blob: db07029cbe583743939fc89b356544401e12c267 [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 Barchard671d1b02021-03-10 15:42:08 -08001703 "src/qs8-gemm/gen/1x8c16-minmax-neon-mlal-padal.c",
Frank Barchardcadd4222021-01-20 16:27:25 -08001704 "src/qs8-gemm/gen/1x16-minmax-neon-mlal-lane.c",
1705 "src/qs8-gemm/gen/1x16-minmax-neon-mull-addw-dup.c",
Frank Barchard8247e212021-02-03 18:12:33 -08001706 "src/qs8-gemm/gen/1x16c2-minmax-neon-mlal-padal-dup.c",
Frank Barchard2302ffd2021-01-22 14:08:02 -08001707 "src/qs8-gemm/gen/1x16c2-minmax-neon-mull-padal-dup.c",
Frank Barchardda78da12021-03-02 14:28:00 -08001708 "src/qs8-gemm/gen/1x16c8-minmax-neon-mlal-padal.c",
Frank Barchard8247e212021-02-03 18:12:33 -08001709 "src/qs8-gemm/gen/1x16c8-minmax-neon-mull-padal.c",
Frank Barchard671d1b02021-03-10 15:42:08 -08001710 "src/qs8-gemm/gen/1x16c16-minmax-neon-mlal-padal.c",
Frank Barchardcadd4222021-01-20 16:27:25 -08001711 "src/qs8-gemm/gen/2x8-minmax-neon-mlal-lane.c",
Frank Barchardec0bf142021-01-15 10:14:05 -08001712 "src/qs8-gemm/gen/2x8-minmax-neon-mull-addw-dup.c",
Frank Barchard8247e212021-02-03 18:12:33 -08001713 "src/qs8-gemm/gen/2x8c2-minmax-neon-mlal-padal-dup.c",
Frank Barchard2302ffd2021-01-22 14:08:02 -08001714 "src/qs8-gemm/gen/2x8c2-minmax-neon-mull-padal-dup.c",
Frank Barchardda78da12021-03-02 14:28:00 -08001715 "src/qs8-gemm/gen/2x8c8-minmax-neon-mlal-padal.c",
Frank Barchard8247e212021-02-03 18:12:33 -08001716 "src/qs8-gemm/gen/2x8c8-minmax-neon-mull-padal.c",
Frank Barchard671d1b02021-03-10 15:42:08 -08001717 "src/qs8-gemm/gen/2x8c16-minmax-neon-mlal-padal.c",
Frank Barchardcadd4222021-01-20 16:27:25 -08001718 "src/qs8-gemm/gen/2x16-minmax-neon-mlal-lane.c",
1719 "src/qs8-gemm/gen/2x16-minmax-neon-mull-addw-dup.c",
Frank Barchard8247e212021-02-03 18:12:33 -08001720 "src/qs8-gemm/gen/2x16c2-minmax-neon-mlal-padal-dup.c",
Frank Barchard2302ffd2021-01-22 14:08:02 -08001721 "src/qs8-gemm/gen/2x16c2-minmax-neon-mull-padal-dup.c",
Frank Barchardda78da12021-03-02 14:28:00 -08001722 "src/qs8-gemm/gen/2x16c8-minmax-neon-mlal-padal.c",
Frank Barchard8247e212021-02-03 18:12:33 -08001723 "src/qs8-gemm/gen/2x16c8-minmax-neon-mull-padal.c",
Frank Barchard671d1b02021-03-10 15:42:08 -08001724 "src/qs8-gemm/gen/2x16c16-minmax-neon-mlal-padal.c",
Frank Barchard2302ffd2021-01-22 14:08:02 -08001725 "src/qs8-gemm/gen/3x8-minmax-neon-mlal-lane.c",
1726 "src/qs8-gemm/gen/3x8-minmax-neon-mull-addw-dup.c",
Frank Barchard8247e212021-02-03 18:12:33 -08001727 "src/qs8-gemm/gen/3x8c2-minmax-neon-mlal-padal-dup.c",
Frank Barchard2302ffd2021-01-22 14:08:02 -08001728 "src/qs8-gemm/gen/3x8c2-minmax-neon-mull-padal-dup.c",
Frank Barchardda78da12021-03-02 14:28:00 -08001729 "src/qs8-gemm/gen/3x8c8-minmax-neon-mlal-padal.c",
Frank Barchard8247e212021-02-03 18:12:33 -08001730 "src/qs8-gemm/gen/3x8c8-minmax-neon-mull-padal.c",
Frank Barchard671d1b02021-03-10 15:42:08 -08001731 "src/qs8-gemm/gen/3x8c16-minmax-neon-mlal-padal.c",
Frank Barchard8247e212021-02-03 18:12:33 -08001732 "src/qs8-gemm/gen/3x16-minmax-neon-mlal-lane.c",
1733 "src/qs8-gemm/gen/3x16-minmax-neon-mull-addw-dup.c",
1734 "src/qs8-gemm/gen/3x16c2-minmax-neon-mlal-padal-dup.c",
1735 "src/qs8-gemm/gen/3x16c2-minmax-neon-mull-padal-dup.c",
Frank Barchardda78da12021-03-02 14:28:00 -08001736 "src/qs8-gemm/gen/3x16c8-minmax-neon-mlal-padal.c",
Frank Barchard8247e212021-02-03 18:12:33 -08001737 "src/qs8-gemm/gen/3x16c8-minmax-neon-mull-padal.c",
Frank Barchard671d1b02021-03-10 15:42:08 -08001738 "src/qs8-gemm/gen/3x16c16-minmax-neon-mlal-padal.c",
Frank Barchardcadd4222021-01-20 16:27:25 -08001739 "src/qs8-gemm/gen/4x8-minmax-neon-mlal-lane.c",
Frank Barchardec0bf142021-01-15 10:14:05 -08001740 "src/qs8-gemm/gen/4x8-minmax-neon-mull-addw-dup.c",
Frank Barchard8247e212021-02-03 18:12:33 -08001741 "src/qs8-gemm/gen/4x8c2-minmax-neon-mlal-padal-dup.c",
Frank Barchard2302ffd2021-01-22 14:08:02 -08001742 "src/qs8-gemm/gen/4x8c2-minmax-neon-mull-padal-dup.c",
Frank Barchardda78da12021-03-02 14:28:00 -08001743 "src/qs8-gemm/gen/4x8c8-minmax-neon-mlal-padal.c",
Frank Barchard8247e212021-02-03 18:12:33 -08001744 "src/qs8-gemm/gen/4x8c8-minmax-neon-mull-padal.c",
Frank Barchard671d1b02021-03-10 15:42:08 -08001745 "src/qs8-gemm/gen/4x8c16-minmax-neon-mlal-padal.c",
Frank Barchardcadd4222021-01-20 16:27:25 -08001746 "src/qs8-gemm/gen/4x16-minmax-neon-mlal-lane.c",
1747 "src/qs8-gemm/gen/4x16-minmax-neon-mull-addw-dup.c",
Frank Barchard8247e212021-02-03 18:12:33 -08001748 "src/qs8-gemm/gen/4x16c2-minmax-neon-mlal-padal-dup.c",
Frank Barchard2302ffd2021-01-22 14:08:02 -08001749 "src/qs8-gemm/gen/4x16c2-minmax-neon-mull-padal-dup.c",
Frank Barchardda78da12021-03-02 14:28:00 -08001750 "src/qs8-gemm/gen/4x16c8-minmax-neon-mlal-padal.c",
Frank Barchard58990122021-01-29 16:09:56 -08001751 "src/qs8-gemm/gen/4x16c8-minmax-neon-mull-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 Barchard671d1b02021-03-10 15:42:08 -08001754 "src/qs8-igemm/gen/1x8-minmax-neon-mull-addw-dup.c",
1755 "src/qs8-igemm/gen/1x8c2-minmax-neon-mlal-padal-dup.c",
1756 "src/qs8-igemm/gen/1x8c2-minmax-neon-mull-padal-dup.c",
Frank Barchardda78da12021-03-02 14:28:00 -08001757 "src/qs8-igemm/gen/1x8c8-minmax-neon-mlal-padal.c",
1758 "src/qs8-igemm/gen/1x8c8-minmax-neon-mull-padal.c",
Frank Barchard671d1b02021-03-10 15:42:08 -08001759 "src/qs8-igemm/gen/1x8c16-minmax-neon-mlal-padal.c",
1760 "src/qs8-igemm/gen/1x16-minmax-neon-mlal-lane.c",
1761 "src/qs8-igemm/gen/1x16-minmax-neon-mull-addw-dup.c",
1762 "src/qs8-igemm/gen/1x16c2-minmax-neon-mlal-padal-dup.c",
1763 "src/qs8-igemm/gen/1x16c2-minmax-neon-mull-padal-dup.c",
1764 "src/qs8-igemm/gen/1x16c8-minmax-neon-mlal-padal.c",
1765 "src/qs8-igemm/gen/1x16c8-minmax-neon-mull-padal.c",
1766 "src/qs8-igemm/gen/1x16c16-minmax-neon-mlal-padal.c",
1767 "src/qs8-igemm/gen/2x8-minmax-neon-mlal-lane.c",
1768 "src/qs8-igemm/gen/2x8-minmax-neon-mull-addw-dup.c",
1769 "src/qs8-igemm/gen/2x8c2-minmax-neon-mlal-padal-dup.c",
1770 "src/qs8-igemm/gen/2x8c2-minmax-neon-mull-padal-dup.c",
Frank Barchardda78da12021-03-02 14:28:00 -08001771 "src/qs8-igemm/gen/2x8c8-minmax-neon-mlal-padal.c",
1772 "src/qs8-igemm/gen/2x8c8-minmax-neon-mull-padal.c",
Frank Barchard671d1b02021-03-10 15:42:08 -08001773 "src/qs8-igemm/gen/2x8c16-minmax-neon-mlal-padal.c",
1774 "src/qs8-igemm/gen/2x16-minmax-neon-mlal-lane.c",
1775 "src/qs8-igemm/gen/2x16-minmax-neon-mull-addw-dup.c",
1776 "src/qs8-igemm/gen/2x16c2-minmax-neon-mlal-padal-dup.c",
1777 "src/qs8-igemm/gen/2x16c2-minmax-neon-mull-padal-dup.c",
1778 "src/qs8-igemm/gen/2x16c8-minmax-neon-mlal-padal.c",
1779 "src/qs8-igemm/gen/2x16c8-minmax-neon-mull-padal.c",
1780 "src/qs8-igemm/gen/2x16c16-minmax-neon-mlal-padal.c",
1781 "src/qs8-igemm/gen/3x8-minmax-neon-mlal-lane.c",
1782 "src/qs8-igemm/gen/3x8-minmax-neon-mull-addw-dup.c",
1783 "src/qs8-igemm/gen/3x8c2-minmax-neon-mlal-padal-dup.c",
1784 "src/qs8-igemm/gen/3x8c2-minmax-neon-mull-padal-dup.c",
Frank Barchardda78da12021-03-02 14:28:00 -08001785 "src/qs8-igemm/gen/3x8c8-minmax-neon-mlal-padal.c",
1786 "src/qs8-igemm/gen/3x8c8-minmax-neon-mull-padal.c",
Frank Barchard671d1b02021-03-10 15:42:08 -08001787 "src/qs8-igemm/gen/3x8c16-minmax-neon-mlal-padal.c",
1788 "src/qs8-igemm/gen/3x16-minmax-neon-mlal-lane.c",
1789 "src/qs8-igemm/gen/3x16-minmax-neon-mull-addw-dup.c",
1790 "src/qs8-igemm/gen/3x16c2-minmax-neon-mlal-padal-dup.c",
1791 "src/qs8-igemm/gen/3x16c2-minmax-neon-mull-padal-dup.c",
1792 "src/qs8-igemm/gen/3x16c8-minmax-neon-mlal-padal.c",
1793 "src/qs8-igemm/gen/3x16c8-minmax-neon-mull-padal.c",
1794 "src/qs8-igemm/gen/3x16c16-minmax-neon-mlal-padal.c",
1795 "src/qs8-igemm/gen/4x8-minmax-neon-mlal-lane.c",
1796 "src/qs8-igemm/gen/4x8-minmax-neon-mull-addw-dup.c",
1797 "src/qs8-igemm/gen/4x8c2-minmax-neon-mlal-padal-dup.c",
1798 "src/qs8-igemm/gen/4x8c2-minmax-neon-mull-padal-dup.c",
Frank Barchardda78da12021-03-02 14:28:00 -08001799 "src/qs8-igemm/gen/4x8c8-minmax-neon-mlal-padal.c",
1800 "src/qs8-igemm/gen/4x8c8-minmax-neon-mull-padal.c",
Frank Barchardda78da12021-03-02 14:28:00 -08001801 "src/qs8-igemm/gen/4x8c16-minmax-neon-mlal-padal.c",
Frank Barchard671d1b02021-03-10 15:42:08 -08001802 "src/qs8-igemm/gen/4x16-minmax-neon-mlal-lane.c",
Frank Barchardda78da12021-03-02 14:28:00 -08001803 "src/qs8-igemm/gen/4x16-minmax-neon-mull-addw-dup.c",
Frank Barchard671d1b02021-03-10 15:42:08 -08001804 "src/qs8-igemm/gen/4x16c2-minmax-neon-mlal-padal-dup.c",
1805 "src/qs8-igemm/gen/4x16c2-minmax-neon-mull-padal-dup.c",
1806 "src/qs8-igemm/gen/4x16c8-minmax-neon-mlal-padal.c",
1807 "src/qs8-igemm/gen/4x16c8-minmax-neon-mull-padal.c",
1808 "src/qs8-igemm/gen/4x16c16-minmax-neon-mlal-padal.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 Barchardb8762632021-03-25 04:00:12 -07003567 "src/qs8-gemm/1x8c8-aarch64-neon-mlal-padal.S",
Frank Barchard1dc9fef2021-03-10 23:50:19 -08003568 "src/qs8-gemm/2x8c8-aarch64-neon-mull-padal.S",
Frank Barchardcbb8e702021-03-12 01:22:55 -08003569 "src/qs8-gemm/2x8c8-aarch64-neon-mlal-padal.S",
Frank Barchard5655cb72021-03-09 20:57:14 -08003570 "src/qs8-gemm/2x8c16-aarch64-neon-mlal-padal.S",
Frank Barchard671d1b02021-03-10 15:42:08 -08003571 "src/qs8-gemm/4x16c4-aarch64-neondot-cortex-a55.S",
Frank Barchardda78da12021-03-02 14:28:00 -08003572 "src/qs8-gemm/4x16c4-aarch64-neondot-ld32.S",
Frank Barchard671d1b02021-03-10 15:42:08 -08003573 "src/qs8-gemm/4x16c4-aarch64-neondot-ld64.S",
Frank Barchard6e35de52021-03-25 04:33:21 -07003574 "src/qs8-igemm/1x8c8-aarch64-neon-mlal-padal.S",
Frank Barchardcbb8e702021-03-12 01:22:55 -08003575 "src/qs8-igemm/2x8c8-aarch64-neon-mlal-padal.S",
Frank Barchard7ca54df2021-03-12 00:09:02 -08003576 "src/qs8-igemm/2x8c16-aarch64-neon-mlal-padal.S",
Frank Barchard89e12f82021-03-08 11:42:44 -08003577 "src/qs8-igemm/4x16c4-aarch64-neondot-cortex-a55.S",
Frank Barchard671d1b02021-03-10 15:42:08 -08003578 "src/qs8-igemm/4x16c4-aarch64-neondot-ld64.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003579]
3580
Marat Dukhan1b354632020-03-23 12:50:22 -07003581INTERNAL_MICROKERNEL_HDRS = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003582 "src/xnnpack/argmaxpool.h",
3583 "src/xnnpack/avgpool.h",
3584 "src/xnnpack/clamp.h",
3585 "src/xnnpack/common.h",
3586 "src/xnnpack/conv.h",
Yury Kartynnike7841862020-11-04 18:22:18 -08003587 "src/xnnpack/depthtospace.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003588 "src/xnnpack/dwconv.h",
Frank Barchard04336c12020-10-22 16:48:55 -07003589 "src/xnnpack/fill.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003590 "src/xnnpack/gavgpool.h",
3591 "src/xnnpack/gemm.h",
3592 "src/xnnpack/hswish.h",
Marat Dukhan660fd192020-03-10 04:55:30 -07003593 "src/xnnpack/ibilinear.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003594 "src/xnnpack/igemm.h",
Marat Dukhancfb31342019-12-05 10:42:57 -08003595 "src/xnnpack/intrinsics-polyfill.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003596 "src/xnnpack/lut.h",
3597 "src/xnnpack/math.h",
3598 "src/xnnpack/maxpool.h",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003599 "src/xnnpack/memory.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003600 "src/xnnpack/packx.h",
3601 "src/xnnpack/pad.h",
3602 "src/xnnpack/params.h",
3603 "src/xnnpack/pavgpool.h",
3604 "src/xnnpack/ppmm.h",
3605 "src/xnnpack/prelu.h",
Marat Dukhan97579532019-10-18 16:40:39 -07003606 "src/xnnpack/raddexpminusmax.h",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07003607 "src/xnnpack/raddextexp.h",
Marat Dukhan97579532019-10-18 16:40:39 -07003608 "src/xnnpack/raddstoreexpminusmax.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003609 "src/xnnpack/rmax.h",
3610 "src/xnnpack/scalar-utils.h",
3611 "src/xnnpack/spmm.h",
3612 "src/xnnpack/unpool.h",
3613 "src/xnnpack/vadd.h",
Marat Dukhan1e782c42019-11-21 17:02:40 -08003614 "src/xnnpack/vbinary.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003615 "src/xnnpack/vmulcaddc.h",
Marat Dukhan05ac8e32019-10-21 15:39:33 -07003616 "src/xnnpack/vscale.h",
Marat Dukhan97579532019-10-18 16:40:39 -07003617 "src/xnnpack/vscaleexpminusmax.h",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07003618 "src/xnnpack/vscaleextexp.h",
Marat Dukhan1e782c42019-11-21 17:02:40 -08003619 "src/xnnpack/vunary.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003620 "src/xnnpack/zip.h",
Marat Dukhan1b354632020-03-23 12:50:22 -07003621]
3622
3623INTERNAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003624 "include/xnnpack.h",
3625 "src/xnnpack/allocator.h",
3626 "src/xnnpack/compute.h",
3627 "src/xnnpack/im2col.h",
3628 "src/xnnpack/indirection.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003629 "src/xnnpack/math-stubs.h",
Chao Mei6ddfc602020-05-13 22:29:36 -07003630 "src/xnnpack/memory-planner.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003631 "src/xnnpack/operator.h",
3632 "src/xnnpack/pack.h",
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07003633 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003634 "src/xnnpack/requantization-stubs.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003635 "src/xnnpack/requantization.h",
Marat Dukhan1d75a542020-02-03 12:23:01 -08003636 "src/xnnpack/subgraph.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003637]
3638
Marat Dukhan1b354632020-03-23 12:50:22 -07003639ACCURACY_EVAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003640 "src/xnnpack/math-stubs.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003641]
3642
Marat Dukhan1b354632020-03-23 12:50:22 -07003643MICROKERNEL_BENCHMARK_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003644 "include/xnnpack.h",
Frank Barchard35db7d02020-10-26 13:37:34 -07003645 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003646]
3647
Marat Dukhan1b354632020-03-23 12:50:22 -07003648MICROKERNEL_TEST_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Frank Barchard35db7d02020-10-26 13:37:34 -07003649 "include/xnnpack.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003650 "src/xnnpack/isa-checks.h",
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07003651 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003652 "src/xnnpack/requantization.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003653]
3654
3655OPERATOR_TEST_PARAMS_HDRS = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003656 "src/xnnpack/common.h",
Frank Barchard04336c12020-10-22 16:48:55 -07003657 "src/xnnpack/params.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003658]
3659
3660WEIGHTS_PACK_HDRS = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003661 "src/xnnpack/compute.h",
Frank Barchard04336c12020-10-22 16:48:55 -07003662 "src/xnnpack/operator.h",
3663 "src/xnnpack/pack.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003664]
3665
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07003666LOGGING_COPTS = select({
3667 # No logging in optimized mode
3668 ":optimized_build": ["-DXNN_LOG_LEVEL=0"],
3669 # Full logging in debug mode
3670 ":debug_build": ["-DXNN_LOG_LEVEL=5"],
3671 # Error-only logging in default (fastbuild) mode
3672 "//conditions:default": ["-DXNN_LOG_LEVEL=2"],
3673})
3674
Marat Dukhan3b59de22020-06-03 20:15:19 -07003675LOGGING_SRCS = select({
3676 # No logging in optimized mode
3677 ":optimized_build": [],
3678 "//conditions:default": [
Marat Dukhanccd3a1d2021-03-29 16:03:12 -07003679 "src/datatype-strings.c",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003680 "src/operator-strings.c",
3681 "src/subgraph-strings.c",
3682 ],
3683})
3684
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07003685LOGGING_HDRS = [
3686 "src/xnnpack/log.h",
3687]
3688
Marat Dukhan08c4a432019-10-03 09:29:21 -07003689xnnpack_cc_library(
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003690 name = "tables",
3691 srcs = TABLE_SRCS,
3692 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003693 gcc_copts = xnnpack_gcc_std_copts(),
3694 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003695)
3696
3697xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003698 name = "scalar_ukernels",
3699 srcs = SCALAR_UKERNELS,
3700 hdrs = INTERNAL_HDRS,
3701 aarch32_copts = ["-marm"],
Marat Dukhan10a38082020-04-17 03:58:35 -07003702 gcc_copts = xnnpack_gcc_std_copts(),
3703 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003704 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003705 ":tables",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003706 "@FP16",
3707 "@FXdiv",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003708 "@pthreadpool",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003709 ],
3710)
3711
3712xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003713 name = "scalar_ukernels_test_mode",
3714 srcs = SCALAR_UKERNELS,
3715 hdrs = INTERNAL_HDRS,
3716 aarch32_copts = ["-marm"],
3717 copts = [
3718 "-UNDEBUG",
3719 "-DXNN_TEST_MODE=1",
3720 ],
3721 gcc_copts = xnnpack_gcc_std_copts(),
3722 msvc_copts = xnnpack_msvc_std_copts(),
3723 deps = [
3724 ":tables",
3725 "@FP16",
3726 "@FXdiv",
3727 "@pthreadpool",
3728 ],
3729)
3730
3731xnnpack_cc_library(
Marat Dukhan436ebe62019-12-04 15:10:12 -08003732 name = "wasm_ukernels",
3733 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003734 gcc_copts = xnnpack_gcc_std_copts(),
3735 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan436ebe62019-12-04 15:10:12 -08003736 wasm_srcs = WASM_UKERNELS,
Marat Dukhan1c6e3892020-06-25 23:56:10 -07003737 wasmsimd_srcs = WASM_UKERNELS + WASMSIMD_UKERNELS,
Marat Dukhan436ebe62019-12-04 15:10:12 -08003738 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003739 ":tables",
Marat Dukhan436ebe62019-12-04 15:10:12 -08003740 "@FP16",
3741 "@FXdiv",
3742 "@pthreadpool",
3743 ],
3744)
3745
3746xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003747 name = "wasm_ukernels_test_mode",
3748 hdrs = INTERNAL_HDRS,
3749 copts = [
3750 "-UNDEBUG",
3751 "-DXNN_TEST_MODE=1",
3752 ],
3753 gcc_copts = xnnpack_gcc_std_copts(),
3754 msvc_copts = xnnpack_msvc_std_copts(),
3755 wasm_srcs = WASM_UKERNELS,
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07003756 wasmsimd_srcs = WASM_UKERNELS + WASMSIMD_UKERNELS,
Marat Dukhan33fcf782020-05-24 14:27:15 -07003757 deps = [
3758 ":tables",
3759 "@FP16",
3760 "@FXdiv",
3761 "@pthreadpool",
3762 ],
3763)
3764
3765xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003766 name = "neon_ukernels",
3767 hdrs = INTERNAL_HDRS,
3768 aarch32_copts = [
3769 "-marm",
Marat Dukhan8853b822020-05-07 12:19:01 -07003770 "-march=armv7-a",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003771 "-mfpu=neon",
3772 ],
3773 aarch32_srcs = NEON_UKERNELS,
3774 aarch64_srcs = NEON_UKERNELS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003775 gcc_copts = xnnpack_gcc_std_copts(),
3776 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08003777 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003778 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003779 "@FP16",
3780 "@pthreadpool",
3781 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003782)
3783
3784xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003785 name = "neon_ukernels_test_mode",
3786 hdrs = INTERNAL_HDRS,
3787 aarch32_copts = [
3788 "-marm",
3789 "-march=armv7-a",
3790 "-mfpu=neon",
3791 ],
3792 aarch32_srcs = NEON_UKERNELS,
3793 aarch64_srcs = NEON_UKERNELS,
3794 copts = [
3795 "-UNDEBUG",
3796 "-DXNN_TEST_MODE=1",
3797 ],
3798 gcc_copts = xnnpack_gcc_std_copts(),
3799 msvc_copts = xnnpack_msvc_std_copts(),
3800 deps = [
3801 ":tables",
3802 "@FP16",
3803 "@pthreadpool",
3804 ],
3805)
3806
3807xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003808 name = "neonfma_ukernels",
3809 hdrs = INTERNAL_HDRS,
3810 aarch32_copts = [
3811 "-marm",
Marat Dukhan8853b822020-05-07 12:19:01 -07003812 "-march=armv7-a",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003813 "-mfpu=neon-vfpv4",
3814 ],
3815 aarch32_srcs = NEONFMA_UKERNELS,
3816 aarch64_srcs = NEONFMA_UKERNELS + AARCH64_NEONFMA_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07003817 apple_aarch32_copts = [
3818 "-mcpu=swift",
3819 "-mtune=generic",
3820 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07003821 gcc_copts = xnnpack_gcc_std_copts(),
3822 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08003823 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003824 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003825 "@FP16",
3826 "@pthreadpool",
3827 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003828)
3829
3830xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003831 name = "neonfma_ukernels_test_mode",
3832 hdrs = INTERNAL_HDRS,
3833 aarch32_copts = [
3834 "-marm",
3835 "-march=armv7-a",
3836 "-mfpu=neon-vfpv4",
3837 ],
3838 aarch32_srcs = NEONFMA_UKERNELS,
3839 aarch64_srcs = NEONFMA_UKERNELS + AARCH64_NEONFMA_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07003840 apple_aarch32_copts = [
3841 "-mcpu=swift",
3842 "-mtune=generic",
3843 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07003844 copts = [
3845 "-UNDEBUG",
3846 "-DXNN_TEST_MODE=1",
3847 ],
3848 gcc_copts = xnnpack_gcc_std_copts(),
3849 msvc_copts = xnnpack_msvc_std_copts(),
3850 deps = [
3851 ":tables",
3852 "@FP16",
3853 "@pthreadpool",
3854 ],
3855)
3856
3857xnnpack_cc_library(
Marat Dukhan8853b822020-05-07 12:19:01 -07003858 name = "neonv8_ukernels",
3859 hdrs = INTERNAL_HDRS,
3860 aarch32_copts = [
3861 "-marm",
3862 "-march=armv8-a",
3863 "-mfpu=neon-fp-armv8",
3864 ],
3865 aarch32_srcs = NEONV8_UKERNELS,
3866 aarch64_srcs = NEONV8_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07003867 apple_aarch32_copts = [
3868 "-mcpu=cyclone",
3869 "-mtune=generic",
3870 ],
Marat Dukhan8853b822020-05-07 12:19:01 -07003871 gcc_copts = xnnpack_gcc_std_copts(),
3872 msvc_copts = xnnpack_msvc_std_copts(),
3873 deps = [
3874 ":tables",
3875 "@FP16",
3876 "@pthreadpool",
3877 ],
3878)
3879
3880xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003881 name = "neonv8_ukernels_test_mode",
3882 hdrs = INTERNAL_HDRS,
3883 aarch32_copts = [
3884 "-marm",
3885 "-march=armv8-a",
3886 "-mfpu=neon-fp-armv8",
3887 ],
3888 aarch32_srcs = NEONV8_UKERNELS,
3889 aarch64_srcs = NEONV8_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07003890 apple_aarch32_copts = [
3891 "-mcpu=cyclone",
3892 "-mtune=generic",
3893 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07003894 copts = [
3895 "-UNDEBUG",
3896 "-DXNN_TEST_MODE=1",
3897 ],
3898 gcc_copts = xnnpack_gcc_std_copts(),
3899 msvc_copts = xnnpack_msvc_std_copts(),
3900 deps = [
3901 ":tables",
3902 "@FP16",
3903 "@pthreadpool",
3904 ],
3905)
3906
3907xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003908 name = "neonfp16arith_ukernels",
3909 hdrs = INTERNAL_HDRS,
3910 aarch64_copts = ["-march=armv8.2-a+fp16"],
3911 aarch64_srcs = AARCH64_NEONFP16ARITH_UKERNELS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003912 gcc_copts = xnnpack_gcc_std_copts(),
3913 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08003914 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003915 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003916 "@FP16",
3917 "@pthreadpool",
3918 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003919)
3920
3921xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003922 name = "neonfp16arith_ukernels_test_mode",
3923 hdrs = INTERNAL_HDRS,
3924 aarch64_copts = ["-march=armv8.2-a+fp16"],
3925 aarch64_srcs = AARCH64_NEONFP16ARITH_UKERNELS,
3926 copts = [
3927 "-UNDEBUG",
3928 "-DXNN_TEST_MODE=1",
3929 ],
3930 gcc_copts = xnnpack_gcc_std_copts(),
3931 msvc_copts = xnnpack_msvc_std_copts(),
3932 deps = [
3933 ":tables",
3934 "@FP16",
3935 "@pthreadpool",
3936 ],
3937)
3938
3939xnnpack_cc_library(
Benoit Jacoba9644732020-08-13 12:48:55 -07003940 name = "neondot_ukernels",
3941 hdrs = INTERNAL_HDRS,
Marat Dukhan799ac752020-08-13 16:08:03 -07003942 aarch32_copts = [
3943 "-marm",
3944 "-march=armv8.2-a+dotprod",
3945 "-mfpu=neon-fp-armv8",
3946 ],
Benoit Jacoba9644732020-08-13 12:48:55 -07003947 aarch32_srcs = NEONDOT_UKERNELS,
3948 aarch64_copts = ["-march=armv8.2-a+dotprod"],
3949 aarch64_srcs = NEONDOT_UKERNELS,
3950 gcc_copts = xnnpack_gcc_std_copts(),
3951 msvc_copts = xnnpack_msvc_std_copts(),
3952 deps = [
3953 ":tables",
3954 "@FP16",
3955 "@pthreadpool",
3956 ],
3957)
3958
3959xnnpack_cc_library(
3960 name = "neondot_ukernels_test_mode",
3961 hdrs = INTERNAL_HDRS,
Marat Dukhan799ac752020-08-13 16:08:03 -07003962 aarch32_copts = [
3963 "-marm",
3964 "-march=armv8.2-a+dotprod",
3965 "-mfpu=neon-fp-armv8",
3966 ],
Benoit Jacoba9644732020-08-13 12:48:55 -07003967 aarch32_srcs = NEONDOT_UKERNELS,
3968 aarch64_copts = ["-march=armv8.2-a+dotprod"],
3969 aarch64_srcs = NEONDOT_UKERNELS,
3970 copts = [
3971 "-UNDEBUG",
3972 "-DXNN_TEST_MODE=1",
3973 ],
3974 gcc_copts = xnnpack_gcc_std_copts(),
3975 msvc_copts = xnnpack_msvc_std_copts(),
3976 deps = [
3977 ":tables",
3978 "@FP16",
3979 "@pthreadpool",
3980 ],
3981)
3982
3983xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003984 name = "sse2_ukernels",
3985 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003986 gcc_copts = xnnpack_gcc_std_copts(),
3987 gcc_x86_copts = ["-msse2"],
3988 msvc_copts = xnnpack_msvc_std_copts(),
3989 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003990 x86_srcs = SSE_UKERNELS + SSE2_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08003991 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003992 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003993 "@FP16",
3994 "@pthreadpool",
3995 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003996)
3997
3998xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003999 name = "sse2_ukernels_test_mode",
4000 hdrs = INTERNAL_HDRS,
4001 copts = [
4002 "-UNDEBUG",
4003 "-DXNN_TEST_MODE=1",
4004 ],
4005 gcc_copts = xnnpack_gcc_std_copts(),
4006 gcc_x86_copts = ["-msse2"],
4007 msvc_copts = xnnpack_msvc_std_copts(),
4008 msvc_x86_32_copts = ["/arch:SSE2"],
4009 x86_srcs = SSE_UKERNELS + SSE2_UKERNELS,
4010 deps = [
4011 ":tables",
4012 "@FP16",
4013 "@pthreadpool",
4014 ],
4015)
4016
4017xnnpack_cc_library(
Marat Dukhanfe7acb62020-03-09 19:30:05 -07004018 name = "ssse3_ukernels",
4019 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004020 gcc_copts = xnnpack_gcc_std_copts(),
4021 gcc_x86_copts = ["-mssse3"],
4022 msvc_copts = xnnpack_msvc_std_copts(),
4023 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhanfe7acb62020-03-09 19:30:05 -07004024 x86_srcs = SSSE3_UKERNELS,
4025 deps = [
4026 ":tables",
4027 "@FP16",
4028 "@pthreadpool",
4029 ],
4030)
4031
4032xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07004033 name = "ssse3_ukernels_test_mode",
4034 hdrs = INTERNAL_HDRS,
4035 copts = [
4036 "-UNDEBUG",
4037 "-DXNN_TEST_MODE=1",
4038 ],
4039 gcc_copts = xnnpack_gcc_std_copts(),
4040 gcc_x86_copts = ["-mssse3"],
4041 msvc_copts = xnnpack_msvc_std_copts(),
4042 msvc_x86_32_copts = ["/arch:SSE2"],
4043 x86_srcs = SSSE3_UKERNELS,
4044 deps = [
4045 ":tables",
4046 "@FP16",
4047 "@pthreadpool",
4048 ],
4049)
4050
4051xnnpack_cc_library(
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08004052 name = "sse41_ukernels",
4053 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004054 gcc_copts = xnnpack_gcc_std_copts(),
4055 gcc_x86_copts = ["-msse4.1"],
4056 msvc_copts = xnnpack_msvc_std_copts(),
4057 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08004058 x86_srcs = SSE41_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08004059 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08004060 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08004061 "@FP16",
4062 "@pthreadpool",
4063 ],
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08004064)
4065
4066xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07004067 name = "sse41_ukernels_test_mode",
4068 hdrs = INTERNAL_HDRS,
4069 copts = [
4070 "-UNDEBUG",
4071 "-DXNN_TEST_MODE=1",
4072 ],
4073 gcc_copts = xnnpack_gcc_std_copts(),
4074 gcc_x86_copts = ["-msse4.1"],
4075 msvc_copts = xnnpack_msvc_std_copts(),
4076 msvc_x86_32_copts = ["/arch:SSE2"],
4077 x86_srcs = SSE41_UKERNELS,
4078 deps = [
4079 ":tables",
4080 "@FP16",
4081 "@pthreadpool",
4082 ],
4083)
4084
4085xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004086 name = "avx_ukernels",
4087 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004088 gcc_copts = xnnpack_gcc_std_copts(),
4089 gcc_x86_copts = ["-mavx"],
4090 msvc_copts = xnnpack_msvc_std_copts(),
4091 msvc_x86_32_copts = ["/arch:AVX"],
4092 msvc_x86_64_copts = ["/arch:AVX"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004093 x86_srcs = AVX_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08004094 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08004095 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08004096 "@FP16",
4097 "@pthreadpool",
4098 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004099)
4100
4101xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07004102 name = "avx_ukernels_test_mode",
4103 hdrs = INTERNAL_HDRS,
4104 copts = [
4105 "-UNDEBUG",
4106 "-DXNN_TEST_MODE=1",
4107 ],
4108 gcc_copts = xnnpack_gcc_std_copts(),
4109 gcc_x86_copts = ["-mavx"],
4110 msvc_copts = xnnpack_msvc_std_copts(),
4111 msvc_x86_32_copts = ["/arch:AVX"],
4112 msvc_x86_64_copts = ["/arch:AVX"],
4113 x86_srcs = AVX_UKERNELS,
4114 deps = [
4115 ":tables",
4116 "@FP16",
4117 "@pthreadpool",
4118 ],
4119)
4120
4121xnnpack_cc_library(
Marat Dukhan1566fee2020-08-02 21:55:41 -07004122 name = "xop_ukernels",
4123 hdrs = INTERNAL_HDRS,
4124 gcc_copts = xnnpack_gcc_std_copts(),
4125 gcc_x86_copts = ["-mxop"],
4126 msvc_copts = xnnpack_msvc_std_copts(),
4127 msvc_x86_32_copts = ["/arch:AVX"],
4128 msvc_x86_64_copts = ["/arch:AVX"],
4129 x86_srcs = XOP_UKERNELS,
4130 deps = [
4131 ":tables",
4132 "@FP16",
4133 "@pthreadpool",
4134 ],
4135)
4136
4137xnnpack_cc_library(
4138 name = "xop_ukernels_test_mode",
4139 hdrs = INTERNAL_HDRS,
4140 copts = [
4141 "-UNDEBUG",
4142 "-DXNN_TEST_MODE=1",
4143 ],
4144 gcc_copts = xnnpack_gcc_std_copts(),
4145 gcc_x86_copts = ["-mxop"],
4146 msvc_copts = xnnpack_msvc_std_copts(),
4147 msvc_x86_32_copts = ["/arch:AVX"],
4148 msvc_x86_64_copts = ["/arch:AVX"],
4149 x86_srcs = XOP_UKERNELS,
4150 deps = [
4151 ":tables",
4152 "@FP16",
4153 "@pthreadpool",
4154 ],
4155)
4156
4157xnnpack_cc_library(
Marat Dukhanfda12b82019-11-21 12:27:59 -08004158 name = "fma3_ukernels",
4159 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004160 gcc_copts = xnnpack_gcc_std_copts(),
4161 gcc_x86_copts = ["-mfma"],
4162 msvc_copts = xnnpack_msvc_std_copts(),
4163 msvc_x86_32_copts = ["/arch:AVX"],
4164 msvc_x86_64_copts = ["/arch:AVX"],
Marat Dukhanfda12b82019-11-21 12:27:59 -08004165 x86_srcs = FMA3_UKERNELS,
4166 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08004167 ":tables",
Marat Dukhanfda12b82019-11-21 12:27:59 -08004168 "@FP16",
4169 "@pthreadpool",
4170 ],
4171)
4172
4173xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07004174 name = "fma3_ukernels_test_mode",
4175 hdrs = INTERNAL_HDRS,
4176 copts = [
4177 "-UNDEBUG",
4178 "-DXNN_TEST_MODE=1",
4179 ],
4180 gcc_copts = xnnpack_gcc_std_copts(),
4181 gcc_x86_copts = ["-mfma"],
4182 msvc_copts = xnnpack_msvc_std_copts(),
4183 msvc_x86_32_copts = ["/arch:AVX"],
4184 msvc_x86_64_copts = ["/arch:AVX"],
4185 x86_srcs = FMA3_UKERNELS,
4186 deps = [
4187 ":tables",
4188 "@FP16",
4189 "@pthreadpool",
4190 ],
4191)
4192
4193xnnpack_cc_library(
Marat Dukhan6adff4e2019-10-14 18:32:07 -07004194 name = "avx2_ukernels",
4195 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004196 gcc_copts = xnnpack_gcc_std_copts(),
4197 gcc_x86_copts = [
Marat Dukhan6adff4e2019-10-14 18:32:07 -07004198 "-mfma",
4199 "-mavx2",
4200 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07004201 msvc_copts = xnnpack_msvc_std_copts(),
4202 msvc_x86_32_copts = ["/arch:AVX2"],
4203 msvc_x86_64_copts = ["/arch:AVX2"],
Marat Dukhan6adff4e2019-10-14 18:32:07 -07004204 x86_srcs = AVX2_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08004205 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08004206 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08004207 "@FP16",
4208 "@pthreadpool",
4209 ],
Marat Dukhan6adff4e2019-10-14 18:32:07 -07004210)
4211
4212xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07004213 name = "avx2_ukernels_test_mode",
4214 hdrs = INTERNAL_HDRS,
4215 copts = [
4216 "-UNDEBUG",
4217 "-DXNN_TEST_MODE=1",
4218 ],
4219 gcc_copts = xnnpack_gcc_std_copts(),
4220 gcc_x86_copts = [
4221 "-mfma",
4222 "-mavx2",
4223 ],
4224 msvc_copts = xnnpack_msvc_std_copts(),
4225 msvc_x86_32_copts = ["/arch:AVX2"],
4226 msvc_x86_64_copts = ["/arch:AVX2"],
4227 x86_srcs = AVX2_UKERNELS,
4228 deps = [
4229 ":tables",
4230 "@FP16",
4231 "@pthreadpool",
4232 ],
4233)
4234
4235xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004236 name = "avx512f_ukernels",
4237 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004238 gcc_copts = xnnpack_gcc_std_copts(),
4239 gcc_x86_copts = ["-mavx512f"],
4240 mingw_copts = ["-fno-asynchronous-unwind-tables"],
4241 msvc_copts = xnnpack_msvc_std_copts(),
4242 msvc_x86_32_copts = ["/arch:AVX512"],
4243 msvc_x86_64_copts = ["/arch:AVX512"],
4244 msys_copts = ["-fno-asynchronous-unwind-tables"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004245 x86_srcs = AVX512F_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08004246 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08004247 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08004248 "@FP16",
4249 "@pthreadpool",
4250 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004251)
4252
4253xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07004254 name = "avx512f_ukernels_test_mode",
4255 hdrs = INTERNAL_HDRS,
4256 copts = [
4257 "-UNDEBUG",
4258 "-DXNN_TEST_MODE=1",
4259 ],
4260 gcc_copts = xnnpack_gcc_std_copts(),
4261 gcc_x86_copts = ["-mavx512f"],
4262 mingw_copts = ["-fno-asynchronous-unwind-tables"],
4263 msvc_copts = xnnpack_msvc_std_copts(),
4264 msvc_x86_32_copts = ["/arch:AVX512"],
4265 msvc_x86_64_copts = ["/arch:AVX512"],
4266 msys_copts = ["-fno-asynchronous-unwind-tables"],
4267 x86_srcs = AVX512F_UKERNELS,
4268 deps = [
4269 ":tables",
4270 "@FP16",
4271 "@pthreadpool",
4272 ],
4273)
4274
4275xnnpack_cc_library(
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07004276 name = "avx512skx_ukernels",
4277 hdrs = INTERNAL_HDRS,
4278 gcc_copts = xnnpack_gcc_std_copts(),
4279 gcc_x86_copts = [
4280 "-mavx512f",
4281 "-mavx512cd",
4282 "-mavx512bw",
4283 "-mavx512dq",
4284 "-mavx512vl",
4285 ],
4286 mingw_copts = ["-fno-asynchronous-unwind-tables"],
4287 msvc_copts = xnnpack_msvc_std_copts(),
4288 msvc_x86_32_copts = ["/arch:AVX512"],
4289 msvc_x86_64_copts = ["/arch:AVX512"],
4290 msys_copts = ["-fno-asynchronous-unwind-tables"],
4291 x86_srcs = AVX512SKX_UKERNELS,
4292 deps = [
4293 ":tables",
4294 "@FP16",
4295 "@pthreadpool",
4296 ],
4297)
4298
4299xnnpack_cc_library(
4300 name = "avx512skx_ukernels_test_mode",
4301 hdrs = INTERNAL_HDRS,
4302 copts = [
4303 "-UNDEBUG",
4304 "-DXNN_TEST_MODE=1",
4305 ],
4306 gcc_copts = xnnpack_gcc_std_copts(),
4307 gcc_x86_copts = [
4308 "-mavx512f",
4309 "-mavx512cd",
4310 "-mavx512bw",
4311 "-mavx512dq",
4312 "-mavx512vl",
4313 ],
4314 mingw_copts = ["-fno-asynchronous-unwind-tables"],
4315 msvc_copts = xnnpack_msvc_std_copts(),
4316 msvc_x86_32_copts = ["/arch:AVX512"],
4317 msvc_x86_64_copts = ["/arch:AVX512"],
4318 msys_copts = ["-fno-asynchronous-unwind-tables"],
4319 x86_srcs = AVX512SKX_UKERNELS,
4320 deps = [
4321 ":tables",
4322 "@FP16",
4323 "@pthreadpool",
4324 ],
4325)
4326
4327xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004328 name = "asm_ukernels",
4329 hdrs = ["src/xnnpack/assembly.h"],
4330 aarch32_srcs = AARCH32_ASM_UKERNELS,
Frank Barchard31bb45b2020-10-06 00:26:33 -07004331 aarch64_copts = ["-march=armv8.2-a+fp16+dotprod"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004332 aarch64_srcs = AARCH64_ASM_UKERNELS,
Frank Barchardbcedc082020-08-17 18:00:51 -07004333 wasm_srcs = WASM32_ASM_UKERNELS,
4334 wasmsimd_srcs = WASM32_ASM_UKERNELS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07004335)
4336
Marat Dukhan3b59de22020-06-03 20:15:19 -07004337xnnpack_cc_library(
4338 name = "logging_utils",
4339 srcs = LOGGING_SRCS,
4340 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
4341 copts = LOGGING_COPTS + [
4342 "-Isrc",
4343 "-Iinclude",
4344 ] + select({
4345 ":debug_build": [],
4346 "//conditions:default": xnnpack_min_size_copts(),
4347 }),
4348 gcc_copts = xnnpack_gcc_std_copts(),
4349 msvc_copts = xnnpack_msvc_std_copts(),
4350 visibility = xnnpack_visibility(),
4351 deps = [
4352 "@FP16",
4353 "@clog",
4354 "@pthreadpool",
4355 ],
4356)
4357
Marat Dukhan08c4a432019-10-03 09:29:21 -07004358xnnpack_aggregate_library(
4359 name = "ukernels",
4360 aarch32_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07004361 ":neon_ukernels",
4362 ":neonfma_ukernels",
Marat Dukhan8853b822020-05-07 12:19:01 -07004363 ":neonv8_ukernels",
Benoit Jacoba9644732020-08-13 12:48:55 -07004364 ":neondot_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004365 ":asm_ukernels",
4366 ],
4367 aarch64_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07004368 ":neon_ukernels",
4369 ":neonfma_ukernels",
Marat Dukhan8853b822020-05-07 12:19:01 -07004370 ":neonv8_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004371 ":neonfp16arith_ukernels",
Benoit Jacoba9644732020-08-13 12:48:55 -07004372 ":neondot_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004373 ":asm_ukernels",
4374 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07004375 generic_deps = [
4376 ":scalar_ukernels",
4377 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07004378 wasm_deps = [
4379 ":wasm_ukernels",
Frank Barchardbcedc082020-08-17 18:00:51 -07004380 ":asm_ukernels",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004381 ],
4382 wasmsimd_deps = [
4383 ":wasm_ukernels",
Frank Barchardbcedc082020-08-17 18:00:51 -07004384 ":asm_ukernels",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004385 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004386 x86_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07004387 ":sse2_ukernels",
Marat Dukhanfe7acb62020-03-09 19:30:05 -07004388 ":ssse3_ukernels",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08004389 ":sse41_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004390 ":avx_ukernels",
Marat Dukhan1566fee2020-08-02 21:55:41 -07004391 ":xop_ukernels",
Marat Dukhanfda12b82019-11-21 12:27:59 -08004392 ":fma3_ukernels",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07004393 ":avx2_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004394 ":avx512f_ukernels",
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07004395 ":avx512skx_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004396 ],
4397)
4398
Marat Dukhan33fcf782020-05-24 14:27:15 -07004399xnnpack_aggregate_library(
4400 name = "ukernels_test_mode",
4401 aarch32_deps = [
4402 ":neon_ukernels_test_mode",
4403 ":neonfma_ukernels_test_mode",
4404 ":neonv8_ukernels_test_mode",
Benoit Jacoba9644732020-08-13 12:48:55 -07004405 ":neondot_ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004406 ":asm_ukernels",
4407 ],
4408 aarch64_deps = [
4409 ":neon_ukernels_test_mode",
4410 ":neonfma_ukernels_test_mode",
4411 ":neonv8_ukernels_test_mode",
4412 ":neonfp16arith_ukernels_test_mode",
Benoit Jacoba9644732020-08-13 12:48:55 -07004413 ":neondot_ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004414 ":asm_ukernels",
4415 ],
4416 generic_deps = [
4417 ":scalar_ukernels_test_mode",
4418 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07004419 wasm_deps = [
4420 ":wasm_ukernels_test_mode",
Frank Barchardbcedc082020-08-17 18:00:51 -07004421 ":asm_ukernels",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004422 ],
4423 wasmsimd_deps = [
4424 ":wasm_ukernels_test_mode",
Frank Barchardbcedc082020-08-17 18:00:51 -07004425 ":asm_ukernels",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004426 ],
4427 x86_deps = [
4428 ":sse2_ukernels_test_mode",
4429 ":ssse3_ukernels_test_mode",
4430 ":sse41_ukernels_test_mode",
4431 ":avx_ukernels_test_mode",
Marat Dukhan1566fee2020-08-02 21:55:41 -07004432 ":xop_ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004433 ":fma3_ukernels_test_mode",
4434 ":avx2_ukernels_test_mode",
4435 ":avx512f_ukernels_test_mode",
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07004436 ":avx512skx_ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004437 ],
4438)
4439
Marat Dukhan86df3842020-06-17 14:03:31 -07004440# This target must be compatible with Android NDK r17 and compile with Clang 6.
4441# It is similar to :ukernels target, but excludes microkernels which require
4442# NEON FP16 arithmetics, NEON Dot Product, and newer intrinsics.
Marat Dukhanae046f52020-06-15 13:16:14 -07004443xnnpack_aggregate_library(
Marat Dukhan86df3842020-06-17 14:03:31 -07004444 name = "ukernels_max_compatibility",
Marat Dukhanae046f52020-06-15 13:16:14 -07004445 aarch32_deps = [
4446 ":neon_ukernels",
4447 ":neonfma_ukernels",
4448 ":neonv8_ukernels",
4449 ":asm_ukernels",
4450 ],
4451 aarch64_deps = [
4452 ":neon_ukernels",
4453 ":neonfma_ukernels",
4454 ":neonv8_ukernels",
4455 ":asm_ukernels",
4456 ],
4457 generic_deps = [
4458 ":scalar_ukernels",
4459 ],
Marat Dukhanae046f52020-06-15 13:16:14 -07004460 wasm_deps = [
4461 ":wasm_ukernels",
4462 ],
4463 wasmsimd_deps = [
4464 ":wasm_ukernels",
4465 ],
4466 x86_deps = [
4467 ":sse2_ukernels",
4468 ":ssse3_ukernels",
4469 ":sse41_ukernels",
4470 ":avx_ukernels",
Marat Dukhan1566fee2020-08-02 21:55:41 -07004471 ":xop_ukernels",
Marat Dukhanae046f52020-06-15 13:16:14 -07004472 ":fma3_ukernels",
4473 ":avx2_ukernels",
4474 ":avx512f_ukernels",
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07004475 ":avx512skx_ukernels",
Marat Dukhanae046f52020-06-15 13:16:14 -07004476 ],
4477)
4478
Marat Dukhan08c4a432019-10-03 09:29:21 -07004479xnnpack_cc_library(
4480 name = "im2col",
4481 srcs = ["src/im2col.c"],
4482 hdrs = [
4483 "src/xnnpack/common.h",
4484 "src/xnnpack/im2col.h",
4485 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07004486 gcc_copts = xnnpack_gcc_std_copts(),
4487 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004488)
4489
4490xnnpack_cc_library(
4491 name = "indirection",
4492 srcs = ["src/indirection.c"],
4493 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004494 gcc_copts = xnnpack_gcc_std_copts(),
4495 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004496 deps = [
4497 "@FP16",
4498 "@FXdiv",
4499 "@pthreadpool",
4500 ],
4501)
4502
4503xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07004504 name = "indirection_test_mode",
4505 srcs = ["src/indirection.c"],
4506 hdrs = INTERNAL_HDRS,
4507 copts = [
4508 "-UNDEBUG",
4509 "-DXNN_TEST_MODE=1",
4510 ],
4511 gcc_copts = xnnpack_gcc_std_copts(),
4512 msvc_copts = xnnpack_msvc_std_copts(),
4513 deps = [
4514 "@FP16",
4515 "@FXdiv",
4516 "@pthreadpool",
4517 ],
4518)
4519
4520xnnpack_cc_library(
Marat Dukhanab582382020-07-06 13:32:08 -07004521 name = "packing",
4522 srcs = ["src/packing.c"],
4523 hdrs = INTERNAL_HDRS,
4524 gcc_copts = xnnpack_gcc_std_copts(),
4525 msvc_copts = xnnpack_msvc_std_copts(),
4526 deps = [
4527 "@FP16",
4528 "@FXdiv",
4529 "@pthreadpool",
4530 ],
4531)
4532
4533xnnpack_cc_library(
4534 name = "packing_test_mode",
4535 srcs = ["src/packing.c"],
4536 hdrs = INTERNAL_HDRS,
4537 copts = [
4538 "-UNDEBUG",
4539 "-DXNN_TEST_MODE=1",
4540 ],
4541 gcc_copts = xnnpack_gcc_std_copts(),
4542 msvc_copts = xnnpack_msvc_std_copts(),
4543 deps = [
4544 "@FP16",
4545 "@FXdiv",
4546 "@pthreadpool",
4547 ],
4548)
4549
4550xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004551 name = "operator_run",
4552 srcs = ["src/operator-run.c"],
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07004553 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004554 copts = LOGGING_COPTS + select({
Marat Dukhan05702cf2020-03-26 15:41:33 -07004555 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4556 "//conditions:default": [],
4557 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07004558 gcc_copts = xnnpack_gcc_std_copts(),
4559 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004560 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07004561 ":logging_utils",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004562 "@FP16",
4563 "@FXdiv",
4564 "@clog",
4565 "@pthreadpool",
4566 ],
4567)
4568
Chao Mei6ddfc602020-05-13 22:29:36 -07004569xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07004570 name = "operator_run_test_mode",
4571 srcs = ["src/operator-run.c"],
4572 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
4573 copts = LOGGING_COPTS + [
4574 "-UNDEBUG",
4575 "-DXNN_TEST_MODE=1",
4576 ] + select({
4577 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4578 "//conditions:default": [],
4579 }),
4580 gcc_copts = xnnpack_gcc_std_copts(),
4581 msvc_copts = xnnpack_msvc_std_copts(),
4582 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07004583 ":logging_utils",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004584 "@FP16",
4585 "@FXdiv",
4586 "@clog",
4587 "@pthreadpool",
4588 ],
4589)
4590
4591xnnpack_cc_library(
Chao Mei6ddfc602020-05-13 22:29:36 -07004592 name = "memory_planner",
4593 srcs = ["src/memory-planner.c"],
4594 hdrs = INTERNAL_HDRS,
4595 defines = select({
4596 ":xnn_enable_memopt_explicit_true": ["XNN_ENABLE_MEMOPT=1"],
4597 ":xnn_enable_memopt_explicit_false": ["XNN_ENABLE_MEMOPT=0"],
4598 "//conditions:default": ["XNN_ENABLE_MEMOPT=1"],
4599 }),
4600 gcc_copts = xnnpack_gcc_std_copts(),
4601 msvc_copts = xnnpack_msvc_std_copts(),
4602 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07004603 ":logging_utils",
Chao Mei6ddfc602020-05-13 22:29:36 -07004604 "@pthreadpool",
4605 ],
4606)
4607
Marat Dukhan33fcf782020-05-24 14:27:15 -07004608xnnpack_cc_library(
4609 name = "memory_planner_test_mode",
4610 srcs = ["src/memory-planner.c"],
4611 hdrs = INTERNAL_HDRS,
4612 copts = [
4613 "-UNDEBUG",
4614 "-DXNN_TEST_MODE=1",
4615 ],
4616 defines = select({
4617 ":xnn_enable_memopt_explicit_true": ["XNN_ENABLE_MEMOPT=1"],
4618 ":xnn_enable_memopt_explicit_false": ["XNN_ENABLE_MEMOPT=0"],
4619 "//conditions:default": ["XNN_ENABLE_MEMOPT=1"],
4620 }),
4621 gcc_copts = xnnpack_gcc_std_copts(),
4622 msvc_copts = xnnpack_msvc_std_copts(),
4623 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07004624 ":logging_utils",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004625 "@pthreadpool",
4626 ],
4627)
4628
Marat Dukhan08c4a432019-10-03 09:29:21 -07004629cc_library(
4630 name = "enable_assembly",
4631 defines = select({
4632 ":xnn_enable_assembly_explicit_true": ["XNN_ENABLE_ASSEMBLY=1"],
4633 ":xnn_enable_assembly_explicit_false": ["XNN_ENABLE_ASSEMBLY=0"],
Frank Barchard810171d2019-10-10 10:34:51 -07004634 "//conditions:default": ["XNN_ENABLE_ASSEMBLY=1"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004635 }),
4636)
4637
Marat Dukhan9de90e02020-06-18 16:04:12 -07004638cc_library(
4639 name = "enable_sparse",
4640 defines = select({
4641 ":xnn_enable_sparse_explicit_true": ["XNN_ENABLE_SPARSE=1"],
4642 ":xnn_enable_sparse_explicit_false": ["XNN_ENABLE_SPARSE=0"],
Marat Dukhanb36582b2020-12-08 11:16:28 -08004643 "//conditions:default": ["XNN_ENABLE_SPARSE=1"],
Marat Dukhan9de90e02020-06-18 16:04:12 -07004644 }),
4645)
4646
Marat Dukhancf056b22019-10-07 10:26:29 -07004647xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004648 name = "operators",
4649 srcs = OPERATOR_SRCS + [
Marat Dukhan04f03be2019-11-19 12:36:47 -08004650 "src/memory.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004651 "src/operator-delete.c",
Marat Dukhancf056b22019-10-07 10:26:29 -07004652 ],
4653 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004654 copts = LOGGING_COPTS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07004655 "-Isrc",
4656 "-Iinclude",
4657 ] + select({
4658 ":debug_build": [],
4659 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07004660 }) + select({
4661 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4662 "//conditions:default": [],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004663 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07004664 gcc_copts = xnnpack_gcc_std_copts(),
4665 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004666 deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07004667 ":indirection",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004668 ":logging_utils",
Marat Dukhanab582382020-07-06 13:32:08 -07004669 ":packing",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004670 "@FP16",
4671 "@FXdiv",
4672 "@clog",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004673 "@pthreadpool",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004674 ],
4675)
4676
Marat Dukhan10a38082020-04-17 03:58:35 -07004677xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07004678 name = "operators_test_mode",
4679 srcs = OPERATOR_SRCS + [
4680 "src/memory.c",
4681 "src/operator-delete.c",
4682 ],
4683 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
4684 copts = LOGGING_COPTS + [
4685 "-Isrc",
4686 "-Iinclude",
4687 "-UNDEBUG",
4688 "-DXNN_TEST_MODE=1",
4689 ] + select({
4690 ":debug_build": [],
4691 "//conditions:default": xnnpack_min_size_copts(),
4692 }) + select({
4693 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4694 "//conditions:default": [],
4695 }),
4696 gcc_copts = xnnpack_gcc_std_copts(),
4697 msvc_copts = xnnpack_msvc_std_copts(),
4698 deps = [
4699 ":indirection_test_mode",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004700 ":logging_utils",
Marat Dukhanab582382020-07-06 13:32:08 -07004701 ":packing_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004702 "@FP16",
4703 "@FXdiv",
4704 "@clog",
4705 "@pthreadpool",
4706 ],
4707)
4708
4709xnnpack_cc_library(
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004710 name = "XNNPACK",
4711 srcs = [
4712 "src/init.c",
Marat Dukhanccfdbd12020-02-03 14:27:45 -08004713 "src/runtime.c",
4714 "src/subgraph.c",
4715 "src/tensor.c",
Marat Dukhanf03da0d2020-06-10 16:00:20 -07004716 ] + SUBGRAPH_SRCS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004717 hdrs = ["include/xnnpack.h"],
4718 copts = LOGGING_COPTS + [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004719 "-Isrc",
4720 "-Iinclude",
4721 ] + select({
4722 ":debug_build": [],
4723 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07004724 }) + select({
4725 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4726 "//conditions:default": [],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004727 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07004728 gcc_copts = xnnpack_gcc_std_copts(),
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004729 includes = ["include"],
Marat Dukhan10a38082020-04-17 03:58:35 -07004730 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004731 visibility = xnnpack_visibility(),
4732 deps = [
4733 ":enable_assembly",
Marat Dukhan9de90e02020-06-18 16:04:12 -07004734 ":enable_sparse",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004735 ":logging_utils",
Chao Mei6ddfc602020-05-13 22:29:36 -07004736 ":memory_planner",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004737 ":operator_run",
4738 ":operators",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004739 ":ukernels",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004740 "@clog",
Marat Dukhanab2946c2020-05-21 20:04:13 -07004741 "@FP16",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004742 "@pthreadpool",
Marat Dukhand343c222019-10-07 09:22:14 -07004743 ] + select({
4744 ":emscripten": [],
4745 "//conditions:default": ["@cpuinfo"],
4746 }),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004747)
4748
Marat Dukhan10a38082020-04-17 03:58:35 -07004749xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07004750 name = "XNNPACK_test_mode",
4751 srcs = [
4752 "src/init.c",
4753 "src/runtime.c",
4754 "src/subgraph.c",
4755 "src/tensor.c",
Marat Dukhanf03da0d2020-06-10 16:00:20 -07004756 ] + SUBGRAPH_SRCS,
Marat Dukhan33fcf782020-05-24 14:27:15 -07004757 hdrs = ["include/xnnpack.h"],
4758 copts = LOGGING_COPTS + [
4759 "-Isrc",
4760 "-Iinclude",
4761 "-UNDEBUG",
4762 "-DXNN_TEST_MODE=1",
4763 ] + select({
4764 ":debug_build": [],
4765 "//conditions:default": xnnpack_min_size_copts(),
4766 }) + select({
4767 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4768 "//conditions:default": [],
4769 }),
4770 gcc_copts = xnnpack_gcc_std_copts(),
4771 includes = ["include"],
4772 msvc_copts = xnnpack_msvc_std_copts(),
4773 visibility = xnnpack_visibility(),
4774 deps = [
4775 ":enable_assembly",
Marat Dukhan9de90e02020-06-18 16:04:12 -07004776 ":enable_sparse",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004777 ":logging_utils",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004778 ":memory_planner_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004779 ":operator_run_test_mode",
4780 ":operators_test_mode",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004781 ":ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004782 "@clog",
4783 "@FP16",
4784 "@pthreadpool",
4785 ] + select({
4786 ":emscripten": [],
4787 "//conditions:default": ["@cpuinfo"],
4788 }),
4789)
4790
4791xnnpack_cc_library(
Marat Dukhanae046f52020-06-15 13:16:14 -07004792 name = "xnnpack_f32",
4793 srcs = [
4794 "src/init.c",
4795 "src/runtime.c",
4796 "src/subgraph.c",
4797 "src/tensor.c",
4798 ] + SUBGRAPH_SRCS,
4799 hdrs = ["include/xnnpack.h"],
4800 copts = LOGGING_COPTS + [
4801 "-Isrc",
4802 "-Iinclude",
4803 ] + select({
4804 ":debug_build": [],
4805 "//conditions:default": xnnpack_min_size_copts(),
4806 }) + select({
4807 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4808 "//conditions:default": [],
4809 }),
4810 defines = [
Marat Dukhan16f1e1a2020-08-04 16:38:22 -07004811 "XNN_NO_QS8_OPERATORS",
Marat Dukhan08b7a972020-07-14 18:17:29 -07004812 "XNN_NO_QU8_OPERATORS",
Marat Dukhanae046f52020-06-15 13:16:14 -07004813 "XNN_NO_U8_OPERATORS",
4814 "XNN_NO_X8_OPERATORS",
4815 "XNN_NO_F16_OPERATORS",
4816 "XNN_NO_X16_OPERATORS",
4817 ],
4818 gcc_copts = xnnpack_gcc_std_copts(),
4819 includes = ["include"],
4820 msvc_copts = xnnpack_msvc_std_copts(),
4821 visibility = xnnpack_visibility(),
4822 deps = [
4823 ":enable_assembly",
Marat Dukhan9de90e02020-06-18 16:04:12 -07004824 ":enable_sparse",
Marat Dukhanae046f52020-06-15 13:16:14 -07004825 ":logging_utils",
4826 ":memory_planner",
4827 ":operator_run",
4828 ":operators",
Marat Dukhan86df3842020-06-17 14:03:31 -07004829 ":ukernels_max_compatibility",
Marat Dukhanae046f52020-06-15 13:16:14 -07004830 "@clog",
4831 "@FP16",
4832 "@pthreadpool",
4833 ] + select({
4834 ":emscripten": [],
4835 "//conditions:default": ["@cpuinfo"],
4836 }),
4837)
4838
Marat Dukhanf0cb70a2021-03-30 15:45:15 -07004839# Specialized XNNPACK version for TensorFlow Lite. Excludes operators currently
4840# not used by the TensorFlow Lite XNNPACK delegate to minimize code size.
Marat Dukhanae046f52020-06-15 13:16:14 -07004841xnnpack_cc_library(
Marat Dukhanf0cb70a2021-03-30 15:45:15 -07004842 name = "xnnpack_for_tflite",
4843 srcs = [
4844 "src/init.c",
4845 "src/runtime.c",
4846 "src/subgraph.c",
4847 "src/tensor.c",
4848 ] + SUBGRAPH_SRCS,
4849 hdrs = ["include/xnnpack.h"],
4850 copts = LOGGING_COPTS + [
4851 "-Isrc",
4852 "-Iinclude",
4853 ] + select({
4854 ":debug_build": [],
4855 "//conditions:default": xnnpack_min_size_copts(),
4856 }) + select({
4857 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4858 "//conditions:default": [],
4859 }),
4860 defines = [
4861 "XNN_NO_QS8_OPERATORS",
4862 "XNN_NO_QU8_OPERATORS",
4863 "XNN_NO_U8_OPERATORS",
4864 "XNN_NO_X8_OPERATORS",
4865 "XNN_NO_F16_OPERATORS",
4866 "XNN_NO_X16_OPERATORS",
4867 ],
4868 gcc_copts = xnnpack_gcc_std_copts(),
4869 includes = ["include"],
4870 msvc_copts = xnnpack_msvc_std_copts(),
4871 visibility = xnnpack_visibility(),
4872 deps = [
4873 ":enable_assembly",
4874 ":enable_sparse",
4875 ":logging_utils",
4876 ":memory_planner",
4877 ":operator_run",
4878 ":operators",
4879 ":ukernels_max_compatibility",
4880 "@clog",
4881 "@FP16",
4882 "@pthreadpool",
4883 ] + select({
4884 ":emscripten": [],
4885 "//conditions:default": ["@cpuinfo"],
4886 }),
4887)
4888
4889# Specialized XNNPACK version for TensorFlow.js. Excludes operators currently
4890# not used by the TensorFlow.js WebAssembly backend to minimize code size.
4891xnnpack_cc_library(
4892 name = "xnnpack_for_tfjs",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004893 srcs = [
4894 "src/init.c",
4895 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07004896 hdrs = ["include/xnnpack.h"],
4897 copts = LOGGING_COPTS + [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004898 "-Isrc",
4899 "-Iinclude",
4900 ] + select({
4901 ":debug_build": [],
4902 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07004903 }) + select({
4904 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4905 "//conditions:default": [],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004906 }),
4907 defines = [
Marat Dukhan16f1e1a2020-08-04 16:38:22 -07004908 "XNN_NO_QS8_OPERATORS",
Marat Dukhan08b7a972020-07-14 18:17:29 -07004909 "XNN_NO_QU8_OPERATORS",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004910 "XNN_NO_U8_OPERATORS",
4911 "XNN_NO_X8_OPERATORS",
Marat Dukhanefc47b82019-11-18 09:25:38 -08004912 "XNN_NO_NCHW_OPERATORS",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004913 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07004914 gcc_copts = xnnpack_gcc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004915 includes = ["include"],
Marat Dukhan10a38082020-04-17 03:58:35 -07004916 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004917 visibility = xnnpack_visibility(),
4918 deps = [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004919 ":enable_assembly",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004920 ":logging_utils",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004921 ":operator_run",
4922 ":operators",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004923 ":ukernels",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004924 "@clog",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004925 "@pthreadpool",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004926 ] + select({
4927 ":emscripten": [],
4928 "//conditions:default": ["@cpuinfo"],
4929 }),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004930)
4931
Marat Dukhancf056b22019-10-07 10:26:29 -07004932xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004933 name = "bench_utils",
4934 srcs = ["bench/utils.cc"],
4935 hdrs = ["bench/utils.h"],
Marat Dukhanbad48fe2019-11-04 10:35:22 -08004936 deps = [
4937 "@com_google_benchmark//:benchmark",
4938 "@cpuinfo",
4939 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004940)
4941
Frank Barchard7e955972019-10-11 10:34:25 -07004942######################### Benchmarks for micro-kernels #########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07004943
4944xnnpack_benchmark(
Marat Dukhan595e1702020-07-31 10:12:52 -07004945 name = "qs8_gemm_bench",
4946 srcs = [
4947 "bench/gemm.h",
4948 "bench/qs8-gemm.cc",
4949 "src/xnnpack/AlignedAllocator.h",
4950 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Frank Barchard31328cb2020-10-12 11:55:18 -07004951 copts = xnnpack_optional_ruy_copts() + xnnpack_optional_gemmlowp_copts(),
4952 deps = MICROKERNEL_BENCHMARK_DEPS + [":packing"] + xnnpack_optional_ruy_deps() + xnnpack_optional_gemmlowp_deps(),
Marat Dukhan595e1702020-07-31 10:12:52 -07004953)
4954
4955xnnpack_benchmark(
Marat Dukhan56bdd4a2020-08-03 19:47:04 -07004956 name = "qs8_requantization_bench",
4957 srcs = [
4958 "bench/qs8-requantization.cc",
Marat Dukhan56bdd4a2020-08-03 19:47:04 -07004959 "src/xnnpack/AlignedAllocator.h",
Frank Barchard04336c12020-10-22 16:48:55 -07004960 "src/xnnpack/requantization-stubs.h",
Marat Dukhan56bdd4a2020-08-03 19:47:04 -07004961 ] + MICROKERNEL_BENCHMARK_HDRS,
4962 deps = MICROKERNEL_BENCHMARK_DEPS,
4963)
4964
4965xnnpack_benchmark(
Marat Dukhan08b7a972020-07-14 18:17:29 -07004966 name = "qu8_gemm_bench",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004967 srcs = [
4968 "bench/gemm.h",
Marat Dukhan08b7a972020-07-14 18:17:29 -07004969 "bench/qu8-gemm.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004970 "src/xnnpack/AlignedAllocator.h",
4971 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004972 copts = xnnpack_optional_ruy_copts() + xnnpack_optional_gemmlowp_copts(),
Marat Dukhanab582382020-07-06 13:32:08 -07004973 deps = MICROKERNEL_BENCHMARK_DEPS + [":packing"] + xnnpack_optional_ruy_deps() + xnnpack_optional_gemmlowp_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004974)
4975
4976xnnpack_benchmark(
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07004977 name = "qu8_requantization_bench",
4978 srcs = [
4979 "bench/qu8-requantization.cc",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07004980 "src/xnnpack/AlignedAllocator.h",
Frank Barchard04336c12020-10-22 16:48:55 -07004981 "src/xnnpack/requantization-stubs.h",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07004982 ] + MICROKERNEL_BENCHMARK_HDRS,
4983 deps = MICROKERNEL_BENCHMARK_DEPS,
4984)
4985
4986xnnpack_benchmark(
Frank Barchard40d20fe2020-05-05 00:37:45 -07004987 name = "f16_igemm_bench",
4988 srcs = [
4989 "bench/f16-igemm.cc",
4990 "bench/conv.h",
4991 "bench/google/conv.h",
4992 "src/xnnpack/AlignedAllocator.h",
4993 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004994 deps = MICROKERNEL_BENCHMARK_DEPS + [
4995 ":indirection",
4996 ":packing",
4997 ],
Frank Barchard40d20fe2020-05-05 00:37:45 -07004998)
4999
5000xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005001 name = "f16_gemm_bench",
5002 srcs = [
5003 "bench/f16-gemm.cc",
5004 "bench/gemm.h",
5005 "src/xnnpack/AlignedAllocator.h",
5006 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005007 deps = MICROKERNEL_BENCHMARK_DEPS + [
5008 ":packing",
5009 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005010)
5011
5012xnnpack_benchmark(
Marat Dukhanbdb56f52020-02-05 21:42:49 -08005013 name = "f16_spmm_bench",
5014 srcs = [
5015 "bench/f16-spmm.cc",
Marat Dukhan1631e3e2020-12-06 19:29:31 -08005016 "bench/spmm.h",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08005017 "src/xnnpack/AlignedAllocator.h",
5018 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanbdb56f52020-02-05 21:42:49 -08005019 deps = MICROKERNEL_BENCHMARK_DEPS,
5020)
5021
5022xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005023 name = "f32_igemm_bench",
5024 srcs = [
5025 "bench/f32-igemm.cc",
5026 "bench/conv.h",
5027 "src/xnnpack/AlignedAllocator.h",
5028 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005029 deps = MICROKERNEL_BENCHMARK_DEPS + [
5030 ":indirection",
5031 ":packing",
5032 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005033)
5034
5035xnnpack_benchmark(
Frank Barchardc2010062020-07-19 10:04:12 -07005036 name = "f16_relu_bench",
5037 srcs = [
5038 "bench/f16-relu.cc",
5039 "src/xnnpack/AlignedAllocator.h",
5040 ] + MICROKERNEL_BENCHMARK_HDRS,
5041 deps = MICROKERNEL_BENCHMARK_DEPS,
5042)
5043
5044xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005045 name = "f32_conv_hwc_bench",
5046 srcs = [
5047 "bench/f32-conv-hwc.cc",
5048 "bench/dconv.h",
5049 "src/xnnpack/AlignedAllocator.h",
5050 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005051 deps = MICROKERNEL_BENCHMARK_DEPS + [
5052 ":packing",
5053 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005054)
5055
5056xnnpack_benchmark(
Marat Dukhan1f29b802020-05-15 23:46:39 -07005057 name = "f32_conv_hwc2chw_bench",
Erich Elsen563df5f2019-10-23 08:02:21 -07005058 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07005059 "bench/f32-conv-hwc2chw.cc",
Erich Elsen563df5f2019-10-23 08:02:21 -07005060 "bench/dconv.h",
5061 "src/xnnpack/AlignedAllocator.h",
5062 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005063 deps = MICROKERNEL_BENCHMARK_DEPS + [
5064 ":packing",
5065 ],
Erich Elsen563df5f2019-10-23 08:02:21 -07005066)
5067
5068xnnpack_benchmark(
Frank Barchard5a599a62020-06-04 20:12:44 -07005069 name = "f16_dwconv_bench",
5070 srcs = [
5071 "bench/f16-dwconv.cc",
5072 "bench/dwconv.h",
5073 "bench/google/dwconv.h",
5074 "src/xnnpack/AlignedAllocator.h",
5075 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005076 deps = MICROKERNEL_BENCHMARK_DEPS + [
5077 ":indirection",
5078 ":packing",
5079 ],
Frank Barchard5a599a62020-06-04 20:12:44 -07005080)
5081
5082xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005083 name = "f32_dwconv_bench",
5084 srcs = [
5085 "bench/f32-dwconv.cc",
5086 "bench/dwconv.h",
5087 "src/xnnpack/AlignedAllocator.h",
5088 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005089 deps = MICROKERNEL_BENCHMARK_DEPS + [
5090 ":indirection",
5091 ":packing",
5092 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005093)
5094
5095xnnpack_benchmark(
Marat Dukhanbf715f92020-10-23 20:17:00 -07005096 name = "f32_dwconv2d_chw_bench",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005097 srcs = [
Marat Dukhanbf715f92020-10-23 20:17:00 -07005098 "bench/f32-dwconv2d-chw.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005099 "bench/dwconv.h",
5100 "src/xnnpack/AlignedAllocator.h",
5101 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005102 deps = MICROKERNEL_BENCHMARK_DEPS + [
5103 ":indirection",
5104 ":packing",
5105 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005106)
5107
5108xnnpack_benchmark(
5109 name = "f32_gemm_bench",
5110 srcs = [
5111 "bench/f32-gemm.cc",
5112 "bench/gemm.h",
5113 "src/xnnpack/AlignedAllocator.h",
5114 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07005115 copts = xnnpack_optional_ruy_copts(),
Marat Dukhanab582382020-07-06 13:32:08 -07005116 deps = MICROKERNEL_BENCHMARK_DEPS + [":packing"] + xnnpack_optional_ruy_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07005117)
5118
5119xnnpack_benchmark(
Marat Dukhana11ca342020-06-25 23:45:07 -07005120 name = "f32_hswish_bench",
5121 srcs = [
5122 "bench/f32-hswish.cc",
5123 "src/xnnpack/AlignedAllocator.h",
5124 ] + MICROKERNEL_BENCHMARK_HDRS,
5125 deps = MICROKERNEL_BENCHMARK_DEPS,
5126)
5127
5128xnnpack_benchmark(
Marat Dukhan4c4eb002019-12-08 21:27:49 -08005129 name = "f32_raddexpminusmax_bench",
5130 srcs = [
5131 "bench/f32-raddexpminusmax.cc",
5132 "src/xnnpack/AlignedAllocator.h",
5133 ] + MICROKERNEL_BENCHMARK_HDRS,
5134 deps = MICROKERNEL_BENCHMARK_DEPS,
5135)
5136
5137xnnpack_benchmark(
5138 name = "f32_raddextexp_bench",
5139 srcs = [
5140 "bench/f32-raddextexp.cc",
5141 "src/xnnpack/AlignedAllocator.h",
5142 ] + MICROKERNEL_BENCHMARK_HDRS,
5143 deps = MICROKERNEL_BENCHMARK_DEPS,
5144)
5145
5146xnnpack_benchmark(
5147 name = "f32_raddstoreexpminusmax_bench",
5148 srcs = [
5149 "bench/f32-raddstoreexpminusmax.cc",
5150 "src/xnnpack/AlignedAllocator.h",
5151 ] + MICROKERNEL_BENCHMARK_HDRS,
5152 deps = MICROKERNEL_BENCHMARK_DEPS,
5153)
5154
5155xnnpack_benchmark(
Frank Barchard3c4a9522020-07-18 16:11:22 -07005156 name = "f32_relu_bench",
5157 srcs = [
5158 "bench/f32-relu.cc",
5159 "src/xnnpack/AlignedAllocator.h",
5160 ] + MICROKERNEL_BENCHMARK_HDRS,
5161 deps = MICROKERNEL_BENCHMARK_DEPS,
5162)
5163
5164xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005165 name = "f32_rmax_bench",
5166 srcs = [
5167 "bench/f32-rmax.cc",
5168 "src/xnnpack/AlignedAllocator.h",
5169 ] + MICROKERNEL_BENCHMARK_HDRS,
5170 deps = MICROKERNEL_BENCHMARK_DEPS,
5171)
5172
5173xnnpack_benchmark(
Marat Dukhan14bec502019-11-18 11:35:31 -08005174 name = "f32_sigmoid_bench",
5175 srcs = [
5176 "bench/f32-sigmoid.cc",
5177 "src/xnnpack/AlignedAllocator.h",
5178 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan14bec502019-11-18 11:35:31 -08005179 deps = MICROKERNEL_BENCHMARK_DEPS,
5180)
5181
5182xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005183 name = "f32_spmm_bench",
5184 srcs = [
5185 "bench/f32-spmm.cc",
Marat Dukhan1631e3e2020-12-06 19:29:31 -08005186 "bench/spmm.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005187 "src/xnnpack/AlignedAllocator.h",
5188 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005189 deps = MICROKERNEL_BENCHMARK_DEPS,
5190)
5191
5192xnnpack_benchmark(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08005193 name = "f32_softmax_bench",
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07005194 srcs = [
Marat Dukhanfd8e6892020-01-27 15:25:25 -08005195 "bench/f32-softmax.cc",
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07005196 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07005197 copts = xnnpack_optional_dnnl_copts(),
Marat Dukhan8d3c6932020-03-06 20:27:27 -08005198 deps = MICROKERNEL_BENCHMARK_DEPS + xnnpack_optional_dnnl_deps(),
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07005199)
5200
5201xnnpack_benchmark(
Marat Dukhaned6baaf2020-12-01 15:07:08 -08005202 name = "f32_velu_bench",
5203 srcs = [
5204 "bench/f32-velu.cc",
5205 "src/xnnpack/AlignedAllocator.h",
5206 ] + MICROKERNEL_BENCHMARK_HDRS,
5207 deps = MICROKERNEL_BENCHMARK_DEPS,
5208)
5209
5210xnnpack_benchmark(
Marat Dukhan4c4eb002019-12-08 21:27:49 -08005211 name = "f32_vscaleexpminusmax_bench",
5212 srcs = [
5213 "bench/f32-vscaleexpminusmax.cc",
5214 "src/xnnpack/AlignedAllocator.h",
5215 ] + MICROKERNEL_BENCHMARK_HDRS,
5216 deps = MICROKERNEL_BENCHMARK_DEPS,
5217)
5218
5219xnnpack_benchmark(
5220 name = "f32_vscaleextexp_bench",
5221 srcs = [
5222 "bench/f32-vscaleextexp.cc",
5223 "src/xnnpack/AlignedAllocator.h",
5224 ] + MICROKERNEL_BENCHMARK_HDRS,
5225 deps = MICROKERNEL_BENCHMARK_DEPS,
5226)
5227
5228xnnpack_benchmark(
Marat Dukhanf4db2f32020-06-30 10:55:30 -07005229 name = "f32_vsqrt_bench",
5230 srcs = [
5231 "bench/f32-vsqrt.cc",
5232 "src/xnnpack/AlignedAllocator.h",
5233 ] + MICROKERNEL_BENCHMARK_HDRS,
5234 deps = MICROKERNEL_BENCHMARK_DEPS,
5235)
5236
5237xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005238 name = "f32_im2col_gemm_bench",
5239 srcs = [
5240 "bench/f32-im2col-gemm.cc",
5241 "bench/conv.h",
5242 "src/xnnpack/AlignedAllocator.h",
5243 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005244 deps = MICROKERNEL_BENCHMARK_DEPS + [
5245 ":im2col",
5246 ":packing",
5247 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005248)
5249
Marat Dukhanfe7acb62020-03-09 19:30:05 -07005250xnnpack_benchmark(
Marat Dukhanffbf96a2020-05-14 02:59:08 -07005251 name = "rounding_bench",
5252 srcs = [
5253 "bench/rounding.cc",
Marat Dukhanffbf96a2020-05-14 02:59:08 -07005254 "src/xnnpack/AlignedAllocator.h",
Frank Barchard04336c12020-10-22 16:48:55 -07005255 "src/xnnpack/math-stubs.h",
Marat Dukhanffbf96a2020-05-14 02:59:08 -07005256 ] + MICROKERNEL_BENCHMARK_HDRS,
5257 deps = MICROKERNEL_BENCHMARK_DEPS,
5258)
5259
Marat Dukhan08c4a432019-10-03 09:29:21 -07005260########################### Benchmarks for operators ###########################
5261
5262xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005263 name = "average_pooling_bench",
5264 srcs = ["bench/average-pooling.cc"],
Marat Dukhan7a16d8b2020-03-11 04:22:44 -07005265 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07005266 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07005267 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07005268)
5269
5270xnnpack_benchmark(
Marat Dukhan87727142020-06-24 15:24:10 -07005271 name = "bankers_rounding_bench",
5272 srcs = ["bench/bankers-rounding.cc"],
5273 copts = xnnpack_optional_tflite_copts(),
5274 tags = ["nowin32"],
5275 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
5276)
5277
5278xnnpack_benchmark(
5279 name = "ceiling_bench",
5280 srcs = ["bench/ceiling.cc"],
5281 copts = xnnpack_optional_tflite_copts(),
5282 tags = ["nowin32"],
5283 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
5284)
5285
5286xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005287 name = "channel_shuffle_bench",
5288 srcs = ["bench/channel-shuffle.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07005289 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005290)
5291
5292xnnpack_benchmark(
5293 name = "convolution_bench",
5294 srcs = ["bench/convolution.cc"],
5295 copts = xnnpack_optional_tflite_copts() + xnnpack_optional_armcl_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07005296 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07005297 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps() + xnnpack_optional_armcl_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07005298)
5299
5300xnnpack_benchmark(
5301 name = "deconvolution_bench",
5302 srcs = ["bench/deconvolution.cc"],
5303 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07005304 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07005305 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07005306)
5307
5308xnnpack_benchmark(
Marat Dukhanb6bd4bc2020-12-01 17:01:40 -08005309 name = "elu_bench",
5310 srcs = ["bench/elu.cc"],
5311 copts = xnnpack_optional_tflite_copts(),
5312 tags = ["nowin32"],
5313 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
5314)
5315
5316xnnpack_benchmark(
Marat Dukhan87727142020-06-24 15:24:10 -07005317 name = "floor_bench",
5318 srcs = ["bench/floor.cc"],
5319 copts = xnnpack_optional_tflite_copts(),
5320 tags = ["nowin32"],
5321 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
5322)
5323
5324xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005325 name = "global_average_pooling_bench",
5326 srcs = ["bench/global-average-pooling.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07005327 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005328)
5329
5330xnnpack_benchmark(
Marat Dukhanad352602020-06-25 21:50:54 -07005331 name = "hardswish_bench",
5332 srcs = ["bench/hardswish.cc"],
5333 copts = xnnpack_optional_tflite_copts(),
5334 tags = ["nowin32"],
5335 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
5336)
5337
5338xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005339 name = "max_pooling_bench",
5340 srcs = ["bench/max-pooling.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07005341 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005342)
5343
5344xnnpack_benchmark(
5345 name = "sigmoid_bench",
5346 srcs = ["bench/sigmoid.cc"],
Marat Dukhanc3b9e862019-11-17 13:18:54 -08005347 copts = xnnpack_optional_tflite_copts(),
Marat Dukhanca2ba702020-04-24 01:31:47 -07005348 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07005349 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07005350)
5351
5352xnnpack_benchmark(
Marat Dukhan95b22432019-10-30 16:30:14 -07005353 name = "prelu_bench",
5354 srcs = ["bench/prelu.cc"],
5355 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07005356 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07005357 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan95b22432019-10-30 16:30:14 -07005358)
5359
5360xnnpack_benchmark(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08005361 name = "softmax_bench",
5362 srcs = ["bench/softmax.cc"],
Marat Dukhan9c0db962020-01-28 12:30:14 -08005363 copts = xnnpack_optional_tflite_copts(),
Marat Dukhanca2ba702020-04-24 01:31:47 -07005364 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07005365 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07005366)
5367
Marat Dukhan87727142020-06-24 15:24:10 -07005368xnnpack_benchmark(
Marat Dukhan6804bbd2020-06-30 19:26:11 -07005369 name = "square_root_bench",
5370 srcs = ["bench/square-root.cc"],
5371 copts = xnnpack_optional_tflite_copts(),
5372 tags = ["nowin32"],
5373 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
5374)
5375
5376xnnpack_benchmark(
Marat Dukhan87727142020-06-24 15:24:10 -07005377 name = "truncation_bench",
5378 srcs = ["bench/truncation.cc"],
5379 deps = OPERATOR_BENCHMARK_DEPS,
5380)
5381
Marat Dukhanc068bb62019-10-04 13:24:39 -07005382############################# End-to-end benchmarks ############################
5383
5384cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07005385 name = "fp32_mobilenet_v1",
5386 srcs = ["models/fp32-mobilenet-v1.cc"],
Marat Dukhanc068bb62019-10-04 13:24:39 -07005387 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08005388 copts = xnnpack_std_cxxopts(),
Marat Dukhanc068bb62019-10-04 13:24:39 -07005389 linkstatic = True,
5390 deps = [
5391 ":XNNPACK",
5392 "@pthreadpool",
5393 ],
5394)
5395
5396cc_library(
Marat Dukhan4cea2322021-03-09 09:35:36 -08005397 name = "fp32_sparse_mobilenet_v1",
5398 srcs = ["models/fp32-sparse-mobilenet-v1.cc"],
5399 hdrs = ["models/models.h"],
5400 copts = xnnpack_std_cxxopts(),
5401 linkstatic = True,
5402 deps = [
5403 ":XNNPACK",
5404 "@pthreadpool",
5405 ],
5406)
5407
5408cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07005409 name = "fp16_mobilenet_v1",
5410 srcs = ["models/fp16-mobilenet-v1.cc"],
5411 hdrs = ["models/models.h"],
5412 copts = xnnpack_std_cxxopts(),
5413 linkstatic = True,
5414 deps = [
5415 ":XNNPACK",
5416 "@FP16",
5417 "@pthreadpool",
5418 ],
5419)
5420
5421cc_library(
Marat Dukhan0743cdf2020-08-04 18:52:07 -07005422 name = "qs8_mobilenet_v1",
5423 srcs = ["models/qs8-mobilenet-v1.cc"],
5424 hdrs = ["models/models.h"],
5425 copts = xnnpack_std_cxxopts(),
5426 linkstatic = True,
5427 deps = [
5428 ":XNNPACK",
5429 "@pthreadpool",
5430 ],
5431)
5432
5433cc_library(
Marat Dukhan70a96182020-09-03 17:13:58 -07005434 name = "qs8_mobilenet_v2",
5435 srcs = ["models/qs8-mobilenet-v2.cc"],
5436 hdrs = ["models/models.h"],
5437 copts = xnnpack_std_cxxopts(),
5438 linkstatic = True,
5439 deps = [
5440 ":XNNPACK",
5441 "@pthreadpool",
5442 ],
5443)
5444
5445cc_library(
Marat Dukhan12a23bb2021-03-08 08:13:21 -08005446 name = "qu8_mobilenet_v1",
5447 srcs = ["models/qu8-mobilenet-v1.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 Dukhan270a2c42020-06-26 16:45:52 -07005458 name = "fp32_mobilenet_v2",
5459 srcs = ["models/fp32-mobilenet-v2.cc"],
Marat Dukhanc068bb62019-10-04 13:24:39 -07005460 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08005461 copts = xnnpack_std_cxxopts(),
Marat Dukhanc068bb62019-10-04 13:24:39 -07005462 linkstatic = True,
5463 deps = [
5464 ":XNNPACK",
5465 "@pthreadpool",
5466 ],
5467)
5468
Marat Dukhanc08cdf52019-12-09 09:17:51 -08005469cc_library(
Marat Dukhan4cea2322021-03-09 09:35:36 -08005470 name = "fp32_sparse_mobilenet_v2",
5471 srcs = ["models/fp32-sparse-mobilenet-v2.cc"],
5472 hdrs = ["models/models.h"],
5473 copts = xnnpack_std_cxxopts(),
5474 linkstatic = True,
5475 deps = [
5476 ":XNNPACK",
5477 "@pthreadpool",
5478 ],
5479)
5480
5481cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07005482 name = "fp16_mobilenet_v2",
5483 srcs = ["models/fp16-mobilenet-v2.cc"],
5484 hdrs = ["models/models.h"],
5485 copts = xnnpack_std_cxxopts(),
5486 linkstatic = True,
5487 deps = [
5488 ":XNNPACK",
5489 "@FP16",
5490 "@pthreadpool",
5491 ],
5492)
5493
5494cc_library(
5495 name = "fp32_mobilenet_v3_large",
5496 srcs = ["models/fp32-mobilenet-v3-large.cc"],
Marat Dukhanc08cdf52019-12-09 09:17:51 -08005497 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08005498 copts = xnnpack_std_cxxopts(),
Marat Dukhanc08cdf52019-12-09 09:17:51 -08005499 linkstatic = True,
5500 deps = [
5501 ":XNNPACK",
5502 "@pthreadpool",
5503 ],
5504)
5505
5506cc_library(
Marat Dukhan4cea2322021-03-09 09:35:36 -08005507 name = "fp32_sparse_mobilenet_v3_large",
5508 srcs = ["models/fp32-sparse-mobilenet-v3-large.cc"],
5509 hdrs = ["models/models.h"],
5510 copts = xnnpack_std_cxxopts(),
5511 linkstatic = True,
5512 deps = [
5513 ":XNNPACK",
5514 "@pthreadpool",
5515 ],
5516)
5517
5518cc_library(
Marat Dukhan66f3ccd2020-09-14 16:09:38 -07005519 name = "fp16_mobilenet_v3_large",
5520 srcs = ["models/fp16-mobilenet-v3-large.cc"],
5521 hdrs = ["models/models.h"],
5522 copts = xnnpack_std_cxxopts(),
5523 linkstatic = True,
5524 deps = [
5525 ":XNNPACK",
5526 "@FP16",
5527 "@pthreadpool",
5528 ],
5529)
5530
5531cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07005532 name = "fp32_mobilenet_v3_small",
5533 srcs = ["models/fp32-mobilenet-v3-small.cc"],
Marat Dukhanc08cdf52019-12-09 09:17:51 -08005534 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08005535 copts = xnnpack_std_cxxopts(),
Marat Dukhanc08cdf52019-12-09 09:17:51 -08005536 linkstatic = True,
5537 deps = [
5538 ":XNNPACK",
5539 "@pthreadpool",
5540 ],
5541)
5542
Marat Dukhan66f3ccd2020-09-14 16:09:38 -07005543cc_library(
Marat Dukhan4cea2322021-03-09 09:35:36 -08005544 name = "fp32_sparse_mobilenet_v3_small",
5545 srcs = ["models/fp32-sparse-mobilenet-v3-small.cc"],
5546 hdrs = ["models/models.h"],
5547 copts = xnnpack_std_cxxopts(),
5548 linkstatic = True,
5549 deps = [
5550 ":XNNPACK",
5551 "@pthreadpool",
5552 ],
5553)
5554
5555cc_library(
Marat Dukhan66f3ccd2020-09-14 16:09:38 -07005556 name = "fp16_mobilenet_v3_small",
5557 srcs = ["models/fp16-mobilenet-v3-small.cc"],
5558 hdrs = ["models/models.h"],
5559 copts = xnnpack_std_cxxopts(),
5560 linkstatic = True,
5561 deps = [
5562 ":XNNPACK",
5563 "@FP16",
5564 "@pthreadpool",
5565 ],
5566)
5567
Marat Dukhanc068bb62019-10-04 13:24:39 -07005568xnnpack_benchmark(
Marat Dukhanef4416e2019-10-31 13:44:40 -07005569 name = "f32_dwconv_e2e_bench",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08005570 srcs = [
5571 "bench/f32-dwconv-e2e.cc",
5572 "bench/end2end.h",
5573 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanef4416e2019-10-31 13:44:40 -07005574 deps = MICROKERNEL_BENCHMARK_DEPS + [
5575 ":XNNPACK",
Marat Dukhan270a2c42020-06-26 16:45:52 -07005576 ":fp32_mobilenet_v1",
5577 ":fp32_mobilenet_v2",
5578 ":fp32_mobilenet_v3_large",
5579 ":fp32_mobilenet_v3_small",
Marat Dukhanef4416e2019-10-31 13:44:40 -07005580 ],
5581)
5582
5583xnnpack_benchmark(
Marat Dukhan5f18d262019-10-31 10:24:14 -07005584 name = "f32_gemm_e2e_bench",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08005585 srcs = [
5586 "bench/f32-gemm-e2e.cc",
5587 "bench/end2end.h",
5588 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan5f18d262019-10-31 10:24:14 -07005589 deps = MICROKERNEL_BENCHMARK_DEPS + [
5590 ":XNNPACK",
Marat Dukhan270a2c42020-06-26 16:45:52 -07005591 ":fp32_mobilenet_v1",
5592 ":fp32_mobilenet_v2",
5593 ":fp32_mobilenet_v3_large",
5594 ":fp32_mobilenet_v3_small",
Marat Dukhan5f18d262019-10-31 10:24:14 -07005595 ],
5596)
5597
5598xnnpack_benchmark(
Frank Barcharddc909cb2021-02-08 13:59:31 -08005599 name = "qs8_gemm_e2e_bench",
5600 srcs = [
5601 "bench/qs8-gemm-e2e.cc",
5602 "bench/end2end.h",
5603 ] + MICROKERNEL_BENCHMARK_HDRS,
5604 deps = MICROKERNEL_BENCHMARK_DEPS + [
5605 ":XNNPACK",
5606 ":qs8_mobilenet_v1",
5607 ":qs8_mobilenet_v2",
5608 ],
5609)
5610
5611xnnpack_benchmark(
Marat Dukhanc068bb62019-10-04 13:24:39 -07005612 name = "end2end_bench",
5613 srcs = ["bench/end2end.cc"],
5614 deps = [
5615 ":XNNPACK",
Frank Barchardc712fa42019-10-31 14:00:21 -07005616 ":bench_utils",
Marat Dukhan270a2c42020-06-26 16:45:52 -07005617 ":fp16_mobilenet_v1",
5618 ":fp16_mobilenet_v2",
Marat Dukhan66f3ccd2020-09-14 16:09:38 -07005619 ":fp16_mobilenet_v3_large",
5620 ":fp16_mobilenet_v3_small",
Marat Dukhan270a2c42020-06-26 16:45:52 -07005621 ":fp32_mobilenet_v1",
5622 ":fp32_mobilenet_v2",
5623 ":fp32_mobilenet_v3_large",
5624 ":fp32_mobilenet_v3_small",
Marat Dukhan4cea2322021-03-09 09:35:36 -08005625 ":fp32_sparse_mobilenet_v1",
5626 ":fp32_sparse_mobilenet_v2",
5627 ":fp32_sparse_mobilenet_v3_large",
5628 ":fp32_sparse_mobilenet_v3_small",
Marat Dukhan0743cdf2020-08-04 18:52:07 -07005629 ":qs8_mobilenet_v1",
Marat Dukhan70a96182020-09-03 17:13:58 -07005630 ":qs8_mobilenet_v2",
Marat Dukhan12a23bb2021-03-08 08:13:21 -08005631 ":qu8_mobilenet_v1",
Marat Dukhanc068bb62019-10-04 13:24:39 -07005632 "@pthreadpool",
5633 ],
5634)
5635
Marat Dukhan6adff4e2019-10-14 18:32:07 -07005636#################### Accuracy evaluation for math functions ####################
5637
5638xnnpack_benchmark(
Marat Dukhancf67ec62020-12-14 09:56:37 -08005639 name = "f32_exp_ulp_eval",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07005640 srcs = [
Marat Dukhancf67ec62020-12-14 09:56:37 -08005641 "eval/f32-exp-ulp.cc",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07005642 "src/xnnpack/AlignedAllocator.h",
5643 ] + ACCURACY_EVAL_HDRS,
Marat Dukhan3a305212020-12-06 19:24:27 -08005644 deps = ACCURACY_EVAL_DEPS + [
5645 ":bench_utils",
5646 "@cpuinfo",
5647 ],
Marat Dukhan6adff4e2019-10-14 18:32:07 -07005648)
5649
Marat Dukhan515c9772019-10-17 18:07:57 -07005650xnnpack_benchmark(
Marat Dukhancf67ec62020-12-14 09:56:37 -08005651 name = "f32_expminus_ulp_eval",
Marat Dukhan515c9772019-10-17 18:07:57 -07005652 srcs = [
Marat Dukhancf67ec62020-12-14 09:56:37 -08005653 "eval/f32-expminus-ulp.cc",
Marat Dukhan515c9772019-10-17 18:07:57 -07005654 "src/xnnpack/AlignedAllocator.h",
5655 ] + ACCURACY_EVAL_HDRS,
Marat Dukhan3a305212020-12-06 19:24:27 -08005656 deps = ACCURACY_EVAL_DEPS + [
5657 ":bench_utils",
5658 "@cpuinfo",
5659 ],
Marat Dukhan515c9772019-10-17 18:07:57 -07005660)
5661
Marat Dukhan98ba4412019-10-23 02:14:28 -07005662xnnpack_benchmark(
Marat Dukhancf67ec62020-12-14 09:56:37 -08005663 name = "f32_expm1minus_ulp_eval",
Marat Dukhana438aca2020-11-20 15:45:01 -08005664 srcs = [
Marat Dukhancf67ec62020-12-14 09:56:37 -08005665 "eval/f32-expm1minus-ulp.cc",
Marat Dukhana438aca2020-11-20 15:45:01 -08005666 "src/xnnpack/AlignedAllocator.h",
5667 ] + ACCURACY_EVAL_HDRS,
Marat Dukhande390d42020-11-29 19:32:18 -08005668 deps = ACCURACY_EVAL_DEPS + [
5669 ":bench_utils",
Marat Dukhan3a305212020-12-06 19:24:27 -08005670 "@cpuinfo",
Marat Dukhande390d42020-11-29 19:32:18 -08005671 ],
Marat Dukhana438aca2020-11-20 15:45:01 -08005672)
5673
5674xnnpack_benchmark(
Marat Dukhancf67ec62020-12-14 09:56:37 -08005675 name = "f32_extexp_ulp_eval",
Marat Dukhan98ba4412019-10-23 02:14:28 -07005676 srcs = [
Marat Dukhancf67ec62020-12-14 09:56:37 -08005677 "eval/f32-extexp-ulp.cc",
Marat Dukhan98ba4412019-10-23 02:14:28 -07005678 "src/xnnpack/AlignedAllocator.h",
5679 ] + ACCURACY_EVAL_HDRS,
Marat Dukhan3a305212020-12-06 19:24:27 -08005680 deps = ACCURACY_EVAL_DEPS + [
5681 ":bench_utils",
5682 "@cpuinfo",
5683 ],
Marat Dukhan98ba4412019-10-23 02:14:28 -07005684)
5685
Marat Dukhanf44f0222020-12-14 11:53:27 -08005686xnnpack_benchmark(
5687 name = "f32_sigmoid_ulp_eval",
5688 srcs = [
5689 "eval/f32-sigmoid-ulp.cc",
5690 "src/xnnpack/AlignedAllocator.h",
5691 ] + ACCURACY_EVAL_HDRS,
5692 deps = ACCURACY_EVAL_DEPS + [
5693 ":bench_utils",
5694 "@cpuinfo",
5695 ],
5696)
5697
5698xnnpack_benchmark(
5699 name = "f32_sqrt_ulp_eval",
5700 srcs = [
5701 "eval/f32-sqrt-ulp.cc",
5702 "src/xnnpack/AlignedAllocator.h",
5703 ] + ACCURACY_EVAL_HDRS,
5704 deps = ACCURACY_EVAL_DEPS + [
5705 ":bench_utils",
5706 "@cpuinfo",
5707 ],
5708)
5709
5710################### Accuracy verification for math functions ##################
5711
5712xnnpack_unit_test(
Marat Dukhanf7291fc2020-12-15 11:02:50 -08005713 name = "f32_exp_eval",
5714 srcs = [
5715 "eval/f32-exp.cc",
5716 "src/xnnpack/AlignedAllocator.h",
5717 "src/xnnpack/math-stubs.h",
5718 ] + MICROKERNEL_TEST_HDRS,
5719 automatic = False,
5720 deps = MICROKERNEL_TEST_DEPS,
5721)
5722
5723xnnpack_unit_test(
Marat Dukhanf44f0222020-12-14 11:53:27 -08005724 name = "f32_expm1minus_eval",
5725 srcs = [
5726 "eval/f32-expm1minus.cc",
5727 "src/xnnpack/AlignedAllocator.h",
5728 "src/xnnpack/math-stubs.h",
5729 ] + MICROKERNEL_TEST_HDRS,
5730 automatic = False,
5731 deps = MICROKERNEL_TEST_DEPS,
5732)
5733
Marat Dukhan8853b822020-05-07 12:19:01 -07005734xnnpack_unit_test(
Marat Dukhand28a5a22020-12-14 15:27:22 -08005735 name = "f32_expminus_eval",
5736 srcs = [
5737 "eval/f32-expminus.cc",
5738 "src/xnnpack/AlignedAllocator.h",
5739 "src/xnnpack/math-stubs.h",
5740 ] + MICROKERNEL_TEST_HDRS,
5741 automatic = False,
5742 deps = MICROKERNEL_TEST_DEPS,
5743)
5744
5745xnnpack_unit_test(
Marat Dukhan8853b822020-05-07 12:19:01 -07005746 name = "f32_roundne_eval",
5747 srcs = [
5748 "eval/f32-roundne.cc",
5749 "src/xnnpack/AlignedAllocator.h",
5750 "src/xnnpack/math-stubs.h",
5751 ] + MICROKERNEL_TEST_HDRS,
Marat Dukhan22eed3d2020-05-11 20:13:37 -07005752 automatic = False,
Marat Dukhan8853b822020-05-07 12:19:01 -07005753 deps = MICROKERNEL_TEST_DEPS,
5754)
5755
Marat Dukhan2dbb9442020-05-12 20:43:43 -07005756xnnpack_unit_test(
Marat Dukhanc9852ba2020-05-13 17:21:29 -07005757 name = "f32_roundd_eval",
5758 srcs = [
5759 "eval/f32-roundd.cc",
5760 "src/xnnpack/AlignedAllocator.h",
5761 "src/xnnpack/math-stubs.h",
5762 ] + MICROKERNEL_TEST_HDRS,
5763 automatic = False,
5764 deps = MICROKERNEL_TEST_DEPS,
5765)
5766
5767xnnpack_unit_test(
5768 name = "f32_roundu_eval",
5769 srcs = [
5770 "eval/f32-roundu.cc",
5771 "src/xnnpack/AlignedAllocator.h",
5772 "src/xnnpack/math-stubs.h",
5773 ] + MICROKERNEL_TEST_HDRS,
5774 automatic = False,
5775 deps = MICROKERNEL_TEST_DEPS,
5776)
5777
5778xnnpack_unit_test(
Marat Dukhan2dbb9442020-05-12 20:43:43 -07005779 name = "f32_roundz_eval",
5780 srcs = [
5781 "eval/f32-roundz.cc",
5782 "src/xnnpack/AlignedAllocator.h",
5783 "src/xnnpack/math-stubs.h",
5784 ] + MICROKERNEL_TEST_HDRS,
Marat Dukhanc9852ba2020-05-13 17:21:29 -07005785 automatic = False,
Marat Dukhan2dbb9442020-05-12 20:43:43 -07005786 deps = MICROKERNEL_TEST_DEPS,
5787)
5788
Marat Dukhan08c4a432019-10-03 09:29:21 -07005789######################### Unit tests for micro-kernels #########################
5790
5791xnnpack_unit_test(
Marat Dukhande06f492020-04-09 00:19:31 -07005792 name = "f16_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005793 srcs = [
Marat Dukhande06f492020-04-09 00:19:31 -07005794 "test/f16-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005795 "test/gemm-microkernel-tester.h",
5796 "src/xnnpack/AlignedAllocator.h",
5797 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005798 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005799)
5800
5801xnnpack_unit_test(
Marat Dukhan355ab432020-04-09 19:01:52 -07005802 name = "f16_spmm_minmax_test",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08005803 srcs = [
Marat Dukhan355ab432020-04-09 19:01:52 -07005804 "test/f16-spmm-minmax.cc",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08005805 "test/spmm-microkernel-tester.h",
5806 "src/xnnpack/AlignedAllocator.h",
5807 ] + MICROKERNEL_TEST_HDRS,
5808 deps = MICROKERNEL_TEST_DEPS,
5809)
5810
5811xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005812 name = "f32_argmaxpool_test",
5813 srcs = [
5814 "test/f32-argmaxpool.cc",
5815 "test/argmaxpool-microkernel-tester.h",
5816 "src/xnnpack/AlignedAllocator.h",
5817 ] + MICROKERNEL_TEST_HDRS,
5818 deps = MICROKERNEL_TEST_DEPS,
5819)
5820
5821xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07005822 name = "f32_avgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005823 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07005824 "test/f32-avgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005825 "test/avgpool-microkernel-tester.h",
5826 "src/xnnpack/AlignedAllocator.h",
5827 ] + MICROKERNEL_TEST_HDRS,
5828 deps = MICROKERNEL_TEST_DEPS,
5829)
5830
5831xnnpack_unit_test(
Marat Dukhan660fd192020-03-10 04:55:30 -07005832 name = "f32_ibilinear_test",
Marat Dukhan35dacfb2019-11-07 19:18:16 -08005833 srcs = [
Marat Dukhan660fd192020-03-10 04:55:30 -07005834 "test/f32-ibilinear.cc",
5835 "test/ibilinear-microkernel-tester.h",
Marat Dukhan35dacfb2019-11-07 19:18:16 -08005836 "src/xnnpack/AlignedAllocator.h",
5837 ] + MICROKERNEL_TEST_HDRS,
5838 deps = MICROKERNEL_TEST_DEPS,
5839)
5840
5841xnnpack_unit_test(
XNNPACK Team21432672020-10-19 19:58:48 -07005842 name = "f32_ibilinear_chw_test",
5843 srcs = [
5844 "test/f32-ibilinear-chw.cc",
XNNPACK Team6be46b22020-10-22 23:34:54 -07005845 "test/ibilinear-microkernel-tester.h",
XNNPACK Team21432672020-10-19 19:58:48 -07005846 "src/xnnpack/AlignedAllocator.h",
5847 ] + MICROKERNEL_TEST_HDRS,
5848 deps = MICROKERNEL_TEST_DEPS,
5849)
5850
5851xnnpack_unit_test(
Frank Barchardb1966592020-05-12 13:47:06 -07005852 name = "f16_clamp_test",
5853 srcs = [
5854 "test/f16-clamp.cc",
5855 "test/clamp-microkernel-tester.h",
5856 ] + MICROKERNEL_TEST_HDRS,
5857 deps = MICROKERNEL_TEST_DEPS,
5858)
5859
5860xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005861 name = "f32_clamp_test",
5862 srcs = [
5863 "test/f32-clamp.cc",
5864 "test/clamp-microkernel-tester.h",
5865 ] + MICROKERNEL_TEST_HDRS,
5866 deps = MICROKERNEL_TEST_DEPS,
5867)
5868
5869xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07005870 name = "f32_igemm_test",
5871 srcs = [
5872 "test/f32-igemm.cc",
5873 "test/gemm-microkernel-tester.h",
5874 "src/xnnpack/AlignedAllocator.h",
5875 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005876 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan163a7e62020-04-09 04:19:26 -07005877)
5878
5879xnnpack_unit_test(
Marat Dukhan467f6362020-05-22 23:21:55 -07005880 name = "f32_igemm_relu_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005881 srcs = [
Marat Dukhan467f6362020-05-22 23:21:55 -07005882 "test/f32-igemm-relu.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005883 "test/gemm-microkernel-tester.h",
5884 "src/xnnpack/AlignedAllocator.h",
5885 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005886 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005887)
5888
5889xnnpack_unit_test(
Marat Dukhane207b7b2020-05-28 16:27:42 -07005890 name = "f32_igemm_minmax_test",
5891 srcs = [
5892 "test/f32-igemm-minmax.cc",
5893 "test/gemm-microkernel-tester.h",
5894 "src/xnnpack/AlignedAllocator.h",
5895 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005896 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhane207b7b2020-05-28 16:27:42 -07005897)
5898
5899xnnpack_unit_test(
Frank Barchardb0e4fae2020-05-04 15:27:51 -07005900 name = "f16_igemm_minmax_test",
5901 srcs = [
5902 "test/f16-igemm-minmax.cc",
5903 "test/gemm-microkernel-tester.h",
5904 "src/xnnpack/AlignedAllocator.h",
5905 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005906 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Frank Barchardb0e4fae2020-05-04 15:27:51 -07005907)
5908
5909xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005910 name = "f32_conv_hwc_test",
5911 srcs = [
5912 "test/f32-conv-hwc.cc",
5913 "test/conv-hwc-microkernel-tester.h",
5914 "src/xnnpack/AlignedAllocator.h",
5915 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005916 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005917)
5918
5919xnnpack_unit_test(
Marat Dukhan1f29b802020-05-15 23:46:39 -07005920 name = "f32_conv_hwc2chw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005921 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07005922 "test/f32-conv-hwc2chw.cc",
5923 "test/conv-hwc2chw-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005924 "src/xnnpack/AlignedAllocator.h",
5925 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005926 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005927)
5928
5929xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07005930 name = "f32_dwconv_test",
5931 srcs = [
5932 "test/f32-dwconv.cc",
5933 "test/dwconv-microkernel-tester.h",
5934 "src/xnnpack/AlignedAllocator.h",
5935 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005936 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan163a7e62020-04-09 04:19:26 -07005937)
5938
5939xnnpack_unit_test(
Frank Barchard5a599a62020-06-04 20:12:44 -07005940 name = "f16_dwconv_minmax_test",
5941 srcs = [
5942 "test/f16-dwconv-minmax.cc",
5943 "test/dwconv-microkernel-tester.h",
5944 "src/xnnpack/AlignedAllocator.h",
5945 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005946 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Frank Barchard5a599a62020-06-04 20:12:44 -07005947)
5948
5949xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07005950 name = "f32_dwconv_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005951 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07005952 "test/f32-dwconv-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005953 "test/dwconv-microkernel-tester.h",
5954 "src/xnnpack/AlignedAllocator.h",
5955 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005956 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005957)
5958
5959xnnpack_unit_test(
Marat Dukhanbf715f92020-10-23 20:17:00 -07005960 name = "f32_dwconv2d_chw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005961 srcs = [
Marat Dukhanbf715f92020-10-23 20:17:00 -07005962 "test/f32-dwconv2d-chw.cc",
5963 "test/dwconv2d-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005964 "src/xnnpack/AlignedAllocator.h",
5965 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005966 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005967)
5968
5969xnnpack_unit_test(
Frank Barchard0bb49a72020-06-04 11:35:11 -07005970 name = "f16_gavgpool_minmax_test",
5971 srcs = [
5972 "test/f16-gavgpool-minmax.cc",
5973 "test/gavgpool-microkernel-tester.h",
5974 "src/xnnpack/AlignedAllocator.h",
5975 ] + MICROKERNEL_TEST_HDRS,
5976 deps = MICROKERNEL_TEST_DEPS,
5977)
5978
5979xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07005980 name = "f32_gavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005981 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07005982 "test/f32-gavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005983 "test/gavgpool-microkernel-tester.h",
5984 "src/xnnpack/AlignedAllocator.h",
5985 ] + MICROKERNEL_TEST_HDRS,
5986 deps = MICROKERNEL_TEST_DEPS,
5987)
5988
5989xnnpack_unit_test(
Marat Dukhan1f29b802020-05-15 23:46:39 -07005990 name = "f32_gavgpool_cw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005991 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07005992 "test/f32-gavgpool-cw.cc",
5993 "test/gavgpool-cw-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005994 "src/xnnpack/AlignedAllocator.h",
5995 ] + MICROKERNEL_TEST_HDRS,
5996 deps = MICROKERNEL_TEST_DEPS,
5997)
5998
5999xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07006000 name = "f32_gemm_test",
6001 srcs = [
6002 "test/f32-gemm.cc",
6003 "test/gemm-microkernel-tester.h",
6004 "src/xnnpack/AlignedAllocator.h",
6005 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07006006 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan163a7e62020-04-09 04:19:26 -07006007)
6008
6009xnnpack_unit_test(
Marat Dukhan467f6362020-05-22 23:21:55 -07006010 name = "f32_gemm_relu_test",
6011 srcs = [
6012 "test/f32-gemm-relu.cc",
6013 "test/gemm-microkernel-tester.h",
6014 "src/xnnpack/AlignedAllocator.h",
6015 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07006016 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan467f6362020-05-22 23:21:55 -07006017)
6018
6019xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07006020 name = "f32_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006021 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07006022 "test/f32-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006023 "test/gemm-microkernel-tester.h",
6024 "src/xnnpack/AlignedAllocator.h",
6025 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07006026 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07006027)
6028
6029xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07006030 name = "f32_gemminc_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006031 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07006032 "test/f32-gemminc-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006033 "test/gemm-microkernel-tester.h",
6034 "src/xnnpack/AlignedAllocator.h",
6035 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07006036 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07006037)
6038
6039xnnpack_unit_test(
Frank Barchardb1966592020-05-12 13:47:06 -07006040 name = "f16_hswish_test",
6041 srcs = [
6042 "test/f16-hswish.cc",
6043 "test/hswish-microkernel-tester.h",
6044 ] + MICROKERNEL_TEST_HDRS,
6045 deps = MICROKERNEL_TEST_DEPS,
6046)
6047
6048xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07006049 name = "f32_hswish_test",
6050 srcs = [
6051 "test/f32-hswish.cc",
6052 "test/hswish-microkernel-tester.h",
6053 ] + MICROKERNEL_TEST_HDRS,
6054 deps = MICROKERNEL_TEST_DEPS,
6055)
6056
6057xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07006058 name = "f32_maxpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006059 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07006060 "test/f32-maxpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006061 "test/maxpool-microkernel-tester.h",
6062 ] + MICROKERNEL_TEST_HDRS,
6063 deps = MICROKERNEL_TEST_DEPS,
6064)
6065
6066xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07006067 name = "f32_pavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006068 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07006069 "test/f32-pavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006070 "test/avgpool-microkernel-tester.h",
6071 "src/xnnpack/AlignedAllocator.h",
6072 ] + MICROKERNEL_TEST_HDRS,
6073 deps = MICROKERNEL_TEST_DEPS,
6074)
6075
6076xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07006077 name = "f32_ppmm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006078 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07006079 "test/f32-ppmm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006080 "test/gemm-microkernel-tester.h",
6081 "src/xnnpack/AlignedAllocator.h",
6082 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07006083 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07006084)
6085
6086xnnpack_unit_test(
Frank Barchardb1966592020-05-12 13:47:06 -07006087 name = "f16_prelu_test",
6088 srcs = [
6089 "test/f16-prelu.cc",
6090 "test/prelu-microkernel-tester.h",
6091 "src/xnnpack/AlignedAllocator.h",
6092 ] + MICROKERNEL_TEST_HDRS,
6093 deps = MICROKERNEL_TEST_DEPS,
6094)
6095
6096xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07006097 name = "f32_prelu_test",
6098 srcs = [
6099 "test/f32-prelu.cc",
6100 "test/prelu-microkernel-tester.h",
6101 "src/xnnpack/AlignedAllocator.h",
6102 ] + MICROKERNEL_TEST_HDRS,
6103 deps = MICROKERNEL_TEST_DEPS,
6104)
6105
6106xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07006107 name = "f32_raddexpminusmax_test",
6108 srcs = [
6109 "test/f32-raddexpminusmax.cc",
6110 "test/raddexpminusmax-microkernel-tester.h",
6111 ] + MICROKERNEL_TEST_HDRS,
6112 deps = MICROKERNEL_TEST_DEPS,
6113)
6114
6115xnnpack_unit_test(
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07006116 name = "f32_raddextexp_test",
6117 srcs = [
6118 "test/f32-raddextexp.cc",
6119 "test/raddextexp-microkernel-tester.h",
6120 ] + MICROKERNEL_TEST_HDRS,
6121 deps = MICROKERNEL_TEST_DEPS,
6122)
6123
6124xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07006125 name = "f32_raddstoreexpminusmax_test",
6126 srcs = [
6127 "test/f32-raddstoreexpminusmax.cc",
6128 "test/raddstoreexpminusmax-microkernel-tester.h",
6129 ] + MICROKERNEL_TEST_HDRS,
6130 deps = MICROKERNEL_TEST_DEPS,
6131)
6132
6133xnnpack_unit_test(
Frank Barchardfb158e22020-07-15 16:10:10 -07006134 name = "f32_relu_test",
6135 srcs = [
6136 "test/f32-relu.cc",
6137 "test/vunary-microkernel-tester.h",
6138 ] + MICROKERNEL_TEST_HDRS,
6139 deps = MICROKERNEL_TEST_DEPS,
6140)
6141
6142xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07006143 name = "f32_rmax_test",
6144 srcs = [
6145 "test/f32-rmax.cc",
6146 "test/rmax-microkernel-tester.h",
6147 ] + MICROKERNEL_TEST_HDRS,
6148 deps = MICROKERNEL_TEST_DEPS,
6149)
6150
6151xnnpack_unit_test(
Marat Dukhan346a9e52019-11-15 09:06:30 -08006152 name = "f32_sigmoid_test",
6153 srcs = [
6154 "test/f32-sigmoid.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08006155 "test/vunary-microkernel-tester.h",
Marat Dukhan346a9e52019-11-15 09:06:30 -08006156 ] + MICROKERNEL_TEST_HDRS,
6157 deps = MICROKERNEL_TEST_DEPS,
6158)
6159
6160xnnpack_unit_test(
Marat Dukhan355ab432020-04-09 19:01:52 -07006161 name = "f32_spmm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006162 srcs = [
Marat Dukhan355ab432020-04-09 19:01:52 -07006163 "test/f32-spmm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006164 "test/spmm-microkernel-tester.h",
6165 "src/xnnpack/AlignedAllocator.h",
6166 ] + MICROKERNEL_TEST_HDRS,
6167 deps = MICROKERNEL_TEST_DEPS,
6168)
6169
6170xnnpack_unit_test(
Frank Barchardd793f6c2020-05-08 13:37:43 -07006171 name = "f16_vadd_minmax_test",
6172 srcs = [
6173 "test/f16-vadd-minmax.cc",
6174 "test/vbinary-microkernel-tester.h",
6175 ] + MICROKERNEL_TEST_HDRS,
6176 deps = MICROKERNEL_TEST_DEPS,
6177)
6178
6179xnnpack_unit_test(
6180 name = "f16_vaddc_minmax_test",
6181 srcs = [
6182 "test/f16-vaddc-minmax.cc",
6183 "test/vbinaryc-microkernel-tester.h",
6184 ] + MICROKERNEL_TEST_HDRS,
6185 deps = MICROKERNEL_TEST_DEPS,
6186)
6187
6188xnnpack_unit_test(
6189 name = "f16_vdiv_minmax_test",
6190 srcs = [
6191 "test/f16-vdiv-minmax.cc",
6192 "test/vbinary-microkernel-tester.h",
6193 ] + MICROKERNEL_TEST_HDRS,
6194 deps = MICROKERNEL_TEST_DEPS,
6195)
6196
6197xnnpack_unit_test(
6198 name = "f16_vdivc_minmax_test",
6199 srcs = [
6200 "test/f16-vdivc-minmax.cc",
6201 "test/vbinaryc-microkernel-tester.h",
6202 ] + MICROKERNEL_TEST_HDRS,
6203 deps = MICROKERNEL_TEST_DEPS,
6204)
6205
6206xnnpack_unit_test(
6207 name = "f16_vrdivc_minmax_test",
6208 srcs = [
6209 "test/f16-vrdivc-minmax.cc",
6210 "test/vbinaryc-microkernel-tester.h",
6211 ] + MICROKERNEL_TEST_HDRS,
6212 deps = MICROKERNEL_TEST_DEPS,
6213)
6214
6215xnnpack_unit_test(
6216 name = "f16_vmax_test",
6217 srcs = [
6218 "test/f16-vmax.cc",
6219 "test/vbinary-microkernel-tester.h",
6220 ] + MICROKERNEL_TEST_HDRS,
6221 deps = MICROKERNEL_TEST_DEPS,
6222)
6223
6224xnnpack_unit_test(
6225 name = "f16_vmaxc_test",
6226 srcs = [
6227 "test/f16-vmaxc.cc",
6228 "test/vbinaryc-microkernel-tester.h",
6229 ] + MICROKERNEL_TEST_HDRS,
6230 deps = MICROKERNEL_TEST_DEPS,
6231)
6232
6233xnnpack_unit_test(
6234 name = "f16_vmin_test",
6235 srcs = [
6236 "test/f16-vmin.cc",
6237 "test/vbinary-microkernel-tester.h",
6238 ] + MICROKERNEL_TEST_HDRS,
6239 deps = MICROKERNEL_TEST_DEPS,
6240)
6241
6242xnnpack_unit_test(
6243 name = "f16_vminc_test",
6244 srcs = [
6245 "test/f16-vminc.cc",
6246 "test/vbinaryc-microkernel-tester.h",
6247 ] + MICROKERNEL_TEST_HDRS,
6248 deps = MICROKERNEL_TEST_DEPS,
6249)
6250
6251xnnpack_unit_test(
6252 name = "f16_vmul_minmax_test",
6253 srcs = [
6254 "test/f16-vmul-minmax.cc",
6255 "test/vbinary-microkernel-tester.h",
6256 ] + MICROKERNEL_TEST_HDRS,
6257 deps = MICROKERNEL_TEST_DEPS,
6258)
6259
6260xnnpack_unit_test(
6261 name = "f16_vmulc_minmax_test",
6262 srcs = [
6263 "test/f16-vmulc-minmax.cc",
6264 "test/vbinaryc-microkernel-tester.h",
6265 ] + MICROKERNEL_TEST_HDRS,
6266 deps = MICROKERNEL_TEST_DEPS,
6267)
6268
6269xnnpack_unit_test(
6270 name = "f16_vsub_minmax_test",
6271 srcs = [
6272 "test/f16-vsub-minmax.cc",
6273 "test/vbinary-microkernel-tester.h",
6274 ] + MICROKERNEL_TEST_HDRS,
6275 deps = MICROKERNEL_TEST_DEPS,
6276)
6277
6278xnnpack_unit_test(
6279 name = "f16_vsubc_minmax_test",
6280 srcs = [
6281 "test/f16-vsubc-minmax.cc",
6282 "test/vbinaryc-microkernel-tester.h",
6283 ] + MICROKERNEL_TEST_HDRS,
6284 deps = MICROKERNEL_TEST_DEPS,
6285)
6286
6287xnnpack_unit_test(
6288 name = "f16_vrsubc_minmax_test",
6289 srcs = [
6290 "test/f16-vrsubc-minmax.cc",
6291 "test/vbinaryc-microkernel-tester.h",
6292 ] + MICROKERNEL_TEST_HDRS,
6293 deps = MICROKERNEL_TEST_DEPS,
6294)
6295
6296xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07006297 name = "f32_vabs_test",
6298 srcs = [
6299 "test/f32-vabs.cc",
6300 "test/vunary-microkernel-tester.h",
6301 ] + MICROKERNEL_TEST_HDRS,
6302 deps = MICROKERNEL_TEST_DEPS,
6303)
6304
6305xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07006306 name = "f32_vadd_test",
6307 srcs = [
6308 "test/f32-vadd.cc",
6309 "test/vbinary-microkernel-tester.h",
6310 ] + MICROKERNEL_TEST_HDRS,
6311 deps = MICROKERNEL_TEST_DEPS,
6312)
6313
6314xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006315 name = "f32_vadd_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006316 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006317 "test/f32-vadd-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08006318 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08006319 ] + MICROKERNEL_TEST_HDRS,
6320 deps = MICROKERNEL_TEST_DEPS,
6321)
6322
6323xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07006324 name = "f32_vadd_relu_test",
6325 srcs = [
6326 "test/f32-vadd-relu.cc",
6327 "test/vbinary-microkernel-tester.h",
6328 ] + MICROKERNEL_TEST_HDRS,
6329 deps = MICROKERNEL_TEST_DEPS,
6330)
6331
6332xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07006333 name = "f32_vaddc_test",
6334 srcs = [
6335 "test/f32-vaddc.cc",
6336 "test/vbinaryc-microkernel-tester.h",
6337 ] + MICROKERNEL_TEST_HDRS,
6338 deps = MICROKERNEL_TEST_DEPS,
6339)
6340
6341xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006342 name = "f32_vaddc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08006343 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006344 "test/f32-vaddc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08006345 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006346 ] + MICROKERNEL_TEST_HDRS,
6347 deps = MICROKERNEL_TEST_DEPS,
6348)
6349
6350xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07006351 name = "f32_vaddc_relu_test",
6352 srcs = [
6353 "test/f32-vaddc-relu.cc",
6354 "test/vbinaryc-microkernel-tester.h",
6355 ] + MICROKERNEL_TEST_HDRS,
6356 deps = MICROKERNEL_TEST_DEPS,
6357)
6358
6359xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07006360 name = "f32_vdiv_test",
6361 srcs = [
6362 "test/f32-vdiv.cc",
6363 "test/vbinary-microkernel-tester.h",
6364 ] + MICROKERNEL_TEST_HDRS,
6365 deps = MICROKERNEL_TEST_DEPS,
6366)
6367
6368xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006369 name = "f32_vdiv_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08006370 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006371 "test/f32-vdiv-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08006372 "test/vbinary-microkernel-tester.h",
6373 ] + MICROKERNEL_TEST_HDRS,
6374 deps = MICROKERNEL_TEST_DEPS,
6375)
6376
6377xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07006378 name = "f32_vdiv_relu_test",
6379 srcs = [
6380 "test/f32-vdiv-relu.cc",
6381 "test/vbinary-microkernel-tester.h",
6382 ] + MICROKERNEL_TEST_HDRS,
6383 deps = MICROKERNEL_TEST_DEPS,
6384)
6385
6386xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07006387 name = "f32_vdivc_test",
6388 srcs = [
6389 "test/f32-vdivc.cc",
6390 "test/vbinaryc-microkernel-tester.h",
6391 ] + MICROKERNEL_TEST_HDRS,
6392 deps = MICROKERNEL_TEST_DEPS,
6393)
6394
6395xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006396 name = "f32_vdivc_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08006397 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006398 "test/f32-vdivc-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08006399 "test/vbinaryc-microkernel-tester.h",
6400 ] + MICROKERNEL_TEST_HDRS,
6401 deps = MICROKERNEL_TEST_DEPS,
6402)
6403
6404xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07006405 name = "f32_vdivc_relu_test",
6406 srcs = [
6407 "test/f32-vdivc-relu.cc",
6408 "test/vbinaryc-microkernel-tester.h",
6409 ] + MICROKERNEL_TEST_HDRS,
6410 deps = MICROKERNEL_TEST_DEPS,
6411)
6412
6413xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07006414 name = "f32_vrdivc_test",
6415 srcs = [
6416 "test/f32-vrdivc.cc",
6417 "test/vbinaryc-microkernel-tester.h",
6418 ] + MICROKERNEL_TEST_HDRS,
6419 deps = MICROKERNEL_TEST_DEPS,
6420)
6421
6422xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006423 name = "f32_vrdivc_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08006424 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006425 "test/f32-vrdivc-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08006426 "test/vbinaryc-microkernel-tester.h",
6427 ] + MICROKERNEL_TEST_HDRS,
6428 deps = MICROKERNEL_TEST_DEPS,
6429)
6430
6431xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07006432 name = "f32_vrdivc_relu_test",
6433 srcs = [
6434 "test/f32-vrdivc-relu.cc",
6435 "test/vbinaryc-microkernel-tester.h",
6436 ] + MICROKERNEL_TEST_HDRS,
6437 deps = MICROKERNEL_TEST_DEPS,
6438)
6439
6440xnnpack_unit_test(
Marat Dukhaned6baaf2020-12-01 15:07:08 -08006441 name = "f32_velu_test",
6442 srcs = [
6443 "test/f32-velu.cc",
6444 "test/vunary-microkernel-tester.h",
6445 ] + MICROKERNEL_TEST_HDRS,
6446 deps = MICROKERNEL_TEST_DEPS,
6447)
6448
6449xnnpack_unit_test(
Marat Dukhan403b7d42019-12-05 12:49:11 -08006450 name = "f32_vmax_test",
6451 srcs = [
6452 "test/f32-vmax.cc",
6453 "test/vbinary-microkernel-tester.h",
6454 ] + MICROKERNEL_TEST_HDRS,
6455 deps = MICROKERNEL_TEST_DEPS,
6456)
6457
6458xnnpack_unit_test(
6459 name = "f32_vmaxc_test",
6460 srcs = [
6461 "test/f32-vmaxc.cc",
6462 "test/vbinaryc-microkernel-tester.h",
6463 ] + MICROKERNEL_TEST_HDRS,
6464 deps = MICROKERNEL_TEST_DEPS,
6465)
6466
6467xnnpack_unit_test(
6468 name = "f32_vmin_test",
6469 srcs = [
6470 "test/f32-vmin.cc",
6471 "test/vbinary-microkernel-tester.h",
6472 ] + MICROKERNEL_TEST_HDRS,
6473 deps = MICROKERNEL_TEST_DEPS,
6474)
6475
6476xnnpack_unit_test(
6477 name = "f32_vminc_test",
6478 srcs = [
6479 "test/f32-vminc.cc",
6480 "test/vbinaryc-microkernel-tester.h",
6481 ] + MICROKERNEL_TEST_HDRS,
6482 deps = MICROKERNEL_TEST_DEPS,
6483)
6484
6485xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07006486 name = "f32_vmul_test",
6487 srcs = [
6488 "test/f32-vmul.cc",
6489 "test/vbinary-microkernel-tester.h",
6490 ] + MICROKERNEL_TEST_HDRS,
6491 deps = MICROKERNEL_TEST_DEPS,
6492)
6493
6494xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006495 name = "f32_vmul_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006496 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006497 "test/f32-vmul-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08006498 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08006499 ] + MICROKERNEL_TEST_HDRS,
6500 deps = MICROKERNEL_TEST_DEPS,
6501)
6502
6503xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07006504 name = "f32_vmul_relu_test",
6505 srcs = [
6506 "test/f32-vmul-relu.cc",
6507 "test/vbinary-microkernel-tester.h",
6508 ] + MICROKERNEL_TEST_HDRS,
6509 deps = MICROKERNEL_TEST_DEPS,
6510)
6511
6512xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07006513 name = "f32_vmulc_test",
6514 srcs = [
6515 "test/f32-vmulc.cc",
6516 "test/vbinaryc-microkernel-tester.h",
6517 ] + MICROKERNEL_TEST_HDRS,
6518 deps = MICROKERNEL_TEST_DEPS,
6519)
6520
6521xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006522 name = "f32_vmulc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08006523 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006524 "test/f32-vmulc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08006525 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006526 ] + MICROKERNEL_TEST_HDRS,
6527 deps = MICROKERNEL_TEST_DEPS,
6528)
6529
6530xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07006531 name = "f32_vmulc_relu_test",
6532 srcs = [
6533 "test/f32-vmulc-relu.cc",
6534 "test/vbinaryc-microkernel-tester.h",
6535 ] + MICROKERNEL_TEST_HDRS,
6536 deps = MICROKERNEL_TEST_DEPS,
6537)
6538
6539xnnpack_unit_test(
Frank Barchard2a1049c2020-06-03 02:31:27 -07006540 name = "f16_vmulcaddc_minmax_test",
6541 srcs = [
6542 "test/f16-vmulcaddc-minmax.cc",
6543 "test/vmulcaddc-microkernel-tester.h",
6544 "src/xnnpack/AlignedAllocator.h",
6545 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07006546 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Frank Barchard2a1049c2020-06-03 02:31:27 -07006547)
6548
6549xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07006550 name = "f32_vmulcaddc_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006551 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07006552 "test/f32-vmulcaddc-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006553 "test/vmulcaddc-microkernel-tester.h",
6554 "src/xnnpack/AlignedAllocator.h",
6555 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07006556 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07006557)
6558
6559xnnpack_unit_test(
Marat Dukhan8cc7efe2020-06-10 16:24:27 -07006560 name = "f32_vlrelu_test",
6561 srcs = [
6562 "test/f32-vlrelu.cc",
6563 "test/vunary-microkernel-tester.h",
6564 ] + MICROKERNEL_TEST_HDRS,
6565 deps = MICROKERNEL_TEST_DEPS,
6566)
6567
6568xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07006569 name = "f32_vneg_test",
6570 srcs = [
6571 "test/f32-vneg.cc",
6572 "test/vunary-microkernel-tester.h",
6573 ] + MICROKERNEL_TEST_HDRS,
6574 deps = MICROKERNEL_TEST_DEPS,
6575)
6576
6577xnnpack_unit_test(
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07006578 name = "f32_vrndne_test",
6579 srcs = [
6580 "test/f32-vrndne.cc",
6581 "test/vunary-microkernel-tester.h",
6582 ] + MICROKERNEL_TEST_HDRS,
6583 deps = MICROKERNEL_TEST_DEPS,
6584)
6585
6586xnnpack_unit_test(
6587 name = "f32_vrndz_test",
6588 srcs = [
6589 "test/f32-vrndz.cc",
6590 "test/vunary-microkernel-tester.h",
6591 ] + MICROKERNEL_TEST_HDRS,
6592 deps = MICROKERNEL_TEST_DEPS,
6593)
6594
6595xnnpack_unit_test(
6596 name = "f32_vrndu_test",
6597 srcs = [
6598 "test/f32-vrndu.cc",
6599 "test/vunary-microkernel-tester.h",
6600 ] + MICROKERNEL_TEST_HDRS,
6601 deps = MICROKERNEL_TEST_DEPS,
6602)
6603
6604xnnpack_unit_test(
6605 name = "f32_vrndd_test",
6606 srcs = [
6607 "test/f32-vrndd.cc",
6608 "test/vunary-microkernel-tester.h",
6609 ] + MICROKERNEL_TEST_HDRS,
6610 deps = MICROKERNEL_TEST_DEPS,
6611)
6612
6613xnnpack_unit_test(
Marat Dukhan05ac8e32019-10-21 15:39:33 -07006614 name = "f32_vscale_test",
6615 srcs = [
6616 "test/f32-vscale.cc",
6617 "test/vscale-microkernel-tester.h",
6618 ] + MICROKERNEL_TEST_HDRS,
6619 deps = MICROKERNEL_TEST_DEPS,
6620)
6621
6622xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07006623 name = "f32_vscaleexpminusmax_test",
6624 srcs = [
6625 "test/f32-vscaleexpminusmax.cc",
6626 "test/vscaleexpminusmax-microkernel-tester.h",
6627 ] + MICROKERNEL_TEST_HDRS,
6628 deps = MICROKERNEL_TEST_DEPS,
6629)
6630
6631xnnpack_unit_test(
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07006632 name = "f32_vscaleextexp_test",
6633 srcs = [
6634 "test/f32-vscaleextexp.cc",
6635 "test/vscaleextexp-microkernel-tester.h",
6636 ] + MICROKERNEL_TEST_HDRS,
6637 deps = MICROKERNEL_TEST_DEPS,
6638)
6639
6640xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07006641 name = "f32_vsqr_test",
6642 srcs = [
6643 "test/f32-vsqr.cc",
6644 "test/vunary-microkernel-tester.h",
6645 ] + MICROKERNEL_TEST_HDRS,
6646 deps = MICROKERNEL_TEST_DEPS,
6647)
6648
6649xnnpack_unit_test(
Marat Dukhan13bafb02020-06-05 00:43:11 -07006650 name = "f32_vsqrdiff_test",
6651 srcs = [
6652 "test/f32-vsqrdiff.cc",
6653 "test/vbinary-microkernel-tester.h",
6654 ] + MICROKERNEL_TEST_HDRS,
6655 deps = MICROKERNEL_TEST_DEPS,
6656)
6657
6658xnnpack_unit_test(
6659 name = "f32_vsqrdiffc_test",
6660 srcs = [
6661 "test/f32-vsqrdiffc.cc",
6662 "test/vbinaryc-microkernel-tester.h",
6663 ] + MICROKERNEL_TEST_HDRS,
6664 deps = MICROKERNEL_TEST_DEPS,
6665)
6666
6667xnnpack_unit_test(
Marat Dukhanf4db2f32020-06-30 10:55:30 -07006668 name = "f32_vsqrt_test",
6669 srcs = [
6670 "test/f32-vsqrt.cc",
6671 "test/vunary-microkernel-tester.h",
6672 ] + MICROKERNEL_TEST_HDRS,
6673 deps = MICROKERNEL_TEST_DEPS,
6674)
6675
6676xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07006677 name = "f32_vsub_test",
6678 srcs = [
6679 "test/f32-vsub.cc",
6680 "test/vbinary-microkernel-tester.h",
6681 ] + MICROKERNEL_TEST_HDRS,
6682 deps = MICROKERNEL_TEST_DEPS,
6683)
6684
6685xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006686 name = "f32_vsub_minmax_test",
Marat Dukhan97579532019-10-18 16:40:39 -07006687 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006688 "test/f32-vsub-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08006689 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08006690 ] + MICROKERNEL_TEST_HDRS,
6691 deps = MICROKERNEL_TEST_DEPS,
6692)
6693
6694xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07006695 name = "f32_vsub_relu_test",
6696 srcs = [
6697 "test/f32-vsub-relu.cc",
6698 "test/vbinary-microkernel-tester.h",
6699 ] + MICROKERNEL_TEST_HDRS,
6700 deps = MICROKERNEL_TEST_DEPS,
6701)
6702
6703xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07006704 name = "f32_vsubc_test",
6705 srcs = [
6706 "test/f32-vsubc.cc",
6707 "test/vbinaryc-microkernel-tester.h",
6708 ] + MICROKERNEL_TEST_HDRS,
6709 deps = MICROKERNEL_TEST_DEPS,
6710)
6711
6712xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006713 name = "f32_vsubc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08006714 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006715 "test/f32-vsubc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08006716 "test/vbinaryc-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08006717 ] + MICROKERNEL_TEST_HDRS,
6718 deps = MICROKERNEL_TEST_DEPS,
6719)
6720
6721xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07006722 name = "f32_vsubc_relu_test",
6723 srcs = [
6724 "test/f32-vsubc-relu.cc",
6725 "test/vbinaryc-microkernel-tester.h",
6726 ] + MICROKERNEL_TEST_HDRS,
6727 deps = MICROKERNEL_TEST_DEPS,
6728)
6729
6730xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07006731 name = "f32_vrsubc_test",
6732 srcs = [
6733 "test/f32-vrsubc.cc",
6734 "test/vbinaryc-microkernel-tester.h",
6735 ] + MICROKERNEL_TEST_HDRS,
6736 deps = MICROKERNEL_TEST_DEPS,
6737)
6738
6739xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006740 name = "f32_vrsubc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08006741 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006742 "test/f32-vrsubc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08006743 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan97579532019-10-18 16:40:39 -07006744 ] + MICROKERNEL_TEST_HDRS,
6745 deps = MICROKERNEL_TEST_DEPS,
6746)
6747
6748xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07006749 name = "f32_vrsubc_relu_test",
6750 srcs = [
6751 "test/f32-vrsubc-relu.cc",
6752 "test/vbinaryc-microkernel-tester.h",
6753 ] + MICROKERNEL_TEST_HDRS,
6754 deps = MICROKERNEL_TEST_DEPS,
6755)
6756
6757xnnpack_unit_test(
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07006758 name = "qs8_dwconv_minmax_test",
6759 srcs = [
6760 "test/qs8-dwconv-minmax.cc",
6761 "test/dwconv-microkernel-tester.h",
6762 "src/xnnpack/AlignedAllocator.h",
6763 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
6764 deps = MICROKERNEL_TEST_DEPS + [":packing"],
6765)
6766
6767xnnpack_unit_test(
Marat Dukhan4ed53f42020-08-06 01:12:55 -07006768 name = "qs8_gavgpool_minmax_test",
6769 srcs = [
6770 "test/qs8-gavgpool-minmax.cc",
6771 "test/gavgpool-microkernel-tester.h",
6772 "src/xnnpack/AlignedAllocator.h",
6773 ] + MICROKERNEL_TEST_HDRS,
6774 deps = MICROKERNEL_TEST_DEPS,
6775)
6776
6777xnnpack_unit_test(
Marat Dukhan595e1702020-07-31 10:12:52 -07006778 name = "qs8_gemm_minmax_test",
6779 srcs = [
6780 "test/qs8-gemm-minmax.cc",
6781 "test/gemm-microkernel-tester.h",
6782 "src/xnnpack/AlignedAllocator.h",
6783 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
6784 deps = MICROKERNEL_TEST_DEPS + [":packing"],
6785)
6786
6787xnnpack_unit_test(
Marat Dukhanf9480682020-07-31 14:50:24 -07006788 name = "qs8_igemm_minmax_test",
6789 srcs = [
6790 "test/qs8-igemm-minmax.cc",
6791 "test/gemm-microkernel-tester.h",
6792 "src/xnnpack/AlignedAllocator.h",
6793 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
6794 deps = MICROKERNEL_TEST_DEPS + [":packing"],
6795)
6796
6797xnnpack_unit_test(
6798 name = "qs8_requantization_test",
6799 srcs = [
6800 "src/xnnpack/requantization-stubs.h",
6801 "test/qs8-requantization.cc",
6802 "test/requantization-tester.h",
6803 ] + MICROKERNEL_TEST_HDRS,
6804 deps = MICROKERNEL_TEST_DEPS,
6805)
6806
6807xnnpack_unit_test(
Marat Dukhand9f3ad42020-08-10 12:30:58 -07006808 name = "qs8_vadd_minmax_test",
6809 srcs = [
6810 "test/qs8-vadd-minmax.cc",
6811 "test/vadd-microkernel-tester.h",
6812 ] + MICROKERNEL_TEST_HDRS,
6813 deps = MICROKERNEL_TEST_DEPS,
6814)
6815
6816xnnpack_unit_test(
Marat Dukhan0270d9f2020-08-11 00:56:46 -07006817 name = "qs8_vaddc_minmax_test",
6818 srcs = [
6819 "test/qs8-vaddc-minmax.cc",
6820 "test/vaddc-microkernel-tester.h",
6821 ] + MICROKERNEL_TEST_HDRS,
6822 deps = MICROKERNEL_TEST_DEPS,
6823)
6824
6825xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07006826 name = "qu8_avgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006827 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07006828 "test/qu8-avgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006829 "test/avgpool-microkernel-tester.h",
6830 "src/xnnpack/AlignedAllocator.h",
6831 ] + MICROKERNEL_TEST_HDRS,
6832 deps = MICROKERNEL_TEST_DEPS,
6833)
6834
6835xnnpack_unit_test(
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07006836 name = "qu8_dwconv_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006837 srcs = [
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07006838 "test/qu8-dwconv-minmax.cc",
6839 "test/dwconv-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006840 "src/xnnpack/AlignedAllocator.h",
6841 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07006842 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07006843)
6844
6845xnnpack_unit_test(
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07006846 name = "qu8_igemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006847 srcs = [
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07006848 "test/qu8-igemm-minmax.cc",
6849 "test/gemm-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006850 "src/xnnpack/AlignedAllocator.h",
6851 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07006852 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07006853)
6854
6855xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07006856 name = "qu8_gavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006857 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07006858 "test/qu8-gavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006859 "test/gavgpool-microkernel-tester.h",
6860 "src/xnnpack/AlignedAllocator.h",
6861 ] + MICROKERNEL_TEST_HDRS,
6862 deps = MICROKERNEL_TEST_DEPS,
6863)
6864
6865xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07006866 name = "qu8_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006867 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07006868 "test/qu8-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006869 "test/gemm-microkernel-tester.h",
6870 "src/xnnpack/AlignedAllocator.h",
6871 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07006872 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07006873)
6874
6875xnnpack_unit_test(
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07006876 name = "qu8_requantization_test",
6877 srcs = [
6878 "src/xnnpack/requantization-stubs.h",
6879 "test/qu8-requantization.cc",
6880 "test/requantization-tester.h",
6881 ] + MICROKERNEL_TEST_HDRS,
6882 deps = MICROKERNEL_TEST_DEPS,
6883)
6884
6885xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07006886 name = "qu8_vadd_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006887 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07006888 "test/qu8-vadd-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006889 "test/vadd-microkernel-tester.h",
6890 ] + MICROKERNEL_TEST_HDRS,
6891 deps = MICROKERNEL_TEST_DEPS,
6892)
6893
6894xnnpack_unit_test(
6895 name = "u8_clamp_test",
6896 srcs = [
6897 "test/u8-clamp.cc",
6898 "test/clamp-microkernel-tester.h",
6899 ] + MICROKERNEL_TEST_HDRS,
6900 deps = MICROKERNEL_TEST_DEPS,
6901)
6902
6903xnnpack_unit_test(
6904 name = "u8_lut32norm_test",
6905 srcs = [
6906 "test/u8-lut32norm.cc",
6907 "test/lut-norm-microkernel-tester.h",
6908 ] + MICROKERNEL_TEST_HDRS,
6909 deps = MICROKERNEL_TEST_DEPS,
6910)
6911
6912xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07006913 name = "u8_maxpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006914 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07006915 "test/u8-maxpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006916 "test/maxpool-microkernel-tester.h",
6917 ] + MICROKERNEL_TEST_HDRS,
6918 deps = MICROKERNEL_TEST_DEPS,
6919)
6920
6921xnnpack_unit_test(
6922 name = "u8_rmax_test",
6923 srcs = [
6924 "test/u8-rmax.cc",
6925 "test/rmax-microkernel-tester.h",
6926 ] + MICROKERNEL_TEST_HDRS,
6927 deps = MICROKERNEL_TEST_DEPS,
6928)
6929
6930xnnpack_unit_test(
Marat Dukhanad71b9a2020-11-20 00:01:51 -08006931 name = "x32_depthtospace2d_chw2hwc_test",
Yury Kartynnike7841862020-11-04 18:22:18 -08006932 srcs = [
Marat Dukhanad71b9a2020-11-20 00:01:51 -08006933 "test/x32-depthtospace2d-chw2hwc.cc",
6934 "test/depthtospace-microkernel-tester.h",
Yury Kartynnike7841862020-11-04 18:22:18 -08006935 ] + MICROKERNEL_TEST_HDRS,
6936 deps = MICROKERNEL_TEST_DEPS,
6937)
6938
6939xnnpack_unit_test(
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -07006940 name = "x32_fill_test",
6941 srcs = [
6942 "test/x32-fill.cc",
6943 "test/fill-microkernel-tester.h",
6944 ] + MICROKERNEL_TEST_HDRS,
6945 deps = MICROKERNEL_TEST_DEPS,
6946)
6947
6948xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07006949 name = "x32_packx_test",
6950 srcs = [
6951 "test/x32-packx.cc",
6952 "test/pack-microkernel-tester.h",
6953 "src/xnnpack/AlignedAllocator.h",
6954 ] + MICROKERNEL_TEST_HDRS,
6955 deps = MICROKERNEL_TEST_DEPS,
6956)
6957
6958xnnpack_unit_test(
6959 name = "x32_pad_test",
6960 srcs = [
6961 "test/x32-pad.cc",
6962 "test/pad-microkernel-tester.h",
6963 ] + MICROKERNEL_TEST_HDRS,
6964 deps = MICROKERNEL_TEST_DEPS,
6965)
6966
6967xnnpack_unit_test(
6968 name = "x32_unpool_test",
6969 srcs = [
6970 "test/x32-unpool.cc",
6971 "test/unpool-microkernel-tester.h",
6972 ] + MICROKERNEL_TEST_HDRS,
6973 deps = MICROKERNEL_TEST_DEPS,
6974)
6975
6976xnnpack_unit_test(
6977 name = "x32_zip_test",
6978 srcs = [
6979 "test/x32-zip.cc",
6980 "test/zip-microkernel-tester.h",
6981 ] + MICROKERNEL_TEST_HDRS,
6982 deps = MICROKERNEL_TEST_DEPS,
6983)
6984
6985xnnpack_unit_test(
6986 name = "x8_lut_test",
6987 srcs = [
6988 "test/x8-lut.cc",
6989 "test/lut-microkernel-tester.h",
6990 ] + MICROKERNEL_TEST_HDRS,
6991 deps = MICROKERNEL_TEST_DEPS,
6992)
6993
6994xnnpack_unit_test(
6995 name = "x8_zip_test",
6996 srcs = [
6997 "test/x8-zip.cc",
6998 "test/zip-microkernel-tester.h",
6999 ] + MICROKERNEL_TEST_HDRS,
7000 deps = MICROKERNEL_TEST_DEPS,
7001)
7002
Marat Dukhan20c3b922020-03-10 03:45:06 -07007003########################## Size tests for the library #########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07007004
7005xnnpack_binary(
Marat Dukhan20c3b922020-03-10 03:45:06 -07007006 name = "operator_size_test",
7007 srcs = ["test/operator-size.c"],
Marat Dukhanf0cb70a2021-03-30 15:45:15 -07007008 deps = [":xnnpack_for_tfjs"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07007009)
7010
Marat Dukhan20c3b922020-03-10 03:45:06 -07007011xnnpack_binary(
7012 name = "subgraph_size_test",
7013 srcs = ["test/subgraph-size.c"],
7014 deps = [":XNNPACK"],
7015)
7016
7017########################### Unit tests for operators ##########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07007018
7019xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07007020 name = "abs_nc_test",
7021 srcs = [
7022 "test/abs-nc.cc",
7023 "test/abs-operator-tester.h",
7024 ],
7025 deps = OPERATOR_TEST_DEPS,
7026)
7027
7028xnnpack_unit_test(
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08007029 name = "add_nd_test",
Artsiom Ablavatski2202c812021-01-22 14:16:43 -08007030 timeout = "moderate",
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08007031 srcs = [
7032 "test/add-nd.cc",
7033 "test/binary-elementwise-operator-tester.h",
7034 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07007035 deps = OPERATOR_TEST_DEPS,
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08007036)
7037
7038xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08007039 name = "argmax_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007040 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08007041 "test/argmax-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007042 "test/argmax-pooling-operator-tester.h",
7043 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07007044 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07007045)
7046
7047xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08007048 name = "average_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007049 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08007050 "test/average-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007051 "test/average-pooling-operator-tester.h",
7052 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07007053 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07007054)
7055
7056xnnpack_unit_test(
Marat Dukhan64e52512020-06-09 13:41:16 -07007057 name = "bankers_rounding_nc_test",
7058 srcs = [
7059 "test/bankers-rounding-nc.cc",
7060 "test/bankers-rounding-operator-tester.h",
7061 ],
7062 deps = OPERATOR_TEST_DEPS,
7063)
7064
7065xnnpack_unit_test(
7066 name = "ceiling_nc_test",
7067 srcs = [
7068 "test/ceiling-nc.cc",
7069 "test/ceiling-operator-tester.h",
7070 ],
7071 deps = OPERATOR_TEST_DEPS,
7072)
7073
7074xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08007075 name = "channel_shuffle_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007076 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08007077 "test/channel-shuffle-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007078 "test/channel-shuffle-operator-tester.h",
7079 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07007080 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07007081)
7082
7083xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08007084 name = "clamp_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007085 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08007086 "test/clamp-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007087 "test/clamp-operator-tester.h",
7088 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07007089 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07007090)
7091
7092xnnpack_unit_test(
Marat Dukhan065b11e2020-05-22 09:49:41 -07007093 name = "constant_pad_nd_test",
7094 srcs = [
7095 "test/constant-pad-nd.cc",
7096 "test/constant-pad-operator-tester.h",
7097 ],
7098 deps = OPERATOR_TEST_DEPS,
7099)
7100
7101xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08007102 name = "convolution_nhwc_test",
Artsiom Ablavatski2202c812021-01-22 14:16:43 -08007103 timeout = "moderate",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007104 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08007105 "test/convolution-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007106 "test/convolution-operator-tester.h",
7107 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07007108 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07007109)
7110
7111xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08007112 name = "convolution_nchw_test",
Artsiom Ablavatski2202c812021-01-22 14:16:43 -08007113 timeout = "moderate",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007114 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08007115 "test/convolution-nchw.cc",
7116 "test/convolution-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007117 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07007118 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07007119)
7120
7121xnnpack_unit_test(
Marat Dukhan4e21b272020-06-04 18:45:01 -07007122 name = "copy_nc_test",
7123 srcs = [
7124 "test/copy-nc.cc",
7125 "test/copy-operator-tester.h",
7126 ],
7127 deps = OPERATOR_TEST_DEPS,
7128)
7129
7130xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08007131 name = "deconvolution_nhwc_test",
Artsiom Ablavatskic1aa2972020-12-08 11:23:34 -08007132 timeout = "moderate",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007133 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08007134 "test/deconvolution-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007135 "test/deconvolution-operator-tester.h",
7136 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07007137 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07007138)
7139
7140xnnpack_unit_test(
Marat Dukhan188d1042020-11-24 23:39:40 -08007141 name = "depth_to_space_nchw2nhwc_test",
Artsiom Ablavatski0f1dc182020-11-05 19:21:50 -08007142 srcs = [
Marat Dukhan188d1042020-11-24 23:39:40 -08007143 "test/depth-to-space-nchw2nhwc.cc",
Artsiom Ablavatski0f1dc182020-11-05 19:21:50 -08007144 "test/depth-to-space-operator-tester.h",
7145 ] + OPERATOR_TEST_PARAMS_HDRS,
7146 deps = OPERATOR_TEST_DEPS,
7147)
7148
7149xnnpack_unit_test(
Marat Dukhan0e521172020-11-25 13:10:04 -08007150 name = "depth_to_space_nhwc_test",
7151 srcs = [
7152 "test/depth-to-space-nhwc.cc",
7153 "test/depth-to-space-operator-tester.h",
7154 ] + OPERATOR_TEST_PARAMS_HDRS,
7155 deps = OPERATOR_TEST_DEPS,
7156)
7157
7158xnnpack_unit_test(
Marat Dukhan69180502019-12-06 15:00:31 -08007159 name = "divide_nd_test",
7160 srcs = [
7161 "test/binary-elementwise-operator-tester.h",
7162 "test/divide-nd.cc",
7163 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07007164 deps = OPERATOR_TEST_DEPS,
Marat Dukhan69180502019-12-06 15:00:31 -08007165)
7166
7167xnnpack_unit_test(
Marat Dukhanb6bd4bc2020-12-01 17:01:40 -08007168 name = "elu_nc_test",
7169 srcs = [
7170 "test/elu-nc.cc",
7171 "test/elu-operator-tester.h",
7172 ],
7173 deps = OPERATOR_TEST_DEPS,
7174)
7175
7176xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08007177 name = "fully_connected_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007178 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08007179 "test/fully-connected-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007180 "test/fully-connected-operator-tester.h",
7181 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07007182 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07007183)
7184
7185xnnpack_unit_test(
Marat Dukhan64e52512020-06-09 13:41:16 -07007186 name = "floor_nc_test",
7187 srcs = [
7188 "test/floor-nc.cc",
7189 "test/floor-operator-tester.h",
7190 ],
7191 deps = OPERATOR_TEST_DEPS,
7192)
7193
7194xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08007195 name = "global_average_pooling_nwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007196 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08007197 "test/global-average-pooling-nwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007198 "test/global-average-pooling-operator-tester.h",
Marat Dukhanef61d022020-06-19 13:54:49 -07007199 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07007200 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07007201)
7202
7203xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08007204 name = "global_average_pooling_ncw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007205 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08007206 "test/global-average-pooling-ncw.cc",
7207 "test/global-average-pooling-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007208 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07007209 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07007210)
7211
7212xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08007213 name = "hardswish_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007214 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08007215 "test/hardswish-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007216 "test/hardswish-operator-tester.h",
7217 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07007218 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07007219)
7220
7221xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08007222 name = "leaky_relu_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007223 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08007224 "test/leaky-relu-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007225 "test/leaky-relu-operator-tester.h",
7226 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07007227 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07007228)
7229
7230xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08007231 name = "max_pooling_nhwc_test",
Artsiom Ablavatski2202c812021-01-22 14:16:43 -08007232 timeout = "moderate",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007233 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08007234 "test/max-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007235 "test/max-pooling-operator-tester.h",
7236 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07007237 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07007238)
7239
7240xnnpack_unit_test(
Marat Dukhan79e7f842019-12-05 14:35:50 -08007241 name = "maximum_nd_test",
7242 srcs = [
7243 "test/binary-elementwise-operator-tester.h",
7244 "test/maximum-nd.cc",
7245 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07007246 deps = OPERATOR_TEST_DEPS,
Marat Dukhan79e7f842019-12-05 14:35:50 -08007247)
7248
7249xnnpack_unit_test(
7250 name = "minimum_nd_test",
7251 srcs = [
7252 "test/binary-elementwise-operator-tester.h",
7253 "test/minimum-nd.cc",
7254 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07007255 deps = OPERATOR_TEST_DEPS,
Marat Dukhan79e7f842019-12-05 14:35:50 -08007256)
7257
7258xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08007259 name = "multiply_nd_test",
Marat Dukhanca2733c2019-11-15 23:21:17 -08007260 srcs = [
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08007261 "test/binary-elementwise-operator-tester.h",
Marat Dukhanefc47b82019-11-18 09:25:38 -08007262 "test/multiply-nd.cc",
Marat Dukhanca2733c2019-11-15 23:21:17 -08007263 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07007264 deps = OPERATOR_TEST_DEPS,
Marat Dukhanca2733c2019-11-15 23:21:17 -08007265)
7266
7267xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07007268 name = "negate_nc_test",
7269 srcs = [
7270 "test/negate-nc.cc",
7271 "test/negate-operator-tester.h",
7272 ],
7273 deps = OPERATOR_TEST_DEPS,
7274)
7275
7276xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08007277 name = "prelu_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007278 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08007279 "test/prelu-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007280 "test/prelu-operator-tester.h",
7281 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07007282 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07007283)
7284
7285xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08007286 name = "resize_bilinear_nhwc_test",
Marat Dukhan69722492019-11-11 19:55:50 -08007287 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08007288 "test/resize-bilinear-nhwc.cc",
Marat Dukhan69722492019-11-11 19:55:50 -08007289 "test/resize-bilinear-operator-tester.h",
7290 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07007291 deps = OPERATOR_TEST_DEPS,
Marat Dukhan69722492019-11-11 19:55:50 -08007292)
7293
7294xnnpack_unit_test(
Artsiom Ablavatski97918102020-10-27 15:52:59 -07007295 name = "resize_bilinear_nchw_test",
7296 srcs = [
7297 "test/resize-bilinear-nchw.cc",
7298 "test/resize-bilinear-operator-tester.h",
7299 ] + OPERATOR_TEST_PARAMS_HDRS,
7300 deps = OPERATOR_TEST_DEPS,
7301)
7302
7303xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08007304 name = "sigmoid_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007305 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08007306 "test/sigmoid-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007307 "test/sigmoid-operator-tester.h",
7308 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07007309 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07007310)
7311
7312xnnpack_unit_test(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08007313 name = "softmax_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007314 srcs = [
Marat Dukhanfd8e6892020-01-27 15:25:25 -08007315 "test/softmax-nc.cc",
7316 "test/softmax-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007317 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07007318 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07007319)
7320
7321xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07007322 name = "square_nc_test",
7323 srcs = [
7324 "test/square-nc.cc",
7325 "test/square-operator-tester.h",
7326 ],
7327 deps = OPERATOR_TEST_DEPS,
7328)
7329
7330xnnpack_unit_test(
Marat Dukhan6804bbd2020-06-30 19:26:11 -07007331 name = "square_root_nc_test",
7332 srcs = [
7333 "test/square-root-nc.cc",
7334 "test/square-root-operator-tester.h",
7335 ],
7336 deps = OPERATOR_TEST_DEPS,
7337)
7338
7339xnnpack_unit_test(
Marat Dukhanf7399262020-06-05 10:58:44 -07007340 name = "squared_difference_nd_test",
7341 srcs = [
7342 "test/binary-elementwise-operator-tester.h",
7343 "test/squared-difference-nd.cc",
7344 ],
7345 deps = OPERATOR_TEST_DEPS,
7346)
7347
7348xnnpack_unit_test(
Marat Dukhan05f3f6d2019-12-03 15:13:53 -08007349 name = "subtract_nd_test",
7350 srcs = [
7351 "test/binary-elementwise-operator-tester.h",
7352 "test/subtract-nd.cc",
7353 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07007354 deps = OPERATOR_TEST_DEPS,
Marat Dukhan05f3f6d2019-12-03 15:13:53 -08007355)
7356
7357xnnpack_unit_test(
Marat Dukhan64e52512020-06-09 13:41:16 -07007358 name = "truncation_nc_test",
7359 srcs = [
7360 "test/truncation-nc.cc",
7361 "test/truncation-operator-tester.h",
7362 ],
7363 deps = OPERATOR_TEST_DEPS,
7364)
7365
7366xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08007367 name = "unpooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007368 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08007369 "test/unpooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007370 "test/unpooling-operator-tester.h",
7371 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07007372 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07007373)
7374
Chao Mei6ddfc602020-05-13 22:29:36 -07007375############################### Misc unit tests ###############################
7376
7377xnnpack_unit_test(
7378 name = "memory_planner_test",
7379 srcs = [
7380 "test/memory-planner-test.cc",
7381 ],
7382 deps = [
7383 ":XNNPACK",
7384 ":memory_planner",
7385 ],
7386)
7387
XNNPACK Teamab8c4c82020-10-09 08:05:51 -07007388xnnpack_unit_test(
7389 name = "subgraph_nchw_test",
7390 srcs = [
7391 "src/xnnpack/subgraph.h",
7392 "test/subgraph-nchw.cc",
7393 "test/subgraph-tester.h",
7394 ],
7395 deps = [
7396 ":XNNPACK",
7397 ],
7398)
7399
Marat Dukhan08c4a432019-10-03 09:29:21 -07007400############################# Build configurations #############################
7401
Marat Dukhanb8642352019-10-30 15:43:02 -07007402# Enables usage of assembly kernels.
Marat Dukhan08c4a432019-10-03 09:29:21 -07007403config_setting(
Marat Dukhanb8642352019-10-30 15:43:02 -07007404 name = "xnn_enable_assembly_explicit_true",
7405 define_values = {"xnn_enable_assembly": "true"},
7406)
7407
7408# Disables usage of assembly kernels.
7409config_setting(
7410 name = "xnn_enable_assembly_explicit_false",
7411 define_values = {"xnn_enable_assembly": "false"},
7412)
7413
Marat Dukhan9de90e02020-06-18 16:04:12 -07007414# Enables usage of sparse inference.
7415config_setting(
7416 name = "xnn_enable_sparse_explicit_true",
7417 define_values = {"xnn_enable_sparse": "true"},
7418)
7419
7420# Disables usage of sparse inference.
7421config_setting(
7422 name = "xnn_enable_sparse_explicit_false",
7423 define_values = {"xnn_enable_sparse": "false"},
7424)
7425
Marat Dukhan05702cf2020-03-26 15:41:33 -07007426# Disables usage of HMP-aware optimizations.
7427config_setting(
7428 name = "xnn_enable_hmp_explicit_false",
7429 define_values = {"xnn_enable_hmp": "false"},
7430)
7431
Chao Mei6ddfc602020-05-13 22:29:36 -07007432# Enable usage of optimized memory allocation
7433config_setting(
7434 name = "xnn_enable_memopt_explicit_true",
7435 define_values = {"xnnpack_enable_memopt": "true"},
7436)
7437
7438# Disable usage of optimized memory allocation
7439config_setting(
7440 name = "xnn_enable_memopt_explicit_false",
7441 define_values = {"xnnpack_enable_memopt": "false"},
7442)
7443
Marat Dukhanb8642352019-10-30 15:43:02 -07007444# Builds with -c dbg
7445config_setting(
7446 name = "debug_build",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007447 values = {
Marat Dukhanb8642352019-10-30 15:43:02 -07007448 "compilation_mode": "dbg",
7449 },
7450)
7451
7452# Builds with -c opt
7453config_setting(
7454 name = "optimized_build",
7455 values = {
7456 "compilation_mode": "opt",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007457 },
7458)
7459
7460config_setting(
Marat Dukhanb8642352019-10-30 15:43:02 -07007461 name = "linux_k8",
7462 values = {"cpu": "k8"},
7463)
7464
7465config_setting(
Marat Dukhan582094e2020-04-30 17:21:25 -07007466 name = "linux_arm",
7467 values = {"cpu": "arm"},
Marat Dukhan4e45e662019-10-03 15:40:24 -07007468)
7469
7470config_setting(
Marat Dukhanf0bd4de2020-06-15 15:53:19 -07007471 name = "linux_armeabi",
7472 values = {"cpu": "armeabi"},
7473)
7474
7475config_setting(
Terry Heo68eef3f2020-04-13 22:53:52 -07007476 name = "linux_armhf",
7477 values = {"cpu": "armhf"},
7478)
7479
7480config_setting(
Marat Dukhana720e932020-06-10 13:01:11 -07007481 name = "linux_armv7a",
7482 values = {"cpu": "armv7a"},
7483)
7484
7485config_setting(
Marat Dukhan582094e2020-04-30 17:21:25 -07007486 name = "linux_aarch64",
7487 values = {"cpu": "aarch64"},
7488)
7489
7490config_setting(
Marat Dukhan08c4a432019-10-03 09:29:21 -07007491 name = "android",
7492 values = {"crosstool_top": "//external:android/crosstool"},
7493)
7494
7495config_setting(
7496 name = "android_armv7",
7497 values = {
7498 "crosstool_top": "//external:android/crosstool",
7499 "cpu": "armeabi-v7a",
7500 },
7501)
7502
7503config_setting(
7504 name = "android_arm64",
7505 values = {
7506 "crosstool_top": "//external:android/crosstool",
7507 "cpu": "arm64-v8a",
7508 },
7509)
7510
7511config_setting(
7512 name = "android_x86",
7513 values = {
7514 "crosstool_top": "//external:android/crosstool",
7515 "cpu": "x86",
7516 },
7517)
7518
7519config_setting(
7520 name = "android_x86_64",
7521 values = {
7522 "crosstool_top": "//external:android/crosstool",
7523 "cpu": "x86_64",
7524 },
7525)
7526
7527config_setting(
Marat Dukhan10a38082020-04-17 03:58:35 -07007528 name = "windows_x86_64",
7529 values = {"cpu": "x64_windows"},
Marat Dukhan9fe932e2020-04-11 17:14:15 -07007530)
7531
7532config_setting(
Marat Dukhan10a38082020-04-17 03:58:35 -07007533 name = "windows_x86_64_clang",
7534 values = {
7535 "compiler": "clang-cl",
7536 "cpu": "x64_windows",
7537 },
7538)
7539
7540config_setting(
7541 name = "windows_x86_64_mingw",
7542 values = {
7543 "compiler": "mingw-gcc",
7544 "cpu": "x64_windows",
7545 },
7546)
7547
7548config_setting(
7549 name = "windows_x86_64_msys",
7550 values = {
7551 "compiler": "msys-gcc",
7552 "cpu": "x64_windows",
7553 },
Marat Dukhan9fe932e2020-04-11 17:14:15 -07007554)
7555
7556config_setting(
Marat Dukhan885ca242019-10-07 09:17:32 -07007557 name = "macos_x86_64",
7558 values = {
7559 "apple_platform_type": "macos",
7560 "cpu": "darwin",
7561 },
7562)
7563
7564config_setting(
Simon Maurerae33ab82021-03-03 23:38:22 +01007565 name = "macos_arm64",
7566 values = {
7567 "apple_platform_type": "macos",
7568 "cpu": "darwin_arm64",
7569 },
7570)
7571
7572config_setting(
Marat Dukhan08c4a432019-10-03 09:29:21 -07007573 name = "emscripten",
XNNPACK Team3bfbdaf2021-03-29 15:26:23 -07007574 values = {"crosstool_top": "@emsdk//emscripten_toolchain:everything"},
Marat Dukhan08c4a432019-10-03 09:29:21 -07007575)
7576
7577config_setting(
7578 name = "emscripten_wasm",
7579 values = {
XNNPACK Team3bfbdaf2021-03-29 15:26:23 -07007580 "crosstool_top": "@emsdk//emscripten_toolchain:everything",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007581 "cpu": "wasm",
7582 },
7583)
7584
7585config_setting(
7586 name = "emscripten_wasmsimd",
7587 values = {
XNNPACK Team3bfbdaf2021-03-29 15:26:23 -07007588 "crosstool_top": "@emsdk//emscripten_toolchain:everything",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007589 "cpu": "wasm",
Marat Dukhan81c62602020-05-29 13:22:49 -07007590 "copt": "-msimd128",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007591 },
7592)
7593
7594config_setting(
Marat Dukhan1498d1d2020-02-11 20:00:05 -08007595 name = "ios_armv7",
7596 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08007597 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08007598 "cpu": "ios_armv7",
7599 },
7600)
7601
7602config_setting(
7603 name = "ios_arm64",
7604 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08007605 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08007606 "cpu": "ios_arm64",
7607 },
7608)
7609
7610config_setting(
7611 name = "ios_arm64e",
7612 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08007613 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08007614 "cpu": "ios_arm64e",
7615 },
7616)
7617
7618config_setting(
7619 name = "ios_x86",
7620 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08007621 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08007622 "cpu": "ios_i386",
7623 },
7624)
7625
7626config_setting(
7627 name = "ios_x86_64",
7628 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08007629 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08007630 "cpu": "ios_x86_64",
7631 },
7632)
7633
7634config_setting(
7635 name = "watchos_armv7k",
7636 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08007637 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08007638 "cpu": "watchos_armv7k",
7639 },
7640)
7641
7642config_setting(
7643 name = "watchos_arm64_32",
7644 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08007645 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08007646 "cpu": "watchos_arm64_32",
7647 },
7648)
7649
7650config_setting(
7651 name = "watchos_x86",
7652 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08007653 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08007654 "cpu": "watchos_i386",
7655 },
7656)
7657
7658config_setting(
7659 name = "watchos_x86_64",
7660 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08007661 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08007662 "cpu": "watchos_x86_64",
7663 },
7664)
7665
7666config_setting(
7667 name = "tvos_arm64",
7668 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08007669 "apple_platform_type": "tvos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08007670 "cpu": "tvos_arm64",
7671 },
7672)
7673
7674config_setting(
7675 name = "tvos_x86_64",
7676 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08007677 "apple_platform_type": "tvos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08007678 "cpu": "tvos_x86_64",
7679 },
7680)