blob: 9db8b0c46c2d531d9c2c5be06225f350bba58c2f [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 Dukhan5fab4092020-06-10 01:28:28 -070091 "src/subgraph/floor.c",
Frank Barchard04336c12020-10-22 16:48:55 -070092 "src/subgraph/fully-connected.c",
Marat Dukhana059b7d2020-06-11 11:41:27 -070093 "src/subgraph/global-average-pooling-2d.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -070094 "src/subgraph/hardswish.c",
Marat Dukhan5bbebac2020-06-10 19:42:15 -070095 "src/subgraph/leaky-relu.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -070096 "src/subgraph/max-pooling-2d.c",
Marat Dukhan9d3a4592020-06-05 16:52:42 -070097 "src/subgraph/maximum2.c",
98 "src/subgraph/minimum2.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -070099 "src/subgraph/multiply2.c",
Marat Dukhan5fab4092020-06-10 01:28:28 -0700100 "src/subgraph/negate.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -0700101 "src/subgraph/prelu.c",
102 "src/subgraph/sigmoid.c",
103 "src/subgraph/softmax.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700104 "src/subgraph/square-root.c",
105 "src/subgraph/square.c",
106 "src/subgraph/squared-difference.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -0700107 "src/subgraph/static-constant-pad.c",
Marat Dukhand27202d2020-07-09 23:43:40 -0700108 "src/subgraph/static-reshape.c",
Marat Dukhanaff24e22020-07-23 01:43:58 -0700109 "src/subgraph/static-resize-bilinear-2d.c",
Marat Dukhan9d3a4592020-06-05 16:52:42 -0700110 "src/subgraph/subtract.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -0700111 "src/subgraph/unpooling-2d.c",
112]
113
Marat Dukhan3a77ea72019-12-23 12:10:24 -0800114TABLE_SRCS = [
115 "src/tables/exp2-k-over-64.c",
116 "src/tables/exp2-k-over-2048.c",
Marat Dukhanc60742b2020-11-23 12:33:27 -0800117 "src/tables/exp2minus-k-over-16.c",
Marat Dukhan1f256fc2020-09-24 21:27:14 -0700118 "src/tables/exp2minus-k-over-64.c",
119 "src/tables/exp2minus-k-over-2048.c",
Marat Dukhan3a77ea72019-12-23 12:10:24 -0800120]
121
Marat Dukhan08c4a432019-10-03 09:29:21 -0700122SCALAR_UKERNELS = [
Marat Dukhan329da642019-11-19 21:44:39 -0800123 "src/f32-argmaxpool/4x-scalar-c1.c",
Marat Dukhan1e782c42019-11-21 17:02:40 -0800124 "src/f32-argmaxpool/9p8x-scalar-c1.c",
Marat Dukhan329da642019-11-19 21:44:39 -0800125 "src/f32-argmaxpool/9x-scalar-c1.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700126 "src/f32-avgpool/9p8x-minmax-scalar-c1.c",
127 "src/f32-avgpool/9x-minmax-scalar-c1.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -0700128 "src/f32-clamp/gen/scalar-x1.c",
129 "src/f32-clamp/gen/scalar-x2.c",
130 "src/f32-clamp/gen/scalar-x4.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700131 "src/f32-conv-hwc/3x3s2p0p1c3x4-scalar-1x1.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700132 "src/f32-conv-hwc/3x3s2p1c3x4-scalar-1x1.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -0700133 "src/f32-conv-hwc2chw/3x3s2p1c3x4-scalar-1x1.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700134 "src/f32-dwconv/gen/up1x4-minmax-scalar-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700135 "src/f32-dwconv/gen/up1x4-minmax-scalar.c",
136 "src/f32-dwconv/gen/up1x4-scalar-acc2.c",
137 "src/f32-dwconv/gen/up1x4-scalar.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700138 "src/f32-dwconv/gen/up1x9-minmax-scalar-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700139 "src/f32-dwconv/gen/up1x9-minmax-scalar.c",
140 "src/f32-dwconv/gen/up1x9-scalar-acc2.c",
141 "src/f32-dwconv/gen/up1x9-scalar.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700142 "src/f32-dwconv/gen/up1x25-minmax-scalar-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700143 "src/f32-dwconv/gen/up1x25-minmax-scalar.c",
144 "src/f32-dwconv/gen/up1x25-scalar-acc2.c",
145 "src/f32-dwconv/gen/up1x25-scalar.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700146 "src/f32-dwconv/gen/up2x4-minmax-scalar-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700147 "src/f32-dwconv/gen/up2x4-minmax-scalar.c",
148 "src/f32-dwconv/gen/up2x4-scalar-acc2.c",
149 "src/f32-dwconv/gen/up2x4-scalar.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700150 "src/f32-dwconv/gen/up2x9-minmax-scalar-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700151 "src/f32-dwconv/gen/up2x9-minmax-scalar.c",
152 "src/f32-dwconv/gen/up2x9-scalar-acc2.c",
153 "src/f32-dwconv/gen/up2x9-scalar.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700154 "src/f32-dwconv/gen/up2x25-minmax-scalar-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700155 "src/f32-dwconv/gen/up2x25-minmax-scalar.c",
156 "src/f32-dwconv/gen/up2x25-scalar-acc2.c",
157 "src/f32-dwconv/gen/up2x25-scalar.c",
Frank Barchard35db7d02020-10-26 13:37:34 -0700158 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-scalar-1x1-acc2.c",
159 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-scalar-1x1-acc3.c",
160 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-scalar-1x1-acc4.c",
Marat Dukhan91249d22020-10-24 12:02:51 -0700161 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-scalar-1x1.c",
Frank Barchard35db7d02020-10-26 13:37:34 -0700162 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-scalar-2x1-acc2.c",
Marat Dukhan91249d22020-10-24 12:02:51 -0700163 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-scalar-2x1.c",
164 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-scalar-3x1.c",
165 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-scalar-4x1.c",
166 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-scalar-5x1.c",
167 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-scalar-6x1.c",
Marat Dukhancf5b3c32020-10-25 19:21:10 -0700168 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-scalar-1x1-acc2.c",
169 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-scalar-1x1-acc3.c",
170 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-scalar-1x1-acc4.c",
Frank Barchard35db7d02020-10-26 13:37:34 -0700171 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-scalar-1x1.c",
Marat Dukhancf5b3c32020-10-25 19:21:10 -0700172 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-scalar-2x1-acc2.c",
Frank Barchard35db7d02020-10-26 13:37:34 -0700173 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-scalar-2x1.c",
174 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-scalar-3x1.c",
175 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-scalar-4x1.c",
Marat Dukhanc4efb002020-10-25 23:14:47 -0700176 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-scalar-1x1-acc2.c",
177 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-scalar-1x1-acc3.c",
178 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-scalar-1x1-acc4.c",
179 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-scalar-1x1-acc5.c",
Frank Barchard35db7d02020-10-26 13:37:34 -0700180 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-scalar-1x1.c",
Marat Dukhanc4efb002020-10-25 23:14:47 -0700181 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-scalar-2x1-acc2.c",
182 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-scalar-2x1-acc3.c",
Frank Barchard35db7d02020-10-26 13:37:34 -0700183 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-scalar-2x1.c",
Marat Dukhanc4efb002020-10-25 23:14:47 -0700184 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-scalar-3x1-acc2.c",
Frank Barchard35db7d02020-10-26 13:37:34 -0700185 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-scalar-3x1.c",
Marat Dukhan29c0c332020-10-28 22:11:00 -0700186 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-scalar-1x1-acc2.c",
187 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-scalar-1x1-acc3.c",
188 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-scalar-1x1-acc4.c",
189 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-scalar-1x1-acc5.c",
190 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-scalar-1x1.c",
191 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-scalar-2x1-acc2.c",
192 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-scalar-2x1-acc3.c",
193 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-scalar-2x1.c",
194 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-scalar-3x1-acc2.c",
195 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-scalar-3x1.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -0700196 "src/f32-gavgpool-cw/scalar-x1.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700197 "src/f32-gavgpool/7p7x-minmax-scalar-c1.c",
198 "src/f32-gavgpool/7x-minmax-scalar-c1.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700199 "src/f32-gemm/gen-inc/1x4inc-minmax-scalar.c",
200 "src/f32-gemm/gen-inc/2x4inc-minmax-scalar.c",
201 "src/f32-gemm/gen-inc/4x4inc-minmax-scalar.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700202 "src/f32-gemm/gen/1x4-minmax-scalar.c",
203 "src/f32-gemm/gen/1x4-relu-scalar.c",
204 "src/f32-gemm/gen/1x4-scalar.c",
205 "src/f32-gemm/gen/2x4-minmax-scalar.c",
206 "src/f32-gemm/gen/2x4-relu-scalar.c",
207 "src/f32-gemm/gen/2x4-scalar.c",
208 "src/f32-gemm/gen/4x2-minmax-scalar.c",
209 "src/f32-gemm/gen/4x2-relu-scalar.c",
210 "src/f32-gemm/gen/4x2-scalar.c",
211 "src/f32-gemm/gen/4x4-minmax-scalar.c",
212 "src/f32-gemm/gen/4x4-relu-scalar.c",
213 "src/f32-gemm/gen/4x4-scalar.c",
Marat Dukhanefaac272020-07-16 16:58:07 -0700214 "src/f32-hswish/gen/hswish-scalar-x1.c",
215 "src/f32-hswish/gen/hswish-scalar-x2.c",
216 "src/f32-hswish/gen/hswish-scalar-x4.c",
XNNPACK Team21432672020-10-19 19:58:48 -0700217 "src/f32-ibilinear-chw/gen/scalar-p1.c",
218 "src/f32-ibilinear-chw/gen/scalar-p2.c",
219 "src/f32-ibilinear-chw/gen/scalar-p4.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700220 "src/f32-ibilinear/gen/scalar-c1.c",
221 "src/f32-ibilinear/gen/scalar-c2.c",
222 "src/f32-ibilinear/gen/scalar-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700223 "src/f32-igemm/gen/1x4-minmax-scalar.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700224 "src/f32-igemm/gen/1x4-relu-scalar.c",
225 "src/f32-igemm/gen/1x4-scalar.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700226 "src/f32-igemm/gen/2x4-minmax-scalar.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700227 "src/f32-igemm/gen/2x4-relu-scalar.c",
228 "src/f32-igemm/gen/2x4-scalar.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700229 "src/f32-igemm/gen/4x2-minmax-scalar.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700230 "src/f32-igemm/gen/4x2-relu-scalar.c",
231 "src/f32-igemm/gen/4x2-scalar.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700232 "src/f32-igemm/gen/4x4-minmax-scalar.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700233 "src/f32-igemm/gen/4x4-relu-scalar.c",
234 "src/f32-igemm/gen/4x4-scalar.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700235 "src/f32-maxpool/9p8x-minmax-scalar-c1.c",
236 "src/f32-pavgpool/9p8x-minmax-scalar-c1.c",
237 "src/f32-pavgpool/9x-minmax-scalar-c1.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700238 "src/f32-ppmm/gen/2x4-minmax-scalar.c",
239 "src/f32-ppmm/gen/3x3-minmax-scalar.c",
240 "src/f32-ppmm/gen/4x2-minmax-scalar.c",
241 "src/f32-ppmm/gen/4x4-minmax-scalar.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800242 "src/f32-prelu/gen/scalar-2x1.c",
243 "src/f32-prelu/gen/scalar-2x4.c",
Marat Dukhanf46f6752020-01-21 11:03:49 -0800244 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x1.c",
Marat Dukhanf46f6752020-01-21 11:03:49 -0800245 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x2-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700246 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x2.c",
Marat Dukhanf46f6752020-01-21 11:03:49 -0800247 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x4-acc2.c",
248 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x4-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700249 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x4.c",
Marat Dukhanf46f6752020-01-21 11:03:49 -0800250 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x1.c",
Marat Dukhanf46f6752020-01-21 11:03:49 -0800251 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x2-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700252 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x2.c",
Marat Dukhanf46f6752020-01-21 11:03:49 -0800253 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x4-acc2.c",
254 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x4-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700255 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x4.c",
Frank Barchardfb158e22020-07-15 16:10:10 -0700256 "src/f32-relu/gen/scalar-x1.c",
257 "src/f32-relu/gen/scalar-x2.c",
258 "src/f32-relu/gen/scalar-x4.c",
Frank Barchard37297a62020-08-27 12:15:10 -0700259 "src/f32-relu/gen/scalar-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700260 "src/f32-rmax/scalar.c",
Marat Dukhan3a77ea72019-12-23 12:10:24 -0800261 "src/f32-sigmoid/gen/scalar-lut64-p2-div-x1.c",
262 "src/f32-sigmoid/gen/scalar-lut64-p2-div-x2.c",
263 "src/f32-sigmoid/gen/scalar-lut64-p2-div-x4.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700264 "src/f32-sigmoid/gen/scalar-lut2048-p1-div-x1.c",
265 "src/f32-sigmoid/gen/scalar-lut2048-p1-div-x2.c",
266 "src/f32-sigmoid/gen/scalar-lut2048-p1-div-x4.c",
Marat Dukhan3a77ea72019-12-23 12:10:24 -0800267 "src/f32-sigmoid/gen/scalar-p5-div-x1.c",
268 "src/f32-sigmoid/gen/scalar-p5-div-x2.c",
269 "src/f32-sigmoid/gen/scalar-p5-div-x4.c",
Marat Dukhan355ab432020-04-09 19:01:52 -0700270 "src/f32-spmm/gen/1x1-minmax-scalar-pipelined.c",
271 "src/f32-spmm/gen/1x1-minmax-scalar.c",
272 "src/f32-spmm/gen/2x1-minmax-scalar-pipelined.c",
273 "src/f32-spmm/gen/2x1-minmax-scalar.c",
274 "src/f32-spmm/gen/4x1-minmax-scalar-pipelined.c",
275 "src/f32-spmm/gen/4x1-minmax-scalar.c",
276 "src/f32-spmm/gen/8x1-minmax-scalar-pipelined.c",
277 "src/f32-spmm/gen/8x1-minmax-scalar.c",
278 "src/f32-spmm/gen/8x2-minmax-scalar.c",
279 "src/f32-spmm/gen/8x4-minmax-scalar.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700280 "src/f32-vbinary/gen/vadd-minmax-scalar-x1.c",
281 "src/f32-vbinary/gen/vadd-minmax-scalar-x2.c",
282 "src/f32-vbinary/gen/vadd-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700283 "src/f32-vbinary/gen/vadd-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700284 "src/f32-vbinary/gen/vadd-relu-scalar-x1.c",
285 "src/f32-vbinary/gen/vadd-relu-scalar-x2.c",
286 "src/f32-vbinary/gen/vadd-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700287 "src/f32-vbinary/gen/vadd-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700288 "src/f32-vbinary/gen/vadd-scalar-x1.c",
289 "src/f32-vbinary/gen/vadd-scalar-x2.c",
290 "src/f32-vbinary/gen/vadd-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700291 "src/f32-vbinary/gen/vadd-scalar-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700292 "src/f32-vbinary/gen/vaddc-minmax-scalar-x1.c",
293 "src/f32-vbinary/gen/vaddc-minmax-scalar-x2.c",
294 "src/f32-vbinary/gen/vaddc-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700295 "src/f32-vbinary/gen/vaddc-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700296 "src/f32-vbinary/gen/vaddc-relu-scalar-x1.c",
297 "src/f32-vbinary/gen/vaddc-relu-scalar-x2.c",
298 "src/f32-vbinary/gen/vaddc-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700299 "src/f32-vbinary/gen/vaddc-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700300 "src/f32-vbinary/gen/vaddc-scalar-x1.c",
301 "src/f32-vbinary/gen/vaddc-scalar-x2.c",
302 "src/f32-vbinary/gen/vaddc-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700303 "src/f32-vbinary/gen/vaddc-scalar-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700304 "src/f32-vbinary/gen/vdiv-minmax-scalar-x1.c",
305 "src/f32-vbinary/gen/vdiv-minmax-scalar-x2.c",
306 "src/f32-vbinary/gen/vdiv-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700307 "src/f32-vbinary/gen/vdiv-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700308 "src/f32-vbinary/gen/vdiv-relu-scalar-x1.c",
309 "src/f32-vbinary/gen/vdiv-relu-scalar-x2.c",
310 "src/f32-vbinary/gen/vdiv-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700311 "src/f32-vbinary/gen/vdiv-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700312 "src/f32-vbinary/gen/vdiv-scalar-x1.c",
313 "src/f32-vbinary/gen/vdiv-scalar-x2.c",
314 "src/f32-vbinary/gen/vdiv-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700315 "src/f32-vbinary/gen/vdiv-scalar-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700316 "src/f32-vbinary/gen/vdivc-minmax-scalar-x1.c",
317 "src/f32-vbinary/gen/vdivc-minmax-scalar-x2.c",
318 "src/f32-vbinary/gen/vdivc-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700319 "src/f32-vbinary/gen/vdivc-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700320 "src/f32-vbinary/gen/vdivc-relu-scalar-x1.c",
321 "src/f32-vbinary/gen/vdivc-relu-scalar-x2.c",
322 "src/f32-vbinary/gen/vdivc-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700323 "src/f32-vbinary/gen/vdivc-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700324 "src/f32-vbinary/gen/vdivc-scalar-x1.c",
325 "src/f32-vbinary/gen/vdivc-scalar-x2.c",
326 "src/f32-vbinary/gen/vdivc-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700327 "src/f32-vbinary/gen/vdivc-scalar-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800328 "src/f32-vbinary/gen/vmax-scalar-x1.c",
329 "src/f32-vbinary/gen/vmax-scalar-x2.c",
330 "src/f32-vbinary/gen/vmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700331 "src/f32-vbinary/gen/vmax-scalar-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800332 "src/f32-vbinary/gen/vmaxc-scalar-x1.c",
333 "src/f32-vbinary/gen/vmaxc-scalar-x2.c",
334 "src/f32-vbinary/gen/vmaxc-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700335 "src/f32-vbinary/gen/vmaxc-scalar-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800336 "src/f32-vbinary/gen/vmin-scalar-x1.c",
337 "src/f32-vbinary/gen/vmin-scalar-x2.c",
338 "src/f32-vbinary/gen/vmin-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700339 "src/f32-vbinary/gen/vmin-scalar-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800340 "src/f32-vbinary/gen/vminc-scalar-x1.c",
341 "src/f32-vbinary/gen/vminc-scalar-x2.c",
342 "src/f32-vbinary/gen/vminc-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700343 "src/f32-vbinary/gen/vminc-scalar-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700344 "src/f32-vbinary/gen/vmul-minmax-scalar-x1.c",
345 "src/f32-vbinary/gen/vmul-minmax-scalar-x2.c",
346 "src/f32-vbinary/gen/vmul-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700347 "src/f32-vbinary/gen/vmul-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700348 "src/f32-vbinary/gen/vmul-relu-scalar-x1.c",
349 "src/f32-vbinary/gen/vmul-relu-scalar-x2.c",
350 "src/f32-vbinary/gen/vmul-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700351 "src/f32-vbinary/gen/vmul-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700352 "src/f32-vbinary/gen/vmul-scalar-x1.c",
353 "src/f32-vbinary/gen/vmul-scalar-x2.c",
354 "src/f32-vbinary/gen/vmul-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700355 "src/f32-vbinary/gen/vmul-scalar-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700356 "src/f32-vbinary/gen/vmulc-minmax-scalar-x1.c",
357 "src/f32-vbinary/gen/vmulc-minmax-scalar-x2.c",
358 "src/f32-vbinary/gen/vmulc-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700359 "src/f32-vbinary/gen/vmulc-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700360 "src/f32-vbinary/gen/vmulc-relu-scalar-x1.c",
361 "src/f32-vbinary/gen/vmulc-relu-scalar-x2.c",
362 "src/f32-vbinary/gen/vmulc-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700363 "src/f32-vbinary/gen/vmulc-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700364 "src/f32-vbinary/gen/vmulc-scalar-x1.c",
365 "src/f32-vbinary/gen/vmulc-scalar-x2.c",
366 "src/f32-vbinary/gen/vmulc-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700367 "src/f32-vbinary/gen/vmulc-scalar-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700368 "src/f32-vbinary/gen/vrdivc-minmax-scalar-x1.c",
369 "src/f32-vbinary/gen/vrdivc-minmax-scalar-x2.c",
370 "src/f32-vbinary/gen/vrdivc-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700371 "src/f32-vbinary/gen/vrdivc-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700372 "src/f32-vbinary/gen/vrdivc-relu-scalar-x1.c",
373 "src/f32-vbinary/gen/vrdivc-relu-scalar-x2.c",
374 "src/f32-vbinary/gen/vrdivc-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700375 "src/f32-vbinary/gen/vrdivc-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700376 "src/f32-vbinary/gen/vrdivc-scalar-x1.c",
377 "src/f32-vbinary/gen/vrdivc-scalar-x2.c",
378 "src/f32-vbinary/gen/vrdivc-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700379 "src/f32-vbinary/gen/vrdivc-scalar-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700380 "src/f32-vbinary/gen/vrsubc-minmax-scalar-x1.c",
381 "src/f32-vbinary/gen/vrsubc-minmax-scalar-x2.c",
382 "src/f32-vbinary/gen/vrsubc-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700383 "src/f32-vbinary/gen/vrsubc-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700384 "src/f32-vbinary/gen/vrsubc-relu-scalar-x1.c",
385 "src/f32-vbinary/gen/vrsubc-relu-scalar-x2.c",
386 "src/f32-vbinary/gen/vrsubc-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700387 "src/f32-vbinary/gen/vrsubc-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700388 "src/f32-vbinary/gen/vrsubc-scalar-x1.c",
389 "src/f32-vbinary/gen/vrsubc-scalar-x2.c",
390 "src/f32-vbinary/gen/vrsubc-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700391 "src/f32-vbinary/gen/vrsubc-scalar-x8.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -0700392 "src/f32-vbinary/gen/vsqrdiff-scalar-x1.c",
393 "src/f32-vbinary/gen/vsqrdiff-scalar-x2.c",
394 "src/f32-vbinary/gen/vsqrdiff-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700395 "src/f32-vbinary/gen/vsqrdiff-scalar-x8.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -0700396 "src/f32-vbinary/gen/vsqrdiffc-scalar-x1.c",
397 "src/f32-vbinary/gen/vsqrdiffc-scalar-x2.c",
398 "src/f32-vbinary/gen/vsqrdiffc-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700399 "src/f32-vbinary/gen/vsqrdiffc-scalar-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700400 "src/f32-vbinary/gen/vsub-minmax-scalar-x1.c",
401 "src/f32-vbinary/gen/vsub-minmax-scalar-x2.c",
402 "src/f32-vbinary/gen/vsub-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700403 "src/f32-vbinary/gen/vsub-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700404 "src/f32-vbinary/gen/vsub-relu-scalar-x1.c",
405 "src/f32-vbinary/gen/vsub-relu-scalar-x2.c",
406 "src/f32-vbinary/gen/vsub-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700407 "src/f32-vbinary/gen/vsub-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700408 "src/f32-vbinary/gen/vsub-scalar-x1.c",
409 "src/f32-vbinary/gen/vsub-scalar-x2.c",
410 "src/f32-vbinary/gen/vsub-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700411 "src/f32-vbinary/gen/vsub-scalar-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700412 "src/f32-vbinary/gen/vsubc-minmax-scalar-x1.c",
413 "src/f32-vbinary/gen/vsubc-minmax-scalar-x2.c",
414 "src/f32-vbinary/gen/vsubc-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700415 "src/f32-vbinary/gen/vsubc-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700416 "src/f32-vbinary/gen/vsubc-relu-scalar-x1.c",
417 "src/f32-vbinary/gen/vsubc-relu-scalar-x2.c",
418 "src/f32-vbinary/gen/vsubc-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700419 "src/f32-vbinary/gen/vsubc-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700420 "src/f32-vbinary/gen/vsubc-scalar-x1.c",
421 "src/f32-vbinary/gen/vsubc-scalar-x2.c",
422 "src/f32-vbinary/gen/vsubc-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700423 "src/f32-vbinary/gen/vsubc-scalar-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700424 "src/f32-vlrelu/gen/vlrelu-scalar-x1.c",
425 "src/f32-vlrelu/gen/vlrelu-scalar-x2.c",
426 "src/f32-vlrelu/gen/vlrelu-scalar-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700427 "src/f32-vmulcaddc/gen/c1-minmax-scalar-2x.c",
428 "src/f32-vmulcaddc/gen/c2-minmax-scalar-2x.c",
429 "src/f32-vmulcaddc/gen/c4-minmax-scalar-2x.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -0700430 "src/f32-vrnd/gen/vrndd-scalar-libm-x1.c",
431 "src/f32-vrnd/gen/vrndd-scalar-libm-x2.c",
432 "src/f32-vrnd/gen/vrndd-scalar-libm-x4.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700433 "src/f32-vrnd/gen/vrndne-scalar-libm-x1.c",
434 "src/f32-vrnd/gen/vrndne-scalar-libm-x2.c",
435 "src/f32-vrnd/gen/vrndne-scalar-libm-x4.c",
436 "src/f32-vrnd/gen/vrndu-scalar-libm-x1.c",
437 "src/f32-vrnd/gen/vrndu-scalar-libm-x2.c",
438 "src/f32-vrnd/gen/vrndu-scalar-libm-x4.c",
439 "src/f32-vrnd/gen/vrndz-scalar-libm-x1.c",
440 "src/f32-vrnd/gen/vrndz-scalar-libm-x2.c",
441 "src/f32-vrnd/gen/vrndz-scalar-libm-x4.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -0700442 "src/f32-vsqrt/gen/scalar-sqrt-x1.c",
443 "src/f32-vsqrt/gen/scalar-sqrt-x2.c",
444 "src/f32-vsqrt/gen/scalar-sqrt-x4.c",
Marat Dukhan5020b962020-06-08 13:30:10 -0700445 "src/f32-vunary/gen/vabs-scalar-x1.c",
446 "src/f32-vunary/gen/vabs-scalar-x2.c",
447 "src/f32-vunary/gen/vabs-scalar-x4.c",
448 "src/f32-vunary/gen/vneg-scalar-x1.c",
449 "src/f32-vunary/gen/vneg-scalar-x2.c",
450 "src/f32-vunary/gen/vneg-scalar-x4.c",
451 "src/f32-vunary/gen/vsqr-scalar-x1.c",
452 "src/f32-vunary/gen/vsqr-scalar-x2.c",
453 "src/f32-vunary/gen/vsqr-scalar-x4.c",
Marat Dukhanc60742b2020-11-23 12:33:27 -0800454 "src/math/expm1minus-scalar-rr2-lut16-p3.c",
455 "src/math/expm1minus-scalar-rr2-lut16-p4.c",
456 "src/math/expm1minus-scalar-rr2-p5.c",
457 "src/math/expm1minus-scalar-rr2-p6.c",
Frank Barchard22136062020-11-24 18:44:46 -0800458 "src/math/expminus-scalar-rr2-lut64-p2.c",
459 "src/math/expminus-scalar-rr2-lut2048-p1.c",
460 "src/math/expminus-scalar-rr2-p5.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -0700461 "src/math/roundd-scalar-addsub.c",
462 "src/math/roundd-scalar-cvt.c",
Marat Dukhanffbf96a2020-05-14 02:59:08 -0700463 "src/math/roundd-scalar-floor.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700464 "src/math/roundne-scalar-addsub.c",
465 "src/math/roundne-scalar-nearbyint.c",
466 "src/math/roundne-scalar-rint.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -0700467 "src/math/roundu-scalar-addsub.c",
Marat Dukhanffbf96a2020-05-14 02:59:08 -0700468 "src/math/roundu-scalar-ceil.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -0700469 "src/math/roundu-scalar-cvt.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -0700470 "src/math/roundz-scalar-addsub.c",
471 "src/math/roundz-scalar-cvt.c",
Marat Dukhanffbf96a2020-05-14 02:59:08 -0700472 "src/math/roundz-scalar-trunc.c",
Marat Dukhanf8475d62020-09-17 15:01:43 -0700473 "src/math/sigmoid-scalar-rr2-lut64-p2-div.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700474 "src/math/sigmoid-scalar-rr2-lut2048-p1-div.c",
Marat Dukhanf8475d62020-09-17 15:01:43 -0700475 "src/math/sigmoid-scalar-rr2-p5-div.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700476 "src/qs8-gemm/gen/8x8c4-minmax-scalar.c",
477 "src/qs8-gemm/gen/12x4c4-minmax-scalar.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -0700478 "src/qs8-requantization/fp32-scalar-lrintf.c",
479 "src/qs8-requantization/fp32-scalar-magic.c",
480 "src/qs8-requantization/precise-scalar-signed64.c",
481 "src/qs8-requantization/precise-scalar-unsigned32.c",
482 "src/qs8-requantization/precise-scalar-unsigned64.c",
483 "src/qs8-requantization/q31-scalar.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -0700484 "src/qu8-avgpool/9p8x-minmax-scalar-c1.c",
485 "src/qu8-avgpool/9x-minmax-scalar-c1.c",
486 "src/qu8-dwconv/up1x9-minmax-scalar.c",
487 "src/qu8-gavgpool/7p7x-minmax-scalar-c1.c",
488 "src/qu8-gavgpool/7x-minmax-scalar-c1.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700489 "src/qu8-gemm/2x2-minmax-scalar.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700490 "src/qu8-gemm/gen/8x8c4-minmax-scalar.c",
491 "src/qu8-gemm/gen/12x4c4-minmax-scalar.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -0700492 "src/qu8-igemm/2x2-minmax-scalar.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -0700493 "src/qu8-requantization/fp32-scalar-lrintf.c",
494 "src/qu8-requantization/fp32-scalar-magic.c",
495 "src/qu8-requantization/precise-scalar-signed64.c",
496 "src/qu8-requantization/precise-scalar-unsigned32.c",
497 "src/qu8-requantization/precise-scalar-unsigned64.c",
498 "src/qu8-requantization/q31-scalar.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -0700499 "src/qu8-vadd/minmax-scalar.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -0700500 "src/u8-clamp/scalar-x4.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700501 "src/u8-lut32norm/scalar.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700502 "src/u8-maxpool/9p8x-minmax-scalar-c1.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700503 "src/u8-rmax/scalar.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700504 "src/x8-lut/scalar.c",
505 "src/x8-zip/x2-scalar.c",
506 "src/x8-zip/x3-scalar.c",
507 "src/x8-zip/x4-scalar.c",
508 "src/x8-zip/xm-scalar.c",
Marat Dukhanad71b9a2020-11-20 00:01:51 -0800509 "src/x32-depthtospace2d-chw2hwc/scalar.c",
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -0700510 "src/x32-fill/scalar-float.c",
511 "src/x32-fill/scalar-int.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700512 "src/x32-packx/x2-scalar.c",
513 "src/x32-packx/x3-scalar.c",
514 "src/x32-packx/x4-scalar.c",
Marat Dukhan63523d42020-05-22 17:07:33 -0700515 "src/x32-pad/scalar-float.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700516 "src/x32-pad/scalar-int.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700517 "src/x32-unpool/scalar.c",
518 "src/x32-zip/x2-scalar.c",
519 "src/x32-zip/x3-scalar.c",
520 "src/x32-zip/x4-scalar.c",
521 "src/x32-zip/xm-scalar.c",
Marat Dukhan048931b2020-11-24 20:53:54 -0800522 "src/xx-copy/memcpy.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700523]
524
Marat Dukhan436ebe62019-12-04 15:10:12 -0800525WASM_UKERNELS = [
Marat Dukhan99936602020-04-11 16:47:01 -0700526 "src/f32-avgpool/9p8x-minmax-wasm-c1.c",
527 "src/f32-avgpool/9x-minmax-wasm-c1.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -0700528 "src/f32-clamp/gen/wasm-x1.c",
529 "src/f32-clamp/gen/wasm-x2.c",
530 "src/f32-clamp/gen/wasm-x4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700531 "src/f32-dwconv/gen/up1x4-minmax-wasm-acc2.c",
532 "src/f32-dwconv/gen/up1x4-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700533 "src/f32-dwconv/gen/up1x4-wasm-acc2.c",
534 "src/f32-dwconv/gen/up1x4-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700535 "src/f32-dwconv/gen/up1x9-minmax-wasm-acc2.c",
536 "src/f32-dwconv/gen/up1x9-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700537 "src/f32-dwconv/gen/up1x9-wasm-acc2.c",
538 "src/f32-dwconv/gen/up1x9-wasm.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700539 "src/f32-dwconv/gen/up1x25-minmax-wasm-acc2.c",
540 "src/f32-dwconv/gen/up1x25-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700541 "src/f32-dwconv/gen/up1x25-wasm-acc2.c",
542 "src/f32-dwconv/gen/up1x25-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700543 "src/f32-dwconv/gen/up2x4-minmax-wasm-acc2.c",
544 "src/f32-dwconv/gen/up2x4-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700545 "src/f32-dwconv/gen/up2x4-wasm-acc2.c",
546 "src/f32-dwconv/gen/up2x4-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700547 "src/f32-dwconv/gen/up2x9-minmax-wasm-acc2.c",
548 "src/f32-dwconv/gen/up2x9-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700549 "src/f32-dwconv/gen/up2x9-wasm-acc2.c",
550 "src/f32-dwconv/gen/up2x9-wasm.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700551 "src/f32-dwconv/gen/up2x25-minmax-wasm-acc2.c",
552 "src/f32-dwconv/gen/up2x25-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700553 "src/f32-dwconv/gen/up2x25-wasm-acc2.c",
554 "src/f32-dwconv/gen/up2x25-wasm.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700555 "src/f32-gavgpool/7p7x-minmax-wasm-c1.c",
556 "src/f32-gavgpool/7x-minmax-wasm-c1.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700557 "src/f32-gemm/gen-inc/1x4inc-minmax-wasm.c",
558 "src/f32-gemm/gen-inc/2x4inc-minmax-wasm.c",
559 "src/f32-gemm/gen-inc/4x4inc-minmax-wasm.c",
560 "src/f32-gemm/gen/1x4-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700561 "src/f32-gemm/gen/1x4-relu-wasm.c",
562 "src/f32-gemm/gen/1x4-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700563 "src/f32-gemm/gen/2x4-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700564 "src/f32-gemm/gen/2x4-relu-wasm.c",
565 "src/f32-gemm/gen/2x4-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700566 "src/f32-gemm/gen/4x2-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700567 "src/f32-gemm/gen/4x2-relu-wasm.c",
568 "src/f32-gemm/gen/4x2-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700569 "src/f32-gemm/gen/4x4-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700570 "src/f32-gemm/gen/4x4-relu-wasm.c",
571 "src/f32-gemm/gen/4x4-wasm.c",
Marat Dukhanefaac272020-07-16 16:58:07 -0700572 "src/f32-hswish/gen/hswish-wasm-x1.c",
573 "src/f32-hswish/gen/hswish-wasm-x2.c",
574 "src/f32-hswish/gen/hswish-wasm-x4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700575 "src/f32-igemm/gen/1x4-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700576 "src/f32-igemm/gen/1x4-relu-wasm.c",
577 "src/f32-igemm/gen/1x4-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700578 "src/f32-igemm/gen/2x4-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700579 "src/f32-igemm/gen/2x4-relu-wasm.c",
580 "src/f32-igemm/gen/2x4-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700581 "src/f32-igemm/gen/4x2-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700582 "src/f32-igemm/gen/4x2-relu-wasm.c",
583 "src/f32-igemm/gen/4x2-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700584 "src/f32-igemm/gen/4x4-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700585 "src/f32-igemm/gen/4x4-relu-wasm.c",
586 "src/f32-igemm/gen/4x4-wasm.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700587 "src/f32-maxpool/9p8x-minmax-wasm-c1.c",
588 "src/f32-pavgpool/9p8x-minmax-wasm-c1.c",
589 "src/f32-pavgpool/9x-minmax-wasm-c1.c",
Marat Dukhan7c1f8082020-06-25 13:26:20 -0700590 "src/f32-prelu/gen/wasm-2x1.c",
591 "src/f32-prelu/gen/wasm-2x4.c",
Frank Barchardfb158e22020-07-15 16:10:10 -0700592 "src/f32-relu/gen/wasm-x1.c",
593 "src/f32-relu/gen/wasm-x2.c",
594 "src/f32-relu/gen/wasm-x4.c",
Frank Barchard37297a62020-08-27 12:15:10 -0700595 "src/f32-relu/gen/wasm-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700596 "src/f32-vbinary/gen/vadd-minmax-wasm-x1.c",
597 "src/f32-vbinary/gen/vadd-minmax-wasm-x2.c",
598 "src/f32-vbinary/gen/vadd-minmax-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700599 "src/f32-vbinary/gen/vadd-minmax-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700600 "src/f32-vbinary/gen/vadd-relu-wasm-x1.c",
601 "src/f32-vbinary/gen/vadd-relu-wasm-x2.c",
602 "src/f32-vbinary/gen/vadd-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700603 "src/f32-vbinary/gen/vadd-relu-wasm-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700604 "src/f32-vbinary/gen/vaddc-minmax-wasm-x1.c",
605 "src/f32-vbinary/gen/vaddc-minmax-wasm-x2.c",
606 "src/f32-vbinary/gen/vaddc-minmax-wasm-x4.c",
607 "src/f32-vbinary/gen/vaddc-minmax-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700608 "src/f32-vbinary/gen/vaddc-relu-wasm-x1.c",
609 "src/f32-vbinary/gen/vaddc-relu-wasm-x2.c",
610 "src/f32-vbinary/gen/vaddc-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700611 "src/f32-vbinary/gen/vaddc-relu-wasm-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700612 "src/f32-vbinary/gen/vdiv-minmax-wasm-x1.c",
613 "src/f32-vbinary/gen/vdiv-minmax-wasm-x2.c",
614 "src/f32-vbinary/gen/vdiv-minmax-wasm-x4.c",
615 "src/f32-vbinary/gen/vdiv-minmax-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700616 "src/f32-vbinary/gen/vdiv-relu-wasm-x1.c",
617 "src/f32-vbinary/gen/vdiv-relu-wasm-x2.c",
618 "src/f32-vbinary/gen/vdiv-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700619 "src/f32-vbinary/gen/vdiv-relu-wasm-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700620 "src/f32-vbinary/gen/vdivc-minmax-wasm-x1.c",
621 "src/f32-vbinary/gen/vdivc-minmax-wasm-x2.c",
622 "src/f32-vbinary/gen/vdivc-minmax-wasm-x4.c",
623 "src/f32-vbinary/gen/vdivc-minmax-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700624 "src/f32-vbinary/gen/vdivc-relu-wasm-x1.c",
625 "src/f32-vbinary/gen/vdivc-relu-wasm-x2.c",
626 "src/f32-vbinary/gen/vdivc-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700627 "src/f32-vbinary/gen/vdivc-relu-wasm-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800628 "src/f32-vbinary/gen/vmax-wasm-x1.c",
629 "src/f32-vbinary/gen/vmax-wasm-x2.c",
630 "src/f32-vbinary/gen/vmax-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700631 "src/f32-vbinary/gen/vmax-wasm-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800632 "src/f32-vbinary/gen/vmaxc-wasm-x1.c",
633 "src/f32-vbinary/gen/vmaxc-wasm-x2.c",
634 "src/f32-vbinary/gen/vmaxc-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700635 "src/f32-vbinary/gen/vmaxc-wasm-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800636 "src/f32-vbinary/gen/vmin-wasm-x1.c",
637 "src/f32-vbinary/gen/vmin-wasm-x2.c",
638 "src/f32-vbinary/gen/vmin-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700639 "src/f32-vbinary/gen/vmin-wasm-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800640 "src/f32-vbinary/gen/vminc-wasm-x1.c",
641 "src/f32-vbinary/gen/vminc-wasm-x2.c",
642 "src/f32-vbinary/gen/vminc-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700643 "src/f32-vbinary/gen/vminc-wasm-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700644 "src/f32-vbinary/gen/vmul-minmax-wasm-x1.c",
645 "src/f32-vbinary/gen/vmul-minmax-wasm-x2.c",
646 "src/f32-vbinary/gen/vmul-minmax-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700647 "src/f32-vbinary/gen/vmul-minmax-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700648 "src/f32-vbinary/gen/vmul-relu-wasm-x1.c",
649 "src/f32-vbinary/gen/vmul-relu-wasm-x2.c",
650 "src/f32-vbinary/gen/vmul-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700651 "src/f32-vbinary/gen/vmul-relu-wasm-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700652 "src/f32-vbinary/gen/vmulc-minmax-wasm-x1.c",
653 "src/f32-vbinary/gen/vmulc-minmax-wasm-x2.c",
654 "src/f32-vbinary/gen/vmulc-minmax-wasm-x4.c",
655 "src/f32-vbinary/gen/vmulc-minmax-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700656 "src/f32-vbinary/gen/vmulc-relu-wasm-x1.c",
657 "src/f32-vbinary/gen/vmulc-relu-wasm-x2.c",
658 "src/f32-vbinary/gen/vmulc-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700659 "src/f32-vbinary/gen/vmulc-relu-wasm-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700660 "src/f32-vbinary/gen/vrdivc-minmax-wasm-x1.c",
661 "src/f32-vbinary/gen/vrdivc-minmax-wasm-x2.c",
662 "src/f32-vbinary/gen/vrdivc-minmax-wasm-x4.c",
663 "src/f32-vbinary/gen/vrdivc-minmax-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700664 "src/f32-vbinary/gen/vrdivc-relu-wasm-x1.c",
665 "src/f32-vbinary/gen/vrdivc-relu-wasm-x2.c",
666 "src/f32-vbinary/gen/vrdivc-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700667 "src/f32-vbinary/gen/vrdivc-relu-wasm-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700668 "src/f32-vbinary/gen/vrsubc-minmax-wasm-x1.c",
669 "src/f32-vbinary/gen/vrsubc-minmax-wasm-x2.c",
670 "src/f32-vbinary/gen/vrsubc-minmax-wasm-x4.c",
671 "src/f32-vbinary/gen/vrsubc-minmax-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700672 "src/f32-vbinary/gen/vrsubc-relu-wasm-x1.c",
673 "src/f32-vbinary/gen/vrsubc-relu-wasm-x2.c",
674 "src/f32-vbinary/gen/vrsubc-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700675 "src/f32-vbinary/gen/vrsubc-relu-wasm-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700676 "src/f32-vbinary/gen/vsub-minmax-wasm-x1.c",
677 "src/f32-vbinary/gen/vsub-minmax-wasm-x2.c",
678 "src/f32-vbinary/gen/vsub-minmax-wasm-x4.c",
679 "src/f32-vbinary/gen/vsub-minmax-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700680 "src/f32-vbinary/gen/vsub-relu-wasm-x1.c",
681 "src/f32-vbinary/gen/vsub-relu-wasm-x2.c",
682 "src/f32-vbinary/gen/vsub-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700683 "src/f32-vbinary/gen/vsub-relu-wasm-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700684 "src/f32-vbinary/gen/vsubc-minmax-wasm-x1.c",
685 "src/f32-vbinary/gen/vsubc-minmax-wasm-x2.c",
686 "src/f32-vbinary/gen/vsubc-minmax-wasm-x4.c",
687 "src/f32-vbinary/gen/vsubc-minmax-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700688 "src/f32-vbinary/gen/vsubc-relu-wasm-x1.c",
689 "src/f32-vbinary/gen/vsubc-relu-wasm-x2.c",
690 "src/f32-vbinary/gen/vsubc-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700691 "src/f32-vbinary/gen/vsubc-relu-wasm-x8.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -0700692 "src/f32-vlrelu/gen/vlrelu-wasm-x1.c",
693 "src/f32-vlrelu/gen/vlrelu-wasm-x2.c",
694 "src/f32-vlrelu/gen/vlrelu-wasm-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700695 "src/f32-vmulcaddc/gen/c1-minmax-wasm-2x.c",
696 "src/f32-vmulcaddc/gen/c2-minmax-wasm-2x.c",
697 "src/f32-vmulcaddc/gen/c4-minmax-wasm-2x.c",
Marat Dukhan436ebe62019-12-04 15:10:12 -0800698]
699
Marat Dukhan290055c2020-06-09 12:24:29 -0700700WASMSIMD_UKERNELS = [
Marat Dukhan40f05522020-07-16 22:33:12 -0700701 "src/f32-argmaxpool/4x-wasmsimd-c4.c",
702 "src/f32-argmaxpool/9p8x-wasmsimd-c4.c",
703 "src/f32-argmaxpool/9x-wasmsimd-c4.c",
Marat Dukhan3b7432d2020-07-16 17:46:32 -0700704 "src/f32-avgpool/9p8x-minmax-wasmsimd-arm-c4.c",
705 "src/f32-avgpool/9p8x-minmax-wasmsimd-x86-c4.c",
706 "src/f32-avgpool/9x-minmax-wasmsimd-arm-c4.c",
707 "src/f32-avgpool/9x-minmax-wasmsimd-x86-c4.c",
Marat Dukhan3fa52c82020-07-08 12:54:33 -0700708 "src/f32-clamp/gen/wasmsimd-arm-x4.c",
709 "src/f32-clamp/gen/wasmsimd-arm-x8.c",
710 "src/f32-clamp/gen/wasmsimd-x86-x4.c",
711 "src/f32-clamp/gen/wasmsimd-x86-x8.c",
Frank Barchard22136062020-11-24 18:44:46 -0800712 "src/f32-conv-hwc2chw/3x3s2p1c3x4-wasmsimd-2x2.c",
Marat Dukhanac014d72020-06-16 08:36:47 -0700713 "src/f32-dwconv/gen/up4x4-minmax-wasmsimd-acc2-arm.c",
Marat Dukhanac014d72020-06-16 08:36:47 -0700714 "src/f32-dwconv/gen/up4x4-minmax-wasmsimd-acc2-x86.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700715 "src/f32-dwconv/gen/up4x4-minmax-wasmsimd-arm.c",
Marat Dukhanac014d72020-06-16 08:36:47 -0700716 "src/f32-dwconv/gen/up4x4-minmax-wasmsimd-x86.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700717 "src/f32-dwconv/gen/up4x4-wasmsimd.c",
718 "src/f32-dwconv/gen/up4x9-minmax-wasmsimd-acc2-arm.c",
Marat Dukhanac014d72020-06-16 08:36:47 -0700719 "src/f32-dwconv/gen/up4x9-minmax-wasmsimd-acc2-x86.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700720 "src/f32-dwconv/gen/up4x9-minmax-wasmsimd-arm.c",
Marat Dukhanac014d72020-06-16 08:36:47 -0700721 "src/f32-dwconv/gen/up4x9-minmax-wasmsimd-x86.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700722 "src/f32-dwconv/gen/up4x9-wasmsimd.c",
723 "src/f32-dwconv/gen/up4x25-minmax-wasmsimd-acc2-arm.c",
724 "src/f32-dwconv/gen/up4x25-minmax-wasmsimd-acc2-x86.c",
725 "src/f32-dwconv/gen/up4x25-minmax-wasmsimd-arm.c",
726 "src/f32-dwconv/gen/up4x25-minmax-wasmsimd-x86.c",
727 "src/f32-dwconv/gen/up4x25-wasmsimd.c",
728 "src/f32-dwconv/gen/up8x4-minmax-wasmsimd-acc2-arm.c",
Marat Dukhanac014d72020-06-16 08:36:47 -0700729 "src/f32-dwconv/gen/up8x4-minmax-wasmsimd-acc2-x86.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700730 "src/f32-dwconv/gen/up8x4-minmax-wasmsimd-arm.c",
Marat Dukhanac014d72020-06-16 08:36:47 -0700731 "src/f32-dwconv/gen/up8x4-minmax-wasmsimd-x86.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700732 "src/f32-dwconv/gen/up8x4-wasmsimd.c",
733 "src/f32-dwconv/gen/up8x9-minmax-wasmsimd-acc2-arm.c",
Marat Dukhanac014d72020-06-16 08:36:47 -0700734 "src/f32-dwconv/gen/up8x9-minmax-wasmsimd-acc2-x86.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700735 "src/f32-dwconv/gen/up8x9-minmax-wasmsimd-arm.c",
Marat Dukhanac014d72020-06-16 08:36:47 -0700736 "src/f32-dwconv/gen/up8x9-minmax-wasmsimd-x86.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700737 "src/f32-dwconv/gen/up8x9-wasmsimd.c",
738 "src/f32-dwconv/gen/up8x25-minmax-wasmsimd-acc2-arm.c",
739 "src/f32-dwconv/gen/up8x25-minmax-wasmsimd-acc2-x86.c",
740 "src/f32-dwconv/gen/up8x25-minmax-wasmsimd-arm.c",
741 "src/f32-dwconv/gen/up8x25-minmax-wasmsimd-x86.c",
742 "src/f32-dwconv/gen/up8x25-wasmsimd.c",
Frank Barchard3b800452020-11-22 12:12:35 -0800743 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-1x4-acc2.c",
744 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-1x4-acc3.c",
745 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-1x4-acc4.c",
746 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-1x4.c",
747 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-2x4-acc2.c",
748 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-2x4.c",
749 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-3x4.c",
750 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-4x4.c",
751 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-5x4.c",
752 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-6x4.c",
753 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-1x4-acc2.c",
754 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-1x4-acc3.c",
755 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-1x4-acc4.c",
756 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-1x4.c",
757 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-2x4-acc2.c",
758 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-2x4.c",
759 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-3x4.c",
760 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-4x4.c",
761 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-5x4.c",
762 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-6x4.c",
Frank Barchard9cef5ea2020-11-18 14:52:08 -0800763 "src/f32-dwconv2d-chw/gen/3x3s2p1-wasmsimd-1x4-acc3-arm.c",
Frank Barcharddb5c32d2020-11-16 23:58:42 -0800764 "src/f32-dwconv2d-chw/gen/3x3s2p1-wasmsimd-1x4-acc3-x86.c",
Frank Barchard9cef5ea2020-11-18 14:52:08 -0800765 "src/f32-dwconv2d-chw/gen/5x5p2-wasmsimd-3x4-arm.c",
Frank Barcharddb5c32d2020-11-16 23:58:42 -0800766 "src/f32-dwconv2d-chw/gen/5x5p2-wasmsimd-3x4-x86.c",
Frank Barchard9cef5ea2020-11-18 14:52:08 -0800767 "src/f32-dwconv2d-chw/gen/5x5s2p2-wasmsimd-1x4-acc2-arm.c",
Frank Barcharddb5c32d2020-11-16 23:58:42 -0800768 "src/f32-dwconv2d-chw/gen/5x5s2p2-wasmsimd-1x4-acc2-x86.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700769 "src/f32-gavgpool-cw/wasmsimd-arm-x4.c",
770 "src/f32-gavgpool-cw/wasmsimd-x86-x4.c",
Marat Dukhanc6016802020-07-16 18:51:28 -0700771 "src/f32-gavgpool/7p7x-minmax-wasmsimd-arm-c4.c",
772 "src/f32-gavgpool/7p7x-minmax-wasmsimd-x86-c4.c",
773 "src/f32-gavgpool/7x-minmax-wasmsimd-arm-c4.c",
774 "src/f32-gavgpool/7x-minmax-wasmsimd-x86-c4.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700775 "src/f32-gemm/gen-inc/1x8inc-minmax-wasmsimd-loadsplat-arm.c",
776 "src/f32-gemm/gen-inc/1x8inc-minmax-wasmsimd-loadsplat-x86.c",
777 "src/f32-gemm/gen-inc/1x8inc-minmax-wasmsimd-splat-arm.c",
778 "src/f32-gemm/gen-inc/1x8inc-minmax-wasmsimd-splat-x86.c",
779 "src/f32-gemm/gen-inc/1x8s4inc-minmax-wasmsimd-arm.c",
780 "src/f32-gemm/gen-inc/1x8s4inc-minmax-wasmsimd-x86.c",
781 "src/f32-gemm/gen-inc/3x8inc-minmax-wasmsimd-loadsplat-arm.c",
782 "src/f32-gemm/gen-inc/3x8inc-minmax-wasmsimd-loadsplat-x86.c",
783 "src/f32-gemm/gen-inc/3x8inc-minmax-wasmsimd-splat-arm.c",
784 "src/f32-gemm/gen-inc/3x8inc-minmax-wasmsimd-splat-x86.c",
785 "src/f32-gemm/gen-inc/3x8s4inc-minmax-wasmsimd-arm.c",
786 "src/f32-gemm/gen-inc/3x8s4inc-minmax-wasmsimd-x86.c",
787 "src/f32-gemm/gen-inc/4x8inc-minmax-wasmsimd-loadsplat-arm.c",
788 "src/f32-gemm/gen-inc/4x8inc-minmax-wasmsimd-loadsplat-x86.c",
789 "src/f32-gemm/gen-inc/4x8inc-minmax-wasmsimd-splat-arm.c",
790 "src/f32-gemm/gen-inc/4x8inc-minmax-wasmsimd-splat-x86.c",
791 "src/f32-gemm/gen-inc/4x8s4inc-minmax-wasmsimd-arm.c",
792 "src/f32-gemm/gen-inc/4x8s4inc-minmax-wasmsimd-x86.c",
793 "src/f32-gemm/gen-inc/5x8inc-minmax-wasmsimd-loadsplat-arm.c",
794 "src/f32-gemm/gen-inc/5x8inc-minmax-wasmsimd-loadsplat-x86.c",
795 "src/f32-gemm/gen-inc/5x8inc-minmax-wasmsimd-splat-arm.c",
796 "src/f32-gemm/gen-inc/5x8inc-minmax-wasmsimd-splat-x86.c",
797 "src/f32-gemm/gen-inc/5x8s4inc-minmax-wasmsimd-arm.c",
798 "src/f32-gemm/gen-inc/5x8s4inc-minmax-wasmsimd-x86.c",
799 "src/f32-gemm/gen-inc/6x8inc-minmax-wasmsimd-loadsplat-arm.c",
800 "src/f32-gemm/gen-inc/6x8inc-minmax-wasmsimd-loadsplat-x86.c",
801 "src/f32-gemm/gen-inc/6x8inc-minmax-wasmsimd-splat-arm.c",
802 "src/f32-gemm/gen-inc/6x8inc-minmax-wasmsimd-splat-x86.c",
803 "src/f32-gemm/gen-inc/6x8s4inc-minmax-wasmsimd-arm.c",
804 "src/f32-gemm/gen-inc/6x8s4inc-minmax-wasmsimd-x86.c",
805 "src/f32-gemm/gen/1x8-minmax-wasmsimd-loadsplat-arm.c",
806 "src/f32-gemm/gen/1x8-minmax-wasmsimd-loadsplat-x86.c",
807 "src/f32-gemm/gen/1x8-minmax-wasmsimd-splat-arm.c",
808 "src/f32-gemm/gen/1x8-minmax-wasmsimd-splat-x86.c",
809 "src/f32-gemm/gen/1x8-relu-wasmsimd-splat.c",
810 "src/f32-gemm/gen/1x8-wasmsimd-splat.c",
811 "src/f32-gemm/gen/1x8s4-minmax-wasmsimd-arm.c",
812 "src/f32-gemm/gen/1x8s4-minmax-wasmsimd-x86.c",
813 "src/f32-gemm/gen/3x8-minmax-wasmsimd-loadsplat-arm.c",
814 "src/f32-gemm/gen/3x8-minmax-wasmsimd-loadsplat-x86.c",
815 "src/f32-gemm/gen/3x8-minmax-wasmsimd-splat-arm.c",
816 "src/f32-gemm/gen/3x8-minmax-wasmsimd-splat-x86.c",
817 "src/f32-gemm/gen/3x8s4-minmax-wasmsimd-arm.c",
818 "src/f32-gemm/gen/3x8s4-minmax-wasmsimd-x86.c",
819 "src/f32-gemm/gen/4x2c4-minmax-wasmsimd-arm.c",
820 "src/f32-gemm/gen/4x2c4-minmax-wasmsimd-x86.c",
821 "src/f32-gemm/gen/4x2c4-relu-wasmsimd.c",
822 "src/f32-gemm/gen/4x2c4-wasmsimd.c",
823 "src/f32-gemm/gen/4x8-minmax-wasmsimd-loadsplat-arm.c",
824 "src/f32-gemm/gen/4x8-minmax-wasmsimd-loadsplat-x86.c",
825 "src/f32-gemm/gen/4x8-minmax-wasmsimd-splat-arm.c",
826 "src/f32-gemm/gen/4x8-minmax-wasmsimd-splat-x86.c",
827 "src/f32-gemm/gen/4x8-relu-wasmsimd-splat.c",
828 "src/f32-gemm/gen/4x8-wasmsimd-splat.c",
829 "src/f32-gemm/gen/4x8s4-minmax-wasmsimd-arm.c",
830 "src/f32-gemm/gen/4x8s4-minmax-wasmsimd-x86.c",
831 "src/f32-gemm/gen/5x8-minmax-wasmsimd-loadsplat-arm.c",
832 "src/f32-gemm/gen/5x8-minmax-wasmsimd-loadsplat-x86.c",
833 "src/f32-gemm/gen/5x8-minmax-wasmsimd-splat-arm.c",
834 "src/f32-gemm/gen/5x8-minmax-wasmsimd-splat-x86.c",
835 "src/f32-gemm/gen/5x8-relu-wasmsimd-splat.c",
836 "src/f32-gemm/gen/5x8-wasmsimd-splat.c",
837 "src/f32-gemm/gen/5x8s4-minmax-wasmsimd-arm.c",
838 "src/f32-gemm/gen/5x8s4-minmax-wasmsimd-x86.c",
839 "src/f32-gemm/gen/6x8-minmax-wasmsimd-loadsplat-arm.c",
840 "src/f32-gemm/gen/6x8-minmax-wasmsimd-loadsplat-x86.c",
841 "src/f32-gemm/gen/6x8-minmax-wasmsimd-splat-arm.c",
842 "src/f32-gemm/gen/6x8-minmax-wasmsimd-splat-x86.c",
843 "src/f32-gemm/gen/6x8s4-minmax-wasmsimd-arm.c",
844 "src/f32-gemm/gen/6x8s4-minmax-wasmsimd-x86.c",
Marat Dukhanefaac272020-07-16 16:58:07 -0700845 "src/f32-hswish/gen/hswish-wasmsimd-x4.c",
846 "src/f32-hswish/gen/hswish-wasmsimd-x8.c",
847 "src/f32-hswish/gen/hswish-wasmsimd-x16.c",
XNNPACK Team965272b2020-10-23 21:10:15 -0700848 "src/f32-ibilinear-chw/gen/wasmsimd-p4.c",
849 "src/f32-ibilinear-chw/gen/wasmsimd-p8.c",
Frank Barchard35db7d02020-10-26 13:37:34 -0700850 "src/f32-ibilinear/gen/wasmsimd-c4.c",
851 "src/f32-ibilinear/gen/wasmsimd-c8.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700852 "src/f32-igemm/gen/1x8-minmax-wasmsimd-loadsplat-arm.c",
853 "src/f32-igemm/gen/1x8-minmax-wasmsimd-loadsplat-x86.c",
854 "src/f32-igemm/gen/1x8-minmax-wasmsimd-splat-arm.c",
855 "src/f32-igemm/gen/1x8-minmax-wasmsimd-splat-x86.c",
856 "src/f32-igemm/gen/1x8-relu-wasmsimd-splat.c",
857 "src/f32-igemm/gen/1x8-wasmsimd-splat.c",
858 "src/f32-igemm/gen/1x8s4-minmax-wasmsimd-arm.c",
859 "src/f32-igemm/gen/1x8s4-minmax-wasmsimd-x86.c",
860 "src/f32-igemm/gen/3x8-minmax-wasmsimd-loadsplat-arm.c",
861 "src/f32-igemm/gen/3x8-minmax-wasmsimd-loadsplat-x86.c",
862 "src/f32-igemm/gen/3x8-minmax-wasmsimd-splat-arm.c",
863 "src/f32-igemm/gen/3x8-minmax-wasmsimd-splat-x86.c",
864 "src/f32-igemm/gen/3x8s4-minmax-wasmsimd-arm.c",
865 "src/f32-igemm/gen/3x8s4-minmax-wasmsimd-x86.c",
866 "src/f32-igemm/gen/4x2c4-minmax-wasmsimd-arm.c",
867 "src/f32-igemm/gen/4x2c4-minmax-wasmsimd-x86.c",
868 "src/f32-igemm/gen/4x2c4-relu-wasmsimd.c",
869 "src/f32-igemm/gen/4x2c4-wasmsimd.c",
870 "src/f32-igemm/gen/4x8-minmax-wasmsimd-loadsplat-arm.c",
871 "src/f32-igemm/gen/4x8-minmax-wasmsimd-loadsplat-x86.c",
872 "src/f32-igemm/gen/4x8-minmax-wasmsimd-splat-arm.c",
873 "src/f32-igemm/gen/4x8-minmax-wasmsimd-splat-x86.c",
874 "src/f32-igemm/gen/4x8-relu-wasmsimd-splat.c",
875 "src/f32-igemm/gen/4x8-wasmsimd-splat.c",
876 "src/f32-igemm/gen/4x8s4-minmax-wasmsimd-arm.c",
877 "src/f32-igemm/gen/4x8s4-minmax-wasmsimd-x86.c",
878 "src/f32-igemm/gen/5x8-minmax-wasmsimd-loadsplat-arm.c",
879 "src/f32-igemm/gen/5x8-minmax-wasmsimd-loadsplat-x86.c",
880 "src/f32-igemm/gen/5x8-minmax-wasmsimd-splat-arm.c",
881 "src/f32-igemm/gen/5x8-minmax-wasmsimd-splat-x86.c",
882 "src/f32-igemm/gen/5x8-relu-wasmsimd-splat.c",
883 "src/f32-igemm/gen/5x8-wasmsimd-splat.c",
884 "src/f32-igemm/gen/5x8s4-minmax-wasmsimd-arm.c",
885 "src/f32-igemm/gen/5x8s4-minmax-wasmsimd-x86.c",
886 "src/f32-igemm/gen/6x8-minmax-wasmsimd-loadsplat-arm.c",
887 "src/f32-igemm/gen/6x8-minmax-wasmsimd-loadsplat-x86.c",
888 "src/f32-igemm/gen/6x8-minmax-wasmsimd-splat-arm.c",
889 "src/f32-igemm/gen/6x8-minmax-wasmsimd-splat-x86.c",
890 "src/f32-igemm/gen/6x8s4-minmax-wasmsimd-arm.c",
891 "src/f32-igemm/gen/6x8s4-minmax-wasmsimd-x86.c",
Marat Dukhanf6e24802020-07-08 22:20:40 -0700892 "src/f32-maxpool/9p8x-minmax-wasmsimd-arm-c4.c",
893 "src/f32-maxpool/9p8x-minmax-wasmsimd-x86-c4.c",
Marat Dukhan1483c532020-07-16 18:08:19 -0700894 "src/f32-pavgpool/9p8x-minmax-wasmsimd-arm-c4.c",
895 "src/f32-pavgpool/9p8x-minmax-wasmsimd-x86-c4.c",
896 "src/f32-pavgpool/9x-minmax-wasmsimd-arm-c4.c",
897 "src/f32-pavgpool/9x-minmax-wasmsimd-x86-c4.c",
Marat Dukhan0d0d8822020-07-23 23:37:56 -0700898 "src/f32-ppmm/gen/4x8-minmax-wasmsimd-splat-arm.c",
899 "src/f32-ppmm/gen/4x8-minmax-wasmsimd-splat-x86.c",
Frank Barcharda5316982020-07-23 13:19:28 -0700900 "src/f32-prelu/gen/wasmsimd-bitselect-1x4.c",
901 "src/f32-prelu/gen/wasmsimd-bitselect-1x8.c",
902 "src/f32-prelu/gen/wasmsimd-bitselect-1x16.c",
Marat Dukhan195f8eb2020-06-25 12:50:57 -0700903 "src/f32-prelu/gen/wasmsimd-bitselect-2x4.c",
904 "src/f32-prelu/gen/wasmsimd-bitselect-2x8.c",
Frank Barcharda5316982020-07-23 13:19:28 -0700905 "src/f32-prelu/gen/wasmsimd-bitselect-2x16.c",
906 "src/f32-prelu/gen/wasmsimd-bitselect-4x4.c",
907 "src/f32-prelu/gen/wasmsimd-bitselect-4x8.c",
908 "src/f32-prelu/gen/wasmsimd-bitselect-4x16.c",
909 "src/f32-prelu/gen/wasmsimd-minmax-1x4.c",
910 "src/f32-prelu/gen/wasmsimd-minmax-1x8.c",
911 "src/f32-prelu/gen/wasmsimd-minmax-1x16.c",
Marat Dukhan195f8eb2020-06-25 12:50:57 -0700912 "src/f32-prelu/gen/wasmsimd-minmax-2x4.c",
913 "src/f32-prelu/gen/wasmsimd-minmax-2x8.c",
Frank Barcharda5316982020-07-23 13:19:28 -0700914 "src/f32-prelu/gen/wasmsimd-minmax-2x16.c",
915 "src/f32-prelu/gen/wasmsimd-minmax-4x4.c",
916 "src/f32-prelu/gen/wasmsimd-minmax-4x8.c",
917 "src/f32-prelu/gen/wasmsimd-minmax-4x16.c",
Marat Dukhan52238f02020-07-16 15:30:28 -0700918 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x4.c",
Marat Dukhan52238f02020-07-16 15:30:28 -0700919 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x8-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700920 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x8.c",
Marat Dukhan52238f02020-07-16 15:30:28 -0700921 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x12-acc2.c",
922 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x12-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700923 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x12.c",
Marat Dukhan52238f02020-07-16 15:30:28 -0700924 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x16-acc2.c",
925 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x16-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700926 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x16.c",
Marat Dukhan52238f02020-07-16 15:30:28 -0700927 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x20-acc2.c",
928 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x20-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700929 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x20.c",
Frank Barchardfb158e22020-07-15 16:10:10 -0700930 "src/f32-relu/gen/wasmsimd-x4.c",
931 "src/f32-relu/gen/wasmsimd-x8.c",
Frank Barchard37297a62020-08-27 12:15:10 -0700932 "src/f32-relu/gen/wasmsimd-x16.c",
Marat Dukhan8c417962020-07-08 12:27:50 -0700933 "src/f32-rmax/wasmsimd-arm.c",
934 "src/f32-rmax/wasmsimd-x86.c",
Marat Dukhand187a5b2020-07-20 01:07:17 -0700935 "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x4.c",
936 "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x8.c",
937 "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x12.c",
938 "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x16.c",
939 "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x20.c",
940 "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x24.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700941 "src/f32-sigmoid/gen/wasmsimd-p5-div-x4.c",
942 "src/f32-sigmoid/gen/wasmsimd-p5-div-x8.c",
943 "src/f32-sigmoid/gen/wasmsimd-p5-div-x12.c",
944 "src/f32-sigmoid/gen/wasmsimd-p5-div-x16.c",
945 "src/f32-sigmoid/gen/wasmsimd-p5-div-x20.c",
946 "src/f32-sigmoid/gen/wasmsimd-p5-div-x24.c",
Frank Barchard8ef44cd2020-11-03 12:30:23 -0800947 "src/f32-spmm/gen/4x1-minmax-wasmsimd-arm-pipelined-x2.c",
948 "src/f32-spmm/gen/4x1-minmax-wasmsimd-arm-pipelined.c",
Frank Barchardbeca6522020-10-30 22:34:35 -0700949 "src/f32-spmm/gen/4x1-minmax-wasmsimd-arm-x2.c",
950 "src/f32-spmm/gen/4x1-minmax-wasmsimd-arm-x4.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700951 "src/f32-spmm/gen/4x1-minmax-wasmsimd-arm.c",
Frank Barchard8ef44cd2020-11-03 12:30:23 -0800952 "src/f32-spmm/gen/4x1-minmax-wasmsimd-x86-pipelined-x2.c",
953 "src/f32-spmm/gen/4x1-minmax-wasmsimd-x86-pipelined.c",
Frank Barchardbeca6522020-10-30 22:34:35 -0700954 "src/f32-spmm/gen/4x1-minmax-wasmsimd-x86-x2.c",
955 "src/f32-spmm/gen/4x1-minmax-wasmsimd-x86-x4.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700956 "src/f32-spmm/gen/4x1-minmax-wasmsimd-x86.c",
Frank Barchard8ef44cd2020-11-03 12:30:23 -0800957 "src/f32-spmm/gen/8x1-minmax-wasmsimd-arm-pipelined-x2.c",
958 "src/f32-spmm/gen/8x1-minmax-wasmsimd-arm-pipelined.c",
Frank Barchardbeca6522020-10-30 22:34:35 -0700959 "src/f32-spmm/gen/8x1-minmax-wasmsimd-arm-x2.c",
960 "src/f32-spmm/gen/8x1-minmax-wasmsimd-arm-x4.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700961 "src/f32-spmm/gen/8x1-minmax-wasmsimd-arm.c",
Frank Barchard8ef44cd2020-11-03 12:30:23 -0800962 "src/f32-spmm/gen/8x1-minmax-wasmsimd-x86-pipelined-x2.c",
963 "src/f32-spmm/gen/8x1-minmax-wasmsimd-x86-pipelined.c",
Frank Barchardbeca6522020-10-30 22:34:35 -0700964 "src/f32-spmm/gen/8x1-minmax-wasmsimd-x86-x2.c",
965 "src/f32-spmm/gen/8x1-minmax-wasmsimd-x86-x4.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700966 "src/f32-spmm/gen/8x1-minmax-wasmsimd-x86.c",
Frank Barchard8ef44cd2020-11-03 12:30:23 -0800967 "src/f32-spmm/gen/16x1-minmax-wasmsimd-arm-pipelined-x2.c",
968 "src/f32-spmm/gen/16x1-minmax-wasmsimd-arm-pipelined.c",
Frank Barchardbeca6522020-10-30 22:34:35 -0700969 "src/f32-spmm/gen/16x1-minmax-wasmsimd-arm-x2.c",
970 "src/f32-spmm/gen/16x1-minmax-wasmsimd-arm-x4.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700971 "src/f32-spmm/gen/16x1-minmax-wasmsimd-arm.c",
Frank Barchard8ef44cd2020-11-03 12:30:23 -0800972 "src/f32-spmm/gen/16x1-minmax-wasmsimd-x86-pipelined-x2.c",
973 "src/f32-spmm/gen/16x1-minmax-wasmsimd-x86-pipelined.c",
Frank Barchardbeca6522020-10-30 22:34:35 -0700974 "src/f32-spmm/gen/16x1-minmax-wasmsimd-x86-x2.c",
975 "src/f32-spmm/gen/16x1-minmax-wasmsimd-x86-x4.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700976 "src/f32-spmm/gen/16x1-minmax-wasmsimd-x86.c",
Frank Barchard8ef44cd2020-11-03 12:30:23 -0800977 "src/f32-spmm/gen/32x1-minmax-wasmsimd-arm-pipelined-x2.c",
978 "src/f32-spmm/gen/32x1-minmax-wasmsimd-arm-pipelined.c",
Frank Barchardbeca6522020-10-30 22:34:35 -0700979 "src/f32-spmm/gen/32x1-minmax-wasmsimd-arm-x2.c",
980 "src/f32-spmm/gen/32x1-minmax-wasmsimd-arm-x4.c",
Frank Barchard846c0c62020-10-26 15:01:39 -0700981 "src/f32-spmm/gen/32x1-minmax-wasmsimd-arm.c",
Frank Barchard8ef44cd2020-11-03 12:30:23 -0800982 "src/f32-spmm/gen/32x1-minmax-wasmsimd-x86-pipelined-x2.c",
983 "src/f32-spmm/gen/32x1-minmax-wasmsimd-x86-pipelined.c",
Frank Barchardbeca6522020-10-30 22:34:35 -0700984 "src/f32-spmm/gen/32x1-minmax-wasmsimd-x86-x2.c",
985 "src/f32-spmm/gen/32x1-minmax-wasmsimd-x86-x4.c",
Frank Barchard846c0c62020-10-26 15:01:39 -0700986 "src/f32-spmm/gen/32x1-minmax-wasmsimd-x86.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700987 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-arm-x4.c",
988 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700989 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-arm-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700990 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-x86-x4.c",
991 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700992 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700993 "src/f32-vbinary/gen/vadd-relu-wasmsimd-x4.c",
994 "src/f32-vbinary/gen/vadd-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700995 "src/f32-vbinary/gen/vadd-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700996 "src/f32-vbinary/gen/vadd-wasmsimd-x4.c",
997 "src/f32-vbinary/gen/vadd-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700998 "src/f32-vbinary/gen/vadd-wasmsimd-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700999 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-arm-x4.c",
1000 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001001 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-arm-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001002 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-x86-x4.c",
1003 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001004 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -07001005 "src/f32-vbinary/gen/vaddc-relu-wasmsimd-x4.c",
1006 "src/f32-vbinary/gen/vaddc-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001007 "src/f32-vbinary/gen/vaddc-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07001008 "src/f32-vbinary/gen/vaddc-wasmsimd-x4.c",
1009 "src/f32-vbinary/gen/vaddc-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001010 "src/f32-vbinary/gen/vaddc-wasmsimd-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -07001011 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-arm-x4.c",
1012 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001013 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-arm-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -07001014 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-x86-x4.c",
1015 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001016 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -07001017 "src/f32-vbinary/gen/vdiv-relu-wasmsimd-x4.c",
1018 "src/f32-vbinary/gen/vdiv-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001019 "src/f32-vbinary/gen/vdiv-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07001020 "src/f32-vbinary/gen/vdiv-wasmsimd-x4.c",
1021 "src/f32-vbinary/gen/vdiv-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001022 "src/f32-vbinary/gen/vdiv-wasmsimd-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001023 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-arm-x4.c",
1024 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001025 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-arm-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001026 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-x86-x4.c",
1027 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001028 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -07001029 "src/f32-vbinary/gen/vdivc-relu-wasmsimd-x4.c",
1030 "src/f32-vbinary/gen/vdivc-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001031 "src/f32-vbinary/gen/vdivc-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07001032 "src/f32-vbinary/gen/vdivc-wasmsimd-x4.c",
1033 "src/f32-vbinary/gen/vdivc-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001034 "src/f32-vbinary/gen/vdivc-wasmsimd-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -07001035 "src/f32-vbinary/gen/vmax-wasmsimd-arm-x4.c",
1036 "src/f32-vbinary/gen/vmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001037 "src/f32-vbinary/gen/vmax-wasmsimd-arm-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -07001038 "src/f32-vbinary/gen/vmax-wasmsimd-x86-x4.c",
1039 "src/f32-vbinary/gen/vmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001040 "src/f32-vbinary/gen/vmax-wasmsimd-x86-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001041 "src/f32-vbinary/gen/vmaxc-wasmsimd-arm-x4.c",
1042 "src/f32-vbinary/gen/vmaxc-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001043 "src/f32-vbinary/gen/vmaxc-wasmsimd-arm-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001044 "src/f32-vbinary/gen/vmaxc-wasmsimd-x86-x4.c",
1045 "src/f32-vbinary/gen/vmaxc-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001046 "src/f32-vbinary/gen/vmaxc-wasmsimd-x86-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -07001047 "src/f32-vbinary/gen/vmin-wasmsimd-arm-x4.c",
1048 "src/f32-vbinary/gen/vmin-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001049 "src/f32-vbinary/gen/vmin-wasmsimd-arm-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -07001050 "src/f32-vbinary/gen/vmin-wasmsimd-x86-x4.c",
1051 "src/f32-vbinary/gen/vmin-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001052 "src/f32-vbinary/gen/vmin-wasmsimd-x86-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001053 "src/f32-vbinary/gen/vminc-wasmsimd-arm-x4.c",
1054 "src/f32-vbinary/gen/vminc-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001055 "src/f32-vbinary/gen/vminc-wasmsimd-arm-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001056 "src/f32-vbinary/gen/vminc-wasmsimd-x86-x4.c",
1057 "src/f32-vbinary/gen/vminc-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001058 "src/f32-vbinary/gen/vminc-wasmsimd-x86-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -07001059 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-arm-x4.c",
1060 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001061 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-arm-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -07001062 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-x86-x4.c",
1063 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001064 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -07001065 "src/f32-vbinary/gen/vmul-relu-wasmsimd-x4.c",
1066 "src/f32-vbinary/gen/vmul-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001067 "src/f32-vbinary/gen/vmul-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07001068 "src/f32-vbinary/gen/vmul-wasmsimd-x4.c",
1069 "src/f32-vbinary/gen/vmul-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001070 "src/f32-vbinary/gen/vmul-wasmsimd-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001071 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-arm-x4.c",
1072 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001073 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-arm-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001074 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-x86-x4.c",
1075 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001076 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -07001077 "src/f32-vbinary/gen/vmulc-relu-wasmsimd-x4.c",
1078 "src/f32-vbinary/gen/vmulc-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001079 "src/f32-vbinary/gen/vmulc-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07001080 "src/f32-vbinary/gen/vmulc-wasmsimd-x4.c",
1081 "src/f32-vbinary/gen/vmulc-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001082 "src/f32-vbinary/gen/vmulc-wasmsimd-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001083 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-arm-x4.c",
1084 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001085 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-arm-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001086 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-x86-x4.c",
1087 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001088 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -07001089 "src/f32-vbinary/gen/vrdivc-relu-wasmsimd-x4.c",
1090 "src/f32-vbinary/gen/vrdivc-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001091 "src/f32-vbinary/gen/vrdivc-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07001092 "src/f32-vbinary/gen/vrdivc-wasmsimd-x4.c",
1093 "src/f32-vbinary/gen/vrdivc-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001094 "src/f32-vbinary/gen/vrdivc-wasmsimd-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001095 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-arm-x4.c",
1096 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001097 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-arm-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001098 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-x86-x4.c",
1099 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001100 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -07001101 "src/f32-vbinary/gen/vrsubc-relu-wasmsimd-x4.c",
1102 "src/f32-vbinary/gen/vrsubc-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001103 "src/f32-vbinary/gen/vrsubc-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07001104 "src/f32-vbinary/gen/vrsubc-wasmsimd-x4.c",
1105 "src/f32-vbinary/gen/vrsubc-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001106 "src/f32-vbinary/gen/vrsubc-wasmsimd-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -07001107 "src/f32-vbinary/gen/vsqrdiff-wasmsimd-x4.c",
1108 "src/f32-vbinary/gen/vsqrdiff-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001109 "src/f32-vbinary/gen/vsqrdiff-wasmsimd-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001110 "src/f32-vbinary/gen/vsqrdiffc-wasmsimd-x4.c",
1111 "src/f32-vbinary/gen/vsqrdiffc-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001112 "src/f32-vbinary/gen/vsqrdiffc-wasmsimd-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -07001113 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-arm-x4.c",
1114 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001115 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-arm-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -07001116 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-x86-x4.c",
1117 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001118 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -07001119 "src/f32-vbinary/gen/vsub-relu-wasmsimd-x4.c",
1120 "src/f32-vbinary/gen/vsub-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001121 "src/f32-vbinary/gen/vsub-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07001122 "src/f32-vbinary/gen/vsub-wasmsimd-x4.c",
1123 "src/f32-vbinary/gen/vsub-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001124 "src/f32-vbinary/gen/vsub-wasmsimd-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001125 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-arm-x4.c",
1126 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001127 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-arm-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001128 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-x86-x4.c",
1129 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001130 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -07001131 "src/f32-vbinary/gen/vsubc-relu-wasmsimd-x4.c",
1132 "src/f32-vbinary/gen/vsubc-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001133 "src/f32-vbinary/gen/vsubc-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07001134 "src/f32-vbinary/gen/vsubc-wasmsimd-x4.c",
1135 "src/f32-vbinary/gen/vsubc-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001136 "src/f32-vbinary/gen/vsubc-wasmsimd-x16.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07001137 "src/f32-vlrelu/gen/vlrelu-wasmsimd-bitselect-x4.c",
1138 "src/f32-vlrelu/gen/vlrelu-wasmsimd-bitselect-x8.c",
1139 "src/f32-vlrelu/gen/vlrelu-wasmsimd-minmax-x4.c",
1140 "src/f32-vlrelu/gen/vlrelu-wasmsimd-minmax-x8.c",
Marat Dukhand816f622020-07-15 10:14:39 -07001141 "src/f32-vmulcaddc/gen/c4-minmax-wasmsimd-arm-2x.c",
Marat Dukhand816f622020-07-15 10:14:39 -07001142 "src/f32-vmulcaddc/gen/c4-minmax-wasmsimd-x86-2x.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001143 "src/f32-vmulcaddc/gen/c8-minmax-wasmsimd-arm-2x.c",
Marat Dukhand816f622020-07-15 10:14:39 -07001144 "src/f32-vmulcaddc/gen/c8-minmax-wasmsimd-x86-2x.c",
Marat Dukhanb82b2cd2020-07-16 02:23:42 -07001145 "src/f32-vrnd/gen/vrndd-wasmsimd-addsub-x4.c",
1146 "src/f32-vrnd/gen/vrndd-wasmsimd-addsub-x8.c",
1147 "src/f32-vrnd/gen/vrndd-wasmsimd-cvt-x4.c",
1148 "src/f32-vrnd/gen/vrndd-wasmsimd-cvt-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001149 "src/f32-vrnd/gen/vrndne-wasmsimd-addsub-x4.c",
1150 "src/f32-vrnd/gen/vrndne-wasmsimd-addsub-x8.c",
1151 "src/f32-vrnd/gen/vrndu-wasmsimd-addsub-x4.c",
1152 "src/f32-vrnd/gen/vrndu-wasmsimd-addsub-x8.c",
1153 "src/f32-vrnd/gen/vrndu-wasmsimd-cvt-x4.c",
1154 "src/f32-vrnd/gen/vrndu-wasmsimd-cvt-x8.c",
1155 "src/f32-vrnd/gen/vrndz-wasmsimd-addsub-x4.c",
1156 "src/f32-vrnd/gen/vrndz-wasmsimd-addsub-x8.c",
1157 "src/f32-vrnd/gen/vrndz-wasmsimd-cvt-x4.c",
1158 "src/f32-vrnd/gen/vrndz-wasmsimd-cvt-x8.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07001159 "src/f32-vsqrt/gen/wasmsimd-sqrt-x4.c",
1160 "src/f32-vsqrt/gen/wasmsimd-sqrt-x8.c",
Marat Dukhan37c83512020-06-29 13:25:53 -07001161 "src/f32-vunary/gen/vabs-wasmsimd-x4.c",
1162 "src/f32-vunary/gen/vabs-wasmsimd-x8.c",
1163 "src/f32-vunary/gen/vneg-wasmsimd-x4.c",
1164 "src/f32-vunary/gen/vneg-wasmsimd-x8.c",
1165 "src/f32-vunary/gen/vsqr-wasmsimd-x4.c",
1166 "src/f32-vunary/gen/vsqr-wasmsimd-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001167 "src/math/roundd-wasmsimd-addsub.c",
1168 "src/math/roundd-wasmsimd-cvt.c",
1169 "src/math/roundne-wasmsimd-addsub.c",
1170 "src/math/roundu-wasmsimd-addsub.c",
1171 "src/math/roundu-wasmsimd-cvt.c",
1172 "src/math/roundz-wasmsimd-addsub.c",
1173 "src/math/roundz-wasmsimd-cvt.c",
1174 "src/math/sigmoid-wasmsimd-rr2-lut64-p2-div.c",
1175 "src/math/sigmoid-wasmsimd-rr2-p5-div.c",
Marat Dukhancc8f34c2020-08-05 16:36:38 -07001176 "src/qs8-dwconv/gen/up8x9-minmax-wasmsimd-mul16.c",
1177 "src/qs8-dwconv/gen/up16x9-minmax-wasmsimd-mul16.c",
1178 "src/qs8-dwconv/gen/up24x9-minmax-wasmsimd-mul16.c",
Marat Dukhanb5e3d172020-08-06 13:29:53 -07001179 "src/qs8-gavgpool/gen/7p7x-minmax-wasmsimd-c8-acc2.c",
1180 "src/qs8-gavgpool/gen/7p7x-minmax-wasmsimd-c16-acc2.c",
1181 "src/qs8-gavgpool/gen/7p7x-minmax-wasmsimd-c24-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001182 "src/qs8-gavgpool/gen/7x-minmax-wasmsimd-c8-acc2.c",
1183 "src/qs8-gavgpool/gen/7x-minmax-wasmsimd-c16-acc2.c",
1184 "src/qs8-gavgpool/gen/7x-minmax-wasmsimd-c24-acc2.c",
Marat Dukhan27203da2020-08-05 15:19:03 -07001185 "src/qs8-gemm/gen/1x4c8-minmax-wasmsimd-ld64.c",
Marat Dukhan27203da2020-08-05 15:19:03 -07001186 "src/qs8-gemm/gen/1x4c8-minmax-wasmsimd-ld128.c",
Marat Dukhan27203da2020-08-05 15:19:03 -07001187 "src/qs8-gemm/gen/1x4c8-xw-minmax-wasmsimd.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001188 "src/qs8-gemm/gen/2x4c8-minmax-wasmsimd-ld64.c",
1189 "src/qs8-gemm/gen/2x4c8-minmax-wasmsimd-ld128.c",
Marat Dukhan27203da2020-08-05 15:19:03 -07001190 "src/qs8-gemm/gen/2x4c8-xw-minmax-wasmsimd.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001191 "src/qs8-gemm/gen/3x4c8-minmax-wasmsimd-ld64.c",
1192 "src/qs8-gemm/gen/3x4c8-minmax-wasmsimd-ld128.c",
Marat Dukhan27203da2020-08-05 15:19:03 -07001193 "src/qs8-gemm/gen/3x4c8-xw-minmax-wasmsimd.c",
1194 "src/qs8-igemm/gen/1x4c8-minmax-wasmsimd-ld64.c",
Marat Dukhan27203da2020-08-05 15:19:03 -07001195 "src/qs8-igemm/gen/1x4c8-minmax-wasmsimd-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001196 "src/qs8-igemm/gen/2x4c8-minmax-wasmsimd-ld64.c",
Marat Dukhan27203da2020-08-05 15:19:03 -07001197 "src/qs8-igemm/gen/2x4c8-minmax-wasmsimd-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001198 "src/qs8-igemm/gen/3x4c8-minmax-wasmsimd-ld64.c",
Marat Dukhan27203da2020-08-05 15:19:03 -07001199 "src/qs8-igemm/gen/3x4c8-minmax-wasmsimd-ld128.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -07001200 "src/qs8-requantization/fp32-wasmsimd.c",
Marat Dukhan138560c2020-08-03 18:57:34 -07001201 "src/qs8-requantization/q31-wasmsimd.c",
Marat Dukhan5df27f82020-09-02 23:59:21 -07001202 "src/qs8-vadd/gen/minmax-wasmsimd-x8.c",
1203 "src/qs8-vadd/gen/minmax-wasmsimd-x16.c",
1204 "src/qs8-vadd/gen/minmax-wasmsimd-x24.c",
1205 "src/qs8-vadd/gen/minmax-wasmsimd-x32.c",
1206 "src/qs8-vaddc/gen/minmax-wasmsimd-x8.c",
1207 "src/qs8-vaddc/gen/minmax-wasmsimd-x16.c",
1208 "src/qs8-vaddc/gen/minmax-wasmsimd-x24.c",
1209 "src/qs8-vaddc/gen/minmax-wasmsimd-x32.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07001210 "src/qu8-requantization/fp32-wasmsimd.c",
Marat Dukhan22de5e72020-08-03 21:53:18 -07001211 "src/qu8-requantization/q31-wasmsimd.c",
Marat Dukhan8ee37012020-07-16 13:17:13 -07001212 "src/x32-fill/wasmsimd.c",
Marat Dukhan66d99e92020-07-16 12:56:21 -07001213 "src/x32-packx/x4-wasmsimd.c",
Marat Dukhan9306ae02020-07-16 15:51:13 -07001214 "src/x32-pad/wasmsimd.c",
Marat Dukhan9d4bfa22020-07-16 19:07:04 -07001215 "src/x32-unpool/wasmsimd.c",
Marat Dukhane3b78762020-07-16 20:02:58 -07001216 "src/x32-zip/x2-wasmsimd.c",
1217 "src/x32-zip/x3-wasmsimd.c",
1218 "src/x32-zip/x4-wasmsimd.c",
1219 "src/x32-zip/xm-wasmsimd.c",
Marat Dukhan290055c2020-06-09 12:24:29 -07001220]
1221
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001222PSIMD_ACCMATH_UKERNELS = [
Marat Dukhan2e23d2b2020-07-29 16:01:37 -07001223 "src/qs8-requantization/fp32-psimd.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001224 "src/qs8-requantization/precise-psimd.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07001225 "src/qu8-requantization/fp32-psimd.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001226 "src/qu8-requantization/precise-psimd.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001227]
1228
Marat Dukhan08c4a432019-10-03 09:29:21 -07001229# ISA-specific micro-kernels
1230NEON_UKERNELS = [
Marat Dukhanef25c6d2020-07-24 00:59:40 -07001231 "src/f32-argmaxpool/4x-neon-c4.c",
1232 "src/f32-argmaxpool/9p8x-neon-c4.c",
1233 "src/f32-argmaxpool/9x-neon-c4.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001234 "src/f32-avgpool/9p8x-minmax-neon-c4.c",
1235 "src/f32-avgpool/9x-minmax-neon-c4.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001236 "src/f32-clamp/gen/neon-x4.c",
1237 "src/f32-clamp/gen/neon-x8.c",
Marat Dukhan56b10cd2020-05-18 09:35:49 -07001238 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neon-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001239 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neon-2x2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001240 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neon-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001241 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neon-2x2.c",
Marat Dukhan56b10cd2020-05-18 09:35:49 -07001242 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neon-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001243 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neon-2x2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001244 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neon-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001245 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neon-2x2.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -07001246 "src/f32-dwconv/gen/up4x4-minmax-neon-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001247 "src/f32-dwconv/gen/up4x4-minmax-neon.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001248 "src/f32-dwconv/gen/up4x9-minmax-neon-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001249 "src/f32-dwconv/gen/up4x9-minmax-neon.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -07001250 "src/f32-dwconv/gen/up4x25-minmax-neon-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001251 "src/f32-dwconv/gen/up4x25-minmax-neon.c",
1252 "src/f32-dwconv/gen/up8x4-minmax-neon-acc2.c",
1253 "src/f32-dwconv/gen/up8x4-minmax-neon.c",
1254 "src/f32-dwconv/gen/up8x9-minmax-neon-acc2.c",
1255 "src/f32-dwconv/gen/up8x9-minmax-neon.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -07001256 "src/f32-dwconv/gen/up8x25-minmax-neon-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001257 "src/f32-dwconv/gen/up8x25-minmax-neon.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001258 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-1x4-acc2.c",
1259 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-1x4-acc3.c",
1260 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-1x4-acc4.c",
Marat Dukhanc581e482020-10-24 01:28:11 -07001261 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-1x4.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001262 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-2x4-acc2.c",
Marat Dukhanc581e482020-10-24 01:28:11 -07001263 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-2x4.c",
1264 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-3x4.c",
1265 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-4x4.c",
1266 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-5x4.c",
1267 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-6x4.c",
Marat Dukhan82f0c322020-10-25 19:17:35 -07001268 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neon-1x4-acc2.c",
1269 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neon-1x4-acc3.c",
1270 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neon-1x4-acc4.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001271 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neon-1x4.c",
Marat Dukhan82f0c322020-10-25 19:17:35 -07001272 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neon-2x4-acc2.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001273 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neon-2x4.c",
1274 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neon-3x4.c",
1275 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neon-4x4.c",
Marat Dukhan149f0ea2020-10-26 12:50:33 -07001276 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-1x4-acc2.c",
1277 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-1x4-acc3.c",
1278 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-1x4-acc4.c",
1279 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-1x4-acc5.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001280 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-1x4.c",
Marat Dukhan149f0ea2020-10-26 12:50:33 -07001281 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-2x4-acc2.c",
1282 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-2x4-acc3.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001283 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-2x4.c",
Marat Dukhan149f0ea2020-10-26 12:50:33 -07001284 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-3x4-acc2.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001285 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-3x4.c",
Marat Dukhan149f0ea2020-10-26 12:50:33 -07001286 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-4x4-acc2.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001287 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-4x4.c",
1288 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-5x4.c",
Marat Dukhan30d4b252020-10-29 16:33:22 -07001289 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neon-1x4-acc2.c",
1290 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neon-1x4-acc3.c",
1291 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neon-1x4-acc4.c",
1292 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neon-1x4-acc5.c",
1293 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neon-1x4.c",
1294 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neon-2x4-acc2.c",
1295 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neon-2x4-acc3.c",
1296 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neon-2x4.c",
Marat Dukhan30d4b252020-10-29 16:33:22 -07001297 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neon-3x4-acc2.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07001298 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neon-3x4.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -07001299 "src/f32-gavgpool-cw/neon-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001300 "src/f32-gavgpool/7p7x-minmax-neon-c4.c",
1301 "src/f32-gavgpool/7x-minmax-neon-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001302 "src/f32-gemm/gen-inc/1x8inc-minmax-neon-dup-ld64.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001303 "src/f32-gemm/gen-inc/1x8inc-minmax-neon-lane-ld64.c",
1304 "src/f32-gemm/gen-inc/1x8s4inc-minmax-neon.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001305 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-dup-ld64.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001306 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-dup-ld128.c",
1307 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-lane-ld64.c",
1308 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-lane-ld128.c",
1309 "src/f32-gemm/gen-inc/4x8s4inc-minmax-neon.c",
1310 "src/f32-gemm/gen-inc/5x8inc-minmax-neon-lane-ld64.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001311 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-dup-ld64.c",
1312 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-dup-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001313 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-lane-ld64.c",
1314 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-lane-ld128.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001315 "src/f32-gemm/gen-inc/6x8s4inc-minmax-neon.c",
1316 "src/f32-gemm/gen-inc/8x8s4inc-minmax-neon.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001317 "src/f32-gemm/gen/1x8-minmax-neon-dup-ld64.c",
1318 "src/f32-gemm/gen/1x8-minmax-neon-lane-ld64.c",
1319 "src/f32-gemm/gen/1x8s4-minmax-neon.c",
1320 "src/f32-gemm/gen/4x2-minmax-neon-lane-ld64.c",
1321 "src/f32-gemm/gen/4x8-minmax-neon-dup-ld64.c",
1322 "src/f32-gemm/gen/4x8-minmax-neon-dup-ld128.c",
1323 "src/f32-gemm/gen/4x8-minmax-neon-lane-ld64.c",
1324 "src/f32-gemm/gen/4x8-minmax-neon-lane-ld128.c",
1325 "src/f32-gemm/gen/4x8s4-minmax-neon.c",
1326 "src/f32-gemm/gen/5x8-minmax-neon-lane-ld64.c",
1327 "src/f32-gemm/gen/6x8-minmax-neon-dup-ld64.c",
1328 "src/f32-gemm/gen/6x8-minmax-neon-dup-ld128.c",
1329 "src/f32-gemm/gen/6x8-minmax-neon-lane-ld64.c",
1330 "src/f32-gemm/gen/6x8-minmax-neon-lane-ld128.c",
1331 "src/f32-gemm/gen/6x8s4-minmax-neon.c",
1332 "src/f32-gemm/gen/8x8s4-minmax-neon.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07001333 "src/f32-hswish/gen/hswish-neon-x4.c",
1334 "src/f32-hswish/gen/hswish-neon-x8.c",
1335 "src/f32-hswish/gen/hswish-neon-x16.c",
Marat Dukhan660fd192020-03-10 04:55:30 -07001336 "src/f32-ibilinear/gen/neon-c4.c",
1337 "src/f32-ibilinear/gen/neon-c8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001338 "src/f32-igemm/gen/1x8-minmax-neon-dup-ld64.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001339 "src/f32-igemm/gen/1x8-minmax-neon-lane-ld64.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001340 "src/f32-igemm/gen/1x8s4-minmax-neon.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001341 "src/f32-igemm/gen/4x2-minmax-neon-lane-ld64.c",
1342 "src/f32-igemm/gen/4x4-minmax-neon-lane-ld64.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001343 "src/f32-igemm/gen/4x8-minmax-neon-dup-ld64.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001344 "src/f32-igemm/gen/4x8-minmax-neon-dup-ld128.c",
1345 "src/f32-igemm/gen/4x8-minmax-neon-lane-ld64.c",
1346 "src/f32-igemm/gen/4x8-minmax-neon-lane-ld128.c",
1347 "src/f32-igemm/gen/4x8s4-minmax-neon.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001348 "src/f32-igemm/gen/6x8-minmax-neon-dup-ld64.c",
1349 "src/f32-igemm/gen/6x8-minmax-neon-dup-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001350 "src/f32-igemm/gen/6x8-minmax-neon-lane-ld64.c",
1351 "src/f32-igemm/gen/6x8-minmax-neon-lane-ld128.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001352 "src/f32-igemm/gen/6x8s4-minmax-neon.c",
1353 "src/f32-igemm/gen/8x8s4-minmax-neon.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001354 "src/f32-maxpool/9p8x-minmax-neon-c4.c",
1355 "src/f32-pavgpool/9p8x-minmax-neon-c4.c",
1356 "src/f32-pavgpool/9x-minmax-neon-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001357 "src/f32-ppmm/gen/4x8-minmax-neon.c",
1358 "src/f32-ppmm/gen/8x8-minmax-neon.c",
Frank Barcharda5316982020-07-23 13:19:28 -07001359 "src/f32-prelu/gen/neon-1x4.c",
1360 "src/f32-prelu/gen/neon-1x8.c",
1361 "src/f32-prelu/gen/neon-1x16.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001362 "src/f32-prelu/gen/neon-2x4.c",
1363 "src/f32-prelu/gen/neon-2x8.c",
Frank Barcharda5316982020-07-23 13:19:28 -07001364 "src/f32-prelu/gen/neon-2x16.c",
1365 "src/f32-prelu/gen/neon-4x4.c",
1366 "src/f32-prelu/gen/neon-4x8.c",
1367 "src/f32-prelu/gen/neon-4x16.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001368 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x4.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001369 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x8-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001370 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x8.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001371 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12-acc2.c",
1372 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001373 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001374 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16-acc2.c",
1375 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001376 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001377 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20-acc2.c",
1378 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001379 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20.c",
1380 "src/f32-raddstoreexpminusmax/gen/neon-p5-x4.c",
1381 "src/f32-raddstoreexpminusmax/gen/neon-p5-x8-acc2.c",
1382 "src/f32-raddstoreexpminusmax/gen/neon-p5-x8.c",
1383 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12-acc2.c",
1384 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12-acc3.c",
1385 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12.c",
1386 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16-acc2.c",
1387 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16-acc4.c",
1388 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16.c",
1389 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20-acc2.c",
1390 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20-acc5.c",
1391 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07001392 "src/f32-relu/gen/neon-x4.c",
1393 "src/f32-relu/gen/neon-x8.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001394 "src/f32-rmax/neon.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001395 "src/f32-sigmoid/gen/neon-frac-p9-p10-nr1recps-x16.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -08001396 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x4.c",
1397 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x8.c",
1398 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x12.c",
1399 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x16.c",
1400 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x20.c",
1401 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x24.c",
1402 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x4.c",
1403 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x8.c",
1404 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x12.c",
1405 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x16.c",
1406 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x20.c",
1407 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x24.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001408 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x4.c",
1409 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x8.c",
1410 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x12.c",
1411 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x16.c",
1412 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x20.c",
1413 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x24.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001414 "src/f32-vbinary/gen/vadd-minmax-neon-x4.c",
1415 "src/f32-vbinary/gen/vadd-minmax-neon-x8.c",
1416 "src/f32-vbinary/gen/vaddc-minmax-neon-x4.c",
1417 "src/f32-vbinary/gen/vaddc-minmax-neon-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -08001418 "src/f32-vbinary/gen/vmax-neon-x4.c",
1419 "src/f32-vbinary/gen/vmax-neon-x8.c",
1420 "src/f32-vbinary/gen/vmaxc-neon-x4.c",
1421 "src/f32-vbinary/gen/vmaxc-neon-x8.c",
1422 "src/f32-vbinary/gen/vmin-neon-x4.c",
1423 "src/f32-vbinary/gen/vmin-neon-x8.c",
1424 "src/f32-vbinary/gen/vminc-neon-x4.c",
1425 "src/f32-vbinary/gen/vminc-neon-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001426 "src/f32-vbinary/gen/vmul-minmax-neon-x4.c",
1427 "src/f32-vbinary/gen/vmul-minmax-neon-x8.c",
1428 "src/f32-vbinary/gen/vmulc-minmax-neon-x4.c",
1429 "src/f32-vbinary/gen/vmulc-minmax-neon-x8.c",
1430 "src/f32-vbinary/gen/vrsubc-minmax-neon-x4.c",
1431 "src/f32-vbinary/gen/vrsubc-minmax-neon-x8.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07001432 "src/f32-vbinary/gen/vsqrdiff-neon-x4.c",
1433 "src/f32-vbinary/gen/vsqrdiff-neon-x8.c",
1434 "src/f32-vbinary/gen/vsqrdiffc-neon-x4.c",
1435 "src/f32-vbinary/gen/vsqrdiffc-neon-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001436 "src/f32-vbinary/gen/vsub-minmax-neon-x4.c",
1437 "src/f32-vbinary/gen/vsub-minmax-neon-x8.c",
1438 "src/f32-vbinary/gen/vsubc-minmax-neon-x4.c",
1439 "src/f32-vbinary/gen/vsubc-minmax-neon-x8.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07001440 "src/f32-vlrelu/gen/vlrelu-neon-x4.c",
1441 "src/f32-vlrelu/gen/vlrelu-neon-x8.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001442 "src/f32-vmulcaddc/gen/c4-minmax-neon-2x.c",
1443 "src/f32-vmulcaddc/gen/c8-minmax-neon-2x.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07001444 "src/f32-vrnd/gen/vrndd-neon-x4.c",
1445 "src/f32-vrnd/gen/vrndd-neon-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001446 "src/f32-vrnd/gen/vrndne-neon-x4.c",
1447 "src/f32-vrnd/gen/vrndne-neon-x8.c",
1448 "src/f32-vrnd/gen/vrndu-neon-x4.c",
1449 "src/f32-vrnd/gen/vrndu-neon-x8.c",
1450 "src/f32-vrnd/gen/vrndz-neon-x4.c",
1451 "src/f32-vrnd/gen/vrndz-neon-x8.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07001452 "src/f32-vunary/gen/vabs-neon-x4.c",
1453 "src/f32-vunary/gen/vabs-neon-x8.c",
1454 "src/f32-vunary/gen/vneg-neon-x4.c",
1455 "src/f32-vunary/gen/vneg-neon-x8.c",
1456 "src/f32-vunary/gen/vsqr-neon-x4.c",
1457 "src/f32-vunary/gen/vsqr-neon-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001458 "src/math/roundd-neon-addsub.c",
1459 "src/math/roundd-neon-cvt.c",
1460 "src/math/roundne-neon-addsub.c",
1461 "src/math/roundu-neon-addsub.c",
1462 "src/math/roundu-neon-cvt.c",
1463 "src/math/roundz-neon-addsub.c",
1464 "src/math/roundz-neon-cvt.c",
1465 "src/math/sigmoid-neon-frac-p9-p10-nr1recps.c",
1466 "src/math/sigmoid-neon-rr1-lut64-p2-nr2recps.c",
1467 "src/math/sigmoid-neon-rr1-lut2048-p1-nr2recps.c",
1468 "src/math/sigmoid-neon-rr1-p5-nr2recps.c",
1469 "src/math/sigmoid-neon-rr2-lut64-p2-nr2recps.c",
1470 "src/math/sigmoid-neon-rr2-lut2048-p1-nr2recps.c",
1471 "src/math/sigmoid-neon-rr2-p5-nr2recps.c",
1472 "src/math/sqrt-neon-nr1rsqrts.c",
1473 "src/math/sqrt-neon-nr2rsqrts.c",
1474 "src/math/sqrt-neon-nr3rsqrts.c",
Marat Dukhan023bcf92020-08-10 12:40:50 -07001475 "src/qs8-dwconv/gen/up8x9-minmax-neon-mul16.c",
1476 "src/qs8-dwconv/gen/up16x9-minmax-neon-mul16.c",
1477 "src/qs8-dwconv/gen/up24x9-minmax-neon-mul16.c",
1478 "src/qs8-dwconv/gen/up32x9-minmax-neon-mul16.c",
Marat Dukhan281262d2020-08-10 13:23:21 -07001479 "src/qs8-gavgpool/gen/7p7x-minmax-neon-c8-acc2.c",
1480 "src/qs8-gavgpool/gen/7p7x-minmax-neon-c16-acc2.c",
1481 "src/qs8-gavgpool/gen/7p7x-minmax-neon-c24-acc2.c",
1482 "src/qs8-gavgpool/gen/7p7x-minmax-neon-c32-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001483 "src/qs8-gavgpool/gen/7x-minmax-neon-c8-acc2.c",
1484 "src/qs8-gavgpool/gen/7x-minmax-neon-c16-acc2.c",
1485 "src/qs8-gavgpool/gen/7x-minmax-neon-c24-acc2.c",
1486 "src/qs8-gavgpool/gen/7x-minmax-neon-c32-acc2.c",
Marat Dukhan40bbafe2020-08-04 02:04:22 -07001487 "src/qs8-gemm/gen/1x8-minmax-neon-mlal-lane.c",
Marat Dukhan40bbafe2020-08-04 02:04:22 -07001488 "src/qs8-gemm/gen/1x16-minmax-neon-mlal-lane.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001489 "src/qs8-gemm/gen/2x8-minmax-neon-mlal-lane.c",
Marat Dukhan40bbafe2020-08-04 02:04:22 -07001490 "src/qs8-gemm/gen/2x16-minmax-neon-mlal-lane.c",
1491 "src/qs8-igemm/gen/1x8-minmax-neon-mlal-lane.c",
Marat Dukhan40bbafe2020-08-04 02:04:22 -07001492 "src/qs8-igemm/gen/1x16-minmax-neon-mlal-lane.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001493 "src/qs8-igemm/gen/2x8-minmax-neon-mlal-lane.c",
Marat Dukhan40bbafe2020-08-04 02:04:22 -07001494 "src/qs8-igemm/gen/2x16-minmax-neon-mlal-lane.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -07001495 "src/qs8-requantization/fp32-neon.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001496 "src/qs8-requantization/precise-neon.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -07001497 "src/qs8-requantization/q31-neon.c",
Marat Dukhanba7b2792020-09-02 14:26:45 -07001498 "src/qs8-vadd/gen/minmax-neon-ld64-x8.c",
1499 "src/qs8-vadd/gen/minmax-neon-ld64-x16.c",
1500 "src/qs8-vadd/gen/minmax-neon-ld64-x24.c",
1501 "src/qs8-vadd/gen/minmax-neon-ld64-x32.c",
1502 "src/qs8-vaddc/gen/minmax-neon-ld64-x8.c",
1503 "src/qs8-vaddc/gen/minmax-neon-ld64-x16.c",
1504 "src/qs8-vaddc/gen/minmax-neon-ld64-x24.c",
1505 "src/qs8-vaddc/gen/minmax-neon-ld64-x32.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -07001506 "src/qu8-avgpool/9p8x-minmax-neon-c8.c",
1507 "src/qu8-avgpool/9x-minmax-neon-c8.c",
1508 "src/qu8-dwconv/up8x9-minmax-neon.c",
1509 "src/qu8-gavgpool/7p7x-minmax-neon-c8.c",
1510 "src/qu8-gavgpool/7x-minmax-neon-c8.c",
1511 "src/qu8-gemm/4x8-minmax-neon.c",
1512 "src/qu8-gemm/8x8-minmax-neon.c",
1513 "src/qu8-igemm/4x8-minmax-neon.c",
1514 "src/qu8-igemm/8x8-minmax-neon.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07001515 "src/qu8-requantization/fp32-neon.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001516 "src/qu8-requantization/precise-neon.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07001517 "src/qu8-requantization/q31-neon.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -07001518 "src/qu8-vadd/minmax-neon.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001519 "src/u8-clamp/neon-x64.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001520 "src/u8-maxpool/9p8x-minmax-neon-c16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001521 "src/u8-rmax/neon.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001522 "src/x8-zip/x2-neon.c",
1523 "src/x8-zip/x3-neon.c",
1524 "src/x8-zip/x4-neon.c",
1525 "src/x8-zip/xm-neon.c",
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -07001526 "src/x32-fill/neon.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001527 "src/x32-packx/x4-neon-st4.c",
Marat Dukhan63523d42020-05-22 17:07:33 -07001528 "src/x32-pad/neon.c",
Marat Dukhan57dccd82020-04-14 00:53:10 -07001529 "src/x32-unpool/neon.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001530 "src/x32-zip/x2-neon.c",
1531 "src/x32-zip/x3-neon.c",
1532 "src/x32-zip/x4-neon.c",
1533 "src/x32-zip/xm-neon.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001534]
1535
1536NEONFMA_UKERNELS = [
Frank Barchard04336c12020-10-22 16:48:55 -07001537 "src/f32-dwconv/gen/up4x4-minmax-neonfma-acc2.c",
1538 "src/f32-dwconv/gen/up4x4-minmax-neonfma.c",
1539 "src/f32-dwconv/gen/up4x9-minmax-neonfma-acc2.c",
1540 "src/f32-dwconv/gen/up4x9-minmax-neonfma.c",
1541 "src/f32-dwconv/gen/up4x25-minmax-neonfma-acc2.c",
1542 "src/f32-dwconv/gen/up4x25-minmax-neonfma.c",
1543 "src/f32-dwconv/gen/up8x4-minmax-neonfma-acc2.c",
1544 "src/f32-dwconv/gen/up8x4-minmax-neonfma.c",
1545 "src/f32-dwconv/gen/up8x9-minmax-neonfma-acc2.c",
1546 "src/f32-dwconv/gen/up8x9-minmax-neonfma.c",
1547 "src/f32-dwconv/gen/up8x25-minmax-neonfma-acc2.c",
1548 "src/f32-dwconv/gen/up8x25-minmax-neonfma.c",
1549 "src/f32-gemm/gen-inc/1x8inc-minmax-neonfma-dup-ld64.c",
1550 "src/f32-gemm/gen-inc/1x8s4inc-minmax-neonfma.c",
1551 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-dup-ld64.c",
1552 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-dup-ld128.c",
1553 "src/f32-gemm/gen-inc/4x8s4inc-minmax-neonfma.c",
1554 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-dup-ld64.c",
1555 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-dup-ld128.c",
1556 "src/f32-gemm/gen-inc/6x8s4inc-minmax-neonfma.c",
1557 "src/f32-gemm/gen-inc/8x8s4inc-minmax-neonfma.c",
1558 "src/f32-gemm/gen/1x8-minmax-neonfma-dup-ld64.c",
1559 "src/f32-gemm/gen/1x8s4-minmax-neonfma.c",
1560 "src/f32-gemm/gen/4x8-minmax-neonfma-dup-ld64.c",
1561 "src/f32-gemm/gen/4x8-minmax-neonfma-dup-ld128.c",
1562 "src/f32-gemm/gen/4x8s4-minmax-neonfma.c",
1563 "src/f32-gemm/gen/6x8-minmax-neonfma-dup-ld64.c",
1564 "src/f32-gemm/gen/6x8-minmax-neonfma-dup-ld128.c",
1565 "src/f32-gemm/gen/6x8s4-minmax-neonfma.c",
1566 "src/f32-gemm/gen/8x8s4-minmax-neonfma.c",
Marat Dukhan660fd192020-03-10 04:55:30 -07001567 "src/f32-ibilinear/gen/neonfma-c4.c",
1568 "src/f32-ibilinear/gen/neonfma-c8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001569 "src/f32-igemm/gen/1x8-minmax-neonfma-dup-ld64.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001570 "src/f32-igemm/gen/1x8s4-minmax-neonfma.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001571 "src/f32-igemm/gen/4x8-minmax-neonfma-dup-ld64.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001572 "src/f32-igemm/gen/4x8-minmax-neonfma-dup-ld128.c",
1573 "src/f32-igemm/gen/4x8s4-minmax-neonfma.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001574 "src/f32-igemm/gen/6x8-minmax-neonfma-dup-ld64.c",
1575 "src/f32-igemm/gen/6x8-minmax-neonfma-dup-ld128.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001576 "src/f32-igemm/gen/6x8s4-minmax-neonfma.c",
1577 "src/f32-igemm/gen/8x8s4-minmax-neonfma.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001578 "src/f32-ppmm/gen/4x8-minmax-neonfma.c",
1579 "src/f32-ppmm/gen/8x8-minmax-neonfma.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001580 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x4.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001581 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x8-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001582 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x8.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001583 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12-acc2.c",
1584 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001585 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001586 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16-acc2.c",
1587 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001588 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001589 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20-acc2.c",
1590 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001591 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20.c",
1592 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x4.c",
1593 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x8-acc2.c",
1594 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x8.c",
1595 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12-acc2.c",
1596 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12-acc3.c",
1597 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12.c",
1598 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16-acc2.c",
1599 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16-acc4.c",
1600 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16.c",
1601 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20-acc2.c",
1602 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20-acc5.c",
1603 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -08001604 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x4.c",
1605 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x8.c",
1606 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x12.c",
1607 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x16.c",
1608 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x20.c",
1609 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x24.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001610 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x4.c",
1611 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x8.c",
1612 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x12.c",
1613 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x16.c",
1614 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x20.c",
1615 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x24.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -08001616 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x4.c",
1617 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x8.c",
1618 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x12.c",
1619 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x16.c",
1620 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x20.c",
1621 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x24.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -08001622 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x4.c",
1623 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x8.c",
1624 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x12.c",
1625 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x16.c",
1626 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x20.c",
1627 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x24.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001628 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x4.c",
1629 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x8.c",
1630 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x12.c",
1631 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x16.c",
1632 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x20.c",
1633 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x24.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -08001634 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x4.c",
1635 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x8.c",
1636 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x12.c",
1637 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x16.c",
1638 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x20.c",
1639 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x24.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001640 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x4.c",
1641 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x8.c",
1642 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x12.c",
1643 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x16.c",
1644 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x20.c",
1645 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x24.c",
1646 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x4.c",
1647 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x8.c",
1648 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x12.c",
1649 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x16.c",
1650 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x20.c",
1651 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x24.c",
1652 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x4.c",
1653 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x8.c",
1654 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x12.c",
1655 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x16.c",
1656 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x20.c",
1657 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x24.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001658 "src/f32-vmulcaddc/gen/c4-minmax-neonfma-2x.c",
1659 "src/f32-vmulcaddc/gen/c8-minmax-neonfma-2x.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07001660 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x4.c",
1661 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x8.c",
1662 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x12.c",
1663 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x16.c",
1664 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x20.c",
1665 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x24.c",
1666 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x28.c",
1667 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x32.c",
1668 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x36.c",
1669 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x40.c",
1670 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x4.c",
1671 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x8.c",
1672 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x12.c",
1673 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x16.c",
1674 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x20.c",
1675 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x24.c",
1676 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x28.c",
1677 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x32.c",
1678 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x36.c",
1679 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x40.c",
Marat Dukhanb7633f22020-11-20 16:34:56 -08001680 "src/math/exp-neonfma-rr2-lut64-p2.c",
1681 "src/math/exp-neonfma-rr2-p5.c",
Marat Dukhan9dd119a2020-11-20 18:20:04 -08001682 "src/math/expminus-neonfma-rr2-lut64-p2.c",
1683 "src/math/expminus-neonfma-rr2-lut2048-p1.c",
1684 "src/math/expminus-neonfma-rr2-p5.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001685 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr1recps1fma.c",
1686 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr2fma.c",
1687 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr2recps.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001688 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr1recps1fma.c",
1689 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr2fma.c",
1690 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr2recps.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001691 "src/math/sigmoid-neonfma-rr1-p5-nr1recps1fma.c",
1692 "src/math/sigmoid-neonfma-rr1-p5-nr2fma.c",
1693 "src/math/sigmoid-neonfma-rr1-p5-nr2recps.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001694 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr1recps1fma.c",
1695 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr2fma.c",
1696 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr2recps.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001697 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr1recps1fma.c",
1698 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr2fma.c",
1699 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr2recps.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001700 "src/math/sigmoid-neonfma-rr2-p5-nr1recps1fma.c",
1701 "src/math/sigmoid-neonfma-rr2-p5-nr2fma.c",
1702 "src/math/sigmoid-neonfma-rr2-p5-nr2recps.c",
Marat Dukhan84000762020-06-29 18:38:43 -07001703 "src/math/sqrt-neonfma-nr1fma.c",
Marat Dukhan84000762020-06-29 18:38:43 -07001704 "src/math/sqrt-neonfma-nr1rsqrts1fma1adj.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001705 "src/math/sqrt-neonfma-nr2fma.c",
1706 "src/math/sqrt-neonfma-nr2fma1adj.c",
1707 "src/math/sqrt-neonfma-nr3fma.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001708]
1709
1710AARCH64_NEONFMA_UKERNELS = [
Marat Dukhan56b10cd2020-05-18 09:35:49 -07001711 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neonfma-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001712 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neonfma-2x2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001713 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neonfma-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001714 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neonfma-2x2.c",
Marat Dukhan56b10cd2020-05-18 09:35:49 -07001715 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neonfma-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001716 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neonfma-2x2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001717 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neonfma-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001718 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neonfma-2x2.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -07001719 "src/f32-conv-hwc2chw/3x3s2p1c3x4-neonfma-2x2.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001720 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-1x4-acc2.c",
1721 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-1x4-acc3.c",
1722 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-1x4-acc4.c",
Marat Dukhan1268a242020-10-24 00:36:32 -07001723 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-1x4.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001724 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-2x4-acc2.c",
Marat Dukhan1268a242020-10-24 00:36:32 -07001725 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-2x4.c",
1726 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-3x4.c",
1727 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-4x4.c",
1728 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-5x4.c",
1729 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-6x4.c",
Marat Dukhan82f0c322020-10-25 19:17:35 -07001730 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neonfma-1x4-acc2.c",
1731 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neonfma-1x4-acc3.c",
1732 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neonfma-1x4-acc4.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001733 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neonfma-1x4.c",
Marat Dukhan82f0c322020-10-25 19:17:35 -07001734 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neonfma-2x4-acc2.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001735 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neonfma-2x4.c",
1736 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neonfma-3x4.c",
1737 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neonfma-4x4.c",
Marat Dukhan149f0ea2020-10-26 12:50:33 -07001738 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-1x4-acc2.c",
1739 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-1x4-acc3.c",
1740 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-1x4-acc4.c",
1741 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-1x4-acc5.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001742 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-1x4.c",
Marat Dukhan149f0ea2020-10-26 12:50:33 -07001743 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-2x4-acc2.c",
1744 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-2x4-acc3.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001745 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-2x4.c",
Marat Dukhan149f0ea2020-10-26 12:50:33 -07001746 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-3x4-acc2.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001747 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-3x4.c",
Marat Dukhan149f0ea2020-10-26 12:50:33 -07001748 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-4x4-acc2.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001749 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-4x4.c",
1750 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-5x4.c",
Marat Dukhan30d4b252020-10-29 16:33:22 -07001751 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neonfma-1x4-acc2.c",
1752 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neonfma-1x4-acc3.c",
1753 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neonfma-1x4-acc4.c",
1754 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neonfma-1x4-acc5.c",
1755 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neonfma-1x4.c",
1756 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neonfma-2x4-acc2.c",
1757 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neonfma-2x4-acc3.c",
1758 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neonfma-2x4.c",
Marat Dukhan30d4b252020-10-29 16:33:22 -07001759 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neonfma-3x4-acc2.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07001760 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neonfma-3x4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001761 "src/f32-gemm/gen-inc/1x8inc-minmax-neonfma-lane-ld64.c",
1762 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-lane-ld64.c",
1763 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-lane-ld128.c",
1764 "src/f32-gemm/gen-inc/5x8inc-minmax-neonfma-lane-ld64.c",
1765 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-lane-ld64.c",
1766 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-lane-ld128.c",
1767 "src/f32-gemm/gen/1x8-minmax-neonfma-lane-ld64.c",
1768 "src/f32-gemm/gen/4x2-minmax-neonfma-lane-ld64.c",
1769 "src/f32-gemm/gen/4x8-minmax-neonfma-lane-ld64.c",
1770 "src/f32-gemm/gen/4x8-minmax-neonfma-lane-ld128.c",
1771 "src/f32-gemm/gen/5x8-minmax-neonfma-lane-ld64.c",
1772 "src/f32-gemm/gen/6x8-minmax-neonfma-lane-ld64.c",
1773 "src/f32-gemm/gen/6x8-minmax-neonfma-lane-ld128.c",
1774 "src/f32-igemm/gen/1x8-minmax-neonfma-lane-ld64.c",
1775 "src/f32-igemm/gen/4x2-minmax-neonfma-lane-ld64.c",
1776 "src/f32-igemm/gen/4x4-minmax-neonfma-lane-ld64.c",
1777 "src/f32-igemm/gen/4x8-minmax-neonfma-lane-ld64.c",
1778 "src/f32-igemm/gen/4x8-minmax-neonfma-lane-ld128.c",
1779 "src/f32-igemm/gen/6x8-minmax-neonfma-lane-ld64.c",
1780 "src/f32-igemm/gen/6x8-minmax-neonfma-lane-ld128.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -08001781 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x4.c",
1782 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x8.c",
1783 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x12.c",
1784 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x16.c",
1785 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x20.c",
1786 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x24.c",
1787 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x4.c",
1788 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x8.c",
1789 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x12.c",
1790 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x16.c",
1791 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x20.c",
1792 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x24.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001793 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x4.c",
1794 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x8.c",
1795 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x12.c",
1796 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x16.c",
1797 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x20.c",
1798 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x24.c",
Marat Dukhan355ab432020-04-09 19:01:52 -07001799 "src/f32-spmm/gen/4x1-minmax-neonfma-pipelined.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07001800 "src/f32-spmm/gen/4x1-minmax-neonfma-x2.c",
Marat Dukhan355ab432020-04-09 19:01:52 -07001801 "src/f32-spmm/gen/4x1-minmax-neonfma.c",
1802 "src/f32-spmm/gen/4x2-minmax-neonfma.c",
1803 "src/f32-spmm/gen/4x4-minmax-neonfma.c",
1804 "src/f32-spmm/gen/8x1-minmax-neonfma-pipelined.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07001805 "src/f32-spmm/gen/8x1-minmax-neonfma-x2.c",
Marat Dukhan355ab432020-04-09 19:01:52 -07001806 "src/f32-spmm/gen/8x1-minmax-neonfma.c",
1807 "src/f32-spmm/gen/8x2-minmax-neonfma.c",
1808 "src/f32-spmm/gen/8x4-minmax-neonfma.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001809 "src/f32-spmm/gen/12x1-minmax-neonfma.c",
1810 "src/f32-spmm/gen/12x2-minmax-neonfma.c",
1811 "src/f32-spmm/gen/12x4-minmax-neonfma.c",
1812 "src/f32-spmm/gen/16x1-minmax-neonfma-pipelined.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07001813 "src/f32-spmm/gen/16x1-minmax-neonfma-x2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001814 "src/f32-spmm/gen/16x1-minmax-neonfma.c",
1815 "src/f32-spmm/gen/16x2-minmax-neonfma.c",
1816 "src/f32-spmm/gen/16x4-minmax-neonfma.c",
Frank Barchard846c0c62020-10-26 15:01:39 -07001817 "src/f32-spmm/gen/32x1-minmax-neonfma-pipelined.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07001818 "src/f32-spmm/gen/32x1-minmax-neonfma-x2.c",
Frank Barchard846c0c62020-10-26 15:01:39 -07001819 "src/f32-spmm/gen/32x1-minmax-neonfma.c",
1820 "src/f32-spmm/gen/32x2-minmax-neonfma.c",
1821 "src/f32-spmm/gen/32x4-minmax-neonfma.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001822 "src/f32-vbinary/gen/vdiv-minmax-neon-x4.c",
1823 "src/f32-vbinary/gen/vdiv-minmax-neon-x8.c",
1824 "src/f32-vbinary/gen/vdivc-minmax-neon-x4.c",
1825 "src/f32-vbinary/gen/vdivc-minmax-neon-x8.c",
1826 "src/f32-vbinary/gen/vrdivc-minmax-neon-x4.c",
1827 "src/f32-vbinary/gen/vrdivc-minmax-neon-x8.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07001828 "src/f32-vsqrt/gen/neon-sqrt-x4.c",
1829 "src/f32-vsqrt/gen/neon-sqrt-x8.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001830 "src/math/sigmoid-neonfma-rr1-lut64-p2-div.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001831 "src/math/sigmoid-neonfma-rr1-lut2048-p1-div.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001832 "src/math/sigmoid-neonfma-rr1-p5-div.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001833 "src/math/sigmoid-neonfma-rr2-lut64-p2-div.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001834 "src/math/sigmoid-neonfma-rr2-lut2048-p1-div.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001835 "src/math/sigmoid-neonfma-rr2-p5-div.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001836]
1837
Marat Dukhan8853b822020-05-07 12:19:01 -07001838NEONV8_UKERNELS = [
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07001839 "src/f32-vrnd/gen/vrndd-neonv8-x4.c",
1840 "src/f32-vrnd/gen/vrndd-neonv8-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001841 "src/f32-vrnd/gen/vrndne-neonv8-x4.c",
1842 "src/f32-vrnd/gen/vrndne-neonv8-x8.c",
1843 "src/f32-vrnd/gen/vrndu-neonv8-x4.c",
1844 "src/f32-vrnd/gen/vrndu-neonv8-x8.c",
1845 "src/f32-vrnd/gen/vrndz-neonv8-x4.c",
1846 "src/f32-vrnd/gen/vrndz-neonv8-x8.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001847 "src/math/roundd-neonv8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001848 "src/math/roundne-neonv8.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001849 "src/math/roundu-neonv8.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07001850 "src/math/roundz-neonv8.c",
Marat Dukhan8853b822020-05-07 12:19:01 -07001851]
1852
Marat Dukhan08c4a432019-10-03 09:29:21 -07001853AARCH64_NEONFP16ARITH_UKERNELS = [
Frank Barchardb1966592020-05-12 13:47:06 -07001854 "src/f16-clamp/gen/neonfp16arith-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001855 "src/f16-clamp/gen/neonfp16arith-x16.c",
Frank Barchard5a599a62020-06-04 20:12:44 -07001856 "src/f16-dwconv/gen/up8x4-minmax-neonfp16arith-acc2.c",
1857 "src/f16-dwconv/gen/up8x4-minmax-neonfp16arith.c",
1858 "src/f16-dwconv/gen/up8x9-minmax-neonfp16arith-acc2.c",
1859 "src/f16-dwconv/gen/up8x9-minmax-neonfp16arith.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001860 "src/f16-dwconv/gen/up8x25-minmax-neonfp16arith-acc2.c",
1861 "src/f16-dwconv/gen/up8x25-minmax-neonfp16arith.c",
1862 "src/f16-dwconv/gen/up16x4-minmax-neonfp16arith-acc2.c",
1863 "src/f16-dwconv/gen/up16x4-minmax-neonfp16arith.c",
1864 "src/f16-dwconv/gen/up16x9-minmax-neonfp16arith-acc2.c",
1865 "src/f16-dwconv/gen/up16x9-minmax-neonfp16arith.c",
1866 "src/f16-dwconv/gen/up16x25-minmax-neonfp16arith-acc2.c",
1867 "src/f16-dwconv/gen/up16x25-minmax-neonfp16arith.c",
Frank Barchard0bb49a72020-06-04 11:35:11 -07001868 "src/f16-gavgpool/7p7x-minmax-neonfp16arith-c8.c",
1869 "src/f16-gavgpool/7x-minmax-neonfp16arith-c8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001870 "src/f16-gemm/gen-inc/1x8inc-minmax-neonfp16arith-ld64.c",
1871 "src/f16-gemm/gen-inc/1x16inc-minmax-neonfp16arith-ld64.c",
1872 "src/f16-gemm/gen-inc/4x8inc-minmax-neonfp16arith-ld64.c",
1873 "src/f16-gemm/gen-inc/4x16inc-minmax-neonfp16arith-ld64.c",
1874 "src/f16-gemm/gen-inc/6x8inc-minmax-neonfp16arith-ld64.c",
1875 "src/f16-gemm/gen-inc/6x16inc-minmax-neonfp16arith-ld64.c",
1876 "src/f16-gemm/gen-inc/8x8inc-minmax-neonfp16arith-ld64.c",
1877 "src/f16-gemm/gen-inc/8x16inc-minmax-neonfp16arith-ld64.c",
1878 "src/f16-gemm/gen/1x8-minmax-neonfp16arith-ld64.c",
1879 "src/f16-gemm/gen/1x16-minmax-neonfp16arith-ld64.c",
1880 "src/f16-gemm/gen/4x8-minmax-neonfp16arith-ld64.c",
1881 "src/f16-gemm/gen/4x16-minmax-neonfp16arith-ld64.c",
1882 "src/f16-gemm/gen/6x8-minmax-neonfp16arith-ld64.c",
1883 "src/f16-gemm/gen/6x16-minmax-neonfp16arith-ld64.c",
1884 "src/f16-gemm/gen/8x8-minmax-neonfp16arith-ld64.c",
1885 "src/f16-gemm/gen/8x16-minmax-neonfp16arith-ld64.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07001886 "src/f16-hswish/gen/hswish-neonfp16arith-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001887 "src/f16-hswish/gen/hswish-neonfp16arith-x16.c",
1888 "src/f16-igemm/gen/1x8-minmax-neonfp16arith-ld64.c",
1889 "src/f16-igemm/gen/1x16-minmax-neonfp16arith-ld64.c",
1890 "src/f16-igemm/gen/4x8-minmax-neonfp16arith-ld64.c",
1891 "src/f16-igemm/gen/4x16-minmax-neonfp16arith-ld64.c",
1892 "src/f16-igemm/gen/6x8-minmax-neonfp16arith-ld64.c",
1893 "src/f16-igemm/gen/6x16-minmax-neonfp16arith-ld64.c",
1894 "src/f16-igemm/gen/8x8-minmax-neonfp16arith-ld64.c",
1895 "src/f16-igemm/gen/8x16-minmax-neonfp16arith-ld64.c",
Frank Barchardb1966592020-05-12 13:47:06 -07001896 "src/f16-prelu/gen/neonfp16arith-2x8.c",
Frank Barcharda5316982020-07-23 13:19:28 -07001897 "src/f16-prelu/gen/neonfp16arith-2x16.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07001898 "src/f16-relu/gen/neonfp16arith-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001899 "src/f16-relu/gen/neonfp16arith-x16.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07001900 "src/f16-spmm/gen/8x1-minmax-neonfp16arith-x2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001901 "src/f16-spmm/gen/8x1-minmax-neonfp16arith.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07001902 "src/f16-spmm/gen/16x1-minmax-neonfp16arith-x2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001903 "src/f16-spmm/gen/16x1-minmax-neonfp16arith.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07001904 "src/f16-spmm/gen/24x1-minmax-neonfp16arith-x2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001905 "src/f16-spmm/gen/24x1-minmax-neonfp16arith.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07001906 "src/f16-spmm/gen/32x1-minmax-neonfp16arith-x2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001907 "src/f16-spmm/gen/32x1-minmax-neonfp16arith.c",
1908 "src/f16-vbinary/gen/vadd-minmax-neonfp16arith-x8.c",
1909 "src/f16-vbinary/gen/vadd-minmax-neonfp16arith-x16.c",
1910 "src/f16-vbinary/gen/vaddc-minmax-neonfp16arith-x8.c",
1911 "src/f16-vbinary/gen/vaddc-minmax-neonfp16arith-x16.c",
1912 "src/f16-vbinary/gen/vdiv-minmax-neonfp16arith-x8.c",
1913 "src/f16-vbinary/gen/vdiv-minmax-neonfp16arith-x16.c",
1914 "src/f16-vbinary/gen/vdivc-minmax-neonfp16arith-x8.c",
1915 "src/f16-vbinary/gen/vdivc-minmax-neonfp16arith-x16.c",
1916 "src/f16-vbinary/gen/vmax-neonfp16arith-x8.c",
1917 "src/f16-vbinary/gen/vmax-neonfp16arith-x16.c",
1918 "src/f16-vbinary/gen/vmaxc-neonfp16arith-x8.c",
1919 "src/f16-vbinary/gen/vmaxc-neonfp16arith-x16.c",
1920 "src/f16-vbinary/gen/vmin-neonfp16arith-x8.c",
1921 "src/f16-vbinary/gen/vmin-neonfp16arith-x16.c",
1922 "src/f16-vbinary/gen/vminc-neonfp16arith-x8.c",
1923 "src/f16-vbinary/gen/vminc-neonfp16arith-x16.c",
1924 "src/f16-vbinary/gen/vmul-minmax-neonfp16arith-x8.c",
1925 "src/f16-vbinary/gen/vmul-minmax-neonfp16arith-x16.c",
1926 "src/f16-vbinary/gen/vmulc-minmax-neonfp16arith-x8.c",
1927 "src/f16-vbinary/gen/vmulc-minmax-neonfp16arith-x16.c",
1928 "src/f16-vbinary/gen/vrdivc-minmax-neonfp16arith-x8.c",
1929 "src/f16-vbinary/gen/vrdivc-minmax-neonfp16arith-x16.c",
1930 "src/f16-vbinary/gen/vrsubc-minmax-neonfp16arith-x8.c",
1931 "src/f16-vbinary/gen/vrsubc-minmax-neonfp16arith-x16.c",
1932 "src/f16-vbinary/gen/vsub-minmax-neonfp16arith-x8.c",
1933 "src/f16-vbinary/gen/vsub-minmax-neonfp16arith-x16.c",
1934 "src/f16-vbinary/gen/vsubc-minmax-neonfp16arith-x8.c",
1935 "src/f16-vbinary/gen/vsubc-minmax-neonfp16arith-x16.c",
1936 "src/f16-vmulcaddc/gen/c8-minmax-neonfp16arith-2x.c",
1937 "src/f16-vmulcaddc/gen/c16-minmax-neonfp16arith-2x.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001938]
1939
Benoit Jacoba9644732020-08-13 12:48:55 -07001940NEONDOT_UKERNELS = [
Marat Dukhand4c83032020-09-09 23:20:12 -07001941 "src/qs8-gemm/gen/1x8c4-minmax-neondot.c",
Marat Dukhand4c83032020-09-09 23:20:12 -07001942 "src/qs8-gemm/gen/1x16c4-minmax-neondot.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001943 "src/qs8-gemm/gen/4x8c4-minmax-neondot.c",
Benoit Jacoba9644732020-08-13 12:48:55 -07001944 "src/qs8-gemm/gen/4x16c4-minmax-neondot.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001945 "src/qs8-gemm/gen/6x8c4-minmax-neondot.c",
Frank Barchard2fa17452020-09-22 09:51:46 -07001946 "src/qs8-gemm/gen/6x16c4-minmax-neondot.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001947 "src/qs8-gemm/gen/8x8c4-minmax-neondot.c",
Frank Barcharda48848f2020-09-24 00:59:24 -07001948 "src/qs8-gemm/gen/8x16c4-minmax-neondot.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001949 "src/qs8-gemm/gen/12x8c4-minmax-neondot.c",
Marat Dukhand4c83032020-09-09 23:20:12 -07001950 "src/qs8-igemm/gen/1x8c4-minmax-neondot.c",
Marat Dukhand4c83032020-09-09 23:20:12 -07001951 "src/qs8-igemm/gen/1x16c4-minmax-neondot.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001952 "src/qs8-igemm/gen/4x8c4-minmax-neondot.c",
Benoit Jacoba05487f2020-08-14 13:37:21 -07001953 "src/qs8-igemm/gen/4x16c4-minmax-neondot.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001954 "src/qs8-igemm/gen/6x8c4-minmax-neondot.c",
Frank Barchard2fa17452020-09-22 09:51:46 -07001955 "src/qs8-igemm/gen/6x16c4-minmax-neondot.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001956 "src/qs8-igemm/gen/8x8c4-minmax-neondot.c",
Frank Barcharda48848f2020-09-24 00:59:24 -07001957 "src/qs8-igemm/gen/8x16c4-minmax-neondot.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001958 "src/qs8-igemm/gen/12x8c4-minmax-neondot.c",
Benoit Jacoba9644732020-08-13 12:48:55 -07001959]
1960
Marat Dukhan08c4a432019-10-03 09:29:21 -07001961SSE_UKERNELS = [
Marat Dukhan99936602020-04-11 16:47:01 -07001962 "src/f32-avgpool/9p8x-minmax-sse-c4.c",
1963 "src/f32-avgpool/9x-minmax-sse-c4.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001964 "src/f32-clamp/gen/sse-x4.c",
1965 "src/f32-clamp/gen/sse-x8.c",
Erich Elsenb1233402020-06-08 15:53:15 -07001966 "src/f32-conv-hwc2chw/3x3s2p1c3x4-sse-1x1.c",
1967 "src/f32-conv-hwc2chw/3x3s2p1c3x4-sse-2x2.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001968 "src/f32-dwconv/gen/up4x4-minmax-sse-acc2.c",
1969 "src/f32-dwconv/gen/up4x4-minmax-sse.c",
1970 "src/f32-dwconv/gen/up4x9-minmax-sse-acc2.c",
1971 "src/f32-dwconv/gen/up4x9-minmax-sse.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001972 "src/f32-dwconv/gen/up4x25-minmax-sse-acc2.c",
1973 "src/f32-dwconv/gen/up4x25-minmax-sse.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001974 "src/f32-dwconv/gen/up8x4-minmax-sse-acc2.c",
1975 "src/f32-dwconv/gen/up8x4-minmax-sse.c",
1976 "src/f32-dwconv/gen/up8x9-minmax-sse-acc2.c",
1977 "src/f32-dwconv/gen/up8x9-minmax-sse.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001978 "src/f32-dwconv/gen/up8x25-minmax-sse-acc2.c",
1979 "src/f32-dwconv/gen/up8x25-minmax-sse.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001980 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-1x4-acc2.c",
1981 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-1x4-acc3.c",
1982 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-1x4-acc4.c",
Marat Dukhan470078a2020-10-23 22:36:52 -07001983 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-1x4.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001984 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-2x4-acc2.c",
Marat Dukhan470078a2020-10-23 22:36:52 -07001985 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-2x4.c",
1986 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-3x4.c",
1987 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-4x4.c",
1988 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-5x4.c",
1989 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-6x4.c",
Marat Dukhan0ff97182020-10-25 19:14:03 -07001990 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-sse-1x4-acc2.c",
1991 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-sse-1x4-acc3.c",
1992 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-sse-1x4-acc4.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001993 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-sse-1x4.c",
Marat Dukhan0ff97182020-10-25 19:14:03 -07001994 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-sse-2x4-acc2.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001995 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-sse-2x4.c",
1996 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-sse-3x4.c",
1997 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-sse-4x4.c",
Marat Dukhand0503892020-10-30 08:22:04 -07001998 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-1x4-acc2.c",
1999 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-1x4-acc3.c",
2000 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-1x4-acc4.c",
2001 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-1x4-acc5.c",
2002 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-1x4.c",
2003 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-2x4-acc2.c",
2004 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-2x4-acc3.c",
2005 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-2x4.c",
2006 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-3x4-acc2.c",
2007 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-3x4.c",
2008 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-4x4-acc2.c",
2009 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-4x4.c",
2010 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-5x4.c",
Marat Dukhanccca2142020-10-30 17:32:45 -07002011 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-sse-1x4-acc2.c",
2012 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-sse-1x4-acc3.c",
2013 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-sse-1x4-acc4.c",
2014 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-sse-1x4-acc5.c",
2015 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-sse-1x4.c",
2016 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-sse-2x4-acc2.c",
2017 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-sse-2x4-acc3.c",
2018 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-sse-2x4.c",
Marat Dukhanccca2142020-10-30 17:32:45 -07002019 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-sse-3x4-acc2.c",
Frank Barchard8ef44cd2020-11-03 12:30:23 -08002020 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-sse-3x4.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -07002021 "src/f32-gavgpool-cw/sse-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -07002022 "src/f32-gavgpool/7p7x-minmax-sse-c4.c",
2023 "src/f32-gavgpool/7x-minmax-sse-c4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002024 "src/f32-gemm/gen-inc/1x8inc-minmax-sse-dup.c",
2025 "src/f32-gemm/gen-inc/1x8inc-minmax-sse-load1.c",
2026 "src/f32-gemm/gen-inc/1x8s4inc-minmax-sse.c",
2027 "src/f32-gemm/gen-inc/4x8inc-minmax-sse-dup.c",
2028 "src/f32-gemm/gen-inc/4x8inc-minmax-sse-load1.c",
2029 "src/f32-gemm/gen-inc/4x8s4inc-minmax-sse.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002030 "src/f32-gemm/gen/1x8-minmax-sse-dup.c",
2031 "src/f32-gemm/gen/1x8-minmax-sse-load1.c",
2032 "src/f32-gemm/gen/1x8s4-minmax-sse.c",
2033 "src/f32-gemm/gen/4x2c4-minmax-sse.c",
2034 "src/f32-gemm/gen/4x8-minmax-sse-dup.c",
2035 "src/f32-gemm/gen/4x8-minmax-sse-load1.c",
2036 "src/f32-gemm/gen/4x8s4-minmax-sse.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07002037 "src/f32-hswish/gen/hswish-sse-x4.c",
2038 "src/f32-hswish/gen/hswish-sse-x8.c",
Marat Dukhan660fd192020-03-10 04:55:30 -07002039 "src/f32-ibilinear/gen/sse-c4.c",
2040 "src/f32-ibilinear/gen/sse-c8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002041 "src/f32-igemm/gen/1x8-minmax-sse-dup.c",
2042 "src/f32-igemm/gen/1x8-minmax-sse-load1.c",
2043 "src/f32-igemm/gen/1x8s4-minmax-sse.c",
2044 "src/f32-igemm/gen/4x2c4-minmax-sse.c",
2045 "src/f32-igemm/gen/4x8-minmax-sse-dup.c",
2046 "src/f32-igemm/gen/4x8-minmax-sse-load1.c",
2047 "src/f32-igemm/gen/4x8s4-minmax-sse.c",
Marat Dukhan99936602020-04-11 16:47:01 -07002048 "src/f32-maxpool/9p8x-minmax-sse-c4.c",
2049 "src/f32-pavgpool/9p8x-minmax-sse-c4.c",
2050 "src/f32-pavgpool/9x-minmax-sse-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002051 "src/f32-ppmm/gen/4x8-minmax-sse.c",
Marat Dukhan39b5e942020-06-24 15:03:48 -07002052 "src/f32-prelu/gen/sse-2x4.c",
2053 "src/f32-prelu/gen/sse-2x8.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07002054 "src/f32-relu/gen/sse-x4.c",
2055 "src/f32-relu/gen/sse-x8.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002056 "src/f32-rmax/sse.c",
Marat Dukhan355ab432020-04-09 19:01:52 -07002057 "src/f32-spmm/gen/4x1-minmax-sse.c",
2058 "src/f32-spmm/gen/8x1-minmax-sse.c",
Erich Elsen6e80fdc2020-06-09 15:35:37 -07002059 "src/f32-spmm/gen/16x1-minmax-sse.c",
Frank Barchard846c0c62020-10-26 15:01:39 -07002060 "src/f32-spmm/gen/32x1-minmax-sse.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002061 "src/f32-vbinary/gen/vadd-minmax-sse-x4.c",
2062 "src/f32-vbinary/gen/vadd-minmax-sse-x8.c",
2063 "src/f32-vbinary/gen/vaddc-minmax-sse-x4.c",
2064 "src/f32-vbinary/gen/vaddc-minmax-sse-x8.c",
2065 "src/f32-vbinary/gen/vdiv-minmax-sse-x4.c",
2066 "src/f32-vbinary/gen/vdiv-minmax-sse-x8.c",
2067 "src/f32-vbinary/gen/vdivc-minmax-sse-x4.c",
2068 "src/f32-vbinary/gen/vdivc-minmax-sse-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -08002069 "src/f32-vbinary/gen/vmax-sse-x4.c",
2070 "src/f32-vbinary/gen/vmax-sse-x8.c",
2071 "src/f32-vbinary/gen/vmaxc-sse-x4.c",
2072 "src/f32-vbinary/gen/vmaxc-sse-x8.c",
2073 "src/f32-vbinary/gen/vmin-sse-x4.c",
2074 "src/f32-vbinary/gen/vmin-sse-x8.c",
2075 "src/f32-vbinary/gen/vminc-sse-x4.c",
2076 "src/f32-vbinary/gen/vminc-sse-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002077 "src/f32-vbinary/gen/vmul-minmax-sse-x4.c",
2078 "src/f32-vbinary/gen/vmul-minmax-sse-x8.c",
2079 "src/f32-vbinary/gen/vmulc-minmax-sse-x4.c",
2080 "src/f32-vbinary/gen/vmulc-minmax-sse-x8.c",
2081 "src/f32-vbinary/gen/vrdivc-minmax-sse-x4.c",
2082 "src/f32-vbinary/gen/vrdivc-minmax-sse-x8.c",
2083 "src/f32-vbinary/gen/vrsubc-minmax-sse-x4.c",
2084 "src/f32-vbinary/gen/vrsubc-minmax-sse-x8.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07002085 "src/f32-vbinary/gen/vsqrdiff-sse-x4.c",
2086 "src/f32-vbinary/gen/vsqrdiff-sse-x8.c",
2087 "src/f32-vbinary/gen/vsqrdiffc-sse-x4.c",
2088 "src/f32-vbinary/gen/vsqrdiffc-sse-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002089 "src/f32-vbinary/gen/vsub-minmax-sse-x4.c",
2090 "src/f32-vbinary/gen/vsub-minmax-sse-x8.c",
2091 "src/f32-vbinary/gen/vsubc-minmax-sse-x4.c",
2092 "src/f32-vbinary/gen/vsubc-minmax-sse-x8.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07002093 "src/f32-vlrelu/gen/vlrelu-sse-x4.c",
2094 "src/f32-vlrelu/gen/vlrelu-sse-x8.c",
Marat Dukhan99936602020-04-11 16:47:01 -07002095 "src/f32-vmulcaddc/gen/c4-minmax-sse-2x.c",
2096 "src/f32-vmulcaddc/gen/c8-minmax-sse-2x.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07002097 "src/f32-vsqrt/gen/sse-sqrt-x4.c",
2098 "src/f32-vsqrt/gen/sse-sqrt-x8.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07002099 "src/f32-vunary/gen/vabs-sse-x4.c",
2100 "src/f32-vunary/gen/vabs-sse-x8.c",
2101 "src/f32-vunary/gen/vneg-sse-x4.c",
2102 "src/f32-vunary/gen/vneg-sse-x8.c",
2103 "src/f32-vunary/gen/vsqr-sse-x4.c",
2104 "src/f32-vunary/gen/vsqr-sse-x8.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07002105 "src/math/roundd-sse-addsub.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002106 "src/math/roundne-sse-addsub.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07002107 "src/math/roundu-sse-addsub.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07002108 "src/math/roundz-sse-addsub.c",
Marat Dukhan84000762020-06-29 18:38:43 -07002109 "src/math/sqrt-sse-hh1mac.c",
2110 "src/math/sqrt-sse-nr1mac.c",
2111 "src/math/sqrt-sse-nr2mac.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002112 "src/x32-fill/sse.c",
2113 "src/x32-packx/x4-sse.c",
2114 "src/x32-pad/sse.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002115]
2116
2117SSE2_UKERNELS = [
Marat Dukhan329da642019-11-19 21:44:39 -08002118 "src/f32-argmaxpool/4x-sse2-c4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002119 "src/f32-argmaxpool/9p8x-sse2-c4.c",
Marat Dukhan329da642019-11-19 21:44:39 -08002120 "src/f32-argmaxpool/9x-sse2-c4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08002121 "src/f32-prelu/gen/sse2-2x4.c",
2122 "src/f32-prelu/gen/sse2-2x8.c",
Marat Dukhanb39689d2020-01-24 13:32:20 -08002123 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x4.c",
Marat Dukhanb39689d2020-01-24 13:32:20 -08002124 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x8-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002125 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x8.c",
Marat Dukhanb39689d2020-01-24 13:32:20 -08002126 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12-acc2.c",
2127 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002128 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12.c",
Marat Dukhanb39689d2020-01-24 13:32:20 -08002129 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16-acc2.c",
2130 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002131 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16.c",
Marat Dukhanb39689d2020-01-24 13:32:20 -08002132 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20-acc2.c",
2133 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002134 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20.c",
Marat Dukhand243c1a2020-09-17 18:48:10 -07002135 "src/f32-sigmoid/gen/sse2-lut64-p2-div-x4.c",
2136 "src/f32-sigmoid/gen/sse2-lut64-p2-div-x8.c",
2137 "src/f32-sigmoid/gen/sse2-lut64-p2-div-x12.c",
2138 "src/f32-sigmoid/gen/sse2-lut64-p2-div-x16.c",
2139 "src/f32-sigmoid/gen/sse2-lut64-p2-div-x20.c",
2140 "src/f32-sigmoid/gen/sse2-lut64-p2-div-x24.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002141 "src/f32-sigmoid/gen/sse2-p5-div-x4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08002142 "src/f32-sigmoid/gen/sse2-p5-div-x8.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002143 "src/f32-sigmoid/gen/sse2-p5-div-x12.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08002144 "src/f32-sigmoid/gen/sse2-p5-div-x16.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002145 "src/f32-sigmoid/gen/sse2-p5-div-x20.c",
2146 "src/f32-sigmoid/gen/sse2-p5-div-x24.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07002147 "src/f32-vlrelu/gen/vlrelu-sse2-x4.c",
2148 "src/f32-vlrelu/gen/vlrelu-sse2-x8.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07002149 "src/f32-vrnd/gen/vrndd-sse2-x4.c",
2150 "src/f32-vrnd/gen/vrndd-sse2-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002151 "src/f32-vrnd/gen/vrndne-sse2-x4.c",
2152 "src/f32-vrnd/gen/vrndne-sse2-x8.c",
2153 "src/f32-vrnd/gen/vrndu-sse2-x4.c",
2154 "src/f32-vrnd/gen/vrndu-sse2-x8.c",
2155 "src/f32-vrnd/gen/vrndz-sse2-x4.c",
2156 "src/f32-vrnd/gen/vrndz-sse2-x8.c",
Marat Dukhanb7633f22020-11-20 16:34:56 -08002157 "src/math/exp-sse2-rr2-lut64-p2.c",
2158 "src/math/exp-sse2-rr2-p5.c",
Marat Dukhand6d321e2020-11-20 18:35:55 -08002159 "src/math/expm1minus-sse2-rr2-p5.c",
Marat Dukhana438aca2020-11-20 15:45:01 -08002160 "src/math/expm1minus-sse2-rr2-p6.c",
Frank Barchard3b800452020-11-22 12:12:35 -08002161 "src/math/expminus-sse2-rr2-p5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002162 "src/math/roundd-sse2-cvt.c",
2163 "src/math/roundne-sse2-cvt.c",
2164 "src/math/roundu-sse2-cvt.c",
2165 "src/math/roundz-sse2-cvt.c",
2166 "src/math/sigmoid-sse2-rr2-lut64-p2-div.c",
2167 "src/math/sigmoid-sse2-rr2-lut64-p2-nr1.c",
2168 "src/math/sigmoid-sse2-rr2-lut64-p2-nr2.c",
2169 "src/math/sigmoid-sse2-rr2-p5-div.c",
2170 "src/math/sigmoid-sse2-rr2-p5-nr1.c",
2171 "src/math/sigmoid-sse2-rr2-p5-nr2.c",
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07002172 "src/qs8-dwconv/gen/up8x9-minmax-sse2-mul16.c",
2173 "src/qs8-dwconv/gen/up16x9-minmax-sse2-mul16.c",
2174 "src/qs8-dwconv/gen/up24x9-minmax-sse2-mul16.c",
Marat Dukhan159688f2020-08-06 10:34:29 -07002175 "src/qs8-gavgpool/gen/7p7x-minmax-sse2-c8-acc2.c",
2176 "src/qs8-gavgpool/gen/7p7x-minmax-sse2-c16-acc2.c",
2177 "src/qs8-gavgpool/gen/7p7x-minmax-sse2-c24-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002178 "src/qs8-gavgpool/gen/7x-minmax-sse2-c8-acc2.c",
2179 "src/qs8-gavgpool/gen/7x-minmax-sse2-c16-acc2.c",
2180 "src/qs8-gavgpool/gen/7x-minmax-sse2-c24-acc2.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002181 "src/qs8-gemm/gen/1x4c2-minmax-sse2-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002182 "src/qs8-gemm/gen/1x4c2-minmax-sse2-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002183 "src/qs8-gemm/gen/1x4c2-xw-minmax-sse2.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002184 "src/qs8-gemm/gen/1x4c8-minmax-sse2-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002185 "src/qs8-gemm/gen/1x4c8-minmax-sse2-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002186 "src/qs8-gemm/gen/1x4c8-xw-minmax-sse2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002187 "src/qs8-gemm/gen/2x4c8-minmax-sse2-ld64.c",
2188 "src/qs8-gemm/gen/2x4c8-minmax-sse2-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002189 "src/qs8-gemm/gen/2x4c8-xw-minmax-sse2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002190 "src/qs8-gemm/gen/3x4c8-minmax-sse2-ld64.c",
2191 "src/qs8-gemm/gen/3x4c8-minmax-sse2-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002192 "src/qs8-gemm/gen/3x4c8-xw-minmax-sse2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002193 "src/qs8-gemm/gen/4x4c2-minmax-sse2-ld64.c",
2194 "src/qs8-gemm/gen/4x4c2-minmax-sse2-ld128.c",
2195 "src/qs8-gemm/gen/4x4c2-xw-minmax-sse2.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002196 "src/qs8-igemm/gen/1x4c2-minmax-sse2-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002197 "src/qs8-igemm/gen/1x4c2-minmax-sse2-ld128.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002198 "src/qs8-igemm/gen/1x4c8-minmax-sse2-ld64.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002199 "src/qs8-igemm/gen/1x4c8-minmax-sse2-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002200 "src/qs8-igemm/gen/2x4c8-minmax-sse2-ld64.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002201 "src/qs8-igemm/gen/2x4c8-minmax-sse2-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002202 "src/qs8-igemm/gen/3x4c8-minmax-sse2-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002203 "src/qs8-igemm/gen/3x4c8-minmax-sse2-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002204 "src/qs8-igemm/gen/4x4c2-minmax-sse2-ld64.c",
2205 "src/qs8-igemm/gen/4x4c2-minmax-sse2-ld128.c",
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07002206 "src/qs8-requantization/fp32-sse2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002207 "src/qs8-requantization/precise-sse2.c",
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07002208 "src/qs8-requantization/q31-sse2.c",
Marat Dukhand9f3ad42020-08-10 12:30:58 -07002209 "src/qs8-vadd/gen/minmax-sse2-mul16-ld64-x8.c",
2210 "src/qs8-vadd/gen/minmax-sse2-mul16-ld64-x16.c",
2211 "src/qs8-vadd/gen/minmax-sse2-mul16-ld64-x24.c",
2212 "src/qs8-vadd/gen/minmax-sse2-mul16-ld64-x32.c",
Marat Dukhan0270d9f2020-08-11 00:56:46 -07002213 "src/qs8-vaddc/gen/minmax-sse2-mul16-ld64-x8.c",
2214 "src/qs8-vaddc/gen/minmax-sse2-mul16-ld64-x16.c",
2215 "src/qs8-vaddc/gen/minmax-sse2-mul16-ld64-x24.c",
2216 "src/qs8-vaddc/gen/minmax-sse2-mul16-ld64-x32.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -07002217 "src/qu8-avgpool/9p8x-minmax-sse2-c8.c",
2218 "src/qu8-avgpool/9x-minmax-sse2-c8.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -07002219 "src/qu8-dwconv/up8x9-minmax-sse2.c",
2220 "src/qu8-gavgpool/7p7x-minmax-sse2-c8.c",
2221 "src/qu8-gavgpool/7x-minmax-sse2-c8.c",
2222 "src/qu8-gemm/2x4c8-minmax-sse2.c",
2223 "src/qu8-gemm/4x4c2-minmax-sse2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002224 "src/qu8-igemm/4x4c2-minmax-sse2.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07002225 "src/qu8-requantization/fp32-sse2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002226 "src/qu8-requantization/precise-sse2.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07002227 "src/qu8-requantization/q31-sse2.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -07002228 "src/qu8-vadd/minmax-sse2.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -07002229 "src/u8-clamp/sse2-x64.c",
Marat Dukhan99936602020-04-11 16:47:01 -07002230 "src/u8-maxpool/9p8x-minmax-sse2-c16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002231 "src/u8-rmax/sse2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002232 "src/x8-zip/x2-sse2.c",
2233 "src/x8-zip/x3-sse2.c",
2234 "src/x8-zip/x4-sse2.c",
2235 "src/x8-zip/xm-sse2.c",
Marat Dukhan57dccd82020-04-14 00:53:10 -07002236 "src/x32-unpool/sse2.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002237 "src/x32-zip/x2-sse2.c",
2238 "src/x32-zip/x3-sse2.c",
2239 "src/x32-zip/x4-sse2.c",
2240 "src/x32-zip/xm-sse2.c",
Marat Dukhanfe7acb62020-03-09 19:30:05 -07002241]
2242
2243SSSE3_UKERNELS = [
Frank Barchard35db7d02020-10-26 13:37:34 -07002244 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-1x4-acc2.c",
2245 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-1x4-acc3.c",
2246 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-1x4-acc4.c",
Marat Dukhan98f2eeb2020-10-23 23:13:41 -07002247 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-1x4.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07002248 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-2x4-acc2.c",
Marat Dukhan98f2eeb2020-10-23 23:13:41 -07002249 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-2x4.c",
2250 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-3x4.c",
2251 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-4x4.c",
2252 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-5x4.c",
2253 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-6x4.c",
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07002254 "src/qs8-dwconv/gen/up8x9-minmax-ssse3-mul16.c",
2255 "src/qs8-dwconv/gen/up16x9-minmax-ssse3-mul16.c",
2256 "src/qs8-dwconv/gen/up24x9-minmax-ssse3-mul16.c",
Marat Dukhan159688f2020-08-06 10:34:29 -07002257 "src/qs8-gavgpool/gen/7p7x-minmax-ssse3-c8-acc2.c",
2258 "src/qs8-gavgpool/gen/7p7x-minmax-ssse3-c16-acc2.c",
2259 "src/qs8-gavgpool/gen/7p7x-minmax-ssse3-c24-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002260 "src/qs8-gavgpool/gen/7x-minmax-ssse3-c8-acc2.c",
2261 "src/qs8-gavgpool/gen/7x-minmax-ssse3-c16-acc2.c",
2262 "src/qs8-gavgpool/gen/7x-minmax-ssse3-c24-acc2.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002263 "src/qs8-gemm/gen/1x4c2-minmax-ssse3-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002264 "src/qs8-gemm/gen/1x4c2-minmax-ssse3-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002265 "src/qs8-gemm/gen/1x4c2-xw-minmax-ssse3.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002266 "src/qs8-gemm/gen/1x4c8-minmax-ssse3-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002267 "src/qs8-gemm/gen/1x4c8-minmax-ssse3-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002268 "src/qs8-gemm/gen/1x4c8-xw-minmax-ssse3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002269 "src/qs8-gemm/gen/2x4c8-minmax-ssse3-ld64.c",
2270 "src/qs8-gemm/gen/2x4c8-minmax-ssse3-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002271 "src/qs8-gemm/gen/2x4c8-xw-minmax-ssse3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002272 "src/qs8-gemm/gen/3x4c8-minmax-ssse3-ld64.c",
2273 "src/qs8-gemm/gen/3x4c8-minmax-ssse3-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002274 "src/qs8-gemm/gen/3x4c8-xw-minmax-ssse3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002275 "src/qs8-gemm/gen/4x4c2-minmax-ssse3-ld64.c",
2276 "src/qs8-gemm/gen/4x4c2-minmax-ssse3-ld128.c",
2277 "src/qs8-gemm/gen/4x4c2-xw-minmax-ssse3.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002278 "src/qs8-igemm/gen/1x4c2-minmax-ssse3-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002279 "src/qs8-igemm/gen/1x4c2-minmax-ssse3-ld128.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002280 "src/qs8-igemm/gen/1x4c8-minmax-ssse3-ld64.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002281 "src/qs8-igemm/gen/1x4c8-minmax-ssse3-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002282 "src/qs8-igemm/gen/2x4c8-minmax-ssse3-ld64.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002283 "src/qs8-igemm/gen/2x4c8-minmax-ssse3-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002284 "src/qs8-igemm/gen/3x4c8-minmax-ssse3-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002285 "src/qs8-igemm/gen/3x4c8-minmax-ssse3-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002286 "src/qs8-igemm/gen/4x4c2-minmax-ssse3-ld64.c",
2287 "src/qs8-igemm/gen/4x4c2-minmax-ssse3-ld128.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -07002288 "src/qs8-requantization/precise-ssse3.c",
2289 "src/qs8-requantization/q31-ssse3.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07002290 "src/qu8-requantization/precise-ssse3.c",
2291 "src/qu8-requantization/q31-ssse3.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002292]
2293
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08002294SSE41_UKERNELS = [
Marat Dukhan40a672f2019-11-25 03:08:22 -08002295 "src/f32-prelu/gen/sse41-2x4.c",
2296 "src/f32-prelu/gen/sse41-2x8.c",
Marat Dukhand243c1a2020-09-17 18:48:10 -07002297 "src/f32-sigmoid/gen/sse41-lut64-p2-div-x4.c",
2298 "src/f32-sigmoid/gen/sse41-lut64-p2-div-x8.c",
2299 "src/f32-sigmoid/gen/sse41-lut64-p2-div-x12.c",
2300 "src/f32-sigmoid/gen/sse41-lut64-p2-div-x16.c",
2301 "src/f32-sigmoid/gen/sse41-lut64-p2-div-x20.c",
2302 "src/f32-sigmoid/gen/sse41-lut64-p2-div-x24.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002303 "src/f32-sigmoid/gen/sse41-p5-div-x4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08002304 "src/f32-sigmoid/gen/sse41-p5-div-x8.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002305 "src/f32-sigmoid/gen/sse41-p5-div-x12.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08002306 "src/f32-sigmoid/gen/sse41-p5-div-x16.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002307 "src/f32-sigmoid/gen/sse41-p5-div-x20.c",
2308 "src/f32-sigmoid/gen/sse41-p5-div-x24.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07002309 "src/f32-vlrelu/gen/vlrelu-sse41-x4.c",
2310 "src/f32-vlrelu/gen/vlrelu-sse41-x8.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07002311 "src/f32-vrnd/gen/vrndd-sse41-x4.c",
2312 "src/f32-vrnd/gen/vrndd-sse41-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002313 "src/f32-vrnd/gen/vrndne-sse41-x4.c",
2314 "src/f32-vrnd/gen/vrndne-sse41-x8.c",
2315 "src/f32-vrnd/gen/vrndu-sse41-x4.c",
2316 "src/f32-vrnd/gen/vrndu-sse41-x8.c",
2317 "src/f32-vrnd/gen/vrndz-sse41-x4.c",
2318 "src/f32-vrnd/gen/vrndz-sse41-x8.c",
2319 "src/math/roundd-sse41.c",
2320 "src/math/roundne-sse41.c",
2321 "src/math/roundu-sse41.c",
2322 "src/math/roundz-sse41.c",
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07002323 "src/qs8-dwconv/gen/up8x9-minmax-sse41-mul16.c",
2324 "src/qs8-dwconv/gen/up16x9-minmax-sse41-mul16.c",
2325 "src/qs8-dwconv/gen/up24x9-minmax-sse41-mul16.c",
Marat Dukhan159688f2020-08-06 10:34:29 -07002326 "src/qs8-gavgpool/gen/7p7x-minmax-sse41-c8-acc2.c",
2327 "src/qs8-gavgpool/gen/7p7x-minmax-sse41-c16-acc2.c",
2328 "src/qs8-gavgpool/gen/7p7x-minmax-sse41-c24-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002329 "src/qs8-gavgpool/gen/7x-minmax-sse41-c8-acc2.c",
2330 "src/qs8-gavgpool/gen/7x-minmax-sse41-c16-acc2.c",
2331 "src/qs8-gavgpool/gen/7x-minmax-sse41-c24-acc2.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002332 "src/qs8-gemm/gen/1x4c2-minmax-sse41-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002333 "src/qs8-gemm/gen/1x4c2-minmax-sse41-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002334 "src/qs8-gemm/gen/1x4c2-xw-minmax-sse41.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002335 "src/qs8-gemm/gen/1x4c8-minmax-sse41-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002336 "src/qs8-gemm/gen/1x4c8-minmax-sse41-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002337 "src/qs8-gemm/gen/1x4c8-xw-minmax-sse41.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002338 "src/qs8-gemm/gen/2x4c8-minmax-sse41-ld64.c",
2339 "src/qs8-gemm/gen/2x4c8-minmax-sse41-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002340 "src/qs8-gemm/gen/2x4c8-xw-minmax-sse41.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002341 "src/qs8-gemm/gen/3x4c8-minmax-sse41-ld64.c",
2342 "src/qs8-gemm/gen/3x4c8-minmax-sse41-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002343 "src/qs8-gemm/gen/3x4c8-xw-minmax-sse41.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002344 "src/qs8-gemm/gen/4x4c2-minmax-sse41-ld64.c",
2345 "src/qs8-gemm/gen/4x4c2-minmax-sse41-ld128.c",
2346 "src/qs8-gemm/gen/4x4c2-xw-minmax-sse41.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002347 "src/qs8-igemm/gen/1x4c2-minmax-sse41-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002348 "src/qs8-igemm/gen/1x4c2-minmax-sse41-ld128.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002349 "src/qs8-igemm/gen/1x4c8-minmax-sse41-ld64.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002350 "src/qs8-igemm/gen/1x4c8-minmax-sse41-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002351 "src/qs8-igemm/gen/2x4c8-minmax-sse41-ld64.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002352 "src/qs8-igemm/gen/2x4c8-minmax-sse41-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002353 "src/qs8-igemm/gen/3x4c8-minmax-sse41-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002354 "src/qs8-igemm/gen/3x4c8-minmax-sse41-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002355 "src/qs8-igemm/gen/4x4c2-minmax-sse41-ld64.c",
2356 "src/qs8-igemm/gen/4x4c2-minmax-sse41-ld128.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -07002357 "src/qs8-requantization/fp32-sse4.c",
2358 "src/qs8-requantization/precise-sse4.c",
2359 "src/qs8-requantization/q31-sse4.c",
Marat Dukhand9f3ad42020-08-10 12:30:58 -07002360 "src/qs8-vadd/gen/minmax-sse41-mul16-ld64-x8.c",
2361 "src/qs8-vadd/gen/minmax-sse41-mul16-ld64-x16.c",
2362 "src/qs8-vadd/gen/minmax-sse41-mul16-ld64-x24.c",
2363 "src/qs8-vadd/gen/minmax-sse41-mul16-ld64-x32.c",
Marat Dukhanbb9225e2020-09-06 22:40:56 -07002364 "src/qs8-vadd/gen/minmax-sse41-mul32-ld32-x8.c",
2365 "src/qs8-vadd/gen/minmax-sse41-mul32-ld32-x16.c",
2366 "src/qs8-vadd/gen/minmax-sse41-mul32-ld32-x24.c",
2367 "src/qs8-vadd/gen/minmax-sse41-mul32-ld32-x32.c",
Marat Dukhan0270d9f2020-08-11 00:56:46 -07002368 "src/qs8-vaddc/gen/minmax-sse41-mul16-ld64-x8.c",
2369 "src/qs8-vaddc/gen/minmax-sse41-mul16-ld64-x16.c",
2370 "src/qs8-vaddc/gen/minmax-sse41-mul16-ld64-x24.c",
2371 "src/qs8-vaddc/gen/minmax-sse41-mul16-ld64-x32.c",
Marat Dukhanbb9225e2020-09-06 22:40:56 -07002372 "src/qs8-vaddc/gen/minmax-sse41-mul32-ld32-x8.c",
2373 "src/qs8-vaddc/gen/minmax-sse41-mul32-ld32-x16.c",
2374 "src/qs8-vaddc/gen/minmax-sse41-mul32-ld32-x24.c",
2375 "src/qs8-vaddc/gen/minmax-sse41-mul32-ld32-x32.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07002376 "src/qu8-requantization/precise-sse4.c",
2377 "src/qu8-requantization/q31-sse4.c",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08002378]
2379
Marat Dukhan08c4a432019-10-03 09:29:21 -07002380AVX_UKERNELS = [
Marat Dukhan5c5fa962020-03-10 18:38:33 -07002381 "src/f32-clamp/gen/avx-x8.c",
2382 "src/f32-clamp/gen/avx-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002383 "src/f32-dwconv/gen/up8x4-minmax-avx-acc2.c",
2384 "src/f32-dwconv/gen/up8x4-minmax-avx.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002385 "src/f32-dwconv/gen/up8x9-minmax-avx-acc2.c",
2386 "src/f32-dwconv/gen/up8x9-minmax-avx.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002387 "src/f32-dwconv/gen/up8x25-minmax-avx-acc2.c",
2388 "src/f32-dwconv/gen/up8x25-minmax-avx.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002389 "src/f32-dwconv/gen/up16x4-minmax-avx-acc2.c",
2390 "src/f32-dwconv/gen/up16x4-minmax-avx.c",
2391 "src/f32-dwconv/gen/up16x9-minmax-avx-acc2.c",
2392 "src/f32-dwconv/gen/up16x9-minmax-avx.c",
2393 "src/f32-dwconv/gen/up16x25-minmax-avx-acc2.c",
2394 "src/f32-dwconv/gen/up16x25-minmax-avx.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002395 "src/f32-gemm/gen-inc/1x8inc-minmax-avx-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002396 "src/f32-gemm/gen-inc/1x16inc-minmax-avx-broadcast.c",
2397 "src/f32-gemm/gen-inc/3x16inc-minmax-avx-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002398 "src/f32-gemm/gen-inc/4x8inc-minmax-avx-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002399 "src/f32-gemm/gen-inc/4x16inc-minmax-avx-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002400 "src/f32-gemm/gen-inc/5x8inc-minmax-avx-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002401 "src/f32-gemm/gen-inc/5x16inc-minmax-avx-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002402 "src/f32-gemm/gen-inc/6x8inc-minmax-avx-broadcast.c",
2403 "src/f32-gemm/gen-inc/7x8inc-minmax-avx-broadcast.c",
2404 "src/f32-gemm/gen/1x8-minmax-avx-broadcast.c",
2405 "src/f32-gemm/gen/1x16-minmax-avx-broadcast.c",
2406 "src/f32-gemm/gen/3x16-minmax-avx-broadcast.c",
2407 "src/f32-gemm/gen/4x8-minmax-avx-broadcast.c",
2408 "src/f32-gemm/gen/4x16-minmax-avx-broadcast.c",
2409 "src/f32-gemm/gen/5x8-minmax-avx-broadcast.c",
2410 "src/f32-gemm/gen/5x16-minmax-avx-broadcast.c",
2411 "src/f32-gemm/gen/6x8-minmax-avx-broadcast.c",
2412 "src/f32-gemm/gen/7x8-minmax-avx-broadcast.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07002413 "src/f32-hswish/gen/hswish-avx-x8.c",
2414 "src/f32-hswish/gen/hswish-avx-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002415 "src/f32-igemm/gen/1x8-minmax-avx-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002416 "src/f32-igemm/gen/1x16-minmax-avx-broadcast.c",
2417 "src/f32-igemm/gen/3x16-minmax-avx-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002418 "src/f32-igemm/gen/4x8-minmax-avx-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002419 "src/f32-igemm/gen/4x16-minmax-avx-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002420 "src/f32-igemm/gen/5x8-minmax-avx-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002421 "src/f32-igemm/gen/5x16-minmax-avx-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002422 "src/f32-igemm/gen/6x8-minmax-avx-broadcast.c",
2423 "src/f32-igemm/gen/7x8-minmax-avx-broadcast.c",
Marat Dukhan90eca0a2020-03-11 00:52:23 -07002424 "src/f32-prelu/gen/avx-2x8.c",
2425 "src/f32-prelu/gen/avx-2x16.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07002426 "src/f32-relu/gen/avx-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002427 "src/f32-relu/gen/avx-x16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002428 "src/f32-rmax/avx.c",
T.J. Alumbaughdc2b29c2020-10-14 13:56:08 -07002429 "src/f32-sigmoid/gen/avx-rr2-p5-div-x8.c",
2430 "src/f32-sigmoid/gen/avx-rr2-p5-div-x16.c",
2431 "src/f32-sigmoid/gen/avx-rr2-p5-div-x24.c",
2432 "src/f32-sigmoid/gen/avx-rr2-p5-div-x32.c",
2433 "src/f32-sigmoid/gen/avx-rr2-p5-div-x40.c",
2434 "src/f32-sigmoid/gen/avx-rr2-p5-div-x48.c",
2435 "src/f32-sigmoid/gen/avx-rr2-p5-div-x56.c",
2436 "src/f32-sigmoid/gen/avx-rr2-p5-div-x64.c",
2437 "src/f32-sigmoid/gen/avx-rr2-p5-div-x72.c",
2438 "src/f32-sigmoid/gen/avx-rr2-p5-div-x80.c",
2439 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x8.c",
2440 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x16.c",
2441 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x24.c",
2442 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x32.c",
2443 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x40.c",
2444 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x48.c",
2445 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x56.c",
2446 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x64.c",
2447 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x72.c",
2448 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x80.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002449 "src/f32-vbinary/gen/vadd-minmax-avx-x8.c",
2450 "src/f32-vbinary/gen/vadd-minmax-avx-x16.c",
2451 "src/f32-vbinary/gen/vaddc-minmax-avx-x8.c",
2452 "src/f32-vbinary/gen/vaddc-minmax-avx-x16.c",
2453 "src/f32-vbinary/gen/vdiv-minmax-avx-x8.c",
2454 "src/f32-vbinary/gen/vdiv-minmax-avx-x16.c",
2455 "src/f32-vbinary/gen/vdivc-minmax-avx-x8.c",
2456 "src/f32-vbinary/gen/vdivc-minmax-avx-x16.c",
Marat Dukhan9a88efe2019-12-10 15:54:24 -08002457 "src/f32-vbinary/gen/vmax-avx-x8.c",
2458 "src/f32-vbinary/gen/vmax-avx-x16.c",
2459 "src/f32-vbinary/gen/vmaxc-avx-x8.c",
2460 "src/f32-vbinary/gen/vmaxc-avx-x16.c",
2461 "src/f32-vbinary/gen/vmin-avx-x8.c",
2462 "src/f32-vbinary/gen/vmin-avx-x16.c",
2463 "src/f32-vbinary/gen/vminc-avx-x8.c",
2464 "src/f32-vbinary/gen/vminc-avx-x16.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002465 "src/f32-vbinary/gen/vmul-minmax-avx-x8.c",
2466 "src/f32-vbinary/gen/vmul-minmax-avx-x16.c",
2467 "src/f32-vbinary/gen/vmulc-minmax-avx-x8.c",
2468 "src/f32-vbinary/gen/vmulc-minmax-avx-x16.c",
2469 "src/f32-vbinary/gen/vrdivc-minmax-avx-x8.c",
2470 "src/f32-vbinary/gen/vrdivc-minmax-avx-x16.c",
2471 "src/f32-vbinary/gen/vrsubc-minmax-avx-x8.c",
2472 "src/f32-vbinary/gen/vrsubc-minmax-avx-x16.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07002473 "src/f32-vbinary/gen/vsqrdiff-avx-x8.c",
2474 "src/f32-vbinary/gen/vsqrdiff-avx-x16.c",
2475 "src/f32-vbinary/gen/vsqrdiffc-avx-x8.c",
2476 "src/f32-vbinary/gen/vsqrdiffc-avx-x16.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002477 "src/f32-vbinary/gen/vsub-minmax-avx-x8.c",
2478 "src/f32-vbinary/gen/vsub-minmax-avx-x16.c",
2479 "src/f32-vbinary/gen/vsubc-minmax-avx-x8.c",
2480 "src/f32-vbinary/gen/vsubc-minmax-avx-x16.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07002481 "src/f32-vlrelu/gen/vlrelu-avx-x8.c",
2482 "src/f32-vlrelu/gen/vlrelu-avx-x16.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07002483 "src/f32-vrnd/gen/vrndd-avx-x8.c",
2484 "src/f32-vrnd/gen/vrndd-avx-x16.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002485 "src/f32-vrnd/gen/vrndne-avx-x8.c",
2486 "src/f32-vrnd/gen/vrndne-avx-x16.c",
2487 "src/f32-vrnd/gen/vrndu-avx-x8.c",
2488 "src/f32-vrnd/gen/vrndu-avx-x16.c",
2489 "src/f32-vrnd/gen/vrndz-avx-x8.c",
2490 "src/f32-vrnd/gen/vrndz-avx-x16.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07002491 "src/f32-vscale/avx-x32.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07002492 "src/f32-vsqrt/gen/avx-sqrt-x8.c",
2493 "src/f32-vsqrt/gen/avx-sqrt-x16.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07002494 "src/f32-vunary/gen/vabs-avx-x8.c",
2495 "src/f32-vunary/gen/vabs-avx-x16.c",
2496 "src/f32-vunary/gen/vneg-avx-x8.c",
2497 "src/f32-vunary/gen/vneg-avx-x16.c",
2498 "src/f32-vunary/gen/vsqr-avx-x8.c",
2499 "src/f32-vunary/gen/vsqr-avx-x16.c",
Marat Dukhanb7633f22020-11-20 16:34:56 -08002500 "src/math/exp-avx-rr2-p5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002501 "src/math/sigmoid-avx-rr2-lut64-p2-div.c",
Marat Dukhan2e4727b2020-09-24 00:48:39 -07002502 "src/math/sigmoid-avx-rr2-p5-div.c",
Marat Dukhan39ab16f2020-10-16 14:30:35 -07002503 "src/math/sigmoid-avx-rr2-p5-nr1.c",
2504 "src/math/sigmoid-avx-rr2-p5-nr2.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002505]
2506
Marat Dukhan1566fee2020-08-02 21:55:41 -07002507XOP_UKERNELS = [
2508 "src/qs8-gemm/gen/1x4c2-minmax-xop-ld64.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002509 "src/qs8-gemm/gen/1x4c2-minmax-xop-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002510 "src/qs8-gemm/gen/1x4c2-xw-minmax-xop.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002511 "src/qs8-gemm/gen/1x4c8-minmax-xop-ld64.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002512 "src/qs8-gemm/gen/1x4c8-minmax-xop-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002513 "src/qs8-gemm/gen/1x4c8-xw-minmax-xop.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002514 "src/qs8-gemm/gen/2x4c8-minmax-xop-ld64.c",
2515 "src/qs8-gemm/gen/2x4c8-minmax-xop-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002516 "src/qs8-gemm/gen/2x4c8-xw-minmax-xop.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002517 "src/qs8-gemm/gen/3x4c8-minmax-xop-ld64.c",
2518 "src/qs8-gemm/gen/3x4c8-minmax-xop-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002519 "src/qs8-gemm/gen/3x4c8-xw-minmax-xop.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002520 "src/qs8-gemm/gen/4x4c2-minmax-xop-ld64.c",
2521 "src/qs8-gemm/gen/4x4c2-minmax-xop-ld128.c",
2522 "src/qs8-gemm/gen/4x4c2-xw-minmax-xop.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002523 "src/qs8-igemm/gen/1x4c2-minmax-xop-ld64.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002524 "src/qs8-igemm/gen/1x4c2-minmax-xop-ld128.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002525 "src/qs8-igemm/gen/1x4c8-minmax-xop-ld64.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002526 "src/qs8-igemm/gen/1x4c8-minmax-xop-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002527 "src/qs8-igemm/gen/2x4c8-minmax-xop-ld64.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002528 "src/qs8-igemm/gen/2x4c8-minmax-xop-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002529 "src/qs8-igemm/gen/3x4c8-minmax-xop-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002530 "src/qs8-igemm/gen/3x4c8-minmax-xop-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002531 "src/qs8-igemm/gen/4x4c2-minmax-xop-ld64.c",
2532 "src/qs8-igemm/gen/4x4c2-minmax-xop-ld128.c",
Marat Dukhanbb9225e2020-09-06 22:40:56 -07002533 "src/qs8-vadd/gen/minmax-xop-mul32-ld32-x8.c",
2534 "src/qs8-vadd/gen/minmax-xop-mul32-ld32-x16.c",
2535 "src/qs8-vadd/gen/minmax-xop-mul32-ld32-x24.c",
2536 "src/qs8-vadd/gen/minmax-xop-mul32-ld32-x32.c",
2537 "src/qs8-vaddc/gen/minmax-xop-mul32-ld32-x8.c",
2538 "src/qs8-vaddc/gen/minmax-xop-mul32-ld32-x16.c",
2539 "src/qs8-vaddc/gen/minmax-xop-mul32-ld32-x24.c",
2540 "src/qs8-vaddc/gen/minmax-xop-mul32-ld32-x32.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002541]
2542
Marat Dukhanfda12b82019-11-21 12:27:59 -08002543FMA3_UKERNELS = [
Marat Dukhan1c587112020-04-08 20:04:28 -07002544 "src/f32-dwconv/gen/up8x4-minmax-fma3-acc2.c",
2545 "src/f32-dwconv/gen/up8x4-minmax-fma3.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002546 "src/f32-dwconv/gen/up8x9-minmax-fma3-acc2.c",
2547 "src/f32-dwconv/gen/up8x9-minmax-fma3.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002548 "src/f32-dwconv/gen/up8x25-minmax-fma3-acc2.c",
2549 "src/f32-dwconv/gen/up8x25-minmax-fma3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002550 "src/f32-dwconv/gen/up16x4-minmax-fma3-acc2.c",
2551 "src/f32-dwconv/gen/up16x4-minmax-fma3.c",
2552 "src/f32-dwconv/gen/up16x9-minmax-fma3-acc2.c",
2553 "src/f32-dwconv/gen/up16x9-minmax-fma3.c",
2554 "src/f32-dwconv/gen/up16x25-minmax-fma3-acc2.c",
2555 "src/f32-dwconv/gen/up16x25-minmax-fma3.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002556 "src/f32-gemm/gen-inc/1x8inc-minmax-fma3-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002557 "src/f32-gemm/gen-inc/1x16inc-minmax-fma3-broadcast.c",
2558 "src/f32-gemm/gen-inc/1x16s4inc-minmax-fma3-broadcast.c",
2559 "src/f32-gemm/gen-inc/3x16inc-minmax-fma3-broadcast.c",
2560 "src/f32-gemm/gen-inc/3x16s4inc-minmax-fma3-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002561 "src/f32-gemm/gen-inc/4x8inc-minmax-fma3-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002562 "src/f32-gemm/gen-inc/4x16inc-minmax-fma3-broadcast.c",
2563 "src/f32-gemm/gen-inc/4x16s4inc-minmax-fma3-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002564 "src/f32-gemm/gen-inc/5x8inc-minmax-fma3-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002565 "src/f32-gemm/gen-inc/5x16inc-minmax-fma3-broadcast.c",
2566 "src/f32-gemm/gen-inc/5x16s4inc-minmax-fma3-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002567 "src/f32-gemm/gen-inc/6x8inc-minmax-fma3-broadcast.c",
2568 "src/f32-gemm/gen-inc/7x8inc-minmax-fma3-broadcast.c",
2569 "src/f32-gemm/gen-inc/8x8inc-minmax-fma3-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002570 "src/f32-gemm/gen/1x8-minmax-fma3-broadcast.c",
2571 "src/f32-gemm/gen/1x16-minmax-fma3-broadcast.c",
2572 "src/f32-gemm/gen/1x16s4-minmax-fma3-broadcast.c",
2573 "src/f32-gemm/gen/3x16-minmax-fma3-broadcast.c",
2574 "src/f32-gemm/gen/3x16s4-minmax-fma3-broadcast.c",
2575 "src/f32-gemm/gen/4x8-minmax-fma3-broadcast.c",
2576 "src/f32-gemm/gen/4x16-minmax-fma3-broadcast.c",
2577 "src/f32-gemm/gen/4x16s4-minmax-fma3-broadcast.c",
2578 "src/f32-gemm/gen/5x8-minmax-fma3-broadcast.c",
2579 "src/f32-gemm/gen/5x16-minmax-fma3-broadcast.c",
2580 "src/f32-gemm/gen/5x16s4-minmax-fma3-broadcast.c",
2581 "src/f32-gemm/gen/6x8-minmax-fma3-broadcast.c",
2582 "src/f32-gemm/gen/7x8-minmax-fma3-broadcast.c",
2583 "src/f32-gemm/gen/8x8-minmax-fma3-broadcast.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07002584 "src/f32-hswish/gen/hswish-fma3-x8.c",
2585 "src/f32-hswish/gen/hswish-fma3-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002586 "src/f32-igemm/gen/1x8-minmax-fma3-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002587 "src/f32-igemm/gen/1x16-minmax-fma3-broadcast.c",
2588 "src/f32-igemm/gen/1x16s4-minmax-fma3-broadcast.c",
2589 "src/f32-igemm/gen/3x16-minmax-fma3-broadcast.c",
2590 "src/f32-igemm/gen/3x16s4-minmax-fma3-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002591 "src/f32-igemm/gen/4x8-minmax-fma3-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002592 "src/f32-igemm/gen/4x16-minmax-fma3-broadcast.c",
2593 "src/f32-igemm/gen/4x16s4-minmax-fma3-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002594 "src/f32-igemm/gen/5x8-minmax-fma3-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002595 "src/f32-igemm/gen/5x16-minmax-fma3-broadcast.c",
2596 "src/f32-igemm/gen/5x16s4-minmax-fma3-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002597 "src/f32-igemm/gen/6x8-minmax-fma3-broadcast.c",
2598 "src/f32-igemm/gen/7x8-minmax-fma3-broadcast.c",
2599 "src/f32-igemm/gen/8x8-minmax-fma3-broadcast.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07002600 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x8.c",
2601 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x16.c",
2602 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x24.c",
2603 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x32.c",
2604 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x40.c",
2605 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x48.c",
2606 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x56.c",
2607 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x64.c",
Marat Dukhan84000762020-06-29 18:38:43 -07002608 "src/math/sqrt-fma3-nr1fma.c",
Marat Dukhan84000762020-06-29 18:38:43 -07002609 "src/math/sqrt-fma3-nr1fma1adj.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002610 "src/math/sqrt-fma3-nr2fma.c",
Marat Dukhanfda12b82019-11-21 12:27:59 -08002611]
2612
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002613AVX2_UKERNELS = [
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002614 "src/f32-raddexpminusmax/gen/avx2-p5-x64-acc2.c",
2615 "src/f32-raddexpminusmax/gen/avx2-p5-x64-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002616 "src/f32-raddexpminusmax/gen/avx2-p5-x64.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002617 "src/f32-raddexpminusmax/gen/avx2-p5-x72-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002618 "src/f32-raddexpminusmax/gen/avx2-p5-x72.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002619 "src/f32-raddexpminusmax/gen/avx2-p5-x80-acc2.c",
2620 "src/f32-raddexpminusmax/gen/avx2-p5-x80-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002621 "src/f32-raddexpminusmax/gen/avx2-p5-x80.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002622 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc2.c",
2623 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc3.c",
2624 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc6.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002625 "src/f32-raddexpminusmax/gen/avx2-p5-x96.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002626 "src/f32-raddextexp/gen/avx2-p5-x64-acc2.c",
2627 "src/f32-raddextexp/gen/avx2-p5-x64-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002628 "src/f32-raddextexp/gen/avx2-p5-x64.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002629 "src/f32-raddextexp/gen/avx2-p5-x72-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002630 "src/f32-raddextexp/gen/avx2-p5-x72.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002631 "src/f32-raddextexp/gen/avx2-p5-x80-acc2.c",
2632 "src/f32-raddextexp/gen/avx2-p5-x80-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002633 "src/f32-raddextexp/gen/avx2-p5-x80.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002634 "src/f32-raddextexp/gen/avx2-p5-x96-acc2.c",
2635 "src/f32-raddextexp/gen/avx2-p5-x96-acc3.c",
2636 "src/f32-raddextexp/gen/avx2-p5-x96-acc6.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002637 "src/f32-raddextexp/gen/avx2-p5-x96.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002638 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64-acc2.c",
2639 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002640 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002641 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x72-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002642 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x72.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002643 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80-acc2.c",
2644 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002645 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002646 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc2.c",
2647 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc3.c",
2648 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc6.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002649 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96.c",
Marat Dukhanfa0a4322020-01-06 16:14:29 -08002650 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x8.c",
2651 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x16.c",
2652 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x24.c",
2653 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x32.c",
2654 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x40.c",
2655 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x48.c",
2656 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x56.c",
2657 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x64.c",
2658 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x72.c",
2659 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x80.c",
2660 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x8.c",
2661 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x16.c",
2662 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x24.c",
2663 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x32.c",
2664 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x40.c",
2665 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x48.c",
2666 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x56.c",
2667 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x64.c",
2668 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x72.c",
2669 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x80.c",
2670 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x8.c",
2671 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x16.c",
2672 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x24.c",
2673 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x32.c",
2674 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x40.c",
2675 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x48.c",
2676 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x56.c",
2677 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x64.c",
2678 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x72.c",
2679 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x80.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002680 "src/f32-vscaleexpminusmax/gen/avx2-p5-x8.c",
2681 "src/f32-vscaleexpminusmax/gen/avx2-p5-x16.c",
2682 "src/f32-vscaleexpminusmax/gen/avx2-p5-x24.c",
2683 "src/f32-vscaleexpminusmax/gen/avx2-p5-x32.c",
2684 "src/f32-vscaleexpminusmax/gen/avx2-p5-x40.c",
2685 "src/f32-vscaleexpminusmax/gen/avx2-p5-x48.c",
2686 "src/f32-vscaleexpminusmax/gen/avx2-p5-x56.c",
2687 "src/f32-vscaleexpminusmax/gen/avx2-p5-x64.c",
2688 "src/f32-vscaleexpminusmax/gen/avx2-p5-x72.c",
2689 "src/f32-vscaleexpminusmax/gen/avx2-p5-x80.c",
2690 "src/f32-vscaleexpminusmax/gen/avx2-p5-x88.c",
2691 "src/f32-vscaleexpminusmax/gen/avx2-p5-x96.c",
2692 "src/f32-vscaleextexp/gen/avx2-p5-x8.c",
2693 "src/f32-vscaleextexp/gen/avx2-p5-x16.c",
2694 "src/f32-vscaleextexp/gen/avx2-p5-x24.c",
2695 "src/f32-vscaleextexp/gen/avx2-p5-x32.c",
2696 "src/f32-vscaleextexp/gen/avx2-p5-x40.c",
2697 "src/f32-vscaleextexp/gen/avx2-p5-x48.c",
2698 "src/f32-vscaleextexp/gen/avx2-p5-x56.c",
2699 "src/f32-vscaleextexp/gen/avx2-p5-x64.c",
2700 "src/f32-vscaleextexp/gen/avx2-p5-x72.c",
2701 "src/f32-vscaleextexp/gen/avx2-p5-x80.c",
2702 "src/f32-vscaleextexp/gen/avx2-p5-x88.c",
2703 "src/f32-vscaleextexp/gen/avx2-p5-x96.c",
Marat Dukhanb7633f22020-11-20 16:34:56 -08002704 "src/math/exp-avx2-rr2-lut8-p3-perm.c",
2705 "src/math/exp-avx2-rr2-lut8-p4-perm.c",
2706 "src/math/exp-avx2-rr2-p5.c",
Marat Dukhan9dd119a2020-11-20 18:20:04 -08002707 "src/math/expminus-avx2-rr2-p5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002708 "src/math/extexp-avx2-p5.c",
2709 "src/math/sigmoid-avx2-rr1-lut64-p2-gather-div.c",
2710 "src/math/sigmoid-avx2-rr1-lut64-p2-gather-nr1fma.c",
2711 "src/math/sigmoid-avx2-rr1-lut64-p2-gather-nr2fma.c",
2712 "src/math/sigmoid-avx2-rr1-lut64-p2-gather-nr2fma1adj.c",
2713 "src/math/sigmoid-avx2-rr1-p5-div.c",
2714 "src/math/sigmoid-avx2-rr1-p5-nr1fma.c",
2715 "src/math/sigmoid-avx2-rr1-p5-nr2fma.c",
2716 "src/math/sigmoid-avx2-rr2-lut64-p2-gather-div.c",
2717 "src/math/sigmoid-avx2-rr2-lut64-p2-gather-nr1fma.c",
2718 "src/math/sigmoid-avx2-rr2-lut64-p2-gather-nr2fma.c",
2719 "src/math/sigmoid-avx2-rr2-lut64-p2-gather-nr2fma1adj.c",
2720 "src/math/sigmoid-avx2-rr2-p5-div.c",
2721 "src/math/sigmoid-avx2-rr2-p5-nr1fma.c",
2722 "src/math/sigmoid-avx2-rr2-p5-nr2fma.c",
Marat Dukhand65a1522020-08-04 19:28:18 -07002723 "src/qs8-dwconv/gen/up8x9-minmax-avx2-mul32.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002724 "src/qs8-dwconv/gen/up16x9-minmax-avx2-mul16.c",
Marat Dukhand65a1522020-08-04 19:28:18 -07002725 "src/qs8-dwconv/gen/up16x9-minmax-avx2-mul32.c",
2726 "src/qs8-dwconv/gen/up24x9-minmax-avx2-mul32.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002727 "src/qs8-dwconv/gen/up32x9-minmax-avx2-mul16.c",
Marat Dukhand65a1522020-08-04 19:28:18 -07002728 "src/qs8-dwconv/gen/up32x9-minmax-avx2-mul32.c",
Marat Dukhan12809522020-08-02 22:23:51 -07002729 "src/qs8-gemm/gen/1x8c8-minmax-avx2.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002730 "src/qs8-gemm/gen/1x8c8-xw-minmax-avx2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002731 "src/qs8-gemm/gen/2x8c8-minmax-avx2.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002732 "src/qs8-gemm/gen/2x8c8-xw-minmax-avx2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002733 "src/qs8-gemm/gen/3x8c8-minmax-avx2.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002734 "src/qs8-gemm/gen/3x8c8-xw-minmax-avx2.c",
Marat Dukhan12809522020-08-02 22:23:51 -07002735 "src/qs8-igemm/gen/1x8c8-minmax-avx2.c",
2736 "src/qs8-igemm/gen/2x8c8-minmax-avx2.c",
2737 "src/qs8-igemm/gen/3x8c8-minmax-avx2.c",
Marat Dukhane6dc0b62020-09-08 23:57:14 -07002738 "src/qs8-vadd/gen/minmax-avx2-mul32-ld64-x8.c",
2739 "src/qs8-vadd/gen/minmax-avx2-mul32-ld64-x16.c",
2740 "src/qs8-vadd/gen/minmax-avx2-mul32-ld64-x24.c",
2741 "src/qs8-vadd/gen/minmax-avx2-mul32-ld64-x32.c",
2742 "src/qs8-vaddc/gen/minmax-avx2-mul32-ld64-x8.c",
2743 "src/qs8-vaddc/gen/minmax-avx2-mul32-ld64-x16.c",
2744 "src/qs8-vaddc/gen/minmax-avx2-mul32-ld64-x24.c",
2745 "src/qs8-vaddc/gen/minmax-avx2-mul32-ld64-x32.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002746]
2747
Marat Dukhan08c4a432019-10-03 09:29:21 -07002748AVX512F_UKERNELS = [
Marat Dukhan5c5fa962020-03-10 18:38:33 -07002749 "src/f32-clamp/gen/avx512f-x16.c",
2750 "src/f32-clamp/gen/avx512f-x32.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002751 "src/f32-dwconv/gen/up16x4-minmax-avx512f-acc2.c",
2752 "src/f32-dwconv/gen/up16x4-minmax-avx512f.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002753 "src/f32-dwconv/gen/up16x9-minmax-avx512f-acc2.c",
2754 "src/f32-dwconv/gen/up16x9-minmax-avx512f.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002755 "src/f32-dwconv/gen/up16x25-minmax-avx512f-acc2.c",
2756 "src/f32-dwconv/gen/up16x25-minmax-avx512f.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002757 "src/f32-dwconv/gen/up32x4-minmax-avx512f-acc2.c",
2758 "src/f32-dwconv/gen/up32x4-minmax-avx512f.c",
2759 "src/f32-dwconv/gen/up32x9-minmax-avx512f-acc2.c",
2760 "src/f32-dwconv/gen/up32x9-minmax-avx512f.c",
2761 "src/f32-dwconv/gen/up32x25-minmax-avx512f-acc2.c",
2762 "src/f32-dwconv/gen/up32x25-minmax-avx512f.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002763 "src/f32-gemm/gen-inc/1x16inc-minmax-avx512f-broadcast.c",
2764 "src/f32-gemm/gen-inc/4x16inc-minmax-avx512f-broadcast.c",
2765 "src/f32-gemm/gen-inc/5x16inc-minmax-avx512f-broadcast.c",
2766 "src/f32-gemm/gen-inc/6x16inc-minmax-avx512f-broadcast.c",
2767 "src/f32-gemm/gen-inc/7x16inc-minmax-avx512f-broadcast.c",
2768 "src/f32-gemm/gen-inc/8x16inc-minmax-avx512f-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002769 "src/f32-gemm/gen/1x16-minmax-avx512f-broadcast.c",
2770 "src/f32-gemm/gen/4x16-minmax-avx512f-broadcast.c",
2771 "src/f32-gemm/gen/5x16-minmax-avx512f-broadcast.c",
2772 "src/f32-gemm/gen/6x16-minmax-avx512f-broadcast.c",
2773 "src/f32-gemm/gen/7x16-minmax-avx512f-broadcast.c",
2774 "src/f32-gemm/gen/8x16-minmax-avx512f-broadcast.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07002775 "src/f32-hswish/gen/hswish-avx512f-x16.c",
2776 "src/f32-hswish/gen/hswish-avx512f-x32.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002777 "src/f32-igemm/gen/1x16-minmax-avx512f-broadcast.c",
2778 "src/f32-igemm/gen/4x16-minmax-avx512f-broadcast.c",
2779 "src/f32-igemm/gen/5x16-minmax-avx512f-broadcast.c",
2780 "src/f32-igemm/gen/6x16-minmax-avx512f-broadcast.c",
2781 "src/f32-igemm/gen/7x16-minmax-avx512f-broadcast.c",
2782 "src/f32-igemm/gen/8x16-minmax-avx512f-broadcast.c",
Marat Dukhan90eca0a2020-03-11 00:52:23 -07002783 "src/f32-prelu/gen/avx512f-2x16.c",
2784 "src/f32-prelu/gen/avx512f-2x32.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002785 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128-acc2.c",
2786 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002787 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002788 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x144-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002789 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x144.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002790 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160-acc2.c",
2791 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002792 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002793 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc2.c",
2794 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc3.c",
2795 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc6.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002796 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002797 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128-acc2.c",
2798 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002799 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002800 "src/f32-raddextexp/gen/avx512f-p5-scalef-x144-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002801 "src/f32-raddextexp/gen/avx512f-p5-scalef-x144.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002802 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160-acc2.c",
2803 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002804 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002805 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc2.c",
2806 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc3.c",
2807 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc6.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002808 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002809 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128-acc2.c",
2810 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002811 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002812 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x144-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002813 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x144.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002814 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160-acc2.c",
2815 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002816 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002817 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc2.c",
2818 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc3.c",
2819 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc6.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002820 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07002821 "src/f32-relu/gen/avx512f-x16.c",
2822 "src/f32-relu/gen/avx512f-x32.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002823 "src/f32-rmax/avx512f.c",
Marat Dukhand9ca7e62020-09-23 23:45:29 -07002824 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x16.c",
2825 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x32.c",
2826 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x48.c",
2827 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x64.c",
2828 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x80.c",
2829 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x96.c",
2830 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x112.c",
2831 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x128.c",
2832 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x16.c",
2833 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x32.c",
2834 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x48.c",
2835 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x64.c",
2836 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x80.c",
2837 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x96.c",
2838 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x112.c",
2839 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002840 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x16.c",
2841 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x32.c",
2842 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x48.c",
2843 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x64.c",
2844 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x80.c",
2845 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x96.c",
2846 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x112.c",
2847 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x128.c",
2848 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x16.c",
2849 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x32.c",
2850 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x48.c",
2851 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x64.c",
2852 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x80.c",
2853 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x96.c",
2854 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x112.c",
2855 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x128.c",
Marat Dukhand9ca7e62020-09-23 23:45:29 -07002856 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x16.c",
2857 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x32.c",
2858 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x48.c",
2859 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x64.c",
2860 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x80.c",
2861 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x96.c",
2862 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x112.c",
2863 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x128.c",
2864 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x16.c",
2865 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x32.c",
2866 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x48.c",
2867 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x64.c",
2868 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x80.c",
2869 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x96.c",
2870 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x112.c",
2871 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x128.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002872 "src/f32-vbinary/gen/vadd-minmax-avx512f-x16.c",
2873 "src/f32-vbinary/gen/vadd-minmax-avx512f-x32.c",
2874 "src/f32-vbinary/gen/vaddc-minmax-avx512f-x16.c",
2875 "src/f32-vbinary/gen/vaddc-minmax-avx512f-x32.c",
2876 "src/f32-vbinary/gen/vdiv-minmax-avx512f-x16.c",
2877 "src/f32-vbinary/gen/vdiv-minmax-avx512f-x32.c",
2878 "src/f32-vbinary/gen/vdivc-minmax-avx512f-x16.c",
2879 "src/f32-vbinary/gen/vdivc-minmax-avx512f-x32.c",
Marat Dukhan9a88efe2019-12-10 15:54:24 -08002880 "src/f32-vbinary/gen/vmax-avx512f-x16.c",
2881 "src/f32-vbinary/gen/vmax-avx512f-x32.c",
2882 "src/f32-vbinary/gen/vmaxc-avx512f-x16.c",
2883 "src/f32-vbinary/gen/vmaxc-avx512f-x32.c",
2884 "src/f32-vbinary/gen/vmin-avx512f-x16.c",
2885 "src/f32-vbinary/gen/vmin-avx512f-x32.c",
2886 "src/f32-vbinary/gen/vminc-avx512f-x16.c",
2887 "src/f32-vbinary/gen/vminc-avx512f-x32.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002888 "src/f32-vbinary/gen/vmul-minmax-avx512f-x16.c",
2889 "src/f32-vbinary/gen/vmul-minmax-avx512f-x32.c",
2890 "src/f32-vbinary/gen/vmulc-minmax-avx512f-x16.c",
2891 "src/f32-vbinary/gen/vmulc-minmax-avx512f-x32.c",
2892 "src/f32-vbinary/gen/vrdivc-minmax-avx512f-x16.c",
2893 "src/f32-vbinary/gen/vrdivc-minmax-avx512f-x32.c",
2894 "src/f32-vbinary/gen/vrsubc-minmax-avx512f-x16.c",
2895 "src/f32-vbinary/gen/vrsubc-minmax-avx512f-x32.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07002896 "src/f32-vbinary/gen/vsqrdiff-avx512f-x16.c",
2897 "src/f32-vbinary/gen/vsqrdiff-avx512f-x32.c",
2898 "src/f32-vbinary/gen/vsqrdiffc-avx512f-x16.c",
2899 "src/f32-vbinary/gen/vsqrdiffc-avx512f-x32.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002900 "src/f32-vbinary/gen/vsub-minmax-avx512f-x16.c",
2901 "src/f32-vbinary/gen/vsub-minmax-avx512f-x32.c",
2902 "src/f32-vbinary/gen/vsubc-minmax-avx512f-x16.c",
2903 "src/f32-vbinary/gen/vsubc-minmax-avx512f-x32.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07002904 "src/f32-vlrelu/gen/vlrelu-avx512f-x16.c",
2905 "src/f32-vlrelu/gen/vlrelu-avx512f-x32.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002906 "src/f32-vrnd/gen/vrndd-avx512f-x16.c",
2907 "src/f32-vrnd/gen/vrndd-avx512f-x32.c",
2908 "src/f32-vrnd/gen/vrndne-avx512f-x16.c",
2909 "src/f32-vrnd/gen/vrndne-avx512f-x32.c",
2910 "src/f32-vrnd/gen/vrndu-avx512f-x16.c",
2911 "src/f32-vrnd/gen/vrndu-avx512f-x32.c",
2912 "src/f32-vrnd/gen/vrndz-avx512f-x16.c",
2913 "src/f32-vrnd/gen/vrndz-avx512f-x32.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07002914 "src/f32-vscale/avx512f-x64.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002915 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x16.c",
2916 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x32.c",
2917 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x48.c",
2918 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x64.c",
2919 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x80.c",
2920 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x96.c",
2921 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x112.c",
2922 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x128.c",
2923 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x144.c",
2924 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x160.c",
2925 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x176.c",
2926 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x192.c",
2927 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x16.c",
2928 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x32.c",
2929 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x48.c",
2930 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x64.c",
2931 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x80.c",
2932 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x96.c",
2933 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x112.c",
2934 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x128.c",
2935 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x144.c",
2936 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x160.c",
2937 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x176.c",
2938 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x192.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07002939 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x16.c",
2940 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x32.c",
2941 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x48.c",
2942 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x64.c",
2943 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x80.c",
2944 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x96.c",
2945 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x112.c",
2946 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x128.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07002947 "src/f32-vunary/gen/vabs-avx512f-x16.c",
2948 "src/f32-vunary/gen/vabs-avx512f-x32.c",
2949 "src/f32-vunary/gen/vneg-avx512f-x16.c",
2950 "src/f32-vunary/gen/vneg-avx512f-x32.c",
2951 "src/f32-vunary/gen/vsqr-avx512f-x16.c",
2952 "src/f32-vunary/gen/vsqr-avx512f-x32.c",
Marat Dukhanb7633f22020-11-20 16:34:56 -08002953 "src/math/exp-avx512f-rr2-lut16-p3-perm-scalef.c",
2954 "src/math/exp-avx512f-rr2-lut16-p3-perm.c",
2955 "src/math/exp-avx512f-rr2-lut32-p2-perm2-scalef.c",
2956 "src/math/exp-avx512f-rr2-lut32-p2-perm2.c",
2957 "src/math/exp-avx512f-rr2-p5-scalef.c",
2958 "src/math/exp-avx512f-rr2-p5.c",
Marat Dukhan98ba4412019-10-23 02:14:28 -07002959 "src/math/extexp-avx512f-p5.c",
Marat Dukhan6a34c5f2020-09-22 21:44:15 -07002960 "src/math/sigmoid-avx512f-rr1-lut16-p3-perm-scalef-div.c",
Marat Dukhan6a34c5f2020-09-22 21:44:15 -07002961 "src/math/sigmoid-avx512f-rr1-lut16-p3-perm-scalef-nr1fma.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002962 "src/math/sigmoid-avx512f-rr1-lut16-p3-perm-scalef-nr1fma1adj.c",
Marat Dukhan6a34c5f2020-09-22 21:44:15 -07002963 "src/math/sigmoid-avx512f-rr1-lut32-p2-perm2-scalef-div.c",
Marat Dukhan6a34c5f2020-09-22 21:44:15 -07002964 "src/math/sigmoid-avx512f-rr1-lut32-p2-perm2-scalef-nr1fma.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002965 "src/math/sigmoid-avx512f-rr1-lut32-p2-perm2-scalef-nr1fma1adj.c",
Marat Dukhan36173d22020-10-15 17:14:26 -07002966 "src/math/sigmoid-avx512f-rr1-lut64-p2-gather-scalef-div.c",
Marat Dukhan36173d22020-10-15 17:14:26 -07002967 "src/math/sigmoid-avx512f-rr1-lut64-p2-gather-scalef-nr1fma.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002968 "src/math/sigmoid-avx512f-rr1-lut64-p2-gather-scalef-nr1fma1adj.c",
2969 "src/math/sigmoid-avx512f-rr1-p5-scalef-div.c",
2970 "src/math/sigmoid-avx512f-rr1-p5-scalef-nr1fma.c",
2971 "src/math/sigmoid-avx512f-rr1-p5-scalef-nr1fma1adj.c",
2972 "src/math/sigmoid-avx512f-rr2-lut16-p3-perm-scalef-div.c",
2973 "src/math/sigmoid-avx512f-rr2-lut16-p3-perm-scalef-nr1fma.c",
2974 "src/math/sigmoid-avx512f-rr2-lut16-p3-perm-scalef-nr1fma1adj.c",
2975 "src/math/sigmoid-avx512f-rr2-lut32-p2-perm2-scalef-div.c",
2976 "src/math/sigmoid-avx512f-rr2-lut32-p2-perm2-scalef-nr1fma.c",
2977 "src/math/sigmoid-avx512f-rr2-lut32-p2-perm2-scalef-nr1fma1adj.c",
Marat Dukhan36173d22020-10-15 17:14:26 -07002978 "src/math/sigmoid-avx512f-rr2-lut64-p2-gather-scalef-div.c",
Marat Dukhan36173d22020-10-15 17:14:26 -07002979 "src/math/sigmoid-avx512f-rr2-lut64-p2-gather-scalef-nr1fma.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002980 "src/math/sigmoid-avx512f-rr2-lut64-p2-gather-scalef-nr1fma1adj.c",
2981 "src/math/sigmoid-avx512f-rr2-p5-scalef-div.c",
2982 "src/math/sigmoid-avx512f-rr2-p5-scalef-nr1fma.c",
2983 "src/math/sigmoid-avx512f-rr2-p5-scalef-nr1fma1adj.c",
Marat Dukhan84000762020-06-29 18:38:43 -07002984 "src/math/sqrt-avx512f-nr1fma.c",
Marat Dukhan84000762020-06-29 18:38:43 -07002985 "src/math/sqrt-avx512f-nr1fma1adj.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002986 "src/math/sqrt-avx512f-nr2fma.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002987]
2988
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07002989AVX512SKX_UKERNELS = [
Marat Dukhan2ffc5e62020-09-06 22:33:38 -07002990 "src/qs8-dwconv/gen/up16x9-minmax-avx512skx-mul32.c",
2991 "src/qs8-dwconv/gen/up32x9-minmax-avx512skx-mul32.c",
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07002992 "src/qs8-gemm/gen/1x16c8-minmax-avx512skx.c",
2993 "src/qs8-gemm/gen/2x16c8-minmax-avx512skx.c",
2994 "src/qs8-gemm/gen/3x16c8-minmax-avx512skx.c",
2995 "src/qs8-gemm/gen/4x16c8-minmax-avx512skx.c",
2996 "src/qs8-igemm/gen/1x16c8-minmax-avx512skx.c",
2997 "src/qs8-igemm/gen/2x16c8-minmax-avx512skx.c",
2998 "src/qs8-igemm/gen/3x16c8-minmax-avx512skx.c",
2999 "src/qs8-igemm/gen/4x16c8-minmax-avx512skx.c",
3000]
3001
Frank Barchardbcedc082020-08-17 18:00:51 -07003002WASM32_ASM_UKERNELS = [
3003 "src/f32-relu/wasm_shr_x1.S",
3004 "src/f32-relu/wasm_shr_x2.S",
3005 "src/f32-relu/wasm_shr_x4.S",
3006]
3007
Marat Dukhan08c4a432019-10-03 09:29:21 -07003008AARCH32_ASM_UKERNELS = [
Marat Dukhan32f93812020-05-17 20:31:21 -07003009 "src/f32-gemm/4x4-aarch32-vfp-ld64.S",
Marat Dukhan3b98f6b2020-05-17 10:09:22 -07003010 "src/f32-gemm/4x4-minmax-aarch32-vfp-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003011 "src/f32-gemm/4x8-minmax-aarch32-neon-cortex-a53.S",
3012 "src/f32-gemm/4x8-minmax-aarch32-neon-cortex-a55.S",
Frank Barchard490febe2020-07-16 18:42:17 -07003013 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-cortex-a7.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003014 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-cortex-a75.S",
Frank Barchard569561d2020-06-17 13:11:12 -07003015 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-ld64.S",
Frank Barchard490febe2020-07-16 18:42:17 -07003016 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-pld-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003017 "src/f32-igemm/4x8-minmax-aarch32-neon-cortex-a53.S",
3018 "src/f32-igemm/4x8-minmax-aarch32-neon-cortex-a55.S",
Frank Barchard490febe2020-07-16 18:42:17 -07003019 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-cortex-a7.S",
3020 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-cortex-a75.S",
3021 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-ld64.S",
3022 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-pld-cortex-a75.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003023]
3024
3025AARCH64_ASM_UKERNELS = [
Frank Barchardbddfbcd2020-04-15 12:32:41 -07003026 "src/f16-gemm/gen-inc/1x8inc-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchard04336c12020-10-22 16:48:55 -07003027 "src/f16-gemm/gen-inc/1x16inc-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchardbddfbcd2020-04-15 12:32:41 -07003028 "src/f16-gemm/gen-inc/4x8inc-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchard04336c12020-10-22 16:48:55 -07003029 "src/f16-gemm/gen-inc/4x16inc-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchardbddfbcd2020-04-15 12:32:41 -07003030 "src/f16-gemm/gen-inc/6x8inc-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchard04336c12020-10-22 16:48:55 -07003031 "src/f16-gemm/gen-inc/6x16inc-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchard3b8e5662020-04-20 12:12:53 -07003032 "src/f16-gemm/gen-inc/8x8inc-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchard04336c12020-10-22 16:48:55 -07003033 "src/f16-gemm/gen/1x8-minmax-aarch64-neonfp16arith-ld64.S",
3034 "src/f16-gemm/gen/1x16-minmax-aarch64-neonfp16arith-ld32.S",
3035 "src/f16-gemm/gen/4x8-minmax-aarch64-neonfp16arith-ld64.S",
3036 "src/f16-gemm/gen/4x16-minmax-aarch64-neonfp16arith-ld32.S",
3037 "src/f16-gemm/gen/6x8-minmax-aarch64-neonfp16arith-ld64.S",
3038 "src/f16-gemm/gen/6x16-minmax-aarch64-neonfp16arith-ld32.S",
3039 "src/f16-gemm/gen/8x8-minmax-aarch64-neonfp16arith-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003040 "src/f32-dwconv/up4x9-minmax-aarch64-neonfma-cortex-a55.S",
3041 "src/f32-dwconv/up4x9-minmax-aarch64-neonfma.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003042 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a53.S",
3043 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a57.S",
3044 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a75.S",
Frank Barchard04336c12020-10-22 16:48:55 -07003045 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-ld64.S",
3046 "src/f32-gemm/gen-inc/1x12inc-minmax-aarch64-neonfma-cortex-a53.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003047 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a53.S",
3048 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a55.S",
3049 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a57.S",
3050 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003051 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-ld64.S",
Frank Barchard04336c12020-10-22 16:48:55 -07003052 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-ld128.S",
3053 "src/f32-gemm/gen-inc/4x12inc-minmax-aarch64-neonfma-cortex-a53.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003054 "src/f32-gemm/gen-inc/5x8inc-minmax-aarch64-neonfma-cortex-a57.S",
3055 "src/f32-gemm/gen-inc/5x8inc-minmax-aarch64-neonfma-cortex-a75.S",
3056 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a53.S",
3057 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a55.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003058 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a57.S",
Frank Barchard04336c12020-10-22 16:48:55 -07003059 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a73.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003060 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003061 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-ld64.S",
Frank Barchard04336c12020-10-22 16:48:55 -07003062 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-ld128.S",
3063 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a53.S",
3064 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a57.S",
3065 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a75.S",
3066 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-ld64.S",
3067 "src/f32-gemm/gen/1x12-minmax-aarch64-neonfma-cortex-a53.S",
3068 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a53.S",
3069 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a55.S",
3070 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a57.S",
3071 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a75.S",
3072 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-ld64.S",
3073 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-ld128.S",
3074 "src/f32-gemm/gen/4x12-minmax-aarch64-neonfma-cortex-a53.S",
3075 "src/f32-gemm/gen/5x8-minmax-aarch64-neonfma-cortex-a57.S",
3076 "src/f32-gemm/gen/5x8-minmax-aarch64-neonfma-cortex-a75.S",
3077 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a53.S",
3078 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a55.S",
3079 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a57.S",
3080 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a73.S",
3081 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a75.S",
3082 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-ld64.S",
3083 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-ld128.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003084 "src/f32-igemm/1x8-minmax-aarch64-neonfma-cortex-a53.S",
Frank Barchard04336c12020-10-22 16:48:55 -07003085 "src/f32-igemm/1x12-minmax-aarch64-neonfma-cortex-a53.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003086 "src/f32-igemm/4x8-minmax-aarch64-neonfma-cortex-a53.S",
3087 "src/f32-igemm/4x8-minmax-aarch64-neonfma-cortex-a55.S",
Frank Barchard04336c12020-10-22 16:48:55 -07003088 "src/f32-igemm/4x12-minmax-aarch64-neonfma-cortex-a53.S",
3089 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a53.S",
3090 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a55.S",
3091 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a73.S",
3092 "src/f32-igemm/gen/1x8-minmax-aarch64-neonfma-cortex-a57.S",
3093 "src/f32-igemm/gen/1x8-minmax-aarch64-neonfma-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003094 "src/f32-igemm/gen/4x8-minmax-aarch64-neonfma-cortex-a57.S",
3095 "src/f32-igemm/gen/4x8-minmax-aarch64-neonfma-cortex-a75.S",
3096 "src/f32-igemm/gen/5x8-minmax-aarch64-neonfma-cortex-a57.S",
3097 "src/f32-igemm/gen/5x8-minmax-aarch64-neonfma-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003098 "src/f32-igemm/gen/6x8-minmax-aarch64-neonfma-cortex-a57.S",
3099 "src/f32-igemm/gen/6x8-minmax-aarch64-neonfma-cortex-a75.S",
Frank Barchardf1fd89e2020-10-06 23:27:11 -07003100 "src/qs8-gemm/1x16c4-aarch64-neondot-ld64.S",
Frank Barchardf2742c42020-10-12 21:16:46 -07003101 "src/qs8-gemm/4x16c4-aarch64-neondot-cortex-a55.S",
Frank Barcharda5237a52020-10-06 16:18:46 -07003102 "src/qs8-gemm/4x16c4-aarch64-neondot-ld64.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003103]
3104
Marat Dukhan1b354632020-03-23 12:50:22 -07003105INTERNAL_MICROKERNEL_HDRS = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003106 "src/xnnpack/argmaxpool.h",
3107 "src/xnnpack/avgpool.h",
3108 "src/xnnpack/clamp.h",
3109 "src/xnnpack/common.h",
3110 "src/xnnpack/conv.h",
Yury Kartynnike7841862020-11-04 18:22:18 -08003111 "src/xnnpack/depthtospace.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003112 "src/xnnpack/dwconv.h",
Frank Barchard04336c12020-10-22 16:48:55 -07003113 "src/xnnpack/fill.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003114 "src/xnnpack/gavgpool.h",
3115 "src/xnnpack/gemm.h",
3116 "src/xnnpack/hswish.h",
Marat Dukhan660fd192020-03-10 04:55:30 -07003117 "src/xnnpack/ibilinear.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003118 "src/xnnpack/igemm.h",
Marat Dukhancfb31342019-12-05 10:42:57 -08003119 "src/xnnpack/intrinsics-polyfill.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003120 "src/xnnpack/lut.h",
3121 "src/xnnpack/math.h",
3122 "src/xnnpack/maxpool.h",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003123 "src/xnnpack/memory.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003124 "src/xnnpack/packx.h",
3125 "src/xnnpack/pad.h",
3126 "src/xnnpack/params.h",
3127 "src/xnnpack/pavgpool.h",
3128 "src/xnnpack/ppmm.h",
3129 "src/xnnpack/prelu.h",
Marat Dukhan97579532019-10-18 16:40:39 -07003130 "src/xnnpack/raddexpminusmax.h",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07003131 "src/xnnpack/raddextexp.h",
Marat Dukhan97579532019-10-18 16:40:39 -07003132 "src/xnnpack/raddstoreexpminusmax.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003133 "src/xnnpack/rmax.h",
3134 "src/xnnpack/scalar-utils.h",
3135 "src/xnnpack/spmm.h",
3136 "src/xnnpack/unpool.h",
3137 "src/xnnpack/vadd.h",
Marat Dukhan1e782c42019-11-21 17:02:40 -08003138 "src/xnnpack/vbinary.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003139 "src/xnnpack/vmulcaddc.h",
Marat Dukhan05ac8e32019-10-21 15:39:33 -07003140 "src/xnnpack/vscale.h",
Marat Dukhan97579532019-10-18 16:40:39 -07003141 "src/xnnpack/vscaleexpminusmax.h",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07003142 "src/xnnpack/vscaleextexp.h",
Marat Dukhan1e782c42019-11-21 17:02:40 -08003143 "src/xnnpack/vunary.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003144 "src/xnnpack/zip.h",
Marat Dukhan1b354632020-03-23 12:50:22 -07003145]
3146
3147INTERNAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003148 "include/xnnpack.h",
3149 "src/xnnpack/allocator.h",
3150 "src/xnnpack/compute.h",
3151 "src/xnnpack/im2col.h",
3152 "src/xnnpack/indirection.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003153 "src/xnnpack/math-stubs.h",
Chao Mei6ddfc602020-05-13 22:29:36 -07003154 "src/xnnpack/memory-planner.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003155 "src/xnnpack/operator.h",
3156 "src/xnnpack/pack.h",
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07003157 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003158 "src/xnnpack/requantization-stubs.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003159 "src/xnnpack/requantization.h",
Marat Dukhan1d75a542020-02-03 12:23:01 -08003160 "src/xnnpack/subgraph.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003161]
3162
Marat Dukhan1b354632020-03-23 12:50:22 -07003163ACCURACY_EVAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003164 "src/xnnpack/math-stubs.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003165]
3166
Marat Dukhan1b354632020-03-23 12:50:22 -07003167MICROKERNEL_BENCHMARK_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003168 "include/xnnpack.h",
Frank Barchard35db7d02020-10-26 13:37:34 -07003169 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003170]
3171
Marat Dukhan1b354632020-03-23 12:50:22 -07003172MICROKERNEL_TEST_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Frank Barchard35db7d02020-10-26 13:37:34 -07003173 "include/xnnpack.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003174 "src/xnnpack/isa-checks.h",
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07003175 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003176 "src/xnnpack/requantization.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003177]
3178
3179OPERATOR_TEST_PARAMS_HDRS = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003180 "src/xnnpack/common.h",
Frank Barchard04336c12020-10-22 16:48:55 -07003181 "src/xnnpack/params.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003182]
3183
3184WEIGHTS_PACK_HDRS = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003185 "src/xnnpack/compute.h",
Frank Barchard04336c12020-10-22 16:48:55 -07003186 "src/xnnpack/operator.h",
3187 "src/xnnpack/pack.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003188]
3189
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07003190LOGGING_COPTS = select({
3191 # No logging in optimized mode
3192 ":optimized_build": ["-DXNN_LOG_LEVEL=0"],
3193 # Full logging in debug mode
3194 ":debug_build": ["-DXNN_LOG_LEVEL=5"],
3195 # Error-only logging in default (fastbuild) mode
3196 "//conditions:default": ["-DXNN_LOG_LEVEL=2"],
3197})
3198
Marat Dukhan3b59de22020-06-03 20:15:19 -07003199LOGGING_SRCS = select({
3200 # No logging in optimized mode
3201 ":optimized_build": [],
3202 "//conditions:default": [
3203 "src/operator-strings.c",
3204 "src/subgraph-strings.c",
3205 ],
3206})
3207
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07003208LOGGING_HDRS = [
3209 "src/xnnpack/log.h",
3210]
3211
Marat Dukhan08c4a432019-10-03 09:29:21 -07003212xnnpack_cc_library(
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003213 name = "tables",
3214 srcs = TABLE_SRCS,
3215 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003216 gcc_copts = xnnpack_gcc_std_copts(),
3217 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003218)
3219
3220xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003221 name = "scalar_ukernels",
3222 srcs = SCALAR_UKERNELS,
3223 hdrs = INTERNAL_HDRS,
3224 aarch32_copts = ["-marm"],
Marat Dukhan10a38082020-04-17 03:58:35 -07003225 gcc_copts = xnnpack_gcc_std_copts(),
3226 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003227 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003228 ":tables",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003229 "@FP16",
3230 "@FXdiv",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003231 "@pthreadpool",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003232 ],
3233)
3234
3235xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003236 name = "scalar_ukernels_test_mode",
3237 srcs = SCALAR_UKERNELS,
3238 hdrs = INTERNAL_HDRS,
3239 aarch32_copts = ["-marm"],
3240 copts = [
3241 "-UNDEBUG",
3242 "-DXNN_TEST_MODE=1",
3243 ],
3244 gcc_copts = xnnpack_gcc_std_copts(),
3245 msvc_copts = xnnpack_msvc_std_copts(),
3246 deps = [
3247 ":tables",
3248 "@FP16",
3249 "@FXdiv",
3250 "@pthreadpool",
3251 ],
3252)
3253
3254xnnpack_cc_library(
Marat Dukhan436ebe62019-12-04 15:10:12 -08003255 name = "wasm_ukernels",
3256 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003257 gcc_copts = xnnpack_gcc_std_copts(),
3258 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan436ebe62019-12-04 15:10:12 -08003259 wasm_srcs = WASM_UKERNELS,
Marat Dukhan1c6e3892020-06-25 23:56:10 -07003260 wasmsimd_srcs = WASM_UKERNELS + WASMSIMD_UKERNELS,
Marat Dukhan436ebe62019-12-04 15:10:12 -08003261 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003262 ":tables",
Marat Dukhan436ebe62019-12-04 15:10:12 -08003263 "@FP16",
3264 "@FXdiv",
3265 "@pthreadpool",
3266 ],
3267)
3268
3269xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003270 name = "wasm_ukernels_test_mode",
3271 hdrs = INTERNAL_HDRS,
3272 copts = [
3273 "-UNDEBUG",
3274 "-DXNN_TEST_MODE=1",
3275 ],
3276 gcc_copts = xnnpack_gcc_std_copts(),
3277 msvc_copts = xnnpack_msvc_std_copts(),
3278 wasm_srcs = WASM_UKERNELS,
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07003279 wasmsimd_srcs = WASM_UKERNELS + WASMSIMD_UKERNELS,
Marat Dukhan33fcf782020-05-24 14:27:15 -07003280 deps = [
3281 ":tables",
3282 "@FP16",
3283 "@FXdiv",
3284 "@pthreadpool",
3285 ],
3286)
3287
3288xnnpack_cc_library(
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08003289 name = "psimd_accmath_ukernels",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08003290 hdrs = INTERNAL_HDRS,
3291 aarch32_copts = [
3292 "-marm",
3293 "-mfpu=neon",
3294 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07003295 gcc_copts = xnnpack_gcc_std_copts(),
3296 gcc_x86_copts = ["-msse2"],
3297 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08003298 optimized_copts = [
3299 "-O3",
3300 ],
Marat Dukhan500b8892020-04-15 17:09:50 -07003301 psimd_srcs = PSIMD_ACCMATH_UKERNELS,
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08003302 deps = [
3303 ":tables",
3304 "@FP16",
3305 "@psimd",
3306 "@pthreadpool",
3307 ],
3308)
3309
3310xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003311 name = "psimd_accmath_ukernels_test_mode",
3312 hdrs = INTERNAL_HDRS,
3313 aarch32_copts = [
3314 "-marm",
3315 "-mfpu=neon",
3316 ],
3317 copts = [
3318 "-UNDEBUG",
3319 "-DXNN_TEST_MODE=1",
3320 ],
3321 gcc_copts = xnnpack_gcc_std_copts(),
3322 gcc_x86_copts = ["-msse2"],
3323 msvc_copts = xnnpack_msvc_std_copts(),
3324 optimized_copts = [
3325 "-O3",
3326 ],
3327 psimd_srcs = PSIMD_ACCMATH_UKERNELS,
3328 deps = [
3329 ":tables",
3330 "@FP16",
3331 "@psimd",
3332 "@pthreadpool",
3333 ],
3334)
3335
3336xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003337 name = "neon_ukernels",
3338 hdrs = INTERNAL_HDRS,
3339 aarch32_copts = [
3340 "-marm",
Marat Dukhan8853b822020-05-07 12:19:01 -07003341 "-march=armv7-a",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003342 "-mfpu=neon",
3343 ],
3344 aarch32_srcs = NEON_UKERNELS,
3345 aarch64_srcs = NEON_UKERNELS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003346 gcc_copts = xnnpack_gcc_std_copts(),
3347 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08003348 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003349 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003350 "@FP16",
3351 "@pthreadpool",
3352 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003353)
3354
3355xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003356 name = "neon_ukernels_test_mode",
3357 hdrs = INTERNAL_HDRS,
3358 aarch32_copts = [
3359 "-marm",
3360 "-march=armv7-a",
3361 "-mfpu=neon",
3362 ],
3363 aarch32_srcs = NEON_UKERNELS,
3364 aarch64_srcs = NEON_UKERNELS,
3365 copts = [
3366 "-UNDEBUG",
3367 "-DXNN_TEST_MODE=1",
3368 ],
3369 gcc_copts = xnnpack_gcc_std_copts(),
3370 msvc_copts = xnnpack_msvc_std_copts(),
3371 deps = [
3372 ":tables",
3373 "@FP16",
3374 "@pthreadpool",
3375 ],
3376)
3377
3378xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003379 name = "neonfma_ukernels",
3380 hdrs = INTERNAL_HDRS,
3381 aarch32_copts = [
3382 "-marm",
Marat Dukhan8853b822020-05-07 12:19:01 -07003383 "-march=armv7-a",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003384 "-mfpu=neon-vfpv4",
3385 ],
3386 aarch32_srcs = NEONFMA_UKERNELS,
3387 aarch64_srcs = NEONFMA_UKERNELS + AARCH64_NEONFMA_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07003388 apple_aarch32_copts = [
3389 "-mcpu=swift",
3390 "-mtune=generic",
3391 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07003392 gcc_copts = xnnpack_gcc_std_copts(),
3393 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08003394 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003395 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003396 "@FP16",
3397 "@pthreadpool",
3398 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003399)
3400
3401xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003402 name = "neonfma_ukernels_test_mode",
3403 hdrs = INTERNAL_HDRS,
3404 aarch32_copts = [
3405 "-marm",
3406 "-march=armv7-a",
3407 "-mfpu=neon-vfpv4",
3408 ],
3409 aarch32_srcs = NEONFMA_UKERNELS,
3410 aarch64_srcs = NEONFMA_UKERNELS + AARCH64_NEONFMA_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07003411 apple_aarch32_copts = [
3412 "-mcpu=swift",
3413 "-mtune=generic",
3414 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07003415 copts = [
3416 "-UNDEBUG",
3417 "-DXNN_TEST_MODE=1",
3418 ],
3419 gcc_copts = xnnpack_gcc_std_copts(),
3420 msvc_copts = xnnpack_msvc_std_copts(),
3421 deps = [
3422 ":tables",
3423 "@FP16",
3424 "@pthreadpool",
3425 ],
3426)
3427
3428xnnpack_cc_library(
Marat Dukhan8853b822020-05-07 12:19:01 -07003429 name = "neonv8_ukernels",
3430 hdrs = INTERNAL_HDRS,
3431 aarch32_copts = [
3432 "-marm",
3433 "-march=armv8-a",
3434 "-mfpu=neon-fp-armv8",
3435 ],
3436 aarch32_srcs = NEONV8_UKERNELS,
3437 aarch64_srcs = NEONV8_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07003438 apple_aarch32_copts = [
3439 "-mcpu=cyclone",
3440 "-mtune=generic",
3441 ],
Marat Dukhan8853b822020-05-07 12:19:01 -07003442 gcc_copts = xnnpack_gcc_std_copts(),
3443 msvc_copts = xnnpack_msvc_std_copts(),
3444 deps = [
3445 ":tables",
3446 "@FP16",
3447 "@pthreadpool",
3448 ],
3449)
3450
3451xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003452 name = "neonv8_ukernels_test_mode",
3453 hdrs = INTERNAL_HDRS,
3454 aarch32_copts = [
3455 "-marm",
3456 "-march=armv8-a",
3457 "-mfpu=neon-fp-armv8",
3458 ],
3459 aarch32_srcs = NEONV8_UKERNELS,
3460 aarch64_srcs = NEONV8_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07003461 apple_aarch32_copts = [
3462 "-mcpu=cyclone",
3463 "-mtune=generic",
3464 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07003465 copts = [
3466 "-UNDEBUG",
3467 "-DXNN_TEST_MODE=1",
3468 ],
3469 gcc_copts = xnnpack_gcc_std_copts(),
3470 msvc_copts = xnnpack_msvc_std_copts(),
3471 deps = [
3472 ":tables",
3473 "@FP16",
3474 "@pthreadpool",
3475 ],
3476)
3477
3478xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003479 name = "neonfp16arith_ukernels",
3480 hdrs = INTERNAL_HDRS,
3481 aarch64_copts = ["-march=armv8.2-a+fp16"],
3482 aarch64_srcs = AARCH64_NEONFP16ARITH_UKERNELS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003483 gcc_copts = xnnpack_gcc_std_copts(),
3484 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08003485 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003486 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003487 "@FP16",
3488 "@pthreadpool",
3489 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003490)
3491
3492xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003493 name = "neonfp16arith_ukernels_test_mode",
3494 hdrs = INTERNAL_HDRS,
3495 aarch64_copts = ["-march=armv8.2-a+fp16"],
3496 aarch64_srcs = AARCH64_NEONFP16ARITH_UKERNELS,
3497 copts = [
3498 "-UNDEBUG",
3499 "-DXNN_TEST_MODE=1",
3500 ],
3501 gcc_copts = xnnpack_gcc_std_copts(),
3502 msvc_copts = xnnpack_msvc_std_copts(),
3503 deps = [
3504 ":tables",
3505 "@FP16",
3506 "@pthreadpool",
3507 ],
3508)
3509
3510xnnpack_cc_library(
Benoit Jacoba9644732020-08-13 12:48:55 -07003511 name = "neondot_ukernels",
3512 hdrs = INTERNAL_HDRS,
Marat Dukhan799ac752020-08-13 16:08:03 -07003513 aarch32_copts = [
3514 "-marm",
3515 "-march=armv8.2-a+dotprod",
3516 "-mfpu=neon-fp-armv8",
3517 ],
Benoit Jacoba9644732020-08-13 12:48:55 -07003518 aarch32_srcs = NEONDOT_UKERNELS,
3519 aarch64_copts = ["-march=armv8.2-a+dotprod"],
3520 aarch64_srcs = NEONDOT_UKERNELS,
3521 gcc_copts = xnnpack_gcc_std_copts(),
3522 msvc_copts = xnnpack_msvc_std_copts(),
3523 deps = [
3524 ":tables",
3525 "@FP16",
3526 "@pthreadpool",
3527 ],
3528)
3529
3530xnnpack_cc_library(
3531 name = "neondot_ukernels_test_mode",
3532 hdrs = INTERNAL_HDRS,
Marat Dukhan799ac752020-08-13 16:08:03 -07003533 aarch32_copts = [
3534 "-marm",
3535 "-march=armv8.2-a+dotprod",
3536 "-mfpu=neon-fp-armv8",
3537 ],
Benoit Jacoba9644732020-08-13 12:48:55 -07003538 aarch32_srcs = NEONDOT_UKERNELS,
3539 aarch64_copts = ["-march=armv8.2-a+dotprod"],
3540 aarch64_srcs = NEONDOT_UKERNELS,
3541 copts = [
3542 "-UNDEBUG",
3543 "-DXNN_TEST_MODE=1",
3544 ],
3545 gcc_copts = xnnpack_gcc_std_copts(),
3546 msvc_copts = xnnpack_msvc_std_copts(),
3547 deps = [
3548 ":tables",
3549 "@FP16",
3550 "@pthreadpool",
3551 ],
3552)
3553
3554xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003555 name = "sse2_ukernels",
3556 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003557 gcc_copts = xnnpack_gcc_std_copts(),
3558 gcc_x86_copts = ["-msse2"],
3559 msvc_copts = xnnpack_msvc_std_copts(),
3560 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003561 x86_srcs = SSE_UKERNELS + SSE2_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08003562 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003563 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003564 "@FP16",
3565 "@pthreadpool",
3566 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003567)
3568
3569xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003570 name = "sse2_ukernels_test_mode",
3571 hdrs = INTERNAL_HDRS,
3572 copts = [
3573 "-UNDEBUG",
3574 "-DXNN_TEST_MODE=1",
3575 ],
3576 gcc_copts = xnnpack_gcc_std_copts(),
3577 gcc_x86_copts = ["-msse2"],
3578 msvc_copts = xnnpack_msvc_std_copts(),
3579 msvc_x86_32_copts = ["/arch:SSE2"],
3580 x86_srcs = SSE_UKERNELS + SSE2_UKERNELS,
3581 deps = [
3582 ":tables",
3583 "@FP16",
3584 "@pthreadpool",
3585 ],
3586)
3587
3588xnnpack_cc_library(
Marat Dukhanfe7acb62020-03-09 19:30:05 -07003589 name = "ssse3_ukernels",
3590 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003591 gcc_copts = xnnpack_gcc_std_copts(),
3592 gcc_x86_copts = ["-mssse3"],
3593 msvc_copts = xnnpack_msvc_std_copts(),
3594 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhanfe7acb62020-03-09 19:30:05 -07003595 x86_srcs = SSSE3_UKERNELS,
3596 deps = [
3597 ":tables",
3598 "@FP16",
3599 "@pthreadpool",
3600 ],
3601)
3602
3603xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003604 name = "ssse3_ukernels_test_mode",
3605 hdrs = INTERNAL_HDRS,
3606 copts = [
3607 "-UNDEBUG",
3608 "-DXNN_TEST_MODE=1",
3609 ],
3610 gcc_copts = xnnpack_gcc_std_copts(),
3611 gcc_x86_copts = ["-mssse3"],
3612 msvc_copts = xnnpack_msvc_std_copts(),
3613 msvc_x86_32_copts = ["/arch:SSE2"],
3614 x86_srcs = SSSE3_UKERNELS,
3615 deps = [
3616 ":tables",
3617 "@FP16",
3618 "@pthreadpool",
3619 ],
3620)
3621
3622xnnpack_cc_library(
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08003623 name = "sse41_ukernels",
3624 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003625 gcc_copts = xnnpack_gcc_std_copts(),
3626 gcc_x86_copts = ["-msse4.1"],
3627 msvc_copts = xnnpack_msvc_std_copts(),
3628 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08003629 x86_srcs = SSE41_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08003630 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003631 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003632 "@FP16",
3633 "@pthreadpool",
3634 ],
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08003635)
3636
3637xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003638 name = "sse41_ukernels_test_mode",
3639 hdrs = INTERNAL_HDRS,
3640 copts = [
3641 "-UNDEBUG",
3642 "-DXNN_TEST_MODE=1",
3643 ],
3644 gcc_copts = xnnpack_gcc_std_copts(),
3645 gcc_x86_copts = ["-msse4.1"],
3646 msvc_copts = xnnpack_msvc_std_copts(),
3647 msvc_x86_32_copts = ["/arch:SSE2"],
3648 x86_srcs = SSE41_UKERNELS,
3649 deps = [
3650 ":tables",
3651 "@FP16",
3652 "@pthreadpool",
3653 ],
3654)
3655
3656xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003657 name = "avx_ukernels",
3658 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003659 gcc_copts = xnnpack_gcc_std_copts(),
3660 gcc_x86_copts = ["-mavx"],
3661 msvc_copts = xnnpack_msvc_std_copts(),
3662 msvc_x86_32_copts = ["/arch:AVX"],
3663 msvc_x86_64_copts = ["/arch:AVX"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003664 x86_srcs = AVX_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08003665 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003666 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003667 "@FP16",
3668 "@pthreadpool",
3669 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003670)
3671
3672xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003673 name = "avx_ukernels_test_mode",
3674 hdrs = INTERNAL_HDRS,
3675 copts = [
3676 "-UNDEBUG",
3677 "-DXNN_TEST_MODE=1",
3678 ],
3679 gcc_copts = xnnpack_gcc_std_copts(),
3680 gcc_x86_copts = ["-mavx"],
3681 msvc_copts = xnnpack_msvc_std_copts(),
3682 msvc_x86_32_copts = ["/arch:AVX"],
3683 msvc_x86_64_copts = ["/arch:AVX"],
3684 x86_srcs = AVX_UKERNELS,
3685 deps = [
3686 ":tables",
3687 "@FP16",
3688 "@pthreadpool",
3689 ],
3690)
3691
3692xnnpack_cc_library(
Marat Dukhan1566fee2020-08-02 21:55:41 -07003693 name = "xop_ukernels",
3694 hdrs = INTERNAL_HDRS,
3695 gcc_copts = xnnpack_gcc_std_copts(),
3696 gcc_x86_copts = ["-mxop"],
3697 msvc_copts = xnnpack_msvc_std_copts(),
3698 msvc_x86_32_copts = ["/arch:AVX"],
3699 msvc_x86_64_copts = ["/arch:AVX"],
3700 x86_srcs = XOP_UKERNELS,
3701 deps = [
3702 ":tables",
3703 "@FP16",
3704 "@pthreadpool",
3705 ],
3706)
3707
3708xnnpack_cc_library(
3709 name = "xop_ukernels_test_mode",
3710 hdrs = INTERNAL_HDRS,
3711 copts = [
3712 "-UNDEBUG",
3713 "-DXNN_TEST_MODE=1",
3714 ],
3715 gcc_copts = xnnpack_gcc_std_copts(),
3716 gcc_x86_copts = ["-mxop"],
3717 msvc_copts = xnnpack_msvc_std_copts(),
3718 msvc_x86_32_copts = ["/arch:AVX"],
3719 msvc_x86_64_copts = ["/arch:AVX"],
3720 x86_srcs = XOP_UKERNELS,
3721 deps = [
3722 ":tables",
3723 "@FP16",
3724 "@pthreadpool",
3725 ],
3726)
3727
3728xnnpack_cc_library(
Marat Dukhanfda12b82019-11-21 12:27:59 -08003729 name = "fma3_ukernels",
3730 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003731 gcc_copts = xnnpack_gcc_std_copts(),
3732 gcc_x86_copts = ["-mfma"],
3733 msvc_copts = xnnpack_msvc_std_copts(),
3734 msvc_x86_32_copts = ["/arch:AVX"],
3735 msvc_x86_64_copts = ["/arch:AVX"],
Marat Dukhanfda12b82019-11-21 12:27:59 -08003736 x86_srcs = FMA3_UKERNELS,
3737 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003738 ":tables",
Marat Dukhanfda12b82019-11-21 12:27:59 -08003739 "@FP16",
3740 "@pthreadpool",
3741 ],
3742)
3743
3744xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003745 name = "fma3_ukernels_test_mode",
3746 hdrs = INTERNAL_HDRS,
3747 copts = [
3748 "-UNDEBUG",
3749 "-DXNN_TEST_MODE=1",
3750 ],
3751 gcc_copts = xnnpack_gcc_std_copts(),
3752 gcc_x86_copts = ["-mfma"],
3753 msvc_copts = xnnpack_msvc_std_copts(),
3754 msvc_x86_32_copts = ["/arch:AVX"],
3755 msvc_x86_64_copts = ["/arch:AVX"],
3756 x86_srcs = FMA3_UKERNELS,
3757 deps = [
3758 ":tables",
3759 "@FP16",
3760 "@pthreadpool",
3761 ],
3762)
3763
3764xnnpack_cc_library(
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003765 name = "avx2_ukernels",
3766 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003767 gcc_copts = xnnpack_gcc_std_copts(),
3768 gcc_x86_copts = [
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003769 "-mfma",
3770 "-mavx2",
3771 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07003772 msvc_copts = xnnpack_msvc_std_copts(),
3773 msvc_x86_32_copts = ["/arch:AVX2"],
3774 msvc_x86_64_copts = ["/arch:AVX2"],
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003775 x86_srcs = AVX2_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08003776 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003777 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003778 "@FP16",
3779 "@pthreadpool",
3780 ],
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003781)
3782
3783xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003784 name = "avx2_ukernels_test_mode",
3785 hdrs = INTERNAL_HDRS,
3786 copts = [
3787 "-UNDEBUG",
3788 "-DXNN_TEST_MODE=1",
3789 ],
3790 gcc_copts = xnnpack_gcc_std_copts(),
3791 gcc_x86_copts = [
3792 "-mfma",
3793 "-mavx2",
3794 ],
3795 msvc_copts = xnnpack_msvc_std_copts(),
3796 msvc_x86_32_copts = ["/arch:AVX2"],
3797 msvc_x86_64_copts = ["/arch:AVX2"],
3798 x86_srcs = AVX2_UKERNELS,
3799 deps = [
3800 ":tables",
3801 "@FP16",
3802 "@pthreadpool",
3803 ],
3804)
3805
3806xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003807 name = "avx512f_ukernels",
3808 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003809 gcc_copts = xnnpack_gcc_std_copts(),
3810 gcc_x86_copts = ["-mavx512f"],
3811 mingw_copts = ["-fno-asynchronous-unwind-tables"],
3812 msvc_copts = xnnpack_msvc_std_copts(),
3813 msvc_x86_32_copts = ["/arch:AVX512"],
3814 msvc_x86_64_copts = ["/arch:AVX512"],
3815 msys_copts = ["-fno-asynchronous-unwind-tables"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003816 x86_srcs = AVX512F_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08003817 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003818 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003819 "@FP16",
3820 "@pthreadpool",
3821 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003822)
3823
3824xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003825 name = "avx512f_ukernels_test_mode",
3826 hdrs = INTERNAL_HDRS,
3827 copts = [
3828 "-UNDEBUG",
3829 "-DXNN_TEST_MODE=1",
3830 ],
3831 gcc_copts = xnnpack_gcc_std_copts(),
3832 gcc_x86_copts = ["-mavx512f"],
3833 mingw_copts = ["-fno-asynchronous-unwind-tables"],
3834 msvc_copts = xnnpack_msvc_std_copts(),
3835 msvc_x86_32_copts = ["/arch:AVX512"],
3836 msvc_x86_64_copts = ["/arch:AVX512"],
3837 msys_copts = ["-fno-asynchronous-unwind-tables"],
3838 x86_srcs = AVX512F_UKERNELS,
3839 deps = [
3840 ":tables",
3841 "@FP16",
3842 "@pthreadpool",
3843 ],
3844)
3845
3846xnnpack_cc_library(
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07003847 name = "avx512skx_ukernels",
3848 hdrs = INTERNAL_HDRS,
3849 gcc_copts = xnnpack_gcc_std_copts(),
3850 gcc_x86_copts = [
3851 "-mavx512f",
3852 "-mavx512cd",
3853 "-mavx512bw",
3854 "-mavx512dq",
3855 "-mavx512vl",
3856 ],
3857 mingw_copts = ["-fno-asynchronous-unwind-tables"],
3858 msvc_copts = xnnpack_msvc_std_copts(),
3859 msvc_x86_32_copts = ["/arch:AVX512"],
3860 msvc_x86_64_copts = ["/arch:AVX512"],
3861 msys_copts = ["-fno-asynchronous-unwind-tables"],
3862 x86_srcs = AVX512SKX_UKERNELS,
3863 deps = [
3864 ":tables",
3865 "@FP16",
3866 "@pthreadpool",
3867 ],
3868)
3869
3870xnnpack_cc_library(
3871 name = "avx512skx_ukernels_test_mode",
3872 hdrs = INTERNAL_HDRS,
3873 copts = [
3874 "-UNDEBUG",
3875 "-DXNN_TEST_MODE=1",
3876 ],
3877 gcc_copts = xnnpack_gcc_std_copts(),
3878 gcc_x86_copts = [
3879 "-mavx512f",
3880 "-mavx512cd",
3881 "-mavx512bw",
3882 "-mavx512dq",
3883 "-mavx512vl",
3884 ],
3885 mingw_copts = ["-fno-asynchronous-unwind-tables"],
3886 msvc_copts = xnnpack_msvc_std_copts(),
3887 msvc_x86_32_copts = ["/arch:AVX512"],
3888 msvc_x86_64_copts = ["/arch:AVX512"],
3889 msys_copts = ["-fno-asynchronous-unwind-tables"],
3890 x86_srcs = AVX512SKX_UKERNELS,
3891 deps = [
3892 ":tables",
3893 "@FP16",
3894 "@pthreadpool",
3895 ],
3896)
3897
3898xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003899 name = "asm_ukernels",
3900 hdrs = ["src/xnnpack/assembly.h"],
3901 aarch32_srcs = AARCH32_ASM_UKERNELS,
Frank Barchard31bb45b2020-10-06 00:26:33 -07003902 aarch64_copts = ["-march=armv8.2-a+fp16+dotprod"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003903 aarch64_srcs = AARCH64_ASM_UKERNELS,
Frank Barchardbcedc082020-08-17 18:00:51 -07003904 wasm_srcs = WASM32_ASM_UKERNELS,
3905 wasmsimd_srcs = WASM32_ASM_UKERNELS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003906)
3907
Marat Dukhan3b59de22020-06-03 20:15:19 -07003908xnnpack_cc_library(
3909 name = "logging_utils",
3910 srcs = LOGGING_SRCS,
3911 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
3912 copts = LOGGING_COPTS + [
3913 "-Isrc",
3914 "-Iinclude",
3915 ] + select({
3916 ":debug_build": [],
3917 "//conditions:default": xnnpack_min_size_copts(),
3918 }),
3919 gcc_copts = xnnpack_gcc_std_copts(),
3920 msvc_copts = xnnpack_msvc_std_copts(),
3921 visibility = xnnpack_visibility(),
3922 deps = [
3923 "@FP16",
3924 "@clog",
3925 "@pthreadpool",
3926 ],
3927)
3928
Marat Dukhan08c4a432019-10-03 09:29:21 -07003929xnnpack_aggregate_library(
3930 name = "ukernels",
3931 aarch32_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003932 ":neon_ukernels",
3933 ":neonfma_ukernels",
Marat Dukhan8853b822020-05-07 12:19:01 -07003934 ":neonv8_ukernels",
Benoit Jacoba9644732020-08-13 12:48:55 -07003935 ":neondot_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003936 ":asm_ukernels",
3937 ],
3938 aarch64_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003939 ":neon_ukernels",
3940 ":neonfma_ukernels",
Marat Dukhan8853b822020-05-07 12:19:01 -07003941 ":neonv8_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003942 ":neonfp16arith_ukernels",
Benoit Jacoba9644732020-08-13 12:48:55 -07003943 ":neondot_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003944 ":asm_ukernels",
3945 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07003946 generic_deps = [
3947 ":scalar_ukernels",
3948 ],
Marat Dukhan500b8892020-04-15 17:09:50 -07003949 psimd_deps = [
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08003950 ":psimd_accmath_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003951 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07003952 wasm_deps = [
3953 ":wasm_ukernels",
Frank Barchardbcedc082020-08-17 18:00:51 -07003954 ":asm_ukernels",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003955 ],
3956 wasmsimd_deps = [
3957 ":wasm_ukernels",
Frank Barchardbcedc082020-08-17 18:00:51 -07003958 ":asm_ukernels",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003959 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003960 x86_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003961 ":sse2_ukernels",
Marat Dukhanfe7acb62020-03-09 19:30:05 -07003962 ":ssse3_ukernels",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08003963 ":sse41_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003964 ":avx_ukernels",
Marat Dukhan1566fee2020-08-02 21:55:41 -07003965 ":xop_ukernels",
Marat Dukhanfda12b82019-11-21 12:27:59 -08003966 ":fma3_ukernels",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003967 ":avx2_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003968 ":avx512f_ukernels",
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07003969 ":avx512skx_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003970 ],
3971)
3972
Marat Dukhan33fcf782020-05-24 14:27:15 -07003973xnnpack_aggregate_library(
3974 name = "ukernels_test_mode",
3975 aarch32_deps = [
3976 ":neon_ukernels_test_mode",
3977 ":neonfma_ukernels_test_mode",
3978 ":neonv8_ukernels_test_mode",
Benoit Jacoba9644732020-08-13 12:48:55 -07003979 ":neondot_ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003980 ":asm_ukernels",
3981 ],
3982 aarch64_deps = [
3983 ":neon_ukernels_test_mode",
3984 ":neonfma_ukernels_test_mode",
3985 ":neonv8_ukernels_test_mode",
3986 ":neonfp16arith_ukernels_test_mode",
Benoit Jacoba9644732020-08-13 12:48:55 -07003987 ":neondot_ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003988 ":asm_ukernels",
3989 ],
3990 generic_deps = [
3991 ":scalar_ukernels_test_mode",
3992 ],
3993 psimd_deps = [
Marat Dukhan33fcf782020-05-24 14:27:15 -07003994 ":psimd_accmath_ukernels_test_mode",
3995 ],
3996 wasm_deps = [
3997 ":wasm_ukernels_test_mode",
Frank Barchardbcedc082020-08-17 18:00:51 -07003998 ":asm_ukernels",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003999 ],
4000 wasmsimd_deps = [
4001 ":wasm_ukernels_test_mode",
Frank Barchardbcedc082020-08-17 18:00:51 -07004002 ":asm_ukernels",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004003 ],
4004 x86_deps = [
4005 ":sse2_ukernels_test_mode",
4006 ":ssse3_ukernels_test_mode",
4007 ":sse41_ukernels_test_mode",
4008 ":avx_ukernels_test_mode",
Marat Dukhan1566fee2020-08-02 21:55:41 -07004009 ":xop_ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004010 ":fma3_ukernels_test_mode",
4011 ":avx2_ukernels_test_mode",
4012 ":avx512f_ukernels_test_mode",
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07004013 ":avx512skx_ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004014 ],
4015)
4016
Marat Dukhan86df3842020-06-17 14:03:31 -07004017# This target must be compatible with Android NDK r17 and compile with Clang 6.
4018# It is similar to :ukernels target, but excludes microkernels which require
4019# NEON FP16 arithmetics, NEON Dot Product, and newer intrinsics.
Marat Dukhanae046f52020-06-15 13:16:14 -07004020xnnpack_aggregate_library(
Marat Dukhan86df3842020-06-17 14:03:31 -07004021 name = "ukernels_max_compatibility",
Marat Dukhanae046f52020-06-15 13:16:14 -07004022 aarch32_deps = [
4023 ":neon_ukernels",
4024 ":neonfma_ukernels",
4025 ":neonv8_ukernels",
4026 ":asm_ukernels",
4027 ],
4028 aarch64_deps = [
4029 ":neon_ukernels",
4030 ":neonfma_ukernels",
4031 ":neonv8_ukernels",
4032 ":asm_ukernels",
4033 ],
4034 generic_deps = [
4035 ":scalar_ukernels",
4036 ],
4037 psimd_deps = [
Marat Dukhanae046f52020-06-15 13:16:14 -07004038 ":psimd_accmath_ukernels",
4039 ],
4040 wasm_deps = [
4041 ":wasm_ukernels",
4042 ],
4043 wasmsimd_deps = [
4044 ":wasm_ukernels",
4045 ],
4046 x86_deps = [
4047 ":sse2_ukernels",
4048 ":ssse3_ukernels",
4049 ":sse41_ukernels",
4050 ":avx_ukernels",
Marat Dukhan1566fee2020-08-02 21:55:41 -07004051 ":xop_ukernels",
Marat Dukhanae046f52020-06-15 13:16:14 -07004052 ":fma3_ukernels",
4053 ":avx2_ukernels",
4054 ":avx512f_ukernels",
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07004055 ":avx512skx_ukernels",
Marat Dukhanae046f52020-06-15 13:16:14 -07004056 ],
4057)
4058
Marat Dukhan08c4a432019-10-03 09:29:21 -07004059xnnpack_cc_library(
4060 name = "im2col",
4061 srcs = ["src/im2col.c"],
4062 hdrs = [
4063 "src/xnnpack/common.h",
4064 "src/xnnpack/im2col.h",
4065 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07004066 gcc_copts = xnnpack_gcc_std_copts(),
4067 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004068)
4069
4070xnnpack_cc_library(
4071 name = "indirection",
4072 srcs = ["src/indirection.c"],
4073 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004074 gcc_copts = xnnpack_gcc_std_copts(),
4075 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004076 deps = [
4077 "@FP16",
4078 "@FXdiv",
4079 "@pthreadpool",
4080 ],
4081)
4082
4083xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07004084 name = "indirection_test_mode",
4085 srcs = ["src/indirection.c"],
4086 hdrs = INTERNAL_HDRS,
4087 copts = [
4088 "-UNDEBUG",
4089 "-DXNN_TEST_MODE=1",
4090 ],
4091 gcc_copts = xnnpack_gcc_std_copts(),
4092 msvc_copts = xnnpack_msvc_std_copts(),
4093 deps = [
4094 "@FP16",
4095 "@FXdiv",
4096 "@pthreadpool",
4097 ],
4098)
4099
4100xnnpack_cc_library(
Marat Dukhanab582382020-07-06 13:32:08 -07004101 name = "packing",
4102 srcs = ["src/packing.c"],
4103 hdrs = INTERNAL_HDRS,
4104 gcc_copts = xnnpack_gcc_std_copts(),
4105 msvc_copts = xnnpack_msvc_std_copts(),
4106 deps = [
4107 "@FP16",
4108 "@FXdiv",
4109 "@pthreadpool",
4110 ],
4111)
4112
4113xnnpack_cc_library(
4114 name = "packing_test_mode",
4115 srcs = ["src/packing.c"],
4116 hdrs = INTERNAL_HDRS,
4117 copts = [
4118 "-UNDEBUG",
4119 "-DXNN_TEST_MODE=1",
4120 ],
4121 gcc_copts = xnnpack_gcc_std_copts(),
4122 msvc_copts = xnnpack_msvc_std_copts(),
4123 deps = [
4124 "@FP16",
4125 "@FXdiv",
4126 "@pthreadpool",
4127 ],
4128)
4129
4130xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004131 name = "operator_run",
4132 srcs = ["src/operator-run.c"],
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07004133 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004134 copts = LOGGING_COPTS + select({
Marat Dukhan05702cf2020-03-26 15:41:33 -07004135 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4136 "//conditions:default": [],
4137 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07004138 gcc_copts = xnnpack_gcc_std_copts(),
4139 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004140 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07004141 ":logging_utils",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004142 "@FP16",
4143 "@FXdiv",
4144 "@clog",
4145 "@pthreadpool",
4146 ],
4147)
4148
Chao Mei6ddfc602020-05-13 22:29:36 -07004149xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07004150 name = "operator_run_test_mode",
4151 srcs = ["src/operator-run.c"],
4152 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
4153 copts = LOGGING_COPTS + [
4154 "-UNDEBUG",
4155 "-DXNN_TEST_MODE=1",
4156 ] + select({
4157 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4158 "//conditions:default": [],
4159 }),
4160 gcc_copts = xnnpack_gcc_std_copts(),
4161 msvc_copts = xnnpack_msvc_std_copts(),
4162 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07004163 ":logging_utils",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004164 "@FP16",
4165 "@FXdiv",
4166 "@clog",
4167 "@pthreadpool",
4168 ],
4169)
4170
4171xnnpack_cc_library(
Chao Mei6ddfc602020-05-13 22:29:36 -07004172 name = "memory_planner",
4173 srcs = ["src/memory-planner.c"],
4174 hdrs = INTERNAL_HDRS,
4175 defines = select({
4176 ":xnn_enable_memopt_explicit_true": ["XNN_ENABLE_MEMOPT=1"],
4177 ":xnn_enable_memopt_explicit_false": ["XNN_ENABLE_MEMOPT=0"],
4178 "//conditions:default": ["XNN_ENABLE_MEMOPT=1"],
4179 }),
4180 gcc_copts = xnnpack_gcc_std_copts(),
4181 msvc_copts = xnnpack_msvc_std_copts(),
4182 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07004183 ":logging_utils",
Chao Mei6ddfc602020-05-13 22:29:36 -07004184 "@pthreadpool",
4185 ],
4186)
4187
Marat Dukhan33fcf782020-05-24 14:27:15 -07004188xnnpack_cc_library(
4189 name = "memory_planner_test_mode",
4190 srcs = ["src/memory-planner.c"],
4191 hdrs = INTERNAL_HDRS,
4192 copts = [
4193 "-UNDEBUG",
4194 "-DXNN_TEST_MODE=1",
4195 ],
4196 defines = select({
4197 ":xnn_enable_memopt_explicit_true": ["XNN_ENABLE_MEMOPT=1"],
4198 ":xnn_enable_memopt_explicit_false": ["XNN_ENABLE_MEMOPT=0"],
4199 "//conditions:default": ["XNN_ENABLE_MEMOPT=1"],
4200 }),
4201 gcc_copts = xnnpack_gcc_std_copts(),
4202 msvc_copts = xnnpack_msvc_std_copts(),
4203 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07004204 ":logging_utils",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004205 "@pthreadpool",
4206 ],
4207)
4208
Marat Dukhan08c4a432019-10-03 09:29:21 -07004209cc_library(
4210 name = "enable_assembly",
4211 defines = select({
4212 ":xnn_enable_assembly_explicit_true": ["XNN_ENABLE_ASSEMBLY=1"],
4213 ":xnn_enable_assembly_explicit_false": ["XNN_ENABLE_ASSEMBLY=0"],
Frank Barchard810171d2019-10-10 10:34:51 -07004214 "//conditions:default": ["XNN_ENABLE_ASSEMBLY=1"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004215 }),
4216)
4217
Marat Dukhan9de90e02020-06-18 16:04:12 -07004218cc_library(
4219 name = "enable_sparse",
4220 defines = select({
4221 ":xnn_enable_sparse_explicit_true": ["XNN_ENABLE_SPARSE=1"],
4222 ":xnn_enable_sparse_explicit_false": ["XNN_ENABLE_SPARSE=0"],
Marat Dukhan8b283aa2020-06-18 16:48:57 -07004223 "//conditions:default": ["XNN_ENABLE_SPARSE=0"],
Marat Dukhan9de90e02020-06-18 16:04:12 -07004224 }),
4225)
4226
Marat Dukhancf056b22019-10-07 10:26:29 -07004227xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004228 name = "operators",
4229 srcs = OPERATOR_SRCS + [
Marat Dukhan04f03be2019-11-19 12:36:47 -08004230 "src/memory.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004231 "src/operator-delete.c",
Marat Dukhancf056b22019-10-07 10:26:29 -07004232 ],
4233 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004234 copts = LOGGING_COPTS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07004235 "-Isrc",
4236 "-Iinclude",
4237 ] + select({
4238 ":debug_build": [],
4239 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07004240 }) + select({
4241 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4242 "//conditions:default": [],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004243 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07004244 gcc_copts = xnnpack_gcc_std_copts(),
4245 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004246 deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07004247 ":indirection",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004248 ":logging_utils",
Marat Dukhanab582382020-07-06 13:32:08 -07004249 ":packing",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004250 "@FP16",
4251 "@FXdiv",
4252 "@clog",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004253 "@pthreadpool",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004254 ],
4255)
4256
Marat Dukhan10a38082020-04-17 03:58:35 -07004257xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07004258 name = "operators_test_mode",
4259 srcs = OPERATOR_SRCS + [
4260 "src/memory.c",
4261 "src/operator-delete.c",
4262 ],
4263 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
4264 copts = LOGGING_COPTS + [
4265 "-Isrc",
4266 "-Iinclude",
4267 "-UNDEBUG",
4268 "-DXNN_TEST_MODE=1",
4269 ] + select({
4270 ":debug_build": [],
4271 "//conditions:default": xnnpack_min_size_copts(),
4272 }) + select({
4273 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4274 "//conditions:default": [],
4275 }),
4276 gcc_copts = xnnpack_gcc_std_copts(),
4277 msvc_copts = xnnpack_msvc_std_copts(),
4278 deps = [
4279 ":indirection_test_mode",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004280 ":logging_utils",
Marat Dukhanab582382020-07-06 13:32:08 -07004281 ":packing_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004282 "@FP16",
4283 "@FXdiv",
4284 "@clog",
4285 "@pthreadpool",
4286 ],
4287)
4288
4289xnnpack_cc_library(
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004290 name = "XNNPACK",
4291 srcs = [
4292 "src/init.c",
Marat Dukhanccfdbd12020-02-03 14:27:45 -08004293 "src/runtime.c",
4294 "src/subgraph.c",
4295 "src/tensor.c",
Marat Dukhanf03da0d2020-06-10 16:00:20 -07004296 ] + SUBGRAPH_SRCS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004297 hdrs = ["include/xnnpack.h"],
4298 copts = LOGGING_COPTS + [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004299 "-Isrc",
4300 "-Iinclude",
4301 ] + select({
4302 ":debug_build": [],
4303 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07004304 }) + select({
4305 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4306 "//conditions:default": [],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004307 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07004308 gcc_copts = xnnpack_gcc_std_copts(),
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004309 includes = ["include"],
Marat Dukhan10a38082020-04-17 03:58:35 -07004310 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004311 visibility = xnnpack_visibility(),
4312 deps = [
4313 ":enable_assembly",
Marat Dukhan9de90e02020-06-18 16:04:12 -07004314 ":enable_sparse",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004315 ":logging_utils",
Chao Mei6ddfc602020-05-13 22:29:36 -07004316 ":memory_planner",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004317 ":operator_run",
4318 ":operators",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004319 ":ukernels",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004320 "@clog",
Marat Dukhanab2946c2020-05-21 20:04:13 -07004321 "@FP16",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004322 "@pthreadpool",
Marat Dukhand343c222019-10-07 09:22:14 -07004323 ] + select({
4324 ":emscripten": [],
4325 "//conditions:default": ["@cpuinfo"],
4326 }),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004327)
4328
Marat Dukhan10a38082020-04-17 03:58:35 -07004329xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07004330 name = "XNNPACK_test_mode",
4331 srcs = [
4332 "src/init.c",
4333 "src/runtime.c",
4334 "src/subgraph.c",
4335 "src/tensor.c",
Marat Dukhanf03da0d2020-06-10 16:00:20 -07004336 ] + SUBGRAPH_SRCS,
Marat Dukhan33fcf782020-05-24 14:27:15 -07004337 hdrs = ["include/xnnpack.h"],
4338 copts = LOGGING_COPTS + [
4339 "-Isrc",
4340 "-Iinclude",
4341 "-UNDEBUG",
4342 "-DXNN_TEST_MODE=1",
4343 ] + select({
4344 ":debug_build": [],
4345 "//conditions:default": xnnpack_min_size_copts(),
4346 }) + select({
4347 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4348 "//conditions:default": [],
4349 }),
4350 gcc_copts = xnnpack_gcc_std_copts(),
4351 includes = ["include"],
4352 msvc_copts = xnnpack_msvc_std_copts(),
4353 visibility = xnnpack_visibility(),
4354 deps = [
4355 ":enable_assembly",
Marat Dukhan9de90e02020-06-18 16:04:12 -07004356 ":enable_sparse",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004357 ":logging_utils",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004358 ":memory_planner_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004359 ":operator_run_test_mode",
4360 ":operators_test_mode",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004361 ":ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004362 "@clog",
4363 "@FP16",
4364 "@pthreadpool",
4365 ] + select({
4366 ":emscripten": [],
4367 "//conditions:default": ["@cpuinfo"],
4368 }),
4369)
4370
4371xnnpack_cc_library(
Marat Dukhanae046f52020-06-15 13:16:14 -07004372 name = "xnnpack_f32",
4373 srcs = [
4374 "src/init.c",
4375 "src/runtime.c",
4376 "src/subgraph.c",
4377 "src/tensor.c",
4378 ] + SUBGRAPH_SRCS,
4379 hdrs = ["include/xnnpack.h"],
4380 copts = LOGGING_COPTS + [
4381 "-Isrc",
4382 "-Iinclude",
4383 ] + select({
4384 ":debug_build": [],
4385 "//conditions:default": xnnpack_min_size_copts(),
4386 }) + select({
4387 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4388 "//conditions:default": [],
4389 }),
4390 defines = [
Marat Dukhan16f1e1a2020-08-04 16:38:22 -07004391 "XNN_NO_QS8_OPERATORS",
Marat Dukhan08b7a972020-07-14 18:17:29 -07004392 "XNN_NO_QU8_OPERATORS",
Marat Dukhanae046f52020-06-15 13:16:14 -07004393 "XNN_NO_U8_OPERATORS",
4394 "XNN_NO_X8_OPERATORS",
4395 "XNN_NO_F16_OPERATORS",
4396 "XNN_NO_X16_OPERATORS",
4397 ],
4398 gcc_copts = xnnpack_gcc_std_copts(),
4399 includes = ["include"],
4400 msvc_copts = xnnpack_msvc_std_copts(),
4401 visibility = xnnpack_visibility(),
4402 deps = [
4403 ":enable_assembly",
Marat Dukhan9de90e02020-06-18 16:04:12 -07004404 ":enable_sparse",
Marat Dukhanae046f52020-06-15 13:16:14 -07004405 ":logging_utils",
4406 ":memory_planner",
4407 ":operator_run",
4408 ":operators",
Marat Dukhan86df3842020-06-17 14:03:31 -07004409 ":ukernels_max_compatibility",
Marat Dukhanae046f52020-06-15 13:16:14 -07004410 "@clog",
4411 "@FP16",
4412 "@pthreadpool",
4413 ] + select({
4414 ":emscripten": [],
4415 "//conditions:default": ["@cpuinfo"],
4416 }),
4417)
4418
4419xnnpack_cc_library(
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004420 name = "xnnpack_operators_nhwc_f32",
4421 srcs = [
4422 "src/init.c",
4423 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07004424 hdrs = ["include/xnnpack.h"],
4425 copts = LOGGING_COPTS + [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004426 "-Isrc",
4427 "-Iinclude",
4428 ] + select({
4429 ":debug_build": [],
4430 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07004431 }) + select({
4432 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4433 "//conditions:default": [],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004434 }),
4435 defines = [
Marat Dukhan16f1e1a2020-08-04 16:38:22 -07004436 "XNN_NO_QS8_OPERATORS",
Marat Dukhan08b7a972020-07-14 18:17:29 -07004437 "XNN_NO_QU8_OPERATORS",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004438 "XNN_NO_U8_OPERATORS",
4439 "XNN_NO_X8_OPERATORS",
Marat Dukhanefc47b82019-11-18 09:25:38 -08004440 "XNN_NO_NCHW_OPERATORS",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004441 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07004442 gcc_copts = xnnpack_gcc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004443 includes = ["include"],
Marat Dukhan10a38082020-04-17 03:58:35 -07004444 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004445 visibility = xnnpack_visibility(),
4446 deps = [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004447 ":enable_assembly",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004448 ":logging_utils",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004449 ":operator_run",
4450 ":operators",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004451 ":ukernels",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004452 "@clog",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004453 "@pthreadpool",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004454 ] + select({
4455 ":emscripten": [],
4456 "//conditions:default": ["@cpuinfo"],
4457 }),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004458)
4459
Marat Dukhancf056b22019-10-07 10:26:29 -07004460xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004461 name = "bench_utils",
4462 srcs = ["bench/utils.cc"],
4463 hdrs = ["bench/utils.h"],
Marat Dukhanbad48fe2019-11-04 10:35:22 -08004464 deps = [
4465 "@com_google_benchmark//:benchmark",
4466 "@cpuinfo",
4467 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004468)
4469
Frank Barchard7e955972019-10-11 10:34:25 -07004470######################### Benchmarks for micro-kernels #########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07004471
4472xnnpack_benchmark(
Marat Dukhan595e1702020-07-31 10:12:52 -07004473 name = "qs8_gemm_bench",
4474 srcs = [
4475 "bench/gemm.h",
4476 "bench/qs8-gemm.cc",
4477 "src/xnnpack/AlignedAllocator.h",
4478 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Frank Barchard31328cb2020-10-12 11:55:18 -07004479 copts = xnnpack_optional_ruy_copts() + xnnpack_optional_gemmlowp_copts(),
4480 deps = MICROKERNEL_BENCHMARK_DEPS + [":packing"] + xnnpack_optional_ruy_deps() + xnnpack_optional_gemmlowp_deps(),
Marat Dukhan595e1702020-07-31 10:12:52 -07004481)
4482
4483xnnpack_benchmark(
Marat Dukhan56bdd4a2020-08-03 19:47:04 -07004484 name = "qs8_requantization_bench",
4485 srcs = [
4486 "bench/qs8-requantization.cc",
Marat Dukhan56bdd4a2020-08-03 19:47:04 -07004487 "src/xnnpack/AlignedAllocator.h",
Frank Barchard04336c12020-10-22 16:48:55 -07004488 "src/xnnpack/requantization-stubs.h",
Marat Dukhan56bdd4a2020-08-03 19:47:04 -07004489 ] + MICROKERNEL_BENCHMARK_HDRS,
4490 deps = MICROKERNEL_BENCHMARK_DEPS,
4491)
4492
4493xnnpack_benchmark(
Marat Dukhan08b7a972020-07-14 18:17:29 -07004494 name = "qu8_gemm_bench",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004495 srcs = [
4496 "bench/gemm.h",
Marat Dukhan08b7a972020-07-14 18:17:29 -07004497 "bench/qu8-gemm.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004498 "src/xnnpack/AlignedAllocator.h",
4499 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004500 copts = xnnpack_optional_ruy_copts() + xnnpack_optional_gemmlowp_copts(),
Marat Dukhanab582382020-07-06 13:32:08 -07004501 deps = MICROKERNEL_BENCHMARK_DEPS + [":packing"] + xnnpack_optional_ruy_deps() + xnnpack_optional_gemmlowp_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004502)
4503
4504xnnpack_benchmark(
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07004505 name = "qu8_requantization_bench",
4506 srcs = [
4507 "bench/qu8-requantization.cc",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07004508 "src/xnnpack/AlignedAllocator.h",
Frank Barchard04336c12020-10-22 16:48:55 -07004509 "src/xnnpack/requantization-stubs.h",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07004510 ] + MICROKERNEL_BENCHMARK_HDRS,
4511 deps = MICROKERNEL_BENCHMARK_DEPS,
4512)
4513
4514xnnpack_benchmark(
Frank Barchard40d20fe2020-05-05 00:37:45 -07004515 name = "f16_igemm_bench",
4516 srcs = [
4517 "bench/f16-igemm.cc",
4518 "bench/conv.h",
4519 "bench/google/conv.h",
4520 "src/xnnpack/AlignedAllocator.h",
4521 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004522 deps = MICROKERNEL_BENCHMARK_DEPS + [
4523 ":indirection",
4524 ":packing",
4525 ],
Frank Barchard40d20fe2020-05-05 00:37:45 -07004526)
4527
4528xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004529 name = "f16_gemm_bench",
4530 srcs = [
4531 "bench/f16-gemm.cc",
4532 "bench/gemm.h",
4533 "src/xnnpack/AlignedAllocator.h",
4534 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004535 deps = MICROKERNEL_BENCHMARK_DEPS + [
4536 ":packing",
4537 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004538)
4539
4540xnnpack_benchmark(
Marat Dukhanbdb56f52020-02-05 21:42:49 -08004541 name = "f16_spmm_bench",
4542 srcs = [
4543 "bench/f16-spmm.cc",
4544 "bench/gemm.h",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08004545 "src/xnnpack/AlignedAllocator.h",
4546 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanbdb56f52020-02-05 21:42:49 -08004547 deps = MICROKERNEL_BENCHMARK_DEPS,
4548)
4549
4550xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004551 name = "f32_igemm_bench",
4552 srcs = [
4553 "bench/f32-igemm.cc",
4554 "bench/conv.h",
4555 "src/xnnpack/AlignedAllocator.h",
4556 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004557 deps = MICROKERNEL_BENCHMARK_DEPS + [
4558 ":indirection",
4559 ":packing",
4560 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004561)
4562
4563xnnpack_benchmark(
Frank Barchardc2010062020-07-19 10:04:12 -07004564 name = "f16_relu_bench",
4565 srcs = [
4566 "bench/f16-relu.cc",
4567 "src/xnnpack/AlignedAllocator.h",
4568 ] + MICROKERNEL_BENCHMARK_HDRS,
4569 deps = MICROKERNEL_BENCHMARK_DEPS,
4570)
4571
4572xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004573 name = "f32_conv_hwc_bench",
4574 srcs = [
4575 "bench/f32-conv-hwc.cc",
4576 "bench/dconv.h",
4577 "src/xnnpack/AlignedAllocator.h",
4578 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004579 deps = MICROKERNEL_BENCHMARK_DEPS + [
4580 ":packing",
4581 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004582)
4583
4584xnnpack_benchmark(
Marat Dukhan1f29b802020-05-15 23:46:39 -07004585 name = "f32_conv_hwc2chw_bench",
Erich Elsen563df5f2019-10-23 08:02:21 -07004586 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07004587 "bench/f32-conv-hwc2chw.cc",
Erich Elsen563df5f2019-10-23 08:02:21 -07004588 "bench/dconv.h",
4589 "src/xnnpack/AlignedAllocator.h",
4590 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004591 deps = MICROKERNEL_BENCHMARK_DEPS + [
4592 ":packing",
4593 ],
Erich Elsen563df5f2019-10-23 08:02:21 -07004594)
4595
4596xnnpack_benchmark(
Frank Barchard5a599a62020-06-04 20:12:44 -07004597 name = "f16_dwconv_bench",
4598 srcs = [
4599 "bench/f16-dwconv.cc",
4600 "bench/dwconv.h",
4601 "bench/google/dwconv.h",
4602 "src/xnnpack/AlignedAllocator.h",
4603 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004604 deps = MICROKERNEL_BENCHMARK_DEPS + [
4605 ":indirection",
4606 ":packing",
4607 ],
Frank Barchard5a599a62020-06-04 20:12:44 -07004608)
4609
4610xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004611 name = "f32_dwconv_bench",
4612 srcs = [
4613 "bench/f32-dwconv.cc",
4614 "bench/dwconv.h",
4615 "src/xnnpack/AlignedAllocator.h",
4616 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004617 deps = MICROKERNEL_BENCHMARK_DEPS + [
4618 ":indirection",
4619 ":packing",
4620 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004621)
4622
4623xnnpack_benchmark(
Marat Dukhanbf715f92020-10-23 20:17:00 -07004624 name = "f32_dwconv2d_chw_bench",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004625 srcs = [
Marat Dukhanbf715f92020-10-23 20:17:00 -07004626 "bench/f32-dwconv2d-chw.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004627 "bench/dwconv.h",
4628 "src/xnnpack/AlignedAllocator.h",
4629 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004630 deps = MICROKERNEL_BENCHMARK_DEPS + [
4631 ":indirection",
4632 ":packing",
4633 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004634)
4635
4636xnnpack_benchmark(
4637 name = "f32_gemm_bench",
4638 srcs = [
4639 "bench/f32-gemm.cc",
4640 "bench/gemm.h",
4641 "src/xnnpack/AlignedAllocator.h",
4642 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004643 copts = xnnpack_optional_ruy_copts(),
Marat Dukhanab582382020-07-06 13:32:08 -07004644 deps = MICROKERNEL_BENCHMARK_DEPS + [":packing"] + xnnpack_optional_ruy_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004645)
4646
4647xnnpack_benchmark(
Marat Dukhana11ca342020-06-25 23:45:07 -07004648 name = "f32_hswish_bench",
4649 srcs = [
4650 "bench/f32-hswish.cc",
4651 "src/xnnpack/AlignedAllocator.h",
4652 ] + MICROKERNEL_BENCHMARK_HDRS,
4653 deps = MICROKERNEL_BENCHMARK_DEPS,
4654)
4655
4656xnnpack_benchmark(
Marat Dukhan4c4eb002019-12-08 21:27:49 -08004657 name = "f32_raddexpminusmax_bench",
4658 srcs = [
4659 "bench/f32-raddexpminusmax.cc",
4660 "src/xnnpack/AlignedAllocator.h",
4661 ] + MICROKERNEL_BENCHMARK_HDRS,
4662 deps = MICROKERNEL_BENCHMARK_DEPS,
4663)
4664
4665xnnpack_benchmark(
4666 name = "f32_raddextexp_bench",
4667 srcs = [
4668 "bench/f32-raddextexp.cc",
4669 "src/xnnpack/AlignedAllocator.h",
4670 ] + MICROKERNEL_BENCHMARK_HDRS,
4671 deps = MICROKERNEL_BENCHMARK_DEPS,
4672)
4673
4674xnnpack_benchmark(
4675 name = "f32_raddstoreexpminusmax_bench",
4676 srcs = [
4677 "bench/f32-raddstoreexpminusmax.cc",
4678 "src/xnnpack/AlignedAllocator.h",
4679 ] + MICROKERNEL_BENCHMARK_HDRS,
4680 deps = MICROKERNEL_BENCHMARK_DEPS,
4681)
4682
4683xnnpack_benchmark(
Frank Barchard3c4a9522020-07-18 16:11:22 -07004684 name = "f32_relu_bench",
4685 srcs = [
4686 "bench/f32-relu.cc",
4687 "src/xnnpack/AlignedAllocator.h",
4688 ] + MICROKERNEL_BENCHMARK_HDRS,
4689 deps = MICROKERNEL_BENCHMARK_DEPS,
4690)
4691
4692xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004693 name = "f32_rmax_bench",
4694 srcs = [
4695 "bench/f32-rmax.cc",
4696 "src/xnnpack/AlignedAllocator.h",
4697 ] + MICROKERNEL_BENCHMARK_HDRS,
4698 deps = MICROKERNEL_BENCHMARK_DEPS,
4699)
4700
4701xnnpack_benchmark(
Marat Dukhan14bec502019-11-18 11:35:31 -08004702 name = "f32_sigmoid_bench",
4703 srcs = [
4704 "bench/f32-sigmoid.cc",
4705 "src/xnnpack/AlignedAllocator.h",
4706 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan14bec502019-11-18 11:35:31 -08004707 deps = MICROKERNEL_BENCHMARK_DEPS,
4708)
4709
4710xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004711 name = "f32_spmm_bench",
4712 srcs = [
4713 "bench/f32-spmm.cc",
4714 "bench/gemm.h",
4715 "src/xnnpack/AlignedAllocator.h",
4716 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07004717 deps = MICROKERNEL_BENCHMARK_DEPS,
4718)
4719
4720xnnpack_benchmark(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08004721 name = "f32_softmax_bench",
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07004722 srcs = [
Marat Dukhanfd8e6892020-01-27 15:25:25 -08004723 "bench/f32-softmax.cc",
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07004724 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004725 copts = xnnpack_optional_dnnl_copts(),
Marat Dukhan8d3c6932020-03-06 20:27:27 -08004726 deps = MICROKERNEL_BENCHMARK_DEPS + xnnpack_optional_dnnl_deps(),
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07004727)
4728
4729xnnpack_benchmark(
Marat Dukhan4c4eb002019-12-08 21:27:49 -08004730 name = "f32_vscaleexpminusmax_bench",
4731 srcs = [
4732 "bench/f32-vscaleexpminusmax.cc",
4733 "src/xnnpack/AlignedAllocator.h",
4734 ] + MICROKERNEL_BENCHMARK_HDRS,
4735 deps = MICROKERNEL_BENCHMARK_DEPS,
4736)
4737
4738xnnpack_benchmark(
4739 name = "f32_vscaleextexp_bench",
4740 srcs = [
4741 "bench/f32-vscaleextexp.cc",
4742 "src/xnnpack/AlignedAllocator.h",
4743 ] + MICROKERNEL_BENCHMARK_HDRS,
4744 deps = MICROKERNEL_BENCHMARK_DEPS,
4745)
4746
4747xnnpack_benchmark(
Marat Dukhanf4db2f32020-06-30 10:55:30 -07004748 name = "f32_vsqrt_bench",
4749 srcs = [
4750 "bench/f32-vsqrt.cc",
4751 "src/xnnpack/AlignedAllocator.h",
4752 ] + MICROKERNEL_BENCHMARK_HDRS,
4753 deps = MICROKERNEL_BENCHMARK_DEPS,
4754)
4755
4756xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004757 name = "f32_im2col_gemm_bench",
4758 srcs = [
4759 "bench/f32-im2col-gemm.cc",
4760 "bench/conv.h",
4761 "src/xnnpack/AlignedAllocator.h",
4762 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004763 deps = MICROKERNEL_BENCHMARK_DEPS + [
4764 ":im2col",
4765 ":packing",
4766 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004767)
4768
Marat Dukhanfe7acb62020-03-09 19:30:05 -07004769xnnpack_benchmark(
Marat Dukhanffbf96a2020-05-14 02:59:08 -07004770 name = "rounding_bench",
4771 srcs = [
4772 "bench/rounding.cc",
Marat Dukhanffbf96a2020-05-14 02:59:08 -07004773 "src/xnnpack/AlignedAllocator.h",
Frank Barchard04336c12020-10-22 16:48:55 -07004774 "src/xnnpack/math-stubs.h",
Marat Dukhanffbf96a2020-05-14 02:59:08 -07004775 ] + MICROKERNEL_BENCHMARK_HDRS,
4776 deps = MICROKERNEL_BENCHMARK_DEPS,
4777)
4778
Marat Dukhan08c4a432019-10-03 09:29:21 -07004779########################### Benchmarks for operators ###########################
4780
4781xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004782 name = "average_pooling_bench",
4783 srcs = ["bench/average-pooling.cc"],
Marat Dukhan7a16d8b2020-03-11 04:22:44 -07004784 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07004785 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004786 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004787)
4788
4789xnnpack_benchmark(
Marat Dukhan87727142020-06-24 15:24:10 -07004790 name = "bankers_rounding_bench",
4791 srcs = ["bench/bankers-rounding.cc"],
4792 copts = xnnpack_optional_tflite_copts(),
4793 tags = ["nowin32"],
4794 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
4795)
4796
4797xnnpack_benchmark(
4798 name = "ceiling_bench",
4799 srcs = ["bench/ceiling.cc"],
4800 copts = xnnpack_optional_tflite_copts(),
4801 tags = ["nowin32"],
4802 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
4803)
4804
4805xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004806 name = "channel_shuffle_bench",
4807 srcs = ["bench/channel-shuffle.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004808 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07004809)
4810
4811xnnpack_benchmark(
4812 name = "convolution_bench",
4813 srcs = ["bench/convolution.cc"],
4814 copts = xnnpack_optional_tflite_copts() + xnnpack_optional_armcl_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07004815 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004816 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps() + xnnpack_optional_armcl_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004817)
4818
4819xnnpack_benchmark(
4820 name = "deconvolution_bench",
4821 srcs = ["bench/deconvolution.cc"],
4822 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07004823 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004824 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004825)
4826
4827xnnpack_benchmark(
Marat Dukhan87727142020-06-24 15:24:10 -07004828 name = "floor_bench",
4829 srcs = ["bench/floor.cc"],
4830 copts = xnnpack_optional_tflite_copts(),
4831 tags = ["nowin32"],
4832 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
4833)
4834
4835xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004836 name = "global_average_pooling_bench",
4837 srcs = ["bench/global-average-pooling.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004838 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07004839)
4840
4841xnnpack_benchmark(
Marat Dukhanad352602020-06-25 21:50:54 -07004842 name = "hardswish_bench",
4843 srcs = ["bench/hardswish.cc"],
4844 copts = xnnpack_optional_tflite_copts(),
4845 tags = ["nowin32"],
4846 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
4847)
4848
4849xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004850 name = "max_pooling_bench",
4851 srcs = ["bench/max-pooling.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004852 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07004853)
4854
4855xnnpack_benchmark(
4856 name = "sigmoid_bench",
4857 srcs = ["bench/sigmoid.cc"],
Marat Dukhanc3b9e862019-11-17 13:18:54 -08004858 copts = xnnpack_optional_tflite_copts(),
Marat Dukhanca2ba702020-04-24 01:31:47 -07004859 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004860 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004861)
4862
4863xnnpack_benchmark(
Marat Dukhan95b22432019-10-30 16:30:14 -07004864 name = "prelu_bench",
4865 srcs = ["bench/prelu.cc"],
4866 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07004867 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004868 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan95b22432019-10-30 16:30:14 -07004869)
4870
4871xnnpack_benchmark(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08004872 name = "softmax_bench",
4873 srcs = ["bench/softmax.cc"],
Marat Dukhan9c0db962020-01-28 12:30:14 -08004874 copts = xnnpack_optional_tflite_copts(),
Marat Dukhanca2ba702020-04-24 01:31:47 -07004875 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004876 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004877)
4878
Marat Dukhan87727142020-06-24 15:24:10 -07004879xnnpack_benchmark(
Marat Dukhan6804bbd2020-06-30 19:26:11 -07004880 name = "square_root_bench",
4881 srcs = ["bench/square-root.cc"],
4882 copts = xnnpack_optional_tflite_copts(),
4883 tags = ["nowin32"],
4884 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
4885)
4886
4887xnnpack_benchmark(
Marat Dukhan87727142020-06-24 15:24:10 -07004888 name = "truncation_bench",
4889 srcs = ["bench/truncation.cc"],
4890 deps = OPERATOR_BENCHMARK_DEPS,
4891)
4892
Marat Dukhanc068bb62019-10-04 13:24:39 -07004893############################# End-to-end benchmarks ############################
4894
4895cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07004896 name = "fp32_mobilenet_v1",
4897 srcs = ["models/fp32-mobilenet-v1.cc"],
Marat Dukhanc068bb62019-10-04 13:24:39 -07004898 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08004899 copts = xnnpack_std_cxxopts(),
Marat Dukhanc068bb62019-10-04 13:24:39 -07004900 linkstatic = True,
4901 deps = [
4902 ":XNNPACK",
4903 "@pthreadpool",
4904 ],
4905)
4906
4907cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07004908 name = "fp16_mobilenet_v1",
4909 srcs = ["models/fp16-mobilenet-v1.cc"],
4910 hdrs = ["models/models.h"],
4911 copts = xnnpack_std_cxxopts(),
4912 linkstatic = True,
4913 deps = [
4914 ":XNNPACK",
4915 "@FP16",
4916 "@pthreadpool",
4917 ],
4918)
4919
4920cc_library(
Marat Dukhan0743cdf2020-08-04 18:52:07 -07004921 name = "qs8_mobilenet_v1",
4922 srcs = ["models/qs8-mobilenet-v1.cc"],
4923 hdrs = ["models/models.h"],
4924 copts = xnnpack_std_cxxopts(),
4925 linkstatic = True,
4926 deps = [
4927 ":XNNPACK",
4928 "@pthreadpool",
4929 ],
4930)
4931
4932cc_library(
Marat Dukhan70a96182020-09-03 17:13:58 -07004933 name = "qs8_mobilenet_v2",
4934 srcs = ["models/qs8-mobilenet-v2.cc"],
4935 hdrs = ["models/models.h"],
4936 copts = xnnpack_std_cxxopts(),
4937 linkstatic = True,
4938 deps = [
4939 ":XNNPACK",
4940 "@pthreadpool",
4941 ],
4942)
4943
4944cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07004945 name = "fp32_mobilenet_v2",
4946 srcs = ["models/fp32-mobilenet-v2.cc"],
Marat Dukhanc068bb62019-10-04 13:24:39 -07004947 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08004948 copts = xnnpack_std_cxxopts(),
Marat Dukhanc068bb62019-10-04 13:24:39 -07004949 linkstatic = True,
4950 deps = [
4951 ":XNNPACK",
4952 "@pthreadpool",
4953 ],
4954)
4955
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004956cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07004957 name = "fp16_mobilenet_v2",
4958 srcs = ["models/fp16-mobilenet-v2.cc"],
4959 hdrs = ["models/models.h"],
4960 copts = xnnpack_std_cxxopts(),
4961 linkstatic = True,
4962 deps = [
4963 ":XNNPACK",
4964 "@FP16",
4965 "@pthreadpool",
4966 ],
4967)
4968
4969cc_library(
4970 name = "fp32_mobilenet_v3_large",
4971 srcs = ["models/fp32-mobilenet-v3-large.cc"],
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004972 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08004973 copts = xnnpack_std_cxxopts(),
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004974 linkstatic = True,
4975 deps = [
4976 ":XNNPACK",
4977 "@pthreadpool",
4978 ],
4979)
4980
4981cc_library(
Marat Dukhan66f3ccd2020-09-14 16:09:38 -07004982 name = "fp16_mobilenet_v3_large",
4983 srcs = ["models/fp16-mobilenet-v3-large.cc"],
4984 hdrs = ["models/models.h"],
4985 copts = xnnpack_std_cxxopts(),
4986 linkstatic = True,
4987 deps = [
4988 ":XNNPACK",
4989 "@FP16",
4990 "@pthreadpool",
4991 ],
4992)
4993
4994cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07004995 name = "fp32_mobilenet_v3_small",
4996 srcs = ["models/fp32-mobilenet-v3-small.cc"],
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004997 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08004998 copts = xnnpack_std_cxxopts(),
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004999 linkstatic = True,
5000 deps = [
5001 ":XNNPACK",
5002 "@pthreadpool",
5003 ],
5004)
5005
Marat Dukhan66f3ccd2020-09-14 16:09:38 -07005006cc_library(
5007 name = "fp16_mobilenet_v3_small",
5008 srcs = ["models/fp16-mobilenet-v3-small.cc"],
5009 hdrs = ["models/models.h"],
5010 copts = xnnpack_std_cxxopts(),
5011 linkstatic = True,
5012 deps = [
5013 ":XNNPACK",
5014 "@FP16",
5015 "@pthreadpool",
5016 ],
5017)
5018
Marat Dukhanc068bb62019-10-04 13:24:39 -07005019xnnpack_benchmark(
Marat Dukhanef4416e2019-10-31 13:44:40 -07005020 name = "f32_dwconv_e2e_bench",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08005021 srcs = [
5022 "bench/f32-dwconv-e2e.cc",
5023 "bench/end2end.h",
5024 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanef4416e2019-10-31 13:44:40 -07005025 deps = MICROKERNEL_BENCHMARK_DEPS + [
5026 ":XNNPACK",
Marat Dukhan270a2c42020-06-26 16:45:52 -07005027 ":fp32_mobilenet_v1",
5028 ":fp32_mobilenet_v2",
5029 ":fp32_mobilenet_v3_large",
5030 ":fp32_mobilenet_v3_small",
Marat Dukhanef4416e2019-10-31 13:44:40 -07005031 ],
5032)
5033
5034xnnpack_benchmark(
Marat Dukhan5f18d262019-10-31 10:24:14 -07005035 name = "f32_gemm_e2e_bench",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08005036 srcs = [
5037 "bench/f32-gemm-e2e.cc",
5038 "bench/end2end.h",
5039 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan5f18d262019-10-31 10:24:14 -07005040 deps = MICROKERNEL_BENCHMARK_DEPS + [
5041 ":XNNPACK",
Marat Dukhan270a2c42020-06-26 16:45:52 -07005042 ":fp32_mobilenet_v1",
5043 ":fp32_mobilenet_v2",
5044 ":fp32_mobilenet_v3_large",
5045 ":fp32_mobilenet_v3_small",
Marat Dukhan5f18d262019-10-31 10:24:14 -07005046 ],
5047)
5048
5049xnnpack_benchmark(
Marat Dukhanc068bb62019-10-04 13:24:39 -07005050 name = "end2end_bench",
5051 srcs = ["bench/end2end.cc"],
5052 deps = [
5053 ":XNNPACK",
Frank Barchardc712fa42019-10-31 14:00:21 -07005054 ":bench_utils",
Marat Dukhan270a2c42020-06-26 16:45:52 -07005055 ":fp16_mobilenet_v1",
5056 ":fp16_mobilenet_v2",
Marat Dukhan66f3ccd2020-09-14 16:09:38 -07005057 ":fp16_mobilenet_v3_large",
5058 ":fp16_mobilenet_v3_small",
Marat Dukhan270a2c42020-06-26 16:45:52 -07005059 ":fp32_mobilenet_v1",
5060 ":fp32_mobilenet_v2",
5061 ":fp32_mobilenet_v3_large",
5062 ":fp32_mobilenet_v3_small",
Marat Dukhan0743cdf2020-08-04 18:52:07 -07005063 ":qs8_mobilenet_v1",
Marat Dukhan70a96182020-09-03 17:13:58 -07005064 ":qs8_mobilenet_v2",
Marat Dukhanc068bb62019-10-04 13:24:39 -07005065 "@pthreadpool",
5066 ],
5067)
5068
Marat Dukhan6adff4e2019-10-14 18:32:07 -07005069#################### Accuracy evaluation for math functions ####################
5070
5071xnnpack_benchmark(
5072 name = "f32_exp_eval",
5073 srcs = [
5074 "eval/f32-exp.cc",
5075 "src/xnnpack/AlignedAllocator.h",
5076 ] + ACCURACY_EVAL_HDRS,
5077 deps = ACCURACY_EVAL_DEPS,
5078)
5079
Marat Dukhan515c9772019-10-17 18:07:57 -07005080xnnpack_benchmark(
5081 name = "f32_expminus_eval",
5082 srcs = [
5083 "eval/f32-expminus.cc",
5084 "src/xnnpack/AlignedAllocator.h",
5085 ] + ACCURACY_EVAL_HDRS,
5086 deps = ACCURACY_EVAL_DEPS,
5087)
5088
Marat Dukhan98ba4412019-10-23 02:14:28 -07005089xnnpack_benchmark(
Marat Dukhana438aca2020-11-20 15:45:01 -08005090 name = "f32_expm1minus_eval",
5091 srcs = [
5092 "eval/f32-expm1minus.cc",
5093 "src/xnnpack/AlignedAllocator.h",
5094 ] + ACCURACY_EVAL_HDRS,
5095 deps = ACCURACY_EVAL_DEPS,
5096)
5097
5098xnnpack_benchmark(
Marat Dukhan98ba4412019-10-23 02:14:28 -07005099 name = "f32_extexp_eval",
5100 srcs = [
5101 "eval/f32-extexp.cc",
5102 "src/xnnpack/AlignedAllocator.h",
5103 ] + ACCURACY_EVAL_HDRS,
5104 deps = ACCURACY_EVAL_DEPS,
5105)
5106
Marat Dukhan8853b822020-05-07 12:19:01 -07005107xnnpack_unit_test(
5108 name = "f32_roundne_eval",
5109 srcs = [
5110 "eval/f32-roundne.cc",
5111 "src/xnnpack/AlignedAllocator.h",
5112 "src/xnnpack/math-stubs.h",
5113 ] + MICROKERNEL_TEST_HDRS,
Marat Dukhan22eed3d2020-05-11 20:13:37 -07005114 automatic = False,
Marat Dukhan8853b822020-05-07 12:19:01 -07005115 deps = MICROKERNEL_TEST_DEPS,
5116)
5117
Marat Dukhan2dbb9442020-05-12 20:43:43 -07005118xnnpack_unit_test(
Marat Dukhanc9852ba2020-05-13 17:21:29 -07005119 name = "f32_roundd_eval",
5120 srcs = [
5121 "eval/f32-roundd.cc",
5122 "src/xnnpack/AlignedAllocator.h",
5123 "src/xnnpack/math-stubs.h",
5124 ] + MICROKERNEL_TEST_HDRS,
5125 automatic = False,
5126 deps = MICROKERNEL_TEST_DEPS,
5127)
5128
5129xnnpack_unit_test(
5130 name = "f32_roundu_eval",
5131 srcs = [
5132 "eval/f32-roundu.cc",
5133 "src/xnnpack/AlignedAllocator.h",
5134 "src/xnnpack/math-stubs.h",
5135 ] + MICROKERNEL_TEST_HDRS,
5136 automatic = False,
5137 deps = MICROKERNEL_TEST_DEPS,
5138)
5139
5140xnnpack_unit_test(
Marat Dukhan2dbb9442020-05-12 20:43:43 -07005141 name = "f32_roundz_eval",
5142 srcs = [
5143 "eval/f32-roundz.cc",
5144 "src/xnnpack/AlignedAllocator.h",
5145 "src/xnnpack/math-stubs.h",
5146 ] + MICROKERNEL_TEST_HDRS,
Marat Dukhanc9852ba2020-05-13 17:21:29 -07005147 automatic = False,
Marat Dukhan2dbb9442020-05-12 20:43:43 -07005148 deps = MICROKERNEL_TEST_DEPS,
5149)
5150
Marat Dukhan346a9e52019-11-15 09:06:30 -08005151xnnpack_benchmark(
5152 name = "f32_sigmoid_eval",
5153 srcs = [
5154 "eval/f32-sigmoid.cc",
5155 "src/xnnpack/AlignedAllocator.h",
5156 ] + ACCURACY_EVAL_HDRS,
5157 deps = ACCURACY_EVAL_DEPS,
5158)
5159
Marat Dukhan84000762020-06-29 18:38:43 -07005160xnnpack_benchmark(
5161 name = "f32_sqrt_eval",
5162 srcs = [
5163 "eval/f32-sqrt.cc",
5164 "src/xnnpack/AlignedAllocator.h",
5165 ] + ACCURACY_EVAL_HDRS,
5166 deps = ACCURACY_EVAL_DEPS + [
5167 ":bench_utils",
5168 ],
5169)
5170
Marat Dukhan08c4a432019-10-03 09:29:21 -07005171######################### Unit tests for micro-kernels #########################
5172
5173xnnpack_unit_test(
Marat Dukhande06f492020-04-09 00:19:31 -07005174 name = "f16_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005175 srcs = [
Marat Dukhande06f492020-04-09 00:19:31 -07005176 "test/f16-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005177 "test/gemm-microkernel-tester.h",
5178 "src/xnnpack/AlignedAllocator.h",
5179 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005180 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005181)
5182
5183xnnpack_unit_test(
Marat Dukhan355ab432020-04-09 19:01:52 -07005184 name = "f16_spmm_minmax_test",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08005185 srcs = [
Marat Dukhan355ab432020-04-09 19:01:52 -07005186 "test/f16-spmm-minmax.cc",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08005187 "test/spmm-microkernel-tester.h",
5188 "src/xnnpack/AlignedAllocator.h",
5189 ] + MICROKERNEL_TEST_HDRS,
5190 deps = MICROKERNEL_TEST_DEPS,
5191)
5192
5193xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005194 name = "f32_argmaxpool_test",
5195 srcs = [
5196 "test/f32-argmaxpool.cc",
5197 "test/argmaxpool-microkernel-tester.h",
5198 "src/xnnpack/AlignedAllocator.h",
5199 ] + MICROKERNEL_TEST_HDRS,
5200 deps = MICROKERNEL_TEST_DEPS,
5201)
5202
5203xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07005204 name = "f32_avgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005205 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07005206 "test/f32-avgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005207 "test/avgpool-microkernel-tester.h",
5208 "src/xnnpack/AlignedAllocator.h",
5209 ] + MICROKERNEL_TEST_HDRS,
5210 deps = MICROKERNEL_TEST_DEPS,
5211)
5212
5213xnnpack_unit_test(
Marat Dukhan660fd192020-03-10 04:55:30 -07005214 name = "f32_ibilinear_test",
Marat Dukhan35dacfb2019-11-07 19:18:16 -08005215 srcs = [
Marat Dukhan660fd192020-03-10 04:55:30 -07005216 "test/f32-ibilinear.cc",
5217 "test/ibilinear-microkernel-tester.h",
Marat Dukhan35dacfb2019-11-07 19:18:16 -08005218 "src/xnnpack/AlignedAllocator.h",
5219 ] + MICROKERNEL_TEST_HDRS,
5220 deps = MICROKERNEL_TEST_DEPS,
5221)
5222
5223xnnpack_unit_test(
XNNPACK Team21432672020-10-19 19:58:48 -07005224 name = "f32_ibilinear_chw_test",
5225 srcs = [
5226 "test/f32-ibilinear-chw.cc",
XNNPACK Team6be46b22020-10-22 23:34:54 -07005227 "test/ibilinear-microkernel-tester.h",
XNNPACK Team21432672020-10-19 19:58:48 -07005228 "src/xnnpack/AlignedAllocator.h",
5229 ] + MICROKERNEL_TEST_HDRS,
5230 deps = MICROKERNEL_TEST_DEPS,
5231)
5232
5233xnnpack_unit_test(
Frank Barchardb1966592020-05-12 13:47:06 -07005234 name = "f16_clamp_test",
5235 srcs = [
5236 "test/f16-clamp.cc",
5237 "test/clamp-microkernel-tester.h",
5238 ] + MICROKERNEL_TEST_HDRS,
5239 deps = MICROKERNEL_TEST_DEPS,
5240)
5241
5242xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005243 name = "f32_clamp_test",
5244 srcs = [
5245 "test/f32-clamp.cc",
5246 "test/clamp-microkernel-tester.h",
5247 ] + MICROKERNEL_TEST_HDRS,
5248 deps = MICROKERNEL_TEST_DEPS,
5249)
5250
5251xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07005252 name = "f32_igemm_test",
5253 srcs = [
5254 "test/f32-igemm.cc",
5255 "test/gemm-microkernel-tester.h",
5256 "src/xnnpack/AlignedAllocator.h",
5257 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005258 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan163a7e62020-04-09 04:19:26 -07005259)
5260
5261xnnpack_unit_test(
Marat Dukhan467f6362020-05-22 23:21:55 -07005262 name = "f32_igemm_relu_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005263 srcs = [
Marat Dukhan467f6362020-05-22 23:21:55 -07005264 "test/f32-igemm-relu.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005265 "test/gemm-microkernel-tester.h",
5266 "src/xnnpack/AlignedAllocator.h",
5267 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005268 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005269)
5270
5271xnnpack_unit_test(
Marat Dukhane207b7b2020-05-28 16:27:42 -07005272 name = "f32_igemm_minmax_test",
5273 srcs = [
5274 "test/f32-igemm-minmax.cc",
5275 "test/gemm-microkernel-tester.h",
5276 "src/xnnpack/AlignedAllocator.h",
5277 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005278 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhane207b7b2020-05-28 16:27:42 -07005279)
5280
5281xnnpack_unit_test(
Frank Barchardb0e4fae2020-05-04 15:27:51 -07005282 name = "f16_igemm_minmax_test",
5283 srcs = [
5284 "test/f16-igemm-minmax.cc",
5285 "test/gemm-microkernel-tester.h",
5286 "src/xnnpack/AlignedAllocator.h",
5287 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005288 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Frank Barchardb0e4fae2020-05-04 15:27:51 -07005289)
5290
5291xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005292 name = "f32_conv_hwc_test",
5293 srcs = [
5294 "test/f32-conv-hwc.cc",
5295 "test/conv-hwc-microkernel-tester.h",
5296 "src/xnnpack/AlignedAllocator.h",
5297 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005298 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005299)
5300
5301xnnpack_unit_test(
Marat Dukhan1f29b802020-05-15 23:46:39 -07005302 name = "f32_conv_hwc2chw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005303 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07005304 "test/f32-conv-hwc2chw.cc",
5305 "test/conv-hwc2chw-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005306 "src/xnnpack/AlignedAllocator.h",
5307 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005308 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005309)
5310
5311xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07005312 name = "f32_dwconv_test",
5313 srcs = [
5314 "test/f32-dwconv.cc",
5315 "test/dwconv-microkernel-tester.h",
5316 "src/xnnpack/AlignedAllocator.h",
5317 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005318 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan163a7e62020-04-09 04:19:26 -07005319)
5320
5321xnnpack_unit_test(
Frank Barchard5a599a62020-06-04 20:12:44 -07005322 name = "f16_dwconv_minmax_test",
5323 srcs = [
5324 "test/f16-dwconv-minmax.cc",
5325 "test/dwconv-microkernel-tester.h",
5326 "src/xnnpack/AlignedAllocator.h",
5327 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005328 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Frank Barchard5a599a62020-06-04 20:12:44 -07005329)
5330
5331xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07005332 name = "f32_dwconv_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005333 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07005334 "test/f32-dwconv-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005335 "test/dwconv-microkernel-tester.h",
5336 "src/xnnpack/AlignedAllocator.h",
5337 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005338 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005339)
5340
5341xnnpack_unit_test(
Marat Dukhanbf715f92020-10-23 20:17:00 -07005342 name = "f32_dwconv2d_chw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005343 srcs = [
Marat Dukhanbf715f92020-10-23 20:17:00 -07005344 "test/f32-dwconv2d-chw.cc",
5345 "test/dwconv2d-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005346 "src/xnnpack/AlignedAllocator.h",
5347 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005348 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005349)
5350
5351xnnpack_unit_test(
Frank Barchard0bb49a72020-06-04 11:35:11 -07005352 name = "f16_gavgpool_minmax_test",
5353 srcs = [
5354 "test/f16-gavgpool-minmax.cc",
5355 "test/gavgpool-microkernel-tester.h",
5356 "src/xnnpack/AlignedAllocator.h",
5357 ] + MICROKERNEL_TEST_HDRS,
5358 deps = MICROKERNEL_TEST_DEPS,
5359)
5360
5361xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07005362 name = "f32_gavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005363 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07005364 "test/f32-gavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005365 "test/gavgpool-microkernel-tester.h",
5366 "src/xnnpack/AlignedAllocator.h",
5367 ] + MICROKERNEL_TEST_HDRS,
5368 deps = MICROKERNEL_TEST_DEPS,
5369)
5370
5371xnnpack_unit_test(
Marat Dukhan1f29b802020-05-15 23:46:39 -07005372 name = "f32_gavgpool_cw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005373 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07005374 "test/f32-gavgpool-cw.cc",
5375 "test/gavgpool-cw-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005376 "src/xnnpack/AlignedAllocator.h",
5377 ] + MICROKERNEL_TEST_HDRS,
5378 deps = MICROKERNEL_TEST_DEPS,
5379)
5380
5381xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07005382 name = "f32_gemm_test",
5383 srcs = [
5384 "test/f32-gemm.cc",
5385 "test/gemm-microkernel-tester.h",
5386 "src/xnnpack/AlignedAllocator.h",
5387 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005388 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan163a7e62020-04-09 04:19:26 -07005389)
5390
5391xnnpack_unit_test(
Marat Dukhan467f6362020-05-22 23:21:55 -07005392 name = "f32_gemm_relu_test",
5393 srcs = [
5394 "test/f32-gemm-relu.cc",
5395 "test/gemm-microkernel-tester.h",
5396 "src/xnnpack/AlignedAllocator.h",
5397 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005398 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan467f6362020-05-22 23:21:55 -07005399)
5400
5401xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07005402 name = "f32_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005403 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07005404 "test/f32-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005405 "test/gemm-microkernel-tester.h",
5406 "src/xnnpack/AlignedAllocator.h",
5407 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005408 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005409)
5410
5411xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07005412 name = "f32_gemminc_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005413 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07005414 "test/f32-gemminc-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005415 "test/gemm-microkernel-tester.h",
5416 "src/xnnpack/AlignedAllocator.h",
5417 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005418 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005419)
5420
5421xnnpack_unit_test(
Frank Barchardb1966592020-05-12 13:47:06 -07005422 name = "f16_hswish_test",
5423 srcs = [
5424 "test/f16-hswish.cc",
5425 "test/hswish-microkernel-tester.h",
5426 ] + MICROKERNEL_TEST_HDRS,
5427 deps = MICROKERNEL_TEST_DEPS,
5428)
5429
5430xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005431 name = "f32_hswish_test",
5432 srcs = [
5433 "test/f32-hswish.cc",
5434 "test/hswish-microkernel-tester.h",
5435 ] + MICROKERNEL_TEST_HDRS,
5436 deps = MICROKERNEL_TEST_DEPS,
5437)
5438
5439xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07005440 name = "f32_maxpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005441 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07005442 "test/f32-maxpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005443 "test/maxpool-microkernel-tester.h",
5444 ] + MICROKERNEL_TEST_HDRS,
5445 deps = MICROKERNEL_TEST_DEPS,
5446)
5447
5448xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07005449 name = "f32_pavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005450 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07005451 "test/f32-pavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005452 "test/avgpool-microkernel-tester.h",
5453 "src/xnnpack/AlignedAllocator.h",
5454 ] + MICROKERNEL_TEST_HDRS,
5455 deps = MICROKERNEL_TEST_DEPS,
5456)
5457
5458xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07005459 name = "f32_ppmm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005460 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07005461 "test/f32-ppmm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005462 "test/gemm-microkernel-tester.h",
5463 "src/xnnpack/AlignedAllocator.h",
5464 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005465 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005466)
5467
5468xnnpack_unit_test(
Frank Barchardb1966592020-05-12 13:47:06 -07005469 name = "f16_prelu_test",
5470 srcs = [
5471 "test/f16-prelu.cc",
5472 "test/prelu-microkernel-tester.h",
5473 "src/xnnpack/AlignedAllocator.h",
5474 ] + MICROKERNEL_TEST_HDRS,
5475 deps = MICROKERNEL_TEST_DEPS,
5476)
5477
5478xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005479 name = "f32_prelu_test",
5480 srcs = [
5481 "test/f32-prelu.cc",
5482 "test/prelu-microkernel-tester.h",
5483 "src/xnnpack/AlignedAllocator.h",
5484 ] + MICROKERNEL_TEST_HDRS,
5485 deps = MICROKERNEL_TEST_DEPS,
5486)
5487
5488xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07005489 name = "f32_raddexpminusmax_test",
5490 srcs = [
5491 "test/f32-raddexpminusmax.cc",
5492 "test/raddexpminusmax-microkernel-tester.h",
5493 ] + MICROKERNEL_TEST_HDRS,
5494 deps = MICROKERNEL_TEST_DEPS,
5495)
5496
5497xnnpack_unit_test(
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07005498 name = "f32_raddextexp_test",
5499 srcs = [
5500 "test/f32-raddextexp.cc",
5501 "test/raddextexp-microkernel-tester.h",
5502 ] + MICROKERNEL_TEST_HDRS,
5503 deps = MICROKERNEL_TEST_DEPS,
5504)
5505
5506xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07005507 name = "f32_raddstoreexpminusmax_test",
5508 srcs = [
5509 "test/f32-raddstoreexpminusmax.cc",
5510 "test/raddstoreexpminusmax-microkernel-tester.h",
5511 ] + MICROKERNEL_TEST_HDRS,
5512 deps = MICROKERNEL_TEST_DEPS,
5513)
5514
5515xnnpack_unit_test(
Frank Barchardfb158e22020-07-15 16:10:10 -07005516 name = "f32_relu_test",
5517 srcs = [
5518 "test/f32-relu.cc",
5519 "test/vunary-microkernel-tester.h",
5520 ] + MICROKERNEL_TEST_HDRS,
5521 deps = MICROKERNEL_TEST_DEPS,
5522)
5523
5524xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005525 name = "f32_rmax_test",
5526 srcs = [
5527 "test/f32-rmax.cc",
5528 "test/rmax-microkernel-tester.h",
5529 ] + MICROKERNEL_TEST_HDRS,
5530 deps = MICROKERNEL_TEST_DEPS,
5531)
5532
5533xnnpack_unit_test(
Marat Dukhan346a9e52019-11-15 09:06:30 -08005534 name = "f32_sigmoid_test",
5535 srcs = [
5536 "test/f32-sigmoid.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005537 "test/vunary-microkernel-tester.h",
Marat Dukhan346a9e52019-11-15 09:06:30 -08005538 ] + MICROKERNEL_TEST_HDRS,
5539 deps = MICROKERNEL_TEST_DEPS,
5540)
5541
5542xnnpack_unit_test(
Marat Dukhan355ab432020-04-09 19:01:52 -07005543 name = "f32_spmm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005544 srcs = [
Marat Dukhan355ab432020-04-09 19:01:52 -07005545 "test/f32-spmm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005546 "test/spmm-microkernel-tester.h",
5547 "src/xnnpack/AlignedAllocator.h",
5548 ] + MICROKERNEL_TEST_HDRS,
5549 deps = MICROKERNEL_TEST_DEPS,
5550)
5551
5552xnnpack_unit_test(
Frank Barchardd793f6c2020-05-08 13:37:43 -07005553 name = "f16_vadd_minmax_test",
5554 srcs = [
5555 "test/f16-vadd-minmax.cc",
5556 "test/vbinary-microkernel-tester.h",
5557 ] + MICROKERNEL_TEST_HDRS,
5558 deps = MICROKERNEL_TEST_DEPS,
5559)
5560
5561xnnpack_unit_test(
5562 name = "f16_vaddc_minmax_test",
5563 srcs = [
5564 "test/f16-vaddc-minmax.cc",
5565 "test/vbinaryc-microkernel-tester.h",
5566 ] + MICROKERNEL_TEST_HDRS,
5567 deps = MICROKERNEL_TEST_DEPS,
5568)
5569
5570xnnpack_unit_test(
5571 name = "f16_vdiv_minmax_test",
5572 srcs = [
5573 "test/f16-vdiv-minmax.cc",
5574 "test/vbinary-microkernel-tester.h",
5575 ] + MICROKERNEL_TEST_HDRS,
5576 deps = MICROKERNEL_TEST_DEPS,
5577)
5578
5579xnnpack_unit_test(
5580 name = "f16_vdivc_minmax_test",
5581 srcs = [
5582 "test/f16-vdivc-minmax.cc",
5583 "test/vbinaryc-microkernel-tester.h",
5584 ] + MICROKERNEL_TEST_HDRS,
5585 deps = MICROKERNEL_TEST_DEPS,
5586)
5587
5588xnnpack_unit_test(
5589 name = "f16_vrdivc_minmax_test",
5590 srcs = [
5591 "test/f16-vrdivc-minmax.cc",
5592 "test/vbinaryc-microkernel-tester.h",
5593 ] + MICROKERNEL_TEST_HDRS,
5594 deps = MICROKERNEL_TEST_DEPS,
5595)
5596
5597xnnpack_unit_test(
5598 name = "f16_vmax_test",
5599 srcs = [
5600 "test/f16-vmax.cc",
5601 "test/vbinary-microkernel-tester.h",
5602 ] + MICROKERNEL_TEST_HDRS,
5603 deps = MICROKERNEL_TEST_DEPS,
5604)
5605
5606xnnpack_unit_test(
5607 name = "f16_vmaxc_test",
5608 srcs = [
5609 "test/f16-vmaxc.cc",
5610 "test/vbinaryc-microkernel-tester.h",
5611 ] + MICROKERNEL_TEST_HDRS,
5612 deps = MICROKERNEL_TEST_DEPS,
5613)
5614
5615xnnpack_unit_test(
5616 name = "f16_vmin_test",
5617 srcs = [
5618 "test/f16-vmin.cc",
5619 "test/vbinary-microkernel-tester.h",
5620 ] + MICROKERNEL_TEST_HDRS,
5621 deps = MICROKERNEL_TEST_DEPS,
5622)
5623
5624xnnpack_unit_test(
5625 name = "f16_vminc_test",
5626 srcs = [
5627 "test/f16-vminc.cc",
5628 "test/vbinaryc-microkernel-tester.h",
5629 ] + MICROKERNEL_TEST_HDRS,
5630 deps = MICROKERNEL_TEST_DEPS,
5631)
5632
5633xnnpack_unit_test(
5634 name = "f16_vmul_minmax_test",
5635 srcs = [
5636 "test/f16-vmul-minmax.cc",
5637 "test/vbinary-microkernel-tester.h",
5638 ] + MICROKERNEL_TEST_HDRS,
5639 deps = MICROKERNEL_TEST_DEPS,
5640)
5641
5642xnnpack_unit_test(
5643 name = "f16_vmulc_minmax_test",
5644 srcs = [
5645 "test/f16-vmulc-minmax.cc",
5646 "test/vbinaryc-microkernel-tester.h",
5647 ] + MICROKERNEL_TEST_HDRS,
5648 deps = MICROKERNEL_TEST_DEPS,
5649)
5650
5651xnnpack_unit_test(
5652 name = "f16_vsub_minmax_test",
5653 srcs = [
5654 "test/f16-vsub-minmax.cc",
5655 "test/vbinary-microkernel-tester.h",
5656 ] + MICROKERNEL_TEST_HDRS,
5657 deps = MICROKERNEL_TEST_DEPS,
5658)
5659
5660xnnpack_unit_test(
5661 name = "f16_vsubc_minmax_test",
5662 srcs = [
5663 "test/f16-vsubc-minmax.cc",
5664 "test/vbinaryc-microkernel-tester.h",
5665 ] + MICROKERNEL_TEST_HDRS,
5666 deps = MICROKERNEL_TEST_DEPS,
5667)
5668
5669xnnpack_unit_test(
5670 name = "f16_vrsubc_minmax_test",
5671 srcs = [
5672 "test/f16-vrsubc-minmax.cc",
5673 "test/vbinaryc-microkernel-tester.h",
5674 ] + MICROKERNEL_TEST_HDRS,
5675 deps = MICROKERNEL_TEST_DEPS,
5676)
5677
5678xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07005679 name = "f32_vabs_test",
5680 srcs = [
5681 "test/f32-vabs.cc",
5682 "test/vunary-microkernel-tester.h",
5683 ] + MICROKERNEL_TEST_HDRS,
5684 deps = MICROKERNEL_TEST_DEPS,
5685)
5686
5687xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005688 name = "f32_vadd_test",
5689 srcs = [
5690 "test/f32-vadd.cc",
5691 "test/vbinary-microkernel-tester.h",
5692 ] + MICROKERNEL_TEST_HDRS,
5693 deps = MICROKERNEL_TEST_DEPS,
5694)
5695
5696xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005697 name = "f32_vadd_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005698 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005699 "test/f32-vadd-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005700 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005701 ] + MICROKERNEL_TEST_HDRS,
5702 deps = MICROKERNEL_TEST_DEPS,
5703)
5704
5705xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005706 name = "f32_vadd_relu_test",
5707 srcs = [
5708 "test/f32-vadd-relu.cc",
5709 "test/vbinary-microkernel-tester.h",
5710 ] + MICROKERNEL_TEST_HDRS,
5711 deps = MICROKERNEL_TEST_DEPS,
5712)
5713
5714xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005715 name = "f32_vaddc_test",
5716 srcs = [
5717 "test/f32-vaddc.cc",
5718 "test/vbinaryc-microkernel-tester.h",
5719 ] + MICROKERNEL_TEST_HDRS,
5720 deps = MICROKERNEL_TEST_DEPS,
5721)
5722
5723xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005724 name = "f32_vaddc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005725 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005726 "test/f32-vaddc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005727 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005728 ] + MICROKERNEL_TEST_HDRS,
5729 deps = MICROKERNEL_TEST_DEPS,
5730)
5731
5732xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005733 name = "f32_vaddc_relu_test",
5734 srcs = [
5735 "test/f32-vaddc-relu.cc",
5736 "test/vbinaryc-microkernel-tester.h",
5737 ] + MICROKERNEL_TEST_HDRS,
5738 deps = MICROKERNEL_TEST_DEPS,
5739)
5740
5741xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005742 name = "f32_vdiv_test",
5743 srcs = [
5744 "test/f32-vdiv.cc",
5745 "test/vbinary-microkernel-tester.h",
5746 ] + MICROKERNEL_TEST_HDRS,
5747 deps = MICROKERNEL_TEST_DEPS,
5748)
5749
5750xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005751 name = "f32_vdiv_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08005752 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005753 "test/f32-vdiv-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08005754 "test/vbinary-microkernel-tester.h",
5755 ] + MICROKERNEL_TEST_HDRS,
5756 deps = MICROKERNEL_TEST_DEPS,
5757)
5758
5759xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005760 name = "f32_vdiv_relu_test",
5761 srcs = [
5762 "test/f32-vdiv-relu.cc",
5763 "test/vbinary-microkernel-tester.h",
5764 ] + MICROKERNEL_TEST_HDRS,
5765 deps = MICROKERNEL_TEST_DEPS,
5766)
5767
5768xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005769 name = "f32_vdivc_test",
5770 srcs = [
5771 "test/f32-vdivc.cc",
5772 "test/vbinaryc-microkernel-tester.h",
5773 ] + MICROKERNEL_TEST_HDRS,
5774 deps = MICROKERNEL_TEST_DEPS,
5775)
5776
5777xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005778 name = "f32_vdivc_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08005779 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005780 "test/f32-vdivc-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08005781 "test/vbinaryc-microkernel-tester.h",
5782 ] + MICROKERNEL_TEST_HDRS,
5783 deps = MICROKERNEL_TEST_DEPS,
5784)
5785
5786xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005787 name = "f32_vdivc_relu_test",
5788 srcs = [
5789 "test/f32-vdivc-relu.cc",
5790 "test/vbinaryc-microkernel-tester.h",
5791 ] + MICROKERNEL_TEST_HDRS,
5792 deps = MICROKERNEL_TEST_DEPS,
5793)
5794
5795xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005796 name = "f32_vrdivc_test",
5797 srcs = [
5798 "test/f32-vrdivc.cc",
5799 "test/vbinaryc-microkernel-tester.h",
5800 ] + MICROKERNEL_TEST_HDRS,
5801 deps = MICROKERNEL_TEST_DEPS,
5802)
5803
5804xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005805 name = "f32_vrdivc_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08005806 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005807 "test/f32-vrdivc-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08005808 "test/vbinaryc-microkernel-tester.h",
5809 ] + MICROKERNEL_TEST_HDRS,
5810 deps = MICROKERNEL_TEST_DEPS,
5811)
5812
5813xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005814 name = "f32_vrdivc_relu_test",
5815 srcs = [
5816 "test/f32-vrdivc-relu.cc",
5817 "test/vbinaryc-microkernel-tester.h",
5818 ] + MICROKERNEL_TEST_HDRS,
5819 deps = MICROKERNEL_TEST_DEPS,
5820)
5821
5822xnnpack_unit_test(
Marat Dukhan403b7d42019-12-05 12:49:11 -08005823 name = "f32_vmax_test",
5824 srcs = [
5825 "test/f32-vmax.cc",
5826 "test/vbinary-microkernel-tester.h",
5827 ] + MICROKERNEL_TEST_HDRS,
5828 deps = MICROKERNEL_TEST_DEPS,
5829)
5830
5831xnnpack_unit_test(
5832 name = "f32_vmaxc_test",
5833 srcs = [
5834 "test/f32-vmaxc.cc",
5835 "test/vbinaryc-microkernel-tester.h",
5836 ] + MICROKERNEL_TEST_HDRS,
5837 deps = MICROKERNEL_TEST_DEPS,
5838)
5839
5840xnnpack_unit_test(
5841 name = "f32_vmin_test",
5842 srcs = [
5843 "test/f32-vmin.cc",
5844 "test/vbinary-microkernel-tester.h",
5845 ] + MICROKERNEL_TEST_HDRS,
5846 deps = MICROKERNEL_TEST_DEPS,
5847)
5848
5849xnnpack_unit_test(
5850 name = "f32_vminc_test",
5851 srcs = [
5852 "test/f32-vminc.cc",
5853 "test/vbinaryc-microkernel-tester.h",
5854 ] + MICROKERNEL_TEST_HDRS,
5855 deps = MICROKERNEL_TEST_DEPS,
5856)
5857
5858xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005859 name = "f32_vmul_test",
5860 srcs = [
5861 "test/f32-vmul.cc",
5862 "test/vbinary-microkernel-tester.h",
5863 ] + MICROKERNEL_TEST_HDRS,
5864 deps = MICROKERNEL_TEST_DEPS,
5865)
5866
5867xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005868 name = "f32_vmul_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005869 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005870 "test/f32-vmul-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005871 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005872 ] + MICROKERNEL_TEST_HDRS,
5873 deps = MICROKERNEL_TEST_DEPS,
5874)
5875
5876xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005877 name = "f32_vmul_relu_test",
5878 srcs = [
5879 "test/f32-vmul-relu.cc",
5880 "test/vbinary-microkernel-tester.h",
5881 ] + MICROKERNEL_TEST_HDRS,
5882 deps = MICROKERNEL_TEST_DEPS,
5883)
5884
5885xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005886 name = "f32_vmulc_test",
5887 srcs = [
5888 "test/f32-vmulc.cc",
5889 "test/vbinaryc-microkernel-tester.h",
5890 ] + MICROKERNEL_TEST_HDRS,
5891 deps = MICROKERNEL_TEST_DEPS,
5892)
5893
5894xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005895 name = "f32_vmulc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005896 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005897 "test/f32-vmulc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005898 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005899 ] + MICROKERNEL_TEST_HDRS,
5900 deps = MICROKERNEL_TEST_DEPS,
5901)
5902
5903xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005904 name = "f32_vmulc_relu_test",
5905 srcs = [
5906 "test/f32-vmulc-relu.cc",
5907 "test/vbinaryc-microkernel-tester.h",
5908 ] + MICROKERNEL_TEST_HDRS,
5909 deps = MICROKERNEL_TEST_DEPS,
5910)
5911
5912xnnpack_unit_test(
Frank Barchard2a1049c2020-06-03 02:31:27 -07005913 name = "f16_vmulcaddc_minmax_test",
5914 srcs = [
5915 "test/f16-vmulcaddc-minmax.cc",
5916 "test/vmulcaddc-microkernel-tester.h",
5917 "src/xnnpack/AlignedAllocator.h",
5918 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005919 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Frank Barchard2a1049c2020-06-03 02:31:27 -07005920)
5921
5922xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07005923 name = "f32_vmulcaddc_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005924 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07005925 "test/f32-vmulcaddc-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005926 "test/vmulcaddc-microkernel-tester.h",
5927 "src/xnnpack/AlignedAllocator.h",
5928 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005929 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005930)
5931
5932xnnpack_unit_test(
Marat Dukhan8cc7efe2020-06-10 16:24:27 -07005933 name = "f32_vlrelu_test",
5934 srcs = [
5935 "test/f32-vlrelu.cc",
5936 "test/vunary-microkernel-tester.h",
5937 ] + MICROKERNEL_TEST_HDRS,
5938 deps = MICROKERNEL_TEST_DEPS,
5939)
5940
5941xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07005942 name = "f32_vneg_test",
5943 srcs = [
5944 "test/f32-vneg.cc",
5945 "test/vunary-microkernel-tester.h",
5946 ] + MICROKERNEL_TEST_HDRS,
5947 deps = MICROKERNEL_TEST_DEPS,
5948)
5949
5950xnnpack_unit_test(
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07005951 name = "f32_vrndne_test",
5952 srcs = [
5953 "test/f32-vrndne.cc",
5954 "test/vunary-microkernel-tester.h",
5955 ] + MICROKERNEL_TEST_HDRS,
5956 deps = MICROKERNEL_TEST_DEPS,
5957)
5958
5959xnnpack_unit_test(
5960 name = "f32_vrndz_test",
5961 srcs = [
5962 "test/f32-vrndz.cc",
5963 "test/vunary-microkernel-tester.h",
5964 ] + MICROKERNEL_TEST_HDRS,
5965 deps = MICROKERNEL_TEST_DEPS,
5966)
5967
5968xnnpack_unit_test(
5969 name = "f32_vrndu_test",
5970 srcs = [
5971 "test/f32-vrndu.cc",
5972 "test/vunary-microkernel-tester.h",
5973 ] + MICROKERNEL_TEST_HDRS,
5974 deps = MICROKERNEL_TEST_DEPS,
5975)
5976
5977xnnpack_unit_test(
5978 name = "f32_vrndd_test",
5979 srcs = [
5980 "test/f32-vrndd.cc",
5981 "test/vunary-microkernel-tester.h",
5982 ] + MICROKERNEL_TEST_HDRS,
5983 deps = MICROKERNEL_TEST_DEPS,
5984)
5985
5986xnnpack_unit_test(
Marat Dukhan05ac8e32019-10-21 15:39:33 -07005987 name = "f32_vscale_test",
5988 srcs = [
5989 "test/f32-vscale.cc",
5990 "test/vscale-microkernel-tester.h",
5991 ] + MICROKERNEL_TEST_HDRS,
5992 deps = MICROKERNEL_TEST_DEPS,
5993)
5994
5995xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07005996 name = "f32_vscaleexpminusmax_test",
5997 srcs = [
5998 "test/f32-vscaleexpminusmax.cc",
5999 "test/vscaleexpminusmax-microkernel-tester.h",
6000 ] + MICROKERNEL_TEST_HDRS,
6001 deps = MICROKERNEL_TEST_DEPS,
6002)
6003
6004xnnpack_unit_test(
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07006005 name = "f32_vscaleextexp_test",
6006 srcs = [
6007 "test/f32-vscaleextexp.cc",
6008 "test/vscaleextexp-microkernel-tester.h",
6009 ] + MICROKERNEL_TEST_HDRS,
6010 deps = MICROKERNEL_TEST_DEPS,
6011)
6012
6013xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07006014 name = "f32_vsqr_test",
6015 srcs = [
6016 "test/f32-vsqr.cc",
6017 "test/vunary-microkernel-tester.h",
6018 ] + MICROKERNEL_TEST_HDRS,
6019 deps = MICROKERNEL_TEST_DEPS,
6020)
6021
6022xnnpack_unit_test(
Marat Dukhan13bafb02020-06-05 00:43:11 -07006023 name = "f32_vsqrdiff_test",
6024 srcs = [
6025 "test/f32-vsqrdiff.cc",
6026 "test/vbinary-microkernel-tester.h",
6027 ] + MICROKERNEL_TEST_HDRS,
6028 deps = MICROKERNEL_TEST_DEPS,
6029)
6030
6031xnnpack_unit_test(
6032 name = "f32_vsqrdiffc_test",
6033 srcs = [
6034 "test/f32-vsqrdiffc.cc",
6035 "test/vbinaryc-microkernel-tester.h",
6036 ] + MICROKERNEL_TEST_HDRS,
6037 deps = MICROKERNEL_TEST_DEPS,
6038)
6039
6040xnnpack_unit_test(
Marat Dukhanf4db2f32020-06-30 10:55:30 -07006041 name = "f32_vsqrt_test",
6042 srcs = [
6043 "test/f32-vsqrt.cc",
6044 "test/vunary-microkernel-tester.h",
6045 ] + MICROKERNEL_TEST_HDRS,
6046 deps = MICROKERNEL_TEST_DEPS,
6047)
6048
6049xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07006050 name = "f32_vsub_test",
6051 srcs = [
6052 "test/f32-vsub.cc",
6053 "test/vbinary-microkernel-tester.h",
6054 ] + MICROKERNEL_TEST_HDRS,
6055 deps = MICROKERNEL_TEST_DEPS,
6056)
6057
6058xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006059 name = "f32_vsub_minmax_test",
Marat Dukhan97579532019-10-18 16:40:39 -07006060 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006061 "test/f32-vsub-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08006062 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08006063 ] + MICROKERNEL_TEST_HDRS,
6064 deps = MICROKERNEL_TEST_DEPS,
6065)
6066
6067xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07006068 name = "f32_vsub_relu_test",
6069 srcs = [
6070 "test/f32-vsub-relu.cc",
6071 "test/vbinary-microkernel-tester.h",
6072 ] + MICROKERNEL_TEST_HDRS,
6073 deps = MICROKERNEL_TEST_DEPS,
6074)
6075
6076xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07006077 name = "f32_vsubc_test",
6078 srcs = [
6079 "test/f32-vsubc.cc",
6080 "test/vbinaryc-microkernel-tester.h",
6081 ] + MICROKERNEL_TEST_HDRS,
6082 deps = MICROKERNEL_TEST_DEPS,
6083)
6084
6085xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006086 name = "f32_vsubc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08006087 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006088 "test/f32-vsubc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08006089 "test/vbinaryc-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08006090 ] + MICROKERNEL_TEST_HDRS,
6091 deps = MICROKERNEL_TEST_DEPS,
6092)
6093
6094xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07006095 name = "f32_vsubc_relu_test",
6096 srcs = [
6097 "test/f32-vsubc-relu.cc",
6098 "test/vbinaryc-microkernel-tester.h",
6099 ] + MICROKERNEL_TEST_HDRS,
6100 deps = MICROKERNEL_TEST_DEPS,
6101)
6102
6103xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07006104 name = "f32_vrsubc_test",
6105 srcs = [
6106 "test/f32-vrsubc.cc",
6107 "test/vbinaryc-microkernel-tester.h",
6108 ] + MICROKERNEL_TEST_HDRS,
6109 deps = MICROKERNEL_TEST_DEPS,
6110)
6111
6112xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006113 name = "f32_vrsubc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08006114 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006115 "test/f32-vrsubc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08006116 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan97579532019-10-18 16:40:39 -07006117 ] + MICROKERNEL_TEST_HDRS,
6118 deps = MICROKERNEL_TEST_DEPS,
6119)
6120
6121xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07006122 name = "f32_vrsubc_relu_test",
6123 srcs = [
6124 "test/f32-vrsubc-relu.cc",
6125 "test/vbinaryc-microkernel-tester.h",
6126 ] + MICROKERNEL_TEST_HDRS,
6127 deps = MICROKERNEL_TEST_DEPS,
6128)
6129
6130xnnpack_unit_test(
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07006131 name = "qs8_dwconv_minmax_test",
6132 srcs = [
6133 "test/qs8-dwconv-minmax.cc",
6134 "test/dwconv-microkernel-tester.h",
6135 "src/xnnpack/AlignedAllocator.h",
6136 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
6137 deps = MICROKERNEL_TEST_DEPS + [":packing"],
6138)
6139
6140xnnpack_unit_test(
Marat Dukhan4ed53f42020-08-06 01:12:55 -07006141 name = "qs8_gavgpool_minmax_test",
6142 srcs = [
6143 "test/qs8-gavgpool-minmax.cc",
6144 "test/gavgpool-microkernel-tester.h",
6145 "src/xnnpack/AlignedAllocator.h",
6146 ] + MICROKERNEL_TEST_HDRS,
6147 deps = MICROKERNEL_TEST_DEPS,
6148)
6149
6150xnnpack_unit_test(
Marat Dukhan595e1702020-07-31 10:12:52 -07006151 name = "qs8_gemm_minmax_test",
6152 srcs = [
6153 "test/qs8-gemm-minmax.cc",
6154 "test/gemm-microkernel-tester.h",
6155 "src/xnnpack/AlignedAllocator.h",
6156 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
6157 deps = MICROKERNEL_TEST_DEPS + [":packing"],
6158)
6159
6160xnnpack_unit_test(
Marat Dukhanf9480682020-07-31 14:50:24 -07006161 name = "qs8_igemm_minmax_test",
6162 srcs = [
6163 "test/qs8-igemm-minmax.cc",
6164 "test/gemm-microkernel-tester.h",
6165 "src/xnnpack/AlignedAllocator.h",
6166 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
6167 deps = MICROKERNEL_TEST_DEPS + [":packing"],
6168)
6169
6170xnnpack_unit_test(
6171 name = "qs8_requantization_test",
6172 srcs = [
6173 "src/xnnpack/requantization-stubs.h",
6174 "test/qs8-requantization.cc",
6175 "test/requantization-tester.h",
6176 ] + MICROKERNEL_TEST_HDRS,
6177 deps = MICROKERNEL_TEST_DEPS,
6178)
6179
6180xnnpack_unit_test(
Marat Dukhand9f3ad42020-08-10 12:30:58 -07006181 name = "qs8_vadd_minmax_test",
6182 srcs = [
6183 "test/qs8-vadd-minmax.cc",
6184 "test/vadd-microkernel-tester.h",
6185 ] + MICROKERNEL_TEST_HDRS,
6186 deps = MICROKERNEL_TEST_DEPS,
6187)
6188
6189xnnpack_unit_test(
Marat Dukhan0270d9f2020-08-11 00:56:46 -07006190 name = "qs8_vaddc_minmax_test",
6191 srcs = [
6192 "test/qs8-vaddc-minmax.cc",
6193 "test/vaddc-microkernel-tester.h",
6194 ] + MICROKERNEL_TEST_HDRS,
6195 deps = MICROKERNEL_TEST_DEPS,
6196)
6197
6198xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07006199 name = "qu8_avgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006200 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07006201 "test/qu8-avgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006202 "test/avgpool-microkernel-tester.h",
6203 "src/xnnpack/AlignedAllocator.h",
6204 ] + MICROKERNEL_TEST_HDRS,
6205 deps = MICROKERNEL_TEST_DEPS,
6206)
6207
6208xnnpack_unit_test(
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07006209 name = "qu8_dwconv_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006210 srcs = [
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07006211 "test/qu8-dwconv-minmax.cc",
6212 "test/dwconv-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006213 "src/xnnpack/AlignedAllocator.h",
6214 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07006215 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07006216)
6217
6218xnnpack_unit_test(
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07006219 name = "qu8_igemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006220 srcs = [
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07006221 "test/qu8-igemm-minmax.cc",
6222 "test/gemm-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006223 "src/xnnpack/AlignedAllocator.h",
6224 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07006225 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07006226)
6227
6228xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07006229 name = "qu8_gavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006230 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07006231 "test/qu8-gavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006232 "test/gavgpool-microkernel-tester.h",
6233 "src/xnnpack/AlignedAllocator.h",
6234 ] + MICROKERNEL_TEST_HDRS,
6235 deps = MICROKERNEL_TEST_DEPS,
6236)
6237
6238xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07006239 name = "qu8_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006240 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07006241 "test/qu8-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006242 "test/gemm-microkernel-tester.h",
6243 "src/xnnpack/AlignedAllocator.h",
6244 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07006245 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07006246)
6247
6248xnnpack_unit_test(
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07006249 name = "qu8_requantization_test",
6250 srcs = [
6251 "src/xnnpack/requantization-stubs.h",
6252 "test/qu8-requantization.cc",
6253 "test/requantization-tester.h",
6254 ] + MICROKERNEL_TEST_HDRS,
6255 deps = MICROKERNEL_TEST_DEPS,
6256)
6257
6258xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07006259 name = "qu8_vadd_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006260 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07006261 "test/qu8-vadd-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006262 "test/vadd-microkernel-tester.h",
6263 ] + MICROKERNEL_TEST_HDRS,
6264 deps = MICROKERNEL_TEST_DEPS,
6265)
6266
6267xnnpack_unit_test(
6268 name = "u8_clamp_test",
6269 srcs = [
6270 "test/u8-clamp.cc",
6271 "test/clamp-microkernel-tester.h",
6272 ] + MICROKERNEL_TEST_HDRS,
6273 deps = MICROKERNEL_TEST_DEPS,
6274)
6275
6276xnnpack_unit_test(
6277 name = "u8_lut32norm_test",
6278 srcs = [
6279 "test/u8-lut32norm.cc",
6280 "test/lut-norm-microkernel-tester.h",
6281 ] + MICROKERNEL_TEST_HDRS,
6282 deps = MICROKERNEL_TEST_DEPS,
6283)
6284
6285xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07006286 name = "u8_maxpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006287 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07006288 "test/u8-maxpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006289 "test/maxpool-microkernel-tester.h",
6290 ] + MICROKERNEL_TEST_HDRS,
6291 deps = MICROKERNEL_TEST_DEPS,
6292)
6293
6294xnnpack_unit_test(
6295 name = "u8_rmax_test",
6296 srcs = [
6297 "test/u8-rmax.cc",
6298 "test/rmax-microkernel-tester.h",
6299 ] + MICROKERNEL_TEST_HDRS,
6300 deps = MICROKERNEL_TEST_DEPS,
6301)
6302
6303xnnpack_unit_test(
Marat Dukhanad71b9a2020-11-20 00:01:51 -08006304 name = "x32_depthtospace2d_chw2hwc_test",
Yury Kartynnike7841862020-11-04 18:22:18 -08006305 srcs = [
Marat Dukhanad71b9a2020-11-20 00:01:51 -08006306 "test/x32-depthtospace2d-chw2hwc.cc",
6307 "test/depthtospace-microkernel-tester.h",
Yury Kartynnike7841862020-11-04 18:22:18 -08006308 ] + MICROKERNEL_TEST_HDRS,
6309 deps = MICROKERNEL_TEST_DEPS,
6310)
6311
6312xnnpack_unit_test(
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -07006313 name = "x32_fill_test",
6314 srcs = [
6315 "test/x32-fill.cc",
6316 "test/fill-microkernel-tester.h",
6317 ] + MICROKERNEL_TEST_HDRS,
6318 deps = MICROKERNEL_TEST_DEPS,
6319)
6320
6321xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07006322 name = "x32_packx_test",
6323 srcs = [
6324 "test/x32-packx.cc",
6325 "test/pack-microkernel-tester.h",
6326 "src/xnnpack/AlignedAllocator.h",
6327 ] + MICROKERNEL_TEST_HDRS,
6328 deps = MICROKERNEL_TEST_DEPS,
6329)
6330
6331xnnpack_unit_test(
6332 name = "x32_pad_test",
6333 srcs = [
6334 "test/x32-pad.cc",
6335 "test/pad-microkernel-tester.h",
6336 ] + MICROKERNEL_TEST_HDRS,
6337 deps = MICROKERNEL_TEST_DEPS,
6338)
6339
6340xnnpack_unit_test(
6341 name = "x32_unpool_test",
6342 srcs = [
6343 "test/x32-unpool.cc",
6344 "test/unpool-microkernel-tester.h",
6345 ] + MICROKERNEL_TEST_HDRS,
6346 deps = MICROKERNEL_TEST_DEPS,
6347)
6348
6349xnnpack_unit_test(
6350 name = "x32_zip_test",
6351 srcs = [
6352 "test/x32-zip.cc",
6353 "test/zip-microkernel-tester.h",
6354 ] + MICROKERNEL_TEST_HDRS,
6355 deps = MICROKERNEL_TEST_DEPS,
6356)
6357
6358xnnpack_unit_test(
6359 name = "x8_lut_test",
6360 srcs = [
6361 "test/x8-lut.cc",
6362 "test/lut-microkernel-tester.h",
6363 ] + MICROKERNEL_TEST_HDRS,
6364 deps = MICROKERNEL_TEST_DEPS,
6365)
6366
6367xnnpack_unit_test(
6368 name = "x8_zip_test",
6369 srcs = [
6370 "test/x8-zip.cc",
6371 "test/zip-microkernel-tester.h",
6372 ] + MICROKERNEL_TEST_HDRS,
6373 deps = MICROKERNEL_TEST_DEPS,
6374)
6375
Marat Dukhan20c3b922020-03-10 03:45:06 -07006376########################## Size tests for the library #########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07006377
6378xnnpack_binary(
Marat Dukhan20c3b922020-03-10 03:45:06 -07006379 name = "operator_size_test",
6380 srcs = ["test/operator-size.c"],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07006381 deps = [":xnnpack_operators_nhwc_f32"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07006382)
6383
Marat Dukhan20c3b922020-03-10 03:45:06 -07006384xnnpack_binary(
6385 name = "subgraph_size_test",
6386 srcs = ["test/subgraph-size.c"],
6387 deps = [":XNNPACK"],
6388)
6389
6390########################### Unit tests for operators ##########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07006391
6392xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07006393 name = "abs_nc_test",
6394 srcs = [
6395 "test/abs-nc.cc",
6396 "test/abs-operator-tester.h",
6397 ],
6398 deps = OPERATOR_TEST_DEPS,
6399)
6400
6401xnnpack_unit_test(
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08006402 name = "add_nd_test",
6403 srcs = [
6404 "test/add-nd.cc",
6405 "test/binary-elementwise-operator-tester.h",
6406 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006407 deps = OPERATOR_TEST_DEPS,
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08006408)
6409
6410xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006411 name = "argmax_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006412 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006413 "test/argmax-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006414 "test/argmax-pooling-operator-tester.h",
6415 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07006416 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006417)
6418
6419xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006420 name = "average_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006421 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006422 "test/average-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006423 "test/average-pooling-operator-tester.h",
6424 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07006425 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006426)
6427
6428xnnpack_unit_test(
Marat Dukhan64e52512020-06-09 13:41:16 -07006429 name = "bankers_rounding_nc_test",
6430 srcs = [
6431 "test/bankers-rounding-nc.cc",
6432 "test/bankers-rounding-operator-tester.h",
6433 ],
6434 deps = OPERATOR_TEST_DEPS,
6435)
6436
6437xnnpack_unit_test(
6438 name = "ceiling_nc_test",
6439 srcs = [
6440 "test/ceiling-nc.cc",
6441 "test/ceiling-operator-tester.h",
6442 ],
6443 deps = OPERATOR_TEST_DEPS,
6444)
6445
6446xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006447 name = "channel_shuffle_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006448 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006449 "test/channel-shuffle-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006450 "test/channel-shuffle-operator-tester.h",
6451 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006452 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006453)
6454
6455xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006456 name = "clamp_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006457 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006458 "test/clamp-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006459 "test/clamp-operator-tester.h",
6460 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006461 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006462)
6463
6464xnnpack_unit_test(
Marat Dukhan065b11e2020-05-22 09:49:41 -07006465 name = "constant_pad_nd_test",
6466 srcs = [
6467 "test/constant-pad-nd.cc",
6468 "test/constant-pad-operator-tester.h",
6469 ],
6470 deps = OPERATOR_TEST_DEPS,
6471)
6472
6473xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006474 name = "convolution_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006475 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006476 "test/convolution-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006477 "test/convolution-operator-tester.h",
6478 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006479 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006480)
6481
6482xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006483 name = "convolution_nchw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006484 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006485 "test/convolution-nchw.cc",
6486 "test/convolution-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006487 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006488 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006489)
6490
6491xnnpack_unit_test(
Marat Dukhan4e21b272020-06-04 18:45:01 -07006492 name = "copy_nc_test",
6493 srcs = [
6494 "test/copy-nc.cc",
6495 "test/copy-operator-tester.h",
6496 ],
6497 deps = OPERATOR_TEST_DEPS,
6498)
6499
6500xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006501 name = "deconvolution_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006502 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006503 "test/deconvolution-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006504 "test/deconvolution-operator-tester.h",
6505 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07006506 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006507)
6508
6509xnnpack_unit_test(
Marat Dukhan188d1042020-11-24 23:39:40 -08006510 name = "depth_to_space_nchw2nhwc_test",
Artsiom Ablavatski0f1dc182020-11-05 19:21:50 -08006511 srcs = [
Marat Dukhan188d1042020-11-24 23:39:40 -08006512 "test/depth-to-space-nchw2nhwc.cc",
Artsiom Ablavatski0f1dc182020-11-05 19:21:50 -08006513 "test/depth-to-space-operator-tester.h",
6514 ] + OPERATOR_TEST_PARAMS_HDRS,
6515 deps = OPERATOR_TEST_DEPS,
6516)
6517
6518xnnpack_unit_test(
Marat Dukhan0e521172020-11-25 13:10:04 -08006519 name = "depth_to_space_nhwc_test",
6520 srcs = [
6521 "test/depth-to-space-nhwc.cc",
6522 "test/depth-to-space-operator-tester.h",
6523 ] + OPERATOR_TEST_PARAMS_HDRS,
6524 deps = OPERATOR_TEST_DEPS,
6525)
6526
6527xnnpack_unit_test(
Marat Dukhan69180502019-12-06 15:00:31 -08006528 name = "divide_nd_test",
6529 srcs = [
6530 "test/binary-elementwise-operator-tester.h",
6531 "test/divide-nd.cc",
6532 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006533 deps = OPERATOR_TEST_DEPS,
Marat Dukhan69180502019-12-06 15:00:31 -08006534)
6535
6536xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006537 name = "fully_connected_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006538 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006539 "test/fully-connected-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006540 "test/fully-connected-operator-tester.h",
6541 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006542 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006543)
6544
6545xnnpack_unit_test(
Marat Dukhan64e52512020-06-09 13:41:16 -07006546 name = "floor_nc_test",
6547 srcs = [
6548 "test/floor-nc.cc",
6549 "test/floor-operator-tester.h",
6550 ],
6551 deps = OPERATOR_TEST_DEPS,
6552)
6553
6554xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006555 name = "global_average_pooling_nwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006556 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006557 "test/global-average-pooling-nwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006558 "test/global-average-pooling-operator-tester.h",
Marat Dukhanef61d022020-06-19 13:54:49 -07006559 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07006560 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006561)
6562
6563xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006564 name = "global_average_pooling_ncw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006565 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006566 "test/global-average-pooling-ncw.cc",
6567 "test/global-average-pooling-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006568 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006569 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006570)
6571
6572xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006573 name = "hardswish_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006574 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006575 "test/hardswish-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006576 "test/hardswish-operator-tester.h",
6577 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006578 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006579)
6580
6581xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006582 name = "leaky_relu_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006583 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006584 "test/leaky-relu-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006585 "test/leaky-relu-operator-tester.h",
6586 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006587 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006588)
6589
6590xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006591 name = "max_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006592 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006593 "test/max-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006594 "test/max-pooling-operator-tester.h",
6595 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07006596 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006597)
6598
6599xnnpack_unit_test(
Marat Dukhan79e7f842019-12-05 14:35:50 -08006600 name = "maximum_nd_test",
6601 srcs = [
6602 "test/binary-elementwise-operator-tester.h",
6603 "test/maximum-nd.cc",
6604 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006605 deps = OPERATOR_TEST_DEPS,
Marat Dukhan79e7f842019-12-05 14:35:50 -08006606)
6607
6608xnnpack_unit_test(
6609 name = "minimum_nd_test",
6610 srcs = [
6611 "test/binary-elementwise-operator-tester.h",
6612 "test/minimum-nd.cc",
6613 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006614 deps = OPERATOR_TEST_DEPS,
Marat Dukhan79e7f842019-12-05 14:35:50 -08006615)
6616
6617xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006618 name = "multiply_nd_test",
Marat Dukhanca2733c2019-11-15 23:21:17 -08006619 srcs = [
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08006620 "test/binary-elementwise-operator-tester.h",
Marat Dukhanefc47b82019-11-18 09:25:38 -08006621 "test/multiply-nd.cc",
Marat Dukhanca2733c2019-11-15 23:21:17 -08006622 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006623 deps = OPERATOR_TEST_DEPS,
Marat Dukhanca2733c2019-11-15 23:21:17 -08006624)
6625
6626xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07006627 name = "negate_nc_test",
6628 srcs = [
6629 "test/negate-nc.cc",
6630 "test/negate-operator-tester.h",
6631 ],
6632 deps = OPERATOR_TEST_DEPS,
6633)
6634
6635xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006636 name = "prelu_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006637 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006638 "test/prelu-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006639 "test/prelu-operator-tester.h",
6640 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07006641 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006642)
6643
6644xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006645 name = "resize_bilinear_nhwc_test",
Marat Dukhan69722492019-11-11 19:55:50 -08006646 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006647 "test/resize-bilinear-nhwc.cc",
Marat Dukhan69722492019-11-11 19:55:50 -08006648 "test/resize-bilinear-operator-tester.h",
6649 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07006650 deps = OPERATOR_TEST_DEPS,
Marat Dukhan69722492019-11-11 19:55:50 -08006651)
6652
6653xnnpack_unit_test(
Artsiom Ablavatski97918102020-10-27 15:52:59 -07006654 name = "resize_bilinear_nchw_test",
6655 srcs = [
6656 "test/resize-bilinear-nchw.cc",
6657 "test/resize-bilinear-operator-tester.h",
6658 ] + OPERATOR_TEST_PARAMS_HDRS,
6659 deps = OPERATOR_TEST_DEPS,
6660)
6661
6662xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006663 name = "sigmoid_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006664 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006665 "test/sigmoid-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006666 "test/sigmoid-operator-tester.h",
6667 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006668 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006669)
6670
6671xnnpack_unit_test(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08006672 name = "softmax_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006673 srcs = [
Marat Dukhanfd8e6892020-01-27 15:25:25 -08006674 "test/softmax-nc.cc",
6675 "test/softmax-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006676 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006677 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006678)
6679
6680xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07006681 name = "square_nc_test",
6682 srcs = [
6683 "test/square-nc.cc",
6684 "test/square-operator-tester.h",
6685 ],
6686 deps = OPERATOR_TEST_DEPS,
6687)
6688
6689xnnpack_unit_test(
Marat Dukhan6804bbd2020-06-30 19:26:11 -07006690 name = "square_root_nc_test",
6691 srcs = [
6692 "test/square-root-nc.cc",
6693 "test/square-root-operator-tester.h",
6694 ],
6695 deps = OPERATOR_TEST_DEPS,
6696)
6697
6698xnnpack_unit_test(
Marat Dukhanf7399262020-06-05 10:58:44 -07006699 name = "squared_difference_nd_test",
6700 srcs = [
6701 "test/binary-elementwise-operator-tester.h",
6702 "test/squared-difference-nd.cc",
6703 ],
6704 deps = OPERATOR_TEST_DEPS,
6705)
6706
6707xnnpack_unit_test(
Marat Dukhan05f3f6d2019-12-03 15:13:53 -08006708 name = "subtract_nd_test",
6709 srcs = [
6710 "test/binary-elementwise-operator-tester.h",
6711 "test/subtract-nd.cc",
6712 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006713 deps = OPERATOR_TEST_DEPS,
Marat Dukhan05f3f6d2019-12-03 15:13:53 -08006714)
6715
6716xnnpack_unit_test(
Marat Dukhan64e52512020-06-09 13:41:16 -07006717 name = "truncation_nc_test",
6718 srcs = [
6719 "test/truncation-nc.cc",
6720 "test/truncation-operator-tester.h",
6721 ],
6722 deps = OPERATOR_TEST_DEPS,
6723)
6724
6725xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006726 name = "unpooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006727 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006728 "test/unpooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006729 "test/unpooling-operator-tester.h",
6730 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006731 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006732)
6733
Chao Mei6ddfc602020-05-13 22:29:36 -07006734############################### Misc unit tests ###############################
6735
6736xnnpack_unit_test(
6737 name = "memory_planner_test",
6738 srcs = [
6739 "test/memory-planner-test.cc",
6740 ],
6741 deps = [
6742 ":XNNPACK",
6743 ":memory_planner",
6744 ],
6745)
6746
XNNPACK Teamab8c4c82020-10-09 08:05:51 -07006747xnnpack_unit_test(
6748 name = "subgraph_nchw_test",
6749 srcs = [
6750 "src/xnnpack/subgraph.h",
6751 "test/subgraph-nchw.cc",
6752 "test/subgraph-tester.h",
6753 ],
6754 deps = [
6755 ":XNNPACK",
6756 ],
6757)
6758
Marat Dukhan08c4a432019-10-03 09:29:21 -07006759############################# Build configurations #############################
6760
Marat Dukhanb8642352019-10-30 15:43:02 -07006761# Enables usage of assembly kernels.
Marat Dukhan08c4a432019-10-03 09:29:21 -07006762config_setting(
Marat Dukhanb8642352019-10-30 15:43:02 -07006763 name = "xnn_enable_assembly_explicit_true",
6764 define_values = {"xnn_enable_assembly": "true"},
6765)
6766
6767# Disables usage of assembly kernels.
6768config_setting(
6769 name = "xnn_enable_assembly_explicit_false",
6770 define_values = {"xnn_enable_assembly": "false"},
6771)
6772
Marat Dukhan9de90e02020-06-18 16:04:12 -07006773# Enables usage of sparse inference.
6774config_setting(
6775 name = "xnn_enable_sparse_explicit_true",
6776 define_values = {"xnn_enable_sparse": "true"},
6777)
6778
6779# Disables usage of sparse inference.
6780config_setting(
6781 name = "xnn_enable_sparse_explicit_false",
6782 define_values = {"xnn_enable_sparse": "false"},
6783)
6784
Marat Dukhan05702cf2020-03-26 15:41:33 -07006785# Disables usage of HMP-aware optimizations.
6786config_setting(
6787 name = "xnn_enable_hmp_explicit_false",
6788 define_values = {"xnn_enable_hmp": "false"},
6789)
6790
Chao Mei6ddfc602020-05-13 22:29:36 -07006791# Enable usage of optimized memory allocation
6792config_setting(
6793 name = "xnn_enable_memopt_explicit_true",
6794 define_values = {"xnnpack_enable_memopt": "true"},
6795)
6796
6797# Disable usage of optimized memory allocation
6798config_setting(
6799 name = "xnn_enable_memopt_explicit_false",
6800 define_values = {"xnnpack_enable_memopt": "false"},
6801)
6802
Marat Dukhanb8642352019-10-30 15:43:02 -07006803# Builds with -c dbg
6804config_setting(
6805 name = "debug_build",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006806 values = {
Marat Dukhanb8642352019-10-30 15:43:02 -07006807 "compilation_mode": "dbg",
6808 },
6809)
6810
6811# Builds with -c opt
6812config_setting(
6813 name = "optimized_build",
6814 values = {
6815 "compilation_mode": "opt",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006816 },
6817)
6818
6819config_setting(
Marat Dukhanb8642352019-10-30 15:43:02 -07006820 name = "linux_k8",
6821 values = {"cpu": "k8"},
6822)
6823
6824config_setting(
Marat Dukhan582094e2020-04-30 17:21:25 -07006825 name = "linux_arm",
6826 values = {"cpu": "arm"},
Marat Dukhan4e45e662019-10-03 15:40:24 -07006827)
6828
6829config_setting(
Marat Dukhanf0bd4de2020-06-15 15:53:19 -07006830 name = "linux_armeabi",
6831 values = {"cpu": "armeabi"},
6832)
6833
6834config_setting(
Terry Heo68eef3f2020-04-13 22:53:52 -07006835 name = "linux_armhf",
6836 values = {"cpu": "armhf"},
6837)
6838
6839config_setting(
Marat Dukhana720e932020-06-10 13:01:11 -07006840 name = "linux_armv7a",
6841 values = {"cpu": "armv7a"},
6842)
6843
6844config_setting(
Marat Dukhan582094e2020-04-30 17:21:25 -07006845 name = "linux_aarch64",
6846 values = {"cpu": "aarch64"},
6847)
6848
6849config_setting(
Marat Dukhan08c4a432019-10-03 09:29:21 -07006850 name = "android",
6851 values = {"crosstool_top": "//external:android/crosstool"},
6852)
6853
6854config_setting(
6855 name = "android_armv7",
6856 values = {
6857 "crosstool_top": "//external:android/crosstool",
6858 "cpu": "armeabi-v7a",
6859 },
6860)
6861
6862config_setting(
6863 name = "android_arm64",
6864 values = {
6865 "crosstool_top": "//external:android/crosstool",
6866 "cpu": "arm64-v8a",
6867 },
6868)
6869
6870config_setting(
6871 name = "android_x86",
6872 values = {
6873 "crosstool_top": "//external:android/crosstool",
6874 "cpu": "x86",
6875 },
6876)
6877
6878config_setting(
6879 name = "android_x86_64",
6880 values = {
6881 "crosstool_top": "//external:android/crosstool",
6882 "cpu": "x86_64",
6883 },
6884)
6885
6886config_setting(
Marat Dukhan10a38082020-04-17 03:58:35 -07006887 name = "windows_x86_64",
6888 values = {"cpu": "x64_windows"},
Marat Dukhan9fe932e2020-04-11 17:14:15 -07006889)
6890
6891config_setting(
Marat Dukhan10a38082020-04-17 03:58:35 -07006892 name = "windows_x86_64_clang",
6893 values = {
6894 "compiler": "clang-cl",
6895 "cpu": "x64_windows",
6896 },
6897)
6898
6899config_setting(
6900 name = "windows_x86_64_mingw",
6901 values = {
6902 "compiler": "mingw-gcc",
6903 "cpu": "x64_windows",
6904 },
6905)
6906
6907config_setting(
6908 name = "windows_x86_64_msys",
6909 values = {
6910 "compiler": "msys-gcc",
6911 "cpu": "x64_windows",
6912 },
Marat Dukhan9fe932e2020-04-11 17:14:15 -07006913)
6914
6915config_setting(
Marat Dukhan885ca242019-10-07 09:17:32 -07006916 name = "macos_x86_64",
6917 values = {
6918 "apple_platform_type": "macos",
6919 "cpu": "darwin",
6920 },
6921)
6922
6923config_setting(
Marat Dukhan08c4a432019-10-03 09:29:21 -07006924 name = "emscripten",
Marat Dukhan1a729ec2019-10-07 09:31:44 -07006925 values = {"crosstool_top": "//toolchain:emscripten"},
Marat Dukhan08c4a432019-10-03 09:29:21 -07006926)
6927
6928config_setting(
6929 name = "emscripten_wasm",
6930 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07006931 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006932 "cpu": "wasm",
6933 },
6934)
6935
6936config_setting(
6937 name = "emscripten_wasmsimd",
6938 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07006939 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006940 "cpu": "wasm",
Marat Dukhan81c62602020-05-29 13:22:49 -07006941 "copt": "-msimd128",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006942 },
6943)
6944
6945config_setting(
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006946 name = "ios_armv7",
6947 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006948 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006949 "cpu": "ios_armv7",
6950 },
6951)
6952
6953config_setting(
6954 name = "ios_arm64",
6955 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006956 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006957 "cpu": "ios_arm64",
6958 },
6959)
6960
6961config_setting(
6962 name = "ios_arm64e",
6963 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006964 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006965 "cpu": "ios_arm64e",
6966 },
6967)
6968
6969config_setting(
6970 name = "ios_x86",
6971 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006972 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006973 "cpu": "ios_i386",
6974 },
6975)
6976
6977config_setting(
6978 name = "ios_x86_64",
6979 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006980 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006981 "cpu": "ios_x86_64",
6982 },
6983)
6984
6985config_setting(
6986 name = "watchos_armv7k",
6987 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006988 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006989 "cpu": "watchos_armv7k",
6990 },
6991)
6992
6993config_setting(
6994 name = "watchos_arm64_32",
6995 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006996 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006997 "cpu": "watchos_arm64_32",
6998 },
6999)
7000
7001config_setting(
7002 name = "watchos_x86",
7003 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08007004 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08007005 "cpu": "watchos_i386",
7006 },
7007)
7008
7009config_setting(
7010 name = "watchos_x86_64",
7011 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08007012 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08007013 "cpu": "watchos_x86_64",
7014 },
7015)
7016
7017config_setting(
7018 name = "tvos_arm64",
7019 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08007020 "apple_platform_type": "tvos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08007021 "cpu": "tvos_arm64",
7022 },
7023)
7024
7025config_setting(
7026 name = "tvos_x86_64",
7027 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08007028 "apple_platform_type": "tvos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08007029 "cpu": "tvos_x86_64",
7030 },
7031)