35 -0800945 "src/f32-gemm/gen-inc/3x8inc-minmax-wasmsimd-arm-loadsplat.c",
946 "src/f32-gemm/gen-inc/3x8inc-minmax-wasmsimd-arm-splat.c",
947 "src/f32-gemm/gen-inc/3x8inc-minmax-wasmsimd-x86-loadsplat.c",
948 "src/f32-gemm/gen-inc/3x8inc-minmax-wasmsimd-x86-splat.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700949 "src/f32-gemm/gen-inc/3x8s4inc-minmax-wasmsimd-arm.c",
950 "src/f32-gemm/gen-inc/3x8s4inc-minmax-wasmsimd-x86.c",
Frank Barchard0725b8d2020-12-07 11:07:35 -0800951 "src/f32-gemm/gen-inc/4x8inc-minmax-wasmsimd-arm-loadsplat.c",
952 "src/f32-gemm/gen-inc/4x8inc-minmax-wasmsimd-arm-splat.c",
953 "src/f32-gemm/gen-inc/4x8inc-minmax-wasmsimd-x86-loadsplat.c",
954 "src/f32-gemm/gen-inc/4x8inc-minmax-wasmsimd-x86-splat.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700955 "src/f32-gemm/gen-inc/4x8s4inc-minmax-wasmsimd-arm.c",
956 "src/f32-gemm/gen-inc/4x8s4inc-minmax-wasmsimd-x86.c",
Frank Barchard0725b8d2020-12-07 11:07:35 -0800957 "src/f32-gemm/gen-inc/5x8inc-minmax-wasmsimd-arm-loadsplat.c",
958 "src/f32-gemm/gen-inc/5x8inc-minmax-wasmsimd-arm-splat.c",
959 "src/f32-gemm/gen-inc/5x8inc-minmax-wasmsimd-x86-loadsplat.c",
960 "src/f32-gemm/gen-inc/5x8inc-minmax-wasmsimd-x86-splat.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700961 "src/f32-gemm/gen-inc/5x8s4inc-minmax-wasmsimd-arm.c",
962 "src/f32-gemm/gen-inc/5x8s4inc-minmax-wasmsimd-x86.c",
Frank Barchard0725b8d2020-12-07 11:07:35 -0800963 "src/f32-gemm/gen-inc/6x8inc-minmax-wasmsimd-arm-loadsplat.c",
964 "src/f32-gemm/gen-inc/6x8inc-minmax-wasmsimd-arm-splat.c",
965 "src/f32-gemm/gen-inc/6x8inc-minmax-wasmsimd-x86-loadsplat.c",
966 "src/f32-gemm/gen-inc/6x8inc-minmax-wasmsimd-x86-splat.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700967 "src/f32-gemm/gen-inc/6x8s4inc-minmax-wasmsimd-arm.c",
968 "src/f32-gemm/gen-inc/6x8s4inc-minmax-wasmsimd-x86.c",
Frank Barchard0725b8d2020-12-07 11:07:35 -0800969 "src/f32-gemm/gen/1x8-minmax-wasmsimd-arm-loadsplat.c",
970 "src/f32-gemm/gen/1x8-minmax-wasmsimd-arm-splat.c",
971 "src/f32-gemm/gen/1x8-minmax-wasmsimd-x86-loadsplat.c",
972 "src/f32-gemm/gen/1x8-minmax-wasmsimd-x86-splat.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700973 "src/f32-gemm/gen/1x8-relu-wasmsimd-splat.c",
974 "src/f32-gemm/gen/1x8-wasmsimd-splat.c",
975 "src/f32-gemm/gen/1x8s4-minmax-wasmsimd-arm.c",
976 "src/f32-gemm/gen/1x8s4-minmax-wasmsimd-x86.c",
Frank Barchard0725b8d2020-12-07 11:07:35 -0800977 "src/f32-gemm/gen/3x8-minmax-wasmsimd-arm-loadsplat.c",
978 "src/f32-gemm/gen/3x8-minmax-wasmsimd-arm-splat.c",
979 "src/f32-gemm/gen/3x8-minmax-wasmsimd-x86-loadsplat.c",
980 "src/f32-gemm/gen/3x8-minmax-wasmsimd-x86-splat.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700981 "src/f32-gemm/gen/3x8s4-minmax-wasmsimd-arm.c",
982 "src/f32-gemm/gen/3x8s4-minmax-wasmsimd-x86.c",
983 "src/f32-gemm/gen/4x2c4-minmax-wasmsimd-arm.c",
984 "src/f32-gemm/gen/4x2c4-minmax-wasmsimd-x86.c",
985 "src/f32-gemm/gen/4x2c4-relu-wasmsimd.c",
986 "src/f32-gemm/gen/4x2c4-wasmsimd.c",
Frank Barchard0725b8d2020-12-07 11:07:35 -0800987 "src/f32-gemm/gen/4x8-minmax-wasmsimd-arm-loadsplat.c",
988 "src/f32-gemm/gen/4x8-minmax-wasmsimd-arm-splat.c",
989 "src/f32-gemm/gen/4x8-minmax-wasmsimd-x86-loadsplat.c",
990 "src/f32-gemm/gen/4x8-minmax-wasmsimd-x86-splat.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700991 "src/f32-gemm/gen/4x8-relu-wasmsimd-splat.c",
992 "src/f32-gemm/gen/4x8-wasmsimd-splat.c",
993 "src/f32-gemm/gen/4x8s4-minmax-wasmsimd-arm.c",
994 "src/f32-gemm/gen/4x8s4-minmax-wasmsimd-x86.c",
Frank Barchard0725b8d2020-12-07 11:07:35 -0800995 "src/f32-gemm/gen/5x8-minmax-wasmsimd-arm-loadsplat.c",
996 "src/f32-gemm/gen/5x8-minmax-wasmsimd-arm-splat.c",
997 "src/f32-gemm/gen/5x8-minmax-wasmsimd-x86-loadsplat.c",
998 "src/f32-gemm/gen/5x8-minmax-wasmsimd-x86-splat.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700999 "src/f32-gemm/gen/5x8-relu-wasmsimd-splat.c",
1000 "src/f32-gemm/gen/5x8-wasmsimd-splat.c",
1001 "src/f32-gemm/gen/5x8s4-minmax-wasmsimd-arm.c",
1002 "src/f32-gemm/gen/5x8s4-minmax-wasmsimd-x86.c",
Frank Barchard0725b8d2020-12-07 11:07:35 -08001003 "src/f32-gemm/gen/6x8-minmax-wasmsimd-arm-loadsplat.c",
1004 "src/f32-gemm/gen/6x8-minmax-wasmsimd-arm-splat.c",
1005 "src/f32-gemm/gen/6x8-minmax-wasmsimd-x86-loadsplat.c",
1006 "src/f32-gemm/gen/6x8-minmax-wasmsimd-x86-splat.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001007 "src/f32-gemm/gen/6x8s4-minmax-wasmsimd-arm.c",
1008 "src/f32-gemm/gen/6x8s4-minmax-wasmsimd-x86.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07001009 "src/f32-hswish/gen/hswish-wasmsimd-x4.c",
1010 "src/f32-hswish/gen/hswish-wasmsimd-x8.c",
1011 "src/f32-hswish/gen/hswish-wasmsimd-x16.c",
XNNPACK Team965272b2020-10-23 21:10:15 -07001012 "src/f32-ibilinear-chw/gen/wasmsimd-p4.c",
1013 "src/f32-ibilinear-chw/gen/wasmsimd-p8.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001014 "src/f32-ibilinear/gen/wasmsimd-c4.c",
1015 "src/f32-ibilinear/gen/wasmsimd-c8.c",
Frank Barchard0725b8d2020-12-07 11:07:35 -08001016 "src/f32-igemm/gen/1x8-minmax-wasmsimd-arm-loadsplat.c",
1017 "src/f32-igemm/gen/1x8-minmax-wasmsimd-arm-splat.c",
1018 "src/f32-igemm/gen/1x8-minmax-wasmsimd-x86-loadsplat.c",
1019 "src/f32-igemm/gen/1x8-minmax-wasmsimd-x86-splat.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001020 "src/f32-igemm/gen/1x8-relu-wasmsimd-splat.c",
1021 "src/f32-igemm/gen/1x8-wasmsimd-splat.c",
1022 "src/f32-igemm/gen/1x8s4-minmax-wasmsimd-arm.c",
1023 "src/f32-igemm/gen/1x8s4-minmax-wasmsimd-x86.c",
Frank Barchard0725b8d2020-12-07 11:07:35 -08001024 "src/f32-igemm/gen/3x8-minmax-wasmsimd-arm-loadsplat.c",
1025 "src/f32-igemm/gen/3x8-minmax-wasmsimd-arm-splat.c",
1026 "src/f32-igemm/gen/3x8-minmax-wasmsimd-x86-loadsplat.c",
1027 "src/f32-igemm/gen/3x8-minmax-wasmsimd-x86-splat.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001028 "src/f32-igemm/gen/3x8s4-minmax-wasmsimd-arm.c",
1029 "src/f32-igemm/gen/3x8s4-minmax-wasmsimd-x86.c",
1030 "src/f32-igemm/gen/4x2c4-minmax-wasmsimd-arm.c",
1031 "src/f32-igemm/gen/4x2c4-minmax-wasmsimd-x86.c",
1032 "src/f32-igemm/gen/4x2c4-relu-wasmsimd.c",
1033 "src/f32-igemm/gen/4x2c4-wasmsimd.c",
Frank Barchard0725b8d2020-12-07 11:07:35 -08001034 "src/f32-igemm/gen/4x8-minmax-wasmsimd-arm-loadsplat.c",
1035 "src/f32-igemm/gen/4x8-minmax-wasmsimd-arm-splat.c",
1036 "src/f32-igemm/gen/4x8-minmax-wasmsimd-x86-loadsplat.c",
1037 "src/f32-igemm/gen/4x8-minmax-wasmsimd-x86-splat.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001038 "src/f32-igemm/gen/4x8-relu-wasmsimd-splat.c",
1039 "src/f32-igemm/gen/4x8-wasmsimd-splat.c",
1040 "src/f32-igemm/gen/4x8s4-minmax-wasmsimd-arm.c",
1041 "src/f32-igemm/gen/4x8s4-minmax-wasmsimd-x86.c",
Frank Barchard0725b8d2020-12-07 11:07:35 -08001042 "src/f32-igemm/gen/5x8-minmax-wasmsimd-arm-loadsplat.c",
1043 "src/f32-igemm/gen/5x8-minmax-wasmsimd-arm-splat.c",
1044 "src/f32-igemm/gen/5x8-minmax-wasmsimd-x86-loadsplat.c",
1045 "src/f32-igemm/gen/5x8-minmax-wasmsimd-x86-splat.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001046 "src/f32-igemm/gen/5x8-relu-wasmsimd-splat.c",
1047 "src/f32-igemm/gen/5x8-wasmsimd-splat.c",
1048 "src/f32-igemm/gen/5x8s4-minmax-wasmsimd-arm.c",
1049 "src/f32-igemm/gen/5x8s4-minmax-wasmsimd-x86.c",
Frank Barchard0725b8d2020-12-07 11:07:35 -08001050 "src/f32-igemm/gen/6x8-minmax-wasmsimd-arm-loadsplat.c",
1051 "src/f32-igemm/gen/6x8-minmax-wasmsimd-arm-splat.c",
1052 "src/f32-igemm/gen/6x8-minmax-wasmsimd-x86-loadsplat.c",
1053 "src/f32-igemm/gen/6x8-minmax-wasmsimd-x86-splat.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001054 "src/f32-igemm/gen/6x8s4-minmax-wasmsimd-arm.c",
1055 "src/f32-igemm/gen/6x8s4-minmax-wasmsimd-x86.c",
Marat Dukhanf6e24802020-07-08 22:20:40 -07001056 "src/f32-maxpool/9p8x-minmax-wasmsimd-arm-c4.c",
1057 "src/f32-maxpool/9p8x-minmax-wasmsimd-x86-c4.c",
Marat Dukhan1483c532020-07-16 18:08:19 -07001058 "src/f32-pavgpool/9p8x-minmax-wasmsimd-arm-c4.c",
1059 "src/f32-pavgpool/9p8x-minmax-wasmsimd-x86-c4.c",
1060 "src/f32-pavgpool/9x-minmax-wasmsimd-arm-c4.c",
1061 "src/f32-pavgpool/9x-minmax-wasmsimd-x86-c4.c",
Frank Barchard0725b8d2020-12-07 11:07:35 -08001062 "src/f32-ppmm/gen/4x8-minmax-wasmsimd-arm-splat.c",
1063 "src/f32-ppmm/gen/4x8-minmax-wasmsimd-x86-splat.c",
Frank Barcharda5316982020-07-23 13:19:28 -07001064 "src/f32-prelu/gen/wasmsimd-bitselect-1x4.c",
1065 "src/f32-prelu/gen/wasmsimd-bitselect-1x8.c",
1066 "src/f32-prelu/gen/wasmsimd-bitselect-1x16.c",
Marat Dukhan195f8eb2020-06-25 12:50:57 -07001067 "src/f32-prelu/gen/wasmsimd-bitselect-2x4.c",
1068 "src/f32-prelu/gen/wasmsimd-bitselect-2x8.c",
Frank Barcharda5316982020-07-23 13:19:28 -07001069 "src/f32-prelu/gen/wasmsimd-bitselect-2x16.c",
1070 "src/f32-prelu/gen/wasmsimd-bitselect-4x4.c",
1071 "src/f32-prelu/gen/wasmsimd-bitselect-4x8.c",
1072 "src/f32-prelu/gen/wasmsimd-bitselect-4x16.c",
1073 "src/f32-prelu/gen/wasmsimd-minmax-1x4.c",
1074 "src/f32-prelu/gen/wasmsimd-minmax-1x8.c",
1075 "src/f32-prelu/gen/wasmsimd-minmax-1x16.c",
Marat Dukhan195f8eb2020-06-25 12:50:57 -07001076 "src/f32-prelu/gen/wasmsimd-minmax-2x4.c",
1077 "src/f32-prelu/gen/wasmsimd-minmax-2x8.c",
Frank Barcharda5316982020-07-23 13:19:28 -07001078 "src/f32-prelu/gen/wasmsimd-minmax-2x16.c",
1079 "src/f32-prelu/gen/wasmsimd-minmax-4x4.c",
1080 "src/f32-prelu/gen/wasmsimd-minmax-4x8.c",
1081 "src/f32-prelu/gen/wasmsimd-minmax-4x16.c",
Marat Dukhan52238f02020-07-16 15:30:28 -07001082 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x4.c",
Marat Dukhan52238f02020-07-16 15:30:28 -07001083 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x8-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001084 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x8.c",
Marat Dukhan52238f02020-07-16 15:30:28 -07001085 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x12-acc2.c",
1086 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x12-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001087 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x12.c",
Marat Dukhan52238f02020-07-16 15:30:28 -07001088 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x16-acc2.c",
1089 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x16-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001090 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x16.c",
Marat Dukhan52238f02020-07-16 15:30:28 -07001091 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x20-acc2.c",
1092 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x20-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001093 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x20.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07001094 "src/f32-relu/gen/wasmsimd-x4.c",
1095 "src/f32-relu/gen/wasmsimd-x8.c",
Frank Barchard37297a62020-08-27 12:15:10 -07001096 "src/f32-relu/gen/wasmsimd-x16.c",
Marat Dukhan8c417962020-07-08 12:27:50 -07001097 "src/f32-rmax/wasmsimd-arm.c",
1098 "src/f32-rmax/wasmsimd-x86.c",
Marat Dukhand187a5b2020-07-20 01:07:17 -07001099 "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x4.c",
1100 "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x8.c",
1101 "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x12.c",
1102 "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x16.c",
1103 "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x20.c",
1104 "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x24.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001105 "src/f32-sigmoid/gen/wasmsimd-p5-div-x4.c",
1106 "src/f32-sigmoid/gen/wasmsimd-p5-div-x8.c",
1107 "src/f32-sigmoid/gen/wasmsimd-p5-div-x12.c",
1108 "src/f32-sigmoid/gen/wasmsimd-p5-div-x16.c",
1109 "src/f32-sigmoid/gen/wasmsimd-p5-div-x20.c",
1110 "src/f32-sigmoid/gen/wasmsimd-p5-div-x24.c",
Frank Barchard8ef44cd2020-11-03 12:30:23 -08001111 "src/f32-spmm/gen/4x1-minmax-wasmsimd-arm-pipelined-x2.c",
1112 "src/f32-spmm/gen/4x1-minmax-wasmsimd-arm-pipelined.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07001113 "src/f32-spmm/gen/4x1-minmax-wasmsimd-arm-x2.c",
1114 "src/f32-spmm/gen/4x1-minmax-wasmsimd-arm-x4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001115 "src/f32-spmm/gen/4x1-minmax-wasmsimd-arm.c",
Frank Barchard8ef44cd2020-11-03 12:30:23 -08001116 "src/f32-spmm/gen/4x1-minmax-wasmsimd-x86-pipelined-x2.c",
1117 "src/f32-spmm/gen/4x1-minmax-wasmsimd-x86-pipelined.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07001118 "src/f32-spmm/gen/4x1-minmax-wasmsimd-x86-x2.c",
1119 "src/f32-spmm/gen/4x1-minmax-wasmsimd-x86-x4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001120 "src/f32-spmm/gen/4x1-minmax-wasmsimd-x86.c",
Frank Barchard8ef44cd2020-11-03 12:30:23 -08001121 "src/f32-spmm/gen/8x1-minmax-wasmsimd-arm-pipelined-x2.c",
1122 "src/f32-spmm/gen/8x1-minmax-wasmsimd-arm-pipelined.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07001123 "src/f32-spmm/gen/8x1-minmax-wasmsimd-arm-x2.c",
1124 "src/f32-spmm/gen/8x1-minmax-wasmsimd-arm-x4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001125 "src/f32-spmm/gen/8x1-minmax-wasmsimd-arm.c",
Frank Barchard8ef44cd2020-11-03 12:30:23 -08001126 "src/f32-spmm/gen/8x1-minmax-wasmsimd-x86-pipelined-x2.c",
1127 "src/f32-spmm/gen/8x1-minmax-wasmsimd-x86-pipelined.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07001128 "src/f32-spmm/gen/8x1-minmax-wasmsimd-x86-x2.c",
1129 "src/f32-spmm/gen/8x1-minmax-wasmsimd-x86-x4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001130 "src/f32-spmm/gen/8x1-minmax-wasmsimd-x86.c",
Frank Barchard8ef44cd2020-11-03 12:30:23 -08001131 "src/f32-spmm/gen/16x1-minmax-wasmsimd-arm-pipelined-x2.c",
1132 "src/f32-spmm/gen/16x1-minmax-wasmsimd-arm-pipelined.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07001133 "src/f32-spmm/gen/16x1-minmax-wasmsimd-arm-x2.c",
1134 "src/f32-spmm/gen/16x1-minmax-wasmsimd-arm-x4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001135 "src/f32-spmm/gen/16x1-minmax-wasmsimd-arm.c",
Frank Barchard8ef44cd2020-11-03 12:30:23 -08001136 "src/f32-spmm/gen/16x1-minmax-wasmsimd-x86-pipelined-x2.c",
1137 "src/f32-spmm/gen/16x1-minmax-wasmsimd-x86-pipelined.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07001138 "src/f32-spmm/gen/16x1-minmax-wasmsimd-x86-x2.c",
1139 "src/f32-spmm/gen/16x1-minmax-wasmsimd-x86-x4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001140 "src/f32-spmm/gen/16x1-minmax-wasmsimd-x86.c",
Frank Barchard8ef44cd2020-11-03 12:30:23 -08001141 "src/f32-spmm/gen/32x1-minmax-wasmsimd-arm-pipelined-x2.c",
1142 "src/f32-spmm/gen/32x1-minmax-wasmsimd-arm-pipelined.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07001143 "src/f32-spmm/gen/32x1-minmax-wasmsimd-arm-x2.c",
1144 "src/f32-spmm/gen/32x1-minmax-wasmsimd-arm-x4.c",
Frank Barchard846c0c62020-10-26 15:01:39 -07001145 "src/f32-spmm/gen/32x1-minmax-wasmsimd-arm.c",
Frank Barchard8ef44cd2020-11-03 12:30:23 -08001146 "src/f32-spmm/gen/32x1-minmax-wasmsimd-x86-pipelined-x2.c",
1147 "src/f32-spmm/gen/32x1-minmax-wasmsimd-x86-pipelined.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07001148 "src/f32-spmm/gen/32x1-minmax-wasmsimd-x86-x2.c",
1149 "src/f32-spmm/gen/32x1-minmax-wasmsimd-x86-x4.c",
Frank Barchard846c0c62020-10-26 15:01:39 -07001150 "src/f32-spmm/gen/32x1-minmax-wasmsimd-x86.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -07001151 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-arm-x4.c",
1152 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001153 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-arm-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -07001154 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-x86-x4.c",
1155 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001156 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -07001157 "src/f32-vbinary/gen/vadd-relu-wasmsimd-x4.c",
1158 "src/f32-vbinary/gen/vadd-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001159 "src/f32-vbinary/gen/vadd-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07001160 "src/f32-vbinary/gen/vadd-wasmsimd-x4.c",
1161 "src/f32-vbinary/gen/vadd-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001162 "src/f32-vbinary/gen/vadd-wasmsimd-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001163 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-arm-x4.c",
1164 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001165 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-arm-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001166 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-x86-x4.c",
1167 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001168 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -07001169 "src/f32-vbinary/gen/vaddc-relu-wasmsimd-x4.c",
1170 "src/f32-vbinary/gen/vaddc-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001171 "src/f32-vbinary/gen/vaddc-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07001172 "src/f32-vbinary/gen/vaddc-wasmsimd-x4.c",
1173 "src/f32-vbinary/gen/vaddc-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001174 "src/f32-vbinary/gen/vaddc-wasmsimd-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -07001175 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-arm-x4.c",
1176 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001177 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-arm-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -07001178 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-x86-x4.c",
1179 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001180 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -07001181 "src/f32-vbinary/gen/vdiv-relu-wasmsimd-x4.c",
1182 "src/f32-vbinary/gen/vdiv-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001183 "src/f32-vbinary/gen/vdiv-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07001184 "src/f32-vbinary/gen/vdiv-wasmsimd-x4.c",
1185 "src/f32-vbinary/gen/vdiv-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001186 "src/f32-vbinary/gen/vdiv-wasmsimd-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001187 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-arm-x4.c",
1188 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001189 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-arm-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001190 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-x86-x4.c",
1191 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001192 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -07001193 "src/f32-vbinary/gen/vdivc-relu-wasmsimd-x4.c",
1194 "src/f32-vbinary/gen/vdivc-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001195 "src/f32-vbinary/gen/vdivc-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07001196 "src/f32-vbinary/gen/vdivc-wasmsimd-x4.c",
1197 "src/f32-vbinary/gen/vdivc-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001198 "src/f32-vbinary/gen/vdivc-wasmsimd-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -07001199 "src/f32-vbinary/gen/vmax-wasmsimd-arm-x4.c",
1200 "src/f32-vbinary/gen/vmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001201 "src/f32-vbinary/gen/vmax-wasmsimd-arm-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -07001202 "src/f32-vbinary/gen/vmax-wasmsimd-x86-x4.c",
1203 "src/f32-vbinary/gen/vmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001204 "src/f32-vbinary/gen/vmax-wasmsimd-x86-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001205 "src/f32-vbinary/gen/vmaxc-wasmsimd-arm-x4.c",
1206 "src/f32-vbinary/gen/vmaxc-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001207 "src/f32-vbinary/gen/vmaxc-wasmsimd-arm-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001208 "src/f32-vbinary/gen/vmaxc-wasmsimd-x86-x4.c",
1209 "src/f32-vbinary/gen/vmaxc-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001210 "src/f32-vbinary/gen/vmaxc-wasmsimd-x86-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -07001211 "src/f32-vbinary/gen/vmin-wasmsimd-arm-x4.c",
1212 "src/f32-vbinary/gen/vmin-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001213 "src/f32-vbinary/gen/vmin-wasmsimd-arm-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -07001214 "src/f32-vbinary/gen/vmin-wasmsimd-x86-x4.c",
1215 "src/f32-vbinary/gen/vmin-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001216 "src/f32-vbinary/gen/vmin-wasmsimd-x86-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001217 "src/f32-vbinary/gen/vminc-wasmsimd-arm-x4.c",
1218 "src/f32-vbinary/gen/vminc-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001219 "src/f32-vbinary/gen/vminc-wasmsimd-arm-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001220 "src/f32-vbinary/gen/vminc-wasmsimd-x86-x4.c",
1221 "src/f32-vbinary/gen/vminc-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001222 "src/f32-vbinary/gen/vminc-wasmsimd-x86-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -07001223 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-arm-x4.c",
1224 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001225 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-arm-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -07001226 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-x86-x4.c",
1227 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001228 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -07001229 "src/f32-vbinary/gen/vmul-relu-wasmsimd-x4.c",
1230 "src/f32-vbinary/gen/vmul-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001231 "src/f32-vbinary/gen/vmul-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07001232 "src/f32-vbinary/gen/vmul-wasmsimd-x4.c",
1233 "src/f32-vbinary/gen/vmul-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001234 "src/f32-vbinary/gen/vmul-wasmsimd-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001235 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-arm-x4.c",
1236 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001237 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-arm-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001238 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-x86-x4.c",
1239 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001240 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -07001241 "src/f32-vbinary/gen/vmulc-relu-wasmsimd-x4.c",
1242 "src/f32-vbinary/gen/vmulc-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001243 "src/f32-vbinary/gen/vmulc-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07001244 "src/f32-vbinary/gen/vmulc-wasmsimd-x4.c",
1245 "src/f32-vbinary/gen/vmulc-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001246 "src/f32-vbinary/gen/vmulc-wasmsimd-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001247 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-arm-x4.c",
1248 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001249 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-arm-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001250 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-x86-x4.c",
1251 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001252 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -07001253 "src/f32-vbinary/gen/vrdivc-relu-wasmsimd-x4.c",
1254 "src/f32-vbinary/gen/vrdivc-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001255 "src/f32-vbinary/gen/vrdivc-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07001256 "src/f32-vbinary/gen/vrdivc-wasmsimd-x4.c",
1257 "src/f32-vbinary/gen/vrdivc-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001258 "src/f32-vbinary/gen/vrdivc-wasmsimd-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001259 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-arm-x4.c",
1260 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001261 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-arm-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001262 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-x86-x4.c",
1263 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001264 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -07001265 "src/f32-vbinary/gen/vrsubc-relu-wasmsimd-x4.c",
1266 "src/f32-vbinary/gen/vrsubc-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001267 "src/f32-vbinary/gen/vrsubc-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07001268 "src/f32-vbinary/gen/vrsubc-wasmsimd-x4.c",
1269 "src/f32-vbinary/gen/vrsubc-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001270 "src/f32-vbinary/gen/vrsubc-wasmsimd-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -07001271 "src/f32-vbinary/gen/vsqrdiff-wasmsimd-x4.c",
1272 "src/f32-vbinary/gen/vsqrdiff-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001273 "src/f32-vbinary/gen/vsqrdiff-wasmsimd-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001274 "src/f32-vbinary/gen/vsqrdiffc-wasmsimd-x4.c",
1275 "src/f32-vbinary/gen/vsqrdiffc-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001276 "src/f32-vbinary/gen/vsqrdiffc-wasmsimd-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -07001277 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-arm-x4.c",
1278 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001279 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-arm-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -07001280 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-x86-x4.c",
1281 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001282 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -07001283 "src/f32-vbinary/gen/vsub-relu-wasmsimd-x4.c",
1284 "src/f32-vbinary/gen/vsub-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001285 "src/f32-vbinary/gen/vsub-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07001286 "src/f32-vbinary/gen/vsub-wasmsimd-x4.c",
1287 "src/f32-vbinary/gen/vsub-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001288 "src/f32-vbinary/gen/vsub-wasmsimd-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001289 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-arm-x4.c",
1290 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001291 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-arm-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001292 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-x86-x4.c",
1293 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001294 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -07001295 "src/f32-vbinary/gen/vsubc-relu-wasmsimd-x4.c",
1296 "src/f32-vbinary/gen/vsubc-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001297 "src/f32-vbinary/gen/vsubc-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07001298 "src/f32-vbinary/gen/vsubc-wasmsimd-x4.c",
1299 "src/f32-vbinary/gen/vsubc-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001300 "src/f32-vbinary/gen/vsubc-wasmsimd-x16.c",
Marat Dukhaned6baaf2020-12-01 15:07:08 -08001301 "src/f32-velu/gen/velu-wasmsimd-arm-rr2-lut16-p3-x4.c",
1302 "src/f32-velu/gen/velu-wasmsimd-arm-rr2-lut16-p3-x8.c",
1303 "src/f32-velu/gen/velu-wasmsimd-arm-rr2-lut16-p3-x12.c",
1304 "src/f32-velu/gen/velu-wasmsimd-arm-rr2-lut16-p3-x16.c",
1305 "src/f32-velu/gen/velu-wasmsimd-arm-rr2-lut16-p3-x20.c",
1306 "src/f32-velu/gen/velu-wasmsimd-arm-rr2-lut16-p3-x24.c",
Marat Dukhaned6baaf2020-12-01 15:07:08 -08001307 "src/f32-velu/gen/velu-wasmsimd-arm-rr2-p6-x4.c",
1308 "src/f32-velu/gen/velu-wasmsimd-arm-rr2-p6-x8.c",
1309 "src/f32-velu/gen/velu-wasmsimd-arm-rr2-p6-x12.c",
1310 "src/f32-velu/gen/velu-wasmsimd-arm-rr2-p6-x16.c",
1311 "src/f32-velu/gen/velu-wasmsimd-arm-rr2-p6-x20.c",
1312 "src/f32-velu/gen/velu-wasmsimd-arm-rr2-p6-x24.c",
Frank Barcharde7223ee2020-12-04 19:04:01 -08001313 "src/f32-velu/gen/velu-wasmsimd-x86-rr2-lut16-p3-x4.c",
1314 "src/f32-velu/gen/velu-wasmsimd-x86-rr2-lut16-p3-x8.c",
1315 "src/f32-velu/gen/velu-wasmsimd-x86-rr2-lut16-p3-x12.c",
1316 "src/f32-velu/gen/velu-wasmsimd-x86-rr2-lut16-p3-x16.c",
1317 "src/f32-velu/gen/velu-wasmsimd-x86-rr2-lut16-p3-x20.c",
1318 "src/f32-velu/gen/velu-wasmsimd-x86-rr2-lut16-p3-x24.c",
Marat Dukhaned6baaf2020-12-01 15:07:08 -08001319 "src/f32-velu/gen/velu-wasmsimd-x86-rr2-p6-x4.c",
1320 "src/f32-velu/gen/velu-wasmsimd-x86-rr2-p6-x8.c",
1321 "src/f32-velu/gen/velu-wasmsimd-x86-rr2-p6-x12.c",
1322 "src/f32-velu/gen/velu-wasmsimd-x86-rr2-p6-x16.c",
1323 "src/f32-velu/gen/velu-wasmsimd-x86-rr2-p6-x20.c",
1324 "src/f32-velu/gen/velu-wasmsimd-x86-rr2-p6-x24.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07001325 "src/f32-vlrelu/gen/vlrelu-wasmsimd-bitselect-x4.c",
1326 "src/f32-vlrelu/gen/vlrelu-wasmsimd-bitselect-x8.c",
1327 "src/f32-vlrelu/gen/vlrelu-wasmsimd-minmax-x4.c",
1328 "src/f32-vlrelu/gen/vlrelu-wasmsimd-minmax-x8.c",
Marat Dukhand816f622020-07-15 10:14:39 -07001329 "src/f32-vmulcaddc/gen/c4-minmax-wasmsimd-arm-2x.c",
Marat Dukhand816f622020-07-15 10:14:39 -07001330 "src/f32-vmulcaddc/gen/c4-minmax-wasmsimd-x86-2x.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001331 "src/f32-vmulcaddc/gen/c8-minmax-wasmsimd-arm-2x.c",
Marat Dukhand816f622020-07-15 10:14:39 -07001332 "src/f32-vmulcaddc/gen/c8-minmax-wasmsimd-x86-2x.c",
Marat Dukhanb82b2cd2020-07-16 02:23:42 -07001333 "src/f32-vrnd/gen/vrndd-wasmsimd-addsub-x4.c",
1334 "src/f32-vrnd/gen/vrndd-wasmsimd-addsub-x8.c",
1335 "src/f32-vrnd/gen/vrndd-wasmsimd-cvt-x4.c",
1336 "src/f32-vrnd/gen/vrndd-wasmsimd-cvt-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001337 "src/f32-vrnd/gen/vrndne-wasmsimd-addsub-x4.c",
1338 "src/f32-vrnd/gen/vrndne-wasmsimd-addsub-x8.c",
1339 "src/f32-vrnd/gen/vrndu-wasmsimd-addsub-x4.c",
1340 "src/f32-vrnd/gen/vrndu-wasmsimd-addsub-x8.c",
1341 "src/f32-vrnd/gen/vrndu-wasmsimd-cvt-x4.c",
1342 "src/f32-vrnd/gen/vrndu-wasmsimd-cvt-x8.c",
1343 "src/f32-vrnd/gen/vrndz-wasmsimd-addsub-x4.c",
1344 "src/f32-vrnd/gen/vrndz-wasmsimd-addsub-x8.c",
1345 "src/f32-vrnd/gen/vrndz-wasmsimd-cvt-x4.c",
1346 "src/f32-vrnd/gen/vrndz-wasmsimd-cvt-x8.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07001347 "src/f32-vsqrt/gen/wasmsimd-sqrt-x4.c",
1348 "src/f32-vsqrt/gen/wasmsimd-sqrt-x8.c",
Marat Dukhan37c83512020-06-29 13:25:53 -07001349 "src/f32-vunary/gen/vabs-wasmsimd-x4.c",
1350 "src/f32-vunary/gen/vabs-wasmsimd-x8.c",
1351 "src/f32-vunary/gen/vneg-wasmsimd-x4.c",
1352 "src/f32-vunary/gen/vneg-wasmsimd-x8.c",
1353 "src/f32-vunary/gen/vsqr-wasmsimd-x4.c",
1354 "src/f32-vunary/gen/vsqr-wasmsimd-x8.c",
Marat Dukhande390d42020-11-29 19:32:18 -08001355 "src/math/expm1minus-wasmsimd-rr2-lut16-p3-andnot.c",
1356 "src/math/expm1minus-wasmsimd-rr2-lut16-p3-max.c",
1357 "src/math/expm1minus-wasmsimd-rr2-p6-andnot.c",
1358 "src/math/expm1minus-wasmsimd-rr2-p6-max.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001359 "src/math/roundd-wasmsimd-addsub.c",
1360 "src/math/roundd-wasmsimd-cvt.c",
1361 "src/math/roundne-wasmsimd-addsub.c",
1362 "src/math/roundu-wasmsimd-addsub.c",
1363 "src/math/roundu-wasmsimd-cvt.c",
1364 "src/math/roundz-wasmsimd-addsub.c",
1365 "src/math/roundz-wasmsimd-cvt.c",
1366 "src/math/sigmoid-wasmsimd-rr2-lut64-p2-div.c",
1367 "src/math/sigmoid-wasmsimd-rr2-p5-div.c",
Marat Dukhancc8f34c2020-08-05 16:36:38 -07001368 "src/qs8-dwconv/gen/up8x9-minmax-wasmsimd-mul16.c",
1369 "src/qs8-dwconv/gen/up16x9-minmax-wasmsimd-mul16.c",
1370 "src/qs8-dwconv/gen/up24x9-minmax-wasmsimd-mul16.c",
Marat Dukhanb5e3d172020-08-06 13:29:53 -07001371 "src/qs8-gavgpool/gen/7p7x-minmax-wasmsimd-c8-acc2.c",
1372 "src/qs8-gavgpool/gen/7p7x-minmax-wasmsimd-c16-acc2.c",
1373 "src/qs8-gavgpool/gen/7p7x-minmax-wasmsimd-c24-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001374 "src/qs8-gavgpool/gen/7x-minmax-wasmsimd-c8-acc2.c",
1375 "src/qs8-gavgpool/gen/7x-minmax-wasmsimd-c16-acc2.c",
1376 "src/qs8-gavgpool/gen/7x-minmax-wasmsimd-c24-acc2.c",
Marat Dukhan27203da2020-08-05 15:19:03 -07001377 "src/qs8-gemm/gen/1x4c8-minmax-wasmsimd-ld64.c",
Marat Dukhan27203da2020-08-05 15:19:03 -07001378 "src/qs8-gemm/gen/1x4c8-minmax-wasmsimd-ld128.c",
Marat Dukhan27203da2020-08-05 15:19:03 -07001379 "src/qs8-gemm/gen/1x4c8-xw-minmax-wasmsimd.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001380 "src/qs8-gemm/gen/2x4c8-minmax-wasmsimd-ld64.c",
1381 "src/qs8-gemm/gen/2x4c8-minmax-wasmsimd-ld128.c",
Marat Dukhan27203da2020-08-05 15:19:03 -07001382 "src/qs8-gemm/gen/2x4c8-xw-minmax-wasmsimd.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001383 "src/qs8-gemm/gen/3x4c8-minmax-wasmsimd-ld64.c",
1384 "src/qs8-gemm/gen/3x4c8-minmax-wasmsimd-ld128.c",
Marat Dukhan27203da2020-08-05 15:19:03 -07001385 "src/qs8-gemm/gen/3x4c8-xw-minmax-wasmsimd.c",
1386 "src/qs8-igemm/gen/1x4c8-minmax-wasmsimd-ld64.c",
Marat Dukhan27203da2020-08-05 15:19:03 -07001387 "src/qs8-igemm/gen/1x4c8-minmax-wasmsimd-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001388 "src/qs8-igemm/gen/2x4c8-minmax-wasmsimd-ld64.c",
Marat Dukhan27203da2020-08-05 15:19:03 -07001389 "src/qs8-igemm/gen/2x4c8-minmax-wasmsimd-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001390 "src/qs8-igemm/gen/3x4c8-minmax-wasmsimd-ld64.c",
Marat Dukhan27203da2020-08-05 15:19:03 -07001391 "src/qs8-igemm/gen/3x4c8-minmax-wasmsimd-ld128.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -07001392 "src/qs8-requantization/fp32-wasmsimd.c",
Marat Dukhan138560c2020-08-03 18:57:34 -07001393 "src/qs8-requantization/q31-wasmsimd.c",
Marat Dukhan5df27f82020-09-02 23:59:21 -07001394 "src/qs8-vadd/gen/minmax-wasmsimd-x8.c",
1395 "src/qs8-vadd/gen/minmax-wasmsimd-x16.c",
1396 "src/qs8-vadd/gen/minmax-wasmsimd-x24.c",
1397 "src/qs8-vadd/gen/minmax-wasmsimd-x32.c",
1398 "src/qs8-vaddc/gen/minmax-wasmsimd-x8.c",
1399 "src/qs8-vaddc/gen/minmax-wasmsimd-x16.c",
1400 "src/qs8-vaddc/gen/minmax-wasmsimd-x24.c",
1401 "src/qs8-vaddc/gen/minmax-wasmsimd-x32.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07001402 "src/qu8-requantization/fp32-wasmsimd.c",
Marat Dukhan22de5e72020-08-03 21:53:18 -07001403 "src/qu8-requantization/q31-wasmsimd.c",
Marat Dukhan8ee37012020-07-16 13:17:13 -07001404 "src/x32-fill/wasmsimd.c",
Marat Dukhan66d99e92020-07-16 12:56:21 -07001405 "src/x32-packx/x4-wasmsimd.c",
Marat Dukhan9306ae02020-07-16 15:51:13 -07001406 "src/x32-pad/wasmsimd.c",
Marat Dukhan9d4bfa22020-07-16 19:07:04 -07001407 "src/x32-unpool/wasmsimd.c",
Marat Dukhane3b78762020-07-16 20:02:58 -07001408 "src/x32-zip/x2-wasmsimd.c",
1409 "src/x32-zip/x3-wasmsimd.c",
1410 "src/x32-zip/x4-wasmsimd.c",
1411 "src/x32-zip/xm-wasmsimd.c",
Marat Dukhan290055c2020-06-09 12:24:29 -07001412]
1413
Marat Dukhan08c4a432019-10-03 09:29:21 -07001414# ISA-specific micro-kernels
1415NEON_UKERNELS = [
Marat Dukhanef25c6d2020-07-24 00:59:40 -07001416 "src/f32-argmaxpool/4x-neon-c4.c",
1417 "src/f32-argmaxpool/9p8x-neon-c4.c",
1418 "src/f32-argmaxpool/9x-neon-c4.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001419 "src/f32-avgpool/9p8x-minmax-neon-c4.c",
1420 "src/f32-avgpool/9x-minmax-neon-c4.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001421 "src/f32-clamp/gen/neon-x4.c",
1422 "src/f32-clamp/gen/neon-x8.c",
Marat Dukhan56b10cd2020-05-18 09:35:49 -07001423 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neon-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001424 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neon-2x2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001425 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neon-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001426 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neon-2x2.c",
Marat Dukhan56b10cd2020-05-18 09:35:49 -07001427 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neon-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001428 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neon-2x2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001429 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neon-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001430 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neon-2x2.c",
Marat Dukhanc7634882020-12-07 15:11:12 -08001431 "src/f32-conv-hwc2chw/3x3s2p1c3x4-neon-2x2.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -07001432 "src/f32-dwconv/gen/up4x4-minmax-neon-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001433 "src/f32-dwconv/gen/up4x4-minmax-neon.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001434 "src/f32-dwconv/gen/up4x9-minmax-neon-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001435 "src/f32-dwconv/gen/up4x9-minmax-neon.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -07001436 "src/f32-dwconv/gen/up4x25-minmax-neon-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001437 "src/f32-dwconv/gen/up4x25-minmax-neon.c",
1438 "src/f32-dwconv/gen/up8x4-minmax-neon-acc2.c",
1439 "src/f32-dwconv/gen/up8x4-minmax-neon.c",
1440 "src/f32-dwconv/gen/up8x9-minmax-neon-acc2.c",
1441 "src/f32-dwconv/gen/up8x9-minmax-neon.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -07001442 "src/f32-dwconv/gen/up8x25-minmax-neon-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001443 "src/f32-dwconv/gen/up8x25-minmax-neon.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001444 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-1x4-acc2.c",
1445 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-1x4-acc3.c",
1446 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-1x4-acc4.c",
Marat Dukhanc581e482020-10-24 01:28:11 -07001447 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-1x4.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001448 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-2x4-acc2.c",
Marat Dukhanc581e482020-10-24 01:28:11 -07001449 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-2x4.c",
1450 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-3x4.c",
1451 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-4x4.c",
1452 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-5x4.c",
1453 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-6x4.c",
Marat Dukhan82f0c322020-10-25 19:17:35 -07001454 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neon-1x4-acc2.c",
1455 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neon-1x4-acc3.c",
1456 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neon-1x4-acc4.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001457 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neon-1x4.c",
Marat Dukhan82f0c322020-10-25 19:17:35 -07001458 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neon-2x4-acc2.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001459 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neon-2x4.c",
1460 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neon-3x4.c",
1461 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neon-4x4.c",
Marat Dukhan149f0ea2020-10-26 12:50:33 -07001462 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-1x4-acc2.c",
1463 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-1x4-acc3.c",
1464 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-1x4-acc4.c",
1465 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-1x4-acc5.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001466 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-1x4.c",
Marat Dukhan149f0ea2020-10-26 12:50:33 -07001467 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-2x4-acc2.c",
1468 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-2x4-acc3.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001469 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-2x4.c",
Marat Dukhan149f0ea2020-10-26 12:50:33 -07001470 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-3x4-acc2.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001471 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-3x4.c",
Marat Dukhan149f0ea2020-10-26 12:50:33 -07001472 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-4x4-acc2.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001473 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-4x4.c",
1474 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-5x4.c",
Marat Dukhan30d4b252020-10-29 16:33:22 -07001475 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neon-1x4-acc2.c",
1476 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neon-1x4-acc3.c",
1477 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neon-1x4-acc4.c",
1478 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neon-1x4-acc5.c",
1479 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neon-1x4.c",
1480 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neon-2x4-acc2.c",
1481 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neon-2x4-acc3.c",
1482 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neon-2x4.c",
Marat Dukhan30d4b252020-10-29 16:33:22 -07001483 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neon-3x4-acc2.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07001484 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neon-3x4.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -07001485 "src/f32-gavgpool-cw/neon-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001486 "src/f32-gavgpool/7p7x-minmax-neon-c4.c",
1487 "src/f32-gavgpool/7x-minmax-neon-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001488 "src/f32-gemm/gen-inc/1x8inc-minmax-neon-dup-ld64.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001489 "src/f32-gemm/gen-inc/1x8inc-minmax-neon-lane-ld64.c",
1490 "src/f32-gemm/gen-inc/1x8s4inc-minmax-neon.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001491 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-dup-ld64.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001492 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-dup-ld128.c",
1493 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-lane-ld64.c",
1494 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-lane-ld128.c",
1495 "src/f32-gemm/gen-inc/4x8s4inc-minmax-neon.c",
1496 "src/f32-gemm/gen-inc/5x8inc-minmax-neon-lane-ld64.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001497 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-dup-ld64.c",
1498 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-dup-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001499 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-lane-ld64.c",
1500 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-lane-ld128.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001501 "src/f32-gemm/gen-inc/6x8s4inc-minmax-neon.c",
1502 "src/f32-gemm/gen-inc/8x8s4inc-minmax-neon.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001503 "src/f32-gemm/gen/1x8-minmax-neon-dup-ld64.c",
1504 "src/f32-gemm/gen/1x8-minmax-neon-lane-ld64.c",
1505 "src/f32-gemm/gen/1x8s4-minmax-neon.c",
1506 "src/f32-gemm/gen/4x2-minmax-neon-lane-ld64.c",
1507 "src/f32-gemm/gen/4x8-minmax-neon-dup-ld64.c",
1508 "src/f32-gemm/gen/4x8-minmax-neon-dup-ld128.c",
1509 "src/f32-gemm/gen/4x8-minmax-neon-lane-ld64.c",
1510 "src/f32-gemm/gen/4x8-minmax-neon-lane-ld128.c",
1511 "src/f32-gemm/gen/4x8s4-minmax-neon.c",
1512 "src/f32-gemm/gen/5x8-minmax-neon-lane-ld64.c",
1513 "src/f32-gemm/gen/6x8-minmax-neon-dup-ld64.c",
1514 "src/f32-gemm/gen/6x8-minmax-neon-dup-ld128.c",
1515 "src/f32-gemm/gen/6x8-minmax-neon-lane-ld64.c",
1516 "src/f32-gemm/gen/6x8-minmax-neon-lane-ld128.c",
1517 "src/f32-gemm/gen/6x8s4-minmax-neon.c",
1518 "src/f32-gemm/gen/8x8s4-minmax-neon.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07001519 "src/f32-hswish/gen/hswish-neon-x4.c",
1520 "src/f32-hswish/gen/hswish-neon-x8.c",
1521 "src/f32-hswish/gen/hswish-neon-x16.c",
Artsiom Ablavatski2202c812021-01-22 14:16:43 -08001522 "src/f32-ibilinear-chw/gen/neon-p4.c",
1523 "src/f32-ibilinear-chw/gen/neon-p8.c",
Frank Barchard8247e212021-02-03 18:12:33 -08001524 "src/f32-ibilinear/gen/neon-c4.c",
1525 "src/f32-ibilinear/gen/neon-c8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001526 "src/f32-igemm/gen/1x8-minmax-neon-dup-ld64.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001527 "src/f32-igemm/gen/1x8-minmax-neon-lane-ld64.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001528 "src/f32-igemm/gen/1x8s4-minmax-neon.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001529 "src/f32-igemm/gen/4x2-minmax-neon-lane-ld64.c",
1530 "src/f32-igemm/gen/4x4-minmax-neon-lane-ld64.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001531 "src/f32-igemm/gen/4x8-minmax-neon-dup-ld64.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001532 "src/f32-igemm/gen/4x8-minmax-neon-dup-ld128.c",
1533 "src/f32-igemm/gen/4x8-minmax-neon-lane-ld64.c",
1534 "src/f32-igemm/gen/4x8-minmax-neon-lane-ld128.c",
1535 "src/f32-igemm/gen/4x8s4-minmax-neon.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001536 "src/f32-igemm/gen/6x8-minmax-neon-dup-ld64.c",
1537 "src/f32-igemm/gen/6x8-minmax-neon-dup-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001538 "src/f32-igemm/gen/6x8-minmax-neon-lane-ld64.c",
1539 "src/f32-igemm/gen/6x8-minmax-neon-lane-ld128.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001540 "src/f32-igemm/gen/6x8s4-minmax-neon.c",
1541 "src/f32-igemm/gen/8x8s4-minmax-neon.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001542 "src/f32-maxpool/9p8x-minmax-neon-c4.c",
1543 "src/f32-pavgpool/9p8x-minmax-neon-c4.c",
1544 "src/f32-pavgpool/9x-minmax-neon-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001545 "src/f32-ppmm/gen/4x8-minmax-neon.c",
1546 "src/f32-ppmm/gen/8x8-minmax-neon.c",
Frank Barcharda5316982020-07-23 13:19:28 -07001547 "src/f32-prelu/gen/neon-1x4.c",
1548 "src/f32-prelu/gen/neon-1x8.c",
1549 "src/f32-prelu/gen/neon-1x16.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001550 "src/f32-prelu/gen/neon-2x4.c",
1551 "src/f32-prelu/gen/neon-2x8.c",
Frank Barcharda5316982020-07-23 13:19:28 -07001552 "src/f32-prelu/gen/neon-2x16.c",
1553 "src/f32-prelu/gen/neon-4x4.c",
1554 "src/f32-prelu/gen/neon-4x8.c",
1555 "src/f32-prelu/gen/neon-4x16.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001556 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x4.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001557 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x8-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001558 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x8.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001559 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12-acc2.c",
1560 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001561 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001562 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16-acc2.c",
1563 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001564 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001565 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20-acc2.c",
1566 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001567 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20.c",
1568 "src/f32-raddstoreexpminusmax/gen/neon-p5-x4.c",
1569 "src/f32-raddstoreexpminusmax/gen/neon-p5-x8-acc2.c",
1570 "src/f32-raddstoreexpminusmax/gen/neon-p5-x8.c",
1571 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12-acc2.c",
1572 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12-acc3.c",
1573 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12.c",
1574 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16-acc2.c",
1575 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16-acc4.c",
1576 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16.c",
1577 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20-acc2.c",
1578 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20-acc5.c",
1579 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07001580 "src/f32-relu/gen/neon-x4.c",
1581 "src/f32-relu/gen/neon-x8.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001582 "src/f32-rmax/neon.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -08001583 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x4.c",
1584 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x8.c",
1585 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x12.c",
1586 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x16.c",
1587 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x20.c",
1588 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x24.c",
1589 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x4.c",
1590 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x8.c",
1591 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x12.c",
1592 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x16.c",
1593 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x20.c",
1594 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x24.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001595 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x4.c",
1596 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x8.c",
1597 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x12.c",
1598 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x16.c",
1599 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x20.c",
1600 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x24.c",
Marat Dukhan5b86c432020-12-06 19:15:03 -08001601 "src/f32-spmm/gen/4x1-minmax-neon-pipelined.c",
1602 "src/f32-spmm/gen/4x1-minmax-neon-x2.c",
1603 "src/f32-spmm/gen/4x1-minmax-neon.c",
1604 "src/f32-spmm/gen/8x1-minmax-neon-pipelined.c",
1605 "src/f32-spmm/gen/8x1-minmax-neon-x2.c",
1606 "src/f32-spmm/gen/8x1-minmax-neon.c",
1607 "src/f32-spmm/gen/12x1-minmax-neon.c",
1608 "src/f32-spmm/gen/16x1-minmax-neon-pipelined.c",
1609 "src/f32-spmm/gen/16x1-minmax-neon-x2.c",
1610 "src/f32-spmm/gen/16x1-minmax-neon.c",
1611 "src/f32-spmm/gen/32x1-minmax-neon-pipelined.c",
1612 "src/f32-spmm/gen/32x1-minmax-neon-x2.c",
1613 "src/f32-spmm/gen/32x1-minmax-neon.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001614 "src/f32-vbinary/gen/vadd-minmax-neon-x4.c",
1615 "src/f32-vbinary/gen/vadd-minmax-neon-x8.c",
1616 "src/f32-vbinary/gen/vaddc-minmax-neon-x4.c",
1617 "src/f32-vbinary/gen/vaddc-minmax-neon-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -08001618 "src/f32-vbinary/gen/vmax-neon-x4.c",
1619 "src/f32-vbinary/gen/vmax-neon-x8.c",
1620 "src/f32-vbinary/gen/vmaxc-neon-x4.c",
1621 "src/f32-vbinary/gen/vmaxc-neon-x8.c",
1622 "src/f32-vbinary/gen/vmin-neon-x4.c",
1623 "src/f32-vbinary/gen/vmin-neon-x8.c",
1624 "src/f32-vbinary/gen/vminc-neon-x4.c",
1625 "src/f32-vbinary/gen/vminc-neon-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001626 "src/f32-vbinary/gen/vmul-minmax-neon-x4.c",
1627 "src/f32-vbinary/gen/vmul-minmax-neon-x8.c",
1628 "src/f32-vbinary/gen/vmulc-minmax-neon-x4.c",
1629 "src/f32-vbinary/gen/vmulc-minmax-neon-x8.c",
1630 "src/f32-vbinary/gen/vrsubc-minmax-neon-x4.c",
1631 "src/f32-vbinary/gen/vrsubc-minmax-neon-x8.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07001632 "src/f32-vbinary/gen/vsqrdiff-neon-x4.c",
1633 "src/f32-vbinary/gen/vsqrdiff-neon-x8.c",
1634 "src/f32-vbinary/gen/vsqrdiffc-neon-x4.c",
1635 "src/f32-vbinary/gen/vsqrdiffc-neon-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001636 "src/f32-vbinary/gen/vsub-minmax-neon-x4.c",
1637 "src/f32-vbinary/gen/vsub-minmax-neon-x8.c",
1638 "src/f32-vbinary/gen/vsubc-minmax-neon-x4.c",
1639 "src/f32-vbinary/gen/vsubc-minmax-neon-x8.c",
Marat Dukhaned6baaf2020-12-01 15:07:08 -08001640 "src/f32-velu/gen/velu-neon-rr2-lut16-p3-x4.c",
1641 "src/f32-velu/gen/velu-neon-rr2-lut16-p3-x8.c",
1642 "src/f32-velu/gen/velu-neon-rr2-lut16-p3-x12.c",
1643 "src/f32-velu/gen/velu-neon-rr2-lut16-p3-x16.c",
1644 "src/f32-velu/gen/velu-neon-rr2-lut16-p3-x20.c",
1645 "src/f32-velu/gen/velu-neon-rr2-lut16-p3-x24.c",
1646 "src/f32-velu/gen/velu-neon-rr2-p6-x4.c",
1647 "src/f32-velu/gen/velu-neon-rr2-p6-x8.c",
1648 "src/f32-velu/gen/velu-neon-rr2-p6-x12.c",
1649 "src/f32-velu/gen/velu-neon-rr2-p6-x16.c",
1650 "src/f32-velu/gen/velu-neon-rr2-p6-x20.c",
1651 "src/f32-velu/gen/velu-neon-rr2-p6-x24.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07001652 "src/f32-vlrelu/gen/vlrelu-neon-x4.c",
1653 "src/f32-vlrelu/gen/vlrelu-neon-x8.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001654 "src/f32-vmulcaddc/gen/c4-minmax-neon-2x.c",
1655 "src/f32-vmulcaddc/gen/c8-minmax-neon-2x.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07001656 "src/f32-vrnd/gen/vrndd-neon-x4.c",
1657 "src/f32-vrnd/gen/vrndd-neon-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001658 "src/f32-vrnd/gen/vrndne-neon-x4.c",
1659 "src/f32-vrnd/gen/vrndne-neon-x8.c",
1660 "src/f32-vrnd/gen/vrndu-neon-x4.c",
1661 "src/f32-vrnd/gen/vrndu-neon-x8.c",
1662 "src/f32-vrnd/gen/vrndz-neon-x4.c",
1663 "src/f32-vrnd/gen/vrndz-neon-x8.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07001664 "src/f32-vunary/gen/vabs-neon-x4.c",
1665 "src/f32-vunary/gen/vabs-neon-x8.c",
1666 "src/f32-vunary/gen/vneg-neon-x4.c",
1667 "src/f32-vunary/gen/vneg-neon-x8.c",
1668 "src/f32-vunary/gen/vsqr-neon-x4.c",
1669 "src/f32-vunary/gen/vsqr-neon-x8.c",
Marat Dukhande390d42020-11-29 19:32:18 -08001670 "src/math/expm1minus-neon-rr2-lut16-p3.c",
1671 "src/math/expm1minus-neon-rr2-p6.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001672 "src/math/roundd-neon-addsub.c",
1673 "src/math/roundd-neon-cvt.c",
1674 "src/math/roundne-neon-addsub.c",
1675 "src/math/roundu-neon-addsub.c",
1676 "src/math/roundu-neon-cvt.c",
1677 "src/math/roundz-neon-addsub.c",
1678 "src/math/roundz-neon-cvt.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001679 "src/math/sigmoid-neon-rr2-lut64-p2-nr2recps.c",
1680 "src/math/sigmoid-neon-rr2-lut2048-p1-nr2recps.c",
1681 "src/math/sigmoid-neon-rr2-p5-nr2recps.c",
1682 "src/math/sqrt-neon-nr1rsqrts.c",
1683 "src/math/sqrt-neon-nr2rsqrts.c",
1684 "src/math/sqrt-neon-nr3rsqrts.c",
Marat Dukhan023bcf92020-08-10 12:40:50 -07001685 "src/qs8-dwconv/gen/up8x9-minmax-neon-mul16.c",
1686 "src/qs8-dwconv/gen/up16x9-minmax-neon-mul16.c",
1687 "src/qs8-dwconv/gen/up24x9-minmax-neon-mul16.c",
1688 "src/qs8-dwconv/gen/up32x9-minmax-neon-mul16.c",
Marat Dukhan281262d2020-08-10 13:23:21 -07001689 "src/qs8-gavgpool/gen/7p7x-minmax-neon-c8-acc2.c",
1690 "src/qs8-gavgpool/gen/7p7x-minmax-neon-c16-acc2.c",
1691 "src/qs8-gavgpool/gen/7p7x-minmax-neon-c24-acc2.c",
1692 "src/qs8-gavgpool/gen/7p7x-minmax-neon-c32-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001693 "src/qs8-gavgpool/gen/7x-minmax-neon-c8-acc2.c",
1694 "src/qs8-gavgpool/gen/7x-minmax-neon-c16-acc2.c",
1695 "src/qs8-gavgpool/gen/7x-minmax-neon-c24-acc2.c",
1696 "src/qs8-gavgpool/gen/7x-minmax-neon-c32-acc2.c",
Frank Barchardec0bf142021-01-15 10:14:05 -08001697 "src/qs8-gemm/gen/1x8-minmax-neon-mlal-lane.c",
Frank Barchardec0bf142021-01-15 10:14:05 -08001698 "src/qs8-gemm/gen/1x8-minmax-neon-mull-addw-dup.c",
Frank Barchard8247e212021-02-03 18:12:33 -08001699 "src/qs8-gemm/gen/1x8c2-minmax-neon-mlal-padal-dup.c",
Frank Barchard2302ffd2021-01-22 14:08:02 -08001700 "src/qs8-gemm/gen/1x8c2-minmax-neon-mull-padal-dup.c",
Frank Barchardda78da12021-03-02 14:28:00 -08001701 "src/qs8-gemm/gen/1x8c8-minmax-neon-mlal-padal.c",
Frank Barchard8247e212021-02-03 18:12:33 -08001702 "src/qs8-gemm/gen/1x8c8-minmax-neon-mull-padal.c",
Frank Barchardcadd4222021-01-20 16:27:25 -08001703 "src/qs8-gemm/gen/1x16-minmax-neon-mlal-lane.c",
1704 "src/qs8-gemm/gen/1x16-minmax-neon-mull-addw-dup.c",
Frank Barchard8247e212021-02-03 18:12:33 -08001705 "src/qs8-gemm/gen/1x16c2-minmax-neon-mlal-padal-dup.c",
Frank Barchard2302ffd2021-01-22 14:08:02 -08001706 "src/qs8-gemm/gen/1x16c2-minmax-neon-mull-padal-dup.c",
Frank Barchardda78da12021-03-02 14:28:00 -08001707 "src/qs8-gemm/gen/1x16c8-minmax-neon-mlal-padal.c",
Frank Barchard8247e212021-02-03 18:12:33 -08001708 "src/qs8-gemm/gen/1x16c8-minmax-neon-mull-padal.c",
Frank Barchardcadd4222021-01-20 16:27:25 -08001709 "src/qs8-gemm/gen/2x8-minmax-neon-mlal-lane.c",
Frank Barchardec0bf142021-01-15 10:14:05 -08001710 "src/qs8-gemm/gen/2x8-minmax-neon-mull-addw-dup.c",
Frank Barchard8247e212021-02-03 18:12:33 -08001711 "src/qs8-gemm/gen/2x8c2-minmax-neon-mlal-padal-dup.c",
Frank Barchard2302ffd2021-01-22 14:08:02 -08001712 "src/qs8-gemm/gen/2x8c2-minmax-neon-mull-padal-dup.c",
Frank Barchardda78da12021-03-02 14:28:00 -08001713 "src/qs8-gemm/gen/2x8c8-minmax-neon-mlal-padal.c",
Frank Barchard8247e212021-02-03 18:12:33 -08001714 "src/qs8-gemm/gen/2x8c8-minmax-neon-mull-padal.c",
Frank Barchardcadd4222021-01-20 16:27:25 -08001715 "src/qs8-gemm/gen/2x16-minmax-neon-mlal-lane.c",
1716 "src/qs8-gemm/gen/2x16-minmax-neon-mull-addw-dup.c",
Frank Barchard8247e212021-02-03 18:12:33 -08001717 "src/qs8-gemm/gen/2x16c2-minmax-neon-mlal-padal-dup.c",
Frank Barchard2302ffd2021-01-22 14:08:02 -08001718 "src/qs8-gemm/gen/2x16c2-minmax-neon-mull-padal-dup.c",
Frank Barchardda78da12021-03-02 14:28:00 -08001719 "src/qs8-gemm/gen/2x16c8-minmax-neon-mlal-padal.c",
Frank Barchard8247e212021-02-03 18:12:33 -08001720 "src/qs8-gemm/gen/2x16c8-minmax-neon-mull-padal.c",
Frank Barchard2302ffd2021-01-22 14:08:02 -08001721 "src/qs8-gemm/gen/3x8-minmax-neon-mlal-lane.c",
1722 "src/qs8-gemm/gen/3x8-minmax-neon-mull-addw-dup.c",
Frank Barchard8247e212021-02-03 18:12:33 -08001723 "src/qs8-gemm/gen/3x8c2-minmax-neon-mlal-padal-dup.c",
Frank Barchard2302ffd2021-01-22 14:08:02 -08001724 "src/qs8-gemm/gen/3x8c2-minmax-neon-mull-padal-dup.c",
Frank Barchardda78da12021-03-02 14:28:00 -08001725 "src/qs8-gemm/gen/3x8c8-minmax-neon-mlal-padal.c",
Frank Barchard8247e212021-02-03 18:12:33 -08001726 "src/qs8-gemm/gen/3x8c8-minmax-neon-mull-padal.c",
1727 "src/qs8-gemm/gen/3x16-minmax-neon-mlal-lane.c",
1728 "src/qs8-gemm/gen/3x16-minmax-neon-mull-addw-dup.c",
1729 "src/qs8-gemm/gen/3x16c2-minmax-neon-mlal-padal-dup.c",
1730 "src/qs8-gemm/gen/3x16c2-minmax-neon-mull-padal-dup.c",
Frank Barchardda78da12021-03-02 14:28:00 -08001731 "src/qs8-gemm/gen/3x16c8-minmax-neon-mlal-padal.c",
Frank Barchard8247e212021-02-03 18:12:33 -08001732 "src/qs8-gemm/gen/3x16c8-minmax-neon-mull-padal.c",
Frank Barchardcadd4222021-01-20 16:27:25 -08001733 "src/qs8-gemm/gen/4x8-minmax-neon-mlal-lane.c",
Frank Barchardec0bf142021-01-15 10:14:05 -08001734 "src/qs8-gemm/gen/4x8-minmax-neon-mull-addw-dup.c",
Frank Barchard8247e212021-02-03 18:12:33 -08001735 "src/qs8-gemm/gen/4x8c2-minmax-neon-mlal-padal-dup.c",
Frank Barchard2302ffd2021-01-22 14:08:02 -08001736 "src/qs8-gemm/gen/4x8c2-minmax-neon-mull-padal-dup.c",
Frank Barchardda78da12021-03-02 14:28:00 -08001737 "src/qs8-gemm/gen/4x8c8-minmax-neon-mlal-padal.c",
Frank Barchard8247e212021-02-03 18:12:33 -08001738 "src/qs8-gemm/gen/4x8c8-minmax-neon-mull-padal.c",
Frank Barchardcadd4222021-01-20 16:27:25 -08001739 "src/qs8-gemm/gen/4x16-minmax-neon-mlal-lane.c",
1740 "src/qs8-gemm/gen/4x16-minmax-neon-mull-addw-dup.c",
Frank Barchard8247e212021-02-03 18:12:33 -08001741 "src/qs8-gemm/gen/4x16c2-minmax-neon-mlal-padal-dup.c",
Frank Barchard2302ffd2021-01-22 14:08:02 -08001742 "src/qs8-gemm/gen/4x16c2-minmax-neon-mull-padal-dup.c",
Frank Barchardda78da12021-03-02 14:28:00 -08001743 "src/qs8-gemm/gen/4x16c8-minmax-neon-mlal-padal.c",
Frank Barchard58990122021-01-29 16:09:56 -08001744 "src/qs8-gemm/gen/4x16c8-minmax-neon-mull-padal.c",
Frank Barchardda78da12021-03-02 14:28:00 -08001745 "src/qs8-gemm/gen/1x8c16-minmax-neon-mlal-padal.c",
1746 "src/qs8-gemm/gen/2x8c16-minmax-neon-mlal-padal.c",
1747 "src/qs8-gemm/gen/3x8c16-minmax-neon-mlal-padal.c",
1748 "src/qs8-gemm/gen/4x8c16-minmax-neon-mlal-padal.c",
1749 "src/qs8-gemm/gen/1x16c16-minmax-neon-mlal-padal.c",
1750 "src/qs8-gemm/gen/2x16c16-minmax-neon-mlal-padal.c",
1751 "src/qs8-gemm/gen/3x16c16-minmax-neon-mlal-padal.c",
Frank Barchard71c4d1a2021-02-19 10:07:23 -08001752 "src/qs8-gemm/gen/4x16c16-minmax-neon-mlal-padal.c",
Frank Barchardec0bf142021-01-15 10:14:05 -08001753 "src/qs8-igemm/gen/1x8-minmax-neon-mlal-lane.c",
Frank Barchardd76a37b2021-03-02 11:36:16 -08001754 "src/qs8-igemm/gen/1x16-minmax-neon-mlal-lane.c",
Frank Barchardd76a37b2021-03-02 11:36:16 -08001755 "src/qs8-igemm/gen/2x8-minmax-neon-mlal-lane.c",
Frank Barchardd76a37b2021-03-02 11:36:16 -08001756 "src/qs8-igemm/gen/2x16-minmax-neon-mlal-lane.c",
Frank Barchardd76a37b2021-03-02 11:36:16 -08001757 "src/qs8-igemm/gen/3x8-minmax-neon-mlal-lane.c",
Frank Barchardd76a37b2021-03-02 11:36:16 -08001758 "src/qs8-igemm/gen/3x16-minmax-neon-mlal-lane.c",
Frank Barchardd76a37b2021-03-02 11:36:16 -08001759 "src/qs8-igemm/gen/4x8-minmax-neon-mlal-lane.c",
Frank Barchardd76a37b2021-03-02 11:36:16 -08001760 "src/qs8-igemm/gen/4x16-minmax-neon-mlal-lane.c",
Frank Barchardda78da12021-03-02 14:28:00 -08001761 "src/qs8-igemm/gen/1x16c8-minmax-neon-mlal-padal.c",
1762 "src/qs8-igemm/gen/1x16c8-minmax-neon-mull-padal.c",
1763 "src/qs8-igemm/gen/1x8c8-minmax-neon-mlal-padal.c",
1764 "src/qs8-igemm/gen/1x8c8-minmax-neon-mull-padal.c",
1765 "src/qs8-igemm/gen/2x16c8-minmax-neon-mlal-padal.c",
1766 "src/qs8-igemm/gen/2x16c8-minmax-neon-mull-padal.c",
1767 "src/qs8-igemm/gen/2x8c8-minmax-neon-mlal-padal.c",
1768 "src/qs8-igemm/gen/2x8c8-minmax-neon-mull-padal.c",
1769 "src/qs8-igemm/gen/3x16c8-minmax-neon-mlal-padal.c",
1770 "src/qs8-igemm/gen/3x16c8-minmax-neon-mull-padal.c",
1771 "src/qs8-igemm/gen/3x8c8-minmax-neon-mlal-padal.c",
1772 "src/qs8-igemm/gen/3x8c8-minmax-neon-mull-padal.c",
1773 "src/qs8-igemm/gen/4x16c8-minmax-neon-mlal-padal.c",
Frank Barchardd76a37b2021-03-02 11:36:16 -08001774 "src/qs8-igemm/gen/4x16c8-minmax-neon-mull-padal.c",
Frank Barchardda78da12021-03-02 14:28:00 -08001775 "src/qs8-igemm/gen/4x8c8-minmax-neon-mlal-padal.c",
1776 "src/qs8-igemm/gen/4x8c8-minmax-neon-mull-padal.c",
1777 "src/qs8-igemm/gen/1x16c16-minmax-neon-mlal-padal.c",
1778 "src/qs8-igemm/gen/1x8c16-minmax-neon-mlal-padal.c",
1779 "src/qs8-igemm/gen/2x16c16-minmax-neon-mlal-padal.c",
1780 "src/qs8-igemm/gen/2x8c16-minmax-neon-mlal-padal.c",
1781 "src/qs8-igemm/gen/3x16c16-minmax-neon-mlal-padal.c",
1782 "src/qs8-igemm/gen/3x8c16-minmax-neon-mlal-padal.c",
Frank Barchardd76a37b2021-03-02 11:36:16 -08001783 "src/qs8-igemm/gen/4x16c16-minmax-neon-mlal-padal.c",
Frank Barchardda78da12021-03-02 14:28:00 -08001784 "src/qs8-igemm/gen/4x8c16-minmax-neon-mlal-padal.c",
1785 "src/qs8-igemm/gen/1x16c2-minmax-neon-mlal-padal-dup.c",
1786 "src/qs8-igemm/gen/1x8c2-minmax-neon-mlal-padal-dup.c",
1787 "src/qs8-igemm/gen/2x16c2-minmax-neon-mlal-padal-dup.c",
1788 "src/qs8-igemm/gen/2x8c2-minmax-neon-mlal-padal-dup.c",
1789 "src/qs8-igemm/gen/3x16c2-minmax-neon-mlal-padal-dup.c",
1790 "src/qs8-igemm/gen/3x8c2-minmax-neon-mlal-padal-dup.c",
1791 "src/qs8-igemm/gen/4x16c2-minmax-neon-mlal-padal-dup.c",
1792 "src/qs8-igemm/gen/4x8c2-minmax-neon-mlal-padal-dup.c",
1793 "src/qs8-igemm/gen/1x16c2-minmax-neon-mull-padal-dup.c",
1794 "src/qs8-igemm/gen/1x8c2-minmax-neon-mull-padal-dup.c",
1795 "src/qs8-igemm/gen/2x16c2-minmax-neon-mull-padal-dup.c",
1796 "src/qs8-igemm/gen/2x8c2-minmax-neon-mull-padal-dup.c",
1797 "src/qs8-igemm/gen/3x16c2-minmax-neon-mull-padal-dup.c",
1798 "src/qs8-igemm/gen/3x8c2-minmax-neon-mull-padal-dup.c",
1799 "src/qs8-igemm/gen/4x16c2-minmax-neon-mull-padal-dup.c",
1800 "src/qs8-igemm/gen/4x8c2-minmax-neon-mull-padal-dup.c",
1801 "src/qs8-igemm/gen/1x16-minmax-neon-mull-addw-dup.c",
1802 "src/qs8-igemm/gen/1x8-minmax-neon-mull-addw-dup.c",
1803 "src/qs8-igemm/gen/2x16-minmax-neon-mull-addw-dup.c",
1804 "src/qs8-igemm/gen/2x8-minmax-neon-mull-addw-dup.c",
1805 "src/qs8-igemm/gen/3x16-minmax-neon-mull-addw-dup.c",
1806 "src/qs8-igemm/gen/3x8-minmax-neon-mull-addw-dup.c",
1807 "src/qs8-igemm/gen/4x16-minmax-neon-mull-addw-dup.c",
1808 "src/qs8-igemm/gen/4x8-minmax-neon-mull-addw-dup.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -07001809 "src/qs8-requantization/fp32-neon.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001810 "src/qs8-requantization/precise-neon.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -07001811 "src/qs8-requantization/q31-neon.c",
Marat Dukhanba7b2792020-09-02 14:26:45 -07001812 "src/qs8-vadd/gen/minmax-neon-ld64-x8.c",
1813 "src/qs8-vadd/gen/minmax-neon-ld64-x16.c",
1814 "src/qs8-vadd/gen/minmax-neon-ld64-x24.c",
1815 "src/qs8-vadd/gen/minmax-neon-ld64-x32.c",
1816 "src/qs8-vaddc/gen/minmax-neon-ld64-x8.c",
1817 "src/qs8-vaddc/gen/minmax-neon-ld64-x16.c",
1818 "src/qs8-vaddc/gen/minmax-neon-ld64-x24.c",
1819 "src/qs8-vaddc/gen/minmax-neon-ld64-x32.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -07001820 "src/qu8-avgpool/9p8x-minmax-neon-c8.c",
1821 "src/qu8-avgpool/9x-minmax-neon-c8.c",
1822 "src/qu8-dwconv/up8x9-minmax-neon.c",
1823 "src/qu8-gavgpool/7p7x-minmax-neon-c8.c",
1824 "src/qu8-gavgpool/7x-minmax-neon-c8.c",
1825 "src/qu8-gemm/4x8-minmax-neon.c",
1826 "src/qu8-gemm/8x8-minmax-neon.c",
1827 "src/qu8-igemm/4x8-minmax-neon.c",
1828 "src/qu8-igemm/8x8-minmax-neon.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07001829 "src/qu8-requantization/fp32-neon.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001830 "src/qu8-requantization/precise-neon.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07001831 "src/qu8-requantization/q31-neon.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -07001832 "src/qu8-vadd/minmax-neon.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001833 "src/u8-clamp/neon-x64.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001834 "src/u8-maxpool/9p8x-minmax-neon-c16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001835 "src/u8-rmax/neon.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001836 "src/x8-zip/x2-neon.c",
1837 "src/x8-zip/x3-neon.c",
1838 "src/x8-zip/x4-neon.c",
1839 "src/x8-zip/xm-neon.c",
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -07001840 "src/x32-fill/neon.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001841 "src/x32-packx/x4-neon-st4.c",
Marat Dukhan63523d42020-05-22 17:07:33 -07001842 "src/x32-pad/neon.c",
Marat Dukhan57dccd82020-04-14 00:53:10 -07001843 "src/x32-unpool/neon.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001844 "src/x32-zip/x2-neon.c",
1845 "src/x32-zip/x3-neon.c",
1846 "src/x32-zip/x4-neon.c",
1847 "src/x32-zip/xm-neon.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001848]
1849
1850NEONFMA_UKERNELS = [
Frank Barchard04336c12020-10-22 16:48:55 -07001851 "src/f32-dwconv/gen/up4x4-minmax-neonfma-acc2.c",
1852 "src/f32-dwconv/gen/up4x4-minmax-neonfma.c",
1853 "src/f32-dwconv/gen/up4x9-minmax-neonfma-acc2.c",
1854 "src/f32-dwconv/gen/up4x9-minmax-neonfma.c",
1855 "src/f32-dwconv/gen/up4x25-minmax-neonfma-acc2.c",
1856 "src/f32-dwconv/gen/up4x25-minmax-neonfma.c",
1857 "src/f32-dwconv/gen/up8x4-minmax-neonfma-acc2.c",
1858 "src/f32-dwconv/gen/up8x4-minmax-neonfma.c",
1859 "src/f32-dwconv/gen/up8x9-minmax-neonfma-acc2.c",
1860 "src/f32-dwconv/gen/up8x9-minmax-neonfma.c",
1861 "src/f32-dwconv/gen/up8x25-minmax-neonfma-acc2.c",
1862 "src/f32-dwconv/gen/up8x25-minmax-neonfma.c",
1863 "src/f32-gemm/gen-inc/1x8inc-minmax-neonfma-dup-ld64.c",
1864 "src/f32-gemm/gen-inc/1x8s4inc-minmax-neonfma.c",
1865 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-dup-ld64.c",
1866 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-dup-ld128.c",
1867 "src/f32-gemm/gen-inc/4x8s4inc-minmax-neonfma.c",
1868 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-dup-ld64.c",
1869 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-dup-ld128.c",
1870 "src/f32-gemm/gen-inc/6x8s4inc-minmax-neonfma.c",
1871 "src/f32-gemm/gen-inc/8x8s4inc-minmax-neonfma.c",
1872 "src/f32-gemm/gen/1x8-minmax-neonfma-dup-ld64.c",
1873 "src/f32-gemm/gen/1x8s4-minmax-neonfma.c",
1874 "src/f32-gemm/gen/4x8-minmax-neonfma-dup-ld64.c",
1875 "src/f32-gemm/gen/4x8-minmax-neonfma-dup-ld128.c",
1876 "src/f32-gemm/gen/4x8s4-minmax-neonfma.c",
1877 "src/f32-gemm/gen/6x8-minmax-neonfma-dup-ld64.c",
1878 "src/f32-gemm/gen/6x8-minmax-neonfma-dup-ld128.c",
1879 "src/f32-gemm/gen/6x8s4-minmax-neonfma.c",
1880 "src/f32-gemm/gen/8x8s4-minmax-neonfma.c",
Artsiom Ablavatski2202c812021-01-22 14:16:43 -08001881 "src/f32-ibilinear-chw/gen/neonfma-p4.c",
1882 "src/f32-ibilinear-chw/gen/neonfma-p8.c",
Frank Barchard8247e212021-02-03 18:12:33 -08001883 "src/f32-ibilinear/gen/neonfma-c4.c",
1884 "src/f32-ibilinear/gen/neonfma-c8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001885 "src/f32-igemm/gen/1x8-minmax-neonfma-dup-ld64.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001886 "src/f32-igemm/gen/1x8s4-minmax-neonfma.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001887 "src/f32-igemm/gen/4x8-minmax-neonfma-dup-ld64.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001888 "src/f32-igemm/gen/4x8-minmax-neonfma-dup-ld128.c",
1889 "src/f32-igemm/gen/4x8s4-minmax-neonfma.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001890 "src/f32-igemm/gen/6x8-minmax-neonfma-dup-ld64.c",
1891 "src/f32-igemm/gen/6x8-minmax-neonfma-dup-ld128.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001892 "src/f32-igemm/gen/6x8s4-minmax-neonfma.c",
1893 "src/f32-igemm/gen/8x8s4-minmax-neonfma.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001894 "src/f32-ppmm/gen/4x8-minmax-neonfma.c",
1895 "src/f32-ppmm/gen/8x8-minmax-neonfma.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001896 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x4.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001897 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x8-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001898 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x8.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001899 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12-acc2.c",
1900 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001901 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001902 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16-acc2.c",
1903 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001904 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001905 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20-acc2.c",
1906 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001907 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20.c",
1908 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x4.c",
1909 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x8-acc2.c",
1910 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x8.c",
1911 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12-acc2.c",
1912 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12-acc3.c",
1913 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12.c",
1914 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16-acc2.c",
1915 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16-acc4.c",
1916 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16.c",
1917 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20-acc2.c",
1918 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20-acc5.c",
1919 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -08001920 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x4.c",
1921 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x8.c",
1922 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x12.c",
1923 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x16.c",
1924 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x20.c",
1925 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x24.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001926 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x4.c",
1927 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x8.c",
1928 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x12.c",
1929 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x16.c",
1930 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x20.c",
1931 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x24.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -08001932 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x4.c",
1933 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x8.c",
1934 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x12.c",
1935 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x16.c",
1936 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x20.c",
1937 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x24.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -08001938 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x4.c",
1939 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x8.c",
1940 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x12.c",
1941 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x16.c",
1942 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x20.c",
1943 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x24.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001944 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x4.c",
1945 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x8.c",
1946 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x12.c",
1947 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x16.c",
1948 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x20.c",
1949 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x24.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -08001950 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x4.c",
1951 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x8.c",
1952 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x12.c",
1953 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x16.c",
1954 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x20.c",
1955 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x24.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001956 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x4.c",
1957 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x8.c",
1958 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x12.c",
1959 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x16.c",
1960 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x20.c",
1961 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x24.c",
1962 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x4.c",
1963 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x8.c",
1964 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x12.c",
1965 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x16.c",
1966 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x20.c",
1967 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x24.c",
1968 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x4.c",
1969 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x8.c",
1970 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x12.c",
1971 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x16.c",
1972 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x20.c",
1973 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x24.c",
Marat Dukhan2fa7a0c2020-12-06 19:09:02 -08001974 "src/f32-spmm/gen/4x1-minmax-neonfma-pipelined.c",
1975 "src/f32-spmm/gen/4x1-minmax-neonfma-x2.c",
1976 "src/f32-spmm/gen/4x1-minmax-neonfma.c",
1977 "src/f32-spmm/gen/8x1-minmax-neonfma-pipelined.c",
1978 "src/f32-spmm/gen/8x1-minmax-neonfma-x2.c",
1979 "src/f32-spmm/gen/8x1-minmax-neonfma.c",
1980 "src/f32-spmm/gen/12x1-minmax-neonfma.c",
1981 "src/f32-spmm/gen/16x1-minmax-neonfma-pipelined.c",
1982 "src/f32-spmm/gen/16x1-minmax-neonfma-x2.c",
1983 "src/f32-spmm/gen/16x1-minmax-neonfma.c",
1984 "src/f32-spmm/gen/32x1-minmax-neonfma-pipelined.c",
1985 "src/f32-spmm/gen/32x1-minmax-neonfma-x2.c",
1986 "src/f32-spmm/gen/32x1-minmax-neonfma.c",
Marat Dukhaned6baaf2020-12-01 15:07:08 -08001987 "src/f32-velu/gen/velu-neonfma-rr1-lut16-p3-x4.c",
1988 "src/f32-velu/gen/velu-neonfma-rr1-lut16-p3-x8.c",
1989 "src/f32-velu/gen/velu-neonfma-rr1-lut16-p3-x12.c",
1990 "src/f32-velu/gen/velu-neonfma-rr1-lut16-p3-x16.c",
1991 "src/f32-velu/gen/velu-neonfma-rr1-lut16-p3-x20.c",
1992 "src/f32-velu/gen/velu-neonfma-rr1-lut16-p3-x24.c",
1993 "src/f32-velu/gen/velu-neonfma-rr1-p6-x4.c",
1994 "src/f32-velu/gen/velu-neonfma-rr1-p6-x8.c",
1995 "src/f32-velu/gen/velu-neonfma-rr1-p6-x12.c",
1996 "src/f32-velu/gen/velu-neonfma-rr1-p6-x16.c",
1997 "src/f32-velu/gen/velu-neonfma-rr1-p6-x20.c",
1998 "src/f32-velu/gen/velu-neonfma-rr1-p6-x24.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001999 "src/f32-vmulcaddc/gen/c4-minmax-neonfma-2x.c",
2000 "src/f32-vmulcaddc/gen/c8-minmax-neonfma-2x.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07002001 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x4.c",
2002 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x8.c",
2003 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x12.c",
2004 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x16.c",
2005 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x20.c",
2006 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x24.c",
2007 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x28.c",
2008 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x32.c",
2009 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x36.c",
2010 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x40.c",
2011 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x4.c",
2012 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x8.c",
2013 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x12.c",
2014 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x16.c",
2015 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x20.c",
2016 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x24.c",
2017 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x28.c",
2018 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x32.c",
2019 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x36.c",
2020 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x40.c",
Marat Dukhanb7633f22020-11-20 16:34:56 -08002021 "src/math/exp-neonfma-rr2-lut64-p2.c",
2022 "src/math/exp-neonfma-rr2-p5.c",
Frank Barcharde7223ee2020-12-04 19:04:01 -08002023 "src/math/expm1minus-neonfma-rr1-lut16-p3.c",
2024 "src/math/expm1minus-neonfma-rr1-p6.c",
Marat Dukhan9dd119a2020-11-20 18:20:04 -08002025 "src/math/expminus-neonfma-rr2-lut64-p2.c",
2026 "src/math/expminus-neonfma-rr2-lut2048-p1.c",
2027 "src/math/expminus-neonfma-rr2-p5.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08002028 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr1recps1fma.c",
2029 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr2fma.c",
2030 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr2recps.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002031 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr1recps1fma.c",
2032 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr2fma.c",
2033 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr2recps.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08002034 "src/math/sigmoid-neonfma-rr1-p5-nr1recps1fma.c",
2035 "src/math/sigmoid-neonfma-rr1-p5-nr2fma.c",
2036 "src/math/sigmoid-neonfma-rr1-p5-nr2recps.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08002037 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr1recps1fma.c",
2038 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr2fma.c",
2039 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr2recps.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002040 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr1recps1fma.c",
2041 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr2fma.c",
2042 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr2recps.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08002043 "src/math/sigmoid-neonfma-rr2-p5-nr1recps1fma.c",
2044 "src/math/sigmoid-neonfma-rr2-p5-nr2fma.c",
2045 "src/math/sigmoid-neonfma-rr2-p5-nr2recps.c",
Marat Dukhan84000762020-06-29 18:38:43 -07002046 "src/math/sqrt-neonfma-nr1fma.c",
Marat Dukhan84000762020-06-29 18:38:43 -07002047 "src/math/sqrt-neonfma-nr1rsqrts1fma1adj.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002048 "src/math/sqrt-neonfma-nr2fma.c",
2049 "src/math/sqrt-neonfma-nr2fma1adj.c",
2050 "src/math/sqrt-neonfma-nr3fma.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002051]
2052
2053AARCH64_NEONFMA_UKERNELS = [
Marat Dukhan56b10cd2020-05-18 09:35:49 -07002054 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neonfma-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07002055 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neonfma-2x2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002056 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neonfma-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07002057 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neonfma-2x2.c",
Marat Dukhan56b10cd2020-05-18 09:35:49 -07002058 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neonfma-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07002059 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neonfma-2x2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002060 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neonfma-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07002061 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neonfma-2x2.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -07002062 "src/f32-conv-hwc2chw/3x3s2p1c3x4-neonfma-2x2.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07002063 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-1x4-acc2.c",
2064 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-1x4-acc3.c",
2065 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-1x4-acc4.c",
Marat Dukhan1268a242020-10-24 00:36:32 -07002066 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-1x4.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07002067 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-2x4-acc2.c",
Marat Dukhan1268a242020-10-24 00:36:32 -07002068 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-2x4.c",
2069 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-3x4.c",
2070 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-4x4.c",
2071 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-5x4.c",
2072 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-6x4.c",
Marat Dukhan82f0c322020-10-25 19:17:35 -07002073 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neonfma-1x4-acc2.c",
2074 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neonfma-1x4-acc3.c",
2075 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neonfma-1x4-acc4.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07002076 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neonfma-1x4.c",
Marat Dukhan82f0c322020-10-25 19:17:35 -07002077 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neonfma-2x4-acc2.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07002078 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neonfma-2x4.c",
2079 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neonfma-3x4.c",
2080 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neonfma-4x4.c",
Marat Dukhan149f0ea2020-10-26 12:50:33 -07002081 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-1x4-acc2.c",
2082 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-1x4-acc3.c",
2083 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-1x4-acc4.c",
2084 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-1x4-acc5.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07002085 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-1x4.c",
Marat Dukhan149f0ea2020-10-26 12:50:33 -07002086 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-2x4-acc2.c",
2087 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-2x4-acc3.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07002088 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-2x4.c",
Marat Dukhan149f0ea2020-10-26 12:50:33 -07002089 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-3x4-acc2.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07002090 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-3x4.c",
Marat Dukhan149f0ea2020-10-26 12:50:33 -07002091 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-4x4-acc2.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07002092 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-4x4.c",
2093 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-5x4.c",
Marat Dukhan30d4b252020-10-29 16:33:22 -07002094 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neonfma-1x4-acc2.c",
2095 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neonfma-1x4-acc3.c",
2096 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neonfma-1x4-acc4.c",
2097 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neonfma-1x4-acc5.c",
2098 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neonfma-1x4.c",
2099 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neonfma-2x4-acc2.c",
2100 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neonfma-2x4-acc3.c",
2101 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neonfma-2x4.c",
Marat Dukhan30d4b252020-10-29 16:33:22 -07002102 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neonfma-3x4-acc2.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07002103 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neonfma-3x4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002104 "src/f32-gemm/gen-inc/1x8inc-minmax-neonfma-lane-ld64.c",
2105 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-lane-ld64.c",
2106 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-lane-ld128.c",
2107 "src/f32-gemm/gen-inc/5x8inc-minmax-neonfma-lane-ld64.c",
2108 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-lane-ld64.c",
2109 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-lane-ld128.c",
2110 "src/f32-gemm/gen/1x8-minmax-neonfma-lane-ld64.c",
2111 "src/f32-gemm/gen/4x2-minmax-neonfma-lane-ld64.c",
2112 "src/f32-gemm/gen/4x8-minmax-neonfma-lane-ld64.c",
2113 "src/f32-gemm/gen/4x8-minmax-neonfma-lane-ld128.c",
2114 "src/f32-gemm/gen/5x8-minmax-neonfma-lane-ld64.c",
2115 "src/f32-gemm/gen/6x8-minmax-neonfma-lane-ld64.c",
2116 "src/f32-gemm/gen/6x8-minmax-neonfma-lane-ld128.c",
2117 "src/f32-igemm/gen/1x8-minmax-neonfma-lane-ld64.c",
2118 "src/f32-igemm/gen/4x2-minmax-neonfma-lane-ld64.c",
2119 "src/f32-igemm/gen/4x4-minmax-neonfma-lane-ld64.c",
2120 "src/f32-igemm/gen/4x8-minmax-neonfma-lane-ld64.c",
2121 "src/f32-igemm/gen/4x8-minmax-neonfma-lane-ld128.c",
2122 "src/f32-igemm/gen/6x8-minmax-neonfma-lane-ld64.c",
2123 "src/f32-igemm/gen/6x8-minmax-neonfma-lane-ld128.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -08002124 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x4.c",
2125 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x8.c",
2126 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x12.c",
2127 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x16.c",
2128 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x20.c",
2129 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x24.c",
2130 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x4.c",
2131 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x8.c",
2132 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x12.c",
2133 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x16.c",
2134 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x20.c",
2135 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x24.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002136 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x4.c",
2137 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x8.c",
2138 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x12.c",
2139 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x16.c",
2140 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x20.c",
2141 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x24.c",
Marat Dukhan355ab432020-04-09 19:01:52 -07002142 "src/f32-spmm/gen/4x2-minmax-neonfma.c",
2143 "src/f32-spmm/gen/4x4-minmax-neonfma.c",
Marat Dukhan355ab432020-04-09 19:01:52 -07002144 "src/f32-spmm/gen/8x2-minmax-neonfma.c",
2145 "src/f32-spmm/gen/8x4-minmax-neonfma.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002146 "src/f32-spmm/gen/12x2-minmax-neonfma.c",
2147 "src/f32-spmm/gen/12x4-minmax-neonfma.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002148 "src/f32-spmm/gen/16x2-minmax-neonfma.c",
2149 "src/f32-spmm/gen/16x4-minmax-neonfma.c",
Frank Barchard846c0c62020-10-26 15:01:39 -07002150 "src/f32-spmm/gen/32x2-minmax-neonfma.c",
2151 "src/f32-spmm/gen/32x4-minmax-neonfma.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002152 "src/f32-vbinary/gen/vdiv-minmax-neon-x4.c",
2153 "src/f32-vbinary/gen/vdiv-minmax-neon-x8.c",
2154 "src/f32-vbinary/gen/vdivc-minmax-neon-x4.c",
2155 "src/f32-vbinary/gen/vdivc-minmax-neon-x8.c",
2156 "src/f32-vbinary/gen/vrdivc-minmax-neon-x4.c",
2157 "src/f32-vbinary/gen/vrdivc-minmax-neon-x8.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07002158 "src/f32-vsqrt/gen/neon-sqrt-x4.c",
2159 "src/f32-vsqrt/gen/neon-sqrt-x8.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08002160 "src/math/sigmoid-neonfma-rr1-lut64-p2-div.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002161 "src/math/sigmoid-neonfma-rr1-lut2048-p1-div.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08002162 "src/math/sigmoid-neonfma-rr1-p5-div.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08002163 "src/math/sigmoid-neonfma-rr2-lut64-p2-div.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002164 "src/math/sigmoid-neonfma-rr2-lut2048-p1-div.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08002165 "src/math/sigmoid-neonfma-rr2-p5-div.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002166]
2167
Marat Dukhan8853b822020-05-07 12:19:01 -07002168NEONV8_UKERNELS = [
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07002169 "src/f32-vrnd/gen/vrndd-neonv8-x4.c",
2170 "src/f32-vrnd/gen/vrndd-neonv8-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002171 "src/f32-vrnd/gen/vrndne-neonv8-x4.c",
2172 "src/f32-vrnd/gen/vrndne-neonv8-x8.c",
2173 "src/f32-vrnd/gen/vrndu-neonv8-x4.c",
2174 "src/f32-vrnd/gen/vrndu-neonv8-x8.c",
2175 "src/f32-vrnd/gen/vrndz-neonv8-x4.c",
2176 "src/f32-vrnd/gen/vrndz-neonv8-x8.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07002177 "src/math/roundd-neonv8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002178 "src/math/roundne-neonv8.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07002179 "src/math/roundu-neonv8.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07002180 "src/math/roundz-neonv8.c",
Marat Dukhan8853b822020-05-07 12:19:01 -07002181]
2182
Marat Dukhan08c4a432019-10-03 09:29:21 -07002183AARCH64_NEONFP16ARITH_UKERNELS = [
Frank Barchardb1966592020-05-12 13:47:06 -07002184 "src/f16-clamp/gen/neonfp16arith-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002185 "src/f16-clamp/gen/neonfp16arith-x16.c",
Frank Barchard5a599a62020-06-04 20:12:44 -07002186 "src/f16-dwconv/gen/up8x4-minmax-neonfp16arith-acc2.c",
2187 "src/f16-dwconv/gen/up8x4-minmax-neonfp16arith.c",
2188 "src/f16-dwconv/gen/up8x9-minmax-neonfp16arith-acc2.c",
2189 "src/f16-dwconv/gen/up8x9-minmax-neonfp16arith.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002190 "src/f16-dwconv/gen/up8x25-minmax-neonfp16arith-acc2.c",
2191 "src/f16-dwconv/gen/up8x25-minmax-neonfp16arith.c",
2192 "src/f16-dwconv/gen/up16x4-minmax-neonfp16arith-acc2.c",
2193 "src/f16-dwconv/gen/up16x4-minmax-neonfp16arith.c",
2194 "src/f16-dwconv/gen/up16x9-minmax-neonfp16arith-acc2.c",
2195 "src/f16-dwconv/gen/up16x9-minmax-neonfp16arith.c",
2196 "src/f16-dwconv/gen/up16x25-minmax-neonfp16arith-acc2.c",
2197 "src/f16-dwconv/gen/up16x25-minmax-neonfp16arith.c",
Frank Barchard0bb49a72020-06-04 11:35:11 -07002198 "src/f16-gavgpool/7p7x-minmax-neonfp16arith-c8.c",
2199 "src/f16-gavgpool/7x-minmax-neonfp16arith-c8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002200 "src/f16-gemm/gen-inc/1x8inc-minmax-neonfp16arith-ld64.c",
2201 "src/f16-gemm/gen-inc/1x16inc-minmax-neonfp16arith-ld64.c",
2202 "src/f16-gemm/gen-inc/4x8inc-minmax-neonfp16arith-ld64.c",
2203 "src/f16-gemm/gen-inc/4x16inc-minmax-neonfp16arith-ld64.c",
2204 "src/f16-gemm/gen-inc/6x8inc-minmax-neonfp16arith-ld64.c",
2205 "src/f16-gemm/gen-inc/6x16inc-minmax-neonfp16arith-ld64.c",
2206 "src/f16-gemm/gen-inc/8x8inc-minmax-neonfp16arith-ld64.c",
2207 "src/f16-gemm/gen-inc/8x16inc-minmax-neonfp16arith-ld64.c",
2208 "src/f16-gemm/gen/1x8-minmax-neonfp16arith-ld64.c",
2209 "src/f16-gemm/gen/1x16-minmax-neonfp16arith-ld64.c",
2210 "src/f16-gemm/gen/4x8-minmax-neonfp16arith-ld64.c",
2211 "src/f16-gemm/gen/4x16-minmax-neonfp16arith-ld64.c",
2212 "src/f16-gemm/gen/6x8-minmax-neonfp16arith-ld64.c",
2213 "src/f16-gemm/gen/6x16-minmax-neonfp16arith-ld64.c",
2214 "src/f16-gemm/gen/8x8-minmax-neonfp16arith-ld64.c",
2215 "src/f16-gemm/gen/8x16-minmax-neonfp16arith-ld64.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07002216 "src/f16-hswish/gen/hswish-neonfp16arith-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002217 "src/f16-hswish/gen/hswish-neonfp16arith-x16.c",
2218 "src/f16-igemm/gen/1x8-minmax-neonfp16arith-ld64.c",
2219 "src/f16-igemm/gen/1x16-minmax-neonfp16arith-ld64.c",
2220 "src/f16-igemm/gen/4x8-minmax-neonfp16arith-ld64.c",
2221 "src/f16-igemm/gen/4x16-minmax-neonfp16arith-ld64.c",
2222 "src/f16-igemm/gen/6x8-minmax-neonfp16arith-ld64.c",
2223 "src/f16-igemm/gen/6x16-minmax-neonfp16arith-ld64.c",
2224 "src/f16-igemm/gen/8x8-minmax-neonfp16arith-ld64.c",
2225 "src/f16-igemm/gen/8x16-minmax-neonfp16arith-ld64.c",
Frank Barchardb1966592020-05-12 13:47:06 -07002226 "src/f16-prelu/gen/neonfp16arith-2x8.c",
Frank Barcharda5316982020-07-23 13:19:28 -07002227 "src/f16-prelu/gen/neonfp16arith-2x16.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07002228 "src/f16-relu/gen/neonfp16arith-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002229 "src/f16-relu/gen/neonfp16arith-x16.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07002230 "src/f16-spmm/gen/8x1-minmax-neonfp16arith-x2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002231 "src/f16-spmm/gen/8x1-minmax-neonfp16arith.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07002232 "src/f16-spmm/gen/16x1-minmax-neonfp16arith-x2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002233 "src/f16-spmm/gen/16x1-minmax-neonfp16arith.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07002234 "src/f16-spmm/gen/24x1-minmax-neonfp16arith-x2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002235 "src/f16-spmm/gen/24x1-minmax-neonfp16arith.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07002236 "src/f16-spmm/gen/32x1-minmax-neonfp16arith-x2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002237 "src/f16-spmm/gen/32x1-minmax-neonfp16arith.c",
2238 "src/f16-vbinary/gen/vadd-minmax-neonfp16arith-x8.c",
2239 "src/f16-vbinary/gen/vadd-minmax-neonfp16arith-x16.c",
2240 "src/f16-vbinary/gen/vaddc-minmax-neonfp16arith-x8.c",
2241 "src/f16-vbinary/gen/vaddc-minmax-neonfp16arith-x16.c",
2242 "src/f16-vbinary/gen/vdiv-minmax-neonfp16arith-x8.c",
2243 "src/f16-vbinary/gen/vdiv-minmax-neonfp16arith-x16.c",
2244 "src/f16-vbinary/gen/vdivc-minmax-neonfp16arith-x8.c",
2245 "src/f16-vbinary/gen/vdivc-minmax-neonfp16arith-x16.c",
2246 "src/f16-vbinary/gen/vmax-neonfp16arith-x8.c",
2247 "src/f16-vbinary/gen/vmax-neonfp16arith-x16.c",
2248 "src/f16-vbinary/gen/vmaxc-neonfp16arith-x8.c",
2249 "src/f16-vbinary/gen/vmaxc-neonfp16arith-x16.c",
2250 "src/f16-vbinary/gen/vmin-neonfp16arith-x8.c",
2251 "src/f16-vbinary/gen/vmin-neonfp16arith-x16.c",
2252 "src/f16-vbinary/gen/vminc-neonfp16arith-x8.c",
2253 "src/f16-vbinary/gen/vminc-neonfp16arith-x16.c",
2254 "src/f16-vbinary/gen/vmul-minmax-neonfp16arith-x8.c",
2255 "src/f16-vbinary/gen/vmul-minmax-neonfp16arith-x16.c",
2256 "src/f16-vbinary/gen/vmulc-minmax-neonfp16arith-x8.c",
2257 "src/f16-vbinary/gen/vmulc-minmax-neonfp16arith-x16.c",
2258 "src/f16-vbinary/gen/vrdivc-minmax-neonfp16arith-x8.c",
2259 "src/f16-vbinary/gen/vrdivc-minmax-neonfp16arith-x16.c",
2260 "src/f16-vbinary/gen/vrsubc-minmax-neonfp16arith-x8.c",
2261 "src/f16-vbinary/gen/vrsubc-minmax-neonfp16arith-x16.c",
2262 "src/f16-vbinary/gen/vsub-minmax-neonfp16arith-x8.c",
2263 "src/f16-vbinary/gen/vsub-minmax-neonfp16arith-x16.c",
2264 "src/f16-vbinary/gen/vsubc-minmax-neonfp16arith-x8.c",
2265 "src/f16-vbinary/gen/vsubc-minmax-neonfp16arith-x16.c",
2266 "src/f16-vmulcaddc/gen/c8-minmax-neonfp16arith-2x.c",
2267 "src/f16-vmulcaddc/gen/c16-minmax-neonfp16arith-2x.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002268]
2269
Benoit Jacoba9644732020-08-13 12:48:55 -07002270NEONDOT_UKERNELS = [
Marat Dukhand4c83032020-09-09 23:20:12 -07002271 "src/qs8-gemm/gen/1x8c4-minmax-neondot.c",
Marat Dukhand4c83032020-09-09 23:20:12 -07002272 "src/qs8-gemm/gen/1x16c4-minmax-neondot.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002273 "src/qs8-gemm/gen/4x8c4-minmax-neondot.c",
Benoit Jacoba9644732020-08-13 12:48:55 -07002274 "src/qs8-gemm/gen/4x16c4-minmax-neondot.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002275 "src/qs8-gemm/gen/6x8c4-minmax-neondot.c",
Frank Barchard2fa17452020-09-22 09:51:46 -07002276 "src/qs8-gemm/gen/6x16c4-minmax-neondot.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002277 "src/qs8-gemm/gen/8x8c4-minmax-neondot.c",
Frank Barcharda48848f2020-09-24 00:59:24 -07002278 "src/qs8-gemm/gen/8x16c4-minmax-neondot.c",
Marat Dukhand4c83032020-09-09 23:20:12 -07002279 "src/qs8-igemm/gen/1x8c4-minmax-neondot.c",
Marat Dukhand4c83032020-09-09 23:20:12 -07002280 "src/qs8-igemm/gen/1x16c4-minmax-neondot.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002281 "src/qs8-igemm/gen/4x8c4-minmax-neondot.c",
Benoit Jacoba05487f2020-08-14 13:37:21 -07002282 "src/qs8-igemm/gen/4x16c4-minmax-neondot.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002283 "src/qs8-igemm/gen/6x8c4-minmax-neondot.c",
Frank Barchard2fa17452020-09-22 09:51:46 -07002284 "src/qs8-igemm/gen/6x16c4-minmax-neondot.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002285 "src/qs8-igemm/gen/8x8c4-minmax-neondot.c",
Frank Barcharda48848f2020-09-24 00:59:24 -07002286 "src/qs8-igemm/gen/8x16c4-minmax-neondot.c",
Benoit Jacoba9644732020-08-13 12:48:55 -07002287]
2288
Marat Dukhan08c4a432019-10-03 09:29:21 -07002289SSE_UKERNELS = [
Marat Dukhan99936602020-04-11 16:47:01 -07002290 "src/f32-avgpool/9p8x-minmax-sse-c4.c",
2291 "src/f32-avgpool/9x-minmax-sse-c4.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -07002292 "src/f32-clamp/gen/sse-x4.c",
2293 "src/f32-clamp/gen/sse-x8.c",
Erich Elsenb1233402020-06-08 15:53:15 -07002294 "src/f32-conv-hwc2chw/3x3s2p1c3x4-sse-1x1.c",
2295 "src/f32-conv-hwc2chw/3x3s2p1c3x4-sse-2x2.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002296 "src/f32-dwconv/gen/up4x4-minmax-sse-acc2.c",
2297 "src/f32-dwconv/gen/up4x4-minmax-sse.c",
2298 "src/f32-dwconv/gen/up4x9-minmax-sse-acc2.c",
2299 "src/f32-dwconv/gen/up4x9-minmax-sse.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002300 "src/f32-dwconv/gen/up4x25-minmax-sse-acc2.c",
2301 "src/f32-dwconv/gen/up4x25-minmax-sse.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002302 "src/f32-dwconv/gen/up8x4-minmax-sse-acc2.c",
2303 "src/f32-dwconv/gen/up8x4-minmax-sse.c",
2304 "src/f32-dwconv/gen/up8x9-minmax-sse-acc2.c",
2305 "src/f32-dwconv/gen/up8x9-minmax-sse.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002306 "src/f32-dwconv/gen/up8x25-minmax-sse-acc2.c",
2307 "src/f32-dwconv/gen/up8x25-minmax-sse.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07002308 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-1x4-acc2.c",
2309 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-1x4-acc3.c",
2310 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-1x4-acc4.c",
Marat Dukhan470078a2020-10-23 22:36:52 -07002311 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-1x4.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07002312 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-2x4-acc2.c",
Marat Dukhan470078a2020-10-23 22:36:52 -07002313 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-2x4.c",
2314 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-3x4.c",
2315 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-4x4.c",
2316 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-5x4.c",
2317 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-6x4.c",
Marat Dukhan0ff97182020-10-25 19:14:03 -07002318 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-sse-1x4-acc2.c",
2319 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-sse-1x4-acc3.c",
2320 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-sse-1x4-acc4.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07002321 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-sse-1x4.c",
Marat Dukhan0ff97182020-10-25 19:14:03 -07002322 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-sse-2x4-acc2.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07002323 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-sse-2x4.c",
2324 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-sse-3x4.c",
2325 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-sse-4x4.c",
Marat Dukhand0503892020-10-30 08:22:04 -07002326 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-1x4-acc2.c",
2327 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-1x4-acc3.c",
2328 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-1x4-acc4.c",
2329 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-1x4-acc5.c",
2330 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-1x4.c",
2331 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-2x4-acc2.c",
2332 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-2x4-acc3.c",
2333 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-2x4.c",
2334 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-3x4-acc2.c",
2335 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-3x4.c",
2336 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-4x4-acc2.c",
2337 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-4x4.c",
2338 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-5x4.c",
Marat Dukhanccca2142020-10-30 17:32:45 -07002339 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-sse-1x4-acc2.c",
2340 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-sse-1x4-acc3.c",
2341 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-sse-1x4-acc4.c",
2342 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-sse-1x4-acc5.c",
2343 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-sse-1x4.c",
2344 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-sse-2x4-acc2.c",
2345 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-sse-2x4-acc3.c",
2346 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-sse-2x4.c",
Marat Dukhanccca2142020-10-30 17:32:45 -07002347 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-sse-3x4-acc2.c",
Frank Barchard8ef44cd2020-11-03 12:30:23 -08002348 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-sse-3x4.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -07002349 "src/f32-gavgpool-cw/sse-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -07002350 "src/f32-gavgpool/7p7x-minmax-sse-c4.c",
2351 "src/f32-gavgpool/7x-minmax-sse-c4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002352 "src/f32-gemm/gen-inc/1x8inc-minmax-sse-dup.c",
2353 "src/f32-gemm/gen-inc/1x8inc-minmax-sse-load1.c",
2354 "src/f32-gemm/gen-inc/1x8s4inc-minmax-sse.c",
Marat Dukhan802fcae2020-12-11 14:37:25 -08002355 "src/f32-gemm/gen-inc/3x8inc-minmax-sse-dup.c",
2356 "src/f32-gemm/gen-inc/3x8inc-minmax-sse-load1.c",
2357 "src/f32-gemm/gen-inc/3x8s4inc-minmax-sse.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002358 "src/f32-gemm/gen-inc/4x8inc-minmax-sse-dup.c",
2359 "src/f32-gemm/gen-inc/4x8inc-minmax-sse-load1.c",
2360 "src/f32-gemm/gen-inc/4x8s4inc-minmax-sse.c",
Marat Dukhan802fcae2020-12-11 14:37:25 -08002361 "src/f32-gemm/gen-inc/5x8inc-minmax-sse-dup.c",
2362 "src/f32-gemm/gen-inc/5x8inc-minmax-sse-load1.c",
2363 "src/f32-gemm/gen-inc/5x8s4inc-minmax-sse.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002364 "src/f32-gemm/gen/1x8-minmax-sse-dup.c",
2365 "src/f32-gemm/gen/1x8-minmax-sse-load1.c",
2366 "src/f32-gemm/gen/1x8s4-minmax-sse.c",
Marat Dukhan802fcae2020-12-11 14:37:25 -08002367 "src/f32-gemm/gen/3x8-minmax-sse-dup.c",
2368 "src/f32-gemm/gen/3x8-minmax-sse-load1.c",
2369 "src/f32-gemm/gen/3x8s4-minmax-sse.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002370 "src/f32-gemm/gen/4x2c4-minmax-sse.c",
2371 "src/f32-gemm/gen/4x8-minmax-sse-dup.c",
2372 "src/f32-gemm/gen/4x8-minmax-sse-load1.c",
2373 "src/f32-gemm/gen/4x8s4-minmax-sse.c",
Marat Dukhan802fcae2020-12-11 14:37:25 -08002374 "src/f32-gemm/gen/5x8-minmax-sse-dup.c",
2375 "src/f32-gemm/gen/5x8-minmax-sse-load1.c",
2376 "src/f32-gemm/gen/5x8s4-minmax-sse.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07002377 "src/f32-hswish/gen/hswish-sse-x4.c",
2378 "src/f32-hswish/gen/hswish-sse-x8.c",
Marat Dukhan660fd192020-03-10 04:55:30 -07002379 "src/f32-ibilinear/gen/sse-c4.c",
2380 "src/f32-ibilinear/gen/sse-c8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002381 "src/f32-igemm/gen/1x8-minmax-sse-dup.c",
2382 "src/f32-igemm/gen/1x8-minmax-sse-load1.c",
2383 "src/f32-igemm/gen/1x8s4-minmax-sse.c",
Marat Dukhan802fcae2020-12-11 14:37:25 -08002384 "src/f32-igemm/gen/3x8-minmax-sse-dup.c",
2385 "src/f32-igemm/gen/3x8-minmax-sse-load1.c",
2386 "src/f32-igemm/gen/3x8s4-minmax-sse.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002387 "src/f32-igemm/gen/4x2c4-minmax-sse.c",
2388 "src/f32-igemm/gen/4x8-minmax-sse-dup.c",
2389 "src/f32-igemm/gen/4x8-minmax-sse-load1.c",
2390 "src/f32-igemm/gen/4x8s4-minmax-sse.c",
Marat Dukhan802fcae2020-12-11 14:37:25 -08002391 "src/f32-igemm/gen/5x8-minmax-sse-dup.c",
2392 "src/f32-igemm/gen/5x8-minmax-sse-load1.c",
2393 "src/f32-igemm/gen/5x8s4-minmax-sse.c",
Marat Dukhan99936602020-04-11 16:47:01 -07002394 "src/f32-maxpool/9p8x-minmax-sse-c4.c",
2395 "src/f32-pavgpool/9p8x-minmax-sse-c4.c",
2396 "src/f32-pavgpool/9x-minmax-sse-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002397 "src/f32-ppmm/gen/4x8-minmax-sse.c",
Marat Dukhan39b5e942020-06-24 15:03:48 -07002398 "src/f32-prelu/gen/sse-2x4.c",
2399 "src/f32-prelu/gen/sse-2x8.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07002400 "src/f32-relu/gen/sse-x4.c",
2401 "src/f32-relu/gen/sse-x8.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002402 "src/f32-rmax/sse.c",
Marat Dukhan355ab432020-04-09 19:01:52 -07002403 "src/f32-spmm/gen/4x1-minmax-sse.c",
2404 "src/f32-spmm/gen/8x1-minmax-sse.c",
Erich Elsen6e80fdc2020-06-09 15:35:37 -07002405 "src/f32-spmm/gen/16x1-minmax-sse.c",
Frank Barchard846c0c62020-10-26 15:01:39 -07002406 "src/f32-spmm/gen/32x1-minmax-sse.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002407 "src/f32-vbinary/gen/vadd-minmax-sse-x4.c",
2408 "src/f32-vbinary/gen/vadd-minmax-sse-x8.c",
2409 "src/f32-vbinary/gen/vaddc-minmax-sse-x4.c",
2410 "src/f32-vbinary/gen/vaddc-minmax-sse-x8.c",
2411 "src/f32-vbinary/gen/vdiv-minmax-sse-x4.c",
2412 "src/f32-vbinary/gen/vdiv-minmax-sse-x8.c",
2413 "src/f32-vbinary/gen/vdivc-minmax-sse-x4.c",
2414 "src/f32-vbinary/gen/vdivc-minmax-sse-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -08002415 "src/f32-vbinary/gen/vmax-sse-x4.c",
2416 "src/f32-vbinary/gen/vmax-sse-x8.c",
2417 "src/f32-vbinary/gen/vmaxc-sse-x4.c",
2418 "src/f32-vbinary/gen/vmaxc-sse-x8.c",
2419 "src/f32-vbinary/gen/vmin-sse-x4.c",
2420 "src/f32-vbinary/gen/vmin-sse-x8.c",
2421 "src/f32-vbinary/gen/vminc-sse-x4.c",
2422 "src/f32-vbinary/gen/vminc-sse-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002423 "src/f32-vbinary/gen/vmul-minmax-sse-x4.c",
2424 "src/f32-vbinary/gen/vmul-minmax-sse-x8.c",
2425 "src/f32-vbinary/gen/vmulc-minmax-sse-x4.c",
2426 "src/f32-vbinary/gen/vmulc-minmax-sse-x8.c",
2427 "src/f32-vbinary/gen/vrdivc-minmax-sse-x4.c",
2428 "src/f32-vbinary/gen/vrdivc-minmax-sse-x8.c",
2429 "src/f32-vbinary/gen/vrsubc-minmax-sse-x4.c",
2430 "src/f32-vbinary/gen/vrsubc-minmax-sse-x8.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07002431 "src/f32-vbinary/gen/vsqrdiff-sse-x4.c",
2432 "src/f32-vbinary/gen/vsqrdiff-sse-x8.c",
2433 "src/f32-vbinary/gen/vsqrdiffc-sse-x4.c",
2434 "src/f32-vbinary/gen/vsqrdiffc-sse-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002435 "src/f32-vbinary/gen/vsub-minmax-sse-x4.c",
2436 "src/f32-vbinary/gen/vsub-minmax-sse-x8.c",
2437 "src/f32-vbinary/gen/vsubc-minmax-sse-x4.c",
2438 "src/f32-vbinary/gen/vsubc-minmax-sse-x8.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07002439 "src/f32-vlrelu/gen/vlrelu-sse-x4.c",
2440 "src/f32-vlrelu/gen/vlrelu-sse-x8.c",
Marat Dukhan99936602020-04-11 16:47:01 -07002441 "src/f32-vmulcaddc/gen/c4-minmax-sse-2x.c",
2442 "src/f32-vmulcaddc/gen/c8-minmax-sse-2x.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07002443 "src/f32-vsqrt/gen/sse-sqrt-x4.c",
2444 "src/f32-vsqrt/gen/sse-sqrt-x8.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07002445 "src/f32-vunary/gen/vabs-sse-x4.c",
2446 "src/f32-vunary/gen/vabs-sse-x8.c",
2447 "src/f32-vunary/gen/vneg-sse-x4.c",
2448 "src/f32-vunary/gen/vneg-sse-x8.c",
2449 "src/f32-vunary/gen/vsqr-sse-x4.c",
2450 "src/f32-vunary/gen/vsqr-sse-x8.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07002451 "src/math/roundd-sse-addsub.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002452 "src/math/roundne-sse-addsub.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07002453 "src/math/roundu-sse-addsub.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07002454 "src/math/roundz-sse-addsub.c",
Marat Dukhan84000762020-06-29 18:38:43 -07002455 "src/math/sqrt-sse-hh1mac.c",
2456 "src/math/sqrt-sse-nr1mac.c",
2457 "src/math/sqrt-sse-nr2mac.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002458 "src/x32-fill/sse.c",
2459 "src/x32-packx/x4-sse.c",
2460 "src/x32-pad/sse.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002461]
2462
2463SSE2_UKERNELS = [
Marat Dukhan329da642019-11-19 21:44:39 -08002464 "src/f32-argmaxpool/4x-sse2-c4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002465 "src/f32-argmaxpool/9p8x-sse2-c4.c",
Marat Dukhan329da642019-11-19 21:44:39 -08002466 "src/f32-argmaxpool/9x-sse2-c4.c",
Marat Dukhan802fcae2020-12-11 14:37:25 -08002467 "src/f32-gemm/gen-inc/1x8inc-minmax-sse2-dup.c",
2468 "src/f32-gemm/gen-inc/3x8inc-minmax-sse2-dup.c",
2469 "src/f32-gemm/gen-inc/4x8inc-minmax-sse2-dup.c",
2470 "src/f32-gemm/gen-inc/5x8inc-minmax-sse2-dup.c",
2471 "src/f32-gemm/gen/1x8-minmax-sse2-dup.c",
2472 "src/f32-gemm/gen/3x8-minmax-sse2-dup.c",
2473 "src/f32-gemm/gen/4x8-minmax-sse2-dup.c",
2474 "src/f32-gemm/gen/5x8-minmax-sse2-dup.c",
2475 "src/f32-igemm/gen/1x8-minmax-sse2-dup.c",
2476 "src/f32-igemm/gen/3x8-minmax-sse2-dup.c",
2477 "src/f32-igemm/gen/4x8-minmax-sse2-dup.c",
2478 "src/f32-igemm/gen/5x8-minmax-sse2-dup.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08002479 "src/f32-prelu/gen/sse2-2x4.c",
2480 "src/f32-prelu/gen/sse2-2x8.c",
Marat Dukhanb39689d2020-01-24 13:32:20 -08002481 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x4.c",
Marat Dukhanb39689d2020-01-24 13:32:20 -08002482 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x8-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002483 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x8.c",
Marat Dukhanb39689d2020-01-24 13:32:20 -08002484 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12-acc2.c",
2485 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002486 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12.c",
Marat Dukhanb39689d2020-01-24 13:32:20 -08002487 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16-acc2.c",
2488 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002489 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16.c",
Marat Dukhanb39689d2020-01-24 13:32:20 -08002490 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20-acc2.c",
2491 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002492 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20.c",
Marat Dukhand243c1a2020-09-17 18:48:10 -07002493 "src/f32-sigmoid/gen/sse2-lut64-p2-div-x4.c",
2494 "src/f32-sigmoid/gen/sse2-lut64-p2-div-x8.c",
2495 "src/f32-sigmoid/gen/sse2-lut64-p2-div-x12.c",
2496 "src/f32-sigmoid/gen/sse2-lut64-p2-div-x16.c",
2497 "src/f32-sigmoid/gen/sse2-lut64-p2-div-x20.c",
2498 "src/f32-sigmoid/gen/sse2-lut64-p2-div-x24.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002499 "src/f32-sigmoid/gen/sse2-p5-div-x4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08002500 "src/f32-sigmoid/gen/sse2-p5-div-x8.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002501 "src/f32-sigmoid/gen/sse2-p5-div-x12.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08002502 "src/f32-sigmoid/gen/sse2-p5-div-x16.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002503 "src/f32-sigmoid/gen/sse2-p5-div-x20.c",
2504 "src/f32-sigmoid/gen/sse2-p5-div-x24.c",
Marat Dukhaned6baaf2020-12-01 15:07:08 -08002505 "src/f32-velu/gen/velu-sse2-rr2-lut16-p3-x4.c",
2506 "src/f32-velu/gen/velu-sse2-rr2-lut16-p3-x8.c",
2507 "src/f32-velu/gen/velu-sse2-rr2-lut16-p3-x12.c",
2508 "src/f32-velu/gen/velu-sse2-rr2-lut16-p3-x16.c",
2509 "src/f32-velu/gen/velu-sse2-rr2-lut16-p3-x20.c",
2510 "src/f32-velu/gen/velu-sse2-rr2-lut16-p3-x24.c",
2511 "src/f32-velu/gen/velu-sse2-rr2-p6-x4.c",
2512 "src/f32-velu/gen/velu-sse2-rr2-p6-x8.c",
2513 "src/f32-velu/gen/velu-sse2-rr2-p6-x12.c",
2514 "src/f32-velu/gen/velu-sse2-rr2-p6-x16.c",
2515 "src/f32-velu/gen/velu-sse2-rr2-p6-x20.c",
2516 "src/f32-velu/gen/velu-sse2-rr2-p6-x24.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07002517 "src/f32-vlrelu/gen/vlrelu-sse2-x4.c",
2518 "src/f32-vlrelu/gen/vlrelu-sse2-x8.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07002519 "src/f32-vrnd/gen/vrndd-sse2-x4.c",
2520 "src/f32-vrnd/gen/vrndd-sse2-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002521 "src/f32-vrnd/gen/vrndne-sse2-x4.c",
2522 "src/f32-vrnd/gen/vrndne-sse2-x8.c",
2523 "src/f32-vrnd/gen/vrndu-sse2-x4.c",
2524 "src/f32-vrnd/gen/vrndu-sse2-x8.c",
2525 "src/f32-vrnd/gen/vrndz-sse2-x4.c",
2526 "src/f32-vrnd/gen/vrndz-sse2-x8.c",
Marat Dukhanb7633f22020-11-20 16:34:56 -08002527 "src/math/exp-sse2-rr2-lut64-p2.c",
2528 "src/math/exp-sse2-rr2-p5.c",
Marat Dukhande390d42020-11-29 19:32:18 -08002529 "src/math/expm1minus-sse2-rr2-lut16-p3.c",
Marat Dukhana438aca2020-11-20 15:45:01 -08002530 "src/math/expm1minus-sse2-rr2-p6.c",
Frank Barchard3b800452020-11-22 12:12:35 -08002531 "src/math/expminus-sse2-rr2-p5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002532 "src/math/roundd-sse2-cvt.c",
2533 "src/math/roundne-sse2-cvt.c",
2534 "src/math/roundu-sse2-cvt.c",
2535 "src/math/roundz-sse2-cvt.c",
2536 "src/math/sigmoid-sse2-rr2-lut64-p2-div.c",
2537 "src/math/sigmoid-sse2-rr2-lut64-p2-nr1.c",
2538 "src/math/sigmoid-sse2-rr2-lut64-p2-nr2.c",
2539 "src/math/sigmoid-sse2-rr2-p5-div.c",
2540 "src/math/sigmoid-sse2-rr2-p5-nr1.c",
2541 "src/math/sigmoid-sse2-rr2-p5-nr2.c",
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07002542 "src/qs8-dwconv/gen/up8x9-minmax-sse2-mul16.c",
2543 "src/qs8-dwconv/gen/up16x9-minmax-sse2-mul16.c",
2544 "src/qs8-dwconv/gen/up24x9-minmax-sse2-mul16.c",
Marat Dukhan159688f2020-08-06 10:34:29 -07002545 "src/qs8-gavgpool/gen/7p7x-minmax-sse2-c8-acc2.c",
2546 "src/qs8-gavgpool/gen/7p7x-minmax-sse2-c16-acc2.c",
2547 "src/qs8-gavgpool/gen/7p7x-minmax-sse2-c24-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002548 "src/qs8-gavgpool/gen/7x-minmax-sse2-c8-acc2.c",
2549 "src/qs8-gavgpool/gen/7x-minmax-sse2-c16-acc2.c",
2550 "src/qs8-gavgpool/gen/7x-minmax-sse2-c24-acc2.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002551 "src/qs8-gemm/gen/1x4c2-minmax-sse2-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002552 "src/qs8-gemm/gen/1x4c2-minmax-sse2-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002553 "src/qs8-gemm/gen/1x4c2-xw-minmax-sse2.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002554 "src/qs8-gemm/gen/1x4c8-minmax-sse2-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002555 "src/qs8-gemm/gen/1x4c8-minmax-sse2-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002556 "src/qs8-gemm/gen/1x4c8-xw-minmax-sse2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002557 "src/qs8-gemm/gen/2x4c8-minmax-sse2-ld64.c",
2558 "src/qs8-gemm/gen/2x4c8-minmax-sse2-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002559 "src/qs8-gemm/gen/2x4c8-xw-minmax-sse2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002560 "src/qs8-gemm/gen/3x4c8-minmax-sse2-ld64.c",
2561 "src/qs8-gemm/gen/3x4c8-minmax-sse2-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002562 "src/qs8-gemm/gen/3x4c8-xw-minmax-sse2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002563 "src/qs8-gemm/gen/4x4c2-minmax-sse2-ld64.c",
2564 "src/qs8-gemm/gen/4x4c2-minmax-sse2-ld128.c",
2565 "src/qs8-gemm/gen/4x4c2-xw-minmax-sse2.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002566 "src/qs8-igemm/gen/1x4c2-minmax-sse2-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002567 "src/qs8-igemm/gen/1x4c2-minmax-sse2-ld128.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002568 "src/qs8-igemm/gen/1x4c8-minmax-sse2-ld64.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002569 "src/qs8-igemm/gen/1x4c8-minmax-sse2-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002570 "src/qs8-igemm/gen/2x4c8-minmax-sse2-ld64.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002571 "src/qs8-igemm/gen/2x4c8-minmax-sse2-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002572 "src/qs8-igemm/gen/3x4c8-minmax-sse2-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002573 "src/qs8-igemm/gen/3x4c8-minmax-sse2-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002574 "src/qs8-igemm/gen/4x4c2-minmax-sse2-ld64.c",
2575 "src/qs8-igemm/gen/4x4c2-minmax-sse2-ld128.c",
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07002576 "src/qs8-requantization/fp32-sse2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002577 "src/qs8-requantization/precise-sse2.c",
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07002578 "src/qs8-requantization/q31-sse2.c",
Marat Dukhand9f3ad42020-08-10 12:30:58 -07002579 "src/qs8-vadd/gen/minmax-sse2-mul16-ld64-x8.c",
2580 "src/qs8-vadd/gen/minmax-sse2-mul16-ld64-x16.c",
2581 "src/qs8-vadd/gen/minmax-sse2-mul16-ld64-x24.c",
2582 "src/qs8-vadd/gen/minmax-sse2-mul16-ld64-x32.c",
Marat Dukhan0270d9f2020-08-11 00:56:46 -07002583 "src/qs8-vaddc/gen/minmax-sse2-mul16-ld64-x8.c",
2584 "src/qs8-vaddc/gen/minmax-sse2-mul16-ld64-x16.c",
2585 "src/qs8-vaddc/gen/minmax-sse2-mul16-ld64-x24.c",
2586 "src/qs8-vaddc/gen/minmax-sse2-mul16-ld64-x32.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -07002587 "src/qu8-avgpool/9p8x-minmax-sse2-c8.c",
2588 "src/qu8-avgpool/9x-minmax-sse2-c8.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -07002589 "src/qu8-dwconv/up8x9-minmax-sse2.c",
2590 "src/qu8-gavgpool/7p7x-minmax-sse2-c8.c",
2591 "src/qu8-gavgpool/7x-minmax-sse2-c8.c",
2592 "src/qu8-gemm/2x4c8-minmax-sse2.c",
2593 "src/qu8-gemm/4x4c2-minmax-sse2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002594 "src/qu8-igemm/4x4c2-minmax-sse2.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07002595 "src/qu8-requantization/fp32-sse2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002596 "src/qu8-requantization/precise-sse2.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07002597 "src/qu8-requantization/q31-sse2.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -07002598 "src/qu8-vadd/minmax-sse2.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -07002599 "src/u8-clamp/sse2-x64.c",
Marat Dukhan99936602020-04-11 16:47:01 -07002600 "src/u8-maxpool/9p8x-minmax-sse2-c16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002601 "src/u8-rmax/sse2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002602 "src/x8-zip/x2-sse2.c",
2603 "src/x8-zip/x3-sse2.c",
2604 "src/x8-zip/x4-sse2.c",
2605 "src/x8-zip/xm-sse2.c",
Marat Dukhan57dccd82020-04-14 00:53:10 -07002606 "src/x32-unpool/sse2.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002607 "src/x32-zip/x2-sse2.c",
2608 "src/x32-zip/x3-sse2.c",
2609 "src/x32-zip/x4-sse2.c",
2610 "src/x32-zip/xm-sse2.c",
Marat Dukhanfe7acb62020-03-09 19:30:05 -07002611]
2612
2613SSSE3_UKERNELS = [
Frank Barchard35db7d02020-10-26 13:37:34 -07002614 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-1x4-acc2.c",
2615 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-1x4-acc3.c",
2616 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-1x4-acc4.c",
Marat Dukhan98f2eeb2020-10-23 23:13:41 -07002617 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-1x4.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07002618 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-2x4-acc2.c",
Marat Dukhan98f2eeb2020-10-23 23:13:41 -07002619 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-2x4.c",
2620 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-3x4.c",
2621 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-4x4.c",
2622 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-5x4.c",
2623 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-6x4.c",
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07002624 "src/qs8-dwconv/gen/up8x9-minmax-ssse3-mul16.c",
2625 "src/qs8-dwconv/gen/up16x9-minmax-ssse3-mul16.c",
2626 "src/qs8-dwconv/gen/up24x9-minmax-ssse3-mul16.c",
Marat Dukhan159688f2020-08-06 10:34:29 -07002627 "src/qs8-gavgpool/gen/7p7x-minmax-ssse3-c8-acc2.c",
2628 "src/qs8-gavgpool/gen/7p7x-minmax-ssse3-c16-acc2.c",
2629 "src/qs8-gavgpool/gen/7p7x-minmax-ssse3-c24-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002630 "src/qs8-gavgpool/gen/7x-minmax-ssse3-c8-acc2.c",
2631 "src/qs8-gavgpool/gen/7x-minmax-ssse3-c16-acc2.c",
2632 "src/qs8-gavgpool/gen/7x-minmax-ssse3-c24-acc2.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002633 "src/qs8-gemm/gen/1x4c2-minmax-ssse3-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002634 "src/qs8-gemm/gen/1x4c2-minmax-ssse3-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002635 "src/qs8-gemm/gen/1x4c2-xw-minmax-ssse3.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002636 "src/qs8-gemm/gen/1x4c8-minmax-ssse3-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002637 "src/qs8-gemm/gen/1x4c8-minmax-ssse3-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002638 "src/qs8-gemm/gen/1x4c8-xw-minmax-ssse3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002639 "src/qs8-gemm/gen/2x4c8-minmax-ssse3-ld64.c",
2640 "src/qs8-gemm/gen/2x4c8-minmax-ssse3-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002641 "src/qs8-gemm/gen/2x4c8-xw-minmax-ssse3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002642 "src/qs8-gemm/gen/3x4c8-minmax-ssse3-ld64.c",
2643 "src/qs8-gemm/gen/3x4c8-minmax-ssse3-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002644 "src/qs8-gemm/gen/3x4c8-xw-minmax-ssse3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002645 "src/qs8-gemm/gen/4x4c2-minmax-ssse3-ld64.c",
2646 "src/qs8-gemm/gen/4x4c2-minmax-ssse3-ld128.c",
2647 "src/qs8-gemm/gen/4x4c2-xw-minmax-ssse3.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002648 "src/qs8-igemm/gen/1x4c2-minmax-ssse3-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002649 "src/qs8-igemm/gen/1x4c2-minmax-ssse3-ld128.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002650 "src/qs8-igemm/gen/1x4c8-minmax-ssse3-ld64.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002651 "src/qs8-igemm/gen/1x4c8-minmax-ssse3-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002652 "src/qs8-igemm/gen/2x4c8-minmax-ssse3-ld64.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002653 "src/qs8-igemm/gen/2x4c8-minmax-ssse3-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002654 "src/qs8-igemm/gen/3x4c8-minmax-ssse3-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002655 "src/qs8-igemm/gen/3x4c8-minmax-ssse3-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002656 "src/qs8-igemm/gen/4x4c2-minmax-ssse3-ld64.c",
2657 "src/qs8-igemm/gen/4x4c2-minmax-ssse3-ld128.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -07002658 "src/qs8-requantization/precise-ssse3.c",
2659 "src/qs8-requantization/q31-ssse3.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07002660 "src/qu8-requantization/precise-ssse3.c",
2661 "src/qu8-requantization/q31-ssse3.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002662]
2663
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08002664SSE41_UKERNELS = [
Marat Dukhan40a672f2019-11-25 03:08:22 -08002665 "src/f32-prelu/gen/sse41-2x4.c",
2666 "src/f32-prelu/gen/sse41-2x8.c",
Marat Dukhand243c1a2020-09-17 18:48:10 -07002667 "src/f32-sigmoid/gen/sse41-lut64-p2-div-x4.c",
2668 "src/f32-sigmoid/gen/sse41-lut64-p2-div-x8.c",
2669 "src/f32-sigmoid/gen/sse41-lut64-p2-div-x12.c",
2670 "src/f32-sigmoid/gen/sse41-lut64-p2-div-x16.c",
2671 "src/f32-sigmoid/gen/sse41-lut64-p2-div-x20.c",
2672 "src/f32-sigmoid/gen/sse41-lut64-p2-div-x24.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002673 "src/f32-sigmoid/gen/sse41-p5-div-x4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08002674 "src/f32-sigmoid/gen/sse41-p5-div-x8.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002675 "src/f32-sigmoid/gen/sse41-p5-div-x12.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08002676 "src/f32-sigmoid/gen/sse41-p5-div-x16.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002677 "src/f32-sigmoid/gen/sse41-p5-div-x20.c",
2678 "src/f32-sigmoid/gen/sse41-p5-div-x24.c",
Marat Dukhaned6baaf2020-12-01 15:07:08 -08002679 "src/f32-velu/gen/velu-sse41-rr2-lut16-p3-x4.c",
2680 "src/f32-velu/gen/velu-sse41-rr2-lut16-p3-x8.c",
2681 "src/f32-velu/gen/velu-sse41-rr2-lut16-p3-x12.c",
2682 "src/f32-velu/gen/velu-sse41-rr2-lut16-p3-x16.c",
2683 "src/f32-velu/gen/velu-sse41-rr2-lut16-p3-x20.c",
2684 "src/f32-velu/gen/velu-sse41-rr2-lut16-p3-x24.c",
2685 "src/f32-velu/gen/velu-sse41-rr2-p6-x4.c",
2686 "src/f32-velu/gen/velu-sse41-rr2-p6-x8.c",
2687 "src/f32-velu/gen/velu-sse41-rr2-p6-x12.c",
2688 "src/f32-velu/gen/velu-sse41-rr2-p6-x16.c",
2689 "src/f32-velu/gen/velu-sse41-rr2-p6-x20.c",
2690 "src/f32-velu/gen/velu-sse41-rr2-p6-x24.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07002691 "src/f32-vlrelu/gen/vlrelu-sse41-x4.c",
2692 "src/f32-vlrelu/gen/vlrelu-sse41-x8.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07002693 "src/f32-vrnd/gen/vrndd-sse41-x4.c",
2694 "src/f32-vrnd/gen/vrndd-sse41-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002695 "src/f32-vrnd/gen/vrndne-sse41-x4.c",
2696 "src/f32-vrnd/gen/vrndne-sse41-x8.c",
2697 "src/f32-vrnd/gen/vrndu-sse41-x4.c",
2698 "src/f32-vrnd/gen/vrndu-sse41-x8.c",
2699 "src/f32-vrnd/gen/vrndz-sse41-x4.c",
2700 "src/f32-vrnd/gen/vrndz-sse41-x8.c",
2701 "src/math/roundd-sse41.c",
2702 "src/math/roundne-sse41.c",
2703 "src/math/roundu-sse41.c",
2704 "src/math/roundz-sse41.c",
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07002705 "src/qs8-dwconv/gen/up8x9-minmax-sse41-mul16.c",
2706 "src/qs8-dwconv/gen/up16x9-minmax-sse41-mul16.c",
2707 "src/qs8-dwconv/gen/up24x9-minmax-sse41-mul16.c",
Marat Dukhan159688f2020-08-06 10:34:29 -07002708 "src/qs8-gavgpool/gen/7p7x-minmax-sse41-c8-acc2.c",
2709 "src/qs8-gavgpool/gen/7p7x-minmax-sse41-c16-acc2.c",
2710 "src/qs8-gavgpool/gen/7p7x-minmax-sse41-c24-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002711 "src/qs8-gavgpool/gen/7x-minmax-sse41-c8-acc2.c",
2712 "src/qs8-gavgpool/gen/7x-minmax-sse41-c16-acc2.c",
2713 "src/qs8-gavgpool/gen/7x-minmax-sse41-c24-acc2.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002714 "src/qs8-gemm/gen/1x4c2-minmax-sse41-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002715 "src/qs8-gemm/gen/1x4c2-minmax-sse41-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002716 "src/qs8-gemm/gen/1x4c2-xw-minmax-sse41.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002717 "src/qs8-gemm/gen/1x4c8-minmax-sse41-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002718 "src/qs8-gemm/gen/1x4c8-minmax-sse41-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002719 "src/qs8-gemm/gen/1x4c8-xw-minmax-sse41.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002720 "src/qs8-gemm/gen/2x4c8-minmax-sse41-ld64.c",
2721 "src/qs8-gemm/gen/2x4c8-minmax-sse41-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002722 "src/qs8-gemm/gen/2x4c8-xw-minmax-sse41.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002723 "src/qs8-gemm/gen/3x4c8-minmax-sse41-ld64.c",
2724 "src/qs8-gemm/gen/3x4c8-minmax-sse41-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002725 "src/qs8-gemm/gen/3x4c8-xw-minmax-sse41.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002726 "src/qs8-gemm/gen/4x4c2-minmax-sse41-ld64.c",
2727 "src/qs8-gemm/gen/4x4c2-minmax-sse41-ld128.c",
2728 "src/qs8-gemm/gen/4x4c2-xw-minmax-sse41.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002729 "src/qs8-igemm/gen/1x4c2-minmax-sse41-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002730 "src/qs8-igemm/gen/1x4c2-minmax-sse41-ld128.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002731 "src/qs8-igemm/gen/1x4c8-minmax-sse41-ld64.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002732 "src/qs8-igemm/gen/1x4c8-minmax-sse41-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002733 "src/qs8-igemm/gen/2x4c8-minmax-sse41-ld64.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002734 "src/qs8-igemm/gen/2x4c8-minmax-sse41-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002735 "src/qs8-igemm/gen/3x4c8-minmax-sse41-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002736 "src/qs8-igemm/gen/3x4c8-minmax-sse41-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002737 "src/qs8-igemm/gen/4x4c2-minmax-sse41-ld64.c",
2738 "src/qs8-igemm/gen/4x4c2-minmax-sse41-ld128.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -07002739 "src/qs8-requantization/fp32-sse4.c",
2740 "src/qs8-requantization/precise-sse4.c",
2741 "src/qs8-requantization/q31-sse4.c",
Marat Dukhand9f3ad42020-08-10 12:30:58 -07002742 "src/qs8-vadd/gen/minmax-sse41-mul16-ld64-x8.c",
2743 "src/qs8-vadd/gen/minmax-sse41-mul16-ld64-x16.c",
2744 "src/qs8-vadd/gen/minmax-sse41-mul16-ld64-x24.c",
2745 "src/qs8-vadd/gen/minmax-sse41-mul16-ld64-x32.c",
Marat Dukhanbb9225e2020-09-06 22:40:56 -07002746 "src/qs8-vadd/gen/minmax-sse41-mul32-ld32-x8.c",
2747 "src/qs8-vadd/gen/minmax-sse41-mul32-ld32-x16.c",
2748 "src/qs8-vadd/gen/minmax-sse41-mul32-ld32-x24.c",
2749 "src/qs8-vadd/gen/minmax-sse41-mul32-ld32-x32.c",
Marat Dukhan0270d9f2020-08-11 00:56:46 -07002750 "src/qs8-vaddc/gen/minmax-sse41-mul16-ld64-x8.c",
2751 "src/qs8-vaddc/gen/minmax-sse41-mul16-ld64-x16.c",
2752 "src/qs8-vaddc/gen/minmax-sse41-mul16-ld64-x24.c",
2753 "src/qs8-vaddc/gen/minmax-sse41-mul16-ld64-x32.c",
Marat Dukhanbb9225e2020-09-06 22:40:56 -07002754 "src/qs8-vaddc/gen/minmax-sse41-mul32-ld32-x8.c",
2755 "src/qs8-vaddc/gen/minmax-sse41-mul32-ld32-x16.c",
2756 "src/qs8-vaddc/gen/minmax-sse41-mul32-ld32-x24.c",
2757 "src/qs8-vaddc/gen/minmax-sse41-mul32-ld32-x32.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07002758 "src/qu8-requantization/precise-sse4.c",
2759 "src/qu8-requantization/q31-sse4.c",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08002760]
2761
Marat Dukhan08c4a432019-10-03 09:29:21 -07002762AVX_UKERNELS = [
Marat Dukhan5c5fa962020-03-10 18:38:33 -07002763 "src/f32-clamp/gen/avx-x8.c",
2764 "src/f32-clamp/gen/avx-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002765 "src/f32-dwconv/gen/up8x4-minmax-avx-acc2.c",
2766 "src/f32-dwconv/gen/up8x4-minmax-avx.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002767 "src/f32-dwconv/gen/up8x9-minmax-avx-acc2.c",
2768 "src/f32-dwconv/gen/up8x9-minmax-avx.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002769 "src/f32-dwconv/gen/up8x25-minmax-avx-acc2.c",
2770 "src/f32-dwconv/gen/up8x25-minmax-avx.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002771 "src/f32-dwconv/gen/up16x4-minmax-avx-acc2.c",
2772 "src/f32-dwconv/gen/up16x4-minmax-avx.c",
2773 "src/f32-dwconv/gen/up16x9-minmax-avx-acc2.c",
2774 "src/f32-dwconv/gen/up16x9-minmax-avx.c",
2775 "src/f32-dwconv/gen/up16x25-minmax-avx-acc2.c",
2776 "src/f32-dwconv/gen/up16x25-minmax-avx.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002777 "src/f32-gemm/gen-inc/1x8inc-minmax-avx-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002778 "src/f32-gemm/gen-inc/1x16inc-minmax-avx-broadcast.c",
2779 "src/f32-gemm/gen-inc/3x16inc-minmax-avx-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002780 "src/f32-gemm/gen-inc/4x8inc-minmax-avx-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002781 "src/f32-gemm/gen-inc/4x16inc-minmax-avx-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002782 "src/f32-gemm/gen-inc/5x8inc-minmax-avx-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002783 "src/f32-gemm/gen-inc/5x16inc-minmax-avx-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002784 "src/f32-gemm/gen-inc/6x8inc-minmax-avx-broadcast.c",
2785 "src/f32-gemm/gen-inc/7x8inc-minmax-avx-broadcast.c",
2786 "src/f32-gemm/gen/1x8-minmax-avx-broadcast.c",
2787 "src/f32-gemm/gen/1x16-minmax-avx-broadcast.c",
2788 "src/f32-gemm/gen/3x16-minmax-avx-broadcast.c",
2789 "src/f32-gemm/gen/4x8-minmax-avx-broadcast.c",
2790 "src/f32-gemm/gen/4x16-minmax-avx-broadcast.c",
2791 "src/f32-gemm/gen/5x8-minmax-avx-broadcast.c",
2792 "src/f32-gemm/gen/5x16-minmax-avx-broadcast.c",
2793 "src/f32-gemm/gen/6x8-minmax-avx-broadcast.c",
2794 "src/f32-gemm/gen/7x8-minmax-avx-broadcast.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07002795 "src/f32-hswish/gen/hswish-avx-x8.c",
2796 "src/f32-hswish/gen/hswish-avx-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002797 "src/f32-igemm/gen/1x8-minmax-avx-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002798 "src/f32-igemm/gen/1x16-minmax-avx-broadcast.c",
2799 "src/f32-igemm/gen/3x16-minmax-avx-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002800 "src/f32-igemm/gen/4x8-minmax-avx-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002801 "src/f32-igemm/gen/4x16-minmax-avx-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002802 "src/f32-igemm/gen/5x8-minmax-avx-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002803 "src/f32-igemm/gen/5x16-minmax-avx-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002804 "src/f32-igemm/gen/6x8-minmax-avx-broadcast.c",
2805 "src/f32-igemm/gen/7x8-minmax-avx-broadcast.c",
Marat Dukhan90eca0a2020-03-11 00:52:23 -07002806 "src/f32-prelu/gen/avx-2x8.c",
2807 "src/f32-prelu/gen/avx-2x16.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07002808 "src/f32-relu/gen/avx-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002809 "src/f32-relu/gen/avx-x16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002810 "src/f32-rmax/avx.c",
T.J. Alumbaughdc2b29c2020-10-14 13:56:08 -07002811 "src/f32-sigmoid/gen/avx-rr2-p5-div-x8.c",
2812 "src/f32-sigmoid/gen/avx-rr2-p5-div-x16.c",
2813 "src/f32-sigmoid/gen/avx-rr2-p5-div-x24.c",
2814 "src/f32-sigmoid/gen/avx-rr2-p5-div-x32.c",
2815 "src/f32-sigmoid/gen/avx-rr2-p5-div-x40.c",
2816 "src/f32-sigmoid/gen/avx-rr2-p5-div-x48.c",
2817 "src/f32-sigmoid/gen/avx-rr2-p5-div-x56.c",
2818 "src/f32-sigmoid/gen/avx-rr2-p5-div-x64.c",
2819 "src/f32-sigmoid/gen/avx-rr2-p5-div-x72.c",
2820 "src/f32-sigmoid/gen/avx-rr2-p5-div-x80.c",
2821 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x8.c",
2822 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x16.c",
2823 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x24.c",
2824 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x32.c",
2825 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x40.c",
2826 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x48.c",
2827 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x56.c",
2828 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x64.c",
2829 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x72.c",
2830 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x80.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002831 "src/f32-vbinary/gen/vadd-minmax-avx-x8.c",
2832 "src/f32-vbinary/gen/vadd-minmax-avx-x16.c",
2833 "src/f32-vbinary/gen/vaddc-minmax-avx-x8.c",
2834 "src/f32-vbinary/gen/vaddc-minmax-avx-x16.c",
2835 "src/f32-vbinary/gen/vdiv-minmax-avx-x8.c",
2836 "src/f32-vbinary/gen/vdiv-minmax-avx-x16.c",
2837 "src/f32-vbinary/gen/vdivc-minmax-avx-x8.c",
2838 "src/f32-vbinary/gen/vdivc-minmax-avx-x16.c",
Marat Dukhan9a88efe2019-12-10 15:54:24 -08002839 "src/f32-vbinary/gen/vmax-avx-x8.c",
2840 "src/f32-vbinary/gen/vmax-avx-x16.c",
2841 "src/f32-vbinary/gen/vmaxc-avx-x8.c",
2842 "src/f32-vbinary/gen/vmaxc-avx-x16.c",
2843 "src/f32-vbinary/gen/vmin-avx-x8.c",
2844 "src/f32-vbinary/gen/vmin-avx-x16.c",
2845 "src/f32-vbinary/gen/vminc-avx-x8.c",
2846 "src/f32-vbinary/gen/vminc-avx-x16.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002847 "src/f32-vbinary/gen/vmul-minmax-avx-x8.c",
2848 "src/f32-vbinary/gen/vmul-minmax-avx-x16.c",
2849 "src/f32-vbinary/gen/vmulc-minmax-avx-x8.c",
2850 "src/f32-vbinary/gen/vmulc-minmax-avx-x16.c",
2851 "src/f32-vbinary/gen/vrdivc-minmax-avx-x8.c",
2852 "src/f32-vbinary/gen/vrdivc-minmax-avx-x16.c",
2853 "src/f32-vbinary/gen/vrsubc-minmax-avx-x8.c",
2854 "src/f32-vbinary/gen/vrsubc-minmax-avx-x16.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07002855 "src/f32-vbinary/gen/vsqrdiff-avx-x8.c",
2856 "src/f32-vbinary/gen/vsqrdiff-avx-x16.c",
2857 "src/f32-vbinary/gen/vsqrdiffc-avx-x8.c",
2858 "src/f32-vbinary/gen/vsqrdiffc-avx-x16.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002859 "src/f32-vbinary/gen/vsub-minmax-avx-x8.c",
2860 "src/f32-vbinary/gen/vsub-minmax-avx-x16.c",
2861 "src/f32-vbinary/gen/vsubc-minmax-avx-x8.c",
2862 "src/f32-vbinary/gen/vsubc-minmax-avx-x16.c",
Marat Dukhaned6baaf2020-12-01 15:07:08 -08002863 "src/f32-velu/gen/velu-avx-rr2-lut4-p4-perm-x8.c",
2864 "src/f32-velu/gen/velu-avx-rr2-lut4-p4-perm-x16.c",
2865 "src/f32-velu/gen/velu-avx-rr2-lut4-p4-perm-x24.c",
2866 "src/f32-velu/gen/velu-avx-rr2-lut4-p4-perm-x32.c",
2867 "src/f32-velu/gen/velu-avx-rr2-lut4-p4-perm-x40.c",
2868 "src/f32-velu/gen/velu-avx-rr2-lut4-p4-perm-x48.c",
2869 "src/f32-velu/gen/velu-avx-rr2-lut16-p3-x8.c",
2870 "src/f32-velu/gen/velu-avx-rr2-lut16-p3-x16.c",
2871 "src/f32-velu/gen/velu-avx-rr2-lut16-p3-x24.c",
2872 "src/f32-velu/gen/velu-avx-rr2-lut16-p3-x32.c",
2873 "src/f32-velu/gen/velu-avx-rr2-lut16-p3-x40.c",
2874 "src/f32-velu/gen/velu-avx-rr2-lut16-p3-x48.c",
2875 "src/f32-velu/gen/velu-avx-rr2-p6-x8.c",
2876 "src/f32-velu/gen/velu-avx-rr2-p6-x16.c",
2877 "src/f32-velu/gen/velu-avx-rr2-p6-x24.c",
2878 "src/f32-velu/gen/velu-avx-rr2-p6-x32.c",
2879 "src/f32-velu/gen/velu-avx-rr2-p6-x40.c",
2880 "src/f32-velu/gen/velu-avx-rr2-p6-x48.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07002881 "src/f32-vlrelu/gen/vlrelu-avx-x8.c",
2882 "src/f32-vlrelu/gen/vlrelu-avx-x16.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07002883 "src/f32-vrnd/gen/vrndd-avx-x8.c",
2884 "src/f32-vrnd/gen/vrndd-avx-x16.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002885 "src/f32-vrnd/gen/vrndne-avx-x8.c",
2886 "src/f32-vrnd/gen/vrndne-avx-x16.c",
2887 "src/f32-vrnd/gen/vrndu-avx-x8.c",
2888 "src/f32-vrnd/gen/vrndu-avx-x16.c",
2889 "src/f32-vrnd/gen/vrndz-avx-x8.c",
2890 "src/f32-vrnd/gen/vrndz-avx-x16.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07002891 "src/f32-vscale/avx-x32.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07002892 "src/f32-vsqrt/gen/avx-sqrt-x8.c",
2893 "src/f32-vsqrt/gen/avx-sqrt-x16.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07002894 "src/f32-vunary/gen/vabs-avx-x8.c",
2895 "src/f32-vunary/gen/vabs-avx-x16.c",
2896 "src/f32-vunary/gen/vneg-avx-x8.c",
2897 "src/f32-vunary/gen/vneg-avx-x16.c",
2898 "src/f32-vunary/gen/vsqr-avx-x8.c",
2899 "src/f32-vunary/gen/vsqr-avx-x16.c",
Marat Dukhanb7633f22020-11-20 16:34:56 -08002900 "src/math/exp-avx-rr2-p5.c",
Marat Dukhande390d42020-11-29 19:32:18 -08002901 "src/math/expm1minus-avx-rr2-lut4-p4-perm.c",
2902 "src/math/expm1minus-avx-rr2-lut16-p3.c",
2903 "src/math/expm1minus-avx-rr2-p6.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002904 "src/math/sigmoid-avx-rr2-lut64-p2-div.c",
Marat Dukhan2e4727b2020-09-24 00:48:39 -07002905 "src/math/sigmoid-avx-rr2-p5-div.c",
Marat Dukhan39ab16f2020-10-16 14:30:35 -07002906 "src/math/sigmoid-avx-rr2-p5-nr1.c",
2907 "src/math/sigmoid-avx-rr2-p5-nr2.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002908]
2909
Marat Dukhan1566fee2020-08-02 21:55:41 -07002910XOP_UKERNELS = [
2911 "src/qs8-gemm/gen/1x4c2-minmax-xop-ld64.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002912 "src/qs8-gemm/gen/1x4c2-minmax-xop-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002913 "src/qs8-gemm/gen/1x4c2-xw-minmax-xop.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002914 "src/qs8-gemm/gen/1x4c8-minmax-xop-ld64.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002915 "src/qs8-gemm/gen/1x4c8-minmax-xop-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002916 "src/qs8-gemm/gen/1x4c8-xw-minmax-xop.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002917 "src/qs8-gemm/gen/2x4c8-minmax-xop-ld64.c",
2918 "src/qs8-gemm/gen/2x4c8-minmax-xop-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002919 "src/qs8-gemm/gen/2x4c8-xw-minmax-xop.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002920 "src/qs8-gemm/gen/3x4c8-minmax-xop-ld64.c",
2921 "src/qs8-gemm/gen/3x4c8-minmax-xop-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002922 "src/qs8-gemm/gen/3x4c8-xw-minmax-xop.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002923 "src/qs8-gemm/gen/4x4c2-minmax-xop-ld64.c",
2924 "src/qs8-gemm/gen/4x4c2-minmax-xop-ld128.c",
2925 "src/qs8-gemm/gen/4x4c2-xw-minmax-xop.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002926 "src/qs8-igemm/gen/1x4c2-minmax-xop-ld64.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002927 "src/qs8-igemm/gen/1x4c2-minmax-xop-ld128.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002928 "src/qs8-igemm/gen/1x4c8-minmax-xop-ld64.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002929 "src/qs8-igemm/gen/1x4c8-minmax-xop-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002930 "src/qs8-igemm/gen/2x4c8-minmax-xop-ld64.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002931 "src/qs8-igemm/gen/2x4c8-minmax-xop-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002932 "src/qs8-igemm/gen/3x4c8-minmax-xop-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002933 "src/qs8-igemm/gen/3x4c8-minmax-xop-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002934 "src/qs8-igemm/gen/4x4c2-minmax-xop-ld64.c",
2935 "src/qs8-igemm/gen/4x4c2-minmax-xop-ld128.c",
Marat Dukhanbb9225e2020-09-06 22:40:56 -07002936 "src/qs8-vadd/gen/minmax-xop-mul32-ld32-x8.c",
2937 "src/qs8-vadd/gen/minmax-xop-mul32-ld32-x16.c",
2938 "src/qs8-vadd/gen/minmax-xop-mul32-ld32-x24.c",
2939 "src/qs8-vadd/gen/minmax-xop-mul32-ld32-x32.c",
2940 "src/qs8-vaddc/gen/minmax-xop-mul32-ld32-x8.c",
2941 "src/qs8-vaddc/gen/minmax-xop-mul32-ld32-x16.c",
2942 "src/qs8-vaddc/gen/minmax-xop-mul32-ld32-x24.c",
2943 "src/qs8-vaddc/gen/minmax-xop-mul32-ld32-x32.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002944]
2945
Marat Dukhanfda12b82019-11-21 12:27:59 -08002946FMA3_UKERNELS = [
Marat Dukhan1c587112020-04-08 20:04:28 -07002947 "src/f32-dwconv/gen/up8x4-minmax-fma3-acc2.c",
2948 "src/f32-dwconv/gen/up8x4-minmax-fma3.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002949 "src/f32-dwconv/gen/up8x9-minmax-fma3-acc2.c",
2950 "src/f32-dwconv/gen/up8x9-minmax-fma3.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002951 "src/f32-dwconv/gen/up8x25-minmax-fma3-acc2.c",
2952 "src/f32-dwconv/gen/up8x25-minmax-fma3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002953 "src/f32-dwconv/gen/up16x4-minmax-fma3-acc2.c",
2954 "src/f32-dwconv/gen/up16x4-minmax-fma3.c",
2955 "src/f32-dwconv/gen/up16x9-minmax-fma3-acc2.c",
2956 "src/f32-dwconv/gen/up16x9-minmax-fma3.c",
2957 "src/f32-dwconv/gen/up16x25-minmax-fma3-acc2.c",
2958 "src/f32-dwconv/gen/up16x25-minmax-fma3.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002959 "src/f32-gemm/gen-inc/1x8inc-minmax-fma3-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002960 "src/f32-gemm/gen-inc/1x16inc-minmax-fma3-broadcast.c",
2961 "src/f32-gemm/gen-inc/1x16s4inc-minmax-fma3-broadcast.c",
2962 "src/f32-gemm/gen-inc/3x16inc-minmax-fma3-broadcast.c",
2963 "src/f32-gemm/gen-inc/3x16s4inc-minmax-fma3-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002964 "src/f32-gemm/gen-inc/4x8inc-minmax-fma3-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002965 "src/f32-gemm/gen-inc/4x16inc-minmax-fma3-broadcast.c",
2966 "src/f32-gemm/gen-inc/4x16s4inc-minmax-fma3-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002967 "src/f32-gemm/gen-inc/5x8inc-minmax-fma3-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002968 "src/f32-gemm/gen-inc/5x16inc-minmax-fma3-broadcast.c",
2969 "src/f32-gemm/gen-inc/5x16s4inc-minmax-fma3-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002970 "src/f32-gemm/gen-inc/6x8inc-minmax-fma3-broadcast.c",
2971 "src/f32-gemm/gen-inc/7x8inc-minmax-fma3-broadcast.c",
2972 "src/f32-gemm/gen-inc/8x8inc-minmax-fma3-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002973 "src/f32-gemm/gen/1x8-minmax-fma3-broadcast.c",
2974 "src/f32-gemm/gen/1x16-minmax-fma3-broadcast.c",
2975 "src/f32-gemm/gen/1x16s4-minmax-fma3-broadcast.c",
2976 "src/f32-gemm/gen/3x16-minmax-fma3-broadcast.c",
2977 "src/f32-gemm/gen/3x16s4-minmax-fma3-broadcast.c",
2978 "src/f32-gemm/gen/4x8-minmax-fma3-broadcast.c",
2979 "src/f32-gemm/gen/4x16-minmax-fma3-broadcast.c",
2980 "src/f32-gemm/gen/4x16s4-minmax-fma3-broadcast.c",
2981 "src/f32-gemm/gen/5x8-minmax-fma3-broadcast.c",
2982 "src/f32-gemm/gen/5x16-minmax-fma3-broadcast.c",
2983 "src/f32-gemm/gen/5x16s4-minmax-fma3-broadcast.c",
2984 "src/f32-gemm/gen/6x8-minmax-fma3-broadcast.c",
2985 "src/f32-gemm/gen/7x8-minmax-fma3-broadcast.c",
2986 "src/f32-gemm/gen/8x8-minmax-fma3-broadcast.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07002987 "src/f32-hswish/gen/hswish-fma3-x8.c",
2988 "src/f32-hswish/gen/hswish-fma3-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002989 "src/f32-igemm/gen/1x8-minmax-fma3-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002990 "src/f32-igemm/gen/1x16-minmax-fma3-broadcast.c",
2991 "src/f32-igemm/gen/1x16s4-minmax-fma3-broadcast.c",
2992 "src/f32-igemm/gen/3x16-minmax-fma3-broadcast.c",
2993 "src/f32-igemm/gen/3x16s4-minmax-fma3-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002994 "src/f32-igemm/gen/4x8-minmax-fma3-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002995 "src/f32-igemm/gen/4x16-minmax-fma3-broadcast.c",
2996 "src/f32-igemm/gen/4x16s4-minmax-fma3-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002997 "src/f32-igemm/gen/5x8-minmax-fma3-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002998 "src/f32-igemm/gen/5x16-minmax-fma3-broadcast.c",
2999 "src/f32-igemm/gen/5x16s4-minmax-fma3-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07003000 "src/f32-igemm/gen/6x8-minmax-fma3-broadcast.c",
3001 "src/f32-igemm/gen/7x8-minmax-fma3-broadcast.c",
3002 "src/f32-igemm/gen/8x8-minmax-fma3-broadcast.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07003003 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x8.c",
3004 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x16.c",
3005 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x24.c",
3006 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x32.c",
3007 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x40.c",
3008 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x48.c",
3009 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x56.c",
3010 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x64.c",
Marat Dukhan84000762020-06-29 18:38:43 -07003011 "src/math/sqrt-fma3-nr1fma.c",
Marat Dukhan84000762020-06-29 18:38:43 -07003012 "src/math/sqrt-fma3-nr1fma1adj.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003013 "src/math/sqrt-fma3-nr2fma.c",
Marat Dukhanfda12b82019-11-21 12:27:59 -08003014]
3015
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003016AVX2_UKERNELS = [
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003017 "src/f32-raddexpminusmax/gen/avx2-p5-x64-acc2.c",
3018 "src/f32-raddexpminusmax/gen/avx2-p5-x64-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003019 "src/f32-raddexpminusmax/gen/avx2-p5-x64.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003020 "src/f32-raddexpminusmax/gen/avx2-p5-x72-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003021 "src/f32-raddexpminusmax/gen/avx2-p5-x72.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003022 "src/f32-raddexpminusmax/gen/avx2-p5-x80-acc2.c",
3023 "src/f32-raddexpminusmax/gen/avx2-p5-x80-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003024 "src/f32-raddexpminusmax/gen/avx2-p5-x80.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003025 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc2.c",
3026 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc3.c",
3027 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc6.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003028 "src/f32-raddexpminusmax/gen/avx2-p5-x96.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003029 "src/f32-raddextexp/gen/avx2-p5-x64-acc2.c",
3030 "src/f32-raddextexp/gen/avx2-p5-x64-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003031 "src/f32-raddextexp/gen/avx2-p5-x64.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003032 "src/f32-raddextexp/gen/avx2-p5-x72-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003033 "src/f32-raddextexp/gen/avx2-p5-x72.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003034 "src/f32-raddextexp/gen/avx2-p5-x80-acc2.c",
3035 "src/f32-raddextexp/gen/avx2-p5-x80-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003036 "src/f32-raddextexp/gen/avx2-p5-x80.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003037 "src/f32-raddextexp/gen/avx2-p5-x96-acc2.c",
3038 "src/f32-raddextexp/gen/avx2-p5-x96-acc3.c",
3039 "src/f32-raddextexp/gen/avx2-p5-x96-acc6.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003040 "src/f32-raddextexp/gen/avx2-p5-x96.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003041 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64-acc2.c",
3042 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003043 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003044 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x72-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003045 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x72.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003046 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80-acc2.c",
3047 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003048 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003049 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc2.c",
3050 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc3.c",
3051 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc6.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003052 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96.c",
Marat Dukhanfa0a4322020-01-06 16:14:29 -08003053 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x8.c",
3054 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x16.c",
3055 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x24.c",
3056 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x32.c",
3057 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x40.c",
3058 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x48.c",
3059 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x56.c",
3060 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x64.c",
3061 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x72.c",
3062 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x80.c",
3063 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x8.c",
3064 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x16.c",
3065 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x24.c",
3066 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x32.c",
3067 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x40.c",
3068 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x48.c",
3069 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x56.c",
3070 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x64.c",
3071 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x72.c",
3072 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x80.c",
3073 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x8.c",
3074 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x16.c",
3075 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x24.c",
3076 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x32.c",
3077 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x40.c",
3078 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x48.c",
3079 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x56.c",
3080 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x64.c",
3081 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x72.c",
3082 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x80.c",
Marat Dukhaned6baaf2020-12-01 15:07:08 -08003083 "src/f32-velu/gen/velu-avx2-rr1-lut4-p4-perm-x8.c",
3084 "src/f32-velu/gen/velu-avx2-rr1-lut4-p4-perm-x16.c",
3085 "src/f32-velu/gen/velu-avx2-rr1-lut4-p4-perm-x24.c",
3086 "src/f32-velu/gen/velu-avx2-rr1-lut4-p4-perm-x32.c",
3087 "src/f32-velu/gen/velu-avx2-rr1-lut4-p4-perm-x40.c",
3088 "src/f32-velu/gen/velu-avx2-rr1-lut4-p4-perm-x48.c",
3089 "src/f32-velu/gen/velu-avx2-rr1-lut4-p4-perm-x56.c",
3090 "src/f32-velu/gen/velu-avx2-rr1-lut4-p4-perm-x64.c",
3091 "src/f32-velu/gen/velu-avx2-rr1-lut4-p4-perm-x72.c",
3092 "src/f32-velu/gen/velu-avx2-rr1-lut4-p4-perm-x80.c",
3093 "src/f32-velu/gen/velu-avx2-rr1-lut8-p4-perm-x8.c",
3094 "src/f32-velu/gen/velu-avx2-rr1-lut8-p4-perm-x16.c",
3095 "src/f32-velu/gen/velu-avx2-rr1-lut8-p4-perm-x24.c",
3096 "src/f32-velu/gen/velu-avx2-rr1-lut8-p4-perm-x32.c",
3097 "src/f32-velu/gen/velu-avx2-rr1-lut8-p4-perm-x40.c",
3098 "src/f32-velu/gen/velu-avx2-rr1-lut8-p4-perm-x48.c",
3099 "src/f32-velu/gen/velu-avx2-rr1-lut8-p4-perm-x56.c",
3100 "src/f32-velu/gen/velu-avx2-rr1-lut8-p4-perm-x64.c",
3101 "src/f32-velu/gen/velu-avx2-rr1-lut8-p4-perm-x72.c",
3102 "src/f32-velu/gen/velu-avx2-rr1-lut8-p4-perm-x80.c",
3103 "src/f32-velu/gen/velu-avx2-rr1-lut16-p3-gather-x8.c",
3104 "src/f32-velu/gen/velu-avx2-rr1-lut16-p3-gather-x16.c",
3105 "src/f32-velu/gen/velu-avx2-rr1-lut16-p3-gather-x24.c",
3106 "src/f32-velu/gen/velu-avx2-rr1-lut16-p3-gather-x32.c",
3107 "src/f32-velu/gen/velu-avx2-rr1-lut16-p3-gather-x40.c",
3108 "src/f32-velu/gen/velu-avx2-rr1-lut16-p3-gather-x48.c",
3109 "src/f32-velu/gen/velu-avx2-rr1-lut16-p3-gather-x56.c",
3110 "src/f32-velu/gen/velu-avx2-rr1-lut16-p3-gather-x64.c",
3111 "src/f32-velu/gen/velu-avx2-rr1-lut16-p3-gather-x72.c",
3112 "src/f32-velu/gen/velu-avx2-rr1-lut16-p3-gather-x80.c",
3113 "src/f32-velu/gen/velu-avx2-rr1-p6-x8.c",
3114 "src/f32-velu/gen/velu-avx2-rr1-p6-x16.c",
3115 "src/f32-velu/gen/velu-avx2-rr1-p6-x24.c",
3116 "src/f32-velu/gen/velu-avx2-rr1-p6-x32.c",
3117 "src/f32-velu/gen/velu-avx2-rr1-p6-x40.c",
3118 "src/f32-velu/gen/velu-avx2-rr1-p6-x48.c",
3119 "src/f32-velu/gen/velu-avx2-rr1-p6-x56.c",
3120 "src/f32-velu/gen/velu-avx2-rr1-p6-x64.c",
3121 "src/f32-velu/gen/velu-avx2-rr1-p6-x72.c",
3122 "src/f32-velu/gen/velu-avx2-rr1-p6-x80.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003123 "src/f32-vscaleexpminusmax/gen/avx2-p5-x8.c",
3124 "src/f32-vscaleexpminusmax/gen/avx2-p5-x16.c",
3125 "src/f32-vscaleexpminusmax/gen/avx2-p5-x24.c",
3126 "src/f32-vscaleexpminusmax/gen/avx2-p5-x32.c",
3127 "src/f32-vscaleexpminusmax/gen/avx2-p5-x40.c",
3128 "src/f32-vscaleexpminusmax/gen/avx2-p5-x48.c",
3129 "src/f32-vscaleexpminusmax/gen/avx2-p5-x56.c",
3130 "src/f32-vscaleexpminusmax/gen/avx2-p5-x64.c",
3131 "src/f32-vscaleexpminusmax/gen/avx2-p5-x72.c",
3132 "src/f32-vscaleexpminusmax/gen/avx2-p5-x80.c",
3133 "src/f32-vscaleexpminusmax/gen/avx2-p5-x88.c",
3134 "src/f32-vscaleexpminusmax/gen/avx2-p5-x96.c",
3135 "src/f32-vscaleextexp/gen/avx2-p5-x8.c",
3136 "src/f32-vscaleextexp/gen/avx2-p5-x16.c",
3137 "src/f32-vscaleextexp/gen/avx2-p5-x24.c",
3138 "src/f32-vscaleextexp/gen/avx2-p5-x32.c",
3139 "src/f32-vscaleextexp/gen/avx2-p5-x40.c",
3140 "src/f32-vscaleextexp/gen/avx2-p5-x48.c",
3141 "src/f32-vscaleextexp/gen/avx2-p5-x56.c",
3142 "src/f32-vscaleextexp/gen/avx2-p5-x64.c",
3143 "src/f32-vscaleextexp/gen/avx2-p5-x72.c",
3144 "src/f32-vscaleextexp/gen/avx2-p5-x80.c",
3145 "src/f32-vscaleextexp/gen/avx2-p5-x88.c",
3146 "src/f32-vscaleextexp/gen/avx2-p5-x96.c",
Marat Dukhanb7633f22020-11-20 16:34:56 -08003147 "src/math/exp-avx2-rr2-lut8-p3-perm.c",
3148 "src/math/exp-avx2-rr2-lut8-p4-perm.c",
3149 "src/math/exp-avx2-rr2-p5.c",
Marat Dukhande390d42020-11-29 19:32:18 -08003150 "src/math/expm1minus-avx2-rr1-lut4-p4-perm.c",
3151 "src/math/expm1minus-avx2-rr1-lut8-p4-perm.c",
3152 "src/math/expm1minus-avx2-rr1-lut16-p3-gather.c",
3153 "src/math/expm1minus-avx2-rr1-p6.c",
Frank Barcharde7223ee2020-12-04 19:04:01 -08003154 "src/math/expminus-avx2-rr2-p5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003155 "src/math/extexp-avx2-p5.c",
3156 "src/math/sigmoid-avx2-rr1-lut64-p2-gather-div.c",
3157 "src/math/sigmoid-avx2-rr1-lut64-p2-gather-nr1fma.c",
3158 "src/math/sigmoid-avx2-rr1-lut64-p2-gather-nr2fma.c",
3159 "src/math/sigmoid-avx2-rr1-lut64-p2-gather-nr2fma1adj.c",
3160 "src/math/sigmoid-avx2-rr1-p5-div.c",
3161 "src/math/sigmoid-avx2-rr1-p5-nr1fma.c",
3162 "src/math/sigmoid-avx2-rr1-p5-nr2fma.c",
3163 "src/math/sigmoid-avx2-rr2-lut64-p2-gather-div.c",
3164 "src/math/sigmoid-avx2-rr2-lut64-p2-gather-nr1fma.c",
3165 "src/math/sigmoid-avx2-rr2-lut64-p2-gather-nr2fma.c",
3166 "src/math/sigmoid-avx2-rr2-lut64-p2-gather-nr2fma1adj.c",
3167 "src/math/sigmoid-avx2-rr2-p5-div.c",
3168 "src/math/sigmoid-avx2-rr2-p5-nr1fma.c",
3169 "src/math/sigmoid-avx2-rr2-p5-nr2fma.c",
Marat Dukhand65a1522020-08-04 19:28:18 -07003170 "src/qs8-dwconv/gen/up8x9-minmax-avx2-mul32.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003171 "src/qs8-dwconv/gen/up16x9-minmax-avx2-mul16.c",
Marat Dukhand65a1522020-08-04 19:28:18 -07003172 "src/qs8-dwconv/gen/up16x9-minmax-avx2-mul32.c",
3173 "src/qs8-dwconv/gen/up24x9-minmax-avx2-mul32.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003174 "src/qs8-dwconv/gen/up32x9-minmax-avx2-mul16.c",
Marat Dukhand65a1522020-08-04 19:28:18 -07003175 "src/qs8-dwconv/gen/up32x9-minmax-avx2-mul32.c",
Marat Dukhan12809522020-08-02 22:23:51 -07003176 "src/qs8-gemm/gen/1x8c8-minmax-avx2.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07003177 "src/qs8-gemm/gen/1x8c8-xw-minmax-avx2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003178 "src/qs8-gemm/gen/2x8c8-minmax-avx2.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07003179 "src/qs8-gemm/gen/2x8c8-xw-minmax-avx2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003180 "src/qs8-gemm/gen/3x8c8-minmax-avx2.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07003181 "src/qs8-gemm/gen/3x8c8-xw-minmax-avx2.c",
Marat Dukhan12809522020-08-02 22:23:51 -07003182 "src/qs8-igemm/gen/1x8c8-minmax-avx2.c",
3183 "src/qs8-igemm/gen/2x8c8-minmax-avx2.c",
3184 "src/qs8-igemm/gen/3x8c8-minmax-avx2.c",
Marat Dukhane6dc0b62020-09-08 23:57:14 -07003185 "src/qs8-vadd/gen/minmax-avx2-mul32-ld64-x8.c",
3186 "src/qs8-vadd/gen/minmax-avx2-mul32-ld64-x16.c",
3187 "src/qs8-vadd/gen/minmax-avx2-mul32-ld64-x24.c",
3188 "src/qs8-vadd/gen/minmax-avx2-mul32-ld64-x32.c",
3189 "src/qs8-vaddc/gen/minmax-avx2-mul32-ld64-x8.c",
3190 "src/qs8-vaddc/gen/minmax-avx2-mul32-ld64-x16.c",
3191 "src/qs8-vaddc/gen/minmax-avx2-mul32-ld64-x24.c",
3192 "src/qs8-vaddc/gen/minmax-avx2-mul32-ld64-x32.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003193]
3194
Marat Dukhan08c4a432019-10-03 09:29:21 -07003195AVX512F_UKERNELS = [
Marat Dukhan5c5fa962020-03-10 18:38:33 -07003196 "src/f32-clamp/gen/avx512f-x16.c",
3197 "src/f32-clamp/gen/avx512f-x32.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07003198 "src/f32-dwconv/gen/up16x4-minmax-avx512f-acc2.c",
3199 "src/f32-dwconv/gen/up16x4-minmax-avx512f.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07003200 "src/f32-dwconv/gen/up16x9-minmax-avx512f-acc2.c",
3201 "src/f32-dwconv/gen/up16x9-minmax-avx512f.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07003202 "src/f32-dwconv/gen/up16x25-minmax-avx512f-acc2.c",
3203 "src/f32-dwconv/gen/up16x25-minmax-avx512f.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003204 "src/f32-dwconv/gen/up32x4-minmax-avx512f-acc2.c",
3205 "src/f32-dwconv/gen/up32x4-minmax-avx512f.c",
3206 "src/f32-dwconv/gen/up32x9-minmax-avx512f-acc2.c",
3207 "src/f32-dwconv/gen/up32x9-minmax-avx512f.c",
3208 "src/f32-dwconv/gen/up32x25-minmax-avx512f-acc2.c",
3209 "src/f32-dwconv/gen/up32x25-minmax-avx512f.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07003210 "src/f32-gemm/gen-inc/1x16inc-minmax-avx512f-broadcast.c",
3211 "src/f32-gemm/gen-inc/4x16inc-minmax-avx512f-broadcast.c",
3212 "src/f32-gemm/gen-inc/5x16inc-minmax-avx512f-broadcast.c",
3213 "src/f32-gemm/gen-inc/6x16inc-minmax-avx512f-broadcast.c",
3214 "src/f32-gemm/gen-inc/7x16inc-minmax-avx512f-broadcast.c",
3215 "src/f32-gemm/gen-inc/8x16inc-minmax-avx512f-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003216 "src/f32-gemm/gen/1x16-minmax-avx512f-broadcast.c",
3217 "src/f32-gemm/gen/4x16-minmax-avx512f-broadcast.c",
3218 "src/f32-gemm/gen/5x16-minmax-avx512f-broadcast.c",
3219 "src/f32-gemm/gen/6x16-minmax-avx512f-broadcast.c",
3220 "src/f32-gemm/gen/7x16-minmax-avx512f-broadcast.c",
3221 "src/f32-gemm/gen/8x16-minmax-avx512f-broadcast.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07003222 "src/f32-hswish/gen/hswish-avx512f-x16.c",
3223 "src/f32-hswish/gen/hswish-avx512f-x32.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07003224 "src/f32-igemm/gen/1x16-minmax-avx512f-broadcast.c",
3225 "src/f32-igemm/gen/4x16-minmax-avx512f-broadcast.c",
3226 "src/f32-igemm/gen/5x16-minmax-avx512f-broadcast.c",
3227 "src/f32-igemm/gen/6x16-minmax-avx512f-broadcast.c",
3228 "src/f32-igemm/gen/7x16-minmax-avx512f-broadcast.c",
3229 "src/f32-igemm/gen/8x16-minmax-avx512f-broadcast.c",
Marat Dukhan90eca0a2020-03-11 00:52:23 -07003230 "src/f32-prelu/gen/avx512f-2x16.c",
3231 "src/f32-prelu/gen/avx512f-2x32.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003232 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128-acc2.c",
3233 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003234 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003235 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x144-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003236 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x144.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003237 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160-acc2.c",
3238 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003239 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003240 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc2.c",
3241 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc3.c",
3242 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc6.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003243 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003244 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128-acc2.c",
3245 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003246 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003247 "src/f32-raddextexp/gen/avx512f-p5-scalef-x144-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003248 "src/f32-raddextexp/gen/avx512f-p5-scalef-x144.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003249 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160-acc2.c",
3250 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003251 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003252 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc2.c",
3253 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc3.c",
3254 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc6.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003255 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003256 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128-acc2.c",
3257 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003258 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003259 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x144-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003260 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x144.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003261 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160-acc2.c",
3262 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003263 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003264 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc2.c",
3265 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc3.c",
3266 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc6.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003267 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07003268 "src/f32-relu/gen/avx512f-x16.c",
3269 "src/f32-relu/gen/avx512f-x32.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003270 "src/f32-rmax/avx512f.c",
Marat Dukhand9ca7e62020-09-23 23:45:29 -07003271 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x16.c",
3272 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x32.c",
3273 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x48.c",
3274 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x64.c",
3275 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x80.c",
3276 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x96.c",
3277 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x112.c",
3278 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x128.c",
3279 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x16.c",
3280 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x32.c",
3281 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x48.c",
3282 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x64.c",
3283 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x80.c",
3284 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x96.c",
3285 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x112.c",
3286 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003287 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x16.c",
3288 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x32.c",
3289 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x48.c",
3290 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x64.c",
3291 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x80.c",
3292 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x96.c",
3293 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x112.c",
3294 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x128.c",
3295 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x16.c",
3296 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x32.c",
3297 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x48.c",
3298 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x64.c",
3299 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x80.c",
3300 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x96.c",
3301 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x112.c",
3302 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x128.c",
Marat Dukhand9ca7e62020-09-23 23:45:29 -07003303 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x16.c",
3304 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x32.c",
3305 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x48.c",
3306 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x64.c",
3307 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x80.c",
3308 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x96.c",
3309 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x112.c",
3310 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x128.c",
3311 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x16.c",
3312 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x32.c",
3313 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x48.c",
3314 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x64.c",
3315 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x80.c",
3316 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x96.c",
3317 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x112.c",
3318 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x128.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07003319 "src/f32-vbinary/gen/vadd-minmax-avx512f-x16.c",
3320 "src/f32-vbinary/gen/vadd-minmax-avx512f-x32.c",
3321 "src/f32-vbinary/gen/vaddc-minmax-avx512f-x16.c",
3322 "src/f32-vbinary/gen/vaddc-minmax-avx512f-x32.c",
3323 "src/f32-vbinary/gen/vdiv-minmax-avx512f-x16.c",
3324 "src/f32-vbinary/gen/vdiv-minmax-avx512f-x32.c",
3325 "src/f32-vbinary/gen/vdivc-minmax-avx512f-x16.c",
3326 "src/f32-vbinary/gen/vdivc-minmax-avx512f-x32.c",
Marat Dukhan9a88efe2019-12-10 15:54:24 -08003327 "src/f32-vbinary/gen/vmax-avx512f-x16.c",
3328 "src/f32-vbinary/gen/vmax-avx512f-x32.c",
3329 "src/f32-vbinary/gen/vmaxc-avx512f-x16.c",
3330 "src/f32-vbinary/gen/vmaxc-avx512f-x32.c",
3331 "src/f32-vbinary/gen/vmin-avx512f-x16.c",
3332 "src/f32-vbinary/gen/vmin-avx512f-x32.c",
3333 "src/f32-vbinary/gen/vminc-avx512f-x16.c",
3334 "src/f32-vbinary/gen/vminc-avx512f-x32.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07003335 "src/f32-vbinary/gen/vmul-minmax-avx512f-x16.c",
3336 "src/f32-vbinary/gen/vmul-minmax-avx512f-x32.c",
3337 "src/f32-vbinary/gen/vmulc-minmax-avx512f-x16.c",
3338 "src/f32-vbinary/gen/vmulc-minmax-avx512f-x32.c",
3339 "src/f32-vbinary/gen/vrdivc-minmax-avx512f-x16.c",
3340 "src/f32-vbinary/gen/vrdivc-minmax-avx512f-x32.c",
3341 "src/f32-vbinary/gen/vrsubc-minmax-avx512f-x16.c",
3342 "src/f32-vbinary/gen/vrsubc-minmax-avx512f-x32.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07003343 "src/f32-vbinary/gen/vsqrdiff-avx512f-x16.c",
3344 "src/f32-vbinary/gen/vsqrdiff-avx512f-x32.c",
3345 "src/f32-vbinary/gen/vsqrdiffc-avx512f-x16.c",
3346 "src/f32-vbinary/gen/vsqrdiffc-avx512f-x32.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07003347 "src/f32-vbinary/gen/vsub-minmax-avx512f-x16.c",
3348 "src/f32-vbinary/gen/vsub-minmax-avx512f-x32.c",
3349 "src/f32-vbinary/gen/vsubc-minmax-avx512f-x16.c",
3350 "src/f32-vbinary/gen/vsubc-minmax-avx512f-x32.c",
Marat Dukhaned6baaf2020-12-01 15:07:08 -08003351 "src/f32-velu/gen/velu-avx512f-rr1-lut16-p3-perm-x16.c",
3352 "src/f32-velu/gen/velu-avx512f-rr1-lut16-p3-perm-x32.c",
3353 "src/f32-velu/gen/velu-avx512f-rr1-lut16-p3-perm-x48.c",
3354 "src/f32-velu/gen/velu-avx512f-rr1-lut16-p3-perm-x64.c",
3355 "src/f32-velu/gen/velu-avx512f-rr1-lut16-p3-perm-x80.c",
3356 "src/f32-velu/gen/velu-avx512f-rr1-lut16-p3-perm-x96.c",
3357 "src/f32-velu/gen/velu-avx512f-rr1-lut16-p3-perm-x112.c",
3358 "src/f32-velu/gen/velu-avx512f-rr1-lut16-p3-perm-x128.c",
3359 "src/f32-velu/gen/velu-avx512f-rr1-p6-x16.c",
3360 "src/f32-velu/gen/velu-avx512f-rr1-p6-x32.c",
3361 "src/f32-velu/gen/velu-avx512f-rr1-p6-x48.c",
3362 "src/f32-velu/gen/velu-avx512f-rr1-p6-x64.c",
3363 "src/f32-velu/gen/velu-avx512f-rr1-p6-x80.c",
3364 "src/f32-velu/gen/velu-avx512f-rr1-p6-x96.c",
3365 "src/f32-velu/gen/velu-avx512f-rr1-p6-x112.c",
3366 "src/f32-velu/gen/velu-avx512f-rr1-p6-x128.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07003367 "src/f32-vlrelu/gen/vlrelu-avx512f-x16.c",
3368 "src/f32-vlrelu/gen/vlrelu-avx512f-x32.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003369 "src/f32-vrnd/gen/vrndd-avx512f-x16.c",
3370 "src/f32-vrnd/gen/vrndd-avx512f-x32.c",
3371 "src/f32-vrnd/gen/vrndne-avx512f-x16.c",
3372 "src/f32-vrnd/gen/vrndne-avx512f-x32.c",
3373 "src/f32-vrnd/gen/vrndu-avx512f-x16.c",
3374 "src/f32-vrnd/gen/vrndu-avx512f-x32.c",
3375 "src/f32-vrnd/gen/vrndz-avx512f-x16.c",
3376 "src/f32-vrnd/gen/vrndz-avx512f-x32.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07003377 "src/f32-vscale/avx512f-x64.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003378 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x16.c",
3379 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x32.c",
3380 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x48.c",
3381 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x64.c",
3382 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x80.c",
3383 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x96.c",
3384 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x112.c",
3385 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x128.c",
3386 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x144.c",
3387 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x160.c",
3388 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x176.c",
3389 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x192.c",
3390 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x16.c",
3391 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x32.c",
3392 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x48.c",
3393 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x64.c",
3394 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x80.c",
3395 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x96.c",
3396 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x112.c",
3397 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x128.c",
3398 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x144.c",
3399 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x160.c",
3400 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x176.c",
3401 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x192.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07003402 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x16.c",
3403 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x32.c",
3404 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x48.c",
3405 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x64.c",
3406 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x80.c",
3407 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x96.c",
3408 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x112.c",
3409 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x128.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07003410 "src/f32-vunary/gen/vabs-avx512f-x16.c",
3411 "src/f32-vunary/gen/vabs-avx512f-x32.c",
3412 "src/f32-vunary/gen/vneg-avx512f-x16.c",
3413 "src/f32-vunary/gen/vneg-avx512f-x32.c",
3414 "src/f32-vunary/gen/vsqr-avx512f-x16.c",
3415 "src/f32-vunary/gen/vsqr-avx512f-x32.c",
Marat Dukhanb7633f22020-11-20 16:34:56 -08003416 "src/math/exp-avx512f-rr2-lut16-p3-perm-scalef.c",
3417 "src/math/exp-avx512f-rr2-lut16-p3-perm.c",
3418 "src/math/exp-avx512f-rr2-lut32-p2-perm2-scalef.c",
3419 "src/math/exp-avx512f-rr2-lut32-p2-perm2.c",
3420 "src/math/exp-avx512f-rr2-p5-scalef.c",
3421 "src/math/exp-avx512f-rr2-p5.c",
Marat Dukhande390d42020-11-29 19:32:18 -08003422 "src/math/expm1minus-avx512f-rr1-lut16-p3-perm.c",
3423 "src/math/expm1minus-avx512f-rr1-p6.c",
Marat Dukhan98ba4412019-10-23 02:14:28 -07003424 "src/math/extexp-avx512f-p5.c",
Marat Dukhan6a34c5f2020-09-22 21:44:15 -07003425 "src/math/sigmoid-avx512f-rr1-lut16-p3-perm-scalef-div.c",
Marat Dukhan6a34c5f2020-09-22 21:44:15 -07003426 "src/math/sigmoid-avx512f-rr1-lut16-p3-perm-scalef-nr1fma.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003427 "src/math/sigmoid-avx512f-rr1-lut16-p3-perm-scalef-nr1fma1adj.c",
Marat Dukhan6a34c5f2020-09-22 21:44:15 -07003428 "src/math/sigmoid-avx512f-rr1-lut32-p2-perm2-scalef-div.c",
Marat Dukhan6a34c5f2020-09-22 21:44:15 -07003429 "src/math/sigmoid-avx512f-rr1-lut32-p2-perm2-scalef-nr1fma.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003430 "src/math/sigmoid-avx512f-rr1-lut32-p2-perm2-scalef-nr1fma1adj.c",
Marat Dukhan36173d22020-10-15 17:14:26 -07003431 "src/math/sigmoid-avx512f-rr1-lut64-p2-gather-scalef-div.c",
Marat Dukhan36173d22020-10-15 17:14:26 -07003432 "src/math/sigmoid-avx512f-rr1-lut64-p2-gather-scalef-nr1fma.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003433 "src/math/sigmoid-avx512f-rr1-lut64-p2-gather-scalef-nr1fma1adj.c",
3434 "src/math/sigmoid-avx512f-rr1-p5-scalef-div.c",
3435 "src/math/sigmoid-avx512f-rr1-p5-scalef-nr1fma.c",
3436 "src/math/sigmoid-avx512f-rr1-p5-scalef-nr1fma1adj.c",
3437 "src/math/sigmoid-avx512f-rr2-lut16-p3-perm-scalef-div.c",
3438 "src/math/sigmoid-avx512f-rr2-lut16-p3-perm-scalef-nr1fma.c",
3439 "src/math/sigmoid-avx512f-rr2-lut16-p3-perm-scalef-nr1fma1adj.c",
3440 "src/math/sigmoid-avx512f-rr2-lut32-p2-perm2-scalef-div.c",
3441 "src/math/sigmoid-avx512f-rr2-lut32-p2-perm2-scalef-nr1fma.c",
3442 "src/math/sigmoid-avx512f-rr2-lut32-p2-perm2-scalef-nr1fma1adj.c",
Marat Dukhan36173d22020-10-15 17:14:26 -07003443 "src/math/sigmoid-avx512f-rr2-lut64-p2-gather-scalef-div.c",
Marat Dukhan36173d22020-10-15 17:14:26 -07003444 "src/math/sigmoid-avx512f-rr2-lut64-p2-gather-scalef-nr1fma.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003445 "src/math/sigmoid-avx512f-rr2-lut64-p2-gather-scalef-nr1fma1adj.c",
3446 "src/math/sigmoid-avx512f-rr2-p5-scalef-div.c",
3447 "src/math/sigmoid-avx512f-rr2-p5-scalef-nr1fma.c",
3448 "src/math/sigmoid-avx512f-rr2-p5-scalef-nr1fma1adj.c",
Marat Dukhan84000762020-06-29 18:38:43 -07003449 "src/math/sqrt-avx512f-nr1fma.c",
Marat Dukhan84000762020-06-29 18:38:43 -07003450 "src/math/sqrt-avx512f-nr1fma1adj.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003451 "src/math/sqrt-avx512f-nr2fma.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003452]
3453
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07003454AVX512SKX_UKERNELS = [
Marat Dukhan2ffc5e62020-09-06 22:33:38 -07003455 "src/qs8-dwconv/gen/up16x9-minmax-avx512skx-mul32.c",
3456 "src/qs8-dwconv/gen/up32x9-minmax-avx512skx-mul32.c",
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07003457 "src/qs8-gemm/gen/1x16c8-minmax-avx512skx.c",
3458 "src/qs8-gemm/gen/2x16c8-minmax-avx512skx.c",
3459 "src/qs8-gemm/gen/3x16c8-minmax-avx512skx.c",
3460 "src/qs8-gemm/gen/4x16c8-minmax-avx512skx.c",
3461 "src/qs8-igemm/gen/1x16c8-minmax-avx512skx.c",
3462 "src/qs8-igemm/gen/2x16c8-minmax-avx512skx.c",
3463 "src/qs8-igemm/gen/3x16c8-minmax-avx512skx.c",
3464 "src/qs8-igemm/gen/4x16c8-minmax-avx512skx.c",
3465]
3466
Frank Barchardbcedc082020-08-17 18:00:51 -07003467WASM32_ASM_UKERNELS = [
3468 "src/f32-relu/wasm_shr_x1.S",
3469 "src/f32-relu/wasm_shr_x2.S",
3470 "src/f32-relu/wasm_shr_x4.S",
3471]
3472
Marat Dukhan08c4a432019-10-03 09:29:21 -07003473AARCH32_ASM_UKERNELS = [
Marat Dukhan32f93812020-05-17 20:31:21 -07003474 "src/f32-gemm/4x4-aarch32-vfp-ld64.S",
Marat Dukhan3b98f6b2020-05-17 10:09:22 -07003475 "src/f32-gemm/4x4-minmax-aarch32-vfp-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003476 "src/f32-gemm/4x8-minmax-aarch32-neon-cortex-a53.S",
3477 "src/f32-gemm/4x8-minmax-aarch32-neon-cortex-a55.S",
Frank Barchard490febe2020-07-16 18:42:17 -07003478 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-cortex-a7.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003479 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-cortex-a75.S",
Frank Barchard569561d2020-06-17 13:11:12 -07003480 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-ld64.S",
Frank Barchard490febe2020-07-16 18:42:17 -07003481 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-pld-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003482 "src/f32-igemm/4x8-minmax-aarch32-neon-cortex-a53.S",
3483 "src/f32-igemm/4x8-minmax-aarch32-neon-cortex-a55.S",
Frank Barchard490febe2020-07-16 18:42:17 -07003484 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-cortex-a7.S",
3485 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-cortex-a75.S",
3486 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-ld64.S",
3487 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-pld-cortex-a75.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003488]
3489
3490AARCH64_ASM_UKERNELS = [
Frank Barchardbddfbcd2020-04-15 12:32:41 -07003491 "src/f16-gemm/gen-inc/1x8inc-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchard04336c12020-10-22 16:48:55 -07003492 "src/f16-gemm/gen-inc/1x16inc-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchardbddfbcd2020-04-15 12:32:41 -07003493 "src/f16-gemm/gen-inc/4x8inc-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchard04336c12020-10-22 16:48:55 -07003494 "src/f16-gemm/gen-inc/4x16inc-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchardbddfbcd2020-04-15 12:32:41 -07003495 "src/f16-gemm/gen-inc/6x8inc-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchard04336c12020-10-22 16:48:55 -07003496 "src/f16-gemm/gen-inc/6x16inc-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchard3b8e5662020-04-20 12:12:53 -07003497 "src/f16-gemm/gen-inc/8x8inc-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchard04336c12020-10-22 16:48:55 -07003498 "src/f16-gemm/gen/1x8-minmax-aarch64-neonfp16arith-ld64.S",
3499 "src/f16-gemm/gen/1x16-minmax-aarch64-neonfp16arith-ld32.S",
3500 "src/f16-gemm/gen/4x8-minmax-aarch64-neonfp16arith-ld64.S",
3501 "src/f16-gemm/gen/4x16-minmax-aarch64-neonfp16arith-ld32.S",
3502 "src/f16-gemm/gen/6x8-minmax-aarch64-neonfp16arith-ld64.S",
3503 "src/f16-gemm/gen/6x16-minmax-aarch64-neonfp16arith-ld32.S",
3504 "src/f16-gemm/gen/8x8-minmax-aarch64-neonfp16arith-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003505 "src/f32-dwconv/up4x9-minmax-aarch64-neonfma-cortex-a55.S",
3506 "src/f32-dwconv/up4x9-minmax-aarch64-neonfma.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003507 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a53.S",
3508 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a57.S",
3509 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a75.S",
Frank Barchard04336c12020-10-22 16:48:55 -07003510 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-ld64.S",
3511 "src/f32-gemm/gen-inc/1x12inc-minmax-aarch64-neonfma-cortex-a53.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003512 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a53.S",
3513 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a55.S",
3514 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a57.S",
3515 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003516 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-ld64.S",
Frank Barchard04336c12020-10-22 16:48:55 -07003517 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-ld128.S",
3518 "src/f32-gemm/gen-inc/4x12inc-minmax-aarch64-neonfma-cortex-a53.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003519 "src/f32-gemm/gen-inc/5x8inc-minmax-aarch64-neonfma-cortex-a57.S",
3520 "src/f32-gemm/gen-inc/5x8inc-minmax-aarch64-neonfma-cortex-a75.S",
3521 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a53.S",
3522 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a55.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003523 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a57.S",
Frank Barchard04336c12020-10-22 16:48:55 -07003524 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a73.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003525 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003526 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-ld64.S",
Frank Barchard04336c12020-10-22 16:48:55 -07003527 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-ld128.S",
3528 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a53.S",
3529 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a57.S",
3530 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a75.S",
3531 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-ld64.S",
3532 "src/f32-gemm/gen/1x12-minmax-aarch64-neonfma-cortex-a53.S",
3533 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a53.S",
3534 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a55.S",
3535 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a57.S",
3536 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a75.S",
3537 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-ld64.S",
3538 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-ld128.S",
3539 "src/f32-gemm/gen/4x12-minmax-aarch64-neonfma-cortex-a53.S",
3540 "src/f32-gemm/gen/5x8-minmax-aarch64-neonfma-cortex-a57.S",
3541 "src/f32-gemm/gen/5x8-minmax-aarch64-neonfma-cortex-a75.S",
3542 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a53.S",
3543 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a55.S",
3544 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a57.S",
3545 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a73.S",
3546 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a75.S",
3547 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-ld64.S",
3548 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-ld128.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003549 "src/f32-igemm/1x8-minmax-aarch64-neonfma-cortex-a53.S",
Frank Barchard04336c12020-10-22 16:48:55 -07003550 "src/f32-igemm/1x12-minmax-aarch64-neonfma-cortex-a53.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003551 "src/f32-igemm/4x8-minmax-aarch64-neonfma-cortex-a53.S",
3552 "src/f32-igemm/4x8-minmax-aarch64-neonfma-cortex-a55.S",
Frank Barchard04336c12020-10-22 16:48:55 -07003553 "src/f32-igemm/4x12-minmax-aarch64-neonfma-cortex-a53.S",
3554 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a53.S",
3555 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a55.S",
3556 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a73.S",
3557 "src/f32-igemm/gen/1x8-minmax-aarch64-neonfma-cortex-a57.S",
3558 "src/f32-igemm/gen/1x8-minmax-aarch64-neonfma-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003559 "src/f32-igemm/gen/4x8-minmax-aarch64-neonfma-cortex-a57.S",
3560 "src/f32-igemm/gen/4x8-minmax-aarch64-neonfma-cortex-a75.S",
3561 "src/f32-igemm/gen/5x8-minmax-aarch64-neonfma-cortex-a57.S",
3562 "src/f32-igemm/gen/5x8-minmax-aarch64-neonfma-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003563 "src/f32-igemm/gen/6x8-minmax-aarch64-neonfma-cortex-a57.S",
3564 "src/f32-igemm/gen/6x8-minmax-aarch64-neonfma-cortex-a75.S",
Frank Barchard4a4be4e2021-03-01 15:00:28 -08003565 "src/qs8-gemm/1x16c4-aarch64-neondot-ld32.S",
Frank Barchardf1fd89e2020-10-06 23:27:11 -07003566 "src/qs8-gemm/1x16c4-aarch64-neondot-ld64.S",
Frank Barchardf2742c42020-10-12 21:16:46 -07003567 "src/qs8-gemm/4x16c4-aarch64-neondot-cortex-a55.S",
Frank Barchardd76a37b2021-03-02 11:36:16 -08003568 "src/qs8-gemm/4x16c4-aarch64-neondot-ld64.S",
Frank Barchardda78da12021-03-02 14:28:00 -08003569 "src/qs8-gemm/4x16c4-aarch64-neondot-ld32.S",
Frank Barchard89e12f82021-03-08 11:42:44 -08003570 "src/qs8-igemm/4x16c4-aarch64-neondot-cortex-a55.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003571]
3572
Marat Dukhan1b354632020-03-23 12:50:22 -07003573INTERNAL_MICROKERNEL_HDRS = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003574 "src/xnnpack/argmaxpool.h",
3575 "src/xnnpack/avgpool.h",
3576 "src/xnnpack/clamp.h",
3577 "src/xnnpack/common.h",
3578 "src/xnnpack/conv.h",
Yury Kartynnike7841862020-11-04 18:22:18 -08003579 "src/xnnpack/depthtospace.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003580 "src/xnnpack/dwconv.h",
Frank Barchard04336c12020-10-22 16:48:55 -07003581 "src/xnnpack/fill.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003582 "src/xnnpack/gavgpool.h",
3583 "src/xnnpack/gemm.h",
3584 "src/xnnpack/hswish.h",
Marat Dukhan660fd192020-03-10 04:55:30 -07003585 "src/xnnpack/ibilinear.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003586 "src/xnnpack/igemm.h",
Marat Dukhancfb31342019-12-05 10:42:57 -08003587 "src/xnnpack/intrinsics-polyfill.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003588 "src/xnnpack/lut.h",
3589 "src/xnnpack/math.h",
3590 "src/xnnpack/maxpool.h",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003591 "src/xnnpack/memory.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003592 "src/xnnpack/packx.h",
3593 "src/xnnpack/pad.h",
3594 "src/xnnpack/params.h",
3595 "src/xnnpack/pavgpool.h",
3596 "src/xnnpack/ppmm.h",
3597 "src/xnnpack/prelu.h",
Marat Dukhan97579532019-10-18 16:40:39 -07003598 "src/xnnpack/raddexpminusmax.h",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07003599 "src/xnnpack/raddextexp.h",
Marat Dukhan97579532019-10-18 16:40:39 -07003600 "src/xnnpack/raddstoreexpminusmax.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003601 "src/xnnpack/rmax.h",
3602 "src/xnnpack/scalar-utils.h",
3603 "src/xnnpack/spmm.h",
3604 "src/xnnpack/unpool.h",
3605 "src/xnnpack/vadd.h",
Marat Dukhan1e782c42019-11-21 17:02:40 -08003606 "src/xnnpack/vbinary.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003607 "src/xnnpack/vmulcaddc.h",
Marat Dukhan05ac8e32019-10-21 15:39:33 -07003608 "src/xnnpack/vscale.h",
Marat Dukhan97579532019-10-18 16:40:39 -07003609 "src/xnnpack/vscaleexpminusmax.h",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07003610 "src/xnnpack/vscaleextexp.h",
Marat Dukhan1e782c42019-11-21 17:02:40 -08003611 "src/xnnpack/vunary.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003612 "src/xnnpack/zip.h",
Marat Dukhan1b354632020-03-23 12:50:22 -07003613]
3614
3615INTERNAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003616 "include/xnnpack.h",
3617 "src/xnnpack/allocator.h",
3618 "src/xnnpack/compute.h",
3619 "src/xnnpack/im2col.h",
3620 "src/xnnpack/indirection.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003621 "src/xnnpack/math-stubs.h",
Chao Mei6ddfc602020-05-13 22:29:36 -07003622 "src/xnnpack/memory-planner.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003623 "src/xnnpack/operator.h",
3624 "src/xnnpack/pack.h",
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07003625 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003626 "src/xnnpack/requantization-stubs.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003627 "src/xnnpack/requantization.h",
Marat Dukhan1d75a542020-02-03 12:23:01 -08003628 "src/xnnpack/subgraph.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003629]
3630
Marat Dukhan1b354632020-03-23 12:50:22 -07003631ACCURACY_EVAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003632 "src/xnnpack/math-stubs.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003633]
3634
Marat Dukhan1b354632020-03-23 12:50:22 -07003635MICROKERNEL_BENCHMARK_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003636 "include/xnnpack.h",
Frank Barchard35db7d02020-10-26 13:37:34 -07003637 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003638]
3639
Marat Dukhan1b354632020-03-23 12:50:22 -07003640MICROKERNEL_TEST_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Frank Barchard35db7d02020-10-26 13:37:34 -07003641 "include/xnnpack.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003642 "src/xnnpack/isa-checks.h",
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07003643 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003644 "src/xnnpack/requantization.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003645]
3646
3647OPERATOR_TEST_PARAMS_HDRS = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003648 "src/xnnpack/common.h",
Frank Barchard04336c12020-10-22 16:48:55 -07003649 "src/xnnpack/params.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003650]
3651
3652WEIGHTS_PACK_HDRS = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003653 "src/xnnpack/compute.h",
Frank Barchard04336c12020-10-22 16:48:55 -07003654 "src/xnnpack/operator.h",
3655 "src/xnnpack/pack.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003656]
3657
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07003658LOGGING_COPTS = select({
3659 # No logging in optimized mode
3660 ":optimized_build": ["-DXNN_LOG_LEVEL=0"],
3661 # Full logging in debug mode
3662 ":debug_build": ["-DXNN_LOG_LEVEL=5"],
3663 # Error-only logging in default (fastbuild) mode
3664 "//conditions:default": ["-DXNN_LOG_LEVEL=2"],
3665})
3666
Marat Dukhan3b59de22020-06-03 20:15:19 -07003667LOGGING_SRCS = select({
3668 # No logging in optimized mode
3669 ":optimized_build": [],
3670 "//conditions:default": [
3671 "src/operator-strings.c",
3672 "src/subgraph-strings.c",
3673 ],
3674})
3675
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07003676LOGGING_HDRS = [
3677 "src/xnnpack/log.h",
3678]
3679
Marat Dukhan08c4a432019-10-03 09:29:21 -07003680xnnpack_cc_library(
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003681 name = "tables",
3682 srcs = TABLE_SRCS,
3683 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003684 gcc_copts = xnnpack_gcc_std_copts(),
3685 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003686)
3687
3688xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003689 name = "scalar_ukernels",
3690 srcs = SCALAR_UKERNELS,
3691 hdrs = INTERNAL_HDRS,
3692 aarch32_copts = ["-marm"],
Marat Dukhan10a38082020-04-17 03:58:35 -07003693 gcc_copts = xnnpack_gcc_std_copts(),
3694 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003695 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003696 ":tables",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003697 "@FP16",
3698 "@FXdiv",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003699 "@pthreadpool",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003700 ],
3701)
3702
3703xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003704 name = "scalar_ukernels_test_mode",
3705 srcs = SCALAR_UKERNELS,
3706 hdrs = INTERNAL_HDRS,
3707 aarch32_copts = ["-marm"],
3708 copts = [
3709 "-UNDEBUG",
3710 "-DXNN_TEST_MODE=1",
3711 ],
3712 gcc_copts = xnnpack_gcc_std_copts(),
3713 msvc_copts = xnnpack_msvc_std_copts(),
3714 deps = [
3715 ":tables",
3716 "@FP16",
3717 "@FXdiv",
3718 "@pthreadpool",
3719 ],
3720)
3721
3722xnnpack_cc_library(
Marat Dukhan436ebe62019-12-04 15:10:12 -08003723 name = "wasm_ukernels",
3724 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003725 gcc_copts = xnnpack_gcc_std_copts(),
3726 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan436ebe62019-12-04 15:10:12 -08003727 wasm_srcs = WASM_UKERNELS,
Marat Dukhan1c6e3892020-06-25 23:56:10 -07003728 wasmsimd_srcs = WASM_UKERNELS + WASMSIMD_UKERNELS,
Marat Dukhan436ebe62019-12-04 15:10:12 -08003729 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003730 ":tables",
Marat Dukhan436ebe62019-12-04 15:10:12 -08003731 "@FP16",
3732 "@FXdiv",
3733 "@pthreadpool",
3734 ],
3735)
3736
3737xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003738 name = "wasm_ukernels_test_mode",
3739 hdrs = INTERNAL_HDRS,
3740 copts = [
3741 "-UNDEBUG",
3742 "-DXNN_TEST_MODE=1",
3743 ],
3744 gcc_copts = xnnpack_gcc_std_copts(),
3745 msvc_copts = xnnpack_msvc_std_copts(),
3746 wasm_srcs = WASM_UKERNELS,
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07003747 wasmsimd_srcs = WASM_UKERNELS + WASMSIMD_UKERNELS,
Marat Dukhan33fcf782020-05-24 14:27:15 -07003748 deps = [
3749 ":tables",
3750 "@FP16",
3751 "@FXdiv",
3752 "@pthreadpool",
3753 ],
3754)
3755
3756xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003757 name = "neon_ukernels",
3758 hdrs = INTERNAL_HDRS,
3759 aarch32_copts = [
3760 "-marm",
Marat Dukhan8853b822020-05-07 12:19:01 -07003761 "-march=armv7-a",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003762 "-mfpu=neon",
3763 ],
3764 aarch32_srcs = NEON_UKERNELS,
3765 aarch64_srcs = NEON_UKERNELS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003766 gcc_copts = xnnpack_gcc_std_copts(),
3767 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08003768 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003769 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003770 "@FP16",
3771 "@pthreadpool",
3772 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003773)
3774
3775xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003776 name = "neon_ukernels_test_mode",
3777 hdrs = INTERNAL_HDRS,
3778 aarch32_copts = [
3779 "-marm",
3780 "-march=armv7-a",
3781 "-mfpu=neon",
3782 ],
3783 aarch32_srcs = NEON_UKERNELS,
3784 aarch64_srcs = NEON_UKERNELS,
3785 copts = [
3786 "-UNDEBUG",
3787 "-DXNN_TEST_MODE=1",
3788 ],
3789 gcc_copts = xnnpack_gcc_std_copts(),
3790 msvc_copts = xnnpack_msvc_std_copts(),
3791 deps = [
3792 ":tables",
3793 "@FP16",
3794 "@pthreadpool",
3795 ],
3796)
3797
3798xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003799 name = "neonfma_ukernels",
3800 hdrs = INTERNAL_HDRS,
3801 aarch32_copts = [
3802 "-marm",
Marat Dukhan8853b822020-05-07 12:19:01 -07003803 "-march=armv7-a",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003804 "-mfpu=neon-vfpv4",
3805 ],
3806 aarch32_srcs = NEONFMA_UKERNELS,
3807 aarch64_srcs = NEONFMA_UKERNELS + AARCH64_NEONFMA_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07003808 apple_aarch32_copts = [
3809 "-mcpu=swift",
3810 "-mtune=generic",
3811 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07003812 gcc_copts = xnnpack_gcc_std_copts(),
3813 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08003814 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003815 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003816 "@FP16",
3817 "@pthreadpool",
3818 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003819)
3820
3821xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003822 name = "neonfma_ukernels_test_mode",
3823 hdrs = INTERNAL_HDRS,
3824 aarch32_copts = [
3825 "-marm",
3826 "-march=armv7-a",
3827 "-mfpu=neon-vfpv4",
3828 ],
3829 aarch32_srcs = NEONFMA_UKERNELS,
3830 aarch64_srcs = NEONFMA_UKERNELS + AARCH64_NEONFMA_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07003831 apple_aarch32_copts = [
3832 "-mcpu=swift",
3833 "-mtune=generic",
3834 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07003835 copts = [
3836 "-UNDEBUG",
3837 "-DXNN_TEST_MODE=1",
3838 ],
3839 gcc_copts = xnnpack_gcc_std_copts(),
3840 msvc_copts = xnnpack_msvc_std_copts(),
3841 deps = [
3842 ":tables",
3843 "@FP16",
3844 "@pthreadpool",
3845 ],
3846)
3847
3848xnnpack_cc_library(
Marat Dukhan8853b822020-05-07 12:19:01 -07003849 name = "neonv8_ukernels",
3850 hdrs = INTERNAL_HDRS,
3851 aarch32_copts = [
3852 "-marm",
3853 "-march=armv8-a",
3854 "-mfpu=neon-fp-armv8",
3855 ],
3856 aarch32_srcs = NEONV8_UKERNELS,
3857 aarch64_srcs = NEONV8_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07003858 apple_aarch32_copts = [
3859 "-mcpu=cyclone",
3860 "-mtune=generic",
3861 ],
Marat Dukhan8853b822020-05-07 12:19:01 -07003862 gcc_copts = xnnpack_gcc_std_copts(),
3863 msvc_copts = xnnpack_msvc_std_copts(),
3864 deps = [
3865 ":tables",
3866 "@FP16",
3867 "@pthreadpool",
3868 ],
3869)
3870
3871xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003872 name = "neonv8_ukernels_test_mode",
3873 hdrs = INTERNAL_HDRS,
3874 aarch32_copts = [
3875 "-marm",
3876 "-march=armv8-a",
3877 "-mfpu=neon-fp-armv8",
3878 ],
3879 aarch32_srcs = NEONV8_UKERNELS,
3880 aarch64_srcs = NEONV8_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07003881 apple_aarch32_copts = [
3882 "-mcpu=cyclone",
3883 "-mtune=generic",
3884 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07003885 copts = [
3886 "-UNDEBUG",
3887 "-DXNN_TEST_MODE=1",
3888 ],
3889 gcc_copts = xnnpack_gcc_std_copts(),
3890 msvc_copts = xnnpack_msvc_std_copts(),
3891 deps = [
3892 ":tables",
3893 "@FP16",
3894 "@pthreadpool",
3895 ],
3896)
3897
3898xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003899 name = "neonfp16arith_ukernels",
3900 hdrs = INTERNAL_HDRS,
3901 aarch64_copts = ["-march=armv8.2-a+fp16"],
3902 aarch64_srcs = AARCH64_NEONFP16ARITH_UKERNELS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003903 gcc_copts = xnnpack_gcc_std_copts(),
3904 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08003905 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003906 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003907 "@FP16",
3908 "@pthreadpool",
3909 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003910)
3911
3912xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003913 name = "neonfp16arith_ukernels_test_mode",
3914 hdrs = INTERNAL_HDRS,
3915 aarch64_copts = ["-march=armv8.2-a+fp16"],
3916 aarch64_srcs = AARCH64_NEONFP16ARITH_UKERNELS,
3917 copts = [
3918 "-UNDEBUG",
3919 "-DXNN_TEST_MODE=1",
3920 ],
3921 gcc_copts = xnnpack_gcc_std_copts(),
3922 msvc_copts = xnnpack_msvc_std_copts(),
3923 deps = [
3924 ":tables",
3925 "@FP16",
3926 "@pthreadpool",
3927 ],
3928)
3929
3930xnnpack_cc_library(
Benoit Jacoba9644732020-08-13 12:48:55 -07003931 name = "neondot_ukernels",
3932 hdrs = INTERNAL_HDRS,
Marat Dukhan799ac752020-08-13 16:08:03 -07003933 aarch32_copts = [
3934 "-marm",
3935 "-march=armv8.2-a+dotprod",
3936 "-mfpu=neon-fp-armv8",
3937 ],
Benoit Jacoba9644732020-08-13 12:48:55 -07003938 aarch32_srcs = NEONDOT_UKERNELS,
3939 aarch64_copts = ["-march=armv8.2-a+dotprod"],
3940 aarch64_srcs = NEONDOT_UKERNELS,
3941 gcc_copts = xnnpack_gcc_std_copts(),
3942 msvc_copts = xnnpack_msvc_std_copts(),
3943 deps = [
3944 ":tables",
3945 "@FP16",
3946 "@pthreadpool",
3947 ],
3948)
3949
3950xnnpack_cc_library(
3951 name = "neondot_ukernels_test_mode",
3952 hdrs = INTERNAL_HDRS,
Marat Dukhan799ac752020-08-13 16:08:03 -07003953 aarch32_copts = [
3954 "-marm",
3955 "-march=armv8.2-a+dotprod",
3956 "-mfpu=neon-fp-armv8",
3957 ],
Benoit Jacoba9644732020-08-13 12:48:55 -07003958 aarch32_srcs = NEONDOT_UKERNELS,
3959 aarch64_copts = ["-march=armv8.2-a+dotprod"],
3960 aarch64_srcs = NEONDOT_UKERNELS,
3961 copts = [
3962 "-UNDEBUG",
3963 "-DXNN_TEST_MODE=1",
3964 ],
3965 gcc_copts = xnnpack_gcc_std_copts(),
3966 msvc_copts = xnnpack_msvc_std_copts(),
3967 deps = [
3968 ":tables",
3969 "@FP16",
3970 "@pthreadpool",
3971 ],
3972)
3973
3974xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003975 name = "sse2_ukernels",
3976 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003977 gcc_copts = xnnpack_gcc_std_copts(),
3978 gcc_x86_copts = ["-msse2"],
3979 msvc_copts = xnnpack_msvc_std_copts(),
3980 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003981 x86_srcs = SSE_UKERNELS + SSE2_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08003982 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003983 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003984 "@FP16",
3985 "@pthreadpool",
3986 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003987)
3988
3989xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003990 name = "sse2_ukernels_test_mode",
3991 hdrs = INTERNAL_HDRS,
3992 copts = [
3993 "-UNDEBUG",
3994 "-DXNN_TEST_MODE=1",
3995 ],
3996 gcc_copts = xnnpack_gcc_std_copts(),
3997 gcc_x86_copts = ["-msse2"],
3998 msvc_copts = xnnpack_msvc_std_copts(),
3999 msvc_x86_32_copts = ["/arch:SSE2"],
4000 x86_srcs = SSE_UKERNELS + SSE2_UKERNELS,
4001 deps = [
4002 ":tables",
4003 "@FP16",
4004 "@pthreadpool",
4005 ],
4006)
4007
4008xnnpack_cc_library(
Marat Dukhanfe7acb62020-03-09 19:30:05 -07004009 name = "ssse3_ukernels",
4010 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004011 gcc_copts = xnnpack_gcc_std_copts(),
4012 gcc_x86_copts = ["-mssse3"],
4013 msvc_copts = xnnpack_msvc_std_copts(),
4014 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhanfe7acb62020-03-09 19:30:05 -07004015 x86_srcs = SSSE3_UKERNELS,
4016 deps = [
4017 ":tables",
4018 "@FP16",
4019 "@pthreadpool",
4020 ],
4021)
4022
4023xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07004024 name = "ssse3_ukernels_test_mode",
4025 hdrs = INTERNAL_HDRS,
4026 copts = [
4027 "-UNDEBUG",
4028 "-DXNN_TEST_MODE=1",
4029 ],
4030 gcc_copts = xnnpack_gcc_std_copts(),
4031 gcc_x86_copts = ["-mssse3"],
4032 msvc_copts = xnnpack_msvc_std_copts(),
4033 msvc_x86_32_copts = ["/arch:SSE2"],
4034 x86_srcs = SSSE3_UKERNELS,
4035 deps = [
4036 ":tables",
4037 "@FP16",
4038 "@pthreadpool",
4039 ],
4040)
4041
4042xnnpack_cc_library(
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08004043 name = "sse41_ukernels",
4044 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004045 gcc_copts = xnnpack_gcc_std_copts(),
4046 gcc_x86_copts = ["-msse4.1"],
4047 msvc_copts = xnnpack_msvc_std_copts(),
4048 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08004049 x86_srcs = SSE41_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08004050 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08004051 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08004052 "@FP16",
4053 "@pthreadpool",
4054 ],
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08004055)
4056
4057xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07004058 name = "sse41_ukernels_test_mode",
4059 hdrs = INTERNAL_HDRS,
4060 copts = [
4061 "-UNDEBUG",
4062 "-DXNN_TEST_MODE=1",
4063 ],
4064 gcc_copts = xnnpack_gcc_std_copts(),
4065 gcc_x86_copts = ["-msse4.1"],
4066 msvc_copts = xnnpack_msvc_std_copts(),
4067 msvc_x86_32_copts = ["/arch:SSE2"],
4068 x86_srcs = SSE41_UKERNELS,
4069 deps = [
4070 ":tables",
4071 "@FP16",
4072 "@pthreadpool",
4073 ],
4074)
4075
4076xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004077 name = "avx_ukernels",
4078 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004079 gcc_copts = xnnpack_gcc_std_copts(),
4080 gcc_x86_copts = ["-mavx"],
4081 msvc_copts = xnnpack_msvc_std_copts(),
4082 msvc_x86_32_copts = ["/arch:AVX"],
4083 msvc_x86_64_copts = ["/arch:AVX"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004084 x86_srcs = AVX_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08004085 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08004086 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08004087 "@FP16",
4088 "@pthreadpool",
4089 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004090)
4091
4092xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07004093 name = "avx_ukernels_test_mode",
4094 hdrs = INTERNAL_HDRS,
4095 copts = [
4096 "-UNDEBUG",
4097 "-DXNN_TEST_MODE=1",
4098 ],
4099 gcc_copts = xnnpack_gcc_std_copts(),
4100 gcc_x86_copts = ["-mavx"],
4101 msvc_copts = xnnpack_msvc_std_copts(),
4102 msvc_x86_32_copts = ["/arch:AVX"],
4103 msvc_x86_64_copts = ["/arch:AVX"],
4104 x86_srcs = AVX_UKERNELS,
4105 deps = [
4106 ":tables",
4107 "@FP16",
4108 "@pthreadpool",
4109 ],
4110)
4111
4112xnnpack_cc_library(
Marat Dukhan1566fee2020-08-02 21:55:41 -07004113 name = "xop_ukernels",
4114 hdrs = INTERNAL_HDRS,
4115 gcc_copts = xnnpack_gcc_std_copts(),
4116 gcc_x86_copts = ["-mxop"],
4117 msvc_copts = xnnpack_msvc_std_copts(),
4118 msvc_x86_32_copts = ["/arch:AVX"],
4119 msvc_x86_64_copts = ["/arch:AVX"],
4120 x86_srcs = XOP_UKERNELS,
4121 deps = [
4122 ":tables",
4123 "@FP16",
4124 "@pthreadpool",
4125 ],
4126)
4127
4128xnnpack_cc_library(
4129 name = "xop_ukernels_test_mode",
4130 hdrs = INTERNAL_HDRS,
4131 copts = [
4132 "-UNDEBUG",
4133 "-DXNN_TEST_MODE=1",
4134 ],
4135 gcc_copts = xnnpack_gcc_std_copts(),
4136 gcc_x86_copts = ["-mxop"],
4137 msvc_copts = xnnpack_msvc_std_copts(),
4138 msvc_x86_32_copts = ["/arch:AVX"],
4139 msvc_x86_64_copts = ["/arch:AVX"],
4140 x86_srcs = XOP_UKERNELS,
4141 deps = [
4142 ":tables",
4143 "@FP16",
4144 "@pthreadpool",
4145 ],
4146)
4147
4148xnnpack_cc_library(
Marat Dukhanfda12b82019-11-21 12:27:59 -08004149 name = "fma3_ukernels",
4150 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004151 gcc_copts = xnnpack_gcc_std_copts(),
4152 gcc_x86_copts = ["-mfma"],
4153 msvc_copts = xnnpack_msvc_std_copts(),
4154 msvc_x86_32_copts = ["/arch:AVX"],
4155 msvc_x86_64_copts = ["/arch:AVX"],
Marat Dukhanfda12b82019-11-21 12:27:59 -08004156 x86_srcs = FMA3_UKERNELS,
4157 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08004158 ":tables",
Marat Dukhanfda12b82019-11-21 12:27:59 -08004159 "@FP16",
4160 "@pthreadpool",
4161 ],
4162)
4163
4164xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07004165 name = "fma3_ukernels_test_mode",
4166 hdrs = INTERNAL_HDRS,
4167 copts = [
4168 "-UNDEBUG",
4169 "-DXNN_TEST_MODE=1",
4170 ],
4171 gcc_copts = xnnpack_gcc_std_copts(),
4172 gcc_x86_copts = ["-mfma"],
4173 msvc_copts = xnnpack_msvc_std_copts(),
4174 msvc_x86_32_copts = ["/arch:AVX"],
4175 msvc_x86_64_copts = ["/arch:AVX"],
4176 x86_srcs = FMA3_UKERNELS,
4177 deps = [
4178 ":tables",
4179 "@FP16",
4180 "@pthreadpool",
4181 ],
4182)
4183
4184xnnpack_cc_library(
Marat Dukhan6adff4e2019-10-14 18:32:07 -07004185 name = "avx2_ukernels",
4186 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004187 gcc_copts = xnnpack_gcc_std_copts(),
4188 gcc_x86_copts = [
Marat Dukhan6adff4e2019-10-14 18:32:07 -07004189 "-mfma",
4190 "-mavx2",
4191 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07004192 msvc_copts = xnnpack_msvc_std_copts(),
4193 msvc_x86_32_copts = ["/arch:AVX2"],
4194 msvc_x86_64_copts = ["/arch:AVX2"],
Marat Dukhan6adff4e2019-10-14 18:32:07 -07004195 x86_srcs = AVX2_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08004196 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08004197 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08004198 "@FP16",
4199 "@pthreadpool",
4200 ],
Marat Dukhan6adff4e2019-10-14 18:32:07 -07004201)
4202
4203xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07004204 name = "avx2_ukernels_test_mode",
4205 hdrs = INTERNAL_HDRS,
4206 copts = [
4207 "-UNDEBUG",
4208 "-DXNN_TEST_MODE=1",
4209 ],
4210 gcc_copts = xnnpack_gcc_std_copts(),
4211 gcc_x86_copts = [
4212 "-mfma",
4213 "-mavx2",
4214 ],
4215 msvc_copts = xnnpack_msvc_std_copts(),
4216 msvc_x86_32_copts = ["/arch:AVX2"],
4217 msvc_x86_64_copts = ["/arch:AVX2"],
4218 x86_srcs = AVX2_UKERNELS,
4219 deps = [
4220 ":tables",
4221 "@FP16",
4222 "@pthreadpool",
4223 ],
4224)
4225
4226xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004227 name = "avx512f_ukernels",
4228 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004229 gcc_copts = xnnpack_gcc_std_copts(),
4230 gcc_x86_copts = ["-mavx512f"],
4231 mingw_copts = ["-fno-asynchronous-unwind-tables"],
4232 msvc_copts = xnnpack_msvc_std_copts(),
4233 msvc_x86_32_copts = ["/arch:AVX512"],
4234 msvc_x86_64_copts = ["/arch:AVX512"],
4235 msys_copts = ["-fno-asynchronous-unwind-tables"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004236 x86_srcs = AVX512F_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08004237 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08004238 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08004239 "@FP16",
4240 "@pthreadpool",
4241 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004242)
4243
4244xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07004245 name = "avx512f_ukernels_test_mode",
4246 hdrs = INTERNAL_HDRS,
4247 copts = [
4248 "-UNDEBUG",
4249 "-DXNN_TEST_MODE=1",
4250 ],
4251 gcc_copts = xnnpack_gcc_std_copts(),
4252 gcc_x86_copts = ["-mavx512f"],
4253 mingw_copts = ["-fno-asynchronous-unwind-tables"],
4254 msvc_copts = xnnpack_msvc_std_copts(),
4255 msvc_x86_32_copts = ["/arch:AVX512"],
4256 msvc_x86_64_copts = ["/arch:AVX512"],
4257 msys_copts = ["-fno-asynchronous-unwind-tables"],
4258 x86_srcs = AVX512F_UKERNELS,
4259 deps = [
4260 ":tables",
4261 "@FP16",
4262 "@pthreadpool",
4263 ],
4264)
4265
4266xnnpack_cc_library(
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07004267 name = "avx512skx_ukernels",
4268 hdrs = INTERNAL_HDRS,
4269 gcc_copts = xnnpack_gcc_std_copts(),
4270 gcc_x86_copts = [
4271 "-mavx512f",
4272 "-mavx512cd",
4273 "-mavx512bw",
4274 "-mavx512dq",
4275 "-mavx512vl",
4276 ],
4277 mingw_copts = ["-fno-asynchronous-unwind-tables"],
4278 msvc_copts = xnnpack_msvc_std_copts(),
4279 msvc_x86_32_copts = ["/arch:AVX512"],
4280 msvc_x86_64_copts = ["/arch:AVX512"],
4281 msys_copts = ["-fno-asynchronous-unwind-tables"],
4282 x86_srcs = AVX512SKX_UKERNELS,
4283 deps = [
4284 ":tables",
4285 "@FP16",
4286 "@pthreadpool",
4287 ],
4288)
4289
4290xnnpack_cc_library(
4291 name = "avx512skx_ukernels_test_mode",
4292 hdrs = INTERNAL_HDRS,
4293 copts = [
4294 "-UNDEBUG",
4295 "-DXNN_TEST_MODE=1",
4296 ],
4297 gcc_copts = xnnpack_gcc_std_copts(),
4298 gcc_x86_copts = [
4299 "-mavx512f",
4300 "-mavx512cd",
4301 "-mavx512bw",
4302 "-mavx512dq",
4303 "-mavx512vl",
4304 ],
4305 mingw_copts = ["-fno-asynchronous-unwind-tables"],
4306 msvc_copts = xnnpack_msvc_std_copts(),
4307 msvc_x86_32_copts = ["/arch:AVX512"],
4308 msvc_x86_64_copts = ["/arch:AVX512"],
4309 msys_copts = ["-fno-asynchronous-unwind-tables"],
4310 x86_srcs = AVX512SKX_UKERNELS,
4311 deps = [
4312 ":tables",
4313 "@FP16",
4314 "@pthreadpool",
4315 ],
4316)
4317
4318xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004319 name = "asm_ukernels",
4320 hdrs = ["src/xnnpack/assembly.h"],
4321 aarch32_srcs = AARCH32_ASM_UKERNELS,
Frank Barchard31bb45b2020-10-06 00:26:33 -07004322 aarch64_copts = ["-march=armv8.2-a+fp16+dotprod"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004323 aarch64_srcs = AARCH64_ASM_UKERNELS,
Frank Barchardbcedc082020-08-17 18:00:51 -07004324 wasm_srcs = WASM32_ASM_UKERNELS,
4325 wasmsimd_srcs = WASM32_ASM_UKERNELS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07004326)
4327
Marat Dukhan3b59de22020-06-03 20:15:19 -07004328xnnpack_cc_library(
4329 name = "logging_utils",
4330 srcs = LOGGING_SRCS,
4331 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
4332 copts = LOGGING_COPTS + [
4333 "-Isrc",
4334 "-Iinclude",
4335 ] + select({
4336 ":debug_build": [],
4337 "//conditions:default": xnnpack_min_size_copts(),
4338 }),
4339 gcc_copts = xnnpack_gcc_std_copts(),
4340 msvc_copts = xnnpack_msvc_std_copts(),
4341 visibility = xnnpack_visibility(),
4342 deps = [
4343 "@FP16",
4344 "@clog",
4345 "@pthreadpool",
4346 ],
4347)
4348
Marat Dukhan08c4a432019-10-03 09:29:21 -07004349xnnpack_aggregate_library(
4350 name = "ukernels",
4351 aarch32_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07004352 ":neon_ukernels",
4353 ":neonfma_ukernels",
Marat Dukhan8853b822020-05-07 12:19:01 -07004354 ":neonv8_ukernels",
Benoit Jacoba9644732020-08-13 12:48:55 -07004355 ":neondot_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004356 ":asm_ukernels",
4357 ],
4358 aarch64_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07004359 ":neon_ukernels",
4360 ":neonfma_ukernels",
Marat Dukhan8853b822020-05-07 12:19:01 -07004361 ":neonv8_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004362 ":neonfp16arith_ukernels",
Benoit Jacoba9644732020-08-13 12:48:55 -07004363 ":neondot_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004364 ":asm_ukernels",
4365 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07004366 generic_deps = [
4367 ":scalar_ukernels",
4368 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07004369 wasm_deps = [
4370 ":wasm_ukernels",
Frank Barchardbcedc082020-08-17 18:00:51 -07004371 ":asm_ukernels",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004372 ],
4373 wasmsimd_deps = [
4374 ":wasm_ukernels",
Frank Barchardbcedc082020-08-17 18:00:51 -07004375 ":asm_ukernels",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004376 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004377 x86_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07004378 ":sse2_ukernels",
Marat Dukhanfe7acb62020-03-09 19:30:05 -07004379 ":ssse3_ukernels",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08004380 ":sse41_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004381 ":avx_ukernels",
Marat Dukhan1566fee2020-08-02 21:55:41 -07004382 ":xop_ukernels",
Marat Dukhanfda12b82019-11-21 12:27:59 -08004383 ":fma3_ukernels",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07004384 ":avx2_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004385 ":avx512f_ukernels",
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07004386 ":avx512skx_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004387 ],
4388)
4389
Marat Dukhan33fcf782020-05-24 14:27:15 -07004390xnnpack_aggregate_library(
4391 name = "ukernels_test_mode",
4392 aarch32_deps = [
4393 ":neon_ukernels_test_mode",
4394 ":neonfma_ukernels_test_mode",
4395 ":neonv8_ukernels_test_mode",
Benoit Jacoba9644732020-08-13 12:48:55 -07004396 ":neondot_ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004397 ":asm_ukernels",
4398 ],
4399 aarch64_deps = [
4400 ":neon_ukernels_test_mode",
4401 ":neonfma_ukernels_test_mode",
4402 ":neonv8_ukernels_test_mode",
4403 ":neonfp16arith_ukernels_test_mode",
Benoit Jacoba9644732020-08-13 12:48:55 -07004404 ":neondot_ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004405 ":asm_ukernels",
4406 ],
4407 generic_deps = [
4408 ":scalar_ukernels_test_mode",
4409 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07004410 wasm_deps = [
4411 ":wasm_ukernels_test_mode",
Frank Barchardbcedc082020-08-17 18:00:51 -07004412 ":asm_ukernels",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004413 ],
4414 wasmsimd_deps = [
4415 ":wasm_ukernels_test_mode",
Frank Barchardbcedc082020-08-17 18:00:51 -07004416 ":asm_ukernels",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004417 ],
4418 x86_deps = [
4419 ":sse2_ukernels_test_mode",
4420 ":ssse3_ukernels_test_mode",
4421 ":sse41_ukernels_test_mode",
4422 ":avx_ukernels_test_mode",
Marat Dukhan1566fee2020-08-02 21:55:41 -07004423 ":xop_ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004424 ":fma3_ukernels_test_mode",
4425 ":avx2_ukernels_test_mode",
4426 ":avx512f_ukernels_test_mode",
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07004427 ":avx512skx_ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004428 ],
4429)
4430
Marat Dukhan86df3842020-06-17 14:03:31 -07004431# This target must be compatible with Android NDK r17 and compile with Clang 6.
4432# It is similar to :ukernels target, but excludes microkernels which require
4433# NEON FP16 arithmetics, NEON Dot Product, and newer intrinsics.
Marat Dukhanae046f52020-06-15 13:16:14 -07004434xnnpack_aggregate_library(
Marat Dukhan86df3842020-06-17 14:03:31 -07004435 name = "ukernels_max_compatibility",
Marat Dukhanae046f52020-06-15 13:16:14 -07004436 aarch32_deps = [
4437 ":neon_ukernels",
4438 ":neonfma_ukernels",
4439 ":neonv8_ukernels",
4440 ":asm_ukernels",
4441 ],
4442 aarch64_deps = [
4443 ":neon_ukernels",
4444 ":neonfma_ukernels",
4445 ":neonv8_ukernels",
4446 ":asm_ukernels",
4447 ],
4448 generic_deps = [
4449 ":scalar_ukernels",
4450 ],
Marat Dukhanae046f52020-06-15 13:16:14 -07004451 wasm_deps = [
4452 ":wasm_ukernels",
4453 ],
4454 wasmsimd_deps = [
4455 ":wasm_ukernels",
4456 ],
4457 x86_deps = [
4458 ":sse2_ukernels",
4459 ":ssse3_ukernels",
4460 ":sse41_ukernels",
4461 ":avx_ukernels",
Marat Dukhan1566fee2020-08-02 21:55:41 -07004462 ":xop_ukernels",
Marat Dukhanae046f52020-06-15 13:16:14 -07004463 ":fma3_ukernels",
4464 ":avx2_ukernels",
4465 ":avx512f_ukernels",
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07004466 ":avx512skx_ukernels",
Marat Dukhanae046f52020-06-15 13:16:14 -07004467 ],
4468)
4469
Marat Dukhan08c4a432019-10-03 09:29:21 -07004470xnnpack_cc_library(
4471 name = "im2col",
4472 srcs = ["src/im2col.c"],
4473 hdrs = [
4474 "src/xnnpack/common.h",
4475 "src/xnnpack/im2col.h",
4476 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07004477 gcc_copts = xnnpack_gcc_std_copts(),
4478 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004479)
4480
4481xnnpack_cc_library(
4482 name = "indirection",
4483 srcs = ["src/indirection.c"],
4484 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004485 gcc_copts = xnnpack_gcc_std_copts(),
4486 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004487 deps = [
4488 "@FP16",
4489 "@FXdiv",
4490 "@pthreadpool",
4491 ],
4492)
4493
4494xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07004495 name = "indirection_test_mode",
4496 srcs = ["src/indirection.c"],
4497 hdrs = INTERNAL_HDRS,
4498 copts = [
4499 "-UNDEBUG",
4500 "-DXNN_TEST_MODE=1",
4501 ],
4502 gcc_copts = xnnpack_gcc_std_copts(),
4503 msvc_copts = xnnpack_msvc_std_copts(),
4504 deps = [
4505 "@FP16",
4506 "@FXdiv",
4507 "@pthreadpool",
4508 ],
4509)
4510
4511xnnpack_cc_library(
Marat Dukhanab582382020-07-06 13:32:08 -07004512 name = "packing",
4513 srcs = ["src/packing.c"],
4514 hdrs = INTERNAL_HDRS,
4515 gcc_copts = xnnpack_gcc_std_copts(),
4516 msvc_copts = xnnpack_msvc_std_copts(),
4517 deps = [
4518 "@FP16",
4519 "@FXdiv",
4520 "@pthreadpool",
4521 ],
4522)
4523
4524xnnpack_cc_library(
4525 name = "packing_test_mode",
4526 srcs = ["src/packing.c"],
4527 hdrs = INTERNAL_HDRS,
4528 copts = [
4529 "-UNDEBUG",
4530 "-DXNN_TEST_MODE=1",
4531 ],
4532 gcc_copts = xnnpack_gcc_std_copts(),
4533 msvc_copts = xnnpack_msvc_std_copts(),
4534 deps = [
4535 "@FP16",
4536 "@FXdiv",
4537 "@pthreadpool",
4538 ],
4539)
4540
4541xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004542 name = "operator_run",
4543 srcs = ["src/operator-run.c"],
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07004544 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004545 copts = LOGGING_COPTS + select({
Marat Dukhan05702cf2020-03-26 15:41:33 -07004546 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4547 "//conditions:default": [],
4548 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07004549 gcc_copts = xnnpack_gcc_std_copts(),
4550 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004551 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07004552 ":logging_utils",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004553 "@FP16",
4554 "@FXdiv",
4555 "@clog",
4556 "@pthreadpool",
4557 ],
4558)
4559
Chao Mei6ddfc602020-05-13 22:29:36 -07004560xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07004561 name = "operator_run_test_mode",
4562 srcs = ["src/operator-run.c"],
4563 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
4564 copts = LOGGING_COPTS + [
4565 "-UNDEBUG",
4566 "-DXNN_TEST_MODE=1",
4567 ] + select({
4568 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4569 "//conditions:default": [],
4570 }),
4571 gcc_copts = xnnpack_gcc_std_copts(),
4572 msvc_copts = xnnpack_msvc_std_copts(),
4573 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07004574 ":logging_utils",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004575 "@FP16",
4576 "@FXdiv",
4577 "@clog",
4578 "@pthreadpool",
4579 ],
4580)
4581
4582xnnpack_cc_library(
Chao Mei6ddfc602020-05-13 22:29:36 -07004583 name = "memory_planner",
4584 srcs = ["src/memory-planner.c"],
4585 hdrs = INTERNAL_HDRS,
4586 defines = select({
4587 ":xnn_enable_memopt_explicit_true": ["XNN_ENABLE_MEMOPT=1"],
4588 ":xnn_enable_memopt_explicit_false": ["XNN_ENABLE_MEMOPT=0"],
4589 "//conditions:default": ["XNN_ENABLE_MEMOPT=1"],
4590 }),
4591 gcc_copts = xnnpack_gcc_std_copts(),
4592 msvc_copts = xnnpack_msvc_std_copts(),
4593 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07004594 ":logging_utils",
Chao Mei6ddfc602020-05-13 22:29:36 -07004595 "@pthreadpool",
4596 ],
4597)
4598
Marat Dukhan33fcf782020-05-24 14:27:15 -07004599xnnpack_cc_library(
4600 name = "memory_planner_test_mode",
4601 srcs = ["src/memory-planner.c"],
4602 hdrs = INTERNAL_HDRS,
4603 copts = [
4604 "-UNDEBUG",
4605 "-DXNN_TEST_MODE=1",
4606 ],
4607 defines = select({
4608 ":xnn_enable_memopt_explicit_true": ["XNN_ENABLE_MEMOPT=1"],
4609 ":xnn_enable_memopt_explicit_false": ["XNN_ENABLE_MEMOPT=0"],
4610 "//conditions:default": ["XNN_ENABLE_MEMOPT=1"],
4611 }),
4612 gcc_copts = xnnpack_gcc_std_copts(),
4613 msvc_copts = xnnpack_msvc_std_copts(),
4614 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07004615 ":logging_utils",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004616 "@pthreadpool",
4617 ],
4618)
4619
Marat Dukhan08c4a432019-10-03 09:29:21 -07004620cc_library(
4621 name = "enable_assembly",
4622 defines = select({
4623 ":xnn_enable_assembly_explicit_true": ["XNN_ENABLE_ASSEMBLY=1"],
4624 ":xnn_enable_assembly_explicit_false": ["XNN_ENABLE_ASSEMBLY=0"],
Frank Barchard810171d2019-10-10 10:34:51 -07004625 "//conditions:default": ["XNN_ENABLE_ASSEMBLY=1"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004626 }),
4627)
4628
Marat Dukhan9de90e02020-06-18 16:04:12 -07004629cc_library(
4630 name = "enable_sparse",
4631 defines = select({
4632 ":xnn_enable_sparse_explicit_true": ["XNN_ENABLE_SPARSE=1"],
4633 ":xnn_enable_sparse_explicit_false": ["XNN_ENABLE_SPARSE=0"],
Marat Dukhanb36582b2020-12-08 11:16:28 -08004634 "//conditions:default": ["XNN_ENABLE_SPARSE=1"],
Marat Dukhan9de90e02020-06-18 16:04:12 -07004635 }),
4636)
4637
Marat Dukhancf056b22019-10-07 10:26:29 -07004638xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004639 name = "operators",
4640 srcs = OPERATOR_SRCS + [
Marat Dukhan04f03be2019-11-19 12:36:47 -08004641 "src/memory.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004642 "src/operator-delete.c",
Marat Dukhancf056b22019-10-07 10:26:29 -07004643 ],
4644 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004645 copts = LOGGING_COPTS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07004646 "-Isrc",
4647 "-Iinclude",
4648 ] + select({
4649 ":debug_build": [],
4650 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07004651 }) + select({
4652 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4653 "//conditions:default": [],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004654 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07004655 gcc_copts = xnnpack_gcc_std_copts(),
4656 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004657 deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07004658 ":indirection",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004659 ":logging_utils",
Marat Dukhanab582382020-07-06 13:32:08 -07004660 ":packing",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004661 "@FP16",
4662 "@FXdiv",
4663 "@clog",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004664 "@pthreadpool",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004665 ],
4666)
4667
Marat Dukhan10a38082020-04-17 03:58:35 -07004668xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07004669 name = "operators_test_mode",
4670 srcs = OPERATOR_SRCS + [
4671 "src/memory.c",
4672 "src/operator-delete.c",
4673 ],
4674 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
4675 copts = LOGGING_COPTS + [
4676 "-Isrc",
4677 "-Iinclude",
4678 "-UNDEBUG",
4679 "-DXNN_TEST_MODE=1",
4680 ] + select({
4681 ":debug_build": [],
4682 "//conditions:default": xnnpack_min_size_copts(),
4683 }) + select({
4684 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4685 "//conditions:default": [],
4686 }),
4687 gcc_copts = xnnpack_gcc_std_copts(),
4688 msvc_copts = xnnpack_msvc_std_copts(),
4689 deps = [
4690 ":indirection_test_mode",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004691 ":logging_utils",
Marat Dukhanab582382020-07-06 13:32:08 -07004692 ":packing_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004693 "@FP16",
4694 "@FXdiv",
4695 "@clog",
4696 "@pthreadpool",
4697 ],
4698)
4699
4700xnnpack_cc_library(
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004701 name = "XNNPACK",
4702 srcs = [
4703 "src/init.c",
Marat Dukhanccfdbd12020-02-03 14:27:45 -08004704 "src/runtime.c",
4705 "src/subgraph.c",
4706 "src/tensor.c",
Marat Dukhanf03da0d2020-06-10 16:00:20 -07004707 ] + SUBGRAPH_SRCS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004708 hdrs = ["include/xnnpack.h"],
4709 copts = LOGGING_COPTS + [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004710 "-Isrc",
4711 "-Iinclude",
4712 ] + select({
4713 ":debug_build": [],
4714 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07004715 }) + select({
4716 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4717 "//conditions:default": [],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004718 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07004719 gcc_copts = xnnpack_gcc_std_copts(),
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004720 includes = ["include"],
Marat Dukhan10a38082020-04-17 03:58:35 -07004721 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004722 visibility = xnnpack_visibility(),
4723 deps = [
4724 ":enable_assembly",
Marat Dukhan9de90e02020-06-18 16:04:12 -07004725 ":enable_sparse",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004726 ":logging_utils",
Chao Mei6ddfc602020-05-13 22:29:36 -07004727 ":memory_planner",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004728 ":operator_run",
4729 ":operators",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004730 ":ukernels",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004731 "@clog",
Marat Dukhanab2946c2020-05-21 20:04:13 -07004732 "@FP16",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004733 "@pthreadpool",
Marat Dukhand343c222019-10-07 09:22:14 -07004734 ] + select({
4735 ":emscripten": [],
4736 "//conditions:default": ["@cpuinfo"],
4737 }),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004738)
4739
Marat Dukhan10a38082020-04-17 03:58:35 -07004740xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07004741 name = "XNNPACK_test_mode",
4742 srcs = [
4743 "src/init.c",
4744 "src/runtime.c",
4745 "src/subgraph.c",
4746 "src/tensor.c",
Marat Dukhanf03da0d2020-06-10 16:00:20 -07004747 ] + SUBGRAPH_SRCS,
Marat Dukhan33fcf782020-05-24 14:27:15 -07004748 hdrs = ["include/xnnpack.h"],
4749 copts = LOGGING_COPTS + [
4750 "-Isrc",
4751 "-Iinclude",
4752 "-UNDEBUG",
4753 "-DXNN_TEST_MODE=1",
4754 ] + select({
4755 ":debug_build": [],
4756 "//conditions:default": xnnpack_min_size_copts(),
4757 }) + select({
4758 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4759 "//conditions:default": [],
4760 }),
4761 gcc_copts = xnnpack_gcc_std_copts(),
4762 includes = ["include"],
4763 msvc_copts = xnnpack_msvc_std_copts(),
4764 visibility = xnnpack_visibility(),
4765 deps = [
4766 ":enable_assembly",
Marat Dukhan9de90e02020-06-18 16:04:12 -07004767 ":enable_sparse",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004768 ":logging_utils",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004769 ":memory_planner_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004770 ":operator_run_test_mode",
4771 ":operators_test_mode",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004772 ":ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004773 "@clog",
4774 "@FP16",
4775 "@pthreadpool",
4776 ] + select({
4777 ":emscripten": [],
4778 "//conditions:default": ["@cpuinfo"],
4779 }),
4780)
4781
4782xnnpack_cc_library(
Marat Dukhanae046f52020-06-15 13:16:14 -07004783 name = "xnnpack_f32",
4784 srcs = [
4785 "src/init.c",
4786 "src/runtime.c",
4787 "src/subgraph.c",
4788 "src/tensor.c",
4789 ] + SUBGRAPH_SRCS,
4790 hdrs = ["include/xnnpack.h"],
4791 copts = LOGGING_COPTS + [
4792 "-Isrc",
4793 "-Iinclude",
4794 ] + select({
4795 ":debug_build": [],
4796 "//conditions:default": xnnpack_min_size_copts(),
4797 }) + select({
4798 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4799 "//conditions:default": [],
4800 }),
4801 defines = [
Marat Dukhan16f1e1a2020-08-04 16:38:22 -07004802 "XNN_NO_QS8_OPERATORS",
Marat Dukhan08b7a972020-07-14 18:17:29 -07004803 "XNN_NO_QU8_OPERATORS",
Marat Dukhanae046f52020-06-15 13:16:14 -07004804 "XNN_NO_U8_OPERATORS",
4805 "XNN_NO_X8_OPERATORS",
4806 "XNN_NO_F16_OPERATORS",
4807 "XNN_NO_X16_OPERATORS",
4808 ],
4809 gcc_copts = xnnpack_gcc_std_copts(),
4810 includes = ["include"],
4811 msvc_copts = xnnpack_msvc_std_copts(),
4812 visibility = xnnpack_visibility(),
4813 deps = [
4814 ":enable_assembly",
Marat Dukhan9de90e02020-06-18 16:04:12 -07004815 ":enable_sparse",
Marat Dukhanae046f52020-06-15 13:16:14 -07004816 ":logging_utils",
4817 ":memory_planner",
4818 ":operator_run",
4819 ":operators",
Marat Dukhan86df3842020-06-17 14:03:31 -07004820 ":ukernels_max_compatibility",
Marat Dukhanae046f52020-06-15 13:16:14 -07004821 "@clog",
4822 "@FP16",
4823 "@pthreadpool",
4824 ] + select({
4825 ":emscripten": [],
4826 "//conditions:default": ["@cpuinfo"],
4827 }),
4828)
4829
4830xnnpack_cc_library(
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004831 name = "xnnpack_operators_nhwc_f32",
4832 srcs = [
4833 "src/init.c",
4834 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07004835 hdrs = ["include/xnnpack.h"],
4836 copts = LOGGING_COPTS + [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004837 "-Isrc",
4838 "-Iinclude",
4839 ] + select({
4840 ":debug_build": [],
4841 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07004842 }) + select({
4843 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4844 "//conditions:default": [],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004845 }),
4846 defines = [
Marat Dukhan16f1e1a2020-08-04 16:38:22 -07004847 "XNN_NO_QS8_OPERATORS",
Marat Dukhan08b7a972020-07-14 18:17:29 -07004848 "XNN_NO_QU8_OPERATORS",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004849 "XNN_NO_U8_OPERATORS",
4850 "XNN_NO_X8_OPERATORS",
Marat Dukhanefc47b82019-11-18 09:25:38 -08004851 "XNN_NO_NCHW_OPERATORS",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004852 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07004853 gcc_copts = xnnpack_gcc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004854 includes = ["include"],
Marat Dukhan10a38082020-04-17 03:58:35 -07004855 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004856 visibility = xnnpack_visibility(),
4857 deps = [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004858 ":enable_assembly",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004859 ":logging_utils",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004860 ":operator_run",
4861 ":operators",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004862 ":ukernels",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004863 "@clog",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004864 "@pthreadpool",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004865 ] + select({
4866 ":emscripten": [],
4867 "//conditions:default": ["@cpuinfo"],
4868 }),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004869)
4870
Marat Dukhancf056b22019-10-07 10:26:29 -07004871xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004872 name = "bench_utils",
4873 srcs = ["bench/utils.cc"],
4874 hdrs = ["bench/utils.h"],
Marat Dukhanbad48fe2019-11-04 10:35:22 -08004875 deps = [
4876 "@com_google_benchmark//:benchmark",
4877 "@cpuinfo",
4878 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004879)
4880
Frank Barchard7e955972019-10-11 10:34:25 -07004881######################### Benchmarks for micro-kernels #########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07004882
4883xnnpack_benchmark(
Marat Dukhan595e1702020-07-31 10:12:52 -07004884 name = "qs8_gemm_bench",
4885 srcs = [
4886 "bench/gemm.h",
4887 "bench/qs8-gemm.cc",
4888 "src/xnnpack/AlignedAllocator.h",
4889 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Frank Barchard31328cb2020-10-12 11:55:18 -07004890 copts = xnnpack_optional_ruy_copts() + xnnpack_optional_gemmlowp_copts(),
4891 deps = MICROKERNEL_BENCHMARK_DEPS + [":packing"] + xnnpack_optional_ruy_deps() + xnnpack_optional_gemmlowp_deps(),
Marat Dukhan595e1702020-07-31 10:12:52 -07004892)
4893
4894xnnpack_benchmark(
Marat Dukhan56bdd4a2020-08-03 19:47:04 -07004895 name = "qs8_requantization_bench",
4896 srcs = [
4897 "bench/qs8-requantization.cc",
Marat Dukhan56bdd4a2020-08-03 19:47:04 -07004898 "src/xnnpack/AlignedAllocator.h",
Frank Barchard04336c12020-10-22 16:48:55 -07004899 "src/xnnpack/requantization-stubs.h",
Marat Dukhan56bdd4a2020-08-03 19:47:04 -07004900 ] + MICROKERNEL_BENCHMARK_HDRS,
4901 deps = MICROKERNEL_BENCHMARK_DEPS,
4902)
4903
4904xnnpack_benchmark(
Marat Dukhan08b7a972020-07-14 18:17:29 -07004905 name = "qu8_gemm_bench",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004906 srcs = [
4907 "bench/gemm.h",
Marat Dukhan08b7a972020-07-14 18:17:29 -07004908 "bench/qu8-gemm.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004909 "src/xnnpack/AlignedAllocator.h",
4910 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004911 copts = xnnpack_optional_ruy_copts() + xnnpack_optional_gemmlowp_copts(),
Marat Dukhanab582382020-07-06 13:32:08 -07004912 deps = MICROKERNEL_BENCHMARK_DEPS + [":packing"] + xnnpack_optional_ruy_deps() + xnnpack_optional_gemmlowp_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004913)
4914
4915xnnpack_benchmark(
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07004916 name = "qu8_requantization_bench",
4917 srcs = [
4918 "bench/qu8-requantization.cc",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07004919 "src/xnnpack/AlignedAllocator.h",
Frank Barchard04336c12020-10-22 16:48:55 -07004920 "src/xnnpack/requantization-stubs.h",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07004921 ] + MICROKERNEL_BENCHMARK_HDRS,
4922 deps = MICROKERNEL_BENCHMARK_DEPS,
4923)
4924
4925xnnpack_benchmark(
Frank Barchard40d20fe2020-05-05 00:37:45 -07004926 name = "f16_igemm_bench",
4927 srcs = [
4928 "bench/f16-igemm.cc",
4929 "bench/conv.h",
4930 "bench/google/conv.h",
4931 "src/xnnpack/AlignedAllocator.h",
4932 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004933 deps = MICROKERNEL_BENCHMARK_DEPS + [
4934 ":indirection",
4935 ":packing",
4936 ],
Frank Barchard40d20fe2020-05-05 00:37:45 -07004937)
4938
4939xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004940 name = "f16_gemm_bench",
4941 srcs = [
4942 "bench/f16-gemm.cc",
4943 "bench/gemm.h",
4944 "src/xnnpack/AlignedAllocator.h",
4945 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004946 deps = MICROKERNEL_BENCHMARK_DEPS + [
4947 ":packing",
4948 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004949)
4950
4951xnnpack_benchmark(
Marat Dukhanbdb56f52020-02-05 21:42:49 -08004952 name = "f16_spmm_bench",
4953 srcs = [
4954 "bench/f16-spmm.cc",
Marat Dukhan1631e3e2020-12-06 19:29:31 -08004955 "bench/spmm.h",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08004956 "src/xnnpack/AlignedAllocator.h",
4957 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanbdb56f52020-02-05 21:42:49 -08004958 deps = MICROKERNEL_BENCHMARK_DEPS,
4959)
4960
4961xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004962 name = "f32_igemm_bench",
4963 srcs = [
4964 "bench/f32-igemm.cc",
4965 "bench/conv.h",
4966 "src/xnnpack/AlignedAllocator.h",
4967 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004968 deps = MICROKERNEL_BENCHMARK_DEPS + [
4969 ":indirection",
4970 ":packing",
4971 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004972)
4973
4974xnnpack_benchmark(
Frank Barchardc2010062020-07-19 10:04:12 -07004975 name = "f16_relu_bench",
4976 srcs = [
4977 "bench/f16-relu.cc",
4978 "src/xnnpack/AlignedAllocator.h",
4979 ] + MICROKERNEL_BENCHMARK_HDRS,
4980 deps = MICROKERNEL_BENCHMARK_DEPS,
4981)
4982
4983xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004984 name = "f32_conv_hwc_bench",
4985 srcs = [
4986 "bench/f32-conv-hwc.cc",
4987 "bench/dconv.h",
4988 "src/xnnpack/AlignedAllocator.h",
4989 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004990 deps = MICROKERNEL_BENCHMARK_DEPS + [
4991 ":packing",
4992 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004993)
4994
4995xnnpack_benchmark(
Marat Dukhan1f29b802020-05-15 23:46:39 -07004996 name = "f32_conv_hwc2chw_bench",
Erich Elsen563df5f2019-10-23 08:02:21 -07004997 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07004998 "bench/f32-conv-hwc2chw.cc",
Erich Elsen563df5f2019-10-23 08:02:21 -07004999 "bench/dconv.h",
5000 "src/xnnpack/AlignedAllocator.h",
5001 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005002 deps = MICROKERNEL_BENCHMARK_DEPS + [
5003 ":packing",
5004 ],
Erich Elsen563df5f2019-10-23 08:02:21 -07005005)
5006
5007xnnpack_benchmark(
Frank Barchard5a599a62020-06-04 20:12:44 -07005008 name = "f16_dwconv_bench",
5009 srcs = [
5010 "bench/f16-dwconv.cc",
5011 "bench/dwconv.h",
5012 "bench/google/dwconv.h",
5013 "src/xnnpack/AlignedAllocator.h",
5014 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005015 deps = MICROKERNEL_BENCHMARK_DEPS + [
5016 ":indirection",
5017 ":packing",
5018 ],
Frank Barchard5a599a62020-06-04 20:12:44 -07005019)
5020
5021xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005022 name = "f32_dwconv_bench",
5023 srcs = [
5024 "bench/f32-dwconv.cc",
5025 "bench/dwconv.h",
5026 "src/xnnpack/AlignedAllocator.h",
5027 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005028 deps = MICROKERNEL_BENCHMARK_DEPS + [
5029 ":indirection",
5030 ":packing",
5031 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005032)
5033
5034xnnpack_benchmark(
Marat Dukhanbf715f92020-10-23 20:17:00 -07005035 name = "f32_dwconv2d_chw_bench",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005036 srcs = [
Marat Dukhanbf715f92020-10-23 20:17:00 -07005037 "bench/f32-dwconv2d-chw.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005038 "bench/dwconv.h",
5039 "src/xnnpack/AlignedAllocator.h",
5040 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005041 deps = MICROKERNEL_BENCHMARK_DEPS + [
5042 ":indirection",
5043 ":packing",
5044 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005045)
5046
5047xnnpack_benchmark(
5048 name = "f32_gemm_bench",
5049 srcs = [
5050 "bench/f32-gemm.cc",
5051 "bench/gemm.h",
5052 "src/xnnpack/AlignedAllocator.h",
5053 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07005054 copts = xnnpack_optional_ruy_copts(),
Marat Dukhanab582382020-07-06 13:32:08 -07005055 deps = MICROKERNEL_BENCHMARK_DEPS + [":packing"] + xnnpack_optional_ruy_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07005056)
5057
5058xnnpack_benchmark(
Marat Dukhana11ca342020-06-25 23:45:07 -07005059 name = "f32_hswish_bench",
5060 srcs = [
5061 "bench/f32-hswish.cc",
5062 "src/xnnpack/AlignedAllocator.h",
5063 ] + MICROKERNEL_BENCHMARK_HDRS,
5064 deps = MICROKERNEL_BENCHMARK_DEPS,
5065)
5066
5067xnnpack_benchmark(
Marat Dukhan4c4eb002019-12-08 21:27:49 -08005068 name = "f32_raddexpminusmax_bench",
5069 srcs = [
5070 "bench/f32-raddexpminusmax.cc",
5071 "src/xnnpack/AlignedAllocator.h",
5072 ] + MICROKERNEL_BENCHMARK_HDRS,
5073 deps = MICROKERNEL_BENCHMARK_DEPS,
5074)
5075
5076xnnpack_benchmark(
5077 name = "f32_raddextexp_bench",
5078 srcs = [
5079 "bench/f32-raddextexp.cc",
5080 "src/xnnpack/AlignedAllocator.h",
5081 ] + MICROKERNEL_BENCHMARK_HDRS,
5082 deps = MICROKERNEL_BENCHMARK_DEPS,
5083)
5084
5085xnnpack_benchmark(
5086 name = "f32_raddstoreexpminusmax_bench",
5087 srcs = [
5088 "bench/f32-raddstoreexpminusmax.cc",
5089 "src/xnnpack/AlignedAllocator.h",
5090 ] + MICROKERNEL_BENCHMARK_HDRS,
5091 deps = MICROKERNEL_BENCHMARK_DEPS,
5092)
5093
5094xnnpack_benchmark(
Frank Barchard3c4a9522020-07-18 16:11:22 -07005095 name = "f32_relu_bench",
5096 srcs = [
5097 "bench/f32-relu.cc",
5098 "src/xnnpack/AlignedAllocator.h",
5099 ] + MICROKERNEL_BENCHMARK_HDRS,
5100 deps = MICROKERNEL_BENCHMARK_DEPS,
5101)
5102
5103xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005104 name = "f32_rmax_bench",
5105 srcs = [
5106 "bench/f32-rmax.cc",
5107 "src/xnnpack/AlignedAllocator.h",
5108 ] + MICROKERNEL_BENCHMARK_HDRS,
5109 deps = MICROKERNEL_BENCHMARK_DEPS,
5110)
5111
5112xnnpack_benchmark(
Marat Dukhan14bec502019-11-18 11:35:31 -08005113 name = "f32_sigmoid_bench",
5114 srcs = [
5115 "bench/f32-sigmoid.cc",
5116 "src/xnnpack/AlignedAllocator.h",
5117 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan14bec502019-11-18 11:35:31 -08005118 deps = MICROKERNEL_BENCHMARK_DEPS,
5119)
5120
5121xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005122 name = "f32_spmm_bench",
5123 srcs = [
5124 "bench/f32-spmm.cc",
Marat Dukhan1631e3e2020-12-06 19:29:31 -08005125 "bench/spmm.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005126 "src/xnnpack/AlignedAllocator.h",
5127 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005128 deps = MICROKERNEL_BENCHMARK_DEPS,
5129)
5130
5131xnnpack_benchmark(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08005132 name = "f32_softmax_bench",
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07005133 srcs = [
Marat Dukhanfd8e6892020-01-27 15:25:25 -08005134 "bench/f32-softmax.cc",
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07005135 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07005136 copts = xnnpack_optional_dnnl_copts(),
Marat Dukhan8d3c6932020-03-06 20:27:27 -08005137 deps = MICROKERNEL_BENCHMARK_DEPS + xnnpack_optional_dnnl_deps(),
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07005138)
5139
5140xnnpack_benchmark(
Marat Dukhaned6baaf2020-12-01 15:07:08 -08005141 name = "f32_velu_bench",
5142 srcs = [
5143 "bench/f32-velu.cc",
5144 "src/xnnpack/AlignedAllocator.h",
5145 ] + MICROKERNEL_BENCHMARK_HDRS,
5146 deps = MICROKERNEL_BENCHMARK_DEPS,
5147)
5148
5149xnnpack_benchmark(
Marat Dukhan4c4eb002019-12-08 21:27:49 -08005150 name = "f32_vscaleexpminusmax_bench",
5151 srcs = [
5152 "bench/f32-vscaleexpminusmax.cc",
5153 "src/xnnpack/AlignedAllocator.h",
5154 ] + MICROKERNEL_BENCHMARK_HDRS,
5155 deps = MICROKERNEL_BENCHMARK_DEPS,
5156)
5157
5158xnnpack_benchmark(
5159 name = "f32_vscaleextexp_bench",
5160 srcs = [
5161 "bench/f32-vscaleextexp.cc",
5162 "src/xnnpack/AlignedAllocator.h",
5163 ] + MICROKERNEL_BENCHMARK_HDRS,
5164 deps = MICROKERNEL_BENCHMARK_DEPS,
5165)
5166
5167xnnpack_benchmark(
Marat Dukhanf4db2f32020-06-30 10:55:30 -07005168 name = "f32_vsqrt_bench",
5169 srcs = [
5170 "bench/f32-vsqrt.cc",
5171 "src/xnnpack/AlignedAllocator.h",
5172 ] + MICROKERNEL_BENCHMARK_HDRS,
5173 deps = MICROKERNEL_BENCHMARK_DEPS,
5174)
5175
5176xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005177 name = "f32_im2col_gemm_bench",
5178 srcs = [
5179 "bench/f32-im2col-gemm.cc",
5180 "bench/conv.h",
5181 "src/xnnpack/AlignedAllocator.h",
5182 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005183 deps = MICROKERNEL_BENCHMARK_DEPS + [
5184 ":im2col",
5185 ":packing",
5186 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005187)
5188
Marat Dukhanfe7acb62020-03-09 19:30:05 -07005189xnnpack_benchmark(
Marat Dukhanffbf96a2020-05-14 02:59:08 -07005190 name = "rounding_bench",
5191 srcs = [
5192 "bench/rounding.cc",
Marat Dukhanffbf96a2020-05-14 02:59:08 -07005193 "src/xnnpack/AlignedAllocator.h",
Frank Barchard04336c12020-10-22 16:48:55 -07005194 "src/xnnpack/math-stubs.h",
Marat Dukhanffbf96a2020-05-14 02:59:08 -07005195 ] + MICROKERNEL_BENCHMARK_HDRS,
5196 deps = MICROKERNEL_BENCHMARK_DEPS,
5197)
5198
Marat Dukhan08c4a432019-10-03 09:29:21 -07005199########################### Benchmarks for operators ###########################
5200
5201xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005202 name = "average_pooling_bench",
5203 srcs = ["bench/average-pooling.cc"],
Marat Dukhan7a16d8b2020-03-11 04:22:44 -07005204 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07005205 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07005206 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07005207)
5208
5209xnnpack_benchmark(
Marat Dukhan87727142020-06-24 15:24:10 -07005210 name = "bankers_rounding_bench",
5211 srcs = ["bench/bankers-rounding.cc"],
5212 copts = xnnpack_optional_tflite_copts(),
5213 tags = ["nowin32"],
5214 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
5215)
5216
5217xnnpack_benchmark(
5218 name = "ceiling_bench",
5219 srcs = ["bench/ceiling.cc"],
5220 copts = xnnpack_optional_tflite_copts(),
5221 tags = ["nowin32"],
5222 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
5223)
5224
5225xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005226 name = "channel_shuffle_bench",
5227 srcs = ["bench/channel-shuffle.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07005228 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005229)
5230
5231xnnpack_benchmark(
5232 name = "convolution_bench",
5233 srcs = ["bench/convolution.cc"],
5234 copts = xnnpack_optional_tflite_copts() + xnnpack_optional_armcl_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07005235 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07005236 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps() + xnnpack_optional_armcl_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07005237)
5238
5239xnnpack_benchmark(
5240 name = "deconvolution_bench",
5241 srcs = ["bench/deconvolution.cc"],
5242 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07005243 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07005244 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07005245)
5246
5247xnnpack_benchmark(
Marat Dukhanb6bd4bc2020-12-01 17:01:40 -08005248 name = "elu_bench",
5249 srcs = ["bench/elu.cc"],
5250 copts = xnnpack_optional_tflite_copts(),
5251 tags = ["nowin32"],
5252 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
5253)
5254
5255xnnpack_benchmark(
Marat Dukhan87727142020-06-24 15:24:10 -07005256 name = "floor_bench",
5257 srcs = ["bench/floor.cc"],
5258 copts = xnnpack_optional_tflite_copts(),
5259 tags = ["nowin32"],
5260 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
5261)
5262
5263xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005264 name = "global_average_pooling_bench",
5265 srcs = ["bench/global-average-pooling.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07005266 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005267)
5268
5269xnnpack_benchmark(
Marat Dukhanad352602020-06-25 21:50:54 -07005270 name = "hardswish_bench",
5271 srcs = ["bench/hardswish.cc"],
5272 copts = xnnpack_optional_tflite_copts(),
5273 tags = ["nowin32"],
5274 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
5275)
5276
5277xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005278 name = "max_pooling_bench",
5279 srcs = ["bench/max-pooling.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07005280 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005281)
5282
5283xnnpack_benchmark(
5284 name = "sigmoid_bench",
5285 srcs = ["bench/sigmoid.cc"],
Marat Dukhanc3b9e862019-11-17 13:18:54 -08005286 copts = xnnpack_optional_tflite_copts(),
Marat Dukhanca2ba702020-04-24 01:31:47 -07005287 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07005288 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07005289)
5290
5291xnnpack_benchmark(
Marat Dukhan95b22432019-10-30 16:30:14 -07005292 name = "prelu_bench",
5293 srcs = ["bench/prelu.cc"],
5294 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07005295 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07005296 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan95b22432019-10-30 16:30:14 -07005297)
5298
5299xnnpack_benchmark(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08005300 name = "softmax_bench",
5301 srcs = ["bench/softmax.cc"],
Marat Dukhan9c0db962020-01-28 12:30:14 -08005302 copts = xnnpack_optional_tflite_copts(),
Marat Dukhanca2ba702020-04-24 01:31:47 -07005303 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07005304 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07005305)
5306
Marat Dukhan87727142020-06-24 15:24:10 -07005307xnnpack_benchmark(
Marat Dukhan6804bbd2020-06-30 19:26:11 -07005308 name = "square_root_bench",
5309 srcs = ["bench/square-root.cc"],
5310 copts = xnnpack_optional_tflite_copts(),
5311 tags = ["nowin32"],
5312 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
5313)
5314
5315xnnpack_benchmark(
Marat Dukhan87727142020-06-24 15:24:10 -07005316 name = "truncation_bench",
5317 srcs = ["bench/truncation.cc"],
5318 deps = OPERATOR_BENCHMARK_DEPS,
5319)
5320
Marat Dukhanc068bb62019-10-04 13:24:39 -07005321############################# End-to-end benchmarks ############################
5322
5323cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07005324 name = "fp32_mobilenet_v1",
5325 srcs = ["models/fp32-mobilenet-v1.cc"],
Marat Dukhanc068bb62019-10-04 13:24:39 -07005326 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08005327 copts = xnnpack_std_cxxopts(),
Marat Dukhanc068bb62019-10-04 13:24:39 -07005328 linkstatic = True,
5329 deps = [
5330 ":XNNPACK",
5331 "@pthreadpool",
5332 ],
5333)
5334
5335cc_library(
Marat Dukhan4cea2322021-03-09 09:35:36 -08005336 name = "fp32_sparse_mobilenet_v1",
5337 srcs = ["models/fp32-sparse-mobilenet-v1.cc"],
5338 hdrs = ["models/models.h"],
5339 copts = xnnpack_std_cxxopts(),
5340 linkstatic = True,
5341 deps = [
5342 ":XNNPACK",
5343 "@pthreadpool",
5344 ],
5345)
5346
5347cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07005348 name = "fp16_mobilenet_v1",
5349 srcs = ["models/fp16-mobilenet-v1.cc"],
5350 hdrs = ["models/models.h"],
5351 copts = xnnpack_std_cxxopts(),
5352 linkstatic = True,
5353 deps = [
5354 ":XNNPACK",
5355 "@FP16",
5356 "@pthreadpool",
5357 ],
5358)
5359
5360cc_library(
Marat Dukhan0743cdf2020-08-04 18:52:07 -07005361 name = "qs8_mobilenet_v1",
5362 srcs = ["models/qs8-mobilenet-v1.cc"],
5363 hdrs = ["models/models.h"],
5364 copts = xnnpack_std_cxxopts(),
5365 linkstatic = True,
5366 deps = [
5367 ":XNNPACK",
5368 "@pthreadpool",
5369 ],
5370)
5371
5372cc_library(
Marat Dukhan70a96182020-09-03 17:13:58 -07005373 name = "qs8_mobilenet_v2",
5374 srcs = ["models/qs8-mobilenet-v2.cc"],
5375 hdrs = ["models/models.h"],
5376 copts = xnnpack_std_cxxopts(),
5377 linkstatic = True,
5378 deps = [
5379 ":XNNPACK",
5380 "@pthreadpool",
5381 ],
5382)
5383
5384cc_library(
Marat Dukhan12a23bb2021-03-08 08:13:21 -08005385 name = "qu8_mobilenet_v1",
5386 srcs = ["models/qu8-mobilenet-v1.cc"],
5387 hdrs = ["models/models.h"],
5388 copts = xnnpack_std_cxxopts(),
5389 linkstatic = True,
5390 deps = [
5391 ":XNNPACK",
5392 "@pthreadpool",
5393 ],
5394)
5395
5396cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07005397 name = "fp32_mobilenet_v2",
5398 srcs = ["models/fp32-mobilenet-v2.cc"],
Marat Dukhanc068bb62019-10-04 13:24:39 -07005399 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08005400 copts = xnnpack_std_cxxopts(),
Marat Dukhanc068bb62019-10-04 13:24:39 -07005401 linkstatic = True,
5402 deps = [
5403 ":XNNPACK",
5404 "@pthreadpool",
5405 ],
5406)
5407
Marat Dukhanc08cdf52019-12-09 09:17:51 -08005408cc_library(
Marat Dukhan4cea2322021-03-09 09:35:36 -08005409 name = "fp32_sparse_mobilenet_v2",
5410 srcs = ["models/fp32-sparse-mobilenet-v2.cc"],
5411 hdrs = ["models/models.h"],
5412 copts = xnnpack_std_cxxopts(),
5413 linkstatic = True,
5414 deps = [
5415 ":XNNPACK",
5416 "@pthreadpool",
5417 ],
5418)
5419
5420cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07005421 name = "fp16_mobilenet_v2",
5422 srcs = ["models/fp16-mobilenet-v2.cc"],
5423 hdrs = ["models/models.h"],
5424 copts = xnnpack_std_cxxopts(),
5425 linkstatic = True,
5426 deps = [
5427 ":XNNPACK",
5428 "@FP16",
5429 "@pthreadpool",
5430 ],
5431)
5432
5433cc_library(
5434 name = "fp32_mobilenet_v3_large",
5435 srcs = ["models/fp32-mobilenet-v3-large.cc"],
Marat Dukhanc08cdf52019-12-09 09:17:51 -08005436 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08005437 copts = xnnpack_std_cxxopts(),
Marat Dukhanc08cdf52019-12-09 09:17:51 -08005438 linkstatic = True,
5439 deps = [
5440 ":XNNPACK",
5441 "@pthreadpool",
5442 ],
5443)
5444
5445cc_library(
Marat Dukhan4cea2322021-03-09 09:35:36 -08005446 name = "fp32_sparse_mobilenet_v3_large",
5447 srcs = ["models/fp32-sparse-mobilenet-v3-large.cc"],
5448 hdrs = ["models/models.h"],
5449 copts = xnnpack_std_cxxopts(),
5450 linkstatic = True,
5451 deps = [
5452 ":XNNPACK",
5453 "@pthreadpool",
5454 ],
5455)
5456
5457cc_library(
Marat Dukhan66f3ccd2020-09-14 16:09:38 -07005458 name = "fp16_mobilenet_v3_large",
5459 srcs = ["models/fp16-mobilenet-v3-large.cc"],
5460 hdrs = ["models/models.h"],
5461 copts = xnnpack_std_cxxopts(),
5462 linkstatic = True,
5463 deps = [
5464 ":XNNPACK",
5465 "@FP16",
5466 "@pthreadpool",
5467 ],
5468)
5469
5470cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07005471 name = "fp32_mobilenet_v3_small",
5472 srcs = ["models/fp32-mobilenet-v3-small.cc"],
Marat Dukhanc08cdf52019-12-09 09:17:51 -08005473 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08005474 copts = xnnpack_std_cxxopts(),
Marat Dukhanc08cdf52019-12-09 09:17:51 -08005475 linkstatic = True,
5476 deps = [
5477 ":XNNPACK",
5478 "@pthreadpool",
5479 ],
5480)
5481
Marat Dukhan66f3ccd2020-09-14 16:09:38 -07005482cc_library(
Marat Dukhan4cea2322021-03-09 09:35:36 -08005483 name = "fp32_sparse_mobilenet_v3_small",
5484 srcs = ["models/fp32-sparse-mobilenet-v3-small.cc"],
5485 hdrs = ["models/models.h"],
5486 copts = xnnpack_std_cxxopts(),
5487 linkstatic = True,
5488 deps = [
5489 ":XNNPACK",
5490 "@pthreadpool",
5491 ],
5492)
5493
5494cc_library(
Marat Dukhan66f3ccd2020-09-14 16:09:38 -07005495 name = "fp16_mobilenet_v3_small",
5496 srcs = ["models/fp16-mobilenet-v3-small.cc"],
5497 hdrs = ["models/models.h"],
5498 copts = xnnpack_std_cxxopts(),
5499 linkstatic = True,
5500 deps = [
5501 ":XNNPACK",
5502 "@FP16",
5503 "@pthreadpool",
5504 ],
5505)
5506
Marat Dukhanc068bb62019-10-04 13:24:39 -07005507xnnpack_benchmark(
Marat Dukhanef4416e2019-10-31 13:44:40 -07005508 name = "f32_dwconv_e2e_bench",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08005509 srcs = [
5510 "bench/f32-dwconv-e2e.cc",
5511 "bench/end2end.h",
5512 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanef4416e2019-10-31 13:44:40 -07005513 deps = MICROKERNEL_BENCHMARK_DEPS + [
5514 ":XNNPACK",
Marat Dukhan270a2c42020-06-26 16:45:52 -07005515 ":fp32_mobilenet_v1",
5516 ":fp32_mobilenet_v2",
5517 ":fp32_mobilenet_v3_large",
5518 ":fp32_mobilenet_v3_small",
Marat Dukhanef4416e2019-10-31 13:44:40 -07005519 ],
5520)
5521
5522xnnpack_benchmark(
Marat Dukhan5f18d262019-10-31 10:24:14 -07005523 name = "f32_gemm_e2e_bench",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08005524 srcs = [
5525 "bench/f32-gemm-e2e.cc",
5526 "bench/end2end.h",
5527 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan5f18d262019-10-31 10:24:14 -07005528 deps = MICROKERNEL_BENCHMARK_DEPS + [
5529 ":XNNPACK",
Marat Dukhan270a2c42020-06-26 16:45:52 -07005530 ":fp32_mobilenet_v1",
5531 ":fp32_mobilenet_v2",
5532 ":fp32_mobilenet_v3_large",
5533 ":fp32_mobilenet_v3_small",
Marat Dukhan5f18d262019-10-31 10:24:14 -07005534 ],
5535)
5536
5537xnnpack_benchmark(
Frank Barcharddc909cb2021-02-08 13:59:31 -08005538 name = "qs8_gemm_e2e_bench",
5539 srcs = [
5540 "bench/qs8-gemm-e2e.cc",
5541 "bench/end2end.h",
5542 ] + MICROKERNEL_BENCHMARK_HDRS,
5543 deps = MICROKERNEL_BENCHMARK_DEPS + [
5544 ":XNNPACK",
5545 ":qs8_mobilenet_v1",
5546 ":qs8_mobilenet_v2",
5547 ],
5548)
5549
5550xnnpack_benchmark(
Marat Dukhanc068bb62019-10-04 13:24:39 -07005551 name = "end2end_bench",
5552 srcs = ["bench/end2end.cc"],
5553 deps = [
5554 ":XNNPACK",
Frank Barchardc712fa42019-10-31 14:00:21 -07005555 ":bench_utils",
Marat Dukhan270a2c42020-06-26 16:45:52 -07005556 ":fp16_mobilenet_v1",
5557 ":fp16_mobilenet_v2",
Marat Dukhan66f3ccd2020-09-14 16:09:38 -07005558 ":fp16_mobilenet_v3_large",
5559 ":fp16_mobilenet_v3_small",
Marat Dukhan270a2c42020-06-26 16:45:52 -07005560 ":fp32_mobilenet_v1",
5561 ":fp32_mobilenet_v2",
5562 ":fp32_mobilenet_v3_large",
5563 ":fp32_mobilenet_v3_small",
Marat Dukhan4cea2322021-03-09 09:35:36 -08005564 ":fp32_sparse_mobilenet_v1",
5565 ":fp32_sparse_mobilenet_v2",
5566 ":fp32_sparse_mobilenet_v3_large",
5567 ":fp32_sparse_mobilenet_v3_small",
Marat Dukhan0743cdf2020-08-04 18:52:07 -07005568 ":qs8_mobilenet_v1",
Marat Dukhan70a96182020-09-03 17:13:58 -07005569 ":qs8_mobilenet_v2",
Marat Dukhan12a23bb2021-03-08 08:13:21 -08005570 ":qu8_mobilenet_v1",
Marat Dukhanc068bb62019-10-04 13:24:39 -07005571 "@pthreadpool",
5572 ],
5573)
5574
Marat Dukhan6adff4e2019-10-14 18:32:07 -07005575#################### Accuracy evaluation for math functions ####################
5576
5577xnnpack_benchmark(
Marat Dukhancf67ec62020-12-14 09:56:37 -08005578 name = "f32_exp_ulp_eval",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07005579 srcs = [
Marat Dukhancf67ec62020-12-14 09:56:37 -08005580 "eval/f32-exp-ulp.cc",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07005581 "src/xnnpack/AlignedAllocator.h",
5582 ] + ACCURACY_EVAL_HDRS,
Marat Dukhan3a305212020-12-06 19:24:27 -08005583 deps = ACCURACY_EVAL_DEPS + [
5584 ":bench_utils",
5585 "@cpuinfo",
5586 ],
Marat Dukhan6adff4e2019-10-14 18:32:07 -07005587)
5588
Marat Dukhan515c9772019-10-17 18:07:57 -07005589xnnpack_benchmark(
Marat Dukhancf67ec62020-12-14 09:56:37 -08005590 name = "f32_expminus_ulp_eval",
Marat Dukhan515c9772019-10-17 18:07:57 -07005591 srcs = [
Marat Dukhancf67ec62020-12-14 09:56:37 -08005592 "eval/f32-expminus-ulp.cc",
Marat Dukhan515c9772019-10-17 18:07:57 -07005593 "src/xnnpack/AlignedAllocator.h",
5594 ] + ACCURACY_EVAL_HDRS,
Marat Dukhan3a305212020-12-06 19:24:27 -08005595 deps = ACCURACY_EVAL_DEPS + [
5596 ":bench_utils",
5597 "@cpuinfo",
5598 ],
Marat Dukhan515c9772019-10-17 18:07:57 -07005599)
5600
Marat Dukhan98ba4412019-10-23 02:14:28 -07005601xnnpack_benchmark(
Marat Dukhancf67ec62020-12-14 09:56:37 -08005602 name = "f32_expm1minus_ulp_eval",
Marat Dukhana438aca2020-11-20 15:45:01 -08005603 srcs = [
Marat Dukhancf67ec62020-12-14 09:56:37 -08005604 "eval/f32-expm1minus-ulp.cc",
Marat Dukhana438aca2020-11-20 15:45:01 -08005605 "src/xnnpack/AlignedAllocator.h",
5606 ] + ACCURACY_EVAL_HDRS,
Marat Dukhande390d42020-11-29 19:32:18 -08005607 deps = ACCURACY_EVAL_DEPS + [
5608 ":bench_utils",
Marat Dukhan3a305212020-12-06 19:24:27 -08005609 "@cpuinfo",
Marat Dukhande390d42020-11-29 19:32:18 -08005610 ],
Marat Dukhana438aca2020-11-20 15:45:01 -08005611)
5612
5613xnnpack_benchmark(
Marat Dukhancf67ec62020-12-14 09:56:37 -08005614 name = "f32_extexp_ulp_eval",
Marat Dukhan98ba4412019-10-23 02:14:28 -07005615 srcs = [
Marat Dukhancf67ec62020-12-14 09:56:37 -08005616 "eval/f32-extexp-ulp.cc",
Marat Dukhan98ba4412019-10-23 02:14:28 -07005617 "src/xnnpack/AlignedAllocator.h",
5618 ] + ACCURACY_EVAL_HDRS,
Marat Dukhan3a305212020-12-06 19:24:27 -08005619 deps = ACCURACY_EVAL_DEPS + [
5620 ":bench_utils",
5621 "@cpuinfo",
5622 ],
Marat Dukhan98ba4412019-10-23 02:14:28 -07005623)
5624
Marat Dukhanf44f0222020-12-14 11:53:27 -08005625xnnpack_benchmark(
5626 name = "f32_sigmoid_ulp_eval",
5627 srcs = [
5628 "eval/f32-sigmoid-ulp.cc",
5629 "src/xnnpack/AlignedAllocator.h",
5630 ] + ACCURACY_EVAL_HDRS,
5631 deps = ACCURACY_EVAL_DEPS + [
5632 ":bench_utils",
5633 "@cpuinfo",
5634 ],
5635)
5636
5637xnnpack_benchmark(
5638 name = "f32_sqrt_ulp_eval",
5639 srcs = [
5640 "eval/f32-sqrt-ulp.cc",
5641 "src/xnnpack/AlignedAllocator.h",
5642 ] + ACCURACY_EVAL_HDRS,
5643 deps = ACCURACY_EVAL_DEPS + [
5644 ":bench_utils",
5645 "@cpuinfo",
5646 ],
5647)
5648
5649################### Accuracy verification for math functions ##################
5650
5651xnnpack_unit_test(
Marat Dukhanf7291fc2020-12-15 11:02:50 -08005652 name = "f32_exp_eval",
5653 srcs = [
5654 "eval/f32-exp.cc",
5655 "src/xnnpack/AlignedAllocator.h",
5656 "src/xnnpack/math-stubs.h",
5657 ] + MICROKERNEL_TEST_HDRS,
5658 automatic = False,
5659 deps = MICROKERNEL_TEST_DEPS,
5660)
5661
5662xnnpack_unit_test(
Marat Dukhanf44f0222020-12-14 11:53:27 -08005663 name = "f32_expm1minus_eval",
5664 srcs = [
5665 "eval/f32-expm1minus.cc",
5666 "src/xnnpack/AlignedAllocator.h",
5667 "src/xnnpack/math-stubs.h",
5668 ] + MICROKERNEL_TEST_HDRS,
5669 automatic = False,
5670 deps = MICROKERNEL_TEST_DEPS,
5671)
5672
Marat Dukhan8853b822020-05-07 12:19:01 -07005673xnnpack_unit_test(
Marat Dukhand28a5a22020-12-14 15:27:22 -08005674 name = "f32_expminus_eval",
5675 srcs = [
5676 "eval/f32-expminus.cc",
5677 "src/xnnpack/AlignedAllocator.h",
5678 "src/xnnpack/math-stubs.h",
5679 ] + MICROKERNEL_TEST_HDRS,
5680 automatic = False,
5681 deps = MICROKERNEL_TEST_DEPS,
5682)
5683
5684xnnpack_unit_test(
Marat Dukhan8853b822020-05-07 12:19:01 -07005685 name = "f32_roundne_eval",
5686 srcs = [
5687 "eval/f32-roundne.cc",
5688 "src/xnnpack/AlignedAllocator.h",
5689 "src/xnnpack/math-stubs.h",
5690 ] + MICROKERNEL_TEST_HDRS,
Marat Dukhan22eed3d2020-05-11 20:13:37 -07005691 automatic = False,
Marat Dukhan8853b822020-05-07 12:19:01 -07005692 deps = MICROKERNEL_TEST_DEPS,
5693)
5694
Marat Dukhan2dbb9442020-05-12 20:43:43 -07005695xnnpack_unit_test(
Marat Dukhanc9852ba2020-05-13 17:21:29 -07005696 name = "f32_roundd_eval",
5697 srcs = [
5698 "eval/f32-roundd.cc",
5699 "src/xnnpack/AlignedAllocator.h",
5700 "src/xnnpack/math-stubs.h",
5701 ] + MICROKERNEL_TEST_HDRS,
5702 automatic = False,
5703 deps = MICROKERNEL_TEST_DEPS,
5704)
5705
5706xnnpack_unit_test(
5707 name = "f32_roundu_eval",
5708 srcs = [
5709 "eval/f32-roundu.cc",
5710 "src/xnnpack/AlignedAllocator.h",
5711 "src/xnnpack/math-stubs.h",
5712 ] + MICROKERNEL_TEST_HDRS,
5713 automatic = False,
5714 deps = MICROKERNEL_TEST_DEPS,
5715)
5716
5717xnnpack_unit_test(
Marat Dukhan2dbb9442020-05-12 20:43:43 -07005718 name = "f32_roundz_eval",
5719 srcs = [
5720 "eval/f32-roundz.cc",
5721 "src/xnnpack/AlignedAllocator.h",
5722 "src/xnnpack/math-stubs.h",
5723 ] + MICROKERNEL_TEST_HDRS,
Marat Dukhanc9852ba2020-05-13 17:21:29 -07005724 automatic = False,
Marat Dukhan2dbb9442020-05-12 20:43:43 -07005725 deps = MICROKERNEL_TEST_DEPS,
5726)
5727
Marat Dukhan08c4a432019-10-03 09:29:21 -07005728######################### Unit tests for micro-kernels #########################
5729
5730xnnpack_unit_test(
Marat Dukhande06f492020-04-09 00:19:31 -07005731 name = "f16_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005732 srcs = [
Marat Dukhande06f492020-04-09 00:19:31 -07005733 "test/f16-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005734 "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 Dukhan08c4a432019-10-03 09:29:21 -07005738)
5739
5740xnnpack_unit_test(
Marat Dukhan355ab432020-04-09 19:01:52 -07005741 name = "f16_spmm_minmax_test",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08005742 srcs = [
Marat Dukhan355ab432020-04-09 19:01:52 -07005743 "test/f16-spmm-minmax.cc",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08005744 "test/spmm-microkernel-tester.h",
5745 "src/xnnpack/AlignedAllocator.h",
5746 ] + MICROKERNEL_TEST_HDRS,
5747 deps = MICROKERNEL_TEST_DEPS,
5748)
5749
5750xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005751 name = "f32_argmaxpool_test",
5752 srcs = [
5753 "test/f32-argmaxpool.cc",
5754 "test/argmaxpool-microkernel-tester.h",
5755 "src/xnnpack/AlignedAllocator.h",
5756 ] + MICROKERNEL_TEST_HDRS,
5757 deps = MICROKERNEL_TEST_DEPS,
5758)
5759
5760xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07005761 name = "f32_avgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005762 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07005763 "test/f32-avgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005764 "test/avgpool-microkernel-tester.h",
5765 "src/xnnpack/AlignedAllocator.h",
5766 ] + MICROKERNEL_TEST_HDRS,
5767 deps = MICROKERNEL_TEST_DEPS,
5768)
5769
5770xnnpack_unit_test(
Marat Dukhan660fd192020-03-10 04:55:30 -07005771 name = "f32_ibilinear_test",
Marat Dukhan35dacfb2019-11-07 19:18:16 -08005772 srcs = [
Marat Dukhan660fd192020-03-10 04:55:30 -07005773 "test/f32-ibilinear.cc",
5774 "test/ibilinear-microkernel-tester.h",
Marat Dukhan35dacfb2019-11-07 19:18:16 -08005775 "src/xnnpack/AlignedAllocator.h",
5776 ] + MICROKERNEL_TEST_HDRS,
5777 deps = MICROKERNEL_TEST_DEPS,
5778)
5779
5780xnnpack_unit_test(
XNNPACK Team21432672020-10-19 19:58:48 -07005781 name = "f32_ibilinear_chw_test",
5782 srcs = [
5783 "test/f32-ibilinear-chw.cc",
XNNPACK Team6be46b22020-10-22 23:34:54 -07005784 "test/ibilinear-microkernel-tester.h",
XNNPACK Team21432672020-10-19 19:58:48 -07005785 "src/xnnpack/AlignedAllocator.h",
5786 ] + MICROKERNEL_TEST_HDRS,
5787 deps = MICROKERNEL_TEST_DEPS,
5788)
5789
5790xnnpack_unit_test(
Frank Barchardb1966592020-05-12 13:47:06 -07005791 name = "f16_clamp_test",
5792 srcs = [
5793 "test/f16-clamp.cc",
5794 "test/clamp-microkernel-tester.h",
5795 ] + MICROKERNEL_TEST_HDRS,
5796 deps = MICROKERNEL_TEST_DEPS,
5797)
5798
5799xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005800 name = "f32_clamp_test",
5801 srcs = [
5802 "test/f32-clamp.cc",
5803 "test/clamp-microkernel-tester.h",
5804 ] + MICROKERNEL_TEST_HDRS,
5805 deps = MICROKERNEL_TEST_DEPS,
5806)
5807
5808xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07005809 name = "f32_igemm_test",
5810 srcs = [
5811 "test/f32-igemm.cc",
5812 "test/gemm-microkernel-tester.h",
5813 "src/xnnpack/AlignedAllocator.h",
5814 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005815 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan163a7e62020-04-09 04:19:26 -07005816)
5817
5818xnnpack_unit_test(
Marat Dukhan467f6362020-05-22 23:21:55 -07005819 name = "f32_igemm_relu_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005820 srcs = [
Marat Dukhan467f6362020-05-22 23:21:55 -07005821 "test/f32-igemm-relu.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005822 "test/gemm-microkernel-tester.h",
5823 "src/xnnpack/AlignedAllocator.h",
5824 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005825 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005826)
5827
5828xnnpack_unit_test(
Marat Dukhane207b7b2020-05-28 16:27:42 -07005829 name = "f32_igemm_minmax_test",
5830 srcs = [
5831 "test/f32-igemm-minmax.cc",
5832 "test/gemm-microkernel-tester.h",
5833 "src/xnnpack/AlignedAllocator.h",
5834 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005835 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhane207b7b2020-05-28 16:27:42 -07005836)
5837
5838xnnpack_unit_test(
Frank Barchardb0e4fae2020-05-04 15:27:51 -07005839 name = "f16_igemm_minmax_test",
5840 srcs = [
5841 "test/f16-igemm-minmax.cc",
5842 "test/gemm-microkernel-tester.h",
5843 "src/xnnpack/AlignedAllocator.h",
5844 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005845 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Frank Barchardb0e4fae2020-05-04 15:27:51 -07005846)
5847
5848xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005849 name = "f32_conv_hwc_test",
5850 srcs = [
5851 "test/f32-conv-hwc.cc",
5852 "test/conv-hwc-microkernel-tester.h",
5853 "src/xnnpack/AlignedAllocator.h",
5854 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005855 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005856)
5857
5858xnnpack_unit_test(
Marat Dukhan1f29b802020-05-15 23:46:39 -07005859 name = "f32_conv_hwc2chw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005860 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07005861 "test/f32-conv-hwc2chw.cc",
5862 "test/conv-hwc2chw-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005863 "src/xnnpack/AlignedAllocator.h",
5864 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005865 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005866)
5867
5868xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07005869 name = "f32_dwconv_test",
5870 srcs = [
5871 "test/f32-dwconv.cc",
5872 "test/dwconv-microkernel-tester.h",
5873 "src/xnnpack/AlignedAllocator.h",
5874 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005875 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan163a7e62020-04-09 04:19:26 -07005876)
5877
5878xnnpack_unit_test(
Frank Barchard5a599a62020-06-04 20:12:44 -07005879 name = "f16_dwconv_minmax_test",
5880 srcs = [
5881 "test/f16-dwconv-minmax.cc",
5882 "test/dwconv-microkernel-tester.h",
5883 "src/xnnpack/AlignedAllocator.h",
5884 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005885 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Frank Barchard5a599a62020-06-04 20:12:44 -07005886)
5887
5888xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07005889 name = "f32_dwconv_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005890 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07005891 "test/f32-dwconv-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005892 "test/dwconv-microkernel-tester.h",
5893 "src/xnnpack/AlignedAllocator.h",
5894 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005895 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005896)
5897
5898xnnpack_unit_test(
Marat Dukhanbf715f92020-10-23 20:17:00 -07005899 name = "f32_dwconv2d_chw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005900 srcs = [
Marat Dukhanbf715f92020-10-23 20:17:00 -07005901 "test/f32-dwconv2d-chw.cc",
5902 "test/dwconv2d-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005903 "src/xnnpack/AlignedAllocator.h",
5904 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005905 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005906)
5907
5908xnnpack_unit_test(
Frank Barchard0bb49a72020-06-04 11:35:11 -07005909 name = "f16_gavgpool_minmax_test",
5910 srcs = [
5911 "test/f16-gavgpool-minmax.cc",
5912 "test/gavgpool-microkernel-tester.h",
5913 "src/xnnpack/AlignedAllocator.h",
5914 ] + MICROKERNEL_TEST_HDRS,
5915 deps = MICROKERNEL_TEST_DEPS,
5916)
5917
5918xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07005919 name = "f32_gavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005920 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07005921 "test/f32-gavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005922 "test/gavgpool-microkernel-tester.h",
5923 "src/xnnpack/AlignedAllocator.h",
5924 ] + MICROKERNEL_TEST_HDRS,
5925 deps = MICROKERNEL_TEST_DEPS,
5926)
5927
5928xnnpack_unit_test(
Marat Dukhan1f29b802020-05-15 23:46:39 -07005929 name = "f32_gavgpool_cw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005930 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07005931 "test/f32-gavgpool-cw.cc",
5932 "test/gavgpool-cw-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005933 "src/xnnpack/AlignedAllocator.h",
5934 ] + MICROKERNEL_TEST_HDRS,
5935 deps = MICROKERNEL_TEST_DEPS,
5936)
5937
5938xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07005939 name = "f32_gemm_test",
5940 srcs = [
5941 "test/f32-gemm.cc",
5942 "test/gemm-microkernel-tester.h",
5943 "src/xnnpack/AlignedAllocator.h",
5944 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005945 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan163a7e62020-04-09 04:19:26 -07005946)
5947
5948xnnpack_unit_test(
Marat Dukhan467f6362020-05-22 23:21:55 -07005949 name = "f32_gemm_relu_test",
5950 srcs = [
5951 "test/f32-gemm-relu.cc",
5952 "test/gemm-microkernel-tester.h",
5953 "src/xnnpack/AlignedAllocator.h",
5954 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005955 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan467f6362020-05-22 23:21:55 -07005956)
5957
5958xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07005959 name = "f32_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005960 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07005961 "test/f32-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005962 "test/gemm-microkernel-tester.h",
5963 "src/xnnpack/AlignedAllocator.h",
5964 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005965 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005966)
5967
5968xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07005969 name = "f32_gemminc_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005970 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07005971 "test/f32-gemminc-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005972 "test/gemm-microkernel-tester.h",
5973 "src/xnnpack/AlignedAllocator.h",
5974 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005975 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005976)
5977
5978xnnpack_unit_test(
Frank Barchardb1966592020-05-12 13:47:06 -07005979 name = "f16_hswish_test",
5980 srcs = [
5981 "test/f16-hswish.cc",
5982 "test/hswish-microkernel-tester.h",
5983 ] + MICROKERNEL_TEST_HDRS,
5984 deps = MICROKERNEL_TEST_DEPS,
5985)
5986
5987xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005988 name = "f32_hswish_test",
5989 srcs = [
5990 "test/f32-hswish.cc",
5991 "test/hswish-microkernel-tester.h",
5992 ] + MICROKERNEL_TEST_HDRS,
5993 deps = MICROKERNEL_TEST_DEPS,
5994)
5995
5996xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07005997 name = "f32_maxpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005998 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07005999 "test/f32-maxpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006000 "test/maxpool-microkernel-tester.h",
6001 ] + MICROKERNEL_TEST_HDRS,
6002 deps = MICROKERNEL_TEST_DEPS,
6003)
6004
6005xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07006006 name = "f32_pavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006007 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07006008 "test/f32-pavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006009 "test/avgpool-microkernel-tester.h",
6010 "src/xnnpack/AlignedAllocator.h",
6011 ] + MICROKERNEL_TEST_HDRS,
6012 deps = MICROKERNEL_TEST_DEPS,
6013)
6014
6015xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07006016 name = "f32_ppmm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006017 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07006018 "test/f32-ppmm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006019 "test/gemm-microkernel-tester.h",
6020 "src/xnnpack/AlignedAllocator.h",
6021 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07006022 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07006023)
6024
6025xnnpack_unit_test(
Frank Barchardb1966592020-05-12 13:47:06 -07006026 name = "f16_prelu_test",
6027 srcs = [
6028 "test/f16-prelu.cc",
6029 "test/prelu-microkernel-tester.h",
6030 "src/xnnpack/AlignedAllocator.h",
6031 ] + MICROKERNEL_TEST_HDRS,
6032 deps = MICROKERNEL_TEST_DEPS,
6033)
6034
6035xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07006036 name = "f32_prelu_test",
6037 srcs = [
6038 "test/f32-prelu.cc",
6039 "test/prelu-microkernel-tester.h",
6040 "src/xnnpack/AlignedAllocator.h",
6041 ] + MICROKERNEL_TEST_HDRS,
6042 deps = MICROKERNEL_TEST_DEPS,
6043)
6044
6045xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07006046 name = "f32_raddexpminusmax_test",
6047 srcs = [
6048 "test/f32-raddexpminusmax.cc",
6049 "test/raddexpminusmax-microkernel-tester.h",
6050 ] + MICROKERNEL_TEST_HDRS,
6051 deps = MICROKERNEL_TEST_DEPS,
6052)
6053
6054xnnpack_unit_test(
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07006055 name = "f32_raddextexp_test",
6056 srcs = [
6057 "test/f32-raddextexp.cc",
6058 "test/raddextexp-microkernel-tester.h",
6059 ] + MICROKERNEL_TEST_HDRS,
6060 deps = MICROKERNEL_TEST_DEPS,
6061)
6062
6063xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07006064 name = "f32_raddstoreexpminusmax_test",
6065 srcs = [
6066 "test/f32-raddstoreexpminusmax.cc",
6067 "test/raddstoreexpminusmax-microkernel-tester.h",
6068 ] + MICROKERNEL_TEST_HDRS,
6069 deps = MICROKERNEL_TEST_DEPS,
6070)
6071
6072xnnpack_unit_test(
Frank Barchardfb158e22020-07-15 16:10:10 -07006073 name = "f32_relu_test",
6074 srcs = [
6075 "test/f32-relu.cc",
6076 "test/vunary-microkernel-tester.h",
6077 ] + MICROKERNEL_TEST_HDRS,
6078 deps = MICROKERNEL_TEST_DEPS,
6079)
6080
6081xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07006082 name = "f32_rmax_test",
6083 srcs = [
6084 "test/f32-rmax.cc",
6085 "test/rmax-microkernel-tester.h",
6086 ] + MICROKERNEL_TEST_HDRS,
6087 deps = MICROKERNEL_TEST_DEPS,
6088)
6089
6090xnnpack_unit_test(
Marat Dukhan346a9e52019-11-15 09:06:30 -08006091 name = "f32_sigmoid_test",
6092 srcs = [
6093 "test/f32-sigmoid.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08006094 "test/vunary-microkernel-tester.h",
Marat Dukhan346a9e52019-11-15 09:06:30 -08006095 ] + MICROKERNEL_TEST_HDRS,
6096 deps = MICROKERNEL_TEST_DEPS,
6097)
6098
6099xnnpack_unit_test(
Marat Dukhan355ab432020-04-09 19:01:52 -07006100 name = "f32_spmm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006101 srcs = [
Marat Dukhan355ab432020-04-09 19:01:52 -07006102 "test/f32-spmm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006103 "test/spmm-microkernel-tester.h",
6104 "src/xnnpack/AlignedAllocator.h",
6105 ] + MICROKERNEL_TEST_HDRS,
6106 deps = MICROKERNEL_TEST_DEPS,
6107)
6108
6109xnnpack_unit_test(
Frank Barchardd793f6c2020-05-08 13:37:43 -07006110 name = "f16_vadd_minmax_test",
6111 srcs = [
6112 "test/f16-vadd-minmax.cc",
6113 "test/vbinary-microkernel-tester.h",
6114 ] + MICROKERNEL_TEST_HDRS,
6115 deps = MICROKERNEL_TEST_DEPS,
6116)
6117
6118xnnpack_unit_test(
6119 name = "f16_vaddc_minmax_test",
6120 srcs = [
6121 "test/f16-vaddc-minmax.cc",
6122 "test/vbinaryc-microkernel-tester.h",
6123 ] + MICROKERNEL_TEST_HDRS,
6124 deps = MICROKERNEL_TEST_DEPS,
6125)
6126
6127xnnpack_unit_test(
6128 name = "f16_vdiv_minmax_test",
6129 srcs = [
6130 "test/f16-vdiv-minmax.cc",
6131 "test/vbinary-microkernel-tester.h",
6132 ] + MICROKERNEL_TEST_HDRS,
6133 deps = MICROKERNEL_TEST_DEPS,
6134)
6135
6136xnnpack_unit_test(
6137 name = "f16_vdivc_minmax_test",
6138 srcs = [
6139 "test/f16-vdivc-minmax.cc",
6140 "test/vbinaryc-microkernel-tester.h",
6141 ] + MICROKERNEL_TEST_HDRS,
6142 deps = MICROKERNEL_TEST_DEPS,
6143)
6144
6145xnnpack_unit_test(
6146 name = "f16_vrdivc_minmax_test",
6147 srcs = [
6148 "test/f16-vrdivc-minmax.cc",
6149 "test/vbinaryc-microkernel-tester.h",
6150 ] + MICROKERNEL_TEST_HDRS,
6151 deps = MICROKERNEL_TEST_DEPS,
6152)
6153
6154xnnpack_unit_test(
6155 name = "f16_vmax_test",
6156 srcs = [
6157 "test/f16-vmax.cc",
6158 "test/vbinary-microkernel-tester.h",
6159 ] + MICROKERNEL_TEST_HDRS,
6160 deps = MICROKERNEL_TEST_DEPS,
6161)
6162
6163xnnpack_unit_test(
6164 name = "f16_vmaxc_test",
6165 srcs = [
6166 "test/f16-vmaxc.cc",
6167 "test/vbinaryc-microkernel-tester.h",
6168 ] + MICROKERNEL_TEST_HDRS,
6169 deps = MICROKERNEL_TEST_DEPS,
6170)
6171
6172xnnpack_unit_test(
6173 name = "f16_vmin_test",
6174 srcs = [
6175 "test/f16-vmin.cc",
6176 "test/vbinary-microkernel-tester.h",
6177 ] + MICROKERNEL_TEST_HDRS,
6178 deps = MICROKERNEL_TEST_DEPS,
6179)
6180
6181xnnpack_unit_test(
6182 name = "f16_vminc_test",
6183 srcs = [
6184 "test/f16-vminc.cc",
6185 "test/vbinaryc-microkernel-tester.h",
6186 ] + MICROKERNEL_TEST_HDRS,
6187 deps = MICROKERNEL_TEST_DEPS,
6188)
6189
6190xnnpack_unit_test(
6191 name = "f16_vmul_minmax_test",
6192 srcs = [
6193 "test/f16-vmul-minmax.cc",
6194 "test/vbinary-microkernel-tester.h",
6195 ] + MICROKERNEL_TEST_HDRS,
6196 deps = MICROKERNEL_TEST_DEPS,
6197)
6198
6199xnnpack_unit_test(
6200 name = "f16_vmulc_minmax_test",
6201 srcs = [
6202 "test/f16-vmulc-minmax.cc",
6203 "test/vbinaryc-microkernel-tester.h",
6204 ] + MICROKERNEL_TEST_HDRS,
6205 deps = MICROKERNEL_TEST_DEPS,
6206)
6207
6208xnnpack_unit_test(
6209 name = "f16_vsub_minmax_test",
6210 srcs = [
6211 "test/f16-vsub-minmax.cc",
6212 "test/vbinary-microkernel-tester.h",
6213 ] + MICROKERNEL_TEST_HDRS,
6214 deps = MICROKERNEL_TEST_DEPS,
6215)
6216
6217xnnpack_unit_test(
6218 name = "f16_vsubc_minmax_test",
6219 srcs = [
6220 "test/f16-vsubc-minmax.cc",
6221 "test/vbinaryc-microkernel-tester.h",
6222 ] + MICROKERNEL_TEST_HDRS,
6223 deps = MICROKERNEL_TEST_DEPS,
6224)
6225
6226xnnpack_unit_test(
6227 name = "f16_vrsubc_minmax_test",
6228 srcs = [
6229 "test/f16-vrsubc-minmax.cc",
6230 "test/vbinaryc-microkernel-tester.h",
6231 ] + MICROKERNEL_TEST_HDRS,
6232 deps = MICROKERNEL_TEST_DEPS,
6233)
6234
6235xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07006236 name = "f32_vabs_test",
6237 srcs = [
6238 "test/f32-vabs.cc",
6239 "test/vunary-microkernel-tester.h",
6240 ] + MICROKERNEL_TEST_HDRS,
6241 deps = MICROKERNEL_TEST_DEPS,
6242)
6243
6244xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07006245 name = "f32_vadd_test",
6246 srcs = [
6247 "test/f32-vadd.cc",
6248 "test/vbinary-microkernel-tester.h",
6249 ] + MICROKERNEL_TEST_HDRS,
6250 deps = MICROKERNEL_TEST_DEPS,
6251)
6252
6253xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006254 name = "f32_vadd_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006255 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006256 "test/f32-vadd-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08006257 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08006258 ] + MICROKERNEL_TEST_HDRS,
6259 deps = MICROKERNEL_TEST_DEPS,
6260)
6261
6262xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07006263 name = "f32_vadd_relu_test",
6264 srcs = [
6265 "test/f32-vadd-relu.cc",
6266 "test/vbinary-microkernel-tester.h",
6267 ] + MICROKERNEL_TEST_HDRS,
6268 deps = MICROKERNEL_TEST_DEPS,
6269)
6270
6271xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07006272 name = "f32_vaddc_test",
6273 srcs = [
6274 "test/f32-vaddc.cc",
6275 "test/vbinaryc-microkernel-tester.h",
6276 ] + MICROKERNEL_TEST_HDRS,
6277 deps = MICROKERNEL_TEST_DEPS,
6278)
6279
6280xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006281 name = "f32_vaddc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08006282 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006283 "test/f32-vaddc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08006284 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006285 ] + MICROKERNEL_TEST_HDRS,
6286 deps = MICROKERNEL_TEST_DEPS,
6287)
6288
6289xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07006290 name = "f32_vaddc_relu_test",
6291 srcs = [
6292 "test/f32-vaddc-relu.cc",
6293 "test/vbinaryc-microkernel-tester.h",
6294 ] + MICROKERNEL_TEST_HDRS,
6295 deps = MICROKERNEL_TEST_DEPS,
6296)
6297
6298xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07006299 name = "f32_vdiv_test",
6300 srcs = [
6301 "test/f32-vdiv.cc",
6302 "test/vbinary-microkernel-tester.h",
6303 ] + MICROKERNEL_TEST_HDRS,
6304 deps = MICROKERNEL_TEST_DEPS,
6305)
6306
6307xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006308 name = "f32_vdiv_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08006309 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006310 "test/f32-vdiv-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08006311 "test/vbinary-microkernel-tester.h",
6312 ] + MICROKERNEL_TEST_HDRS,
6313 deps = MICROKERNEL_TEST_DEPS,
6314)
6315
6316xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07006317 name = "f32_vdiv_relu_test",
6318 srcs = [
6319 "test/f32-vdiv-relu.cc",
6320 "test/vbinary-microkernel-tester.h",
6321 ] + MICROKERNEL_TEST_HDRS,
6322 deps = MICROKERNEL_TEST_DEPS,
6323)
6324
6325xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07006326 name = "f32_vdivc_test",
6327 srcs = [
6328 "test/f32-vdivc.cc",
6329 "test/vbinaryc-microkernel-tester.h",
6330 ] + MICROKERNEL_TEST_HDRS,
6331 deps = MICROKERNEL_TEST_DEPS,
6332)
6333
6334xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006335 name = "f32_vdivc_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08006336 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006337 "test/f32-vdivc-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08006338 "test/vbinaryc-microkernel-tester.h",
6339 ] + MICROKERNEL_TEST_HDRS,
6340 deps = MICROKERNEL_TEST_DEPS,
6341)
6342
6343xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07006344 name = "f32_vdivc_relu_test",
6345 srcs = [
6346 "test/f32-vdivc-relu.cc",
6347 "test/vbinaryc-microkernel-tester.h",
6348 ] + MICROKERNEL_TEST_HDRS,
6349 deps = MICROKERNEL_TEST_DEPS,
6350)
6351
6352xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07006353 name = "f32_vrdivc_test",
6354 srcs = [
6355 "test/f32-vrdivc.cc",
6356 "test/vbinaryc-microkernel-tester.h",
6357 ] + MICROKERNEL_TEST_HDRS,
6358 deps = MICROKERNEL_TEST_DEPS,
6359)
6360
6361xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006362 name = "f32_vrdivc_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08006363 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006364 "test/f32-vrdivc-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08006365 "test/vbinaryc-microkernel-tester.h",
6366 ] + MICROKERNEL_TEST_HDRS,
6367 deps = MICROKERNEL_TEST_DEPS,
6368)
6369
6370xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07006371 name = "f32_vrdivc_relu_test",
6372 srcs = [
6373 "test/f32-vrdivc-relu.cc",
6374 "test/vbinaryc-microkernel-tester.h",
6375 ] + MICROKERNEL_TEST_HDRS,
6376 deps = MICROKERNEL_TEST_DEPS,
6377)
6378
6379xnnpack_unit_test(
Marat Dukhaned6baaf2020-12-01 15:07:08 -08006380 name = "f32_velu_test",
6381 srcs = [
6382 "test/f32-velu.cc",
6383 "test/vunary-microkernel-tester.h",
6384 ] + MICROKERNEL_TEST_HDRS,
6385 deps = MICROKERNEL_TEST_DEPS,
6386)
6387
6388xnnpack_unit_test(
Marat Dukhan403b7d42019-12-05 12:49:11 -08006389 name = "f32_vmax_test",
6390 srcs = [
6391 "test/f32-vmax.cc",
6392 "test/vbinary-microkernel-tester.h",
6393 ] + MICROKERNEL_TEST_HDRS,
6394 deps = MICROKERNEL_TEST_DEPS,
6395)
6396
6397xnnpack_unit_test(
6398 name = "f32_vmaxc_test",
6399 srcs = [
6400 "test/f32-vmaxc.cc",
6401 "test/vbinaryc-microkernel-tester.h",
6402 ] + MICROKERNEL_TEST_HDRS,
6403 deps = MICROKERNEL_TEST_DEPS,
6404)
6405
6406xnnpack_unit_test(
6407 name = "f32_vmin_test",
6408 srcs = [
6409 "test/f32-vmin.cc",
6410 "test/vbinary-microkernel-tester.h",
6411 ] + MICROKERNEL_TEST_HDRS,
6412 deps = MICROKERNEL_TEST_DEPS,
6413)
6414
6415xnnpack_unit_test(
6416 name = "f32_vminc_test",
6417 srcs = [
6418 "test/f32-vminc.cc",
6419 "test/vbinaryc-microkernel-tester.h",
6420 ] + MICROKERNEL_TEST_HDRS,
6421 deps = MICROKERNEL_TEST_DEPS,
6422)
6423
6424xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07006425 name = "f32_vmul_test",
6426 srcs = [
6427 "test/f32-vmul.cc",
6428 "test/vbinary-microkernel-tester.h",
6429 ] + MICROKERNEL_TEST_HDRS,
6430 deps = MICROKERNEL_TEST_DEPS,
6431)
6432
6433xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006434 name = "f32_vmul_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006435 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006436 "test/f32-vmul-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08006437 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08006438 ] + MICROKERNEL_TEST_HDRS,
6439 deps = MICROKERNEL_TEST_DEPS,
6440)
6441
6442xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07006443 name = "f32_vmul_relu_test",
6444 srcs = [
6445 "test/f32-vmul-relu.cc",
6446 "test/vbinary-microkernel-tester.h",
6447 ] + MICROKERNEL_TEST_HDRS,
6448 deps = MICROKERNEL_TEST_DEPS,
6449)
6450
6451xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07006452 name = "f32_vmulc_test",
6453 srcs = [
6454 "test/f32-vmulc.cc",
6455 "test/vbinaryc-microkernel-tester.h",
6456 ] + MICROKERNEL_TEST_HDRS,
6457 deps = MICROKERNEL_TEST_DEPS,
6458)
6459
6460xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006461 name = "f32_vmulc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08006462 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006463 "test/f32-vmulc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08006464 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006465 ] + MICROKERNEL_TEST_HDRS,
6466 deps = MICROKERNEL_TEST_DEPS,
6467)
6468
6469xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07006470 name = "f32_vmulc_relu_test",
6471 srcs = [
6472 "test/f32-vmulc-relu.cc",
6473 "test/vbinaryc-microkernel-tester.h",
6474 ] + MICROKERNEL_TEST_HDRS,
6475 deps = MICROKERNEL_TEST_DEPS,
6476)
6477
6478xnnpack_unit_test(
Frank Barchard2a1049c2020-06-03 02:31:27 -07006479 name = "f16_vmulcaddc_minmax_test",
6480 srcs = [
6481 "test/f16-vmulcaddc-minmax.cc",
6482 "test/vmulcaddc-microkernel-tester.h",
6483 "src/xnnpack/AlignedAllocator.h",
6484 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07006485 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Frank Barchard2a1049c2020-06-03 02:31:27 -07006486)
6487
6488xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07006489 name = "f32_vmulcaddc_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006490 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07006491 "test/f32-vmulcaddc-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006492 "test/vmulcaddc-microkernel-tester.h",
6493 "src/xnnpack/AlignedAllocator.h",
6494 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07006495 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07006496)
6497
6498xnnpack_unit_test(
Marat Dukhan8cc7efe2020-06-10 16:24:27 -07006499 name = "f32_vlrelu_test",
6500 srcs = [
6501 "test/f32-vlrelu.cc",
6502 "test/vunary-microkernel-tester.h",
6503 ] + MICROKERNEL_TEST_HDRS,
6504 deps = MICROKERNEL_TEST_DEPS,
6505)
6506
6507xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07006508 name = "f32_vneg_test",
6509 srcs = [
6510 "test/f32-vneg.cc",
6511 "test/vunary-microkernel-tester.h",
6512 ] + MICROKERNEL_TEST_HDRS,
6513 deps = MICROKERNEL_TEST_DEPS,
6514)
6515
6516xnnpack_unit_test(
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07006517 name = "f32_vrndne_test",
6518 srcs = [
6519 "test/f32-vrndne.cc",
6520 "test/vunary-microkernel-tester.h",
6521 ] + MICROKERNEL_TEST_HDRS,
6522 deps = MICROKERNEL_TEST_DEPS,
6523)
6524
6525xnnpack_unit_test(
6526 name = "f32_vrndz_test",
6527 srcs = [
6528 "test/f32-vrndz.cc",
6529 "test/vunary-microkernel-tester.h",
6530 ] + MICROKERNEL_TEST_HDRS,
6531 deps = MICROKERNEL_TEST_DEPS,
6532)
6533
6534xnnpack_unit_test(
6535 name = "f32_vrndu_test",
6536 srcs = [
6537 "test/f32-vrndu.cc",
6538 "test/vunary-microkernel-tester.h",
6539 ] + MICROKERNEL_TEST_HDRS,
6540 deps = MICROKERNEL_TEST_DEPS,
6541)
6542
6543xnnpack_unit_test(
6544 name = "f32_vrndd_test",
6545 srcs = [
6546 "test/f32-vrndd.cc",
6547 "test/vunary-microkernel-tester.h",
6548 ] + MICROKERNEL_TEST_HDRS,
6549 deps = MICROKERNEL_TEST_DEPS,
6550)
6551
6552xnnpack_unit_test(
Marat Dukhan05ac8e32019-10-21 15:39:33 -07006553 name = "f32_vscale_test",
6554 srcs = [
6555 "test/f32-vscale.cc",
6556 "test/vscale-microkernel-tester.h",
6557 ] + MICROKERNEL_TEST_HDRS,
6558 deps = MICROKERNEL_TEST_DEPS,
6559)
6560
6561xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07006562 name = "f32_vscaleexpminusmax_test",
6563 srcs = [
6564 "test/f32-vscaleexpminusmax.cc",
6565 "test/vscaleexpminusmax-microkernel-tester.h",
6566 ] + MICROKERNEL_TEST_HDRS,
6567 deps = MICROKERNEL_TEST_DEPS,
6568)
6569
6570xnnpack_unit_test(
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07006571 name = "f32_vscaleextexp_test",
6572 srcs = [
6573 "test/f32-vscaleextexp.cc",
6574 "test/vscaleextexp-microkernel-tester.h",
6575 ] + MICROKERNEL_TEST_HDRS,
6576 deps = MICROKERNEL_TEST_DEPS,
6577)
6578
6579xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07006580 name = "f32_vsqr_test",
6581 srcs = [
6582 "test/f32-vsqr.cc",
6583 "test/vunary-microkernel-tester.h",
6584 ] + MICROKERNEL_TEST_HDRS,
6585 deps = MICROKERNEL_TEST_DEPS,
6586)
6587
6588xnnpack_unit_test(
Marat Dukhan13bafb02020-06-05 00:43:11 -07006589 name = "f32_vsqrdiff_test",
6590 srcs = [
6591 "test/f32-vsqrdiff.cc",
6592 "test/vbinary-microkernel-tester.h",
6593 ] + MICROKERNEL_TEST_HDRS,
6594 deps = MICROKERNEL_TEST_DEPS,
6595)
6596
6597xnnpack_unit_test(
6598 name = "f32_vsqrdiffc_test",
6599 srcs = [
6600 "test/f32-vsqrdiffc.cc",
6601 "test/vbinaryc-microkernel-tester.h",
6602 ] + MICROKERNEL_TEST_HDRS,
6603 deps = MICROKERNEL_TEST_DEPS,
6604)
6605
6606xnnpack_unit_test(
Marat Dukhanf4db2f32020-06-30 10:55:30 -07006607 name = "f32_vsqrt_test",
6608 srcs = [
6609 "test/f32-vsqrt.cc",
6610 "test/vunary-microkernel-tester.h",
6611 ] + MICROKERNEL_TEST_HDRS,
6612 deps = MICROKERNEL_TEST_DEPS,
6613)
6614
6615xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07006616 name = "f32_vsub_test",
6617 srcs = [
6618 "test/f32-vsub.cc",
6619 "test/vbinary-microkernel-tester.h",
6620 ] + MICROKERNEL_TEST_HDRS,
6621 deps = MICROKERNEL_TEST_DEPS,
6622)
6623
6624xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006625 name = "f32_vsub_minmax_test",
Marat Dukhan97579532019-10-18 16:40:39 -07006626 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006627 "test/f32-vsub-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08006628 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08006629 ] + MICROKERNEL_TEST_HDRS,
6630 deps = MICROKERNEL_TEST_DEPS,
6631)
6632
6633xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07006634 name = "f32_vsub_relu_test",
6635 srcs = [
6636 "test/f32-vsub-relu.cc",
6637 "test/vbinary-microkernel-tester.h",
6638 ] + MICROKERNEL_TEST_HDRS,
6639 deps = MICROKERNEL_TEST_DEPS,
6640)
6641
6642xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07006643 name = "f32_vsubc_test",
6644 srcs = [
6645 "test/f32-vsubc.cc",
6646 "test/vbinaryc-microkernel-tester.h",
6647 ] + MICROKERNEL_TEST_HDRS,
6648 deps = MICROKERNEL_TEST_DEPS,
6649)
6650
6651xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006652 name = "f32_vsubc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08006653 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006654 "test/f32-vsubc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08006655 "test/vbinaryc-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08006656 ] + MICROKERNEL_TEST_HDRS,
6657 deps = MICROKERNEL_TEST_DEPS,
6658)
6659
6660xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07006661 name = "f32_vsubc_relu_test",
6662 srcs = [
6663 "test/f32-vsubc-relu.cc",
6664 "test/vbinaryc-microkernel-tester.h",
6665 ] + MICROKERNEL_TEST_HDRS,
6666 deps = MICROKERNEL_TEST_DEPS,
6667)
6668
6669xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07006670 name = "f32_vrsubc_test",
6671 srcs = [
6672 "test/f32-vrsubc.cc",
6673 "test/vbinaryc-microkernel-tester.h",
6674 ] + MICROKERNEL_TEST_HDRS,
6675 deps = MICROKERNEL_TEST_DEPS,
6676)
6677
6678xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006679 name = "f32_vrsubc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08006680 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006681 "test/f32-vrsubc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08006682 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan97579532019-10-18 16:40:39 -07006683 ] + MICROKERNEL_TEST_HDRS,
6684 deps = MICROKERNEL_TEST_DEPS,
6685)
6686
6687xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07006688 name = "f32_vrsubc_relu_test",
6689 srcs = [
6690 "test/f32-vrsubc-relu.cc",
6691 "test/vbinaryc-microkernel-tester.h",
6692 ] + MICROKERNEL_TEST_HDRS,
6693 deps = MICROKERNEL_TEST_DEPS,
6694)
6695
6696xnnpack_unit_test(
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07006697 name = "qs8_dwconv_minmax_test",
6698 srcs = [
6699 "test/qs8-dwconv-minmax.cc",
6700 "test/dwconv-microkernel-tester.h",
6701 "src/xnnpack/AlignedAllocator.h",
6702 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
6703 deps = MICROKERNEL_TEST_DEPS + [":packing"],
6704)
6705
6706xnnpack_unit_test(
Marat Dukhan4ed53f42020-08-06 01:12:55 -07006707 name = "qs8_gavgpool_minmax_test",
6708 srcs = [
6709 "test/qs8-gavgpool-minmax.cc",
6710 "test/gavgpool-microkernel-tester.h",
6711 "src/xnnpack/AlignedAllocator.h",
6712 ] + MICROKERNEL_TEST_HDRS,
6713 deps = MICROKERNEL_TEST_DEPS,
6714)
6715
6716xnnpack_unit_test(
Marat Dukhan595e1702020-07-31 10:12:52 -07006717 name = "qs8_gemm_minmax_test",
6718 srcs = [
6719 "test/qs8-gemm-minmax.cc",
6720 "test/gemm-microkernel-tester.h",
6721 "src/xnnpack/AlignedAllocator.h",
6722 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
6723 deps = MICROKERNEL_TEST_DEPS + [":packing"],
6724)
6725
6726xnnpack_unit_test(
Marat Dukhanf9480682020-07-31 14:50:24 -07006727 name = "qs8_igemm_minmax_test",
6728 srcs = [
6729 "test/qs8-igemm-minmax.cc",
6730 "test/gemm-microkernel-tester.h",
6731 "src/xnnpack/AlignedAllocator.h",
6732 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
6733 deps = MICROKERNEL_TEST_DEPS + [":packing"],
6734)
6735
6736xnnpack_unit_test(
6737 name = "qs8_requantization_test",
6738 srcs = [
6739 "src/xnnpack/requantization-stubs.h",
6740 "test/qs8-requantization.cc",
6741 "test/requantization-tester.h",
6742 ] + MICROKERNEL_TEST_HDRS,
6743 deps = MICROKERNEL_TEST_DEPS,
6744)
6745
6746xnnpack_unit_test(
Marat Dukhand9f3ad42020-08-10 12:30:58 -07006747 name = "qs8_vadd_minmax_test",
6748 srcs = [
6749 "test/qs8-vadd-minmax.cc",
6750 "test/vadd-microkernel-tester.h",
6751 ] + MICROKERNEL_TEST_HDRS,
6752 deps = MICROKERNEL_TEST_DEPS,
6753)
6754
6755xnnpack_unit_test(
Marat Dukhan0270d9f2020-08-11 00:56:46 -07006756 name = "qs8_vaddc_minmax_test",
6757 srcs = [
6758 "test/qs8-vaddc-minmax.cc",
6759 "test/vaddc-microkernel-tester.h",
6760 ] + MICROKERNEL_TEST_HDRS,
6761 deps = MICROKERNEL_TEST_DEPS,
6762)
6763
6764xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07006765 name = "qu8_avgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006766 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07006767 "test/qu8-avgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006768 "test/avgpool-microkernel-tester.h",
6769 "src/xnnpack/AlignedAllocator.h",
6770 ] + MICROKERNEL_TEST_HDRS,
6771 deps = MICROKERNEL_TEST_DEPS,
6772)
6773
6774xnnpack_unit_test(
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07006775 name = "qu8_dwconv_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006776 srcs = [
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07006777 "test/qu8-dwconv-minmax.cc",
6778 "test/dwconv-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006779 "src/xnnpack/AlignedAllocator.h",
6780 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07006781 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07006782)
6783
6784xnnpack_unit_test(
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07006785 name = "qu8_igemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006786 srcs = [
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07006787 "test/qu8-igemm-minmax.cc",
6788 "test/gemm-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006789 "src/xnnpack/AlignedAllocator.h",
6790 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07006791 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07006792)
6793
6794xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07006795 name = "qu8_gavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006796 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07006797 "test/qu8-gavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006798 "test/gavgpool-microkernel-tester.h",
6799 "src/xnnpack/AlignedAllocator.h",
6800 ] + MICROKERNEL_TEST_HDRS,
6801 deps = MICROKERNEL_TEST_DEPS,
6802)
6803
6804xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07006805 name = "qu8_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006806 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07006807 "test/qu8-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006808 "test/gemm-microkernel-tester.h",
6809 "src/xnnpack/AlignedAllocator.h",
6810 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07006811 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07006812)
6813
6814xnnpack_unit_test(
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07006815 name = "qu8_requantization_test",
6816 srcs = [
6817 "src/xnnpack/requantization-stubs.h",
6818 "test/qu8-requantization.cc",
6819 "test/requantization-tester.h",
6820 ] + MICROKERNEL_TEST_HDRS,
6821 deps = MICROKERNEL_TEST_DEPS,
6822)
6823
6824xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07006825 name = "qu8_vadd_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006826 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07006827 "test/qu8-vadd-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006828 "test/vadd-microkernel-tester.h",
6829 ] + MICROKERNEL_TEST_HDRS,
6830 deps = MICROKERNEL_TEST_DEPS,
6831)
6832
6833xnnpack_unit_test(
6834 name = "u8_clamp_test",
6835 srcs = [
6836 "test/u8-clamp.cc",
6837 "test/clamp-microkernel-tester.h",
6838 ] + MICROKERNEL_TEST_HDRS,
6839 deps = MICROKERNEL_TEST_DEPS,
6840)
6841
6842xnnpack_unit_test(
6843 name = "u8_lut32norm_test",
6844 srcs = [
6845 "test/u8-lut32norm.cc",
6846 "test/lut-norm-microkernel-tester.h",
6847 ] + MICROKERNEL_TEST_HDRS,
6848 deps = MICROKERNEL_TEST_DEPS,
6849)
6850
6851xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07006852 name = "u8_maxpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006853 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07006854 "test/u8-maxpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006855 "test/maxpool-microkernel-tester.h",
6856 ] + MICROKERNEL_TEST_HDRS,
6857 deps = MICROKERNEL_TEST_DEPS,
6858)
6859
6860xnnpack_unit_test(
6861 name = "u8_rmax_test",
6862 srcs = [
6863 "test/u8-rmax.cc",
6864 "test/rmax-microkernel-tester.h",
6865 ] + MICROKERNEL_TEST_HDRS,
6866 deps = MICROKERNEL_TEST_DEPS,
6867)
6868
6869xnnpack_unit_test(
Marat Dukhanad71b9a2020-11-20 00:01:51 -08006870 name = "x32_depthtospace2d_chw2hwc_test",
Yury Kartynnike7841862020-11-04 18:22:18 -08006871 srcs = [
Marat Dukhanad71b9a2020-11-20 00:01:51 -08006872 "test/x32-depthtospace2d-chw2hwc.cc",
6873 "test/depthtospace-microkernel-tester.h",
Yury Kartynnike7841862020-11-04 18:22:18 -08006874 ] + MICROKERNEL_TEST_HDRS,
6875 deps = MICROKERNEL_TEST_DEPS,
6876)
6877
6878xnnpack_unit_test(
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -07006879 name = "x32_fill_test",
6880 srcs = [
6881 "test/x32-fill.cc",
6882 "test/fill-microkernel-tester.h",
6883 ] + MICROKERNEL_TEST_HDRS,
6884 deps = MICROKERNEL_TEST_DEPS,
6885)
6886
6887xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07006888 name = "x32_packx_test",
6889 srcs = [
6890 "test/x32-packx.cc",
6891 "test/pack-microkernel-tester.h",
6892 "src/xnnpack/AlignedAllocator.h",
6893 ] + MICROKERNEL_TEST_HDRS,
6894 deps = MICROKERNEL_TEST_DEPS,
6895)
6896
6897xnnpack_unit_test(
6898 name = "x32_pad_test",
6899 srcs = [
6900 "test/x32-pad.cc",
6901 "test/pad-microkernel-tester.h",
6902 ] + MICROKERNEL_TEST_HDRS,
6903 deps = MICROKERNEL_TEST_DEPS,
6904)
6905
6906xnnpack_unit_test(
6907 name = "x32_unpool_test",
6908 srcs = [
6909 "test/x32-unpool.cc",
6910 "test/unpool-microkernel-tester.h",
6911 ] + MICROKERNEL_TEST_HDRS,
6912 deps = MICROKERNEL_TEST_DEPS,
6913)
6914
6915xnnpack_unit_test(
6916 name = "x32_zip_test",
6917 srcs = [
6918 "test/x32-zip.cc",
6919 "test/zip-microkernel-tester.h",
6920 ] + MICROKERNEL_TEST_HDRS,
6921 deps = MICROKERNEL_TEST_DEPS,
6922)
6923
6924xnnpack_unit_test(
6925 name = "x8_lut_test",
6926 srcs = [
6927 "test/x8-lut.cc",
6928 "test/lut-microkernel-tester.h",
6929 ] + MICROKERNEL_TEST_HDRS,
6930 deps = MICROKERNEL_TEST_DEPS,
6931)
6932
6933xnnpack_unit_test(
6934 name = "x8_zip_test",
6935 srcs = [
6936 "test/x8-zip.cc",
6937 "test/zip-microkernel-tester.h",
6938 ] + MICROKERNEL_TEST_HDRS,
6939 deps = MICROKERNEL_TEST_DEPS,
6940)
6941
Marat Dukhan20c3b922020-03-10 03:45:06 -07006942########################## Size tests for the library #########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07006943
6944xnnpack_binary(
Marat Dukhan20c3b922020-03-10 03:45:06 -07006945 name = "operator_size_test",
6946 srcs = ["test/operator-size.c"],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07006947 deps = [":xnnpack_operators_nhwc_f32"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07006948)
6949
Marat Dukhan20c3b922020-03-10 03:45:06 -07006950xnnpack_binary(
6951 name = "subgraph_size_test",
6952 srcs = ["test/subgraph-size.c"],
6953 deps = [":XNNPACK"],
6954)
6955
6956########################### Unit tests for operators ##########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07006957
6958xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07006959 name = "abs_nc_test",
6960 srcs = [
6961 "test/abs-nc.cc",
6962 "test/abs-operator-tester.h",
6963 ],
6964 deps = OPERATOR_TEST_DEPS,
6965)
6966
6967xnnpack_unit_test(
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08006968 name = "add_nd_test",
Artsiom Ablavatski2202c812021-01-22 14:16:43 -08006969 timeout = "moderate",
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08006970 srcs = [
6971 "test/add-nd.cc",
6972 "test/binary-elementwise-operator-tester.h",
6973 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006974 deps = OPERATOR_TEST_DEPS,
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08006975)
6976
6977xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006978 name = "argmax_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006979 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006980 "test/argmax-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006981 "test/argmax-pooling-operator-tester.h",
6982 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07006983 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006984)
6985
6986xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006987 name = "average_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006988 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006989 "test/average-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006990 "test/average-pooling-operator-tester.h",
6991 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07006992 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006993)
6994
6995xnnpack_unit_test(
Marat Dukhan64e52512020-06-09 13:41:16 -07006996 name = "bankers_rounding_nc_test",
6997 srcs = [
6998 "test/bankers-rounding-nc.cc",
6999 "test/bankers-rounding-operator-tester.h",
7000 ],
7001 deps = OPERATOR_TEST_DEPS,
7002)
7003
7004xnnpack_unit_test(
7005 name = "ceiling_nc_test",
7006 srcs = [
7007 "test/ceiling-nc.cc",
7008 "test/ceiling-operator-tester.h",
7009 ],
7010 deps = OPERATOR_TEST_DEPS,
7011)
7012
7013xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08007014 name = "channel_shuffle_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007015 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08007016 "test/channel-shuffle-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007017 "test/channel-shuffle-operator-tester.h",
7018 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07007019 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07007020)
7021
7022xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08007023 name = "clamp_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007024 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08007025 "test/clamp-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007026 "test/clamp-operator-tester.h",
7027 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07007028 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07007029)
7030
7031xnnpack_unit_test(
Marat Dukhan065b11e2020-05-22 09:49:41 -07007032 name = "constant_pad_nd_test",
7033 srcs = [
7034 "test/constant-pad-nd.cc",
7035 "test/constant-pad-operator-tester.h",
7036 ],
7037 deps = OPERATOR_TEST_DEPS,
7038)
7039
7040xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08007041 name = "convolution_nhwc_test",
Artsiom Ablavatski2202c812021-01-22 14:16:43 -08007042 timeout = "moderate",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007043 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08007044 "test/convolution-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007045 "test/convolution-operator-tester.h",
7046 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07007047 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07007048)
7049
7050xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08007051 name = "convolution_nchw_test",
Artsiom Ablavatski2202c812021-01-22 14:16:43 -08007052 timeout = "moderate",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007053 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08007054 "test/convolution-nchw.cc",
7055 "test/convolution-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007056 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07007057 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07007058)
7059
7060xnnpack_unit_test(
Marat Dukhan4e21b272020-06-04 18:45:01 -07007061 name = "copy_nc_test",
7062 srcs = [
7063 "test/copy-nc.cc",
7064 "test/copy-operator-tester.h",
7065 ],
7066 deps = OPERATOR_TEST_DEPS,
7067)
7068
7069xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08007070 name = "deconvolution_nhwc_test",
Artsiom Ablavatskic1aa2972020-12-08 11:23:34 -08007071 timeout = "moderate",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007072 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08007073 "test/deconvolution-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007074 "test/deconvolution-operator-tester.h",
7075 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07007076 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07007077)
7078
7079xnnpack_unit_test(
Marat Dukhan188d1042020-11-24 23:39:40 -08007080 name = "depth_to_space_nchw2nhwc_test",
Artsiom Ablavatski0f1dc182020-11-05 19:21:50 -08007081 srcs = [
Marat Dukhan188d1042020-11-24 23:39:40 -08007082 "test/depth-to-space-nchw2nhwc.cc",
Artsiom Ablavatski0f1dc182020-11-05 19:21:50 -08007083 "test/depth-to-space-operator-tester.h",
7084 ] + OPERATOR_TEST_PARAMS_HDRS,
7085 deps = OPERATOR_TEST_DEPS,
7086)
7087
7088xnnpack_unit_test(
Marat Dukhan0e521172020-11-25 13:10:04 -08007089 name = "depth_to_space_nhwc_test",
7090 srcs = [
7091 "test/depth-to-space-nhwc.cc",
7092 "test/depth-to-space-operator-tester.h",
7093 ] + OPERATOR_TEST_PARAMS_HDRS,
7094 deps = OPERATOR_TEST_DEPS,
7095)
7096
7097xnnpack_unit_test(
Marat Dukhan69180502019-12-06 15:00:31 -08007098 name = "divide_nd_test",
7099 srcs = [
7100 "test/binary-elementwise-operator-tester.h",
7101 "test/divide-nd.cc",
7102 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07007103 deps = OPERATOR_TEST_DEPS,
Marat Dukhan69180502019-12-06 15:00:31 -08007104)
7105
7106xnnpack_unit_test(
Marat Dukhanb6bd4bc2020-12-01 17:01:40 -08007107 name = "elu_nc_test",
7108 srcs = [
7109 "test/elu-nc.cc",
7110 "test/elu-operator-tester.h",
7111 ],
7112 deps = OPERATOR_TEST_DEPS,
7113)
7114
7115xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08007116 name = "fully_connected_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007117 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08007118 "test/fully-connected-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007119 "test/fully-connected-operator-tester.h",
7120 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07007121 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07007122)
7123
7124xnnpack_unit_test(
Marat Dukhan64e52512020-06-09 13:41:16 -07007125 name = "floor_nc_test",
7126 srcs = [
7127 "test/floor-nc.cc",
7128 "test/floor-operator-tester.h",
7129 ],
7130 deps = OPERATOR_TEST_DEPS,
7131)
7132
7133xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08007134 name = "global_average_pooling_nwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007135 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08007136 "test/global-average-pooling-nwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007137 "test/global-average-pooling-operator-tester.h",
Marat Dukhanef61d022020-06-19 13:54:49 -07007138 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07007139 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07007140)
7141
7142xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08007143 name = "global_average_pooling_ncw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007144 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08007145 "test/global-average-pooling-ncw.cc",
7146 "test/global-average-pooling-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007147 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07007148 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07007149)
7150
7151xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08007152 name = "hardswish_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007153 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08007154 "test/hardswish-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007155 "test/hardswish-operator-tester.h",
7156 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07007157 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07007158)
7159
7160xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08007161 name = "leaky_relu_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007162 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08007163 "test/leaky-relu-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007164 "test/leaky-relu-operator-tester.h",
7165 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07007166 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07007167)
7168
7169xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08007170 name = "max_pooling_nhwc_test",
Artsiom Ablavatski2202c812021-01-22 14:16:43 -08007171 timeout = "moderate",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007172 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08007173 "test/max-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007174 "test/max-pooling-operator-tester.h",
7175 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07007176 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07007177)
7178
7179xnnpack_unit_test(
Marat Dukhan79e7f842019-12-05 14:35:50 -08007180 name = "maximum_nd_test",
7181 srcs = [
7182 "test/binary-elementwise-operator-tester.h",
7183 "test/maximum-nd.cc",
7184 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07007185 deps = OPERATOR_TEST_DEPS,
Marat Dukhan79e7f842019-12-05 14:35:50 -08007186)
7187
7188xnnpack_unit_test(
7189 name = "minimum_nd_test",
7190 srcs = [
7191 "test/binary-elementwise-operator-tester.h",
7192 "test/minimum-nd.cc",
7193 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07007194 deps = OPERATOR_TEST_DEPS,
Marat Dukhan79e7f842019-12-05 14:35:50 -08007195)
7196
7197xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08007198 name = "multiply_nd_test",
Marat Dukhanca2733c2019-11-15 23:21:17 -08007199 srcs = [
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08007200 "test/binary-elementwise-operator-tester.h",
Marat Dukhanefc47b82019-11-18 09:25:38 -08007201 "test/multiply-nd.cc",
Marat Dukhanca2733c2019-11-15 23:21:17 -08007202 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07007203 deps = OPERATOR_TEST_DEPS,
Marat Dukhanca2733c2019-11-15 23:21:17 -08007204)
7205
7206xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07007207 name = "negate_nc_test",
7208 srcs = [
7209 "test/negate-nc.cc",
7210 "test/negate-operator-tester.h",
7211 ],
7212 deps = OPERATOR_TEST_DEPS,
7213)
7214
7215xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08007216 name = "prelu_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007217 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08007218 "test/prelu-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007219 "test/prelu-operator-tester.h",
7220 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07007221 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07007222)
7223
7224xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08007225 name = "resize_bilinear_nhwc_test",
Marat Dukhan69722492019-11-11 19:55:50 -08007226 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08007227 "test/resize-bilinear-nhwc.cc",
Marat Dukhan69722492019-11-11 19:55:50 -08007228 "test/resize-bilinear-operator-tester.h",
7229 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07007230 deps = OPERATOR_TEST_DEPS,
Marat Dukhan69722492019-11-11 19:55:50 -08007231)
7232
7233xnnpack_unit_test(
Artsiom Ablavatski97918102020-10-27 15:52:59 -07007234 name = "resize_bilinear_nchw_test",
7235 srcs = [
7236 "test/resize-bilinear-nchw.cc",
7237 "test/resize-bilinear-operator-tester.h",
7238 ] + OPERATOR_TEST_PARAMS_HDRS,
7239 deps = OPERATOR_TEST_DEPS,
7240)
7241
7242xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08007243 name = "sigmoid_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007244 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08007245 "test/sigmoid-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007246 "test/sigmoid-operator-tester.h",
7247 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07007248 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07007249)
7250
7251xnnpack_unit_test(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08007252 name = "softmax_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007253 srcs = [
Marat Dukhanfd8e6892020-01-27 15:25:25 -08007254 "test/softmax-nc.cc",
7255 "test/softmax-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007256 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07007257 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07007258)
7259
7260xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07007261 name = "square_nc_test",
7262 srcs = [
7263 "test/square-nc.cc",
7264 "test/square-operator-tester.h",
7265 ],
7266 deps = OPERATOR_TEST_DEPS,
7267)
7268
7269xnnpack_unit_test(
Marat Dukhan6804bbd2020-06-30 19:26:11 -07007270 name = "square_root_nc_test",
7271 srcs = [
7272 "test/square-root-nc.cc",
7273 "test/square-root-operator-tester.h",
7274 ],
7275 deps = OPERATOR_TEST_DEPS,
7276)
7277
7278xnnpack_unit_test(
Marat Dukhanf7399262020-06-05 10:58:44 -07007279 name = "squared_difference_nd_test",
7280 srcs = [
7281 "test/binary-elementwise-operator-tester.h",
7282 "test/squared-difference-nd.cc",
7283 ],
7284 deps = OPERATOR_TEST_DEPS,
7285)
7286
7287xnnpack_unit_test(
Marat Dukhan05f3f6d2019-12-03 15:13:53 -08007288 name = "subtract_nd_test",
7289 srcs = [
7290 "test/binary-elementwise-operator-tester.h",
7291 "test/subtract-nd.cc",
7292 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07007293 deps = OPERATOR_TEST_DEPS,
Marat Dukhan05f3f6d2019-12-03 15:13:53 -08007294)
7295
7296xnnpack_unit_test(
Marat Dukhan64e52512020-06-09 13:41:16 -07007297 name = "truncation_nc_test",
7298 srcs = [
7299 "test/truncation-nc.cc",
7300 "test/truncation-operator-tester.h",
7301 ],
7302 deps = OPERATOR_TEST_DEPS,
7303)
7304
7305xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08007306 name = "unpooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007307 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08007308 "test/unpooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007309 "test/unpooling-operator-tester.h",
7310 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07007311 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07007312)
7313
Chao Mei6ddfc602020-05-13 22:29:36 -07007314############################### Misc unit tests ###############################
7315
7316xnnpack_unit_test(
7317 name = "memory_planner_test",
7318 srcs = [
7319 "test/memory-planner-test.cc",
7320 ],
7321 deps = [
7322 ":XNNPACK",
7323 ":memory_planner",
7324 ],
7325)
7326
XNNPACK Teamab8c4c82020-10-09 08:05:51 -07007327xnnpack_unit_test(
7328 name = "subgraph_nchw_test",
7329 srcs = [
7330 "src/xnnpack/subgraph.h",
7331 "test/subgraph-nchw.cc",
7332 "test/subgraph-tester.h",
7333 ],
7334 deps = [
7335 ":XNNPACK",
7336 ],
7337)
7338
Marat Dukhan08c4a432019-10-03 09:29:21 -07007339############################# Build configurations #############################
7340
Marat Dukhanb8642352019-10-30 15:43:02 -07007341# Enables usage of assembly kernels.
Marat Dukhan08c4a432019-10-03 09:29:21 -07007342config_setting(
Marat Dukhanb8642352019-10-30 15:43:02 -07007343 name = "xnn_enable_assembly_explicit_true",
7344 define_values = {"xnn_enable_assembly": "true"},
7345)
7346
7347# Disables usage of assembly kernels.
7348config_setting(
7349 name = "xnn_enable_assembly_explicit_false",
7350 define_values = {"xnn_enable_assembly": "false"},
7351)
7352
Marat Dukhan9de90e02020-06-18 16:04:12 -07007353# Enables usage of sparse inference.
7354config_setting(
7355 name = "xnn_enable_sparse_explicit_true",
7356 define_values = {"xnn_enable_sparse": "true"},
7357)
7358
7359# Disables usage of sparse inference.
7360config_setting(
7361 name = "xnn_enable_sparse_explicit_false",
7362 define_values = {"xnn_enable_sparse": "false"},
7363)
7364
Marat Dukhan05702cf2020-03-26 15:41:33 -07007365# Disables usage of HMP-aware optimizations.
7366config_setting(
7367 name = "xnn_enable_hmp_explicit_false",
7368 define_values = {"xnn_enable_hmp": "false"},
7369)
7370
Chao Mei6ddfc602020-05-13 22:29:36 -07007371# Enable usage of optimized memory allocation
7372config_setting(
7373 name = "xnn_enable_memopt_explicit_true",
7374 define_values = {"xnnpack_enable_memopt": "true"},
7375)
7376
7377# Disable usage of optimized memory allocation
7378config_setting(
7379 name = "xnn_enable_memopt_explicit_false",
7380 define_values = {"xnnpack_enable_memopt": "false"},
7381)
7382
Marat Dukhanb8642352019-10-30 15:43:02 -07007383# Builds with -c dbg
7384config_setting(
7385 name = "debug_build",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007386 values = {
Marat Dukhanb8642352019-10-30 15:43:02 -07007387 "compilation_mode": "dbg",
7388 },
7389)
7390
7391# Builds with -c opt
7392config_setting(
7393 name = "optimized_build",
7394 values = {
7395 "compilation_mode": "opt",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007396 },
7397)
7398
7399config_setting(
Marat Dukhanb8642352019-10-30 15:43:02 -07007400 name = "linux_k8",
7401 values = {"cpu": "k8"},
7402)
7403
7404config_setting(
Marat Dukhan582094e2020-04-30 17:21:25 -07007405 name = "linux_arm",
7406 values = {"cpu": "arm"},
Marat Dukhan4e45e662019-10-03 15:40:24 -07007407)
7408
7409config_setting(
Marat Dukhanf0bd4de2020-06-15 15:53:19 -07007410 name = "linux_armeabi",
7411 values = {"cpu": "armeabi"},
7412)
7413
7414config_setting(
Terry Heo68eef3f2020-04-13 22:53:52 -07007415 name = "linux_armhf",
7416 values = {"cpu": "armhf"},
7417)
7418
7419config_setting(
Marat Dukhana720e932020-06-10 13:01:11 -07007420 name = "linux_armv7a",
7421 values = {"cpu": "armv7a"},
7422)
7423
7424config_setting(
Marat Dukhan582094e2020-04-30 17:21:25 -07007425 name = "linux_aarch64",
7426 values = {"cpu": "aarch64"},
7427)
7428
7429config_setting(
Marat Dukhan08c4a432019-10-03 09:29:21 -07007430 name = "android",
7431 values = {"crosstool_top": "//external:android/crosstool"},
7432)
7433
7434config_setting(
7435 name = "android_armv7",
7436 values = {
7437 "crosstool_top": "//external:android/crosstool",
7438 "cpu": "armeabi-v7a",
7439 },
7440)
7441
7442config_setting(
7443 name = "android_arm64",
7444 values = {
7445 "crosstool_top": "//external:android/crosstool",
7446 "cpu": "arm64-v8a",
7447 },
7448)
7449
7450config_setting(
7451 name = "android_x86",
7452 values = {
7453 "crosstool_top": "//external:android/crosstool",
7454 "cpu": "x86",
7455 },
7456)
7457
7458config_setting(
7459 name = "android_x86_64",
7460 values = {
7461 "crosstool_top": "//external:android/crosstool",
7462 "cpu": "x86_64",
7463 },
7464)
7465
7466config_setting(
Marat Dukhan10a38082020-04-17 03:58:35 -07007467 name = "windows_x86_64",
7468 values = {"cpu": "x64_windows"},
Marat Dukhan9fe932e2020-04-11 17:14:15 -07007469)
7470
7471config_setting(
Marat Dukhan10a38082020-04-17 03:58:35 -07007472 name = "windows_x86_64_clang",
7473 values = {
7474 "compiler": "clang-cl",
7475 "cpu": "x64_windows",
7476 },
7477)
7478
7479config_setting(
7480 name = "windows_x86_64_mingw",
7481 values = {
7482 "compiler": "mingw-gcc",
7483 "cpu": "x64_windows",
7484 },
7485)
7486
7487config_setting(
7488 name = "windows_x86_64_msys",
7489 values = {
7490 "compiler": "msys-gcc",
7491 "cpu": "x64_windows",
7492 },
Marat Dukhan9fe932e2020-04-11 17:14:15 -07007493)
7494
7495config_setting(
Marat Dukhan885ca242019-10-07 09:17:32 -07007496 name = "macos_x86_64",
7497 values = {
7498 "apple_platform_type": "macos",
7499 "cpu": "darwin",
7500 },
7501)
7502
7503config_setting(
Simon Maurerae33ab82021-03-03 23:38:22 +01007504 name = "macos_arm64",
7505 values = {
7506 "apple_platform_type": "macos",
7507 "cpu": "darwin_arm64",
7508 },
7509)
7510
7511config_setting(
Marat Dukhan08c4a432019-10-03 09:29:21 -07007512 name = "emscripten",
XNNPACK Teamf69ac5d2021-03-03 16:47:33 -08007513 values = {"crosstool_top": "//emscripten_toolchain:everything"},
Marat Dukhan08c4a432019-10-03 09:29:21 -07007514)
7515
7516config_setting(
7517 name = "emscripten_wasm",
7518 values = {
XNNPACK Teamf69ac5d2021-03-03 16:47:33 -08007519 "crosstool_top": "//emscripten_toolchain:everything",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007520 "cpu": "wasm",
7521 },
7522)
7523
7524config_setting(
7525 name = "emscripten_wasmsimd",
7526 values = {
XNNPACK Teamf69ac5d2021-03-03 16:47:33 -08007527 "crosstool_top": "//emscripten_toolchain:everything",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007528 "cpu": "wasm",
Marat Dukhan81c62602020-05-29 13:22:49 -07007529 "copt": "-msimd128",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007530 },
7531)
7532
7533config_setting(
Marat Dukhan1498d1d2020-02-11 20:00:05 -08007534 name = "ios_armv7",
7535 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08007536 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08007537 "cpu": "ios_armv7",
7538 },
7539)
7540
7541config_setting(
7542 name = "ios_arm64",
7543 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08007544 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08007545 "cpu": "ios_arm64",
7546 },
7547)
7548
7549config_setting(
7550 name = "ios_arm64e",
7551 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08007552 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08007553 "cpu": "ios_arm64e",
7554 },
7555)
7556
7557config_setting(
7558 name = "ios_x86",
7559 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08007560 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08007561 "cpu": "ios_i386",
7562 },
7563)
7564
7565config_setting(
7566 name = "ios_x86_64",
7567 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08007568 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08007569 "cpu": "ios_x86_64",
7570 },
7571)
7572
7573config_setting(
7574 name = "watchos_armv7k",
7575 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08007576 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08007577 "cpu": "watchos_armv7k",
7578 },
7579)
7580
7581config_setting(
7582 name = "watchos_arm64_32",
7583 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08007584 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08007585 "cpu": "watchos_arm64_32",
7586 },
7587)
7588
7589config_setting(
7590 name = "watchos_x86",
7591 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08007592 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08007593 "cpu": "watchos_i386",
7594 },
7595)
7596
7597config_setting(
7598 name = "watchos_x86_64",
7599 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08007600 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08007601 "cpu": "watchos_x86_64",
7602 },
7603)
7604
7605config_setting(
7606 name = "tvos_arm64",
7607 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08007608 "apple_platform_type": "tvos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08007609 "cpu": "tvos_arm64",
7610 },
7611)
7612
7613config_setting(
7614 name = "tvos_x86_64",
7615 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08007616 "apple_platform_type": "tvos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08007617 "cpu": "tvos_x86_64",
7618 },
7619)