blob: 3918df0419361811052ae7e8271ee6dfc85d0c04 [file] [log] [blame]
Frank Barchard1a953052020-11-16 18:44:58 -08001# Copyright 2020 Google LLC
Marat Dukhan08c4a432019-10-03 09:29:21 -07002#
3# This source code is licensed under the BSD-style license found in the
4# LICENSE file in the root directory of this source tree.
5#
6# Description:
7# XNNPACK - optimized floating-point neural network operators library
8
Marat Dukhan10a38082020-04-17 03:58:35 -07009load(":build_defs.bzl", "xnnpack_aggregate_library", "xnnpack_benchmark", "xnnpack_binary", "xnnpack_cc_library", "xnnpack_gcc_std_copts", "xnnpack_min_size_copts", "xnnpack_msvc_std_copts", "xnnpack_optional_armcl_copts", "xnnpack_optional_armcl_deps", "xnnpack_optional_dnnl_copts", "xnnpack_optional_dnnl_deps", "xnnpack_optional_gemmlowp_copts", "xnnpack_optional_gemmlowp_deps", "xnnpack_optional_ruy_copts", "xnnpack_optional_ruy_deps", "xnnpack_optional_tflite_copts", "xnnpack_optional_tflite_deps", "xnnpack_std_cxxopts", "xnnpack_unit_test", "xnnpack_visibility")
Marat Dukhan69c3f2c2019-11-06 12:30:01 -080010
Marat Dukhan08c4a432019-10-03 09:29:21 -070011licenses(["notice"])
12
13exports_files(["LICENSE"])
14
Marat Dukhan1b354632020-03-23 12:50:22 -070015OPERATOR_BENCHMARK_DEPS = [
16 ":XNNPACK",
17 ":bench_utils",
18 "@cpuinfo",
Frank Barchard0c849732020-06-12 13:31:32 -070019 "@FP16",
Marat Dukhan1b354632020-03-23 12:50:22 -070020 "@pthreadpool",
21]
22
Marat Dukhan08c4a432019-10-03 09:29:21 -070023MICROKERNEL_BENCHMARK_DEPS = [
24 ":ukernels",
25 ":bench_utils",
Frank Barchard7e955972019-10-11 10:34:25 -070026 ":enable_assembly",
Marat Dukhan08c4a432019-10-03 09:29:21 -070027 "@cpuinfo",
28 "@FP16",
29 "@pthreadpool",
30]
31
Marat Dukhan6adff4e2019-10-14 18:32:07 -070032ACCURACY_EVAL_DEPS = [
33 ":XNNPACK",
34 ":ukernels",
35 "@FP16",
36 "@pthreadpool",
37]
38
Marat Dukhan08c4a432019-10-03 09:29:21 -070039MICROKERNEL_TEST_DEPS = [
Marat Dukhan33fcf782020-05-24 14:27:15 -070040 ":ukernels_test_mode",
Frank Barchard7e955972019-10-11 10:34:25 -070041 ":enable_assembly",
Marat Dukhan08c4a432019-10-03 09:29:21 -070042 "@cpuinfo",
43 "@FP16",
44 "@pthreadpool",
45]
46
Marat Dukhan1b354632020-03-23 12:50:22 -070047OPERATOR_TEST_DEPS = [
Marat Dukhan33fcf782020-05-24 14:27:15 -070048 ":XNNPACK_test_mode",
Marat Dukhan1b354632020-03-23 12:50:22 -070049 "@pthreadpool",
50 "@FP16",
51]
52
Marat Dukhan08c4a432019-10-03 09:29:21 -070053OPERATOR_SRCS = [
Marat Dukhane8265432020-04-28 18:42:59 -070054 "src/operators/argmax-pooling-nhwc.c",
55 "src/operators/average-pooling-nhwc.c",
56 "src/operators/binary-elementwise-nd.c",
Marat Dukhane8265432020-04-28 18:42:59 -070057 "src/operators/channel-shuffle-nc.c",
Marat Dukhan065b11e2020-05-22 09:49:41 -070058 "src/operators/constant-pad-nd.c",
Marat Dukhane8265432020-04-28 18:42:59 -070059 "src/operators/convolution-nchw.c",
60 "src/operators/convolution-nhwc.c",
61 "src/operators/deconvolution-nhwc.c",
Marat Dukhan13b68f22020-11-12 11:55:19 -080062 "src/operators/depth-to-space-nchw2nhwc.c",
Marat Dukhan0e521172020-11-25 13:10:04 -080063 "src/operators/depth-to-space-nhwc.c",
Marat Dukhane8265432020-04-28 18:42:59 -070064 "src/operators/fully-connected-nc.c",
65 "src/operators/global-average-pooling-ncw.c",
66 "src/operators/global-average-pooling-nwc.c",
Marat Dukhane8265432020-04-28 18:42:59 -070067 "src/operators/leaky-relu-nc.c",
68 "src/operators/max-pooling-nhwc.c",
69 "src/operators/prelu-nc.c",
Artsiom Ablavatski97918102020-10-27 15:52:59 -070070 "src/operators/resize-bilinear-nchw.c",
Marat Dukhane8265432020-04-28 18:42:59 -070071 "src/operators/resize-bilinear-nhwc.c",
72 "src/operators/sigmoid-nc.c",
73 "src/operators/softmax-nc.c",
Marat Dukhanc3065f52020-06-04 13:33:32 -070074 "src/operators/unary-elementwise-nc.c",
Marat Dukhane8265432020-04-28 18:42:59 -070075 "src/operators/unpooling-nhwc.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -070076]
77
Marat Dukhan95e8b7a2020-06-03 12:46:26 -070078SUBGRAPH_SRCS = [
Marat Dukhan5fab4092020-06-10 01:28:28 -070079 "src/subgraph/abs.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -070080 "src/subgraph/add2.c",
81 "src/subgraph/argmax-pooling-2d.c",
82 "src/subgraph/average-pooling-2d.c",
Marat Dukhan5fab4092020-06-10 01:28:28 -070083 "src/subgraph/bankers-rounding.c",
84 "src/subgraph/ceiling.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -070085 "src/subgraph/clamp.c",
86 "src/subgraph/convolution-2d.c",
87 "src/subgraph/deconvolution-2d.c",
Artsiom Ablavatskibbe85062020-11-05 14:07:37 -080088 "src/subgraph/depth-to-space.c",
Frank Barchard9cef5ea2020-11-18 14:52:08 -080089 "src/subgraph/depthwise-convolution-2d.c",
Marat Dukhan9d3a4592020-06-05 16:52:42 -070090 "src/subgraph/divide.c",
Marat Dukhana1600202020-12-01 22:17:16 -080091 "src/subgraph/elu.c",
Marat Dukhan5fab4092020-06-10 01:28:28 -070092 "src/subgraph/floor.c",
Frank Barchard04336c12020-10-22 16:48:55 -070093 "src/subgraph/fully-connected.c",
Marat Dukhana059b7d2020-06-11 11:41:27 -070094 "src/subgraph/global-average-pooling-2d.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -070095 "src/subgraph/hardswish.c",
Marat Dukhan5bbebac2020-06-10 19:42:15 -070096 "src/subgraph/leaky-relu.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -070097 "src/subgraph/max-pooling-2d.c",
Marat Dukhan9d3a4592020-06-05 16:52:42 -070098 "src/subgraph/maximum2.c",
99 "src/subgraph/minimum2.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -0700100 "src/subgraph/multiply2.c",
Marat Dukhan5fab4092020-06-10 01:28:28 -0700101 "src/subgraph/negate.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -0700102 "src/subgraph/prelu.c",
103 "src/subgraph/sigmoid.c",
104 "src/subgraph/softmax.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700105 "src/subgraph/square-root.c",
106 "src/subgraph/square.c",
107 "src/subgraph/squared-difference.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -0700108 "src/subgraph/static-constant-pad.c",
Marat Dukhand27202d2020-07-09 23:43:40 -0700109 "src/subgraph/static-reshape.c",
Marat Dukhanaff24e22020-07-23 01:43:58 -0700110 "src/subgraph/static-resize-bilinear-2d.c",
Marat Dukhan9d3a4592020-06-05 16:52:42 -0700111 "src/subgraph/subtract.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -0700112 "src/subgraph/unpooling-2d.c",
113]
114
Marat Dukhan3a77ea72019-12-23 12:10:24 -0800115TABLE_SRCS = [
116 "src/tables/exp2-k-over-64.c",
117 "src/tables/exp2-k-over-2048.c",
Marat Dukhande390d42020-11-29 19:32:18 -0800118 "src/tables/exp2minus-k-over-4.c",
119 "src/tables/exp2minus-k-over-8.c",
Marat Dukhanc60742b2020-11-23 12:33:27 -0800120 "src/tables/exp2minus-k-over-16.c",
Marat Dukhan1f256fc2020-09-24 21:27:14 -0700121 "src/tables/exp2minus-k-over-64.c",
122 "src/tables/exp2minus-k-over-2048.c",
Marat Dukhan3a77ea72019-12-23 12:10:24 -0800123]
124
Marat Dukhan08c4a432019-10-03 09:29:21 -0700125SCALAR_UKERNELS = [
Marat Dukhan329da642019-11-19 21:44:39 -0800126 "src/f32-argmaxpool/4x-scalar-c1.c",
Marat Dukhan1e782c42019-11-21 17:02:40 -0800127 "src/f32-argmaxpool/9p8x-scalar-c1.c",
Marat Dukhan329da642019-11-19 21:44:39 -0800128 "src/f32-argmaxpool/9x-scalar-c1.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700129 "src/f32-avgpool/9p8x-minmax-scalar-c1.c",
130 "src/f32-avgpool/9x-minmax-scalar-c1.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -0700131 "src/f32-clamp/gen/scalar-x1.c",
132 "src/f32-clamp/gen/scalar-x2.c",
133 "src/f32-clamp/gen/scalar-x4.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700134 "src/f32-conv-hwc/3x3s2p0p1c3x4-scalar-1x1.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700135 "src/f32-conv-hwc/3x3s2p1c3x4-scalar-1x1.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -0700136 "src/f32-conv-hwc2chw/3x3s2p1c3x4-scalar-1x1.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700137 "src/f32-dwconv/gen/up1x4-minmax-scalar-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700138 "src/f32-dwconv/gen/up1x4-minmax-scalar.c",
139 "src/f32-dwconv/gen/up1x4-scalar-acc2.c",
140 "src/f32-dwconv/gen/up1x4-scalar.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700141 "src/f32-dwconv/gen/up1x9-minmax-scalar-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700142 "src/f32-dwconv/gen/up1x9-minmax-scalar.c",
143 "src/f32-dwconv/gen/up1x9-scalar-acc2.c",
144 "src/f32-dwconv/gen/up1x9-scalar.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700145 "src/f32-dwconv/gen/up1x25-minmax-scalar-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700146 "src/f32-dwconv/gen/up1x25-minmax-scalar.c",
147 "src/f32-dwconv/gen/up1x25-scalar-acc2.c",
148 "src/f32-dwconv/gen/up1x25-scalar.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700149 "src/f32-dwconv/gen/up2x4-minmax-scalar-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700150 "src/f32-dwconv/gen/up2x4-minmax-scalar.c",
151 "src/f32-dwconv/gen/up2x4-scalar-acc2.c",
152 "src/f32-dwconv/gen/up2x4-scalar.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700153 "src/f32-dwconv/gen/up2x9-minmax-scalar-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700154 "src/f32-dwconv/gen/up2x9-minmax-scalar.c",
155 "src/f32-dwconv/gen/up2x9-scalar-acc2.c",
156 "src/f32-dwconv/gen/up2x9-scalar.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700157 "src/f32-dwconv/gen/up2x25-minmax-scalar-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700158 "src/f32-dwconv/gen/up2x25-minmax-scalar.c",
159 "src/f32-dwconv/gen/up2x25-scalar-acc2.c",
160 "src/f32-dwconv/gen/up2x25-scalar.c",
Frank Barchard35db7d02020-10-26 13:37:34 -0700161 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-scalar-1x1-acc2.c",
162 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-scalar-1x1-acc3.c",
163 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-scalar-1x1-acc4.c",
Marat Dukhan91249d22020-10-24 12:02:51 -0700164 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-scalar-1x1.c",
Frank Barchard35db7d02020-10-26 13:37:34 -0700165 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-scalar-2x1-acc2.c",
Marat Dukhan91249d22020-10-24 12:02:51 -0700166 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-scalar-2x1.c",
167 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-scalar-3x1.c",
168 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-scalar-4x1.c",
169 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-scalar-5x1.c",
170 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-scalar-6x1.c",
Marat Dukhancf5b3c32020-10-25 19:21:10 -0700171 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-scalar-1x1-acc2.c",
172 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-scalar-1x1-acc3.c",
173 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-scalar-1x1-acc4.c",
Frank Barchard35db7d02020-10-26 13:37:34 -0700174 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-scalar-1x1.c",
Marat Dukhancf5b3c32020-10-25 19:21:10 -0700175 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-scalar-2x1-acc2.c",
Frank Barchard35db7d02020-10-26 13:37:34 -0700176 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-scalar-2x1.c",
177 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-scalar-3x1.c",
178 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-scalar-4x1.c",
Marat Dukhanc4efb002020-10-25 23:14:47 -0700179 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-scalar-1x1-acc2.c",
180 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-scalar-1x1-acc3.c",
181 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-scalar-1x1-acc4.c",
182 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-scalar-1x1-acc5.c",
Frank Barchard35db7d02020-10-26 13:37:34 -0700183 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-scalar-1x1.c",
Marat Dukhanc4efb002020-10-25 23:14:47 -0700184 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-scalar-2x1-acc2.c",
185 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-scalar-2x1-acc3.c",
Frank Barchard35db7d02020-10-26 13:37:34 -0700186 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-scalar-2x1.c",
Marat Dukhanc4efb002020-10-25 23:14:47 -0700187 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-scalar-3x1-acc2.c",
Frank Barchard35db7d02020-10-26 13:37:34 -0700188 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-scalar-3x1.c",
Marat Dukhan29c0c332020-10-28 22:11:00 -0700189 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-scalar-1x1-acc2.c",
190 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-scalar-1x1-acc3.c",
191 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-scalar-1x1-acc4.c",
192 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-scalar-1x1-acc5.c",
193 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-scalar-1x1.c",
194 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-scalar-2x1-acc2.c",
195 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-scalar-2x1-acc3.c",
196 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-scalar-2x1.c",
197 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-scalar-3x1-acc2.c",
198 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-scalar-3x1.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -0700199 "src/f32-gavgpool-cw/scalar-x1.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700200 "src/f32-gavgpool/7p7x-minmax-scalar-c1.c",
201 "src/f32-gavgpool/7x-minmax-scalar-c1.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700202 "src/f32-gemm/gen-inc/1x4inc-minmax-scalar.c",
203 "src/f32-gemm/gen-inc/2x4inc-minmax-scalar.c",
204 "src/f32-gemm/gen-inc/4x4inc-minmax-scalar.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700205 "src/f32-gemm/gen/1x4-minmax-scalar.c",
206 "src/f32-gemm/gen/1x4-relu-scalar.c",
207 "src/f32-gemm/gen/1x4-scalar.c",
208 "src/f32-gemm/gen/2x4-minmax-scalar.c",
209 "src/f32-gemm/gen/2x4-relu-scalar.c",
210 "src/f32-gemm/gen/2x4-scalar.c",
211 "src/f32-gemm/gen/4x2-minmax-scalar.c",
212 "src/f32-gemm/gen/4x2-relu-scalar.c",
213 "src/f32-gemm/gen/4x2-scalar.c",
214 "src/f32-gemm/gen/4x4-minmax-scalar.c",
215 "src/f32-gemm/gen/4x4-relu-scalar.c",
216 "src/f32-gemm/gen/4x4-scalar.c",
Marat Dukhanefaac272020-07-16 16:58:07 -0700217 "src/f32-hswish/gen/hswish-scalar-x1.c",
218 "src/f32-hswish/gen/hswish-scalar-x2.c",
219 "src/f32-hswish/gen/hswish-scalar-x4.c",
XNNPACK Team21432672020-10-19 19:58:48 -0700220 "src/f32-ibilinear-chw/gen/scalar-p1.c",
221 "src/f32-ibilinear-chw/gen/scalar-p2.c",
222 "src/f32-ibilinear-chw/gen/scalar-p4.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700223 "src/f32-ibilinear/gen/scalar-c1.c",
224 "src/f32-ibilinear/gen/scalar-c2.c",
225 "src/f32-ibilinear/gen/scalar-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700226 "src/f32-igemm/gen/1x4-minmax-scalar.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700227 "src/f32-igemm/gen/1x4-relu-scalar.c",
228 "src/f32-igemm/gen/1x4-scalar.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700229 "src/f32-igemm/gen/2x4-minmax-scalar.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700230 "src/f32-igemm/gen/2x4-relu-scalar.c",
231 "src/f32-igemm/gen/2x4-scalar.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700232 "src/f32-igemm/gen/4x2-minmax-scalar.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700233 "src/f32-igemm/gen/4x2-relu-scalar.c",
234 "src/f32-igemm/gen/4x2-scalar.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700235 "src/f32-igemm/gen/4x4-minmax-scalar.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700236 "src/f32-igemm/gen/4x4-relu-scalar.c",
237 "src/f32-igemm/gen/4x4-scalar.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700238 "src/f32-maxpool/9p8x-minmax-scalar-c1.c",
239 "src/f32-pavgpool/9p8x-minmax-scalar-c1.c",
240 "src/f32-pavgpool/9x-minmax-scalar-c1.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700241 "src/f32-ppmm/gen/2x4-minmax-scalar.c",
242 "src/f32-ppmm/gen/3x3-minmax-scalar.c",
243 "src/f32-ppmm/gen/4x2-minmax-scalar.c",
244 "src/f32-ppmm/gen/4x4-minmax-scalar.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800245 "src/f32-prelu/gen/scalar-2x1.c",
246 "src/f32-prelu/gen/scalar-2x4.c",
Marat Dukhanf46f6752020-01-21 11:03:49 -0800247 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x1.c",
Marat Dukhanf46f6752020-01-21 11:03:49 -0800248 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x2-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700249 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x2.c",
Marat Dukhanf46f6752020-01-21 11:03:49 -0800250 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x4-acc2.c",
251 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x4-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700252 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x4.c",
Marat Dukhanf46f6752020-01-21 11:03:49 -0800253 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x1.c",
Marat Dukhanf46f6752020-01-21 11:03:49 -0800254 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x2-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700255 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x2.c",
Marat Dukhanf46f6752020-01-21 11:03:49 -0800256 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x4-acc2.c",
257 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x4-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700258 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x4.c",
Frank Barchardfb158e22020-07-15 16:10:10 -0700259 "src/f32-relu/gen/scalar-x1.c",
260 "src/f32-relu/gen/scalar-x2.c",
261 "src/f32-relu/gen/scalar-x4.c",
Frank Barchard37297a62020-08-27 12:15:10 -0700262 "src/f32-relu/gen/scalar-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700263 "src/f32-rmax/scalar.c",
Marat Dukhan3a77ea72019-12-23 12:10:24 -0800264 "src/f32-sigmoid/gen/scalar-lut64-p2-div-x1.c",
265 "src/f32-sigmoid/gen/scalar-lut64-p2-div-x2.c",
266 "src/f32-sigmoid/gen/scalar-lut64-p2-div-x4.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700267 "src/f32-sigmoid/gen/scalar-lut2048-p1-div-x1.c",
268 "src/f32-sigmoid/gen/scalar-lut2048-p1-div-x2.c",
269 "src/f32-sigmoid/gen/scalar-lut2048-p1-div-x4.c",
Marat Dukhan3a77ea72019-12-23 12:10:24 -0800270 "src/f32-sigmoid/gen/scalar-p5-div-x1.c",
271 "src/f32-sigmoid/gen/scalar-p5-div-x2.c",
272 "src/f32-sigmoid/gen/scalar-p5-div-x4.c",
Marat Dukhan355ab432020-04-09 19:01:52 -0700273 "src/f32-spmm/gen/1x1-minmax-scalar-pipelined.c",
274 "src/f32-spmm/gen/1x1-minmax-scalar.c",
275 "src/f32-spmm/gen/2x1-minmax-scalar-pipelined.c",
276 "src/f32-spmm/gen/2x1-minmax-scalar.c",
277 "src/f32-spmm/gen/4x1-minmax-scalar-pipelined.c",
278 "src/f32-spmm/gen/4x1-minmax-scalar.c",
279 "src/f32-spmm/gen/8x1-minmax-scalar-pipelined.c",
280 "src/f32-spmm/gen/8x1-minmax-scalar.c",
281 "src/f32-spmm/gen/8x2-minmax-scalar.c",
282 "src/f32-spmm/gen/8x4-minmax-scalar.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700283 "src/f32-vbinary/gen/vadd-minmax-scalar-x1.c",
284 "src/f32-vbinary/gen/vadd-minmax-scalar-x2.c",
285 "src/f32-vbinary/gen/vadd-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700286 "src/f32-vbinary/gen/vadd-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700287 "src/f32-vbinary/gen/vadd-relu-scalar-x1.c",
288 "src/f32-vbinary/gen/vadd-relu-scalar-x2.c",
289 "src/f32-vbinary/gen/vadd-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700290 "src/f32-vbinary/gen/vadd-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700291 "src/f32-vbinary/gen/vadd-scalar-x1.c",
292 "src/f32-vbinary/gen/vadd-scalar-x2.c",
293 "src/f32-vbinary/gen/vadd-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700294 "src/f32-vbinary/gen/vadd-scalar-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700295 "src/f32-vbinary/gen/vaddc-minmax-scalar-x1.c",
296 "src/f32-vbinary/gen/vaddc-minmax-scalar-x2.c",
297 "src/f32-vbinary/gen/vaddc-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700298 "src/f32-vbinary/gen/vaddc-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700299 "src/f32-vbinary/gen/vaddc-relu-scalar-x1.c",
300 "src/f32-vbinary/gen/vaddc-relu-scalar-x2.c",
301 "src/f32-vbinary/gen/vaddc-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700302 "src/f32-vbinary/gen/vaddc-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700303 "src/f32-vbinary/gen/vaddc-scalar-x1.c",
304 "src/f32-vbinary/gen/vaddc-scalar-x2.c",
305 "src/f32-vbinary/gen/vaddc-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700306 "src/f32-vbinary/gen/vaddc-scalar-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700307 "src/f32-vbinary/gen/vdiv-minmax-scalar-x1.c",
308 "src/f32-vbinary/gen/vdiv-minmax-scalar-x2.c",
309 "src/f32-vbinary/gen/vdiv-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700310 "src/f32-vbinary/gen/vdiv-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700311 "src/f32-vbinary/gen/vdiv-relu-scalar-x1.c",
312 "src/f32-vbinary/gen/vdiv-relu-scalar-x2.c",
313 "src/f32-vbinary/gen/vdiv-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700314 "src/f32-vbinary/gen/vdiv-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700315 "src/f32-vbinary/gen/vdiv-scalar-x1.c",
316 "src/f32-vbinary/gen/vdiv-scalar-x2.c",
317 "src/f32-vbinary/gen/vdiv-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700318 "src/f32-vbinary/gen/vdiv-scalar-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700319 "src/f32-vbinary/gen/vdivc-minmax-scalar-x1.c",
320 "src/f32-vbinary/gen/vdivc-minmax-scalar-x2.c",
321 "src/f32-vbinary/gen/vdivc-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700322 "src/f32-vbinary/gen/vdivc-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700323 "src/f32-vbinary/gen/vdivc-relu-scalar-x1.c",
324 "src/f32-vbinary/gen/vdivc-relu-scalar-x2.c",
325 "src/f32-vbinary/gen/vdivc-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700326 "src/f32-vbinary/gen/vdivc-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700327 "src/f32-vbinary/gen/vdivc-scalar-x1.c",
328 "src/f32-vbinary/gen/vdivc-scalar-x2.c",
329 "src/f32-vbinary/gen/vdivc-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700330 "src/f32-vbinary/gen/vdivc-scalar-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800331 "src/f32-vbinary/gen/vmax-scalar-x1.c",
332 "src/f32-vbinary/gen/vmax-scalar-x2.c",
333 "src/f32-vbinary/gen/vmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700334 "src/f32-vbinary/gen/vmax-scalar-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800335 "src/f32-vbinary/gen/vmaxc-scalar-x1.c",
336 "src/f32-vbinary/gen/vmaxc-scalar-x2.c",
337 "src/f32-vbinary/gen/vmaxc-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700338 "src/f32-vbinary/gen/vmaxc-scalar-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800339 "src/f32-vbinary/gen/vmin-scalar-x1.c",
340 "src/f32-vbinary/gen/vmin-scalar-x2.c",
341 "src/f32-vbinary/gen/vmin-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700342 "src/f32-vbinary/gen/vmin-scalar-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800343 "src/f32-vbinary/gen/vminc-scalar-x1.c",
344 "src/f32-vbinary/gen/vminc-scalar-x2.c",
345 "src/f32-vbinary/gen/vminc-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700346 "src/f32-vbinary/gen/vminc-scalar-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700347 "src/f32-vbinary/gen/vmul-minmax-scalar-x1.c",
348 "src/f32-vbinary/gen/vmul-minmax-scalar-x2.c",
349 "src/f32-vbinary/gen/vmul-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700350 "src/f32-vbinary/gen/vmul-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700351 "src/f32-vbinary/gen/vmul-relu-scalar-x1.c",
352 "src/f32-vbinary/gen/vmul-relu-scalar-x2.c",
353 "src/f32-vbinary/gen/vmul-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700354 "src/f32-vbinary/gen/vmul-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700355 "src/f32-vbinary/gen/vmul-scalar-x1.c",
356 "src/f32-vbinary/gen/vmul-scalar-x2.c",
357 "src/f32-vbinary/gen/vmul-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700358 "src/f32-vbinary/gen/vmul-scalar-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700359 "src/f32-vbinary/gen/vmulc-minmax-scalar-x1.c",
360 "src/f32-vbinary/gen/vmulc-minmax-scalar-x2.c",
361 "src/f32-vbinary/gen/vmulc-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700362 "src/f32-vbinary/gen/vmulc-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700363 "src/f32-vbinary/gen/vmulc-relu-scalar-x1.c",
364 "src/f32-vbinary/gen/vmulc-relu-scalar-x2.c",
365 "src/f32-vbinary/gen/vmulc-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700366 "src/f32-vbinary/gen/vmulc-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700367 "src/f32-vbinary/gen/vmulc-scalar-x1.c",
368 "src/f32-vbinary/gen/vmulc-scalar-x2.c",
369 "src/f32-vbinary/gen/vmulc-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700370 "src/f32-vbinary/gen/vmulc-scalar-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700371 "src/f32-vbinary/gen/vrdivc-minmax-scalar-x1.c",
372 "src/f32-vbinary/gen/vrdivc-minmax-scalar-x2.c",
373 "src/f32-vbinary/gen/vrdivc-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700374 "src/f32-vbinary/gen/vrdivc-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700375 "src/f32-vbinary/gen/vrdivc-relu-scalar-x1.c",
376 "src/f32-vbinary/gen/vrdivc-relu-scalar-x2.c",
377 "src/f32-vbinary/gen/vrdivc-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700378 "src/f32-vbinary/gen/vrdivc-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700379 "src/f32-vbinary/gen/vrdivc-scalar-x1.c",
380 "src/f32-vbinary/gen/vrdivc-scalar-x2.c",
381 "src/f32-vbinary/gen/vrdivc-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700382 "src/f32-vbinary/gen/vrdivc-scalar-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700383 "src/f32-vbinary/gen/vrsubc-minmax-scalar-x1.c",
384 "src/f32-vbinary/gen/vrsubc-minmax-scalar-x2.c",
385 "src/f32-vbinary/gen/vrsubc-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700386 "src/f32-vbinary/gen/vrsubc-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700387 "src/f32-vbinary/gen/vrsubc-relu-scalar-x1.c",
388 "src/f32-vbinary/gen/vrsubc-relu-scalar-x2.c",
389 "src/f32-vbinary/gen/vrsubc-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700390 "src/f32-vbinary/gen/vrsubc-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700391 "src/f32-vbinary/gen/vrsubc-scalar-x1.c",
392 "src/f32-vbinary/gen/vrsubc-scalar-x2.c",
393 "src/f32-vbinary/gen/vrsubc-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700394 "src/f32-vbinary/gen/vrsubc-scalar-x8.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -0700395 "src/f32-vbinary/gen/vsqrdiff-scalar-x1.c",
396 "src/f32-vbinary/gen/vsqrdiff-scalar-x2.c",
397 "src/f32-vbinary/gen/vsqrdiff-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700398 "src/f32-vbinary/gen/vsqrdiff-scalar-x8.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -0700399 "src/f32-vbinary/gen/vsqrdiffc-scalar-x1.c",
400 "src/f32-vbinary/gen/vsqrdiffc-scalar-x2.c",
401 "src/f32-vbinary/gen/vsqrdiffc-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700402 "src/f32-vbinary/gen/vsqrdiffc-scalar-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700403 "src/f32-vbinary/gen/vsub-minmax-scalar-x1.c",
404 "src/f32-vbinary/gen/vsub-minmax-scalar-x2.c",
405 "src/f32-vbinary/gen/vsub-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700406 "src/f32-vbinary/gen/vsub-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700407 "src/f32-vbinary/gen/vsub-relu-scalar-x1.c",
408 "src/f32-vbinary/gen/vsub-relu-scalar-x2.c",
409 "src/f32-vbinary/gen/vsub-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700410 "src/f32-vbinary/gen/vsub-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700411 "src/f32-vbinary/gen/vsub-scalar-x1.c",
412 "src/f32-vbinary/gen/vsub-scalar-x2.c",
413 "src/f32-vbinary/gen/vsub-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700414 "src/f32-vbinary/gen/vsub-scalar-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700415 "src/f32-vbinary/gen/vsubc-minmax-scalar-x1.c",
416 "src/f32-vbinary/gen/vsubc-minmax-scalar-x2.c",
417 "src/f32-vbinary/gen/vsubc-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700418 "src/f32-vbinary/gen/vsubc-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700419 "src/f32-vbinary/gen/vsubc-relu-scalar-x1.c",
420 "src/f32-vbinary/gen/vsubc-relu-scalar-x2.c",
421 "src/f32-vbinary/gen/vsubc-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700422 "src/f32-vbinary/gen/vsubc-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700423 "src/f32-vbinary/gen/vsubc-scalar-x1.c",
424 "src/f32-vbinary/gen/vsubc-scalar-x2.c",
425 "src/f32-vbinary/gen/vsubc-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700426 "src/f32-vbinary/gen/vsubc-scalar-x8.c",
Marat Dukhaned6baaf2020-12-01 15:07:08 -0800427 "src/f32-velu/gen/velu-scalar-rr2-lut16-p3-x1.c",
428 "src/f32-velu/gen/velu-scalar-rr2-lut16-p3-x2.c",
429 "src/f32-velu/gen/velu-scalar-rr2-lut16-p3-x3.c",
430 "src/f32-velu/gen/velu-scalar-rr2-lut16-p3-x4.c",
431 "src/f32-velu/gen/velu-scalar-rr2-lut16-p3-x5.c",
432 "src/f32-velu/gen/velu-scalar-rr2-lut16-p3-x6.c",
433 "src/f32-velu/gen/velu-scalar-rr2-p6-x1.c",
434 "src/f32-velu/gen/velu-scalar-rr2-p6-x2.c",
435 "src/f32-velu/gen/velu-scalar-rr2-p6-x3.c",
436 "src/f32-velu/gen/velu-scalar-rr2-p6-x4.c",
437 "src/f32-velu/gen/velu-scalar-rr2-p6-x5.c",
438 "src/f32-velu/gen/velu-scalar-rr2-p6-x6.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700439 "src/f32-vlrelu/gen/vlrelu-scalar-x1.c",
440 "src/f32-vlrelu/gen/vlrelu-scalar-x2.c",
441 "src/f32-vlrelu/gen/vlrelu-scalar-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700442 "src/f32-vmulcaddc/gen/c1-minmax-scalar-2x.c",
443 "src/f32-vmulcaddc/gen/c2-minmax-scalar-2x.c",
444 "src/f32-vmulcaddc/gen/c4-minmax-scalar-2x.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -0700445 "src/f32-vrnd/gen/vrndd-scalar-libm-x1.c",
446 "src/f32-vrnd/gen/vrndd-scalar-libm-x2.c",
447 "src/f32-vrnd/gen/vrndd-scalar-libm-x4.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700448 "src/f32-vrnd/gen/vrndne-scalar-libm-x1.c",
449 "src/f32-vrnd/gen/vrndne-scalar-libm-x2.c",
450 "src/f32-vrnd/gen/vrndne-scalar-libm-x4.c",
451 "src/f32-vrnd/gen/vrndu-scalar-libm-x1.c",
452 "src/f32-vrnd/gen/vrndu-scalar-libm-x2.c",
453 "src/f32-vrnd/gen/vrndu-scalar-libm-x4.c",
454 "src/f32-vrnd/gen/vrndz-scalar-libm-x1.c",
455 "src/f32-vrnd/gen/vrndz-scalar-libm-x2.c",
456 "src/f32-vrnd/gen/vrndz-scalar-libm-x4.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -0700457 "src/f32-vsqrt/gen/scalar-sqrt-x1.c",
458 "src/f32-vsqrt/gen/scalar-sqrt-x2.c",
459 "src/f32-vsqrt/gen/scalar-sqrt-x4.c",
Marat Dukhan5020b962020-06-08 13:30:10 -0700460 "src/f32-vunary/gen/vabs-scalar-x1.c",
461 "src/f32-vunary/gen/vabs-scalar-x2.c",
462 "src/f32-vunary/gen/vabs-scalar-x4.c",
463 "src/f32-vunary/gen/vneg-scalar-x1.c",
464 "src/f32-vunary/gen/vneg-scalar-x2.c",
465 "src/f32-vunary/gen/vneg-scalar-x4.c",
466 "src/f32-vunary/gen/vsqr-scalar-x1.c",
467 "src/f32-vunary/gen/vsqr-scalar-x2.c",
468 "src/f32-vunary/gen/vsqr-scalar-x4.c",
Marat Dukhande390d42020-11-29 19:32:18 -0800469 "src/math/expm1minus-scalar-rr2-lut4-p4.c",
470 "src/math/expm1minus-scalar-rr2-lut8-p3.c",
471 "src/math/expm1minus-scalar-rr2-lut8-p4.c",
Marat Dukhanc60742b2020-11-23 12:33:27 -0800472 "src/math/expm1minus-scalar-rr2-lut16-p3.c",
473 "src/math/expm1minus-scalar-rr2-lut16-p4.c",
474 "src/math/expm1minus-scalar-rr2-p5.c",
475 "src/math/expm1minus-scalar-rr2-p6.c",
Frank Barchard22136062020-11-24 18:44:46 -0800476 "src/math/expminus-scalar-rr2-lut64-p2.c",
477 "src/math/expminus-scalar-rr2-lut2048-p1.c",
478 "src/math/expminus-scalar-rr2-p5.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -0700479 "src/math/roundd-scalar-addsub.c",
480 "src/math/roundd-scalar-cvt.c",
Marat Dukhanffbf96a2020-05-14 02:59:08 -0700481 "src/math/roundd-scalar-floor.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700482 "src/math/roundne-scalar-addsub.c",
483 "src/math/roundne-scalar-nearbyint.c",
484 "src/math/roundne-scalar-rint.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -0700485 "src/math/roundu-scalar-addsub.c",
Marat Dukhanffbf96a2020-05-14 02:59:08 -0700486 "src/math/roundu-scalar-ceil.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -0700487 "src/math/roundu-scalar-cvt.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -0700488 "src/math/roundz-scalar-addsub.c",
489 "src/math/roundz-scalar-cvt.c",
Marat Dukhanffbf96a2020-05-14 02:59:08 -0700490 "src/math/roundz-scalar-trunc.c",
Marat Dukhanf8475d62020-09-17 15:01:43 -0700491 "src/math/sigmoid-scalar-rr2-lut64-p2-div.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700492 "src/math/sigmoid-scalar-rr2-lut2048-p1-div.c",
Marat Dukhanf8475d62020-09-17 15:01:43 -0700493 "src/math/sigmoid-scalar-rr2-p5-div.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700494 "src/qs8-gemm/gen/8x8c4-minmax-scalar.c",
495 "src/qs8-gemm/gen/12x4c4-minmax-scalar.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -0700496 "src/qs8-requantization/fp32-scalar-lrintf.c",
497 "src/qs8-requantization/fp32-scalar-magic.c",
498 "src/qs8-requantization/precise-scalar-signed64.c",
499 "src/qs8-requantization/precise-scalar-unsigned32.c",
500 "src/qs8-requantization/precise-scalar-unsigned64.c",
501 "src/qs8-requantization/q31-scalar.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -0700502 "src/qu8-avgpool/9p8x-minmax-scalar-c1.c",
503 "src/qu8-avgpool/9x-minmax-scalar-c1.c",
504 "src/qu8-dwconv/up1x9-minmax-scalar.c",
505 "src/qu8-gavgpool/7p7x-minmax-scalar-c1.c",
506 "src/qu8-gavgpool/7x-minmax-scalar-c1.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700507 "src/qu8-gemm/2x2-minmax-scalar.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700508 "src/qu8-gemm/gen/8x8c4-minmax-scalar.c",
509 "src/qu8-gemm/gen/12x4c4-minmax-scalar.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -0700510 "src/qu8-igemm/2x2-minmax-scalar.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -0700511 "src/qu8-requantization/fp32-scalar-lrintf.c",
512 "src/qu8-requantization/fp32-scalar-magic.c",
513 "src/qu8-requantization/precise-scalar-signed64.c",
514 "src/qu8-requantization/precise-scalar-unsigned32.c",
515 "src/qu8-requantization/precise-scalar-unsigned64.c",
516 "src/qu8-requantization/q31-scalar.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -0700517 "src/qu8-vadd/minmax-scalar.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -0700518 "src/u8-clamp/scalar-x4.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700519 "src/u8-lut32norm/scalar.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700520 "src/u8-maxpool/9p8x-minmax-scalar-c1.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700521 "src/u8-rmax/scalar.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700522 "src/x8-lut/scalar.c",
523 "src/x8-zip/x2-scalar.c",
524 "src/x8-zip/x3-scalar.c",
525 "src/x8-zip/x4-scalar.c",
526 "src/x8-zip/xm-scalar.c",
Marat Dukhanad71b9a2020-11-20 00:01:51 -0800527 "src/x32-depthtospace2d-chw2hwc/scalar.c",
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -0700528 "src/x32-fill/scalar-float.c",
529 "src/x32-fill/scalar-int.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700530 "src/x32-packx/x2-scalar.c",
531 "src/x32-packx/x3-scalar.c",
532 "src/x32-packx/x4-scalar.c",
Marat Dukhan63523d42020-05-22 17:07:33 -0700533 "src/x32-pad/scalar-float.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700534 "src/x32-pad/scalar-int.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700535 "src/x32-unpool/scalar.c",
536 "src/x32-zip/x2-scalar.c",
537 "src/x32-zip/x3-scalar.c",
538 "src/x32-zip/x4-scalar.c",
539 "src/x32-zip/xm-scalar.c",
Marat Dukhan048931b2020-11-24 20:53:54 -0800540 "src/xx-copy/memcpy.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700541]
542
Marat Dukhan436ebe62019-12-04 15:10:12 -0800543WASM_UKERNELS = [
Marat Dukhan99936602020-04-11 16:47:01 -0700544 "src/f32-avgpool/9p8x-minmax-wasm-c1.c",
545 "src/f32-avgpool/9x-minmax-wasm-c1.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -0700546 "src/f32-clamp/gen/wasm-x1.c",
547 "src/f32-clamp/gen/wasm-x2.c",
548 "src/f32-clamp/gen/wasm-x4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700549 "src/f32-dwconv/gen/up1x4-minmax-wasm-acc2.c",
550 "src/f32-dwconv/gen/up1x4-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700551 "src/f32-dwconv/gen/up1x4-wasm-acc2.c",
552 "src/f32-dwconv/gen/up1x4-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700553 "src/f32-dwconv/gen/up1x9-minmax-wasm-acc2.c",
554 "src/f32-dwconv/gen/up1x9-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700555 "src/f32-dwconv/gen/up1x9-wasm-acc2.c",
556 "src/f32-dwconv/gen/up1x9-wasm.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700557 "src/f32-dwconv/gen/up1x25-minmax-wasm-acc2.c",
558 "src/f32-dwconv/gen/up1x25-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700559 "src/f32-dwconv/gen/up1x25-wasm-acc2.c",
560 "src/f32-dwconv/gen/up1x25-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700561 "src/f32-dwconv/gen/up2x4-minmax-wasm-acc2.c",
562 "src/f32-dwconv/gen/up2x4-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700563 "src/f32-dwconv/gen/up2x4-wasm-acc2.c",
564 "src/f32-dwconv/gen/up2x4-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700565 "src/f32-dwconv/gen/up2x9-minmax-wasm-acc2.c",
566 "src/f32-dwconv/gen/up2x9-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700567 "src/f32-dwconv/gen/up2x9-wasm-acc2.c",
568 "src/f32-dwconv/gen/up2x9-wasm.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700569 "src/f32-dwconv/gen/up2x25-minmax-wasm-acc2.c",
570 "src/f32-dwconv/gen/up2x25-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700571 "src/f32-dwconv/gen/up2x25-wasm-acc2.c",
572 "src/f32-dwconv/gen/up2x25-wasm.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700573 "src/f32-gavgpool/7p7x-minmax-wasm-c1.c",
574 "src/f32-gavgpool/7x-minmax-wasm-c1.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700575 "src/f32-gemm/gen-inc/1x4inc-minmax-wasm.c",
576 "src/f32-gemm/gen-inc/2x4inc-minmax-wasm.c",
577 "src/f32-gemm/gen-inc/4x4inc-minmax-wasm.c",
578 "src/f32-gemm/gen/1x4-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700579 "src/f32-gemm/gen/1x4-relu-wasm.c",
580 "src/f32-gemm/gen/1x4-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700581 "src/f32-gemm/gen/2x4-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700582 "src/f32-gemm/gen/2x4-relu-wasm.c",
583 "src/f32-gemm/gen/2x4-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700584 "src/f32-gemm/gen/4x2-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700585 "src/f32-gemm/gen/4x2-relu-wasm.c",
586 "src/f32-gemm/gen/4x2-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700587 "src/f32-gemm/gen/4x4-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700588 "src/f32-gemm/gen/4x4-relu-wasm.c",
589 "src/f32-gemm/gen/4x4-wasm.c",
Marat Dukhanefaac272020-07-16 16:58:07 -0700590 "src/f32-hswish/gen/hswish-wasm-x1.c",
591 "src/f32-hswish/gen/hswish-wasm-x2.c",
592 "src/f32-hswish/gen/hswish-wasm-x4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700593 "src/f32-igemm/gen/1x4-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700594 "src/f32-igemm/gen/1x4-relu-wasm.c",
595 "src/f32-igemm/gen/1x4-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700596 "src/f32-igemm/gen/2x4-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700597 "src/f32-igemm/gen/2x4-relu-wasm.c",
598 "src/f32-igemm/gen/2x4-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700599 "src/f32-igemm/gen/4x2-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700600 "src/f32-igemm/gen/4x2-relu-wasm.c",
601 "src/f32-igemm/gen/4x2-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700602 "src/f32-igemm/gen/4x4-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700603 "src/f32-igemm/gen/4x4-relu-wasm.c",
604 "src/f32-igemm/gen/4x4-wasm.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700605 "src/f32-maxpool/9p8x-minmax-wasm-c1.c",
606 "src/f32-pavgpool/9p8x-minmax-wasm-c1.c",
607 "src/f32-pavgpool/9x-minmax-wasm-c1.c",
Marat Dukhan7c1f8082020-06-25 13:26:20 -0700608 "src/f32-prelu/gen/wasm-2x1.c",
609 "src/f32-prelu/gen/wasm-2x4.c",
Frank Barchardfb158e22020-07-15 16:10:10 -0700610 "src/f32-relu/gen/wasm-x1.c",
611 "src/f32-relu/gen/wasm-x2.c",
612 "src/f32-relu/gen/wasm-x4.c",
Frank Barchard37297a62020-08-27 12:15:10 -0700613 "src/f32-relu/gen/wasm-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700614 "src/f32-vbinary/gen/vadd-minmax-wasm-x1.c",
615 "src/f32-vbinary/gen/vadd-minmax-wasm-x2.c",
616 "src/f32-vbinary/gen/vadd-minmax-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700617 "src/f32-vbinary/gen/vadd-minmax-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700618 "src/f32-vbinary/gen/vadd-relu-wasm-x1.c",
619 "src/f32-vbinary/gen/vadd-relu-wasm-x2.c",
620 "src/f32-vbinary/gen/vadd-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700621 "src/f32-vbinary/gen/vadd-relu-wasm-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700622 "src/f32-vbinary/gen/vaddc-minmax-wasm-x1.c",
623 "src/f32-vbinary/gen/vaddc-minmax-wasm-x2.c",
624 "src/f32-vbinary/gen/vaddc-minmax-wasm-x4.c",
625 "src/f32-vbinary/gen/vaddc-minmax-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700626 "src/f32-vbinary/gen/vaddc-relu-wasm-x1.c",
627 "src/f32-vbinary/gen/vaddc-relu-wasm-x2.c",
628 "src/f32-vbinary/gen/vaddc-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700629 "src/f32-vbinary/gen/vaddc-relu-wasm-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700630 "src/f32-vbinary/gen/vdiv-minmax-wasm-x1.c",
631 "src/f32-vbinary/gen/vdiv-minmax-wasm-x2.c",
632 "src/f32-vbinary/gen/vdiv-minmax-wasm-x4.c",
633 "src/f32-vbinary/gen/vdiv-minmax-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700634 "src/f32-vbinary/gen/vdiv-relu-wasm-x1.c",
635 "src/f32-vbinary/gen/vdiv-relu-wasm-x2.c",
636 "src/f32-vbinary/gen/vdiv-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700637 "src/f32-vbinary/gen/vdiv-relu-wasm-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700638 "src/f32-vbinary/gen/vdivc-minmax-wasm-x1.c",
639 "src/f32-vbinary/gen/vdivc-minmax-wasm-x2.c",
640 "src/f32-vbinary/gen/vdivc-minmax-wasm-x4.c",
641 "src/f32-vbinary/gen/vdivc-minmax-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700642 "src/f32-vbinary/gen/vdivc-relu-wasm-x1.c",
643 "src/f32-vbinary/gen/vdivc-relu-wasm-x2.c",
644 "src/f32-vbinary/gen/vdivc-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700645 "src/f32-vbinary/gen/vdivc-relu-wasm-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800646 "src/f32-vbinary/gen/vmax-wasm-x1.c",
647 "src/f32-vbinary/gen/vmax-wasm-x2.c",
648 "src/f32-vbinary/gen/vmax-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700649 "src/f32-vbinary/gen/vmax-wasm-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800650 "src/f32-vbinary/gen/vmaxc-wasm-x1.c",
651 "src/f32-vbinary/gen/vmaxc-wasm-x2.c",
652 "src/f32-vbinary/gen/vmaxc-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700653 "src/f32-vbinary/gen/vmaxc-wasm-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800654 "src/f32-vbinary/gen/vmin-wasm-x1.c",
655 "src/f32-vbinary/gen/vmin-wasm-x2.c",
656 "src/f32-vbinary/gen/vmin-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700657 "src/f32-vbinary/gen/vmin-wasm-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800658 "src/f32-vbinary/gen/vminc-wasm-x1.c",
659 "src/f32-vbinary/gen/vminc-wasm-x2.c",
660 "src/f32-vbinary/gen/vminc-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700661 "src/f32-vbinary/gen/vminc-wasm-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700662 "src/f32-vbinary/gen/vmul-minmax-wasm-x1.c",
663 "src/f32-vbinary/gen/vmul-minmax-wasm-x2.c",
664 "src/f32-vbinary/gen/vmul-minmax-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700665 "src/f32-vbinary/gen/vmul-minmax-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700666 "src/f32-vbinary/gen/vmul-relu-wasm-x1.c",
667 "src/f32-vbinary/gen/vmul-relu-wasm-x2.c",
668 "src/f32-vbinary/gen/vmul-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700669 "src/f32-vbinary/gen/vmul-relu-wasm-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700670 "src/f32-vbinary/gen/vmulc-minmax-wasm-x1.c",
671 "src/f32-vbinary/gen/vmulc-minmax-wasm-x2.c",
672 "src/f32-vbinary/gen/vmulc-minmax-wasm-x4.c",
673 "src/f32-vbinary/gen/vmulc-minmax-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700674 "src/f32-vbinary/gen/vmulc-relu-wasm-x1.c",
675 "src/f32-vbinary/gen/vmulc-relu-wasm-x2.c",
676 "src/f32-vbinary/gen/vmulc-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700677 "src/f32-vbinary/gen/vmulc-relu-wasm-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700678 "src/f32-vbinary/gen/vrdivc-minmax-wasm-x1.c",
679 "src/f32-vbinary/gen/vrdivc-minmax-wasm-x2.c",
680 "src/f32-vbinary/gen/vrdivc-minmax-wasm-x4.c",
681 "src/f32-vbinary/gen/vrdivc-minmax-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700682 "src/f32-vbinary/gen/vrdivc-relu-wasm-x1.c",
683 "src/f32-vbinary/gen/vrdivc-relu-wasm-x2.c",
684 "src/f32-vbinary/gen/vrdivc-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700685 "src/f32-vbinary/gen/vrdivc-relu-wasm-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700686 "src/f32-vbinary/gen/vrsubc-minmax-wasm-x1.c",
687 "src/f32-vbinary/gen/vrsubc-minmax-wasm-x2.c",
688 "src/f32-vbinary/gen/vrsubc-minmax-wasm-x4.c",
689 "src/f32-vbinary/gen/vrsubc-minmax-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700690 "src/f32-vbinary/gen/vrsubc-relu-wasm-x1.c",
691 "src/f32-vbinary/gen/vrsubc-relu-wasm-x2.c",
692 "src/f32-vbinary/gen/vrsubc-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700693 "src/f32-vbinary/gen/vrsubc-relu-wasm-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700694 "src/f32-vbinary/gen/vsub-minmax-wasm-x1.c",
695 "src/f32-vbinary/gen/vsub-minmax-wasm-x2.c",
696 "src/f32-vbinary/gen/vsub-minmax-wasm-x4.c",
697 "src/f32-vbinary/gen/vsub-minmax-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700698 "src/f32-vbinary/gen/vsub-relu-wasm-x1.c",
699 "src/f32-vbinary/gen/vsub-relu-wasm-x2.c",
700 "src/f32-vbinary/gen/vsub-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700701 "src/f32-vbinary/gen/vsub-relu-wasm-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700702 "src/f32-vbinary/gen/vsubc-minmax-wasm-x1.c",
703 "src/f32-vbinary/gen/vsubc-minmax-wasm-x2.c",
704 "src/f32-vbinary/gen/vsubc-minmax-wasm-x4.c",
705 "src/f32-vbinary/gen/vsubc-minmax-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700706 "src/f32-vbinary/gen/vsubc-relu-wasm-x1.c",
707 "src/f32-vbinary/gen/vsubc-relu-wasm-x2.c",
708 "src/f32-vbinary/gen/vsubc-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700709 "src/f32-vbinary/gen/vsubc-relu-wasm-x8.c",
Marat Dukhaned6baaf2020-12-01 15:07:08 -0800710 "src/f32-velu/gen/velu-wasm-rr2-lut16-p3-x1.c",
711 "src/f32-velu/gen/velu-wasm-rr2-lut16-p3-x2.c",
712 "src/f32-velu/gen/velu-wasm-rr2-lut16-p3-x3.c",
713 "src/f32-velu/gen/velu-wasm-rr2-lut16-p3-x4.c",
714 "src/f32-velu/gen/velu-wasm-rr2-lut16-p3-x5.c",
715 "src/f32-velu/gen/velu-wasm-rr2-lut16-p3-x6.c",
716 "src/f32-velu/gen/velu-wasm-rr2-p6-x1.c",
717 "src/f32-velu/gen/velu-wasm-rr2-p6-x2.c",
718 "src/f32-velu/gen/velu-wasm-rr2-p6-x3.c",
719 "src/f32-velu/gen/velu-wasm-rr2-p6-x4.c",
720 "src/f32-velu/gen/velu-wasm-rr2-p6-x5.c",
721 "src/f32-velu/gen/velu-wasm-rr2-p6-x6.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -0700722 "src/f32-vlrelu/gen/vlrelu-wasm-x1.c",
723 "src/f32-vlrelu/gen/vlrelu-wasm-x2.c",
724 "src/f32-vlrelu/gen/vlrelu-wasm-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700725 "src/f32-vmulcaddc/gen/c1-minmax-wasm-2x.c",
726 "src/f32-vmulcaddc/gen/c2-minmax-wasm-2x.c",
727 "src/f32-vmulcaddc/gen/c4-minmax-wasm-2x.c",
Marat Dukhan436ebe62019-12-04 15:10:12 -0800728]
729
Marat Dukhan290055c2020-06-09 12:24:29 -0700730WASMSIMD_UKERNELS = [
Marat Dukhan40f05522020-07-16 22:33:12 -0700731 "src/f32-argmaxpool/4x-wasmsimd-c4.c",
732 "src/f32-argmaxpool/9p8x-wasmsimd-c4.c",
733 "src/f32-argmaxpool/9x-wasmsimd-c4.c",
Marat Dukhan3b7432d2020-07-16 17:46:32 -0700734 "src/f32-avgpool/9p8x-minmax-wasmsimd-arm-c4.c",
735 "src/f32-avgpool/9p8x-minmax-wasmsimd-x86-c4.c",
736 "src/f32-avgpool/9x-minmax-wasmsimd-arm-c4.c",
737 "src/f32-avgpool/9x-minmax-wasmsimd-x86-c4.c",
Marat Dukhan3fa52c82020-07-08 12:54:33 -0700738 "src/f32-clamp/gen/wasmsimd-arm-x4.c",
739 "src/f32-clamp/gen/wasmsimd-arm-x8.c",
740 "src/f32-clamp/gen/wasmsimd-x86-x4.c",
741 "src/f32-clamp/gen/wasmsimd-x86-x8.c",
Frank Barchard22136062020-11-24 18:44:46 -0800742 "src/f32-conv-hwc2chw/3x3s2p1c3x4-wasmsimd-2x2.c",
Marat Dukhanac014d72020-06-16 08:36:47 -0700743 "src/f32-dwconv/gen/up4x4-minmax-wasmsimd-acc2-arm.c",
Marat Dukhanac014d72020-06-16 08:36:47 -0700744 "src/f32-dwconv/gen/up4x4-minmax-wasmsimd-acc2-x86.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700745 "src/f32-dwconv/gen/up4x4-minmax-wasmsimd-arm.c",
Marat Dukhanac014d72020-06-16 08:36:47 -0700746 "src/f32-dwconv/gen/up4x4-minmax-wasmsimd-x86.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700747 "src/f32-dwconv/gen/up4x4-wasmsimd.c",
748 "src/f32-dwconv/gen/up4x9-minmax-wasmsimd-acc2-arm.c",
Marat Dukhanac014d72020-06-16 08:36:47 -0700749 "src/f32-dwconv/gen/up4x9-minmax-wasmsimd-acc2-x86.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700750 "src/f32-dwconv/gen/up4x9-minmax-wasmsimd-arm.c",
Marat Dukhanac014d72020-06-16 08:36:47 -0700751 "src/f32-dwconv/gen/up4x9-minmax-wasmsimd-x86.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700752 "src/f32-dwconv/gen/up4x9-wasmsimd.c",
753 "src/f32-dwconv/gen/up4x25-minmax-wasmsimd-acc2-arm.c",
754 "src/f32-dwconv/gen/up4x25-minmax-wasmsimd-acc2-x86.c",
755 "src/f32-dwconv/gen/up4x25-minmax-wasmsimd-arm.c",
756 "src/f32-dwconv/gen/up4x25-minmax-wasmsimd-x86.c",
757 "src/f32-dwconv/gen/up4x25-wasmsimd.c",
758 "src/f32-dwconv/gen/up8x4-minmax-wasmsimd-acc2-arm.c",
Marat Dukhanac014d72020-06-16 08:36:47 -0700759 "src/f32-dwconv/gen/up8x4-minmax-wasmsimd-acc2-x86.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700760 "src/f32-dwconv/gen/up8x4-minmax-wasmsimd-arm.c",
Marat Dukhanac014d72020-06-16 08:36:47 -0700761 "src/f32-dwconv/gen/up8x4-minmax-wasmsimd-x86.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700762 "src/f32-dwconv/gen/up8x4-wasmsimd.c",
763 "src/f32-dwconv/gen/up8x9-minmax-wasmsimd-acc2-arm.c",
Marat Dukhanac014d72020-06-16 08:36:47 -0700764 "src/f32-dwconv/gen/up8x9-minmax-wasmsimd-acc2-x86.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700765 "src/f32-dwconv/gen/up8x9-minmax-wasmsimd-arm.c",
Marat Dukhanac014d72020-06-16 08:36:47 -0700766 "src/f32-dwconv/gen/up8x9-minmax-wasmsimd-x86.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700767 "src/f32-dwconv/gen/up8x9-wasmsimd.c",
768 "src/f32-dwconv/gen/up8x25-minmax-wasmsimd-acc2-arm.c",
769 "src/f32-dwconv/gen/up8x25-minmax-wasmsimd-acc2-x86.c",
770 "src/f32-dwconv/gen/up8x25-minmax-wasmsimd-arm.c",
771 "src/f32-dwconv/gen/up8x25-minmax-wasmsimd-x86.c",
772 "src/f32-dwconv/gen/up8x25-wasmsimd.c",
Frank Barchard3b800452020-11-22 12:12:35 -0800773 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-1x4-acc2.c",
774 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-1x4-acc3.c",
775 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-1x4-acc4.c",
776 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-1x4.c",
777 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-2x4-acc2.c",
778 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-2x4.c",
779 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-3x4.c",
780 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-4x4.c",
781 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-5x4.c",
782 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-6x4.c",
783 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-1x4-acc2.c",
784 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-1x4-acc3.c",
785 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-1x4-acc4.c",
786 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-1x4.c",
787 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-2x4-acc2.c",
788 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-2x4.c",
789 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-3x4.c",
790 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-4x4.c",
791 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-5x4.c",
792 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-6x4.c",
Frank Barchardb88d0112020-12-04 09:07:13 -0800793 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-arm-1x4-acc2.c",
794 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-arm-1x4-acc3.c",
795 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-arm-1x4-acc4.c",
796 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-arm-1x4.c",
797 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-arm-2x4-acc2.c",
798 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-arm-2x4.c",
799 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-arm-3x4.c",
800 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-arm-4x4.c",
801 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-x86-1x4-acc2.c",
802 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-x86-1x4-acc3.c",
803 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-x86-1x4-acc4.c",
804 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-x86-1x4.c",
805 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-x86-2x4-acc2.c",
806 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-x86-2x4.c",
807 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-x86-3x4.c",
808 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-x86-4x4.c",
Frank Barchard20a07412020-11-30 23:30:00 -0800809 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-1x4-acc2.c",
810 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-1x4-acc3.c",
811 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-1x4-acc4.c",
812 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-1x4-acc5.c",
813 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-1x4.c",
814 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-2x4-acc2.c",
815 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-2x4-acc3.c",
816 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-2x4.c",
817 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-3x4-acc2.c",
818 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-3x4.c",
819 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-4x4-acc2.c",
820 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-4x4.c",
821 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-5x4.c",
822 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-1x4-acc2.c",
823 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-1x4-acc3.c",
824 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-1x4-acc4.c",
825 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-1x4-acc5.c",
826 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-1x4.c",
827 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-2x4-acc2.c",
828 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-2x4-acc3.c",
829 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-2x4.c",
830 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-3x4-acc2.c",
831 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-3x4.c",
832 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-4x4-acc2.c",
833 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-4x4.c",
834 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-5x4.c",
Frank Barchard729f07b2020-12-03 16:14:33 -0800835 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-arm-1x4-acc2.c",
836 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-arm-1x4-acc3.c",
837 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-arm-1x4-acc4.c",
838 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-arm-1x4-acc5.c",
839 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-arm-1x4.c",
840 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-arm-2x4-acc2.c",
841 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-arm-2x4-acc3.c",
842 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-arm-2x4.c",
843 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-arm-3x4-acc2.c",
844 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-arm-3x4.c",
845 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-x86-1x4-acc2.c",
846 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-x86-1x4-acc3.c",
847 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-x86-1x4-acc4.c",
848 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-x86-1x4-acc5.c",
849 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-x86-1x4.c",
850 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-x86-2x4-acc2.c",
851 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-x86-2x4-acc3.c",
852 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-x86-2x4.c",
853 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-x86-3x4-acc2.c",
854 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-x86-3x4.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700855 "src/f32-gavgpool-cw/wasmsimd-arm-x4.c",
856 "src/f32-gavgpool-cw/wasmsimd-x86-x4.c",
Marat Dukhanc6016802020-07-16 18:51:28 -0700857 "src/f32-gavgpool/7p7x-minmax-wasmsimd-arm-c4.c",
858 "src/f32-gavgpool/7p7x-minmax-wasmsimd-x86-c4.c",
859 "src/f32-gavgpool/7x-minmax-wasmsimd-arm-c4.c",
860 "src/f32-gavgpool/7x-minmax-wasmsimd-x86-c4.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700861 "src/f32-gemm/gen-inc/1x8inc-minmax-wasmsimd-loadsplat-arm.c",
862 "src/f32-gemm/gen-inc/1x8inc-minmax-wasmsimd-loadsplat-x86.c",
863 "src/f32-gemm/gen-inc/1x8inc-minmax-wasmsimd-splat-arm.c",
864 "src/f32-gemm/gen-inc/1x8inc-minmax-wasmsimd-splat-x86.c",
865 "src/f32-gemm/gen-inc/1x8s4inc-minmax-wasmsimd-arm.c",
866 "src/f32-gemm/gen-inc/1x8s4inc-minmax-wasmsimd-x86.c",
867 "src/f32-gemm/gen-inc/3x8inc-minmax-wasmsimd-loadsplat-arm.c",
868 "src/f32-gemm/gen-inc/3x8inc-minmax-wasmsimd-loadsplat-x86.c",
869 "src/f32-gemm/gen-inc/3x8inc-minmax-wasmsimd-splat-arm.c",
870 "src/f32-gemm/gen-inc/3x8inc-minmax-wasmsimd-splat-x86.c",
871 "src/f32-gemm/gen-inc/3x8s4inc-minmax-wasmsimd-arm.c",
872 "src/f32-gemm/gen-inc/3x8s4inc-minmax-wasmsimd-x86.c",
873 "src/f32-gemm/gen-inc/4x8inc-minmax-wasmsimd-loadsplat-arm.c",
874 "src/f32-gemm/gen-inc/4x8inc-minmax-wasmsimd-loadsplat-x86.c",
875 "src/f32-gemm/gen-inc/4x8inc-minmax-wasmsimd-splat-arm.c",
876 "src/f32-gemm/gen-inc/4x8inc-minmax-wasmsimd-splat-x86.c",
877 "src/f32-gemm/gen-inc/4x8s4inc-minmax-wasmsimd-arm.c",
878 "src/f32-gemm/gen-inc/4x8s4inc-minmax-wasmsimd-x86.c",
879 "src/f32-gemm/gen-inc/5x8inc-minmax-wasmsimd-loadsplat-arm.c",
880 "src/f32-gemm/gen-inc/5x8inc-minmax-wasmsimd-loadsplat-x86.c",
881 "src/f32-gemm/gen-inc/5x8inc-minmax-wasmsimd-splat-arm.c",
882 "src/f32-gemm/gen-inc/5x8inc-minmax-wasmsimd-splat-x86.c",
883 "src/f32-gemm/gen-inc/5x8s4inc-minmax-wasmsimd-arm.c",
884 "src/f32-gemm/gen-inc/5x8s4inc-minmax-wasmsimd-x86.c",
885 "src/f32-gemm/gen-inc/6x8inc-minmax-wasmsimd-loadsplat-arm.c",
886 "src/f32-gemm/gen-inc/6x8inc-minmax-wasmsimd-loadsplat-x86.c",
887 "src/f32-gemm/gen-inc/6x8inc-minmax-wasmsimd-splat-arm.c",
888 "src/f32-gemm/gen-inc/6x8inc-minmax-wasmsimd-splat-x86.c",
889 "src/f32-gemm/gen-inc/6x8s4inc-minmax-wasmsimd-arm.c",
890 "src/f32-gemm/gen-inc/6x8s4inc-minmax-wasmsimd-x86.c",
891 "src/f32-gemm/gen/1x8-minmax-wasmsimd-loadsplat-arm.c",
892 "src/f32-gemm/gen/1x8-minmax-wasmsimd-loadsplat-x86.c",
893 "src/f32-gemm/gen/1x8-minmax-wasmsimd-splat-arm.c",
894 "src/f32-gemm/gen/1x8-minmax-wasmsimd-splat-x86.c",
895 "src/f32-gemm/gen/1x8-relu-wasmsimd-splat.c",
896 "src/f32-gemm/gen/1x8-wasmsimd-splat.c",
897 "src/f32-gemm/gen/1x8s4-minmax-wasmsimd-arm.c",
898 "src/f32-gemm/gen/1x8s4-minmax-wasmsimd-x86.c",
899 "src/f32-gemm/gen/3x8-minmax-wasmsimd-loadsplat-arm.c",
900 "src/f32-gemm/gen/3x8-minmax-wasmsimd-loadsplat-x86.c",
901 "src/f32-gemm/gen/3x8-minmax-wasmsimd-splat-arm.c",
902 "src/f32-gemm/gen/3x8-minmax-wasmsimd-splat-x86.c",
903 "src/f32-gemm/gen/3x8s4-minmax-wasmsimd-arm.c",
904 "src/f32-gemm/gen/3x8s4-minmax-wasmsimd-x86.c",
905 "src/f32-gemm/gen/4x2c4-minmax-wasmsimd-arm.c",
906 "src/f32-gemm/gen/4x2c4-minmax-wasmsimd-x86.c",
907 "src/f32-gemm/gen/4x2c4-relu-wasmsimd.c",
908 "src/f32-gemm/gen/4x2c4-wasmsimd.c",
909 "src/f32-gemm/gen/4x8-minmax-wasmsimd-loadsplat-arm.c",
910 "src/f32-gemm/gen/4x8-minmax-wasmsimd-loadsplat-x86.c",
911 "src/f32-gemm/gen/4x8-minmax-wasmsimd-splat-arm.c",
912 "src/f32-gemm/gen/4x8-minmax-wasmsimd-splat-x86.c",
913 "src/f32-gemm/gen/4x8-relu-wasmsimd-splat.c",
914 "src/f32-gemm/gen/4x8-wasmsimd-splat.c",
915 "src/f32-gemm/gen/4x8s4-minmax-wasmsimd-arm.c",
916 "src/f32-gemm/gen/4x8s4-minmax-wasmsimd-x86.c",
917 "src/f32-gemm/gen/5x8-minmax-wasmsimd-loadsplat-arm.c",
918 "src/f32-gemm/gen/5x8-minmax-wasmsimd-loadsplat-x86.c",
919 "src/f32-gemm/gen/5x8-minmax-wasmsimd-splat-arm.c",
920 "src/f32-gemm/gen/5x8-minmax-wasmsimd-splat-x86.c",
921 "src/f32-gemm/gen/5x8-relu-wasmsimd-splat.c",
922 "src/f32-gemm/gen/5x8-wasmsimd-splat.c",
923 "src/f32-gemm/gen/5x8s4-minmax-wasmsimd-arm.c",
924 "src/f32-gemm/gen/5x8s4-minmax-wasmsimd-x86.c",
925 "src/f32-gemm/gen/6x8-minmax-wasmsimd-loadsplat-arm.c",
926 "src/f32-gemm/gen/6x8-minmax-wasmsimd-loadsplat-x86.c",
927 "src/f32-gemm/gen/6x8-minmax-wasmsimd-splat-arm.c",
928 "src/f32-gemm/gen/6x8-minmax-wasmsimd-splat-x86.c",
929 "src/f32-gemm/gen/6x8s4-minmax-wasmsimd-arm.c",
930 "src/f32-gemm/gen/6x8s4-minmax-wasmsimd-x86.c",
Marat Dukhanefaac272020-07-16 16:58:07 -0700931 "src/f32-hswish/gen/hswish-wasmsimd-x4.c",
932 "src/f32-hswish/gen/hswish-wasmsimd-x8.c",
933 "src/f32-hswish/gen/hswish-wasmsimd-x16.c",
XNNPACK Team965272b2020-10-23 21:10:15 -0700934 "src/f32-ibilinear-chw/gen/wasmsimd-p4.c",
935 "src/f32-ibilinear-chw/gen/wasmsimd-p8.c",
Frank Barchard35db7d02020-10-26 13:37:34 -0700936 "src/f32-ibilinear/gen/wasmsimd-c4.c",
937 "src/f32-ibilinear/gen/wasmsimd-c8.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700938 "src/f32-igemm/gen/1x8-minmax-wasmsimd-loadsplat-arm.c",
939 "src/f32-igemm/gen/1x8-minmax-wasmsimd-loadsplat-x86.c",
940 "src/f32-igemm/gen/1x8-minmax-wasmsimd-splat-arm.c",
941 "src/f32-igemm/gen/1x8-minmax-wasmsimd-splat-x86.c",
942 "src/f32-igemm/gen/1x8-relu-wasmsimd-splat.c",
943 "src/f32-igemm/gen/1x8-wasmsimd-splat.c",
944 "src/f32-igemm/gen/1x8s4-minmax-wasmsimd-arm.c",
945 "src/f32-igemm/gen/1x8s4-minmax-wasmsimd-x86.c",
946 "src/f32-igemm/gen/3x8-minmax-wasmsimd-loadsplat-arm.c",
947 "src/f32-igemm/gen/3x8-minmax-wasmsimd-loadsplat-x86.c",
948 "src/f32-igemm/gen/3x8-minmax-wasmsimd-splat-arm.c",
949 "src/f32-igemm/gen/3x8-minmax-wasmsimd-splat-x86.c",
950 "src/f32-igemm/gen/3x8s4-minmax-wasmsimd-arm.c",
951 "src/f32-igemm/gen/3x8s4-minmax-wasmsimd-x86.c",
952 "src/f32-igemm/gen/4x2c4-minmax-wasmsimd-arm.c",
953 "src/f32-igemm/gen/4x2c4-minmax-wasmsimd-x86.c",
954 "src/f32-igemm/gen/4x2c4-relu-wasmsimd.c",
955 "src/f32-igemm/gen/4x2c4-wasmsimd.c",
956 "src/f32-igemm/gen/4x8-minmax-wasmsimd-loadsplat-arm.c",
957 "src/f32-igemm/gen/4x8-minmax-wasmsimd-loadsplat-x86.c",
958 "src/f32-igemm/gen/4x8-minmax-wasmsimd-splat-arm.c",
959 "src/f32-igemm/gen/4x8-minmax-wasmsimd-splat-x86.c",
960 "src/f32-igemm/gen/4x8-relu-wasmsimd-splat.c",
961 "src/f32-igemm/gen/4x8-wasmsimd-splat.c",
962 "src/f32-igemm/gen/4x8s4-minmax-wasmsimd-arm.c",
963 "src/f32-igemm/gen/4x8s4-minmax-wasmsimd-x86.c",
964 "src/f32-igemm/gen/5x8-minmax-wasmsimd-loadsplat-arm.c",
965 "src/f32-igemm/gen/5x8-minmax-wasmsimd-loadsplat-x86.c",
966 "src/f32-igemm/gen/5x8-minmax-wasmsimd-splat-arm.c",
967 "src/f32-igemm/gen/5x8-minmax-wasmsimd-splat-x86.c",
968 "src/f32-igemm/gen/5x8-relu-wasmsimd-splat.c",
969 "src/f32-igemm/gen/5x8-wasmsimd-splat.c",
970 "src/f32-igemm/gen/5x8s4-minmax-wasmsimd-arm.c",
971 "src/f32-igemm/gen/5x8s4-minmax-wasmsimd-x86.c",
972 "src/f32-igemm/gen/6x8-minmax-wasmsimd-loadsplat-arm.c",
973 "src/f32-igemm/gen/6x8-minmax-wasmsimd-loadsplat-x86.c",
974 "src/f32-igemm/gen/6x8-minmax-wasmsimd-splat-arm.c",
975 "src/f32-igemm/gen/6x8-minmax-wasmsimd-splat-x86.c",
976 "src/f32-igemm/gen/6x8s4-minmax-wasmsimd-arm.c",
977 "src/f32-igemm/gen/6x8s4-minmax-wasmsimd-x86.c",
Marat Dukhanf6e24802020-07-08 22:20:40 -0700978 "src/f32-maxpool/9p8x-minmax-wasmsimd-arm-c4.c",
979 "src/f32-maxpool/9p8x-minmax-wasmsimd-x86-c4.c",
Marat Dukhan1483c532020-07-16 18:08:19 -0700980 "src/f32-pavgpool/9p8x-minmax-wasmsimd-arm-c4.c",
981 "src/f32-pavgpool/9p8x-minmax-wasmsimd-x86-c4.c",
982 "src/f32-pavgpool/9x-minmax-wasmsimd-arm-c4.c",
983 "src/f32-pavgpool/9x-minmax-wasmsimd-x86-c4.c",
Marat Dukhan0d0d8822020-07-23 23:37:56 -0700984 "src/f32-ppmm/gen/4x8-minmax-wasmsimd-splat-arm.c",
985 "src/f32-ppmm/gen/4x8-minmax-wasmsimd-splat-x86.c",
Frank Barcharda5316982020-07-23 13:19:28 -0700986 "src/f32-prelu/gen/wasmsimd-bitselect-1x4.c",
987 "src/f32-prelu/gen/wasmsimd-bitselect-1x8.c",
988 "src/f32-prelu/gen/wasmsimd-bitselect-1x16.c",
Marat Dukhan195f8eb2020-06-25 12:50:57 -0700989 "src/f32-prelu/gen/wasmsimd-bitselect-2x4.c",
990 "src/f32-prelu/gen/wasmsimd-bitselect-2x8.c",
Frank Barcharda5316982020-07-23 13:19:28 -0700991 "src/f32-prelu/gen/wasmsimd-bitselect-2x16.c",
992 "src/f32-prelu/gen/wasmsimd-bitselect-4x4.c",
993 "src/f32-prelu/gen/wasmsimd-bitselect-4x8.c",
994 "src/f32-prelu/gen/wasmsimd-bitselect-4x16.c",
995 "src/f32-prelu/gen/wasmsimd-minmax-1x4.c",
996 "src/f32-prelu/gen/wasmsimd-minmax-1x8.c",
997 "src/f32-prelu/gen/wasmsimd-minmax-1x16.c",
Marat Dukhan195f8eb2020-06-25 12:50:57 -0700998 "src/f32-prelu/gen/wasmsimd-minmax-2x4.c",
999 "src/f32-prelu/gen/wasmsimd-minmax-2x8.c",
Frank Barcharda5316982020-07-23 13:19:28 -07001000 "src/f32-prelu/gen/wasmsimd-minmax-2x16.c",
1001 "src/f32-prelu/gen/wasmsimd-minmax-4x4.c",
1002 "src/f32-prelu/gen/wasmsimd-minmax-4x8.c",
1003 "src/f32-prelu/gen/wasmsimd-minmax-4x16.c",
Marat Dukhan52238f02020-07-16 15:30:28 -07001004 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x4.c",
Marat Dukhan52238f02020-07-16 15:30:28 -07001005 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x8-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001006 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x8.c",
Marat Dukhan52238f02020-07-16 15:30:28 -07001007 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x12-acc2.c",
1008 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x12-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001009 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x12.c",
Marat Dukhan52238f02020-07-16 15:30:28 -07001010 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x16-acc2.c",
1011 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x16-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001012 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x16.c",
Marat Dukhan52238f02020-07-16 15:30:28 -07001013 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x20-acc2.c",
1014 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x20-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001015 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x20.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07001016 "src/f32-relu/gen/wasmsimd-x4.c",
1017 "src/f32-relu/gen/wasmsimd-x8.c",
Frank Barchard37297a62020-08-27 12:15:10 -07001018 "src/f32-relu/gen/wasmsimd-x16.c",
Marat Dukhan8c417962020-07-08 12:27:50 -07001019 "src/f32-rmax/wasmsimd-arm.c",
1020 "src/f32-rmax/wasmsimd-x86.c",
Marat Dukhand187a5b2020-07-20 01:07:17 -07001021 "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x4.c",
1022 "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x8.c",
1023 "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x12.c",
1024 "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x16.c",
1025 "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x20.c",
1026 "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x24.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001027 "src/f32-sigmoid/gen/wasmsimd-p5-div-x4.c",
1028 "src/f32-sigmoid/gen/wasmsimd-p5-div-x8.c",
1029 "src/f32-sigmoid/gen/wasmsimd-p5-div-x12.c",
1030 "src/f32-sigmoid/gen/wasmsimd-p5-div-x16.c",
1031 "src/f32-sigmoid/gen/wasmsimd-p5-div-x20.c",
1032 "src/f32-sigmoid/gen/wasmsimd-p5-div-x24.c",
Frank Barchard8ef44cd2020-11-03 12:30:23 -08001033 "src/f32-spmm/gen/4x1-minmax-wasmsimd-arm-pipelined-x2.c",
1034 "src/f32-spmm/gen/4x1-minmax-wasmsimd-arm-pipelined.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07001035 "src/f32-spmm/gen/4x1-minmax-wasmsimd-arm-x2.c",
1036 "src/f32-spmm/gen/4x1-minmax-wasmsimd-arm-x4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001037 "src/f32-spmm/gen/4x1-minmax-wasmsimd-arm.c",
Frank Barchard8ef44cd2020-11-03 12:30:23 -08001038 "src/f32-spmm/gen/4x1-minmax-wasmsimd-x86-pipelined-x2.c",
1039 "src/f32-spmm/gen/4x1-minmax-wasmsimd-x86-pipelined.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07001040 "src/f32-spmm/gen/4x1-minmax-wasmsimd-x86-x2.c",
1041 "src/f32-spmm/gen/4x1-minmax-wasmsimd-x86-x4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001042 "src/f32-spmm/gen/4x1-minmax-wasmsimd-x86.c",
Frank Barchard8ef44cd2020-11-03 12:30:23 -08001043 "src/f32-spmm/gen/8x1-minmax-wasmsimd-arm-pipelined-x2.c",
1044 "src/f32-spmm/gen/8x1-minmax-wasmsimd-arm-pipelined.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07001045 "src/f32-spmm/gen/8x1-minmax-wasmsimd-arm-x2.c",
1046 "src/f32-spmm/gen/8x1-minmax-wasmsimd-arm-x4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001047 "src/f32-spmm/gen/8x1-minmax-wasmsimd-arm.c",
Frank Barchard8ef44cd2020-11-03 12:30:23 -08001048 "src/f32-spmm/gen/8x1-minmax-wasmsimd-x86-pipelined-x2.c",
1049 "src/f32-spmm/gen/8x1-minmax-wasmsimd-x86-pipelined.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07001050 "src/f32-spmm/gen/8x1-minmax-wasmsimd-x86-x2.c",
1051 "src/f32-spmm/gen/8x1-minmax-wasmsimd-x86-x4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001052 "src/f32-spmm/gen/8x1-minmax-wasmsimd-x86.c",
Frank Barchard8ef44cd2020-11-03 12:30:23 -08001053 "src/f32-spmm/gen/16x1-minmax-wasmsimd-arm-pipelined-x2.c",
1054 "src/f32-spmm/gen/16x1-minmax-wasmsimd-arm-pipelined.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07001055 "src/f32-spmm/gen/16x1-minmax-wasmsimd-arm-x2.c",
1056 "src/f32-spmm/gen/16x1-minmax-wasmsimd-arm-x4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001057 "src/f32-spmm/gen/16x1-minmax-wasmsimd-arm.c",
Frank Barchard8ef44cd2020-11-03 12:30:23 -08001058 "src/f32-spmm/gen/16x1-minmax-wasmsimd-x86-pipelined-x2.c",
1059 "src/f32-spmm/gen/16x1-minmax-wasmsimd-x86-pipelined.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07001060 "src/f32-spmm/gen/16x1-minmax-wasmsimd-x86-x2.c",
1061 "src/f32-spmm/gen/16x1-minmax-wasmsimd-x86-x4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001062 "src/f32-spmm/gen/16x1-minmax-wasmsimd-x86.c",
Frank Barchard8ef44cd2020-11-03 12:30:23 -08001063 "src/f32-spmm/gen/32x1-minmax-wasmsimd-arm-pipelined-x2.c",
1064 "src/f32-spmm/gen/32x1-minmax-wasmsimd-arm-pipelined.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07001065 "src/f32-spmm/gen/32x1-minmax-wasmsimd-arm-x2.c",
1066 "src/f32-spmm/gen/32x1-minmax-wasmsimd-arm-x4.c",
Frank Barchard846c0c62020-10-26 15:01:39 -07001067 "src/f32-spmm/gen/32x1-minmax-wasmsimd-arm.c",
Frank Barchard8ef44cd2020-11-03 12:30:23 -08001068 "src/f32-spmm/gen/32x1-minmax-wasmsimd-x86-pipelined-x2.c",
1069 "src/f32-spmm/gen/32x1-minmax-wasmsimd-x86-pipelined.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07001070 "src/f32-spmm/gen/32x1-minmax-wasmsimd-x86-x2.c",
1071 "src/f32-spmm/gen/32x1-minmax-wasmsimd-x86-x4.c",
Frank Barchard846c0c62020-10-26 15:01:39 -07001072 "src/f32-spmm/gen/32x1-minmax-wasmsimd-x86.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -07001073 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-arm-x4.c",
1074 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001075 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-arm-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -07001076 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-x86-x4.c",
1077 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001078 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -07001079 "src/f32-vbinary/gen/vadd-relu-wasmsimd-x4.c",
1080 "src/f32-vbinary/gen/vadd-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001081 "src/f32-vbinary/gen/vadd-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07001082 "src/f32-vbinary/gen/vadd-wasmsimd-x4.c",
1083 "src/f32-vbinary/gen/vadd-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001084 "src/f32-vbinary/gen/vadd-wasmsimd-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001085 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-arm-x4.c",
1086 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001087 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-arm-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001088 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-x86-x4.c",
1089 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001090 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -07001091 "src/f32-vbinary/gen/vaddc-relu-wasmsimd-x4.c",
1092 "src/f32-vbinary/gen/vaddc-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001093 "src/f32-vbinary/gen/vaddc-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07001094 "src/f32-vbinary/gen/vaddc-wasmsimd-x4.c",
1095 "src/f32-vbinary/gen/vaddc-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001096 "src/f32-vbinary/gen/vaddc-wasmsimd-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -07001097 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-arm-x4.c",
1098 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001099 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-arm-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -07001100 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-x86-x4.c",
1101 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001102 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -07001103 "src/f32-vbinary/gen/vdiv-relu-wasmsimd-x4.c",
1104 "src/f32-vbinary/gen/vdiv-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001105 "src/f32-vbinary/gen/vdiv-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07001106 "src/f32-vbinary/gen/vdiv-wasmsimd-x4.c",
1107 "src/f32-vbinary/gen/vdiv-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001108 "src/f32-vbinary/gen/vdiv-wasmsimd-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001109 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-arm-x4.c",
1110 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001111 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-arm-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001112 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-x86-x4.c",
1113 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001114 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -07001115 "src/f32-vbinary/gen/vdivc-relu-wasmsimd-x4.c",
1116 "src/f32-vbinary/gen/vdivc-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001117 "src/f32-vbinary/gen/vdivc-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07001118 "src/f32-vbinary/gen/vdivc-wasmsimd-x4.c",
1119 "src/f32-vbinary/gen/vdivc-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001120 "src/f32-vbinary/gen/vdivc-wasmsimd-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -07001121 "src/f32-vbinary/gen/vmax-wasmsimd-arm-x4.c",
1122 "src/f32-vbinary/gen/vmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001123 "src/f32-vbinary/gen/vmax-wasmsimd-arm-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -07001124 "src/f32-vbinary/gen/vmax-wasmsimd-x86-x4.c",
1125 "src/f32-vbinary/gen/vmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001126 "src/f32-vbinary/gen/vmax-wasmsimd-x86-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001127 "src/f32-vbinary/gen/vmaxc-wasmsimd-arm-x4.c",
1128 "src/f32-vbinary/gen/vmaxc-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001129 "src/f32-vbinary/gen/vmaxc-wasmsimd-arm-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001130 "src/f32-vbinary/gen/vmaxc-wasmsimd-x86-x4.c",
1131 "src/f32-vbinary/gen/vmaxc-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001132 "src/f32-vbinary/gen/vmaxc-wasmsimd-x86-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -07001133 "src/f32-vbinary/gen/vmin-wasmsimd-arm-x4.c",
1134 "src/f32-vbinary/gen/vmin-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001135 "src/f32-vbinary/gen/vmin-wasmsimd-arm-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -07001136 "src/f32-vbinary/gen/vmin-wasmsimd-x86-x4.c",
1137 "src/f32-vbinary/gen/vmin-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001138 "src/f32-vbinary/gen/vmin-wasmsimd-x86-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001139 "src/f32-vbinary/gen/vminc-wasmsimd-arm-x4.c",
1140 "src/f32-vbinary/gen/vminc-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001141 "src/f32-vbinary/gen/vminc-wasmsimd-arm-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001142 "src/f32-vbinary/gen/vminc-wasmsimd-x86-x4.c",
1143 "src/f32-vbinary/gen/vminc-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001144 "src/f32-vbinary/gen/vminc-wasmsimd-x86-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -07001145 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-arm-x4.c",
1146 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001147 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-arm-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -07001148 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-x86-x4.c",
1149 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001150 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -07001151 "src/f32-vbinary/gen/vmul-relu-wasmsimd-x4.c",
1152 "src/f32-vbinary/gen/vmul-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001153 "src/f32-vbinary/gen/vmul-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07001154 "src/f32-vbinary/gen/vmul-wasmsimd-x4.c",
1155 "src/f32-vbinary/gen/vmul-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001156 "src/f32-vbinary/gen/vmul-wasmsimd-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001157 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-arm-x4.c",
1158 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001159 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-arm-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001160 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-x86-x4.c",
1161 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001162 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -07001163 "src/f32-vbinary/gen/vmulc-relu-wasmsimd-x4.c",
1164 "src/f32-vbinary/gen/vmulc-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001165 "src/f32-vbinary/gen/vmulc-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07001166 "src/f32-vbinary/gen/vmulc-wasmsimd-x4.c",
1167 "src/f32-vbinary/gen/vmulc-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001168 "src/f32-vbinary/gen/vmulc-wasmsimd-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001169 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-arm-x4.c",
1170 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001171 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-arm-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001172 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-x86-x4.c",
1173 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001174 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -07001175 "src/f32-vbinary/gen/vrdivc-relu-wasmsimd-x4.c",
1176 "src/f32-vbinary/gen/vrdivc-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001177 "src/f32-vbinary/gen/vrdivc-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07001178 "src/f32-vbinary/gen/vrdivc-wasmsimd-x4.c",
1179 "src/f32-vbinary/gen/vrdivc-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001180 "src/f32-vbinary/gen/vrdivc-wasmsimd-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001181 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-arm-x4.c",
1182 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001183 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-arm-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001184 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-x86-x4.c",
1185 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001186 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -07001187 "src/f32-vbinary/gen/vrsubc-relu-wasmsimd-x4.c",
1188 "src/f32-vbinary/gen/vrsubc-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001189 "src/f32-vbinary/gen/vrsubc-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07001190 "src/f32-vbinary/gen/vrsubc-wasmsimd-x4.c",
1191 "src/f32-vbinary/gen/vrsubc-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001192 "src/f32-vbinary/gen/vrsubc-wasmsimd-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -07001193 "src/f32-vbinary/gen/vsqrdiff-wasmsimd-x4.c",
1194 "src/f32-vbinary/gen/vsqrdiff-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001195 "src/f32-vbinary/gen/vsqrdiff-wasmsimd-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001196 "src/f32-vbinary/gen/vsqrdiffc-wasmsimd-x4.c",
1197 "src/f32-vbinary/gen/vsqrdiffc-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001198 "src/f32-vbinary/gen/vsqrdiffc-wasmsimd-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -07001199 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-arm-x4.c",
1200 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001201 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-arm-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -07001202 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-x86-x4.c",
1203 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001204 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -07001205 "src/f32-vbinary/gen/vsub-relu-wasmsimd-x4.c",
1206 "src/f32-vbinary/gen/vsub-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001207 "src/f32-vbinary/gen/vsub-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07001208 "src/f32-vbinary/gen/vsub-wasmsimd-x4.c",
1209 "src/f32-vbinary/gen/vsub-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001210 "src/f32-vbinary/gen/vsub-wasmsimd-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001211 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-arm-x4.c",
1212 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001213 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-arm-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001214 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-x86-x4.c",
1215 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001216 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -07001217 "src/f32-vbinary/gen/vsubc-relu-wasmsimd-x4.c",
1218 "src/f32-vbinary/gen/vsubc-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001219 "src/f32-vbinary/gen/vsubc-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07001220 "src/f32-vbinary/gen/vsubc-wasmsimd-x4.c",
1221 "src/f32-vbinary/gen/vsubc-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001222 "src/f32-vbinary/gen/vsubc-wasmsimd-x16.c",
Marat Dukhaned6baaf2020-12-01 15:07:08 -08001223 "src/f32-velu/gen/velu-wasmsimd-arm-rr2-lut16-p3-x4.c",
1224 "src/f32-velu/gen/velu-wasmsimd-arm-rr2-lut16-p3-x8.c",
1225 "src/f32-velu/gen/velu-wasmsimd-arm-rr2-lut16-p3-x12.c",
1226 "src/f32-velu/gen/velu-wasmsimd-arm-rr2-lut16-p3-x16.c",
1227 "src/f32-velu/gen/velu-wasmsimd-arm-rr2-lut16-p3-x20.c",
1228 "src/f32-velu/gen/velu-wasmsimd-arm-rr2-lut16-p3-x24.c",
Marat Dukhaned6baaf2020-12-01 15:07:08 -08001229 "src/f32-velu/gen/velu-wasmsimd-arm-rr2-p6-x4.c",
1230 "src/f32-velu/gen/velu-wasmsimd-arm-rr2-p6-x8.c",
1231 "src/f32-velu/gen/velu-wasmsimd-arm-rr2-p6-x12.c",
1232 "src/f32-velu/gen/velu-wasmsimd-arm-rr2-p6-x16.c",
1233 "src/f32-velu/gen/velu-wasmsimd-arm-rr2-p6-x20.c",
1234 "src/f32-velu/gen/velu-wasmsimd-arm-rr2-p6-x24.c",
Frank Barcharde7223ee2020-12-04 19:04:01 -08001235 "src/f32-velu/gen/velu-wasmsimd-x86-rr2-lut16-p3-x4.c",
1236 "src/f32-velu/gen/velu-wasmsimd-x86-rr2-lut16-p3-x8.c",
1237 "src/f32-velu/gen/velu-wasmsimd-x86-rr2-lut16-p3-x12.c",
1238 "src/f32-velu/gen/velu-wasmsimd-x86-rr2-lut16-p3-x16.c",
1239 "src/f32-velu/gen/velu-wasmsimd-x86-rr2-lut16-p3-x20.c",
1240 "src/f32-velu/gen/velu-wasmsimd-x86-rr2-lut16-p3-x24.c",
Marat Dukhaned6baaf2020-12-01 15:07:08 -08001241 "src/f32-velu/gen/velu-wasmsimd-x86-rr2-p6-x4.c",
1242 "src/f32-velu/gen/velu-wasmsimd-x86-rr2-p6-x8.c",
1243 "src/f32-velu/gen/velu-wasmsimd-x86-rr2-p6-x12.c",
1244 "src/f32-velu/gen/velu-wasmsimd-x86-rr2-p6-x16.c",
1245 "src/f32-velu/gen/velu-wasmsimd-x86-rr2-p6-x20.c",
1246 "src/f32-velu/gen/velu-wasmsimd-x86-rr2-p6-x24.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07001247 "src/f32-vlrelu/gen/vlrelu-wasmsimd-bitselect-x4.c",
1248 "src/f32-vlrelu/gen/vlrelu-wasmsimd-bitselect-x8.c",
1249 "src/f32-vlrelu/gen/vlrelu-wasmsimd-minmax-x4.c",
1250 "src/f32-vlrelu/gen/vlrelu-wasmsimd-minmax-x8.c",
Marat Dukhand816f622020-07-15 10:14:39 -07001251 "src/f32-vmulcaddc/gen/c4-minmax-wasmsimd-arm-2x.c",
Marat Dukhand816f622020-07-15 10:14:39 -07001252 "src/f32-vmulcaddc/gen/c4-minmax-wasmsimd-x86-2x.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001253 "src/f32-vmulcaddc/gen/c8-minmax-wasmsimd-arm-2x.c",
Marat Dukhand816f622020-07-15 10:14:39 -07001254 "src/f32-vmulcaddc/gen/c8-minmax-wasmsimd-x86-2x.c",
Marat Dukhanb82b2cd2020-07-16 02:23:42 -07001255 "src/f32-vrnd/gen/vrndd-wasmsimd-addsub-x4.c",
1256 "src/f32-vrnd/gen/vrndd-wasmsimd-addsub-x8.c",
1257 "src/f32-vrnd/gen/vrndd-wasmsimd-cvt-x4.c",
1258 "src/f32-vrnd/gen/vrndd-wasmsimd-cvt-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001259 "src/f32-vrnd/gen/vrndne-wasmsimd-addsub-x4.c",
1260 "src/f32-vrnd/gen/vrndne-wasmsimd-addsub-x8.c",
1261 "src/f32-vrnd/gen/vrndu-wasmsimd-addsub-x4.c",
1262 "src/f32-vrnd/gen/vrndu-wasmsimd-addsub-x8.c",
1263 "src/f32-vrnd/gen/vrndu-wasmsimd-cvt-x4.c",
1264 "src/f32-vrnd/gen/vrndu-wasmsimd-cvt-x8.c",
1265 "src/f32-vrnd/gen/vrndz-wasmsimd-addsub-x4.c",
1266 "src/f32-vrnd/gen/vrndz-wasmsimd-addsub-x8.c",
1267 "src/f32-vrnd/gen/vrndz-wasmsimd-cvt-x4.c",
1268 "src/f32-vrnd/gen/vrndz-wasmsimd-cvt-x8.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07001269 "src/f32-vsqrt/gen/wasmsimd-sqrt-x4.c",
1270 "src/f32-vsqrt/gen/wasmsimd-sqrt-x8.c",
Marat Dukhan37c83512020-06-29 13:25:53 -07001271 "src/f32-vunary/gen/vabs-wasmsimd-x4.c",
1272 "src/f32-vunary/gen/vabs-wasmsimd-x8.c",
1273 "src/f32-vunary/gen/vneg-wasmsimd-x4.c",
1274 "src/f32-vunary/gen/vneg-wasmsimd-x8.c",
1275 "src/f32-vunary/gen/vsqr-wasmsimd-x4.c",
1276 "src/f32-vunary/gen/vsqr-wasmsimd-x8.c",
Marat Dukhande390d42020-11-29 19:32:18 -08001277 "src/math/expm1minus-wasmsimd-rr2-lut16-p3-andnot.c",
1278 "src/math/expm1minus-wasmsimd-rr2-lut16-p3-max.c",
1279 "src/math/expm1minus-wasmsimd-rr2-p6-andnot.c",
1280 "src/math/expm1minus-wasmsimd-rr2-p6-max.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001281 "src/math/roundd-wasmsimd-addsub.c",
1282 "src/math/roundd-wasmsimd-cvt.c",
1283 "src/math/roundne-wasmsimd-addsub.c",
1284 "src/math/roundu-wasmsimd-addsub.c",
1285 "src/math/roundu-wasmsimd-cvt.c",
1286 "src/math/roundz-wasmsimd-addsub.c",
1287 "src/math/roundz-wasmsimd-cvt.c",
1288 "src/math/sigmoid-wasmsimd-rr2-lut64-p2-div.c",
1289 "src/math/sigmoid-wasmsimd-rr2-p5-div.c",
Marat Dukhancc8f34c2020-08-05 16:36:38 -07001290 "src/qs8-dwconv/gen/up8x9-minmax-wasmsimd-mul16.c",
1291 "src/qs8-dwconv/gen/up16x9-minmax-wasmsimd-mul16.c",
1292 "src/qs8-dwconv/gen/up24x9-minmax-wasmsimd-mul16.c",
Marat Dukhanb5e3d172020-08-06 13:29:53 -07001293 "src/qs8-gavgpool/gen/7p7x-minmax-wasmsimd-c8-acc2.c",
1294 "src/qs8-gavgpool/gen/7p7x-minmax-wasmsimd-c16-acc2.c",
1295 "src/qs8-gavgpool/gen/7p7x-minmax-wasmsimd-c24-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001296 "src/qs8-gavgpool/gen/7x-minmax-wasmsimd-c8-acc2.c",
1297 "src/qs8-gavgpool/gen/7x-minmax-wasmsimd-c16-acc2.c",
1298 "src/qs8-gavgpool/gen/7x-minmax-wasmsimd-c24-acc2.c",
Marat Dukhan27203da2020-08-05 15:19:03 -07001299 "src/qs8-gemm/gen/1x4c8-minmax-wasmsimd-ld64.c",
Marat Dukhan27203da2020-08-05 15:19:03 -07001300 "src/qs8-gemm/gen/1x4c8-minmax-wasmsimd-ld128.c",
Marat Dukhan27203da2020-08-05 15:19:03 -07001301 "src/qs8-gemm/gen/1x4c8-xw-minmax-wasmsimd.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001302 "src/qs8-gemm/gen/2x4c8-minmax-wasmsimd-ld64.c",
1303 "src/qs8-gemm/gen/2x4c8-minmax-wasmsimd-ld128.c",
Marat Dukhan27203da2020-08-05 15:19:03 -07001304 "src/qs8-gemm/gen/2x4c8-xw-minmax-wasmsimd.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001305 "src/qs8-gemm/gen/3x4c8-minmax-wasmsimd-ld64.c",
1306 "src/qs8-gemm/gen/3x4c8-minmax-wasmsimd-ld128.c",
Marat Dukhan27203da2020-08-05 15:19:03 -07001307 "src/qs8-gemm/gen/3x4c8-xw-minmax-wasmsimd.c",
1308 "src/qs8-igemm/gen/1x4c8-minmax-wasmsimd-ld64.c",
Marat Dukhan27203da2020-08-05 15:19:03 -07001309 "src/qs8-igemm/gen/1x4c8-minmax-wasmsimd-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001310 "src/qs8-igemm/gen/2x4c8-minmax-wasmsimd-ld64.c",
Marat Dukhan27203da2020-08-05 15:19:03 -07001311 "src/qs8-igemm/gen/2x4c8-minmax-wasmsimd-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001312 "src/qs8-igemm/gen/3x4c8-minmax-wasmsimd-ld64.c",
Marat Dukhan27203da2020-08-05 15:19:03 -07001313 "src/qs8-igemm/gen/3x4c8-minmax-wasmsimd-ld128.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -07001314 "src/qs8-requantization/fp32-wasmsimd.c",
Marat Dukhan138560c2020-08-03 18:57:34 -07001315 "src/qs8-requantization/q31-wasmsimd.c",
Marat Dukhan5df27f82020-09-02 23:59:21 -07001316 "src/qs8-vadd/gen/minmax-wasmsimd-x8.c",
1317 "src/qs8-vadd/gen/minmax-wasmsimd-x16.c",
1318 "src/qs8-vadd/gen/minmax-wasmsimd-x24.c",
1319 "src/qs8-vadd/gen/minmax-wasmsimd-x32.c",
1320 "src/qs8-vaddc/gen/minmax-wasmsimd-x8.c",
1321 "src/qs8-vaddc/gen/minmax-wasmsimd-x16.c",
1322 "src/qs8-vaddc/gen/minmax-wasmsimd-x24.c",
1323 "src/qs8-vaddc/gen/minmax-wasmsimd-x32.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07001324 "src/qu8-requantization/fp32-wasmsimd.c",
Marat Dukhan22de5e72020-08-03 21:53:18 -07001325 "src/qu8-requantization/q31-wasmsimd.c",
Marat Dukhan8ee37012020-07-16 13:17:13 -07001326 "src/x32-fill/wasmsimd.c",
Marat Dukhan66d99e92020-07-16 12:56:21 -07001327 "src/x32-packx/x4-wasmsimd.c",
Marat Dukhan9306ae02020-07-16 15:51:13 -07001328 "src/x32-pad/wasmsimd.c",
Marat Dukhan9d4bfa22020-07-16 19:07:04 -07001329 "src/x32-unpool/wasmsimd.c",
Marat Dukhane3b78762020-07-16 20:02:58 -07001330 "src/x32-zip/x2-wasmsimd.c",
1331 "src/x32-zip/x3-wasmsimd.c",
1332 "src/x32-zip/x4-wasmsimd.c",
1333 "src/x32-zip/xm-wasmsimd.c",
Marat Dukhan290055c2020-06-09 12:24:29 -07001334]
1335
Marat Dukhan102e9ea2020-12-05 22:35:18 -08001336PSIMD_ACCMATH_UKERNELS = [
1337 "src/qs8-requantization/fp32-psimd.c",
1338 "src/qs8-requantization/precise-psimd.c",
1339 "src/qu8-requantization/fp32-psimd.c",
1340 "src/qu8-requantization/precise-psimd.c",
1341]
1342
Marat Dukhan08c4a432019-10-03 09:29:21 -07001343# ISA-specific micro-kernels
1344NEON_UKERNELS = [
Marat Dukhanef25c6d2020-07-24 00:59:40 -07001345 "src/f32-argmaxpool/4x-neon-c4.c",
1346 "src/f32-argmaxpool/9p8x-neon-c4.c",
1347 "src/f32-argmaxpool/9x-neon-c4.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001348 "src/f32-avgpool/9p8x-minmax-neon-c4.c",
1349 "src/f32-avgpool/9x-minmax-neon-c4.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001350 "src/f32-clamp/gen/neon-x4.c",
1351 "src/f32-clamp/gen/neon-x8.c",
Marat Dukhan56b10cd2020-05-18 09:35:49 -07001352 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neon-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001353 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neon-2x2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001354 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neon-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001355 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neon-2x2.c",
Marat Dukhan56b10cd2020-05-18 09:35:49 -07001356 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neon-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001357 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neon-2x2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001358 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neon-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001359 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neon-2x2.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -07001360 "src/f32-dwconv/gen/up4x4-minmax-neon-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001361 "src/f32-dwconv/gen/up4x4-minmax-neon.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001362 "src/f32-dwconv/gen/up4x9-minmax-neon-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001363 "src/f32-dwconv/gen/up4x9-minmax-neon.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -07001364 "src/f32-dwconv/gen/up4x25-minmax-neon-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001365 "src/f32-dwconv/gen/up4x25-minmax-neon.c",
1366 "src/f32-dwconv/gen/up8x4-minmax-neon-acc2.c",
1367 "src/f32-dwconv/gen/up8x4-minmax-neon.c",
1368 "src/f32-dwconv/gen/up8x9-minmax-neon-acc2.c",
1369 "src/f32-dwconv/gen/up8x9-minmax-neon.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -07001370 "src/f32-dwconv/gen/up8x25-minmax-neon-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001371 "src/f32-dwconv/gen/up8x25-minmax-neon.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001372 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-1x4-acc2.c",
1373 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-1x4-acc3.c",
1374 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-1x4-acc4.c",
Marat Dukhanc581e482020-10-24 01:28:11 -07001375 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-1x4.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001376 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-2x4-acc2.c",
Marat Dukhanc581e482020-10-24 01:28:11 -07001377 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-2x4.c",
1378 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-3x4.c",
1379 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-4x4.c",
1380 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-5x4.c",
1381 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-6x4.c",
Marat Dukhan82f0c322020-10-25 19:17:35 -07001382 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neon-1x4-acc2.c",
1383 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neon-1x4-acc3.c",
1384 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neon-1x4-acc4.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001385 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neon-1x4.c",
Marat Dukhan82f0c322020-10-25 19:17:35 -07001386 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neon-2x4-acc2.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001387 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neon-2x4.c",
1388 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neon-3x4.c",
1389 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neon-4x4.c",
Marat Dukhan149f0ea2020-10-26 12:50:33 -07001390 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-1x4-acc2.c",
1391 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-1x4-acc3.c",
1392 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-1x4-acc4.c",
1393 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-1x4-acc5.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001394 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-1x4.c",
Marat Dukhan149f0ea2020-10-26 12:50:33 -07001395 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-2x4-acc2.c",
1396 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-2x4-acc3.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001397 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-2x4.c",
Marat Dukhan149f0ea2020-10-26 12:50:33 -07001398 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-3x4-acc2.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001399 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-3x4.c",
Marat Dukhan149f0ea2020-10-26 12:50:33 -07001400 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-4x4-acc2.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001401 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-4x4.c",
1402 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-5x4.c",
Marat Dukhan30d4b252020-10-29 16:33:22 -07001403 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neon-1x4-acc2.c",
1404 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neon-1x4-acc3.c",
1405 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neon-1x4-acc4.c",
1406 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neon-1x4-acc5.c",
1407 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neon-1x4.c",
1408 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neon-2x4-acc2.c",
1409 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neon-2x4-acc3.c",
1410 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neon-2x4.c",
Marat Dukhan30d4b252020-10-29 16:33:22 -07001411 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neon-3x4-acc2.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07001412 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neon-3x4.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -07001413 "src/f32-gavgpool-cw/neon-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001414 "src/f32-gavgpool/7p7x-minmax-neon-c4.c",
1415 "src/f32-gavgpool/7x-minmax-neon-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001416 "src/f32-gemm/gen-inc/1x8inc-minmax-neon-dup-ld64.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001417 "src/f32-gemm/gen-inc/1x8inc-minmax-neon-lane-ld64.c",
1418 "src/f32-gemm/gen-inc/1x8s4inc-minmax-neon.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001419 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-dup-ld64.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001420 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-dup-ld128.c",
1421 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-lane-ld64.c",
1422 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-lane-ld128.c",
1423 "src/f32-gemm/gen-inc/4x8s4inc-minmax-neon.c",
1424 "src/f32-gemm/gen-inc/5x8inc-minmax-neon-lane-ld64.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001425 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-dup-ld64.c",
1426 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-dup-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001427 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-lane-ld64.c",
1428 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-lane-ld128.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001429 "src/f32-gemm/gen-inc/6x8s4inc-minmax-neon.c",
1430 "src/f32-gemm/gen-inc/8x8s4inc-minmax-neon.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001431 "src/f32-gemm/gen/1x8-minmax-neon-dup-ld64.c",
1432 "src/f32-gemm/gen/1x8-minmax-neon-lane-ld64.c",
1433 "src/f32-gemm/gen/1x8s4-minmax-neon.c",
1434 "src/f32-gemm/gen/4x2-minmax-neon-lane-ld64.c",
1435 "src/f32-gemm/gen/4x8-minmax-neon-dup-ld64.c",
1436 "src/f32-gemm/gen/4x8-minmax-neon-dup-ld128.c",
1437 "src/f32-gemm/gen/4x8-minmax-neon-lane-ld64.c",
1438 "src/f32-gemm/gen/4x8-minmax-neon-lane-ld128.c",
1439 "src/f32-gemm/gen/4x8s4-minmax-neon.c",
1440 "src/f32-gemm/gen/5x8-minmax-neon-lane-ld64.c",
1441 "src/f32-gemm/gen/6x8-minmax-neon-dup-ld64.c",
1442 "src/f32-gemm/gen/6x8-minmax-neon-dup-ld128.c",
1443 "src/f32-gemm/gen/6x8-minmax-neon-lane-ld64.c",
1444 "src/f32-gemm/gen/6x8-minmax-neon-lane-ld128.c",
1445 "src/f32-gemm/gen/6x8s4-minmax-neon.c",
1446 "src/f32-gemm/gen/8x8s4-minmax-neon.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07001447 "src/f32-hswish/gen/hswish-neon-x4.c",
1448 "src/f32-hswish/gen/hswish-neon-x8.c",
1449 "src/f32-hswish/gen/hswish-neon-x16.c",
Marat Dukhan660fd192020-03-10 04:55:30 -07001450 "src/f32-ibilinear/gen/neon-c4.c",
1451 "src/f32-ibilinear/gen/neon-c8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001452 "src/f32-igemm/gen/1x8-minmax-neon-dup-ld64.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001453 "src/f32-igemm/gen/1x8-minmax-neon-lane-ld64.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001454 "src/f32-igemm/gen/1x8s4-minmax-neon.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001455 "src/f32-igemm/gen/4x2-minmax-neon-lane-ld64.c",
1456 "src/f32-igemm/gen/4x4-minmax-neon-lane-ld64.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001457 "src/f32-igemm/gen/4x8-minmax-neon-dup-ld64.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001458 "src/f32-igemm/gen/4x8-minmax-neon-dup-ld128.c",
1459 "src/f32-igemm/gen/4x8-minmax-neon-lane-ld64.c",
1460 "src/f32-igemm/gen/4x8-minmax-neon-lane-ld128.c",
1461 "src/f32-igemm/gen/4x8s4-minmax-neon.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001462 "src/f32-igemm/gen/6x8-minmax-neon-dup-ld64.c",
1463 "src/f32-igemm/gen/6x8-minmax-neon-dup-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001464 "src/f32-igemm/gen/6x8-minmax-neon-lane-ld64.c",
1465 "src/f32-igemm/gen/6x8-minmax-neon-lane-ld128.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001466 "src/f32-igemm/gen/6x8s4-minmax-neon.c",
1467 "src/f32-igemm/gen/8x8s4-minmax-neon.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001468 "src/f32-maxpool/9p8x-minmax-neon-c4.c",
1469 "src/f32-pavgpool/9p8x-minmax-neon-c4.c",
1470 "src/f32-pavgpool/9x-minmax-neon-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001471 "src/f32-ppmm/gen/4x8-minmax-neon.c",
1472 "src/f32-ppmm/gen/8x8-minmax-neon.c",
Frank Barcharda5316982020-07-23 13:19:28 -07001473 "src/f32-prelu/gen/neon-1x4.c",
1474 "src/f32-prelu/gen/neon-1x8.c",
1475 "src/f32-prelu/gen/neon-1x16.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001476 "src/f32-prelu/gen/neon-2x4.c",
1477 "src/f32-prelu/gen/neon-2x8.c",
Frank Barcharda5316982020-07-23 13:19:28 -07001478 "src/f32-prelu/gen/neon-2x16.c",
1479 "src/f32-prelu/gen/neon-4x4.c",
1480 "src/f32-prelu/gen/neon-4x8.c",
1481 "src/f32-prelu/gen/neon-4x16.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001482 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x4.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001483 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x8-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001484 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x8.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001485 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12-acc2.c",
1486 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001487 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001488 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16-acc2.c",
1489 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001490 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001491 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20-acc2.c",
1492 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001493 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20.c",
1494 "src/f32-raddstoreexpminusmax/gen/neon-p5-x4.c",
1495 "src/f32-raddstoreexpminusmax/gen/neon-p5-x8-acc2.c",
1496 "src/f32-raddstoreexpminusmax/gen/neon-p5-x8.c",
1497 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12-acc2.c",
1498 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12-acc3.c",
1499 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12.c",
1500 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16-acc2.c",
1501 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16-acc4.c",
1502 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16.c",
1503 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20-acc2.c",
1504 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20-acc5.c",
1505 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07001506 "src/f32-relu/gen/neon-x4.c",
1507 "src/f32-relu/gen/neon-x8.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001508 "src/f32-rmax/neon.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -08001509 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x4.c",
1510 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x8.c",
1511 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x12.c",
1512 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x16.c",
1513 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x20.c",
1514 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x24.c",
1515 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x4.c",
1516 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x8.c",
1517 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x12.c",
1518 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x16.c",
1519 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x20.c",
1520 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x24.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001521 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x4.c",
1522 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x8.c",
1523 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x12.c",
1524 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x16.c",
1525 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x20.c",
1526 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x24.c",
Marat Dukhan5b86c432020-12-06 19:15:03 -08001527 "src/f32-spmm/gen/4x1-minmax-neon-pipelined.c",
1528 "src/f32-spmm/gen/4x1-minmax-neon-x2.c",
1529 "src/f32-spmm/gen/4x1-minmax-neon.c",
1530 "src/f32-spmm/gen/8x1-minmax-neon-pipelined.c",
1531 "src/f32-spmm/gen/8x1-minmax-neon-x2.c",
1532 "src/f32-spmm/gen/8x1-minmax-neon.c",
1533 "src/f32-spmm/gen/12x1-minmax-neon.c",
1534 "src/f32-spmm/gen/16x1-minmax-neon-pipelined.c",
1535 "src/f32-spmm/gen/16x1-minmax-neon-x2.c",
1536 "src/f32-spmm/gen/16x1-minmax-neon.c",
1537 "src/f32-spmm/gen/32x1-minmax-neon-pipelined.c",
1538 "src/f32-spmm/gen/32x1-minmax-neon-x2.c",
1539 "src/f32-spmm/gen/32x1-minmax-neon.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001540 "src/f32-vbinary/gen/vadd-minmax-neon-x4.c",
1541 "src/f32-vbinary/gen/vadd-minmax-neon-x8.c",
1542 "src/f32-vbinary/gen/vaddc-minmax-neon-x4.c",
1543 "src/f32-vbinary/gen/vaddc-minmax-neon-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -08001544 "src/f32-vbinary/gen/vmax-neon-x4.c",
1545 "src/f32-vbinary/gen/vmax-neon-x8.c",
1546 "src/f32-vbinary/gen/vmaxc-neon-x4.c",
1547 "src/f32-vbinary/gen/vmaxc-neon-x8.c",
1548 "src/f32-vbinary/gen/vmin-neon-x4.c",
1549 "src/f32-vbinary/gen/vmin-neon-x8.c",
1550 "src/f32-vbinary/gen/vminc-neon-x4.c",
1551 "src/f32-vbinary/gen/vminc-neon-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001552 "src/f32-vbinary/gen/vmul-minmax-neon-x4.c",
1553 "src/f32-vbinary/gen/vmul-minmax-neon-x8.c",
1554 "src/f32-vbinary/gen/vmulc-minmax-neon-x4.c",
1555 "src/f32-vbinary/gen/vmulc-minmax-neon-x8.c",
1556 "src/f32-vbinary/gen/vrsubc-minmax-neon-x4.c",
1557 "src/f32-vbinary/gen/vrsubc-minmax-neon-x8.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07001558 "src/f32-vbinary/gen/vsqrdiff-neon-x4.c",
1559 "src/f32-vbinary/gen/vsqrdiff-neon-x8.c",
1560 "src/f32-vbinary/gen/vsqrdiffc-neon-x4.c",
1561 "src/f32-vbinary/gen/vsqrdiffc-neon-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001562 "src/f32-vbinary/gen/vsub-minmax-neon-x4.c",
1563 "src/f32-vbinary/gen/vsub-minmax-neon-x8.c",
1564 "src/f32-vbinary/gen/vsubc-minmax-neon-x4.c",
1565 "src/f32-vbinary/gen/vsubc-minmax-neon-x8.c",
Marat Dukhaned6baaf2020-12-01 15:07:08 -08001566 "src/f32-velu/gen/velu-neon-rr2-lut16-p3-x4.c",
1567 "src/f32-velu/gen/velu-neon-rr2-lut16-p3-x8.c",
1568 "src/f32-velu/gen/velu-neon-rr2-lut16-p3-x12.c",
1569 "src/f32-velu/gen/velu-neon-rr2-lut16-p3-x16.c",
1570 "src/f32-velu/gen/velu-neon-rr2-lut16-p3-x20.c",
1571 "src/f32-velu/gen/velu-neon-rr2-lut16-p3-x24.c",
1572 "src/f32-velu/gen/velu-neon-rr2-p6-x4.c",
1573 "src/f32-velu/gen/velu-neon-rr2-p6-x8.c",
1574 "src/f32-velu/gen/velu-neon-rr2-p6-x12.c",
1575 "src/f32-velu/gen/velu-neon-rr2-p6-x16.c",
1576 "src/f32-velu/gen/velu-neon-rr2-p6-x20.c",
1577 "src/f32-velu/gen/velu-neon-rr2-p6-x24.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07001578 "src/f32-vlrelu/gen/vlrelu-neon-x4.c",
1579 "src/f32-vlrelu/gen/vlrelu-neon-x8.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001580 "src/f32-vmulcaddc/gen/c4-minmax-neon-2x.c",
1581 "src/f32-vmulcaddc/gen/c8-minmax-neon-2x.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07001582 "src/f32-vrnd/gen/vrndd-neon-x4.c",
1583 "src/f32-vrnd/gen/vrndd-neon-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001584 "src/f32-vrnd/gen/vrndne-neon-x4.c",
1585 "src/f32-vrnd/gen/vrndne-neon-x8.c",
1586 "src/f32-vrnd/gen/vrndu-neon-x4.c",
1587 "src/f32-vrnd/gen/vrndu-neon-x8.c",
1588 "src/f32-vrnd/gen/vrndz-neon-x4.c",
1589 "src/f32-vrnd/gen/vrndz-neon-x8.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07001590 "src/f32-vunary/gen/vabs-neon-x4.c",
1591 "src/f32-vunary/gen/vabs-neon-x8.c",
1592 "src/f32-vunary/gen/vneg-neon-x4.c",
1593 "src/f32-vunary/gen/vneg-neon-x8.c",
1594 "src/f32-vunary/gen/vsqr-neon-x4.c",
1595 "src/f32-vunary/gen/vsqr-neon-x8.c",
Marat Dukhande390d42020-11-29 19:32:18 -08001596 "src/math/expm1minus-neon-rr2-lut16-p3.c",
1597 "src/math/expm1minus-neon-rr2-p6.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001598 "src/math/roundd-neon-addsub.c",
1599 "src/math/roundd-neon-cvt.c",
1600 "src/math/roundne-neon-addsub.c",
1601 "src/math/roundu-neon-addsub.c",
1602 "src/math/roundu-neon-cvt.c",
1603 "src/math/roundz-neon-addsub.c",
1604 "src/math/roundz-neon-cvt.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001605 "src/math/sigmoid-neon-rr2-lut64-p2-nr2recps.c",
1606 "src/math/sigmoid-neon-rr2-lut2048-p1-nr2recps.c",
1607 "src/math/sigmoid-neon-rr2-p5-nr2recps.c",
1608 "src/math/sqrt-neon-nr1rsqrts.c",
1609 "src/math/sqrt-neon-nr2rsqrts.c",
1610 "src/math/sqrt-neon-nr3rsqrts.c",
Marat Dukhan023bcf92020-08-10 12:40:50 -07001611 "src/qs8-dwconv/gen/up8x9-minmax-neon-mul16.c",
1612 "src/qs8-dwconv/gen/up16x9-minmax-neon-mul16.c",
1613 "src/qs8-dwconv/gen/up24x9-minmax-neon-mul16.c",
1614 "src/qs8-dwconv/gen/up32x9-minmax-neon-mul16.c",
Marat Dukhan281262d2020-08-10 13:23:21 -07001615 "src/qs8-gavgpool/gen/7p7x-minmax-neon-c8-acc2.c",
1616 "src/qs8-gavgpool/gen/7p7x-minmax-neon-c16-acc2.c",
1617 "src/qs8-gavgpool/gen/7p7x-minmax-neon-c24-acc2.c",
1618 "src/qs8-gavgpool/gen/7p7x-minmax-neon-c32-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001619 "src/qs8-gavgpool/gen/7x-minmax-neon-c8-acc2.c",
1620 "src/qs8-gavgpool/gen/7x-minmax-neon-c16-acc2.c",
1621 "src/qs8-gavgpool/gen/7x-minmax-neon-c24-acc2.c",
1622 "src/qs8-gavgpool/gen/7x-minmax-neon-c32-acc2.c",
Marat Dukhan40bbafe2020-08-04 02:04:22 -07001623 "src/qs8-gemm/gen/1x8-minmax-neon-mlal-lane.c",
Marat Dukhan40bbafe2020-08-04 02:04:22 -07001624 "src/qs8-gemm/gen/1x16-minmax-neon-mlal-lane.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001625 "src/qs8-gemm/gen/2x8-minmax-neon-mlal-lane.c",
Marat Dukhan40bbafe2020-08-04 02:04:22 -07001626 "src/qs8-gemm/gen/2x16-minmax-neon-mlal-lane.c",
1627 "src/qs8-igemm/gen/1x8-minmax-neon-mlal-lane.c",
Marat Dukhan40bbafe2020-08-04 02:04:22 -07001628 "src/qs8-igemm/gen/1x16-minmax-neon-mlal-lane.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001629 "src/qs8-igemm/gen/2x8-minmax-neon-mlal-lane.c",
Marat Dukhan40bbafe2020-08-04 02:04:22 -07001630 "src/qs8-igemm/gen/2x16-minmax-neon-mlal-lane.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -07001631 "src/qs8-requantization/fp32-neon.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001632 "src/qs8-requantization/precise-neon.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -07001633 "src/qs8-requantization/q31-neon.c",
Marat Dukhanba7b2792020-09-02 14:26:45 -07001634 "src/qs8-vadd/gen/minmax-neon-ld64-x8.c",
1635 "src/qs8-vadd/gen/minmax-neon-ld64-x16.c",
1636 "src/qs8-vadd/gen/minmax-neon-ld64-x24.c",
1637 "src/qs8-vadd/gen/minmax-neon-ld64-x32.c",
1638 "src/qs8-vaddc/gen/minmax-neon-ld64-x8.c",
1639 "src/qs8-vaddc/gen/minmax-neon-ld64-x16.c",
1640 "src/qs8-vaddc/gen/minmax-neon-ld64-x24.c",
1641 "src/qs8-vaddc/gen/minmax-neon-ld64-x32.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -07001642 "src/qu8-avgpool/9p8x-minmax-neon-c8.c",
1643 "src/qu8-avgpool/9x-minmax-neon-c8.c",
1644 "src/qu8-dwconv/up8x9-minmax-neon.c",
1645 "src/qu8-gavgpool/7p7x-minmax-neon-c8.c",
1646 "src/qu8-gavgpool/7x-minmax-neon-c8.c",
1647 "src/qu8-gemm/4x8-minmax-neon.c",
1648 "src/qu8-gemm/8x8-minmax-neon.c",
1649 "src/qu8-igemm/4x8-minmax-neon.c",
1650 "src/qu8-igemm/8x8-minmax-neon.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07001651 "src/qu8-requantization/fp32-neon.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001652 "src/qu8-requantization/precise-neon.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07001653 "src/qu8-requantization/q31-neon.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -07001654 "src/qu8-vadd/minmax-neon.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001655 "src/u8-clamp/neon-x64.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001656 "src/u8-maxpool/9p8x-minmax-neon-c16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001657 "src/u8-rmax/neon.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001658 "src/x8-zip/x2-neon.c",
1659 "src/x8-zip/x3-neon.c",
1660 "src/x8-zip/x4-neon.c",
1661 "src/x8-zip/xm-neon.c",
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -07001662 "src/x32-fill/neon.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001663 "src/x32-packx/x4-neon-st4.c",
Marat Dukhan63523d42020-05-22 17:07:33 -07001664 "src/x32-pad/neon.c",
Marat Dukhan57dccd82020-04-14 00:53:10 -07001665 "src/x32-unpool/neon.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001666 "src/x32-zip/x2-neon.c",
1667 "src/x32-zip/x3-neon.c",
1668 "src/x32-zip/x4-neon.c",
1669 "src/x32-zip/xm-neon.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001670]
1671
1672NEONFMA_UKERNELS = [
Frank Barchard04336c12020-10-22 16:48:55 -07001673 "src/f32-dwconv/gen/up4x4-minmax-neonfma-acc2.c",
1674 "src/f32-dwconv/gen/up4x4-minmax-neonfma.c",
1675 "src/f32-dwconv/gen/up4x9-minmax-neonfma-acc2.c",
1676 "src/f32-dwconv/gen/up4x9-minmax-neonfma.c",
1677 "src/f32-dwconv/gen/up4x25-minmax-neonfma-acc2.c",
1678 "src/f32-dwconv/gen/up4x25-minmax-neonfma.c",
1679 "src/f32-dwconv/gen/up8x4-minmax-neonfma-acc2.c",
1680 "src/f32-dwconv/gen/up8x4-minmax-neonfma.c",
1681 "src/f32-dwconv/gen/up8x9-minmax-neonfma-acc2.c",
1682 "src/f32-dwconv/gen/up8x9-minmax-neonfma.c",
1683 "src/f32-dwconv/gen/up8x25-minmax-neonfma-acc2.c",
1684 "src/f32-dwconv/gen/up8x25-minmax-neonfma.c",
1685 "src/f32-gemm/gen-inc/1x8inc-minmax-neonfma-dup-ld64.c",
1686 "src/f32-gemm/gen-inc/1x8s4inc-minmax-neonfma.c",
1687 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-dup-ld64.c",
1688 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-dup-ld128.c",
1689 "src/f32-gemm/gen-inc/4x8s4inc-minmax-neonfma.c",
1690 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-dup-ld64.c",
1691 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-dup-ld128.c",
1692 "src/f32-gemm/gen-inc/6x8s4inc-minmax-neonfma.c",
1693 "src/f32-gemm/gen-inc/8x8s4inc-minmax-neonfma.c",
1694 "src/f32-gemm/gen/1x8-minmax-neonfma-dup-ld64.c",
1695 "src/f32-gemm/gen/1x8s4-minmax-neonfma.c",
1696 "src/f32-gemm/gen/4x8-minmax-neonfma-dup-ld64.c",
1697 "src/f32-gemm/gen/4x8-minmax-neonfma-dup-ld128.c",
1698 "src/f32-gemm/gen/4x8s4-minmax-neonfma.c",
1699 "src/f32-gemm/gen/6x8-minmax-neonfma-dup-ld64.c",
1700 "src/f32-gemm/gen/6x8-minmax-neonfma-dup-ld128.c",
1701 "src/f32-gemm/gen/6x8s4-minmax-neonfma.c",
1702 "src/f32-gemm/gen/8x8s4-minmax-neonfma.c",
Marat Dukhan660fd192020-03-10 04:55:30 -07001703 "src/f32-ibilinear/gen/neonfma-c4.c",
1704 "src/f32-ibilinear/gen/neonfma-c8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001705 "src/f32-igemm/gen/1x8-minmax-neonfma-dup-ld64.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001706 "src/f32-igemm/gen/1x8s4-minmax-neonfma.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001707 "src/f32-igemm/gen/4x8-minmax-neonfma-dup-ld64.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001708 "src/f32-igemm/gen/4x8-minmax-neonfma-dup-ld128.c",
1709 "src/f32-igemm/gen/4x8s4-minmax-neonfma.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001710 "src/f32-igemm/gen/6x8-minmax-neonfma-dup-ld64.c",
1711 "src/f32-igemm/gen/6x8-minmax-neonfma-dup-ld128.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001712 "src/f32-igemm/gen/6x8s4-minmax-neonfma.c",
1713 "src/f32-igemm/gen/8x8s4-minmax-neonfma.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001714 "src/f32-ppmm/gen/4x8-minmax-neonfma.c",
1715 "src/f32-ppmm/gen/8x8-minmax-neonfma.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001716 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x4.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001717 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x8-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001718 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x8.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001719 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12-acc2.c",
1720 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001721 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001722 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16-acc2.c",
1723 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001724 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001725 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20-acc2.c",
1726 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001727 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20.c",
1728 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x4.c",
1729 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x8-acc2.c",
1730 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x8.c",
1731 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12-acc2.c",
1732 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12-acc3.c",
1733 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12.c",
1734 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16-acc2.c",
1735 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16-acc4.c",
1736 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16.c",
1737 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20-acc2.c",
1738 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20-acc5.c",
1739 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -08001740 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x4.c",
1741 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x8.c",
1742 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x12.c",
1743 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x16.c",
1744 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x20.c",
1745 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x24.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001746 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x4.c",
1747 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x8.c",
1748 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x12.c",
1749 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x16.c",
1750 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x20.c",
1751 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x24.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -08001752 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x4.c",
1753 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x8.c",
1754 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x12.c",
1755 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x16.c",
1756 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x20.c",
1757 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x24.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -08001758 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x4.c",
1759 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x8.c",
1760 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x12.c",
1761 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x16.c",
1762 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x20.c",
1763 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x24.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001764 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x4.c",
1765 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x8.c",
1766 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x12.c",
1767 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x16.c",
1768 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x20.c",
1769 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x24.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -08001770 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x4.c",
1771 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x8.c",
1772 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x12.c",
1773 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x16.c",
1774 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x20.c",
1775 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x24.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001776 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x4.c",
1777 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x8.c",
1778 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x12.c",
1779 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x16.c",
1780 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x20.c",
1781 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x24.c",
1782 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x4.c",
1783 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x8.c",
1784 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x12.c",
1785 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x16.c",
1786 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x20.c",
1787 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x24.c",
1788 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x4.c",
1789 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x8.c",
1790 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x12.c",
1791 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x16.c",
1792 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x20.c",
1793 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x24.c",
Marat Dukhan2fa7a0c2020-12-06 19:09:02 -08001794 "src/f32-spmm/gen/4x1-minmax-neonfma-pipelined.c",
1795 "src/f32-spmm/gen/4x1-minmax-neonfma-x2.c",
1796 "src/f32-spmm/gen/4x1-minmax-neonfma.c",
1797 "src/f32-spmm/gen/8x1-minmax-neonfma-pipelined.c",
1798 "src/f32-spmm/gen/8x1-minmax-neonfma-x2.c",
1799 "src/f32-spmm/gen/8x1-minmax-neonfma.c",
1800 "src/f32-spmm/gen/12x1-minmax-neonfma.c",
1801 "src/f32-spmm/gen/16x1-minmax-neonfma-pipelined.c",
1802 "src/f32-spmm/gen/16x1-minmax-neonfma-x2.c",
1803 "src/f32-spmm/gen/16x1-minmax-neonfma.c",
1804 "src/f32-spmm/gen/32x1-minmax-neonfma-pipelined.c",
1805 "src/f32-spmm/gen/32x1-minmax-neonfma-x2.c",
1806 "src/f32-spmm/gen/32x1-minmax-neonfma.c",
Marat Dukhaned6baaf2020-12-01 15:07:08 -08001807 "src/f32-velu/gen/velu-neonfma-rr1-lut16-p3-x4.c",
1808 "src/f32-velu/gen/velu-neonfma-rr1-lut16-p3-x8.c",
1809 "src/f32-velu/gen/velu-neonfma-rr1-lut16-p3-x12.c",
1810 "src/f32-velu/gen/velu-neonfma-rr1-lut16-p3-x16.c",
1811 "src/f32-velu/gen/velu-neonfma-rr1-lut16-p3-x20.c",
1812 "src/f32-velu/gen/velu-neonfma-rr1-lut16-p3-x24.c",
1813 "src/f32-velu/gen/velu-neonfma-rr1-p6-x4.c",
1814 "src/f32-velu/gen/velu-neonfma-rr1-p6-x8.c",
1815 "src/f32-velu/gen/velu-neonfma-rr1-p6-x12.c",
1816 "src/f32-velu/gen/velu-neonfma-rr1-p6-x16.c",
1817 "src/f32-velu/gen/velu-neonfma-rr1-p6-x20.c",
1818 "src/f32-velu/gen/velu-neonfma-rr1-p6-x24.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001819 "src/f32-vmulcaddc/gen/c4-minmax-neonfma-2x.c",
1820 "src/f32-vmulcaddc/gen/c8-minmax-neonfma-2x.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07001821 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x4.c",
1822 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x8.c",
1823 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x12.c",
1824 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x16.c",
1825 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x20.c",
1826 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x24.c",
1827 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x28.c",
1828 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x32.c",
1829 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x36.c",
1830 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x40.c",
1831 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x4.c",
1832 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x8.c",
1833 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x12.c",
1834 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x16.c",
1835 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x20.c",
1836 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x24.c",
1837 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x28.c",
1838 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x32.c",
1839 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x36.c",
1840 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x40.c",
Marat Dukhanb7633f22020-11-20 16:34:56 -08001841 "src/math/exp-neonfma-rr2-lut64-p2.c",
1842 "src/math/exp-neonfma-rr2-p5.c",
Frank Barcharde7223ee2020-12-04 19:04:01 -08001843 "src/math/expm1minus-neonfma-rr1-lut16-p3.c",
1844 "src/math/expm1minus-neonfma-rr1-p6.c",
Marat Dukhan9dd119a2020-11-20 18:20:04 -08001845 "src/math/expminus-neonfma-rr2-lut64-p2.c",
1846 "src/math/expminus-neonfma-rr2-lut2048-p1.c",
1847 "src/math/expminus-neonfma-rr2-p5.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001848 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr1recps1fma.c",
1849 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr2fma.c",
1850 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr2recps.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001851 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr1recps1fma.c",
1852 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr2fma.c",
1853 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr2recps.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001854 "src/math/sigmoid-neonfma-rr1-p5-nr1recps1fma.c",
1855 "src/math/sigmoid-neonfma-rr1-p5-nr2fma.c",
1856 "src/math/sigmoid-neonfma-rr1-p5-nr2recps.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001857 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr1recps1fma.c",
1858 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr2fma.c",
1859 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr2recps.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001860 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr1recps1fma.c",
1861 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr2fma.c",
1862 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr2recps.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001863 "src/math/sigmoid-neonfma-rr2-p5-nr1recps1fma.c",
1864 "src/math/sigmoid-neonfma-rr2-p5-nr2fma.c",
1865 "src/math/sigmoid-neonfma-rr2-p5-nr2recps.c",
Marat Dukhan84000762020-06-29 18:38:43 -07001866 "src/math/sqrt-neonfma-nr1fma.c",
Marat Dukhan84000762020-06-29 18:38:43 -07001867 "src/math/sqrt-neonfma-nr1rsqrts1fma1adj.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001868 "src/math/sqrt-neonfma-nr2fma.c",
1869 "src/math/sqrt-neonfma-nr2fma1adj.c",
1870 "src/math/sqrt-neonfma-nr3fma.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001871]
1872
1873AARCH64_NEONFMA_UKERNELS = [
Marat Dukhan56b10cd2020-05-18 09:35:49 -07001874 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neonfma-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001875 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neonfma-2x2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001876 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neonfma-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001877 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neonfma-2x2.c",
Marat Dukhan56b10cd2020-05-18 09:35:49 -07001878 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neonfma-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001879 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neonfma-2x2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001880 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neonfma-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001881 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neonfma-2x2.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -07001882 "src/f32-conv-hwc2chw/3x3s2p1c3x4-neonfma-2x2.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001883 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-1x4-acc2.c",
1884 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-1x4-acc3.c",
1885 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-1x4-acc4.c",
Marat Dukhan1268a242020-10-24 00:36:32 -07001886 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-1x4.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001887 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-2x4-acc2.c",
Marat Dukhan1268a242020-10-24 00:36:32 -07001888 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-2x4.c",
1889 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-3x4.c",
1890 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-4x4.c",
1891 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-5x4.c",
1892 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-6x4.c",
Marat Dukhan82f0c322020-10-25 19:17:35 -07001893 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neonfma-1x4-acc2.c",
1894 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neonfma-1x4-acc3.c",
1895 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neonfma-1x4-acc4.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001896 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neonfma-1x4.c",
Marat Dukhan82f0c322020-10-25 19:17:35 -07001897 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neonfma-2x4-acc2.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001898 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neonfma-2x4.c",
1899 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neonfma-3x4.c",
1900 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neonfma-4x4.c",
Marat Dukhan149f0ea2020-10-26 12:50:33 -07001901 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-1x4-acc2.c",
1902 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-1x4-acc3.c",
1903 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-1x4-acc4.c",
1904 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-1x4-acc5.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001905 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-1x4.c",
Marat Dukhan149f0ea2020-10-26 12:50:33 -07001906 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-2x4-acc2.c",
1907 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-2x4-acc3.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001908 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-2x4.c",
Marat Dukhan149f0ea2020-10-26 12:50:33 -07001909 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-3x4-acc2.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001910 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-3x4.c",
Marat Dukhan149f0ea2020-10-26 12:50:33 -07001911 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-4x4-acc2.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001912 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-4x4.c",
1913 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-5x4.c",
Marat Dukhan30d4b252020-10-29 16:33:22 -07001914 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neonfma-1x4-acc2.c",
1915 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neonfma-1x4-acc3.c",
1916 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neonfma-1x4-acc4.c",
1917 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neonfma-1x4-acc5.c",
1918 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neonfma-1x4.c",
1919 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neonfma-2x4-acc2.c",
1920 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neonfma-2x4-acc3.c",
1921 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neonfma-2x4.c",
Marat Dukhan30d4b252020-10-29 16:33:22 -07001922 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neonfma-3x4-acc2.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07001923 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neonfma-3x4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001924 "src/f32-gemm/gen-inc/1x8inc-minmax-neonfma-lane-ld64.c",
1925 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-lane-ld64.c",
1926 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-lane-ld128.c",
1927 "src/f32-gemm/gen-inc/5x8inc-minmax-neonfma-lane-ld64.c",
1928 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-lane-ld64.c",
1929 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-lane-ld128.c",
1930 "src/f32-gemm/gen/1x8-minmax-neonfma-lane-ld64.c",
1931 "src/f32-gemm/gen/4x2-minmax-neonfma-lane-ld64.c",
1932 "src/f32-gemm/gen/4x8-minmax-neonfma-lane-ld64.c",
1933 "src/f32-gemm/gen/4x8-minmax-neonfma-lane-ld128.c",
1934 "src/f32-gemm/gen/5x8-minmax-neonfma-lane-ld64.c",
1935 "src/f32-gemm/gen/6x8-minmax-neonfma-lane-ld64.c",
1936 "src/f32-gemm/gen/6x8-minmax-neonfma-lane-ld128.c",
1937 "src/f32-igemm/gen/1x8-minmax-neonfma-lane-ld64.c",
1938 "src/f32-igemm/gen/4x2-minmax-neonfma-lane-ld64.c",
1939 "src/f32-igemm/gen/4x4-minmax-neonfma-lane-ld64.c",
1940 "src/f32-igemm/gen/4x8-minmax-neonfma-lane-ld64.c",
1941 "src/f32-igemm/gen/4x8-minmax-neonfma-lane-ld128.c",
1942 "src/f32-igemm/gen/6x8-minmax-neonfma-lane-ld64.c",
1943 "src/f32-igemm/gen/6x8-minmax-neonfma-lane-ld128.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -08001944 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x4.c",
1945 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x8.c",
1946 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x12.c",
1947 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x16.c",
1948 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x20.c",
1949 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x24.c",
1950 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x4.c",
1951 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x8.c",
1952 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x12.c",
1953 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x16.c",
1954 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x20.c",
1955 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x24.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001956 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x4.c",
1957 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x8.c",
1958 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x12.c",
1959 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x16.c",
1960 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x20.c",
1961 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x24.c",
Marat Dukhan355ab432020-04-09 19:01:52 -07001962 "src/f32-spmm/gen/4x2-minmax-neonfma.c",
1963 "src/f32-spmm/gen/4x4-minmax-neonfma.c",
Marat Dukhan355ab432020-04-09 19:01:52 -07001964 "src/f32-spmm/gen/8x2-minmax-neonfma.c",
1965 "src/f32-spmm/gen/8x4-minmax-neonfma.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001966 "src/f32-spmm/gen/12x2-minmax-neonfma.c",
1967 "src/f32-spmm/gen/12x4-minmax-neonfma.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001968 "src/f32-spmm/gen/16x2-minmax-neonfma.c",
1969 "src/f32-spmm/gen/16x4-minmax-neonfma.c",
Frank Barchard846c0c62020-10-26 15:01:39 -07001970 "src/f32-spmm/gen/32x2-minmax-neonfma.c",
1971 "src/f32-spmm/gen/32x4-minmax-neonfma.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001972 "src/f32-vbinary/gen/vdiv-minmax-neon-x4.c",
1973 "src/f32-vbinary/gen/vdiv-minmax-neon-x8.c",
1974 "src/f32-vbinary/gen/vdivc-minmax-neon-x4.c",
1975 "src/f32-vbinary/gen/vdivc-minmax-neon-x8.c",
1976 "src/f32-vbinary/gen/vrdivc-minmax-neon-x4.c",
1977 "src/f32-vbinary/gen/vrdivc-minmax-neon-x8.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07001978 "src/f32-vsqrt/gen/neon-sqrt-x4.c",
1979 "src/f32-vsqrt/gen/neon-sqrt-x8.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001980 "src/math/sigmoid-neonfma-rr1-lut64-p2-div.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001981 "src/math/sigmoid-neonfma-rr1-lut2048-p1-div.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001982 "src/math/sigmoid-neonfma-rr1-p5-div.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001983 "src/math/sigmoid-neonfma-rr2-lut64-p2-div.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001984 "src/math/sigmoid-neonfma-rr2-lut2048-p1-div.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001985 "src/math/sigmoid-neonfma-rr2-p5-div.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001986]
1987
Marat Dukhan8853b822020-05-07 12:19:01 -07001988NEONV8_UKERNELS = [
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07001989 "src/f32-vrnd/gen/vrndd-neonv8-x4.c",
1990 "src/f32-vrnd/gen/vrndd-neonv8-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001991 "src/f32-vrnd/gen/vrndne-neonv8-x4.c",
1992 "src/f32-vrnd/gen/vrndne-neonv8-x8.c",
1993 "src/f32-vrnd/gen/vrndu-neonv8-x4.c",
1994 "src/f32-vrnd/gen/vrndu-neonv8-x8.c",
1995 "src/f32-vrnd/gen/vrndz-neonv8-x4.c",
1996 "src/f32-vrnd/gen/vrndz-neonv8-x8.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001997 "src/math/roundd-neonv8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001998 "src/math/roundne-neonv8.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001999 "src/math/roundu-neonv8.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07002000 "src/math/roundz-neonv8.c",
Marat Dukhan8853b822020-05-07 12:19:01 -07002001]
2002
Marat Dukhan08c4a432019-10-03 09:29:21 -07002003AARCH64_NEONFP16ARITH_UKERNELS = [
Frank Barchardb1966592020-05-12 13:47:06 -07002004 "src/f16-clamp/gen/neonfp16arith-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002005 "src/f16-clamp/gen/neonfp16arith-x16.c",
Frank Barchard5a599a62020-06-04 20:12:44 -07002006 "src/f16-dwconv/gen/up8x4-minmax-neonfp16arith-acc2.c",
2007 "src/f16-dwconv/gen/up8x4-minmax-neonfp16arith.c",
2008 "src/f16-dwconv/gen/up8x9-minmax-neonfp16arith-acc2.c",
2009 "src/f16-dwconv/gen/up8x9-minmax-neonfp16arith.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002010 "src/f16-dwconv/gen/up8x25-minmax-neonfp16arith-acc2.c",
2011 "src/f16-dwconv/gen/up8x25-minmax-neonfp16arith.c",
2012 "src/f16-dwconv/gen/up16x4-minmax-neonfp16arith-acc2.c",
2013 "src/f16-dwconv/gen/up16x4-minmax-neonfp16arith.c",
2014 "src/f16-dwconv/gen/up16x9-minmax-neonfp16arith-acc2.c",
2015 "src/f16-dwconv/gen/up16x9-minmax-neonfp16arith.c",
2016 "src/f16-dwconv/gen/up16x25-minmax-neonfp16arith-acc2.c",
2017 "src/f16-dwconv/gen/up16x25-minmax-neonfp16arith.c",
Frank Barchard0bb49a72020-06-04 11:35:11 -07002018 "src/f16-gavgpool/7p7x-minmax-neonfp16arith-c8.c",
2019 "src/f16-gavgpool/7x-minmax-neonfp16arith-c8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002020 "src/f16-gemm/gen-inc/1x8inc-minmax-neonfp16arith-ld64.c",
2021 "src/f16-gemm/gen-inc/1x16inc-minmax-neonfp16arith-ld64.c",
2022 "src/f16-gemm/gen-inc/4x8inc-minmax-neonfp16arith-ld64.c",
2023 "src/f16-gemm/gen-inc/4x16inc-minmax-neonfp16arith-ld64.c",
2024 "src/f16-gemm/gen-inc/6x8inc-minmax-neonfp16arith-ld64.c",
2025 "src/f16-gemm/gen-inc/6x16inc-minmax-neonfp16arith-ld64.c",
2026 "src/f16-gemm/gen-inc/8x8inc-minmax-neonfp16arith-ld64.c",
2027 "src/f16-gemm/gen-inc/8x16inc-minmax-neonfp16arith-ld64.c",
2028 "src/f16-gemm/gen/1x8-minmax-neonfp16arith-ld64.c",
2029 "src/f16-gemm/gen/1x16-minmax-neonfp16arith-ld64.c",
2030 "src/f16-gemm/gen/4x8-minmax-neonfp16arith-ld64.c",
2031 "src/f16-gemm/gen/4x16-minmax-neonfp16arith-ld64.c",
2032 "src/f16-gemm/gen/6x8-minmax-neonfp16arith-ld64.c",
2033 "src/f16-gemm/gen/6x16-minmax-neonfp16arith-ld64.c",
2034 "src/f16-gemm/gen/8x8-minmax-neonfp16arith-ld64.c",
2035 "src/f16-gemm/gen/8x16-minmax-neonfp16arith-ld64.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07002036 "src/f16-hswish/gen/hswish-neonfp16arith-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002037 "src/f16-hswish/gen/hswish-neonfp16arith-x16.c",
2038 "src/f16-igemm/gen/1x8-minmax-neonfp16arith-ld64.c",
2039 "src/f16-igemm/gen/1x16-minmax-neonfp16arith-ld64.c",
2040 "src/f16-igemm/gen/4x8-minmax-neonfp16arith-ld64.c",
2041 "src/f16-igemm/gen/4x16-minmax-neonfp16arith-ld64.c",
2042 "src/f16-igemm/gen/6x8-minmax-neonfp16arith-ld64.c",
2043 "src/f16-igemm/gen/6x16-minmax-neonfp16arith-ld64.c",
2044 "src/f16-igemm/gen/8x8-minmax-neonfp16arith-ld64.c",
2045 "src/f16-igemm/gen/8x16-minmax-neonfp16arith-ld64.c",
Frank Barchardb1966592020-05-12 13:47:06 -07002046 "src/f16-prelu/gen/neonfp16arith-2x8.c",
Frank Barcharda5316982020-07-23 13:19:28 -07002047 "src/f16-prelu/gen/neonfp16arith-2x16.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07002048 "src/f16-relu/gen/neonfp16arith-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002049 "src/f16-relu/gen/neonfp16arith-x16.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07002050 "src/f16-spmm/gen/8x1-minmax-neonfp16arith-x2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002051 "src/f16-spmm/gen/8x1-minmax-neonfp16arith.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07002052 "src/f16-spmm/gen/16x1-minmax-neonfp16arith-x2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002053 "src/f16-spmm/gen/16x1-minmax-neonfp16arith.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07002054 "src/f16-spmm/gen/24x1-minmax-neonfp16arith-x2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002055 "src/f16-spmm/gen/24x1-minmax-neonfp16arith.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07002056 "src/f16-spmm/gen/32x1-minmax-neonfp16arith-x2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002057 "src/f16-spmm/gen/32x1-minmax-neonfp16arith.c",
2058 "src/f16-vbinary/gen/vadd-minmax-neonfp16arith-x8.c",
2059 "src/f16-vbinary/gen/vadd-minmax-neonfp16arith-x16.c",
2060 "src/f16-vbinary/gen/vaddc-minmax-neonfp16arith-x8.c",
2061 "src/f16-vbinary/gen/vaddc-minmax-neonfp16arith-x16.c",
2062 "src/f16-vbinary/gen/vdiv-minmax-neonfp16arith-x8.c",
2063 "src/f16-vbinary/gen/vdiv-minmax-neonfp16arith-x16.c",
2064 "src/f16-vbinary/gen/vdivc-minmax-neonfp16arith-x8.c",
2065 "src/f16-vbinary/gen/vdivc-minmax-neonfp16arith-x16.c",
2066 "src/f16-vbinary/gen/vmax-neonfp16arith-x8.c",
2067 "src/f16-vbinary/gen/vmax-neonfp16arith-x16.c",
2068 "src/f16-vbinary/gen/vmaxc-neonfp16arith-x8.c",
2069 "src/f16-vbinary/gen/vmaxc-neonfp16arith-x16.c",
2070 "src/f16-vbinary/gen/vmin-neonfp16arith-x8.c",
2071 "src/f16-vbinary/gen/vmin-neonfp16arith-x16.c",
2072 "src/f16-vbinary/gen/vminc-neonfp16arith-x8.c",
2073 "src/f16-vbinary/gen/vminc-neonfp16arith-x16.c",
2074 "src/f16-vbinary/gen/vmul-minmax-neonfp16arith-x8.c",
2075 "src/f16-vbinary/gen/vmul-minmax-neonfp16arith-x16.c",
2076 "src/f16-vbinary/gen/vmulc-minmax-neonfp16arith-x8.c",
2077 "src/f16-vbinary/gen/vmulc-minmax-neonfp16arith-x16.c",
2078 "src/f16-vbinary/gen/vrdivc-minmax-neonfp16arith-x8.c",
2079 "src/f16-vbinary/gen/vrdivc-minmax-neonfp16arith-x16.c",
2080 "src/f16-vbinary/gen/vrsubc-minmax-neonfp16arith-x8.c",
2081 "src/f16-vbinary/gen/vrsubc-minmax-neonfp16arith-x16.c",
2082 "src/f16-vbinary/gen/vsub-minmax-neonfp16arith-x8.c",
2083 "src/f16-vbinary/gen/vsub-minmax-neonfp16arith-x16.c",
2084 "src/f16-vbinary/gen/vsubc-minmax-neonfp16arith-x8.c",
2085 "src/f16-vbinary/gen/vsubc-minmax-neonfp16arith-x16.c",
2086 "src/f16-vmulcaddc/gen/c8-minmax-neonfp16arith-2x.c",
2087 "src/f16-vmulcaddc/gen/c16-minmax-neonfp16arith-2x.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002088]
2089
Benoit Jacoba9644732020-08-13 12:48:55 -07002090NEONDOT_UKERNELS = [
Marat Dukhand4c83032020-09-09 23:20:12 -07002091 "src/qs8-gemm/gen/1x8c4-minmax-neondot.c",
Marat Dukhand4c83032020-09-09 23:20:12 -07002092 "src/qs8-gemm/gen/1x16c4-minmax-neondot.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002093 "src/qs8-gemm/gen/4x8c4-minmax-neondot.c",
Benoit Jacoba9644732020-08-13 12:48:55 -07002094 "src/qs8-gemm/gen/4x16c4-minmax-neondot.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002095 "src/qs8-gemm/gen/6x8c4-minmax-neondot.c",
Frank Barchard2fa17452020-09-22 09:51:46 -07002096 "src/qs8-gemm/gen/6x16c4-minmax-neondot.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002097 "src/qs8-gemm/gen/8x8c4-minmax-neondot.c",
Frank Barcharda48848f2020-09-24 00:59:24 -07002098 "src/qs8-gemm/gen/8x16c4-minmax-neondot.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002099 "src/qs8-gemm/gen/12x8c4-minmax-neondot.c",
Marat Dukhand4c83032020-09-09 23:20:12 -07002100 "src/qs8-igemm/gen/1x8c4-minmax-neondot.c",
Marat Dukhand4c83032020-09-09 23:20:12 -07002101 "src/qs8-igemm/gen/1x16c4-minmax-neondot.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002102 "src/qs8-igemm/gen/4x8c4-minmax-neondot.c",
Benoit Jacoba05487f2020-08-14 13:37:21 -07002103 "src/qs8-igemm/gen/4x16c4-minmax-neondot.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002104 "src/qs8-igemm/gen/6x8c4-minmax-neondot.c",
Frank Barchard2fa17452020-09-22 09:51:46 -07002105 "src/qs8-igemm/gen/6x16c4-minmax-neondot.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002106 "src/qs8-igemm/gen/8x8c4-minmax-neondot.c",
Frank Barcharda48848f2020-09-24 00:59:24 -07002107 "src/qs8-igemm/gen/8x16c4-minmax-neondot.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002108 "src/qs8-igemm/gen/12x8c4-minmax-neondot.c",
Benoit Jacoba9644732020-08-13 12:48:55 -07002109]
2110
Marat Dukhan08c4a432019-10-03 09:29:21 -07002111SSE_UKERNELS = [
Marat Dukhan99936602020-04-11 16:47:01 -07002112 "src/f32-avgpool/9p8x-minmax-sse-c4.c",
2113 "src/f32-avgpool/9x-minmax-sse-c4.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -07002114 "src/f32-clamp/gen/sse-x4.c",
2115 "src/f32-clamp/gen/sse-x8.c",
Erich Elsenb1233402020-06-08 15:53:15 -07002116 "src/f32-conv-hwc2chw/3x3s2p1c3x4-sse-1x1.c",
2117 "src/f32-conv-hwc2chw/3x3s2p1c3x4-sse-2x2.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002118 "src/f32-dwconv/gen/up4x4-minmax-sse-acc2.c",
2119 "src/f32-dwconv/gen/up4x4-minmax-sse.c",
2120 "src/f32-dwconv/gen/up4x9-minmax-sse-acc2.c",
2121 "src/f32-dwconv/gen/up4x9-minmax-sse.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002122 "src/f32-dwconv/gen/up4x25-minmax-sse-acc2.c",
2123 "src/f32-dwconv/gen/up4x25-minmax-sse.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002124 "src/f32-dwconv/gen/up8x4-minmax-sse-acc2.c",
2125 "src/f32-dwconv/gen/up8x4-minmax-sse.c",
2126 "src/f32-dwconv/gen/up8x9-minmax-sse-acc2.c",
2127 "src/f32-dwconv/gen/up8x9-minmax-sse.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002128 "src/f32-dwconv/gen/up8x25-minmax-sse-acc2.c",
2129 "src/f32-dwconv/gen/up8x25-minmax-sse.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07002130 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-1x4-acc2.c",
2131 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-1x4-acc3.c",
2132 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-1x4-acc4.c",
Marat Dukhan470078a2020-10-23 22:36:52 -07002133 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-1x4.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07002134 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-2x4-acc2.c",
Marat Dukhan470078a2020-10-23 22:36:52 -07002135 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-2x4.c",
2136 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-3x4.c",
2137 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-4x4.c",
2138 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-5x4.c",
2139 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-6x4.c",
Marat Dukhan0ff97182020-10-25 19:14:03 -07002140 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-sse-1x4-acc2.c",
2141 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-sse-1x4-acc3.c",
2142 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-sse-1x4-acc4.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07002143 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-sse-1x4.c",
Marat Dukhan0ff97182020-10-25 19:14:03 -07002144 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-sse-2x4-acc2.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07002145 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-sse-2x4.c",
2146 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-sse-3x4.c",
2147 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-sse-4x4.c",
Marat Dukhand0503892020-10-30 08:22:04 -07002148 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-1x4-acc2.c",
2149 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-1x4-acc3.c",
2150 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-1x4-acc4.c",
2151 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-1x4-acc5.c",
2152 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-1x4.c",
2153 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-2x4-acc2.c",
2154 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-2x4-acc3.c",
2155 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-2x4.c",
2156 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-3x4-acc2.c",
2157 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-3x4.c",
2158 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-4x4-acc2.c",
2159 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-4x4.c",
2160 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-5x4.c",
Marat Dukhanccca2142020-10-30 17:32:45 -07002161 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-sse-1x4-acc2.c",
2162 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-sse-1x4-acc3.c",
2163 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-sse-1x4-acc4.c",
2164 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-sse-1x4-acc5.c",
2165 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-sse-1x4.c",
2166 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-sse-2x4-acc2.c",
2167 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-sse-2x4-acc3.c",
2168 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-sse-2x4.c",
Marat Dukhanccca2142020-10-30 17:32:45 -07002169 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-sse-3x4-acc2.c",
Frank Barchard8ef44cd2020-11-03 12:30:23 -08002170 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-sse-3x4.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -07002171 "src/f32-gavgpool-cw/sse-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -07002172 "src/f32-gavgpool/7p7x-minmax-sse-c4.c",
2173 "src/f32-gavgpool/7x-minmax-sse-c4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002174 "src/f32-gemm/gen-inc/1x8inc-minmax-sse-dup.c",
2175 "src/f32-gemm/gen-inc/1x8inc-minmax-sse-load1.c",
2176 "src/f32-gemm/gen-inc/1x8s4inc-minmax-sse.c",
2177 "src/f32-gemm/gen-inc/4x8inc-minmax-sse-dup.c",
2178 "src/f32-gemm/gen-inc/4x8inc-minmax-sse-load1.c",
2179 "src/f32-gemm/gen-inc/4x8s4inc-minmax-sse.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002180 "src/f32-gemm/gen/1x8-minmax-sse-dup.c",
2181 "src/f32-gemm/gen/1x8-minmax-sse-load1.c",
2182 "src/f32-gemm/gen/1x8s4-minmax-sse.c",
2183 "src/f32-gemm/gen/4x2c4-minmax-sse.c",
2184 "src/f32-gemm/gen/4x8-minmax-sse-dup.c",
2185 "src/f32-gemm/gen/4x8-minmax-sse-load1.c",
2186 "src/f32-gemm/gen/4x8s4-minmax-sse.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07002187 "src/f32-hswish/gen/hswish-sse-x4.c",
2188 "src/f32-hswish/gen/hswish-sse-x8.c",
Marat Dukhan660fd192020-03-10 04:55:30 -07002189 "src/f32-ibilinear/gen/sse-c4.c",
2190 "src/f32-ibilinear/gen/sse-c8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002191 "src/f32-igemm/gen/1x8-minmax-sse-dup.c",
2192 "src/f32-igemm/gen/1x8-minmax-sse-load1.c",
2193 "src/f32-igemm/gen/1x8s4-minmax-sse.c",
2194 "src/f32-igemm/gen/4x2c4-minmax-sse.c",
2195 "src/f32-igemm/gen/4x8-minmax-sse-dup.c",
2196 "src/f32-igemm/gen/4x8-minmax-sse-load1.c",
2197 "src/f32-igemm/gen/4x8s4-minmax-sse.c",
Marat Dukhan99936602020-04-11 16:47:01 -07002198 "src/f32-maxpool/9p8x-minmax-sse-c4.c",
2199 "src/f32-pavgpool/9p8x-minmax-sse-c4.c",
2200 "src/f32-pavgpool/9x-minmax-sse-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002201 "src/f32-ppmm/gen/4x8-minmax-sse.c",
Marat Dukhan39b5e942020-06-24 15:03:48 -07002202 "src/f32-prelu/gen/sse-2x4.c",
2203 "src/f32-prelu/gen/sse-2x8.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07002204 "src/f32-relu/gen/sse-x4.c",
2205 "src/f32-relu/gen/sse-x8.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002206 "src/f32-rmax/sse.c",
Marat Dukhan355ab432020-04-09 19:01:52 -07002207 "src/f32-spmm/gen/4x1-minmax-sse.c",
2208 "src/f32-spmm/gen/8x1-minmax-sse.c",
Erich Elsen6e80fdc2020-06-09 15:35:37 -07002209 "src/f32-spmm/gen/16x1-minmax-sse.c",
Frank Barchard846c0c62020-10-26 15:01:39 -07002210 "src/f32-spmm/gen/32x1-minmax-sse.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002211 "src/f32-vbinary/gen/vadd-minmax-sse-x4.c",
2212 "src/f32-vbinary/gen/vadd-minmax-sse-x8.c",
2213 "src/f32-vbinary/gen/vaddc-minmax-sse-x4.c",
2214 "src/f32-vbinary/gen/vaddc-minmax-sse-x8.c",
2215 "src/f32-vbinary/gen/vdiv-minmax-sse-x4.c",
2216 "src/f32-vbinary/gen/vdiv-minmax-sse-x8.c",
2217 "src/f32-vbinary/gen/vdivc-minmax-sse-x4.c",
2218 "src/f32-vbinary/gen/vdivc-minmax-sse-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -08002219 "src/f32-vbinary/gen/vmax-sse-x4.c",
2220 "src/f32-vbinary/gen/vmax-sse-x8.c",
2221 "src/f32-vbinary/gen/vmaxc-sse-x4.c",
2222 "src/f32-vbinary/gen/vmaxc-sse-x8.c",
2223 "src/f32-vbinary/gen/vmin-sse-x4.c",
2224 "src/f32-vbinary/gen/vmin-sse-x8.c",
2225 "src/f32-vbinary/gen/vminc-sse-x4.c",
2226 "src/f32-vbinary/gen/vminc-sse-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002227 "src/f32-vbinary/gen/vmul-minmax-sse-x4.c",
2228 "src/f32-vbinary/gen/vmul-minmax-sse-x8.c",
2229 "src/f32-vbinary/gen/vmulc-minmax-sse-x4.c",
2230 "src/f32-vbinary/gen/vmulc-minmax-sse-x8.c",
2231 "src/f32-vbinary/gen/vrdivc-minmax-sse-x4.c",
2232 "src/f32-vbinary/gen/vrdivc-minmax-sse-x8.c",
2233 "src/f32-vbinary/gen/vrsubc-minmax-sse-x4.c",
2234 "src/f32-vbinary/gen/vrsubc-minmax-sse-x8.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07002235 "src/f32-vbinary/gen/vsqrdiff-sse-x4.c",
2236 "src/f32-vbinary/gen/vsqrdiff-sse-x8.c",
2237 "src/f32-vbinary/gen/vsqrdiffc-sse-x4.c",
2238 "src/f32-vbinary/gen/vsqrdiffc-sse-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002239 "src/f32-vbinary/gen/vsub-minmax-sse-x4.c",
2240 "src/f32-vbinary/gen/vsub-minmax-sse-x8.c",
2241 "src/f32-vbinary/gen/vsubc-minmax-sse-x4.c",
2242 "src/f32-vbinary/gen/vsubc-minmax-sse-x8.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07002243 "src/f32-vlrelu/gen/vlrelu-sse-x4.c",
2244 "src/f32-vlrelu/gen/vlrelu-sse-x8.c",
Marat Dukhan99936602020-04-11 16:47:01 -07002245 "src/f32-vmulcaddc/gen/c4-minmax-sse-2x.c",
2246 "src/f32-vmulcaddc/gen/c8-minmax-sse-2x.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07002247 "src/f32-vsqrt/gen/sse-sqrt-x4.c",
2248 "src/f32-vsqrt/gen/sse-sqrt-x8.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07002249 "src/f32-vunary/gen/vabs-sse-x4.c",
2250 "src/f32-vunary/gen/vabs-sse-x8.c",
2251 "src/f32-vunary/gen/vneg-sse-x4.c",
2252 "src/f32-vunary/gen/vneg-sse-x8.c",
2253 "src/f32-vunary/gen/vsqr-sse-x4.c",
2254 "src/f32-vunary/gen/vsqr-sse-x8.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07002255 "src/math/roundd-sse-addsub.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002256 "src/math/roundne-sse-addsub.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07002257 "src/math/roundu-sse-addsub.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07002258 "src/math/roundz-sse-addsub.c",
Marat Dukhan84000762020-06-29 18:38:43 -07002259 "src/math/sqrt-sse-hh1mac.c",
2260 "src/math/sqrt-sse-nr1mac.c",
2261 "src/math/sqrt-sse-nr2mac.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002262 "src/x32-fill/sse.c",
2263 "src/x32-packx/x4-sse.c",
2264 "src/x32-pad/sse.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002265]
2266
2267SSE2_UKERNELS = [
Marat Dukhan329da642019-11-19 21:44:39 -08002268 "src/f32-argmaxpool/4x-sse2-c4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002269 "src/f32-argmaxpool/9p8x-sse2-c4.c",
Marat Dukhan329da642019-11-19 21:44:39 -08002270 "src/f32-argmaxpool/9x-sse2-c4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08002271 "src/f32-prelu/gen/sse2-2x4.c",
2272 "src/f32-prelu/gen/sse2-2x8.c",
Marat Dukhanb39689d2020-01-24 13:32:20 -08002273 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x4.c",
Marat Dukhanb39689d2020-01-24 13:32:20 -08002274 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x8-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002275 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x8.c",
Marat Dukhanb39689d2020-01-24 13:32:20 -08002276 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12-acc2.c",
2277 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002278 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12.c",
Marat Dukhanb39689d2020-01-24 13:32:20 -08002279 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16-acc2.c",
2280 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002281 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16.c",
Marat Dukhanb39689d2020-01-24 13:32:20 -08002282 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20-acc2.c",
2283 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002284 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20.c",
Marat Dukhand243c1a2020-09-17 18:48:10 -07002285 "src/f32-sigmoid/gen/sse2-lut64-p2-div-x4.c",
2286 "src/f32-sigmoid/gen/sse2-lut64-p2-div-x8.c",
2287 "src/f32-sigmoid/gen/sse2-lut64-p2-div-x12.c",
2288 "src/f32-sigmoid/gen/sse2-lut64-p2-div-x16.c",
2289 "src/f32-sigmoid/gen/sse2-lut64-p2-div-x20.c",
2290 "src/f32-sigmoid/gen/sse2-lut64-p2-div-x24.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002291 "src/f32-sigmoid/gen/sse2-p5-div-x4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08002292 "src/f32-sigmoid/gen/sse2-p5-div-x8.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002293 "src/f32-sigmoid/gen/sse2-p5-div-x12.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08002294 "src/f32-sigmoid/gen/sse2-p5-div-x16.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002295 "src/f32-sigmoid/gen/sse2-p5-div-x20.c",
2296 "src/f32-sigmoid/gen/sse2-p5-div-x24.c",
Marat Dukhaned6baaf2020-12-01 15:07:08 -08002297 "src/f32-velu/gen/velu-sse2-rr2-lut16-p3-x4.c",
2298 "src/f32-velu/gen/velu-sse2-rr2-lut16-p3-x8.c",
2299 "src/f32-velu/gen/velu-sse2-rr2-lut16-p3-x12.c",
2300 "src/f32-velu/gen/velu-sse2-rr2-lut16-p3-x16.c",
2301 "src/f32-velu/gen/velu-sse2-rr2-lut16-p3-x20.c",
2302 "src/f32-velu/gen/velu-sse2-rr2-lut16-p3-x24.c",
2303 "src/f32-velu/gen/velu-sse2-rr2-p6-x4.c",
2304 "src/f32-velu/gen/velu-sse2-rr2-p6-x8.c",
2305 "src/f32-velu/gen/velu-sse2-rr2-p6-x12.c",
2306 "src/f32-velu/gen/velu-sse2-rr2-p6-x16.c",
2307 "src/f32-velu/gen/velu-sse2-rr2-p6-x20.c",
2308 "src/f32-velu/gen/velu-sse2-rr2-p6-x24.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07002309 "src/f32-vlrelu/gen/vlrelu-sse2-x4.c",
2310 "src/f32-vlrelu/gen/vlrelu-sse2-x8.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07002311 "src/f32-vrnd/gen/vrndd-sse2-x4.c",
2312 "src/f32-vrnd/gen/vrndd-sse2-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002313 "src/f32-vrnd/gen/vrndne-sse2-x4.c",
2314 "src/f32-vrnd/gen/vrndne-sse2-x8.c",
2315 "src/f32-vrnd/gen/vrndu-sse2-x4.c",
2316 "src/f32-vrnd/gen/vrndu-sse2-x8.c",
2317 "src/f32-vrnd/gen/vrndz-sse2-x4.c",
2318 "src/f32-vrnd/gen/vrndz-sse2-x8.c",
Marat Dukhanb7633f22020-11-20 16:34:56 -08002319 "src/math/exp-sse2-rr2-lut64-p2.c",
2320 "src/math/exp-sse2-rr2-p5.c",
Marat Dukhande390d42020-11-29 19:32:18 -08002321 "src/math/expm1minus-sse2-rr2-lut16-p3.c",
Marat Dukhana438aca2020-11-20 15:45:01 -08002322 "src/math/expm1minus-sse2-rr2-p6.c",
Frank Barchard3b800452020-11-22 12:12:35 -08002323 "src/math/expminus-sse2-rr2-p5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002324 "src/math/roundd-sse2-cvt.c",
2325 "src/math/roundne-sse2-cvt.c",
2326 "src/math/roundu-sse2-cvt.c",
2327 "src/math/roundz-sse2-cvt.c",
2328 "src/math/sigmoid-sse2-rr2-lut64-p2-div.c",
2329 "src/math/sigmoid-sse2-rr2-lut64-p2-nr1.c",
2330 "src/math/sigmoid-sse2-rr2-lut64-p2-nr2.c",
2331 "src/math/sigmoid-sse2-rr2-p5-div.c",
2332 "src/math/sigmoid-sse2-rr2-p5-nr1.c",
2333 "src/math/sigmoid-sse2-rr2-p5-nr2.c",
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07002334 "src/qs8-dwconv/gen/up8x9-minmax-sse2-mul16.c",
2335 "src/qs8-dwconv/gen/up16x9-minmax-sse2-mul16.c",
2336 "src/qs8-dwconv/gen/up24x9-minmax-sse2-mul16.c",
Marat Dukhan159688f2020-08-06 10:34:29 -07002337 "src/qs8-gavgpool/gen/7p7x-minmax-sse2-c8-acc2.c",
2338 "src/qs8-gavgpool/gen/7p7x-minmax-sse2-c16-acc2.c",
2339 "src/qs8-gavgpool/gen/7p7x-minmax-sse2-c24-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002340 "src/qs8-gavgpool/gen/7x-minmax-sse2-c8-acc2.c",
2341 "src/qs8-gavgpool/gen/7x-minmax-sse2-c16-acc2.c",
2342 "src/qs8-gavgpool/gen/7x-minmax-sse2-c24-acc2.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002343 "src/qs8-gemm/gen/1x4c2-minmax-sse2-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002344 "src/qs8-gemm/gen/1x4c2-minmax-sse2-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002345 "src/qs8-gemm/gen/1x4c2-xw-minmax-sse2.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002346 "src/qs8-gemm/gen/1x4c8-minmax-sse2-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002347 "src/qs8-gemm/gen/1x4c8-minmax-sse2-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002348 "src/qs8-gemm/gen/1x4c8-xw-minmax-sse2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002349 "src/qs8-gemm/gen/2x4c8-minmax-sse2-ld64.c",
2350 "src/qs8-gemm/gen/2x4c8-minmax-sse2-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002351 "src/qs8-gemm/gen/2x4c8-xw-minmax-sse2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002352 "src/qs8-gemm/gen/3x4c8-minmax-sse2-ld64.c",
2353 "src/qs8-gemm/gen/3x4c8-minmax-sse2-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002354 "src/qs8-gemm/gen/3x4c8-xw-minmax-sse2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002355 "src/qs8-gemm/gen/4x4c2-minmax-sse2-ld64.c",
2356 "src/qs8-gemm/gen/4x4c2-minmax-sse2-ld128.c",
2357 "src/qs8-gemm/gen/4x4c2-xw-minmax-sse2.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002358 "src/qs8-igemm/gen/1x4c2-minmax-sse2-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002359 "src/qs8-igemm/gen/1x4c2-minmax-sse2-ld128.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002360 "src/qs8-igemm/gen/1x4c8-minmax-sse2-ld64.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002361 "src/qs8-igemm/gen/1x4c8-minmax-sse2-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002362 "src/qs8-igemm/gen/2x4c8-minmax-sse2-ld64.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002363 "src/qs8-igemm/gen/2x4c8-minmax-sse2-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002364 "src/qs8-igemm/gen/3x4c8-minmax-sse2-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002365 "src/qs8-igemm/gen/3x4c8-minmax-sse2-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002366 "src/qs8-igemm/gen/4x4c2-minmax-sse2-ld64.c",
2367 "src/qs8-igemm/gen/4x4c2-minmax-sse2-ld128.c",
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07002368 "src/qs8-requantization/fp32-sse2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002369 "src/qs8-requantization/precise-sse2.c",
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07002370 "src/qs8-requantization/q31-sse2.c",
Marat Dukhand9f3ad42020-08-10 12:30:58 -07002371 "src/qs8-vadd/gen/minmax-sse2-mul16-ld64-x8.c",
2372 "src/qs8-vadd/gen/minmax-sse2-mul16-ld64-x16.c",
2373 "src/qs8-vadd/gen/minmax-sse2-mul16-ld64-x24.c",
2374 "src/qs8-vadd/gen/minmax-sse2-mul16-ld64-x32.c",
Marat Dukhan0270d9f2020-08-11 00:56:46 -07002375 "src/qs8-vaddc/gen/minmax-sse2-mul16-ld64-x8.c",
2376 "src/qs8-vaddc/gen/minmax-sse2-mul16-ld64-x16.c",
2377 "src/qs8-vaddc/gen/minmax-sse2-mul16-ld64-x24.c",
2378 "src/qs8-vaddc/gen/minmax-sse2-mul16-ld64-x32.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -07002379 "src/qu8-avgpool/9p8x-minmax-sse2-c8.c",
2380 "src/qu8-avgpool/9x-minmax-sse2-c8.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -07002381 "src/qu8-dwconv/up8x9-minmax-sse2.c",
2382 "src/qu8-gavgpool/7p7x-minmax-sse2-c8.c",
2383 "src/qu8-gavgpool/7x-minmax-sse2-c8.c",
2384 "src/qu8-gemm/2x4c8-minmax-sse2.c",
2385 "src/qu8-gemm/4x4c2-minmax-sse2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002386 "src/qu8-igemm/4x4c2-minmax-sse2.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07002387 "src/qu8-requantization/fp32-sse2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002388 "src/qu8-requantization/precise-sse2.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07002389 "src/qu8-requantization/q31-sse2.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -07002390 "src/qu8-vadd/minmax-sse2.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -07002391 "src/u8-clamp/sse2-x64.c",
Marat Dukhan99936602020-04-11 16:47:01 -07002392 "src/u8-maxpool/9p8x-minmax-sse2-c16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002393 "src/u8-rmax/sse2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002394 "src/x8-zip/x2-sse2.c",
2395 "src/x8-zip/x3-sse2.c",
2396 "src/x8-zip/x4-sse2.c",
2397 "src/x8-zip/xm-sse2.c",
Marat Dukhan57dccd82020-04-14 00:53:10 -07002398 "src/x32-unpool/sse2.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002399 "src/x32-zip/x2-sse2.c",
2400 "src/x32-zip/x3-sse2.c",
2401 "src/x32-zip/x4-sse2.c",
2402 "src/x32-zip/xm-sse2.c",
Marat Dukhanfe7acb62020-03-09 19:30:05 -07002403]
2404
2405SSSE3_UKERNELS = [
Frank Barchard35db7d02020-10-26 13:37:34 -07002406 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-1x4-acc2.c",
2407 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-1x4-acc3.c",
2408 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-1x4-acc4.c",
Marat Dukhan98f2eeb2020-10-23 23:13:41 -07002409 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-1x4.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07002410 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-2x4-acc2.c",
Marat Dukhan98f2eeb2020-10-23 23:13:41 -07002411 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-2x4.c",
2412 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-3x4.c",
2413 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-4x4.c",
2414 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-5x4.c",
2415 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-6x4.c",
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07002416 "src/qs8-dwconv/gen/up8x9-minmax-ssse3-mul16.c",
2417 "src/qs8-dwconv/gen/up16x9-minmax-ssse3-mul16.c",
2418 "src/qs8-dwconv/gen/up24x9-minmax-ssse3-mul16.c",
Marat Dukhan159688f2020-08-06 10:34:29 -07002419 "src/qs8-gavgpool/gen/7p7x-minmax-ssse3-c8-acc2.c",
2420 "src/qs8-gavgpool/gen/7p7x-minmax-ssse3-c16-acc2.c",
2421 "src/qs8-gavgpool/gen/7p7x-minmax-ssse3-c24-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002422 "src/qs8-gavgpool/gen/7x-minmax-ssse3-c8-acc2.c",
2423 "src/qs8-gavgpool/gen/7x-minmax-ssse3-c16-acc2.c",
2424 "src/qs8-gavgpool/gen/7x-minmax-ssse3-c24-acc2.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002425 "src/qs8-gemm/gen/1x4c2-minmax-ssse3-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002426 "src/qs8-gemm/gen/1x4c2-minmax-ssse3-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002427 "src/qs8-gemm/gen/1x4c2-xw-minmax-ssse3.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002428 "src/qs8-gemm/gen/1x4c8-minmax-ssse3-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002429 "src/qs8-gemm/gen/1x4c8-minmax-ssse3-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002430 "src/qs8-gemm/gen/1x4c8-xw-minmax-ssse3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002431 "src/qs8-gemm/gen/2x4c8-minmax-ssse3-ld64.c",
2432 "src/qs8-gemm/gen/2x4c8-minmax-ssse3-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002433 "src/qs8-gemm/gen/2x4c8-xw-minmax-ssse3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002434 "src/qs8-gemm/gen/3x4c8-minmax-ssse3-ld64.c",
2435 "src/qs8-gemm/gen/3x4c8-minmax-ssse3-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002436 "src/qs8-gemm/gen/3x4c8-xw-minmax-ssse3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002437 "src/qs8-gemm/gen/4x4c2-minmax-ssse3-ld64.c",
2438 "src/qs8-gemm/gen/4x4c2-minmax-ssse3-ld128.c",
2439 "src/qs8-gemm/gen/4x4c2-xw-minmax-ssse3.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002440 "src/qs8-igemm/gen/1x4c2-minmax-ssse3-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002441 "src/qs8-igemm/gen/1x4c2-minmax-ssse3-ld128.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002442 "src/qs8-igemm/gen/1x4c8-minmax-ssse3-ld64.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002443 "src/qs8-igemm/gen/1x4c8-minmax-ssse3-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002444 "src/qs8-igemm/gen/2x4c8-minmax-ssse3-ld64.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002445 "src/qs8-igemm/gen/2x4c8-minmax-ssse3-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002446 "src/qs8-igemm/gen/3x4c8-minmax-ssse3-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002447 "src/qs8-igemm/gen/3x4c8-minmax-ssse3-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002448 "src/qs8-igemm/gen/4x4c2-minmax-ssse3-ld64.c",
2449 "src/qs8-igemm/gen/4x4c2-minmax-ssse3-ld128.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -07002450 "src/qs8-requantization/precise-ssse3.c",
2451 "src/qs8-requantization/q31-ssse3.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07002452 "src/qu8-requantization/precise-ssse3.c",
2453 "src/qu8-requantization/q31-ssse3.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002454]
2455
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08002456SSE41_UKERNELS = [
Marat Dukhan40a672f2019-11-25 03:08:22 -08002457 "src/f32-prelu/gen/sse41-2x4.c",
2458 "src/f32-prelu/gen/sse41-2x8.c",
Marat Dukhand243c1a2020-09-17 18:48:10 -07002459 "src/f32-sigmoid/gen/sse41-lut64-p2-div-x4.c",
2460 "src/f32-sigmoid/gen/sse41-lut64-p2-div-x8.c",
2461 "src/f32-sigmoid/gen/sse41-lut64-p2-div-x12.c",
2462 "src/f32-sigmoid/gen/sse41-lut64-p2-div-x16.c",
2463 "src/f32-sigmoid/gen/sse41-lut64-p2-div-x20.c",
2464 "src/f32-sigmoid/gen/sse41-lut64-p2-div-x24.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002465 "src/f32-sigmoid/gen/sse41-p5-div-x4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08002466 "src/f32-sigmoid/gen/sse41-p5-div-x8.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002467 "src/f32-sigmoid/gen/sse41-p5-div-x12.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08002468 "src/f32-sigmoid/gen/sse41-p5-div-x16.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002469 "src/f32-sigmoid/gen/sse41-p5-div-x20.c",
2470 "src/f32-sigmoid/gen/sse41-p5-div-x24.c",
Marat Dukhaned6baaf2020-12-01 15:07:08 -08002471 "src/f32-velu/gen/velu-sse41-rr2-lut16-p3-x4.c",
2472 "src/f32-velu/gen/velu-sse41-rr2-lut16-p3-x8.c",
2473 "src/f32-velu/gen/velu-sse41-rr2-lut16-p3-x12.c",
2474 "src/f32-velu/gen/velu-sse41-rr2-lut16-p3-x16.c",
2475 "src/f32-velu/gen/velu-sse41-rr2-lut16-p3-x20.c",
2476 "src/f32-velu/gen/velu-sse41-rr2-lut16-p3-x24.c",
2477 "src/f32-velu/gen/velu-sse41-rr2-p6-x4.c",
2478 "src/f32-velu/gen/velu-sse41-rr2-p6-x8.c",
2479 "src/f32-velu/gen/velu-sse41-rr2-p6-x12.c",
2480 "src/f32-velu/gen/velu-sse41-rr2-p6-x16.c",
2481 "src/f32-velu/gen/velu-sse41-rr2-p6-x20.c",
2482 "src/f32-velu/gen/velu-sse41-rr2-p6-x24.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07002483 "src/f32-vlrelu/gen/vlrelu-sse41-x4.c",
2484 "src/f32-vlrelu/gen/vlrelu-sse41-x8.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07002485 "src/f32-vrnd/gen/vrndd-sse41-x4.c",
2486 "src/f32-vrnd/gen/vrndd-sse41-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002487 "src/f32-vrnd/gen/vrndne-sse41-x4.c",
2488 "src/f32-vrnd/gen/vrndne-sse41-x8.c",
2489 "src/f32-vrnd/gen/vrndu-sse41-x4.c",
2490 "src/f32-vrnd/gen/vrndu-sse41-x8.c",
2491 "src/f32-vrnd/gen/vrndz-sse41-x4.c",
2492 "src/f32-vrnd/gen/vrndz-sse41-x8.c",
2493 "src/math/roundd-sse41.c",
2494 "src/math/roundne-sse41.c",
2495 "src/math/roundu-sse41.c",
2496 "src/math/roundz-sse41.c",
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07002497 "src/qs8-dwconv/gen/up8x9-minmax-sse41-mul16.c",
2498 "src/qs8-dwconv/gen/up16x9-minmax-sse41-mul16.c",
2499 "src/qs8-dwconv/gen/up24x9-minmax-sse41-mul16.c",
Marat Dukhan159688f2020-08-06 10:34:29 -07002500 "src/qs8-gavgpool/gen/7p7x-minmax-sse41-c8-acc2.c",
2501 "src/qs8-gavgpool/gen/7p7x-minmax-sse41-c16-acc2.c",
2502 "src/qs8-gavgpool/gen/7p7x-minmax-sse41-c24-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002503 "src/qs8-gavgpool/gen/7x-minmax-sse41-c8-acc2.c",
2504 "src/qs8-gavgpool/gen/7x-minmax-sse41-c16-acc2.c",
2505 "src/qs8-gavgpool/gen/7x-minmax-sse41-c24-acc2.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002506 "src/qs8-gemm/gen/1x4c2-minmax-sse41-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002507 "src/qs8-gemm/gen/1x4c2-minmax-sse41-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002508 "src/qs8-gemm/gen/1x4c2-xw-minmax-sse41.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002509 "src/qs8-gemm/gen/1x4c8-minmax-sse41-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002510 "src/qs8-gemm/gen/1x4c8-minmax-sse41-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002511 "src/qs8-gemm/gen/1x4c8-xw-minmax-sse41.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002512 "src/qs8-gemm/gen/2x4c8-minmax-sse41-ld64.c",
2513 "src/qs8-gemm/gen/2x4c8-minmax-sse41-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002514 "src/qs8-gemm/gen/2x4c8-xw-minmax-sse41.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002515 "src/qs8-gemm/gen/3x4c8-minmax-sse41-ld64.c",
2516 "src/qs8-gemm/gen/3x4c8-minmax-sse41-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002517 "src/qs8-gemm/gen/3x4c8-xw-minmax-sse41.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002518 "src/qs8-gemm/gen/4x4c2-minmax-sse41-ld64.c",
2519 "src/qs8-gemm/gen/4x4c2-minmax-sse41-ld128.c",
2520 "src/qs8-gemm/gen/4x4c2-xw-minmax-sse41.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002521 "src/qs8-igemm/gen/1x4c2-minmax-sse41-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002522 "src/qs8-igemm/gen/1x4c2-minmax-sse41-ld128.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002523 "src/qs8-igemm/gen/1x4c8-minmax-sse41-ld64.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002524 "src/qs8-igemm/gen/1x4c8-minmax-sse41-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002525 "src/qs8-igemm/gen/2x4c8-minmax-sse41-ld64.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002526 "src/qs8-igemm/gen/2x4c8-minmax-sse41-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002527 "src/qs8-igemm/gen/3x4c8-minmax-sse41-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002528 "src/qs8-igemm/gen/3x4c8-minmax-sse41-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002529 "src/qs8-igemm/gen/4x4c2-minmax-sse41-ld64.c",
2530 "src/qs8-igemm/gen/4x4c2-minmax-sse41-ld128.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -07002531 "src/qs8-requantization/fp32-sse4.c",
2532 "src/qs8-requantization/precise-sse4.c",
2533 "src/qs8-requantization/q31-sse4.c",
Marat Dukhand9f3ad42020-08-10 12:30:58 -07002534 "src/qs8-vadd/gen/minmax-sse41-mul16-ld64-x8.c",
2535 "src/qs8-vadd/gen/minmax-sse41-mul16-ld64-x16.c",
2536 "src/qs8-vadd/gen/minmax-sse41-mul16-ld64-x24.c",
2537 "src/qs8-vadd/gen/minmax-sse41-mul16-ld64-x32.c",
Marat Dukhanbb9225e2020-09-06 22:40:56 -07002538 "src/qs8-vadd/gen/minmax-sse41-mul32-ld32-x8.c",
2539 "src/qs8-vadd/gen/minmax-sse41-mul32-ld32-x16.c",
2540 "src/qs8-vadd/gen/minmax-sse41-mul32-ld32-x24.c",
2541 "src/qs8-vadd/gen/minmax-sse41-mul32-ld32-x32.c",
Marat Dukhan0270d9f2020-08-11 00:56:46 -07002542 "src/qs8-vaddc/gen/minmax-sse41-mul16-ld64-x8.c",
2543 "src/qs8-vaddc/gen/minmax-sse41-mul16-ld64-x16.c",
2544 "src/qs8-vaddc/gen/minmax-sse41-mul16-ld64-x24.c",
2545 "src/qs8-vaddc/gen/minmax-sse41-mul16-ld64-x32.c",
Marat Dukhanbb9225e2020-09-06 22:40:56 -07002546 "src/qs8-vaddc/gen/minmax-sse41-mul32-ld32-x8.c",
2547 "src/qs8-vaddc/gen/minmax-sse41-mul32-ld32-x16.c",
2548 "src/qs8-vaddc/gen/minmax-sse41-mul32-ld32-x24.c",
2549 "src/qs8-vaddc/gen/minmax-sse41-mul32-ld32-x32.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07002550 "src/qu8-requantization/precise-sse4.c",
2551 "src/qu8-requantization/q31-sse4.c",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08002552]
2553
Marat Dukhan08c4a432019-10-03 09:29:21 -07002554AVX_UKERNELS = [
Marat Dukhan5c5fa962020-03-10 18:38:33 -07002555 "src/f32-clamp/gen/avx-x8.c",
2556 "src/f32-clamp/gen/avx-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002557 "src/f32-dwconv/gen/up8x4-minmax-avx-acc2.c",
2558 "src/f32-dwconv/gen/up8x4-minmax-avx.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002559 "src/f32-dwconv/gen/up8x9-minmax-avx-acc2.c",
2560 "src/f32-dwconv/gen/up8x9-minmax-avx.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002561 "src/f32-dwconv/gen/up8x25-minmax-avx-acc2.c",
2562 "src/f32-dwconv/gen/up8x25-minmax-avx.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002563 "src/f32-dwconv/gen/up16x4-minmax-avx-acc2.c",
2564 "src/f32-dwconv/gen/up16x4-minmax-avx.c",
2565 "src/f32-dwconv/gen/up16x9-minmax-avx-acc2.c",
2566 "src/f32-dwconv/gen/up16x9-minmax-avx.c",
2567 "src/f32-dwconv/gen/up16x25-minmax-avx-acc2.c",
2568 "src/f32-dwconv/gen/up16x25-minmax-avx.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002569 "src/f32-gemm/gen-inc/1x8inc-minmax-avx-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002570 "src/f32-gemm/gen-inc/1x16inc-minmax-avx-broadcast.c",
2571 "src/f32-gemm/gen-inc/3x16inc-minmax-avx-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002572 "src/f32-gemm/gen-inc/4x8inc-minmax-avx-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002573 "src/f32-gemm/gen-inc/4x16inc-minmax-avx-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002574 "src/f32-gemm/gen-inc/5x8inc-minmax-avx-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002575 "src/f32-gemm/gen-inc/5x16inc-minmax-avx-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002576 "src/f32-gemm/gen-inc/6x8inc-minmax-avx-broadcast.c",
2577 "src/f32-gemm/gen-inc/7x8inc-minmax-avx-broadcast.c",
2578 "src/f32-gemm/gen/1x8-minmax-avx-broadcast.c",
2579 "src/f32-gemm/gen/1x16-minmax-avx-broadcast.c",
2580 "src/f32-gemm/gen/3x16-minmax-avx-broadcast.c",
2581 "src/f32-gemm/gen/4x8-minmax-avx-broadcast.c",
2582 "src/f32-gemm/gen/4x16-minmax-avx-broadcast.c",
2583 "src/f32-gemm/gen/5x8-minmax-avx-broadcast.c",
2584 "src/f32-gemm/gen/5x16-minmax-avx-broadcast.c",
2585 "src/f32-gemm/gen/6x8-minmax-avx-broadcast.c",
2586 "src/f32-gemm/gen/7x8-minmax-avx-broadcast.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07002587 "src/f32-hswish/gen/hswish-avx-x8.c",
2588 "src/f32-hswish/gen/hswish-avx-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002589 "src/f32-igemm/gen/1x8-minmax-avx-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002590 "src/f32-igemm/gen/1x16-minmax-avx-broadcast.c",
2591 "src/f32-igemm/gen/3x16-minmax-avx-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002592 "src/f32-igemm/gen/4x8-minmax-avx-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002593 "src/f32-igemm/gen/4x16-minmax-avx-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002594 "src/f32-igemm/gen/5x8-minmax-avx-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002595 "src/f32-igemm/gen/5x16-minmax-avx-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002596 "src/f32-igemm/gen/6x8-minmax-avx-broadcast.c",
2597 "src/f32-igemm/gen/7x8-minmax-avx-broadcast.c",
Marat Dukhan90eca0a2020-03-11 00:52:23 -07002598 "src/f32-prelu/gen/avx-2x8.c",
2599 "src/f32-prelu/gen/avx-2x16.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07002600 "src/f32-relu/gen/avx-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002601 "src/f32-relu/gen/avx-x16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002602 "src/f32-rmax/avx.c",
T.J. Alumbaughdc2b29c2020-10-14 13:56:08 -07002603 "src/f32-sigmoid/gen/avx-rr2-p5-div-x8.c",
2604 "src/f32-sigmoid/gen/avx-rr2-p5-div-x16.c",
2605 "src/f32-sigmoid/gen/avx-rr2-p5-div-x24.c",
2606 "src/f32-sigmoid/gen/avx-rr2-p5-div-x32.c",
2607 "src/f32-sigmoid/gen/avx-rr2-p5-div-x40.c",
2608 "src/f32-sigmoid/gen/avx-rr2-p5-div-x48.c",
2609 "src/f32-sigmoid/gen/avx-rr2-p5-div-x56.c",
2610 "src/f32-sigmoid/gen/avx-rr2-p5-div-x64.c",
2611 "src/f32-sigmoid/gen/avx-rr2-p5-div-x72.c",
2612 "src/f32-sigmoid/gen/avx-rr2-p5-div-x80.c",
2613 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x8.c",
2614 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x16.c",
2615 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x24.c",
2616 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x32.c",
2617 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x40.c",
2618 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x48.c",
2619 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x56.c",
2620 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x64.c",
2621 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x72.c",
2622 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x80.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002623 "src/f32-vbinary/gen/vadd-minmax-avx-x8.c",
2624 "src/f32-vbinary/gen/vadd-minmax-avx-x16.c",
2625 "src/f32-vbinary/gen/vaddc-minmax-avx-x8.c",
2626 "src/f32-vbinary/gen/vaddc-minmax-avx-x16.c",
2627 "src/f32-vbinary/gen/vdiv-minmax-avx-x8.c",
2628 "src/f32-vbinary/gen/vdiv-minmax-avx-x16.c",
2629 "src/f32-vbinary/gen/vdivc-minmax-avx-x8.c",
2630 "src/f32-vbinary/gen/vdivc-minmax-avx-x16.c",
Marat Dukhan9a88efe2019-12-10 15:54:24 -08002631 "src/f32-vbinary/gen/vmax-avx-x8.c",
2632 "src/f32-vbinary/gen/vmax-avx-x16.c",
2633 "src/f32-vbinary/gen/vmaxc-avx-x8.c",
2634 "src/f32-vbinary/gen/vmaxc-avx-x16.c",
2635 "src/f32-vbinary/gen/vmin-avx-x8.c",
2636 "src/f32-vbinary/gen/vmin-avx-x16.c",
2637 "src/f32-vbinary/gen/vminc-avx-x8.c",
2638 "src/f32-vbinary/gen/vminc-avx-x16.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002639 "src/f32-vbinary/gen/vmul-minmax-avx-x8.c",
2640 "src/f32-vbinary/gen/vmul-minmax-avx-x16.c",
2641 "src/f32-vbinary/gen/vmulc-minmax-avx-x8.c",
2642 "src/f32-vbinary/gen/vmulc-minmax-avx-x16.c",
2643 "src/f32-vbinary/gen/vrdivc-minmax-avx-x8.c",
2644 "src/f32-vbinary/gen/vrdivc-minmax-avx-x16.c",
2645 "src/f32-vbinary/gen/vrsubc-minmax-avx-x8.c",
2646 "src/f32-vbinary/gen/vrsubc-minmax-avx-x16.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07002647 "src/f32-vbinary/gen/vsqrdiff-avx-x8.c",
2648 "src/f32-vbinary/gen/vsqrdiff-avx-x16.c",
2649 "src/f32-vbinary/gen/vsqrdiffc-avx-x8.c",
2650 "src/f32-vbinary/gen/vsqrdiffc-avx-x16.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002651 "src/f32-vbinary/gen/vsub-minmax-avx-x8.c",
2652 "src/f32-vbinary/gen/vsub-minmax-avx-x16.c",
2653 "src/f32-vbinary/gen/vsubc-minmax-avx-x8.c",
2654 "src/f32-vbinary/gen/vsubc-minmax-avx-x16.c",
Marat Dukhaned6baaf2020-12-01 15:07:08 -08002655 "src/f32-velu/gen/velu-avx-rr2-lut4-p4-perm-x8.c",
2656 "src/f32-velu/gen/velu-avx-rr2-lut4-p4-perm-x16.c",
2657 "src/f32-velu/gen/velu-avx-rr2-lut4-p4-perm-x24.c",
2658 "src/f32-velu/gen/velu-avx-rr2-lut4-p4-perm-x32.c",
2659 "src/f32-velu/gen/velu-avx-rr2-lut4-p4-perm-x40.c",
2660 "src/f32-velu/gen/velu-avx-rr2-lut4-p4-perm-x48.c",
2661 "src/f32-velu/gen/velu-avx-rr2-lut16-p3-x8.c",
2662 "src/f32-velu/gen/velu-avx-rr2-lut16-p3-x16.c",
2663 "src/f32-velu/gen/velu-avx-rr2-lut16-p3-x24.c",
2664 "src/f32-velu/gen/velu-avx-rr2-lut16-p3-x32.c",
2665 "src/f32-velu/gen/velu-avx-rr2-lut16-p3-x40.c",
2666 "src/f32-velu/gen/velu-avx-rr2-lut16-p3-x48.c",
2667 "src/f32-velu/gen/velu-avx-rr2-p6-x8.c",
2668 "src/f32-velu/gen/velu-avx-rr2-p6-x16.c",
2669 "src/f32-velu/gen/velu-avx-rr2-p6-x24.c",
2670 "src/f32-velu/gen/velu-avx-rr2-p6-x32.c",
2671 "src/f32-velu/gen/velu-avx-rr2-p6-x40.c",
2672 "src/f32-velu/gen/velu-avx-rr2-p6-x48.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07002673 "src/f32-vlrelu/gen/vlrelu-avx-x8.c",
2674 "src/f32-vlrelu/gen/vlrelu-avx-x16.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07002675 "src/f32-vrnd/gen/vrndd-avx-x8.c",
2676 "src/f32-vrnd/gen/vrndd-avx-x16.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002677 "src/f32-vrnd/gen/vrndne-avx-x8.c",
2678 "src/f32-vrnd/gen/vrndne-avx-x16.c",
2679 "src/f32-vrnd/gen/vrndu-avx-x8.c",
2680 "src/f32-vrnd/gen/vrndu-avx-x16.c",
2681 "src/f32-vrnd/gen/vrndz-avx-x8.c",
2682 "src/f32-vrnd/gen/vrndz-avx-x16.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07002683 "src/f32-vscale/avx-x32.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07002684 "src/f32-vsqrt/gen/avx-sqrt-x8.c",
2685 "src/f32-vsqrt/gen/avx-sqrt-x16.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07002686 "src/f32-vunary/gen/vabs-avx-x8.c",
2687 "src/f32-vunary/gen/vabs-avx-x16.c",
2688 "src/f32-vunary/gen/vneg-avx-x8.c",
2689 "src/f32-vunary/gen/vneg-avx-x16.c",
2690 "src/f32-vunary/gen/vsqr-avx-x8.c",
2691 "src/f32-vunary/gen/vsqr-avx-x16.c",
Marat Dukhanb7633f22020-11-20 16:34:56 -08002692 "src/math/exp-avx-rr2-p5.c",
Marat Dukhande390d42020-11-29 19:32:18 -08002693 "src/math/expm1minus-avx-rr2-lut4-p4-perm.c",
2694 "src/math/expm1minus-avx-rr2-lut16-p3.c",
2695 "src/math/expm1minus-avx-rr2-p6.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002696 "src/math/sigmoid-avx-rr2-lut64-p2-div.c",
Marat Dukhan2e4727b2020-09-24 00:48:39 -07002697 "src/math/sigmoid-avx-rr2-p5-div.c",
Marat Dukhan39ab16f2020-10-16 14:30:35 -07002698 "src/math/sigmoid-avx-rr2-p5-nr1.c",
2699 "src/math/sigmoid-avx-rr2-p5-nr2.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002700]
2701
Marat Dukhan1566fee2020-08-02 21:55:41 -07002702XOP_UKERNELS = [
2703 "src/qs8-gemm/gen/1x4c2-minmax-xop-ld64.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002704 "src/qs8-gemm/gen/1x4c2-minmax-xop-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002705 "src/qs8-gemm/gen/1x4c2-xw-minmax-xop.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002706 "src/qs8-gemm/gen/1x4c8-minmax-xop-ld64.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002707 "src/qs8-gemm/gen/1x4c8-minmax-xop-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002708 "src/qs8-gemm/gen/1x4c8-xw-minmax-xop.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002709 "src/qs8-gemm/gen/2x4c8-minmax-xop-ld64.c",
2710 "src/qs8-gemm/gen/2x4c8-minmax-xop-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002711 "src/qs8-gemm/gen/2x4c8-xw-minmax-xop.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002712 "src/qs8-gemm/gen/3x4c8-minmax-xop-ld64.c",
2713 "src/qs8-gemm/gen/3x4c8-minmax-xop-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002714 "src/qs8-gemm/gen/3x4c8-xw-minmax-xop.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002715 "src/qs8-gemm/gen/4x4c2-minmax-xop-ld64.c",
2716 "src/qs8-gemm/gen/4x4c2-minmax-xop-ld128.c",
2717 "src/qs8-gemm/gen/4x4c2-xw-minmax-xop.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002718 "src/qs8-igemm/gen/1x4c2-minmax-xop-ld64.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002719 "src/qs8-igemm/gen/1x4c2-minmax-xop-ld128.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002720 "src/qs8-igemm/gen/1x4c8-minmax-xop-ld64.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002721 "src/qs8-igemm/gen/1x4c8-minmax-xop-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002722 "src/qs8-igemm/gen/2x4c8-minmax-xop-ld64.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002723 "src/qs8-igemm/gen/2x4c8-minmax-xop-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002724 "src/qs8-igemm/gen/3x4c8-minmax-xop-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002725 "src/qs8-igemm/gen/3x4c8-minmax-xop-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002726 "src/qs8-igemm/gen/4x4c2-minmax-xop-ld64.c",
2727 "src/qs8-igemm/gen/4x4c2-minmax-xop-ld128.c",
Marat Dukhanbb9225e2020-09-06 22:40:56 -07002728 "src/qs8-vadd/gen/minmax-xop-mul32-ld32-x8.c",
2729 "src/qs8-vadd/gen/minmax-xop-mul32-ld32-x16.c",
2730 "src/qs8-vadd/gen/minmax-xop-mul32-ld32-x24.c",
2731 "src/qs8-vadd/gen/minmax-xop-mul32-ld32-x32.c",
2732 "src/qs8-vaddc/gen/minmax-xop-mul32-ld32-x8.c",
2733 "src/qs8-vaddc/gen/minmax-xop-mul32-ld32-x16.c",
2734 "src/qs8-vaddc/gen/minmax-xop-mul32-ld32-x24.c",
2735 "src/qs8-vaddc/gen/minmax-xop-mul32-ld32-x32.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002736]
2737
Marat Dukhanfda12b82019-11-21 12:27:59 -08002738FMA3_UKERNELS = [
Marat Dukhan1c587112020-04-08 20:04:28 -07002739 "src/f32-dwconv/gen/up8x4-minmax-fma3-acc2.c",
2740 "src/f32-dwconv/gen/up8x4-minmax-fma3.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002741 "src/f32-dwconv/gen/up8x9-minmax-fma3-acc2.c",
2742 "src/f32-dwconv/gen/up8x9-minmax-fma3.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002743 "src/f32-dwconv/gen/up8x25-minmax-fma3-acc2.c",
2744 "src/f32-dwconv/gen/up8x25-minmax-fma3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002745 "src/f32-dwconv/gen/up16x4-minmax-fma3-acc2.c",
2746 "src/f32-dwconv/gen/up16x4-minmax-fma3.c",
2747 "src/f32-dwconv/gen/up16x9-minmax-fma3-acc2.c",
2748 "src/f32-dwconv/gen/up16x9-minmax-fma3.c",
2749 "src/f32-dwconv/gen/up16x25-minmax-fma3-acc2.c",
2750 "src/f32-dwconv/gen/up16x25-minmax-fma3.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002751 "src/f32-gemm/gen-inc/1x8inc-minmax-fma3-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002752 "src/f32-gemm/gen-inc/1x16inc-minmax-fma3-broadcast.c",
2753 "src/f32-gemm/gen-inc/1x16s4inc-minmax-fma3-broadcast.c",
2754 "src/f32-gemm/gen-inc/3x16inc-minmax-fma3-broadcast.c",
2755 "src/f32-gemm/gen-inc/3x16s4inc-minmax-fma3-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002756 "src/f32-gemm/gen-inc/4x8inc-minmax-fma3-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002757 "src/f32-gemm/gen-inc/4x16inc-minmax-fma3-broadcast.c",
2758 "src/f32-gemm/gen-inc/4x16s4inc-minmax-fma3-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002759 "src/f32-gemm/gen-inc/5x8inc-minmax-fma3-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002760 "src/f32-gemm/gen-inc/5x16inc-minmax-fma3-broadcast.c",
2761 "src/f32-gemm/gen-inc/5x16s4inc-minmax-fma3-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002762 "src/f32-gemm/gen-inc/6x8inc-minmax-fma3-broadcast.c",
2763 "src/f32-gemm/gen-inc/7x8inc-minmax-fma3-broadcast.c",
2764 "src/f32-gemm/gen-inc/8x8inc-minmax-fma3-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002765 "src/f32-gemm/gen/1x8-minmax-fma3-broadcast.c",
2766 "src/f32-gemm/gen/1x16-minmax-fma3-broadcast.c",
2767 "src/f32-gemm/gen/1x16s4-minmax-fma3-broadcast.c",
2768 "src/f32-gemm/gen/3x16-minmax-fma3-broadcast.c",
2769 "src/f32-gemm/gen/3x16s4-minmax-fma3-broadcast.c",
2770 "src/f32-gemm/gen/4x8-minmax-fma3-broadcast.c",
2771 "src/f32-gemm/gen/4x16-minmax-fma3-broadcast.c",
2772 "src/f32-gemm/gen/4x16s4-minmax-fma3-broadcast.c",
2773 "src/f32-gemm/gen/5x8-minmax-fma3-broadcast.c",
2774 "src/f32-gemm/gen/5x16-minmax-fma3-broadcast.c",
2775 "src/f32-gemm/gen/5x16s4-minmax-fma3-broadcast.c",
2776 "src/f32-gemm/gen/6x8-minmax-fma3-broadcast.c",
2777 "src/f32-gemm/gen/7x8-minmax-fma3-broadcast.c",
2778 "src/f32-gemm/gen/8x8-minmax-fma3-broadcast.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07002779 "src/f32-hswish/gen/hswish-fma3-x8.c",
2780 "src/f32-hswish/gen/hswish-fma3-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002781 "src/f32-igemm/gen/1x8-minmax-fma3-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002782 "src/f32-igemm/gen/1x16-minmax-fma3-broadcast.c",
2783 "src/f32-igemm/gen/1x16s4-minmax-fma3-broadcast.c",
2784 "src/f32-igemm/gen/3x16-minmax-fma3-broadcast.c",
2785 "src/f32-igemm/gen/3x16s4-minmax-fma3-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002786 "src/f32-igemm/gen/4x8-minmax-fma3-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002787 "src/f32-igemm/gen/4x16-minmax-fma3-broadcast.c",
2788 "src/f32-igemm/gen/4x16s4-minmax-fma3-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002789 "src/f32-igemm/gen/5x8-minmax-fma3-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002790 "src/f32-igemm/gen/5x16-minmax-fma3-broadcast.c",
2791 "src/f32-igemm/gen/5x16s4-minmax-fma3-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002792 "src/f32-igemm/gen/6x8-minmax-fma3-broadcast.c",
2793 "src/f32-igemm/gen/7x8-minmax-fma3-broadcast.c",
2794 "src/f32-igemm/gen/8x8-minmax-fma3-broadcast.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07002795 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x8.c",
2796 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x16.c",
2797 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x24.c",
2798 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x32.c",
2799 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x40.c",
2800 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x48.c",
2801 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x56.c",
2802 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x64.c",
Marat Dukhan84000762020-06-29 18:38:43 -07002803 "src/math/sqrt-fma3-nr1fma.c",
Marat Dukhan84000762020-06-29 18:38:43 -07002804 "src/math/sqrt-fma3-nr1fma1adj.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002805 "src/math/sqrt-fma3-nr2fma.c",
Marat Dukhanfda12b82019-11-21 12:27:59 -08002806]
2807
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002808AVX2_UKERNELS = [
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002809 "src/f32-raddexpminusmax/gen/avx2-p5-x64-acc2.c",
2810 "src/f32-raddexpminusmax/gen/avx2-p5-x64-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002811 "src/f32-raddexpminusmax/gen/avx2-p5-x64.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002812 "src/f32-raddexpminusmax/gen/avx2-p5-x72-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002813 "src/f32-raddexpminusmax/gen/avx2-p5-x72.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002814 "src/f32-raddexpminusmax/gen/avx2-p5-x80-acc2.c",
2815 "src/f32-raddexpminusmax/gen/avx2-p5-x80-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002816 "src/f32-raddexpminusmax/gen/avx2-p5-x80.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002817 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc2.c",
2818 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc3.c",
2819 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc6.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002820 "src/f32-raddexpminusmax/gen/avx2-p5-x96.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002821 "src/f32-raddextexp/gen/avx2-p5-x64-acc2.c",
2822 "src/f32-raddextexp/gen/avx2-p5-x64-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002823 "src/f32-raddextexp/gen/avx2-p5-x64.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002824 "src/f32-raddextexp/gen/avx2-p5-x72-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002825 "src/f32-raddextexp/gen/avx2-p5-x72.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002826 "src/f32-raddextexp/gen/avx2-p5-x80-acc2.c",
2827 "src/f32-raddextexp/gen/avx2-p5-x80-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002828 "src/f32-raddextexp/gen/avx2-p5-x80.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002829 "src/f32-raddextexp/gen/avx2-p5-x96-acc2.c",
2830 "src/f32-raddextexp/gen/avx2-p5-x96-acc3.c",
2831 "src/f32-raddextexp/gen/avx2-p5-x96-acc6.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002832 "src/f32-raddextexp/gen/avx2-p5-x96.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002833 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64-acc2.c",
2834 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002835 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002836 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x72-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002837 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x72.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002838 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80-acc2.c",
2839 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002840 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002841 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc2.c",
2842 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc3.c",
2843 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc6.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002844 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96.c",
Marat Dukhanfa0a4322020-01-06 16:14:29 -08002845 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x8.c",
2846 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x16.c",
2847 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x24.c",
2848 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x32.c",
2849 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x40.c",
2850 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x48.c",
2851 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x56.c",
2852 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x64.c",
2853 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x72.c",
2854 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x80.c",
2855 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x8.c",
2856 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x16.c",
2857 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x24.c",
2858 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x32.c",
2859 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x40.c",
2860 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x48.c",
2861 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x56.c",
2862 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x64.c",
2863 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x72.c",
2864 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x80.c",
2865 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x8.c",
2866 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x16.c",
2867 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x24.c",
2868 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x32.c",
2869 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x40.c",
2870 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x48.c",
2871 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x56.c",
2872 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x64.c",
2873 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x72.c",
2874 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x80.c",
Marat Dukhaned6baaf2020-12-01 15:07:08 -08002875 "src/f32-velu/gen/velu-avx2-rr1-lut4-p4-perm-x8.c",
2876 "src/f32-velu/gen/velu-avx2-rr1-lut4-p4-perm-x16.c",
2877 "src/f32-velu/gen/velu-avx2-rr1-lut4-p4-perm-x24.c",
2878 "src/f32-velu/gen/velu-avx2-rr1-lut4-p4-perm-x32.c",
2879 "src/f32-velu/gen/velu-avx2-rr1-lut4-p4-perm-x40.c",
2880 "src/f32-velu/gen/velu-avx2-rr1-lut4-p4-perm-x48.c",
2881 "src/f32-velu/gen/velu-avx2-rr1-lut4-p4-perm-x56.c",
2882 "src/f32-velu/gen/velu-avx2-rr1-lut4-p4-perm-x64.c",
2883 "src/f32-velu/gen/velu-avx2-rr1-lut4-p4-perm-x72.c",
2884 "src/f32-velu/gen/velu-avx2-rr1-lut4-p4-perm-x80.c",
2885 "src/f32-velu/gen/velu-avx2-rr1-lut8-p4-perm-x8.c",
2886 "src/f32-velu/gen/velu-avx2-rr1-lut8-p4-perm-x16.c",
2887 "src/f32-velu/gen/velu-avx2-rr1-lut8-p4-perm-x24.c",
2888 "src/f32-velu/gen/velu-avx2-rr1-lut8-p4-perm-x32.c",
2889 "src/f32-velu/gen/velu-avx2-rr1-lut8-p4-perm-x40.c",
2890 "src/f32-velu/gen/velu-avx2-rr1-lut8-p4-perm-x48.c",
2891 "src/f32-velu/gen/velu-avx2-rr1-lut8-p4-perm-x56.c",
2892 "src/f32-velu/gen/velu-avx2-rr1-lut8-p4-perm-x64.c",
2893 "src/f32-velu/gen/velu-avx2-rr1-lut8-p4-perm-x72.c",
2894 "src/f32-velu/gen/velu-avx2-rr1-lut8-p4-perm-x80.c",
2895 "src/f32-velu/gen/velu-avx2-rr1-lut16-p3-gather-x8.c",
2896 "src/f32-velu/gen/velu-avx2-rr1-lut16-p3-gather-x16.c",
2897 "src/f32-velu/gen/velu-avx2-rr1-lut16-p3-gather-x24.c",
2898 "src/f32-velu/gen/velu-avx2-rr1-lut16-p3-gather-x32.c",
2899 "src/f32-velu/gen/velu-avx2-rr1-lut16-p3-gather-x40.c",
2900 "src/f32-velu/gen/velu-avx2-rr1-lut16-p3-gather-x48.c",
2901 "src/f32-velu/gen/velu-avx2-rr1-lut16-p3-gather-x56.c",
2902 "src/f32-velu/gen/velu-avx2-rr1-lut16-p3-gather-x64.c",
2903 "src/f32-velu/gen/velu-avx2-rr1-lut16-p3-gather-x72.c",
2904 "src/f32-velu/gen/velu-avx2-rr1-lut16-p3-gather-x80.c",
2905 "src/f32-velu/gen/velu-avx2-rr1-p6-x8.c",
2906 "src/f32-velu/gen/velu-avx2-rr1-p6-x16.c",
2907 "src/f32-velu/gen/velu-avx2-rr1-p6-x24.c",
2908 "src/f32-velu/gen/velu-avx2-rr1-p6-x32.c",
2909 "src/f32-velu/gen/velu-avx2-rr1-p6-x40.c",
2910 "src/f32-velu/gen/velu-avx2-rr1-p6-x48.c",
2911 "src/f32-velu/gen/velu-avx2-rr1-p6-x56.c",
2912 "src/f32-velu/gen/velu-avx2-rr1-p6-x64.c",
2913 "src/f32-velu/gen/velu-avx2-rr1-p6-x72.c",
2914 "src/f32-velu/gen/velu-avx2-rr1-p6-x80.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002915 "src/f32-vscaleexpminusmax/gen/avx2-p5-x8.c",
2916 "src/f32-vscaleexpminusmax/gen/avx2-p5-x16.c",
2917 "src/f32-vscaleexpminusmax/gen/avx2-p5-x24.c",
2918 "src/f32-vscaleexpminusmax/gen/avx2-p5-x32.c",
2919 "src/f32-vscaleexpminusmax/gen/avx2-p5-x40.c",
2920 "src/f32-vscaleexpminusmax/gen/avx2-p5-x48.c",
2921 "src/f32-vscaleexpminusmax/gen/avx2-p5-x56.c",
2922 "src/f32-vscaleexpminusmax/gen/avx2-p5-x64.c",
2923 "src/f32-vscaleexpminusmax/gen/avx2-p5-x72.c",
2924 "src/f32-vscaleexpminusmax/gen/avx2-p5-x80.c",
2925 "src/f32-vscaleexpminusmax/gen/avx2-p5-x88.c",
2926 "src/f32-vscaleexpminusmax/gen/avx2-p5-x96.c",
2927 "src/f32-vscaleextexp/gen/avx2-p5-x8.c",
2928 "src/f32-vscaleextexp/gen/avx2-p5-x16.c",
2929 "src/f32-vscaleextexp/gen/avx2-p5-x24.c",
2930 "src/f32-vscaleextexp/gen/avx2-p5-x32.c",
2931 "src/f32-vscaleextexp/gen/avx2-p5-x40.c",
2932 "src/f32-vscaleextexp/gen/avx2-p5-x48.c",
2933 "src/f32-vscaleextexp/gen/avx2-p5-x56.c",
2934 "src/f32-vscaleextexp/gen/avx2-p5-x64.c",
2935 "src/f32-vscaleextexp/gen/avx2-p5-x72.c",
2936 "src/f32-vscaleextexp/gen/avx2-p5-x80.c",
2937 "src/f32-vscaleextexp/gen/avx2-p5-x88.c",
2938 "src/f32-vscaleextexp/gen/avx2-p5-x96.c",
Marat Dukhanb7633f22020-11-20 16:34:56 -08002939 "src/math/exp-avx2-rr2-lut8-p3-perm.c",
2940 "src/math/exp-avx2-rr2-lut8-p4-perm.c",
2941 "src/math/exp-avx2-rr2-p5.c",
Marat Dukhande390d42020-11-29 19:32:18 -08002942 "src/math/expm1minus-avx2-rr1-lut4-p4-perm.c",
2943 "src/math/expm1minus-avx2-rr1-lut8-p4-perm.c",
2944 "src/math/expm1minus-avx2-rr1-lut16-p3-gather.c",
2945 "src/math/expm1minus-avx2-rr1-p6.c",
Frank Barcharde7223ee2020-12-04 19:04:01 -08002946 "src/math/expminus-avx2-rr2-p5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002947 "src/math/extexp-avx2-p5.c",
2948 "src/math/sigmoid-avx2-rr1-lut64-p2-gather-div.c",
2949 "src/math/sigmoid-avx2-rr1-lut64-p2-gather-nr1fma.c",
2950 "src/math/sigmoid-avx2-rr1-lut64-p2-gather-nr2fma.c",
2951 "src/math/sigmoid-avx2-rr1-lut64-p2-gather-nr2fma1adj.c",
2952 "src/math/sigmoid-avx2-rr1-p5-div.c",
2953 "src/math/sigmoid-avx2-rr1-p5-nr1fma.c",
2954 "src/math/sigmoid-avx2-rr1-p5-nr2fma.c",
2955 "src/math/sigmoid-avx2-rr2-lut64-p2-gather-div.c",
2956 "src/math/sigmoid-avx2-rr2-lut64-p2-gather-nr1fma.c",
2957 "src/math/sigmoid-avx2-rr2-lut64-p2-gather-nr2fma.c",
2958 "src/math/sigmoid-avx2-rr2-lut64-p2-gather-nr2fma1adj.c",
2959 "src/math/sigmoid-avx2-rr2-p5-div.c",
2960 "src/math/sigmoid-avx2-rr2-p5-nr1fma.c",
2961 "src/math/sigmoid-avx2-rr2-p5-nr2fma.c",
Marat Dukhand65a1522020-08-04 19:28:18 -07002962 "src/qs8-dwconv/gen/up8x9-minmax-avx2-mul32.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002963 "src/qs8-dwconv/gen/up16x9-minmax-avx2-mul16.c",
Marat Dukhand65a1522020-08-04 19:28:18 -07002964 "src/qs8-dwconv/gen/up16x9-minmax-avx2-mul32.c",
2965 "src/qs8-dwconv/gen/up24x9-minmax-avx2-mul32.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002966 "src/qs8-dwconv/gen/up32x9-minmax-avx2-mul16.c",
Marat Dukhand65a1522020-08-04 19:28:18 -07002967 "src/qs8-dwconv/gen/up32x9-minmax-avx2-mul32.c",
Marat Dukhan12809522020-08-02 22:23:51 -07002968 "src/qs8-gemm/gen/1x8c8-minmax-avx2.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002969 "src/qs8-gemm/gen/1x8c8-xw-minmax-avx2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002970 "src/qs8-gemm/gen/2x8c8-minmax-avx2.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002971 "src/qs8-gemm/gen/2x8c8-xw-minmax-avx2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002972 "src/qs8-gemm/gen/3x8c8-minmax-avx2.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002973 "src/qs8-gemm/gen/3x8c8-xw-minmax-avx2.c",
Marat Dukhan12809522020-08-02 22:23:51 -07002974 "src/qs8-igemm/gen/1x8c8-minmax-avx2.c",
2975 "src/qs8-igemm/gen/2x8c8-minmax-avx2.c",
2976 "src/qs8-igemm/gen/3x8c8-minmax-avx2.c",
Marat Dukhane6dc0b62020-09-08 23:57:14 -07002977 "src/qs8-vadd/gen/minmax-avx2-mul32-ld64-x8.c",
2978 "src/qs8-vadd/gen/minmax-avx2-mul32-ld64-x16.c",
2979 "src/qs8-vadd/gen/minmax-avx2-mul32-ld64-x24.c",
2980 "src/qs8-vadd/gen/minmax-avx2-mul32-ld64-x32.c",
2981 "src/qs8-vaddc/gen/minmax-avx2-mul32-ld64-x8.c",
2982 "src/qs8-vaddc/gen/minmax-avx2-mul32-ld64-x16.c",
2983 "src/qs8-vaddc/gen/minmax-avx2-mul32-ld64-x24.c",
2984 "src/qs8-vaddc/gen/minmax-avx2-mul32-ld64-x32.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002985]
2986
Marat Dukhan08c4a432019-10-03 09:29:21 -07002987AVX512F_UKERNELS = [
Marat Dukhan5c5fa962020-03-10 18:38:33 -07002988 "src/f32-clamp/gen/avx512f-x16.c",
2989 "src/f32-clamp/gen/avx512f-x32.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002990 "src/f32-dwconv/gen/up16x4-minmax-avx512f-acc2.c",
2991 "src/f32-dwconv/gen/up16x4-minmax-avx512f.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002992 "src/f32-dwconv/gen/up16x9-minmax-avx512f-acc2.c",
2993 "src/f32-dwconv/gen/up16x9-minmax-avx512f.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002994 "src/f32-dwconv/gen/up16x25-minmax-avx512f-acc2.c",
2995 "src/f32-dwconv/gen/up16x25-minmax-avx512f.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002996 "src/f32-dwconv/gen/up32x4-minmax-avx512f-acc2.c",
2997 "src/f32-dwconv/gen/up32x4-minmax-avx512f.c",
2998 "src/f32-dwconv/gen/up32x9-minmax-avx512f-acc2.c",
2999 "src/f32-dwconv/gen/up32x9-minmax-avx512f.c",
3000 "src/f32-dwconv/gen/up32x25-minmax-avx512f-acc2.c",
3001 "src/f32-dwconv/gen/up32x25-minmax-avx512f.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07003002 "src/f32-gemm/gen-inc/1x16inc-minmax-avx512f-broadcast.c",
3003 "src/f32-gemm/gen-inc/4x16inc-minmax-avx512f-broadcast.c",
3004 "src/f32-gemm/gen-inc/5x16inc-minmax-avx512f-broadcast.c",
3005 "src/f32-gemm/gen-inc/6x16inc-minmax-avx512f-broadcast.c",
3006 "src/f32-gemm/gen-inc/7x16inc-minmax-avx512f-broadcast.c",
3007 "src/f32-gemm/gen-inc/8x16inc-minmax-avx512f-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003008 "src/f32-gemm/gen/1x16-minmax-avx512f-broadcast.c",
3009 "src/f32-gemm/gen/4x16-minmax-avx512f-broadcast.c",
3010 "src/f32-gemm/gen/5x16-minmax-avx512f-broadcast.c",
3011 "src/f32-gemm/gen/6x16-minmax-avx512f-broadcast.c",
3012 "src/f32-gemm/gen/7x16-minmax-avx512f-broadcast.c",
3013 "src/f32-gemm/gen/8x16-minmax-avx512f-broadcast.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07003014 "src/f32-hswish/gen/hswish-avx512f-x16.c",
3015 "src/f32-hswish/gen/hswish-avx512f-x32.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07003016 "src/f32-igemm/gen/1x16-minmax-avx512f-broadcast.c",
3017 "src/f32-igemm/gen/4x16-minmax-avx512f-broadcast.c",
3018 "src/f32-igemm/gen/5x16-minmax-avx512f-broadcast.c",
3019 "src/f32-igemm/gen/6x16-minmax-avx512f-broadcast.c",
3020 "src/f32-igemm/gen/7x16-minmax-avx512f-broadcast.c",
3021 "src/f32-igemm/gen/8x16-minmax-avx512f-broadcast.c",
Marat Dukhan90eca0a2020-03-11 00:52:23 -07003022 "src/f32-prelu/gen/avx512f-2x16.c",
3023 "src/f32-prelu/gen/avx512f-2x32.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003024 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128-acc2.c",
3025 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003026 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003027 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x144-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003028 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x144.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003029 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160-acc2.c",
3030 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003031 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003032 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc2.c",
3033 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc3.c",
3034 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc6.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003035 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003036 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128-acc2.c",
3037 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003038 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003039 "src/f32-raddextexp/gen/avx512f-p5-scalef-x144-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003040 "src/f32-raddextexp/gen/avx512f-p5-scalef-x144.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003041 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160-acc2.c",
3042 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003043 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003044 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc2.c",
3045 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc3.c",
3046 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc6.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003047 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003048 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128-acc2.c",
3049 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003050 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003051 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x144-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003052 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x144.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003053 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160-acc2.c",
3054 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003055 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003056 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc2.c",
3057 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc3.c",
3058 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc6.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003059 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07003060 "src/f32-relu/gen/avx512f-x16.c",
3061 "src/f32-relu/gen/avx512f-x32.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003062 "src/f32-rmax/avx512f.c",
Marat Dukhand9ca7e62020-09-23 23:45:29 -07003063 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x16.c",
3064 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x32.c",
3065 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x48.c",
3066 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x64.c",
3067 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x80.c",
3068 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x96.c",
3069 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x112.c",
3070 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x128.c",
3071 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x16.c",
3072 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x32.c",
3073 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x48.c",
3074 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x64.c",
3075 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x80.c",
3076 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x96.c",
3077 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x112.c",
3078 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003079 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x16.c",
3080 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x32.c",
3081 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x48.c",
3082 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x64.c",
3083 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x80.c",
3084 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x96.c",
3085 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x112.c",
3086 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x128.c",
3087 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x16.c",
3088 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x32.c",
3089 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x48.c",
3090 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x64.c",
3091 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x80.c",
3092 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x96.c",
3093 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x112.c",
3094 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x128.c",
Marat Dukhand9ca7e62020-09-23 23:45:29 -07003095 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x16.c",
3096 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x32.c",
3097 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x48.c",
3098 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x64.c",
3099 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x80.c",
3100 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x96.c",
3101 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x112.c",
3102 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x128.c",
3103 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x16.c",
3104 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x32.c",
3105 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x48.c",
3106 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x64.c",
3107 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x80.c",
3108 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x96.c",
3109 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x112.c",
3110 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x128.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07003111 "src/f32-vbinary/gen/vadd-minmax-avx512f-x16.c",
3112 "src/f32-vbinary/gen/vadd-minmax-avx512f-x32.c",
3113 "src/f32-vbinary/gen/vaddc-minmax-avx512f-x16.c",
3114 "src/f32-vbinary/gen/vaddc-minmax-avx512f-x32.c",
3115 "src/f32-vbinary/gen/vdiv-minmax-avx512f-x16.c",
3116 "src/f32-vbinary/gen/vdiv-minmax-avx512f-x32.c",
3117 "src/f32-vbinary/gen/vdivc-minmax-avx512f-x16.c",
3118 "src/f32-vbinary/gen/vdivc-minmax-avx512f-x32.c",
Marat Dukhan9a88efe2019-12-10 15:54:24 -08003119 "src/f32-vbinary/gen/vmax-avx512f-x16.c",
3120 "src/f32-vbinary/gen/vmax-avx512f-x32.c",
3121 "src/f32-vbinary/gen/vmaxc-avx512f-x16.c",
3122 "src/f32-vbinary/gen/vmaxc-avx512f-x32.c",
3123 "src/f32-vbinary/gen/vmin-avx512f-x16.c",
3124 "src/f32-vbinary/gen/vmin-avx512f-x32.c",
3125 "src/f32-vbinary/gen/vminc-avx512f-x16.c",
3126 "src/f32-vbinary/gen/vminc-avx512f-x32.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07003127 "src/f32-vbinary/gen/vmul-minmax-avx512f-x16.c",
3128 "src/f32-vbinary/gen/vmul-minmax-avx512f-x32.c",
3129 "src/f32-vbinary/gen/vmulc-minmax-avx512f-x16.c",
3130 "src/f32-vbinary/gen/vmulc-minmax-avx512f-x32.c",
3131 "src/f32-vbinary/gen/vrdivc-minmax-avx512f-x16.c",
3132 "src/f32-vbinary/gen/vrdivc-minmax-avx512f-x32.c",
3133 "src/f32-vbinary/gen/vrsubc-minmax-avx512f-x16.c",
3134 "src/f32-vbinary/gen/vrsubc-minmax-avx512f-x32.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07003135 "src/f32-vbinary/gen/vsqrdiff-avx512f-x16.c",
3136 "src/f32-vbinary/gen/vsqrdiff-avx512f-x32.c",
3137 "src/f32-vbinary/gen/vsqrdiffc-avx512f-x16.c",
3138 "src/f32-vbinary/gen/vsqrdiffc-avx512f-x32.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07003139 "src/f32-vbinary/gen/vsub-minmax-avx512f-x16.c",
3140 "src/f32-vbinary/gen/vsub-minmax-avx512f-x32.c",
3141 "src/f32-vbinary/gen/vsubc-minmax-avx512f-x16.c",
3142 "src/f32-vbinary/gen/vsubc-minmax-avx512f-x32.c",
Marat Dukhaned6baaf2020-12-01 15:07:08 -08003143 "src/f32-velu/gen/velu-avx512f-rr1-lut16-p3-perm-x16.c",
3144 "src/f32-velu/gen/velu-avx512f-rr1-lut16-p3-perm-x32.c",
3145 "src/f32-velu/gen/velu-avx512f-rr1-lut16-p3-perm-x48.c",
3146 "src/f32-velu/gen/velu-avx512f-rr1-lut16-p3-perm-x64.c",
3147 "src/f32-velu/gen/velu-avx512f-rr1-lut16-p3-perm-x80.c",
3148 "src/f32-velu/gen/velu-avx512f-rr1-lut16-p3-perm-x96.c",
3149 "src/f32-velu/gen/velu-avx512f-rr1-lut16-p3-perm-x112.c",
3150 "src/f32-velu/gen/velu-avx512f-rr1-lut16-p3-perm-x128.c",
3151 "src/f32-velu/gen/velu-avx512f-rr1-p6-x16.c",
3152 "src/f32-velu/gen/velu-avx512f-rr1-p6-x32.c",
3153 "src/f32-velu/gen/velu-avx512f-rr1-p6-x48.c",
3154 "src/f32-velu/gen/velu-avx512f-rr1-p6-x64.c",
3155 "src/f32-velu/gen/velu-avx512f-rr1-p6-x80.c",
3156 "src/f32-velu/gen/velu-avx512f-rr1-p6-x96.c",
3157 "src/f32-velu/gen/velu-avx512f-rr1-p6-x112.c",
3158 "src/f32-velu/gen/velu-avx512f-rr1-p6-x128.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07003159 "src/f32-vlrelu/gen/vlrelu-avx512f-x16.c",
3160 "src/f32-vlrelu/gen/vlrelu-avx512f-x32.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003161 "src/f32-vrnd/gen/vrndd-avx512f-x16.c",
3162 "src/f32-vrnd/gen/vrndd-avx512f-x32.c",
3163 "src/f32-vrnd/gen/vrndne-avx512f-x16.c",
3164 "src/f32-vrnd/gen/vrndne-avx512f-x32.c",
3165 "src/f32-vrnd/gen/vrndu-avx512f-x16.c",
3166 "src/f32-vrnd/gen/vrndu-avx512f-x32.c",
3167 "src/f32-vrnd/gen/vrndz-avx512f-x16.c",
3168 "src/f32-vrnd/gen/vrndz-avx512f-x32.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07003169 "src/f32-vscale/avx512f-x64.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003170 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x16.c",
3171 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x32.c",
3172 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x48.c",
3173 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x64.c",
3174 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x80.c",
3175 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x96.c",
3176 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x112.c",
3177 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x128.c",
3178 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x144.c",
3179 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x160.c",
3180 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x176.c",
3181 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x192.c",
3182 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x16.c",
3183 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x32.c",
3184 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x48.c",
3185 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x64.c",
3186 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x80.c",
3187 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x96.c",
3188 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x112.c",
3189 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x128.c",
3190 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x144.c",
3191 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x160.c",
3192 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x176.c",
3193 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x192.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07003194 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x16.c",
3195 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x32.c",
3196 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x48.c",
3197 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x64.c",
3198 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x80.c",
3199 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x96.c",
3200 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x112.c",
3201 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x128.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07003202 "src/f32-vunary/gen/vabs-avx512f-x16.c",
3203 "src/f32-vunary/gen/vabs-avx512f-x32.c",
3204 "src/f32-vunary/gen/vneg-avx512f-x16.c",
3205 "src/f32-vunary/gen/vneg-avx512f-x32.c",
3206 "src/f32-vunary/gen/vsqr-avx512f-x16.c",
3207 "src/f32-vunary/gen/vsqr-avx512f-x32.c",
Marat Dukhanb7633f22020-11-20 16:34:56 -08003208 "src/math/exp-avx512f-rr2-lut16-p3-perm-scalef.c",
3209 "src/math/exp-avx512f-rr2-lut16-p3-perm.c",
3210 "src/math/exp-avx512f-rr2-lut32-p2-perm2-scalef.c",
3211 "src/math/exp-avx512f-rr2-lut32-p2-perm2.c",
3212 "src/math/exp-avx512f-rr2-p5-scalef.c",
3213 "src/math/exp-avx512f-rr2-p5.c",
Marat Dukhande390d42020-11-29 19:32:18 -08003214 "src/math/expm1minus-avx512f-rr1-lut16-p3-perm.c",
3215 "src/math/expm1minus-avx512f-rr1-p6.c",
Marat Dukhan98ba4412019-10-23 02:14:28 -07003216 "src/math/extexp-avx512f-p5.c",
Marat Dukhan6a34c5f2020-09-22 21:44:15 -07003217 "src/math/sigmoid-avx512f-rr1-lut16-p3-perm-scalef-div.c",
Marat Dukhan6a34c5f2020-09-22 21:44:15 -07003218 "src/math/sigmoid-avx512f-rr1-lut16-p3-perm-scalef-nr1fma.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003219 "src/math/sigmoid-avx512f-rr1-lut16-p3-perm-scalef-nr1fma1adj.c",
Marat Dukhan6a34c5f2020-09-22 21:44:15 -07003220 "src/math/sigmoid-avx512f-rr1-lut32-p2-perm2-scalef-div.c",
Marat Dukhan6a34c5f2020-09-22 21:44:15 -07003221 "src/math/sigmoid-avx512f-rr1-lut32-p2-perm2-scalef-nr1fma.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003222 "src/math/sigmoid-avx512f-rr1-lut32-p2-perm2-scalef-nr1fma1adj.c",
Marat Dukhan36173d22020-10-15 17:14:26 -07003223 "src/math/sigmoid-avx512f-rr1-lut64-p2-gather-scalef-div.c",
Marat Dukhan36173d22020-10-15 17:14:26 -07003224 "src/math/sigmoid-avx512f-rr1-lut64-p2-gather-scalef-nr1fma.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003225 "src/math/sigmoid-avx512f-rr1-lut64-p2-gather-scalef-nr1fma1adj.c",
3226 "src/math/sigmoid-avx512f-rr1-p5-scalef-div.c",
3227 "src/math/sigmoid-avx512f-rr1-p5-scalef-nr1fma.c",
3228 "src/math/sigmoid-avx512f-rr1-p5-scalef-nr1fma1adj.c",
3229 "src/math/sigmoid-avx512f-rr2-lut16-p3-perm-scalef-div.c",
3230 "src/math/sigmoid-avx512f-rr2-lut16-p3-perm-scalef-nr1fma.c",
3231 "src/math/sigmoid-avx512f-rr2-lut16-p3-perm-scalef-nr1fma1adj.c",
3232 "src/math/sigmoid-avx512f-rr2-lut32-p2-perm2-scalef-div.c",
3233 "src/math/sigmoid-avx512f-rr2-lut32-p2-perm2-scalef-nr1fma.c",
3234 "src/math/sigmoid-avx512f-rr2-lut32-p2-perm2-scalef-nr1fma1adj.c",
Marat Dukhan36173d22020-10-15 17:14:26 -07003235 "src/math/sigmoid-avx512f-rr2-lut64-p2-gather-scalef-div.c",
Marat Dukhan36173d22020-10-15 17:14:26 -07003236 "src/math/sigmoid-avx512f-rr2-lut64-p2-gather-scalef-nr1fma.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003237 "src/math/sigmoid-avx512f-rr2-lut64-p2-gather-scalef-nr1fma1adj.c",
3238 "src/math/sigmoid-avx512f-rr2-p5-scalef-div.c",
3239 "src/math/sigmoid-avx512f-rr2-p5-scalef-nr1fma.c",
3240 "src/math/sigmoid-avx512f-rr2-p5-scalef-nr1fma1adj.c",
Marat Dukhan84000762020-06-29 18:38:43 -07003241 "src/math/sqrt-avx512f-nr1fma.c",
Marat Dukhan84000762020-06-29 18:38:43 -07003242 "src/math/sqrt-avx512f-nr1fma1adj.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003243 "src/math/sqrt-avx512f-nr2fma.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003244]
3245
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07003246AVX512SKX_UKERNELS = [
Marat Dukhan2ffc5e62020-09-06 22:33:38 -07003247 "src/qs8-dwconv/gen/up16x9-minmax-avx512skx-mul32.c",
3248 "src/qs8-dwconv/gen/up32x9-minmax-avx512skx-mul32.c",
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07003249 "src/qs8-gemm/gen/1x16c8-minmax-avx512skx.c",
3250 "src/qs8-gemm/gen/2x16c8-minmax-avx512skx.c",
3251 "src/qs8-gemm/gen/3x16c8-minmax-avx512skx.c",
3252 "src/qs8-gemm/gen/4x16c8-minmax-avx512skx.c",
3253 "src/qs8-igemm/gen/1x16c8-minmax-avx512skx.c",
3254 "src/qs8-igemm/gen/2x16c8-minmax-avx512skx.c",
3255 "src/qs8-igemm/gen/3x16c8-minmax-avx512skx.c",
3256 "src/qs8-igemm/gen/4x16c8-minmax-avx512skx.c",
3257]
3258
Frank Barchardbcedc082020-08-17 18:00:51 -07003259WASM32_ASM_UKERNELS = [
3260 "src/f32-relu/wasm_shr_x1.S",
3261 "src/f32-relu/wasm_shr_x2.S",
3262 "src/f32-relu/wasm_shr_x4.S",
3263]
3264
Marat Dukhan08c4a432019-10-03 09:29:21 -07003265AARCH32_ASM_UKERNELS = [
Marat Dukhan32f93812020-05-17 20:31:21 -07003266 "src/f32-gemm/4x4-aarch32-vfp-ld64.S",
Marat Dukhan3b98f6b2020-05-17 10:09:22 -07003267 "src/f32-gemm/4x4-minmax-aarch32-vfp-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003268 "src/f32-gemm/4x8-minmax-aarch32-neon-cortex-a53.S",
3269 "src/f32-gemm/4x8-minmax-aarch32-neon-cortex-a55.S",
Frank Barchard490febe2020-07-16 18:42:17 -07003270 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-cortex-a7.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003271 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-cortex-a75.S",
Frank Barchard569561d2020-06-17 13:11:12 -07003272 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-ld64.S",
Frank Barchard490febe2020-07-16 18:42:17 -07003273 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-pld-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003274 "src/f32-igemm/4x8-minmax-aarch32-neon-cortex-a53.S",
3275 "src/f32-igemm/4x8-minmax-aarch32-neon-cortex-a55.S",
Frank Barchard490febe2020-07-16 18:42:17 -07003276 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-cortex-a7.S",
3277 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-cortex-a75.S",
3278 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-ld64.S",
3279 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-pld-cortex-a75.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003280]
3281
3282AARCH64_ASM_UKERNELS = [
Frank Barchardbddfbcd2020-04-15 12:32:41 -07003283 "src/f16-gemm/gen-inc/1x8inc-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchard04336c12020-10-22 16:48:55 -07003284 "src/f16-gemm/gen-inc/1x16inc-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchardbddfbcd2020-04-15 12:32:41 -07003285 "src/f16-gemm/gen-inc/4x8inc-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchard04336c12020-10-22 16:48:55 -07003286 "src/f16-gemm/gen-inc/4x16inc-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchardbddfbcd2020-04-15 12:32:41 -07003287 "src/f16-gemm/gen-inc/6x8inc-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchard04336c12020-10-22 16:48:55 -07003288 "src/f16-gemm/gen-inc/6x16inc-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchard3b8e5662020-04-20 12:12:53 -07003289 "src/f16-gemm/gen-inc/8x8inc-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchard04336c12020-10-22 16:48:55 -07003290 "src/f16-gemm/gen/1x8-minmax-aarch64-neonfp16arith-ld64.S",
3291 "src/f16-gemm/gen/1x16-minmax-aarch64-neonfp16arith-ld32.S",
3292 "src/f16-gemm/gen/4x8-minmax-aarch64-neonfp16arith-ld64.S",
3293 "src/f16-gemm/gen/4x16-minmax-aarch64-neonfp16arith-ld32.S",
3294 "src/f16-gemm/gen/6x8-minmax-aarch64-neonfp16arith-ld64.S",
3295 "src/f16-gemm/gen/6x16-minmax-aarch64-neonfp16arith-ld32.S",
3296 "src/f16-gemm/gen/8x8-minmax-aarch64-neonfp16arith-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003297 "src/f32-dwconv/up4x9-minmax-aarch64-neonfma-cortex-a55.S",
3298 "src/f32-dwconv/up4x9-minmax-aarch64-neonfma.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003299 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a53.S",
3300 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a57.S",
3301 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a75.S",
Frank Barchard04336c12020-10-22 16:48:55 -07003302 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-ld64.S",
3303 "src/f32-gemm/gen-inc/1x12inc-minmax-aarch64-neonfma-cortex-a53.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003304 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a53.S",
3305 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a55.S",
3306 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a57.S",
3307 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003308 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-ld64.S",
Frank Barchard04336c12020-10-22 16:48:55 -07003309 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-ld128.S",
3310 "src/f32-gemm/gen-inc/4x12inc-minmax-aarch64-neonfma-cortex-a53.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003311 "src/f32-gemm/gen-inc/5x8inc-minmax-aarch64-neonfma-cortex-a57.S",
3312 "src/f32-gemm/gen-inc/5x8inc-minmax-aarch64-neonfma-cortex-a75.S",
3313 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a53.S",
3314 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a55.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003315 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a57.S",
Frank Barchard04336c12020-10-22 16:48:55 -07003316 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a73.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003317 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003318 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-ld64.S",
Frank Barchard04336c12020-10-22 16:48:55 -07003319 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-ld128.S",
3320 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a53.S",
3321 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a57.S",
3322 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a75.S",
3323 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-ld64.S",
3324 "src/f32-gemm/gen/1x12-minmax-aarch64-neonfma-cortex-a53.S",
3325 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a53.S",
3326 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a55.S",
3327 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a57.S",
3328 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a75.S",
3329 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-ld64.S",
3330 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-ld128.S",
3331 "src/f32-gemm/gen/4x12-minmax-aarch64-neonfma-cortex-a53.S",
3332 "src/f32-gemm/gen/5x8-minmax-aarch64-neonfma-cortex-a57.S",
3333 "src/f32-gemm/gen/5x8-minmax-aarch64-neonfma-cortex-a75.S",
3334 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a53.S",
3335 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a55.S",
3336 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a57.S",
3337 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a73.S",
3338 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a75.S",
3339 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-ld64.S",
3340 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-ld128.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003341 "src/f32-igemm/1x8-minmax-aarch64-neonfma-cortex-a53.S",
Frank Barchard04336c12020-10-22 16:48:55 -07003342 "src/f32-igemm/1x12-minmax-aarch64-neonfma-cortex-a53.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003343 "src/f32-igemm/4x8-minmax-aarch64-neonfma-cortex-a53.S",
3344 "src/f32-igemm/4x8-minmax-aarch64-neonfma-cortex-a55.S",
Frank Barchard04336c12020-10-22 16:48:55 -07003345 "src/f32-igemm/4x12-minmax-aarch64-neonfma-cortex-a53.S",
3346 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a53.S",
3347 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a55.S",
3348 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a73.S",
3349 "src/f32-igemm/gen/1x8-minmax-aarch64-neonfma-cortex-a57.S",
3350 "src/f32-igemm/gen/1x8-minmax-aarch64-neonfma-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003351 "src/f32-igemm/gen/4x8-minmax-aarch64-neonfma-cortex-a57.S",
3352 "src/f32-igemm/gen/4x8-minmax-aarch64-neonfma-cortex-a75.S",
3353 "src/f32-igemm/gen/5x8-minmax-aarch64-neonfma-cortex-a57.S",
3354 "src/f32-igemm/gen/5x8-minmax-aarch64-neonfma-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003355 "src/f32-igemm/gen/6x8-minmax-aarch64-neonfma-cortex-a57.S",
3356 "src/f32-igemm/gen/6x8-minmax-aarch64-neonfma-cortex-a75.S",
Frank Barchardf1fd89e2020-10-06 23:27:11 -07003357 "src/qs8-gemm/1x16c4-aarch64-neondot-ld64.S",
Frank Barchardf2742c42020-10-12 21:16:46 -07003358 "src/qs8-gemm/4x16c4-aarch64-neondot-cortex-a55.S",
Frank Barcharda5237a52020-10-06 16:18:46 -07003359 "src/qs8-gemm/4x16c4-aarch64-neondot-ld64.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003360]
3361
Marat Dukhan1b354632020-03-23 12:50:22 -07003362INTERNAL_MICROKERNEL_HDRS = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003363 "src/xnnpack/argmaxpool.h",
3364 "src/xnnpack/avgpool.h",
3365 "src/xnnpack/clamp.h",
3366 "src/xnnpack/common.h",
3367 "src/xnnpack/conv.h",
Yury Kartynnike7841862020-11-04 18:22:18 -08003368 "src/xnnpack/depthtospace.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003369 "src/xnnpack/dwconv.h",
Frank Barchard04336c12020-10-22 16:48:55 -07003370 "src/xnnpack/fill.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003371 "src/xnnpack/gavgpool.h",
3372 "src/xnnpack/gemm.h",
3373 "src/xnnpack/hswish.h",
Marat Dukhan660fd192020-03-10 04:55:30 -07003374 "src/xnnpack/ibilinear.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003375 "src/xnnpack/igemm.h",
Marat Dukhancfb31342019-12-05 10:42:57 -08003376 "src/xnnpack/intrinsics-polyfill.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003377 "src/xnnpack/lut.h",
3378 "src/xnnpack/math.h",
3379 "src/xnnpack/maxpool.h",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003380 "src/xnnpack/memory.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003381 "src/xnnpack/packx.h",
3382 "src/xnnpack/pad.h",
3383 "src/xnnpack/params.h",
3384 "src/xnnpack/pavgpool.h",
3385 "src/xnnpack/ppmm.h",
3386 "src/xnnpack/prelu.h",
Marat Dukhan97579532019-10-18 16:40:39 -07003387 "src/xnnpack/raddexpminusmax.h",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07003388 "src/xnnpack/raddextexp.h",
Marat Dukhan97579532019-10-18 16:40:39 -07003389 "src/xnnpack/raddstoreexpminusmax.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003390 "src/xnnpack/rmax.h",
3391 "src/xnnpack/scalar-utils.h",
3392 "src/xnnpack/spmm.h",
3393 "src/xnnpack/unpool.h",
3394 "src/xnnpack/vadd.h",
Marat Dukhan1e782c42019-11-21 17:02:40 -08003395 "src/xnnpack/vbinary.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003396 "src/xnnpack/vmulcaddc.h",
Marat Dukhan05ac8e32019-10-21 15:39:33 -07003397 "src/xnnpack/vscale.h",
Marat Dukhan97579532019-10-18 16:40:39 -07003398 "src/xnnpack/vscaleexpminusmax.h",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07003399 "src/xnnpack/vscaleextexp.h",
Marat Dukhan1e782c42019-11-21 17:02:40 -08003400 "src/xnnpack/vunary.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003401 "src/xnnpack/zip.h",
Marat Dukhan1b354632020-03-23 12:50:22 -07003402]
3403
3404INTERNAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003405 "include/xnnpack.h",
3406 "src/xnnpack/allocator.h",
3407 "src/xnnpack/compute.h",
3408 "src/xnnpack/im2col.h",
3409 "src/xnnpack/indirection.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003410 "src/xnnpack/math-stubs.h",
Chao Mei6ddfc602020-05-13 22:29:36 -07003411 "src/xnnpack/memory-planner.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003412 "src/xnnpack/operator.h",
3413 "src/xnnpack/pack.h",
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07003414 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003415 "src/xnnpack/requantization-stubs.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003416 "src/xnnpack/requantization.h",
Marat Dukhan1d75a542020-02-03 12:23:01 -08003417 "src/xnnpack/subgraph.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003418]
3419
Marat Dukhan1b354632020-03-23 12:50:22 -07003420ACCURACY_EVAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003421 "src/xnnpack/math-stubs.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003422]
3423
Marat Dukhan1b354632020-03-23 12:50:22 -07003424MICROKERNEL_BENCHMARK_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003425 "include/xnnpack.h",
Frank Barchard35db7d02020-10-26 13:37:34 -07003426 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003427]
3428
Marat Dukhan1b354632020-03-23 12:50:22 -07003429MICROKERNEL_TEST_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Frank Barchard35db7d02020-10-26 13:37:34 -07003430 "include/xnnpack.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003431 "src/xnnpack/isa-checks.h",
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07003432 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003433 "src/xnnpack/requantization.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003434]
3435
3436OPERATOR_TEST_PARAMS_HDRS = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003437 "src/xnnpack/common.h",
Frank Barchard04336c12020-10-22 16:48:55 -07003438 "src/xnnpack/params.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003439]
3440
3441WEIGHTS_PACK_HDRS = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003442 "src/xnnpack/compute.h",
Frank Barchard04336c12020-10-22 16:48:55 -07003443 "src/xnnpack/operator.h",
3444 "src/xnnpack/pack.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003445]
3446
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07003447LOGGING_COPTS = select({
3448 # No logging in optimized mode
3449 ":optimized_build": ["-DXNN_LOG_LEVEL=0"],
3450 # Full logging in debug mode
3451 ":debug_build": ["-DXNN_LOG_LEVEL=5"],
3452 # Error-only logging in default (fastbuild) mode
3453 "//conditions:default": ["-DXNN_LOG_LEVEL=2"],
3454})
3455
Marat Dukhan3b59de22020-06-03 20:15:19 -07003456LOGGING_SRCS = select({
3457 # No logging in optimized mode
3458 ":optimized_build": [],
3459 "//conditions:default": [
3460 "src/operator-strings.c",
3461 "src/subgraph-strings.c",
3462 ],
3463})
3464
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07003465LOGGING_HDRS = [
3466 "src/xnnpack/log.h",
3467]
3468
Marat Dukhan08c4a432019-10-03 09:29:21 -07003469xnnpack_cc_library(
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003470 name = "tables",
3471 srcs = TABLE_SRCS,
3472 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003473 gcc_copts = xnnpack_gcc_std_copts(),
3474 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003475)
3476
3477xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003478 name = "scalar_ukernels",
3479 srcs = SCALAR_UKERNELS,
3480 hdrs = INTERNAL_HDRS,
3481 aarch32_copts = ["-marm"],
Marat Dukhan10a38082020-04-17 03:58:35 -07003482 gcc_copts = xnnpack_gcc_std_copts(),
3483 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003484 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003485 ":tables",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003486 "@FP16",
3487 "@FXdiv",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003488 "@pthreadpool",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003489 ],
3490)
3491
3492xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003493 name = "scalar_ukernels_test_mode",
3494 srcs = SCALAR_UKERNELS,
3495 hdrs = INTERNAL_HDRS,
3496 aarch32_copts = ["-marm"],
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 "@FXdiv",
3507 "@pthreadpool",
3508 ],
3509)
3510
3511xnnpack_cc_library(
Marat Dukhan436ebe62019-12-04 15:10:12 -08003512 name = "wasm_ukernels",
3513 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003514 gcc_copts = xnnpack_gcc_std_copts(),
3515 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan436ebe62019-12-04 15:10:12 -08003516 wasm_srcs = WASM_UKERNELS,
Marat Dukhan1c6e3892020-06-25 23:56:10 -07003517 wasmsimd_srcs = WASM_UKERNELS + WASMSIMD_UKERNELS,
Marat Dukhan436ebe62019-12-04 15:10:12 -08003518 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003519 ":tables",
Marat Dukhan436ebe62019-12-04 15:10:12 -08003520 "@FP16",
3521 "@FXdiv",
3522 "@pthreadpool",
3523 ],
3524)
3525
3526xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003527 name = "wasm_ukernels_test_mode",
3528 hdrs = INTERNAL_HDRS,
3529 copts = [
3530 "-UNDEBUG",
3531 "-DXNN_TEST_MODE=1",
3532 ],
3533 gcc_copts = xnnpack_gcc_std_copts(),
3534 msvc_copts = xnnpack_msvc_std_copts(),
3535 wasm_srcs = WASM_UKERNELS,
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07003536 wasmsimd_srcs = WASM_UKERNELS + WASMSIMD_UKERNELS,
Marat Dukhan33fcf782020-05-24 14:27:15 -07003537 deps = [
3538 ":tables",
3539 "@FP16",
3540 "@FXdiv",
3541 "@pthreadpool",
3542 ],
3543)
3544
3545xnnpack_cc_library(
Marat Dukhan102e9ea2020-12-05 22:35:18 -08003546 name = "psimd_accmath_ukernels",
3547 hdrs = INTERNAL_HDRS,
3548 aarch32_copts = [
3549 "-marm",
3550 "-mfpu=neon",
3551 ],
3552 gcc_copts = xnnpack_gcc_std_copts(),
3553 gcc_x86_copts = ["-msse2"],
3554 msvc_copts = xnnpack_msvc_std_copts(),
3555 optimized_copts = [
3556 "-O3",
3557 ],
3558 psimd_srcs = PSIMD_ACCMATH_UKERNELS,
3559 deps = [
3560 ":tables",
3561 "@FP16",
3562 "@psimd",
3563 "@pthreadpool",
3564 ],
3565)
3566
3567xnnpack_cc_library(
3568 name = "psimd_accmath_ukernels_test_mode",
3569 hdrs = INTERNAL_HDRS,
3570 aarch32_copts = [
3571 "-marm",
3572 "-mfpu=neon",
3573 ],
3574 copts = [
3575 "-UNDEBUG",
3576 "-DXNN_TEST_MODE=1",
3577 ],
3578 gcc_copts = xnnpack_gcc_std_copts(),
3579 gcc_x86_copts = ["-msse2"],
3580 msvc_copts = xnnpack_msvc_std_copts(),
3581 optimized_copts = [
3582 "-O3",
3583 ],
3584 psimd_srcs = PSIMD_ACCMATH_UKERNELS,
3585 deps = [
3586 ":tables",
3587 "@FP16",
3588 "@psimd",
3589 "@pthreadpool",
3590 ],
3591)
3592
3593xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003594 name = "neon_ukernels",
3595 hdrs = INTERNAL_HDRS,
3596 aarch32_copts = [
3597 "-marm",
Marat Dukhan8853b822020-05-07 12:19:01 -07003598 "-march=armv7-a",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003599 "-mfpu=neon",
3600 ],
3601 aarch32_srcs = NEON_UKERNELS,
3602 aarch64_srcs = NEON_UKERNELS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003603 gcc_copts = xnnpack_gcc_std_copts(),
3604 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08003605 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003606 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003607 "@FP16",
3608 "@pthreadpool",
3609 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003610)
3611
3612xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003613 name = "neon_ukernels_test_mode",
3614 hdrs = INTERNAL_HDRS,
3615 aarch32_copts = [
3616 "-marm",
3617 "-march=armv7-a",
3618 "-mfpu=neon",
3619 ],
3620 aarch32_srcs = NEON_UKERNELS,
3621 aarch64_srcs = NEON_UKERNELS,
3622 copts = [
3623 "-UNDEBUG",
3624 "-DXNN_TEST_MODE=1",
3625 ],
3626 gcc_copts = xnnpack_gcc_std_copts(),
3627 msvc_copts = xnnpack_msvc_std_copts(),
3628 deps = [
3629 ":tables",
3630 "@FP16",
3631 "@pthreadpool",
3632 ],
3633)
3634
3635xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003636 name = "neonfma_ukernels",
3637 hdrs = INTERNAL_HDRS,
3638 aarch32_copts = [
3639 "-marm",
Marat Dukhan8853b822020-05-07 12:19:01 -07003640 "-march=armv7-a",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003641 "-mfpu=neon-vfpv4",
3642 ],
3643 aarch32_srcs = NEONFMA_UKERNELS,
3644 aarch64_srcs = NEONFMA_UKERNELS + AARCH64_NEONFMA_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07003645 apple_aarch32_copts = [
3646 "-mcpu=swift",
3647 "-mtune=generic",
3648 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07003649 gcc_copts = xnnpack_gcc_std_copts(),
3650 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08003651 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003652 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003653 "@FP16",
3654 "@pthreadpool",
3655 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003656)
3657
3658xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003659 name = "neonfma_ukernels_test_mode",
3660 hdrs = INTERNAL_HDRS,
3661 aarch32_copts = [
3662 "-marm",
3663 "-march=armv7-a",
3664 "-mfpu=neon-vfpv4",
3665 ],
3666 aarch32_srcs = NEONFMA_UKERNELS,
3667 aarch64_srcs = NEONFMA_UKERNELS + AARCH64_NEONFMA_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07003668 apple_aarch32_copts = [
3669 "-mcpu=swift",
3670 "-mtune=generic",
3671 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07003672 copts = [
3673 "-UNDEBUG",
3674 "-DXNN_TEST_MODE=1",
3675 ],
3676 gcc_copts = xnnpack_gcc_std_copts(),
3677 msvc_copts = xnnpack_msvc_std_copts(),
3678 deps = [
3679 ":tables",
3680 "@FP16",
3681 "@pthreadpool",
3682 ],
3683)
3684
3685xnnpack_cc_library(
Marat Dukhan8853b822020-05-07 12:19:01 -07003686 name = "neonv8_ukernels",
3687 hdrs = INTERNAL_HDRS,
3688 aarch32_copts = [
3689 "-marm",
3690 "-march=armv8-a",
3691 "-mfpu=neon-fp-armv8",
3692 ],
3693 aarch32_srcs = NEONV8_UKERNELS,
3694 aarch64_srcs = NEONV8_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07003695 apple_aarch32_copts = [
3696 "-mcpu=cyclone",
3697 "-mtune=generic",
3698 ],
Marat Dukhan8853b822020-05-07 12:19:01 -07003699 gcc_copts = xnnpack_gcc_std_copts(),
3700 msvc_copts = xnnpack_msvc_std_copts(),
3701 deps = [
3702 ":tables",
3703 "@FP16",
3704 "@pthreadpool",
3705 ],
3706)
3707
3708xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003709 name = "neonv8_ukernels_test_mode",
3710 hdrs = INTERNAL_HDRS,
3711 aarch32_copts = [
3712 "-marm",
3713 "-march=armv8-a",
3714 "-mfpu=neon-fp-armv8",
3715 ],
3716 aarch32_srcs = NEONV8_UKERNELS,
3717 aarch64_srcs = NEONV8_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07003718 apple_aarch32_copts = [
3719 "-mcpu=cyclone",
3720 "-mtune=generic",
3721 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07003722 copts = [
3723 "-UNDEBUG",
3724 "-DXNN_TEST_MODE=1",
3725 ],
3726 gcc_copts = xnnpack_gcc_std_copts(),
3727 msvc_copts = xnnpack_msvc_std_copts(),
3728 deps = [
3729 ":tables",
3730 "@FP16",
3731 "@pthreadpool",
3732 ],
3733)
3734
3735xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003736 name = "neonfp16arith_ukernels",
3737 hdrs = INTERNAL_HDRS,
3738 aarch64_copts = ["-march=armv8.2-a+fp16"],
3739 aarch64_srcs = AARCH64_NEONFP16ARITH_UKERNELS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003740 gcc_copts = xnnpack_gcc_std_copts(),
3741 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08003742 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003743 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003744 "@FP16",
3745 "@pthreadpool",
3746 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003747)
3748
3749xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003750 name = "neonfp16arith_ukernels_test_mode",
3751 hdrs = INTERNAL_HDRS,
3752 aarch64_copts = ["-march=armv8.2-a+fp16"],
3753 aarch64_srcs = AARCH64_NEONFP16ARITH_UKERNELS,
3754 copts = [
3755 "-UNDEBUG",
3756 "-DXNN_TEST_MODE=1",
3757 ],
3758 gcc_copts = xnnpack_gcc_std_copts(),
3759 msvc_copts = xnnpack_msvc_std_copts(),
3760 deps = [
3761 ":tables",
3762 "@FP16",
3763 "@pthreadpool",
3764 ],
3765)
3766
3767xnnpack_cc_library(
Benoit Jacoba9644732020-08-13 12:48:55 -07003768 name = "neondot_ukernels",
3769 hdrs = INTERNAL_HDRS,
Marat Dukhan799ac752020-08-13 16:08:03 -07003770 aarch32_copts = [
3771 "-marm",
3772 "-march=armv8.2-a+dotprod",
3773 "-mfpu=neon-fp-armv8",
3774 ],
Benoit Jacoba9644732020-08-13 12:48:55 -07003775 aarch32_srcs = NEONDOT_UKERNELS,
3776 aarch64_copts = ["-march=armv8.2-a+dotprod"],
3777 aarch64_srcs = NEONDOT_UKERNELS,
3778 gcc_copts = xnnpack_gcc_std_copts(),
3779 msvc_copts = xnnpack_msvc_std_copts(),
3780 deps = [
3781 ":tables",
3782 "@FP16",
3783 "@pthreadpool",
3784 ],
3785)
3786
3787xnnpack_cc_library(
3788 name = "neondot_ukernels_test_mode",
3789 hdrs = INTERNAL_HDRS,
Marat Dukhan799ac752020-08-13 16:08:03 -07003790 aarch32_copts = [
3791 "-marm",
3792 "-march=armv8.2-a+dotprod",
3793 "-mfpu=neon-fp-armv8",
3794 ],
Benoit Jacoba9644732020-08-13 12:48:55 -07003795 aarch32_srcs = NEONDOT_UKERNELS,
3796 aarch64_copts = ["-march=armv8.2-a+dotprod"],
3797 aarch64_srcs = NEONDOT_UKERNELS,
3798 copts = [
3799 "-UNDEBUG",
3800 "-DXNN_TEST_MODE=1",
3801 ],
3802 gcc_copts = xnnpack_gcc_std_copts(),
3803 msvc_copts = xnnpack_msvc_std_copts(),
3804 deps = [
3805 ":tables",
3806 "@FP16",
3807 "@pthreadpool",
3808 ],
3809)
3810
3811xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003812 name = "sse2_ukernels",
3813 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003814 gcc_copts = xnnpack_gcc_std_copts(),
3815 gcc_x86_copts = ["-msse2"],
3816 msvc_copts = xnnpack_msvc_std_copts(),
3817 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003818 x86_srcs = SSE_UKERNELS + SSE2_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08003819 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003820 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003821 "@FP16",
3822 "@pthreadpool",
3823 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003824)
3825
3826xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003827 name = "sse2_ukernels_test_mode",
3828 hdrs = INTERNAL_HDRS,
3829 copts = [
3830 "-UNDEBUG",
3831 "-DXNN_TEST_MODE=1",
3832 ],
3833 gcc_copts = xnnpack_gcc_std_copts(),
3834 gcc_x86_copts = ["-msse2"],
3835 msvc_copts = xnnpack_msvc_std_copts(),
3836 msvc_x86_32_copts = ["/arch:SSE2"],
3837 x86_srcs = SSE_UKERNELS + SSE2_UKERNELS,
3838 deps = [
3839 ":tables",
3840 "@FP16",
3841 "@pthreadpool",
3842 ],
3843)
3844
3845xnnpack_cc_library(
Marat Dukhanfe7acb62020-03-09 19:30:05 -07003846 name = "ssse3_ukernels",
3847 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003848 gcc_copts = xnnpack_gcc_std_copts(),
3849 gcc_x86_copts = ["-mssse3"],
3850 msvc_copts = xnnpack_msvc_std_copts(),
3851 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhanfe7acb62020-03-09 19:30:05 -07003852 x86_srcs = SSSE3_UKERNELS,
3853 deps = [
3854 ":tables",
3855 "@FP16",
3856 "@pthreadpool",
3857 ],
3858)
3859
3860xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003861 name = "ssse3_ukernels_test_mode",
3862 hdrs = INTERNAL_HDRS,
3863 copts = [
3864 "-UNDEBUG",
3865 "-DXNN_TEST_MODE=1",
3866 ],
3867 gcc_copts = xnnpack_gcc_std_copts(),
3868 gcc_x86_copts = ["-mssse3"],
3869 msvc_copts = xnnpack_msvc_std_copts(),
3870 msvc_x86_32_copts = ["/arch:SSE2"],
3871 x86_srcs = SSSE3_UKERNELS,
3872 deps = [
3873 ":tables",
3874 "@FP16",
3875 "@pthreadpool",
3876 ],
3877)
3878
3879xnnpack_cc_library(
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08003880 name = "sse41_ukernels",
3881 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003882 gcc_copts = xnnpack_gcc_std_copts(),
3883 gcc_x86_copts = ["-msse4.1"],
3884 msvc_copts = xnnpack_msvc_std_copts(),
3885 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08003886 x86_srcs = SSE41_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08003887 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003888 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003889 "@FP16",
3890 "@pthreadpool",
3891 ],
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08003892)
3893
3894xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003895 name = "sse41_ukernels_test_mode",
3896 hdrs = INTERNAL_HDRS,
3897 copts = [
3898 "-UNDEBUG",
3899 "-DXNN_TEST_MODE=1",
3900 ],
3901 gcc_copts = xnnpack_gcc_std_copts(),
3902 gcc_x86_copts = ["-msse4.1"],
3903 msvc_copts = xnnpack_msvc_std_copts(),
3904 msvc_x86_32_copts = ["/arch:SSE2"],
3905 x86_srcs = SSE41_UKERNELS,
3906 deps = [
3907 ":tables",
3908 "@FP16",
3909 "@pthreadpool",
3910 ],
3911)
3912
3913xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003914 name = "avx_ukernels",
3915 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003916 gcc_copts = xnnpack_gcc_std_copts(),
3917 gcc_x86_copts = ["-mavx"],
3918 msvc_copts = xnnpack_msvc_std_copts(),
3919 msvc_x86_32_copts = ["/arch:AVX"],
3920 msvc_x86_64_copts = ["/arch:AVX"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003921 x86_srcs = AVX_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08003922 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003923 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003924 "@FP16",
3925 "@pthreadpool",
3926 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003927)
3928
3929xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003930 name = "avx_ukernels_test_mode",
3931 hdrs = INTERNAL_HDRS,
3932 copts = [
3933 "-UNDEBUG",
3934 "-DXNN_TEST_MODE=1",
3935 ],
3936 gcc_copts = xnnpack_gcc_std_copts(),
3937 gcc_x86_copts = ["-mavx"],
3938 msvc_copts = xnnpack_msvc_std_copts(),
3939 msvc_x86_32_copts = ["/arch:AVX"],
3940 msvc_x86_64_copts = ["/arch:AVX"],
3941 x86_srcs = AVX_UKERNELS,
3942 deps = [
3943 ":tables",
3944 "@FP16",
3945 "@pthreadpool",
3946 ],
3947)
3948
3949xnnpack_cc_library(
Marat Dukhan1566fee2020-08-02 21:55:41 -07003950 name = "xop_ukernels",
3951 hdrs = INTERNAL_HDRS,
3952 gcc_copts = xnnpack_gcc_std_copts(),
3953 gcc_x86_copts = ["-mxop"],
3954 msvc_copts = xnnpack_msvc_std_copts(),
3955 msvc_x86_32_copts = ["/arch:AVX"],
3956 msvc_x86_64_copts = ["/arch:AVX"],
3957 x86_srcs = XOP_UKERNELS,
3958 deps = [
3959 ":tables",
3960 "@FP16",
3961 "@pthreadpool",
3962 ],
3963)
3964
3965xnnpack_cc_library(
3966 name = "xop_ukernels_test_mode",
3967 hdrs = INTERNAL_HDRS,
3968 copts = [
3969 "-UNDEBUG",
3970 "-DXNN_TEST_MODE=1",
3971 ],
3972 gcc_copts = xnnpack_gcc_std_copts(),
3973 gcc_x86_copts = ["-mxop"],
3974 msvc_copts = xnnpack_msvc_std_copts(),
3975 msvc_x86_32_copts = ["/arch:AVX"],
3976 msvc_x86_64_copts = ["/arch:AVX"],
3977 x86_srcs = XOP_UKERNELS,
3978 deps = [
3979 ":tables",
3980 "@FP16",
3981 "@pthreadpool",
3982 ],
3983)
3984
3985xnnpack_cc_library(
Marat Dukhanfda12b82019-11-21 12:27:59 -08003986 name = "fma3_ukernels",
3987 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003988 gcc_copts = xnnpack_gcc_std_copts(),
3989 gcc_x86_copts = ["-mfma"],
3990 msvc_copts = xnnpack_msvc_std_copts(),
3991 msvc_x86_32_copts = ["/arch:AVX"],
3992 msvc_x86_64_copts = ["/arch:AVX"],
Marat Dukhanfda12b82019-11-21 12:27:59 -08003993 x86_srcs = FMA3_UKERNELS,
3994 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003995 ":tables",
Marat Dukhanfda12b82019-11-21 12:27:59 -08003996 "@FP16",
3997 "@pthreadpool",
3998 ],
3999)
4000
4001xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07004002 name = "fma3_ukernels_test_mode",
4003 hdrs = INTERNAL_HDRS,
4004 copts = [
4005 "-UNDEBUG",
4006 "-DXNN_TEST_MODE=1",
4007 ],
4008 gcc_copts = xnnpack_gcc_std_copts(),
4009 gcc_x86_copts = ["-mfma"],
4010 msvc_copts = xnnpack_msvc_std_copts(),
4011 msvc_x86_32_copts = ["/arch:AVX"],
4012 msvc_x86_64_copts = ["/arch:AVX"],
4013 x86_srcs = FMA3_UKERNELS,
4014 deps = [
4015 ":tables",
4016 "@FP16",
4017 "@pthreadpool",
4018 ],
4019)
4020
4021xnnpack_cc_library(
Marat Dukhan6adff4e2019-10-14 18:32:07 -07004022 name = "avx2_ukernels",
4023 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004024 gcc_copts = xnnpack_gcc_std_copts(),
4025 gcc_x86_copts = [
Marat Dukhan6adff4e2019-10-14 18:32:07 -07004026 "-mfma",
4027 "-mavx2",
4028 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07004029 msvc_copts = xnnpack_msvc_std_copts(),
4030 msvc_x86_32_copts = ["/arch:AVX2"],
4031 msvc_x86_64_copts = ["/arch:AVX2"],
Marat Dukhan6adff4e2019-10-14 18:32:07 -07004032 x86_srcs = AVX2_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08004033 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08004034 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08004035 "@FP16",
4036 "@pthreadpool",
4037 ],
Marat Dukhan6adff4e2019-10-14 18:32:07 -07004038)
4039
4040xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07004041 name = "avx2_ukernels_test_mode",
4042 hdrs = INTERNAL_HDRS,
4043 copts = [
4044 "-UNDEBUG",
4045 "-DXNN_TEST_MODE=1",
4046 ],
4047 gcc_copts = xnnpack_gcc_std_copts(),
4048 gcc_x86_copts = [
4049 "-mfma",
4050 "-mavx2",
4051 ],
4052 msvc_copts = xnnpack_msvc_std_copts(),
4053 msvc_x86_32_copts = ["/arch:AVX2"],
4054 msvc_x86_64_copts = ["/arch:AVX2"],
4055 x86_srcs = AVX2_UKERNELS,
4056 deps = [
4057 ":tables",
4058 "@FP16",
4059 "@pthreadpool",
4060 ],
4061)
4062
4063xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004064 name = "avx512f_ukernels",
4065 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004066 gcc_copts = xnnpack_gcc_std_copts(),
4067 gcc_x86_copts = ["-mavx512f"],
4068 mingw_copts = ["-fno-asynchronous-unwind-tables"],
4069 msvc_copts = xnnpack_msvc_std_copts(),
4070 msvc_x86_32_copts = ["/arch:AVX512"],
4071 msvc_x86_64_copts = ["/arch:AVX512"],
4072 msys_copts = ["-fno-asynchronous-unwind-tables"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004073 x86_srcs = AVX512F_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08004074 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08004075 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08004076 "@FP16",
4077 "@pthreadpool",
4078 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004079)
4080
4081xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07004082 name = "avx512f_ukernels_test_mode",
4083 hdrs = INTERNAL_HDRS,
4084 copts = [
4085 "-UNDEBUG",
4086 "-DXNN_TEST_MODE=1",
4087 ],
4088 gcc_copts = xnnpack_gcc_std_copts(),
4089 gcc_x86_copts = ["-mavx512f"],
4090 mingw_copts = ["-fno-asynchronous-unwind-tables"],
4091 msvc_copts = xnnpack_msvc_std_copts(),
4092 msvc_x86_32_copts = ["/arch:AVX512"],
4093 msvc_x86_64_copts = ["/arch:AVX512"],
4094 msys_copts = ["-fno-asynchronous-unwind-tables"],
4095 x86_srcs = AVX512F_UKERNELS,
4096 deps = [
4097 ":tables",
4098 "@FP16",
4099 "@pthreadpool",
4100 ],
4101)
4102
4103xnnpack_cc_library(
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07004104 name = "avx512skx_ukernels",
4105 hdrs = INTERNAL_HDRS,
4106 gcc_copts = xnnpack_gcc_std_copts(),
4107 gcc_x86_copts = [
4108 "-mavx512f",
4109 "-mavx512cd",
4110 "-mavx512bw",
4111 "-mavx512dq",
4112 "-mavx512vl",
4113 ],
4114 mingw_copts = ["-fno-asynchronous-unwind-tables"],
4115 msvc_copts = xnnpack_msvc_std_copts(),
4116 msvc_x86_32_copts = ["/arch:AVX512"],
4117 msvc_x86_64_copts = ["/arch:AVX512"],
4118 msys_copts = ["-fno-asynchronous-unwind-tables"],
4119 x86_srcs = AVX512SKX_UKERNELS,
4120 deps = [
4121 ":tables",
4122 "@FP16",
4123 "@pthreadpool",
4124 ],
4125)
4126
4127xnnpack_cc_library(
4128 name = "avx512skx_ukernels_test_mode",
4129 hdrs = INTERNAL_HDRS,
4130 copts = [
4131 "-UNDEBUG",
4132 "-DXNN_TEST_MODE=1",
4133 ],
4134 gcc_copts = xnnpack_gcc_std_copts(),
4135 gcc_x86_copts = [
4136 "-mavx512f",
4137 "-mavx512cd",
4138 "-mavx512bw",
4139 "-mavx512dq",
4140 "-mavx512vl",
4141 ],
4142 mingw_copts = ["-fno-asynchronous-unwind-tables"],
4143 msvc_copts = xnnpack_msvc_std_copts(),
4144 msvc_x86_32_copts = ["/arch:AVX512"],
4145 msvc_x86_64_copts = ["/arch:AVX512"],
4146 msys_copts = ["-fno-asynchronous-unwind-tables"],
4147 x86_srcs = AVX512SKX_UKERNELS,
4148 deps = [
4149 ":tables",
4150 "@FP16",
4151 "@pthreadpool",
4152 ],
4153)
4154
4155xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004156 name = "asm_ukernels",
4157 hdrs = ["src/xnnpack/assembly.h"],
4158 aarch32_srcs = AARCH32_ASM_UKERNELS,
Frank Barchard31bb45b2020-10-06 00:26:33 -07004159 aarch64_copts = ["-march=armv8.2-a+fp16+dotprod"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004160 aarch64_srcs = AARCH64_ASM_UKERNELS,
Frank Barchardbcedc082020-08-17 18:00:51 -07004161 wasm_srcs = WASM32_ASM_UKERNELS,
4162 wasmsimd_srcs = WASM32_ASM_UKERNELS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07004163)
4164
Marat Dukhan3b59de22020-06-03 20:15:19 -07004165xnnpack_cc_library(
4166 name = "logging_utils",
4167 srcs = LOGGING_SRCS,
4168 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
4169 copts = LOGGING_COPTS + [
4170 "-Isrc",
4171 "-Iinclude",
4172 ] + select({
4173 ":debug_build": [],
4174 "//conditions:default": xnnpack_min_size_copts(),
4175 }),
4176 gcc_copts = xnnpack_gcc_std_copts(),
4177 msvc_copts = xnnpack_msvc_std_copts(),
4178 visibility = xnnpack_visibility(),
4179 deps = [
4180 "@FP16",
4181 "@clog",
4182 "@pthreadpool",
4183 ],
4184)
4185
Marat Dukhan08c4a432019-10-03 09:29:21 -07004186xnnpack_aggregate_library(
4187 name = "ukernels",
4188 aarch32_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07004189 ":neon_ukernels",
4190 ":neonfma_ukernels",
Marat Dukhan8853b822020-05-07 12:19:01 -07004191 ":neonv8_ukernels",
Benoit Jacoba9644732020-08-13 12:48:55 -07004192 ":neondot_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004193 ":asm_ukernels",
4194 ],
4195 aarch64_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07004196 ":neon_ukernels",
4197 ":neonfma_ukernels",
Marat Dukhan8853b822020-05-07 12:19:01 -07004198 ":neonv8_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004199 ":neonfp16arith_ukernels",
Benoit Jacoba9644732020-08-13 12:48:55 -07004200 ":neondot_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004201 ":asm_ukernels",
4202 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07004203 generic_deps = [
4204 ":scalar_ukernels",
4205 ],
Marat Dukhan102e9ea2020-12-05 22:35:18 -08004206 psimd_deps = [
4207 ":psimd_accmath_ukernels",
4208 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07004209 wasm_deps = [
4210 ":wasm_ukernels",
Frank Barchardbcedc082020-08-17 18:00:51 -07004211 ":asm_ukernels",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004212 ],
4213 wasmsimd_deps = [
4214 ":wasm_ukernels",
Frank Barchardbcedc082020-08-17 18:00:51 -07004215 ":asm_ukernels",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004216 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004217 x86_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07004218 ":sse2_ukernels",
Marat Dukhanfe7acb62020-03-09 19:30:05 -07004219 ":ssse3_ukernels",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08004220 ":sse41_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004221 ":avx_ukernels",
Marat Dukhan1566fee2020-08-02 21:55:41 -07004222 ":xop_ukernels",
Marat Dukhanfda12b82019-11-21 12:27:59 -08004223 ":fma3_ukernels",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07004224 ":avx2_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004225 ":avx512f_ukernels",
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07004226 ":avx512skx_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004227 ],
4228)
4229
Marat Dukhan33fcf782020-05-24 14:27:15 -07004230xnnpack_aggregate_library(
4231 name = "ukernels_test_mode",
4232 aarch32_deps = [
4233 ":neon_ukernels_test_mode",
4234 ":neonfma_ukernels_test_mode",
4235 ":neonv8_ukernels_test_mode",
Benoit Jacoba9644732020-08-13 12:48:55 -07004236 ":neondot_ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004237 ":asm_ukernels",
4238 ],
4239 aarch64_deps = [
4240 ":neon_ukernels_test_mode",
4241 ":neonfma_ukernels_test_mode",
4242 ":neonv8_ukernels_test_mode",
4243 ":neonfp16arith_ukernels_test_mode",
Benoit Jacoba9644732020-08-13 12:48:55 -07004244 ":neondot_ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004245 ":asm_ukernels",
4246 ],
4247 generic_deps = [
4248 ":scalar_ukernels_test_mode",
4249 ],
Marat Dukhan102e9ea2020-12-05 22:35:18 -08004250 psimd_deps = [
4251 ":psimd_accmath_ukernels_test_mode",
4252 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07004253 wasm_deps = [
4254 ":wasm_ukernels_test_mode",
Frank Barchardbcedc082020-08-17 18:00:51 -07004255 ":asm_ukernels",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004256 ],
4257 wasmsimd_deps = [
4258 ":wasm_ukernels_test_mode",
Frank Barchardbcedc082020-08-17 18:00:51 -07004259 ":asm_ukernels",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004260 ],
4261 x86_deps = [
4262 ":sse2_ukernels_test_mode",
4263 ":ssse3_ukernels_test_mode",
4264 ":sse41_ukernels_test_mode",
4265 ":avx_ukernels_test_mode",
Marat Dukhan1566fee2020-08-02 21:55:41 -07004266 ":xop_ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004267 ":fma3_ukernels_test_mode",
4268 ":avx2_ukernels_test_mode",
4269 ":avx512f_ukernels_test_mode",
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07004270 ":avx512skx_ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004271 ],
4272)
4273
Marat Dukhan86df3842020-06-17 14:03:31 -07004274# This target must be compatible with Android NDK r17 and compile with Clang 6.
4275# It is similar to :ukernels target, but excludes microkernels which require
4276# NEON FP16 arithmetics, NEON Dot Product, and newer intrinsics.
Marat Dukhanae046f52020-06-15 13:16:14 -07004277xnnpack_aggregate_library(
Marat Dukhan86df3842020-06-17 14:03:31 -07004278 name = "ukernels_max_compatibility",
Marat Dukhanae046f52020-06-15 13:16:14 -07004279 aarch32_deps = [
4280 ":neon_ukernels",
4281 ":neonfma_ukernels",
4282 ":neonv8_ukernels",
4283 ":asm_ukernels",
4284 ],
4285 aarch64_deps = [
4286 ":neon_ukernels",
4287 ":neonfma_ukernels",
4288 ":neonv8_ukernels",
4289 ":asm_ukernels",
4290 ],
4291 generic_deps = [
4292 ":scalar_ukernels",
4293 ],
Marat Dukhan102e9ea2020-12-05 22:35:18 -08004294 psimd_deps = [
4295 ":psimd_accmath_ukernels",
4296 ],
Marat Dukhanae046f52020-06-15 13:16:14 -07004297 wasm_deps = [
4298 ":wasm_ukernels",
4299 ],
4300 wasmsimd_deps = [
4301 ":wasm_ukernels",
4302 ],
4303 x86_deps = [
4304 ":sse2_ukernels",
4305 ":ssse3_ukernels",
4306 ":sse41_ukernels",
4307 ":avx_ukernels",
Marat Dukhan1566fee2020-08-02 21:55:41 -07004308 ":xop_ukernels",
Marat Dukhanae046f52020-06-15 13:16:14 -07004309 ":fma3_ukernels",
4310 ":avx2_ukernels",
4311 ":avx512f_ukernels",
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07004312 ":avx512skx_ukernels",
Marat Dukhanae046f52020-06-15 13:16:14 -07004313 ],
4314)
4315
Marat Dukhan08c4a432019-10-03 09:29:21 -07004316xnnpack_cc_library(
4317 name = "im2col",
4318 srcs = ["src/im2col.c"],
4319 hdrs = [
4320 "src/xnnpack/common.h",
4321 "src/xnnpack/im2col.h",
4322 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07004323 gcc_copts = xnnpack_gcc_std_copts(),
4324 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004325)
4326
4327xnnpack_cc_library(
4328 name = "indirection",
4329 srcs = ["src/indirection.c"],
4330 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004331 gcc_copts = xnnpack_gcc_std_copts(),
4332 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004333 deps = [
4334 "@FP16",
4335 "@FXdiv",
4336 "@pthreadpool",
4337 ],
4338)
4339
4340xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07004341 name = "indirection_test_mode",
4342 srcs = ["src/indirection.c"],
4343 hdrs = INTERNAL_HDRS,
4344 copts = [
4345 "-UNDEBUG",
4346 "-DXNN_TEST_MODE=1",
4347 ],
4348 gcc_copts = xnnpack_gcc_std_copts(),
4349 msvc_copts = xnnpack_msvc_std_copts(),
4350 deps = [
4351 "@FP16",
4352 "@FXdiv",
4353 "@pthreadpool",
4354 ],
4355)
4356
4357xnnpack_cc_library(
Marat Dukhanab582382020-07-06 13:32:08 -07004358 name = "packing",
4359 srcs = ["src/packing.c"],
4360 hdrs = INTERNAL_HDRS,
4361 gcc_copts = xnnpack_gcc_std_copts(),
4362 msvc_copts = xnnpack_msvc_std_copts(),
4363 deps = [
4364 "@FP16",
4365 "@FXdiv",
4366 "@pthreadpool",
4367 ],
4368)
4369
4370xnnpack_cc_library(
4371 name = "packing_test_mode",
4372 srcs = ["src/packing.c"],
4373 hdrs = INTERNAL_HDRS,
4374 copts = [
4375 "-UNDEBUG",
4376 "-DXNN_TEST_MODE=1",
4377 ],
4378 gcc_copts = xnnpack_gcc_std_copts(),
4379 msvc_copts = xnnpack_msvc_std_copts(),
4380 deps = [
4381 "@FP16",
4382 "@FXdiv",
4383 "@pthreadpool",
4384 ],
4385)
4386
4387xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004388 name = "operator_run",
4389 srcs = ["src/operator-run.c"],
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07004390 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004391 copts = LOGGING_COPTS + select({
Marat Dukhan05702cf2020-03-26 15:41:33 -07004392 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4393 "//conditions:default": [],
4394 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07004395 gcc_copts = xnnpack_gcc_std_copts(),
4396 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004397 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07004398 ":logging_utils",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004399 "@FP16",
4400 "@FXdiv",
4401 "@clog",
4402 "@pthreadpool",
4403 ],
4404)
4405
Chao Mei6ddfc602020-05-13 22:29:36 -07004406xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07004407 name = "operator_run_test_mode",
4408 srcs = ["src/operator-run.c"],
4409 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
4410 copts = LOGGING_COPTS + [
4411 "-UNDEBUG",
4412 "-DXNN_TEST_MODE=1",
4413 ] + select({
4414 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4415 "//conditions:default": [],
4416 }),
4417 gcc_copts = xnnpack_gcc_std_copts(),
4418 msvc_copts = xnnpack_msvc_std_copts(),
4419 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07004420 ":logging_utils",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004421 "@FP16",
4422 "@FXdiv",
4423 "@clog",
4424 "@pthreadpool",
4425 ],
4426)
4427
4428xnnpack_cc_library(
Chao Mei6ddfc602020-05-13 22:29:36 -07004429 name = "memory_planner",
4430 srcs = ["src/memory-planner.c"],
4431 hdrs = INTERNAL_HDRS,
4432 defines = select({
4433 ":xnn_enable_memopt_explicit_true": ["XNN_ENABLE_MEMOPT=1"],
4434 ":xnn_enable_memopt_explicit_false": ["XNN_ENABLE_MEMOPT=0"],
4435 "//conditions:default": ["XNN_ENABLE_MEMOPT=1"],
4436 }),
4437 gcc_copts = xnnpack_gcc_std_copts(),
4438 msvc_copts = xnnpack_msvc_std_copts(),
4439 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07004440 ":logging_utils",
Chao Mei6ddfc602020-05-13 22:29:36 -07004441 "@pthreadpool",
4442 ],
4443)
4444
Marat Dukhan33fcf782020-05-24 14:27:15 -07004445xnnpack_cc_library(
4446 name = "memory_planner_test_mode",
4447 srcs = ["src/memory-planner.c"],
4448 hdrs = INTERNAL_HDRS,
4449 copts = [
4450 "-UNDEBUG",
4451 "-DXNN_TEST_MODE=1",
4452 ],
4453 defines = select({
4454 ":xnn_enable_memopt_explicit_true": ["XNN_ENABLE_MEMOPT=1"],
4455 ":xnn_enable_memopt_explicit_false": ["XNN_ENABLE_MEMOPT=0"],
4456 "//conditions:default": ["XNN_ENABLE_MEMOPT=1"],
4457 }),
4458 gcc_copts = xnnpack_gcc_std_copts(),
4459 msvc_copts = xnnpack_msvc_std_copts(),
4460 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07004461 ":logging_utils",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004462 "@pthreadpool",
4463 ],
4464)
4465
Marat Dukhan08c4a432019-10-03 09:29:21 -07004466cc_library(
4467 name = "enable_assembly",
4468 defines = select({
4469 ":xnn_enable_assembly_explicit_true": ["XNN_ENABLE_ASSEMBLY=1"],
4470 ":xnn_enable_assembly_explicit_false": ["XNN_ENABLE_ASSEMBLY=0"],
Frank Barchard810171d2019-10-10 10:34:51 -07004471 "//conditions:default": ["XNN_ENABLE_ASSEMBLY=1"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004472 }),
4473)
4474
Marat Dukhan9de90e02020-06-18 16:04:12 -07004475cc_library(
4476 name = "enable_sparse",
4477 defines = select({
4478 ":xnn_enable_sparse_explicit_true": ["XNN_ENABLE_SPARSE=1"],
4479 ":xnn_enable_sparse_explicit_false": ["XNN_ENABLE_SPARSE=0"],
Marat Dukhan8b283aa2020-06-18 16:48:57 -07004480 "//conditions:default": ["XNN_ENABLE_SPARSE=0"],
Marat Dukhan9de90e02020-06-18 16:04:12 -07004481 }),
4482)
4483
Marat Dukhancf056b22019-10-07 10:26:29 -07004484xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004485 name = "operators",
4486 srcs = OPERATOR_SRCS + [
Marat Dukhan04f03be2019-11-19 12:36:47 -08004487 "src/memory.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004488 "src/operator-delete.c",
Marat Dukhancf056b22019-10-07 10:26:29 -07004489 ],
4490 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004491 copts = LOGGING_COPTS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07004492 "-Isrc",
4493 "-Iinclude",
4494 ] + select({
4495 ":debug_build": [],
4496 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07004497 }) + select({
4498 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4499 "//conditions:default": [],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004500 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07004501 gcc_copts = xnnpack_gcc_std_copts(),
4502 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004503 deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07004504 ":indirection",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004505 ":logging_utils",
Marat Dukhanab582382020-07-06 13:32:08 -07004506 ":packing",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004507 "@FP16",
4508 "@FXdiv",
4509 "@clog",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004510 "@pthreadpool",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004511 ],
4512)
4513
Marat Dukhan10a38082020-04-17 03:58:35 -07004514xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07004515 name = "operators_test_mode",
4516 srcs = OPERATOR_SRCS + [
4517 "src/memory.c",
4518 "src/operator-delete.c",
4519 ],
4520 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
4521 copts = LOGGING_COPTS + [
4522 "-Isrc",
4523 "-Iinclude",
4524 "-UNDEBUG",
4525 "-DXNN_TEST_MODE=1",
4526 ] + select({
4527 ":debug_build": [],
4528 "//conditions:default": xnnpack_min_size_copts(),
4529 }) + select({
4530 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4531 "//conditions:default": [],
4532 }),
4533 gcc_copts = xnnpack_gcc_std_copts(),
4534 msvc_copts = xnnpack_msvc_std_copts(),
4535 deps = [
4536 ":indirection_test_mode",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004537 ":logging_utils",
Marat Dukhanab582382020-07-06 13:32:08 -07004538 ":packing_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004539 "@FP16",
4540 "@FXdiv",
4541 "@clog",
4542 "@pthreadpool",
4543 ],
4544)
4545
4546xnnpack_cc_library(
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004547 name = "XNNPACK",
4548 srcs = [
4549 "src/init.c",
Marat Dukhanccfdbd12020-02-03 14:27:45 -08004550 "src/runtime.c",
4551 "src/subgraph.c",
4552 "src/tensor.c",
Marat Dukhanf03da0d2020-06-10 16:00:20 -07004553 ] + SUBGRAPH_SRCS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004554 hdrs = ["include/xnnpack.h"],
4555 copts = LOGGING_COPTS + [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004556 "-Isrc",
4557 "-Iinclude",
4558 ] + select({
4559 ":debug_build": [],
4560 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07004561 }) + select({
4562 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4563 "//conditions:default": [],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004564 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07004565 gcc_copts = xnnpack_gcc_std_copts(),
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004566 includes = ["include"],
Marat Dukhan10a38082020-04-17 03:58:35 -07004567 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004568 visibility = xnnpack_visibility(),
4569 deps = [
4570 ":enable_assembly",
Marat Dukhan9de90e02020-06-18 16:04:12 -07004571 ":enable_sparse",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004572 ":logging_utils",
Chao Mei6ddfc602020-05-13 22:29:36 -07004573 ":memory_planner",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004574 ":operator_run",
4575 ":operators",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004576 ":ukernels",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004577 "@clog",
Marat Dukhanab2946c2020-05-21 20:04:13 -07004578 "@FP16",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004579 "@pthreadpool",
Marat Dukhand343c222019-10-07 09:22:14 -07004580 ] + select({
4581 ":emscripten": [],
4582 "//conditions:default": ["@cpuinfo"],
4583 }),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004584)
4585
Marat Dukhan10a38082020-04-17 03:58:35 -07004586xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07004587 name = "XNNPACK_test_mode",
4588 srcs = [
4589 "src/init.c",
4590 "src/runtime.c",
4591 "src/subgraph.c",
4592 "src/tensor.c",
Marat Dukhanf03da0d2020-06-10 16:00:20 -07004593 ] + SUBGRAPH_SRCS,
Marat Dukhan33fcf782020-05-24 14:27:15 -07004594 hdrs = ["include/xnnpack.h"],
4595 copts = LOGGING_COPTS + [
4596 "-Isrc",
4597 "-Iinclude",
4598 "-UNDEBUG",
4599 "-DXNN_TEST_MODE=1",
4600 ] + select({
4601 ":debug_build": [],
4602 "//conditions:default": xnnpack_min_size_copts(),
4603 }) + select({
4604 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4605 "//conditions:default": [],
4606 }),
4607 gcc_copts = xnnpack_gcc_std_copts(),
4608 includes = ["include"],
4609 msvc_copts = xnnpack_msvc_std_copts(),
4610 visibility = xnnpack_visibility(),
4611 deps = [
4612 ":enable_assembly",
Marat Dukhan9de90e02020-06-18 16:04:12 -07004613 ":enable_sparse",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004614 ":logging_utils",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004615 ":memory_planner_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004616 ":operator_run_test_mode",
4617 ":operators_test_mode",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004618 ":ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004619 "@clog",
4620 "@FP16",
4621 "@pthreadpool",
4622 ] + select({
4623 ":emscripten": [],
4624 "//conditions:default": ["@cpuinfo"],
4625 }),
4626)
4627
4628xnnpack_cc_library(
Marat Dukhanae046f52020-06-15 13:16:14 -07004629 name = "xnnpack_f32",
4630 srcs = [
4631 "src/init.c",
4632 "src/runtime.c",
4633 "src/subgraph.c",
4634 "src/tensor.c",
4635 ] + SUBGRAPH_SRCS,
4636 hdrs = ["include/xnnpack.h"],
4637 copts = LOGGING_COPTS + [
4638 "-Isrc",
4639 "-Iinclude",
4640 ] + select({
4641 ":debug_build": [],
4642 "//conditions:default": xnnpack_min_size_copts(),
4643 }) + select({
4644 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4645 "//conditions:default": [],
4646 }),
4647 defines = [
Marat Dukhan16f1e1a2020-08-04 16:38:22 -07004648 "XNN_NO_QS8_OPERATORS",
Marat Dukhan08b7a972020-07-14 18:17:29 -07004649 "XNN_NO_QU8_OPERATORS",
Marat Dukhanae046f52020-06-15 13:16:14 -07004650 "XNN_NO_U8_OPERATORS",
4651 "XNN_NO_X8_OPERATORS",
4652 "XNN_NO_F16_OPERATORS",
4653 "XNN_NO_X16_OPERATORS",
4654 ],
4655 gcc_copts = xnnpack_gcc_std_copts(),
4656 includes = ["include"],
4657 msvc_copts = xnnpack_msvc_std_copts(),
4658 visibility = xnnpack_visibility(),
4659 deps = [
4660 ":enable_assembly",
Marat Dukhan9de90e02020-06-18 16:04:12 -07004661 ":enable_sparse",
Marat Dukhanae046f52020-06-15 13:16:14 -07004662 ":logging_utils",
4663 ":memory_planner",
4664 ":operator_run",
4665 ":operators",
Marat Dukhan86df3842020-06-17 14:03:31 -07004666 ":ukernels_max_compatibility",
Marat Dukhanae046f52020-06-15 13:16:14 -07004667 "@clog",
4668 "@FP16",
4669 "@pthreadpool",
4670 ] + select({
4671 ":emscripten": [],
4672 "//conditions:default": ["@cpuinfo"],
4673 }),
4674)
4675
4676xnnpack_cc_library(
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004677 name = "xnnpack_operators_nhwc_f32",
4678 srcs = [
4679 "src/init.c",
4680 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07004681 hdrs = ["include/xnnpack.h"],
4682 copts = LOGGING_COPTS + [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004683 "-Isrc",
4684 "-Iinclude",
4685 ] + select({
4686 ":debug_build": [],
4687 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07004688 }) + select({
4689 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4690 "//conditions:default": [],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004691 }),
4692 defines = [
Marat Dukhan16f1e1a2020-08-04 16:38:22 -07004693 "XNN_NO_QS8_OPERATORS",
Marat Dukhan08b7a972020-07-14 18:17:29 -07004694 "XNN_NO_QU8_OPERATORS",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004695 "XNN_NO_U8_OPERATORS",
4696 "XNN_NO_X8_OPERATORS",
Marat Dukhanefc47b82019-11-18 09:25:38 -08004697 "XNN_NO_NCHW_OPERATORS",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004698 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07004699 gcc_copts = xnnpack_gcc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004700 includes = ["include"],
Marat Dukhan10a38082020-04-17 03:58:35 -07004701 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004702 visibility = xnnpack_visibility(),
4703 deps = [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004704 ":enable_assembly",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004705 ":logging_utils",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004706 ":operator_run",
4707 ":operators",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004708 ":ukernels",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004709 "@clog",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004710 "@pthreadpool",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004711 ] + select({
4712 ":emscripten": [],
4713 "//conditions:default": ["@cpuinfo"],
4714 }),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004715)
4716
Marat Dukhancf056b22019-10-07 10:26:29 -07004717xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004718 name = "bench_utils",
4719 srcs = ["bench/utils.cc"],
4720 hdrs = ["bench/utils.h"],
Marat Dukhanbad48fe2019-11-04 10:35:22 -08004721 deps = [
4722 "@com_google_benchmark//:benchmark",
4723 "@cpuinfo",
4724 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004725)
4726
Frank Barchard7e955972019-10-11 10:34:25 -07004727######################### Benchmarks for micro-kernels #########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07004728
4729xnnpack_benchmark(
Marat Dukhan595e1702020-07-31 10:12:52 -07004730 name = "qs8_gemm_bench",
4731 srcs = [
4732 "bench/gemm.h",
4733 "bench/qs8-gemm.cc",
4734 "src/xnnpack/AlignedAllocator.h",
4735 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Frank Barchard31328cb2020-10-12 11:55:18 -07004736 copts = xnnpack_optional_ruy_copts() + xnnpack_optional_gemmlowp_copts(),
4737 deps = MICROKERNEL_BENCHMARK_DEPS + [":packing"] + xnnpack_optional_ruy_deps() + xnnpack_optional_gemmlowp_deps(),
Marat Dukhan595e1702020-07-31 10:12:52 -07004738)
4739
4740xnnpack_benchmark(
Marat Dukhan56bdd4a2020-08-03 19:47:04 -07004741 name = "qs8_requantization_bench",
4742 srcs = [
4743 "bench/qs8-requantization.cc",
Marat Dukhan56bdd4a2020-08-03 19:47:04 -07004744 "src/xnnpack/AlignedAllocator.h",
Frank Barchard04336c12020-10-22 16:48:55 -07004745 "src/xnnpack/requantization-stubs.h",
Marat Dukhan56bdd4a2020-08-03 19:47:04 -07004746 ] + MICROKERNEL_BENCHMARK_HDRS,
4747 deps = MICROKERNEL_BENCHMARK_DEPS,
4748)
4749
4750xnnpack_benchmark(
Marat Dukhan08b7a972020-07-14 18:17:29 -07004751 name = "qu8_gemm_bench",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004752 srcs = [
4753 "bench/gemm.h",
Marat Dukhan08b7a972020-07-14 18:17:29 -07004754 "bench/qu8-gemm.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004755 "src/xnnpack/AlignedAllocator.h",
4756 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004757 copts = xnnpack_optional_ruy_copts() + xnnpack_optional_gemmlowp_copts(),
Marat Dukhanab582382020-07-06 13:32:08 -07004758 deps = MICROKERNEL_BENCHMARK_DEPS + [":packing"] + xnnpack_optional_ruy_deps() + xnnpack_optional_gemmlowp_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004759)
4760
4761xnnpack_benchmark(
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07004762 name = "qu8_requantization_bench",
4763 srcs = [
4764 "bench/qu8-requantization.cc",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07004765 "src/xnnpack/AlignedAllocator.h",
Frank Barchard04336c12020-10-22 16:48:55 -07004766 "src/xnnpack/requantization-stubs.h",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07004767 ] + MICROKERNEL_BENCHMARK_HDRS,
4768 deps = MICROKERNEL_BENCHMARK_DEPS,
4769)
4770
4771xnnpack_benchmark(
Frank Barchard40d20fe2020-05-05 00:37:45 -07004772 name = "f16_igemm_bench",
4773 srcs = [
4774 "bench/f16-igemm.cc",
4775 "bench/conv.h",
4776 "bench/google/conv.h",
4777 "src/xnnpack/AlignedAllocator.h",
4778 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004779 deps = MICROKERNEL_BENCHMARK_DEPS + [
4780 ":indirection",
4781 ":packing",
4782 ],
Frank Barchard40d20fe2020-05-05 00:37:45 -07004783)
4784
4785xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004786 name = "f16_gemm_bench",
4787 srcs = [
4788 "bench/f16-gemm.cc",
4789 "bench/gemm.h",
4790 "src/xnnpack/AlignedAllocator.h",
4791 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004792 deps = MICROKERNEL_BENCHMARK_DEPS + [
4793 ":packing",
4794 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004795)
4796
4797xnnpack_benchmark(
Marat Dukhanbdb56f52020-02-05 21:42:49 -08004798 name = "f16_spmm_bench",
4799 srcs = [
4800 "bench/f16-spmm.cc",
Marat Dukhan1631e3e2020-12-06 19:29:31 -08004801 "bench/spmm.h",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08004802 "src/xnnpack/AlignedAllocator.h",
4803 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanbdb56f52020-02-05 21:42:49 -08004804 deps = MICROKERNEL_BENCHMARK_DEPS,
4805)
4806
4807xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004808 name = "f32_igemm_bench",
4809 srcs = [
4810 "bench/f32-igemm.cc",
4811 "bench/conv.h",
4812 "src/xnnpack/AlignedAllocator.h",
4813 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004814 deps = MICROKERNEL_BENCHMARK_DEPS + [
4815 ":indirection",
4816 ":packing",
4817 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004818)
4819
4820xnnpack_benchmark(
Frank Barchardc2010062020-07-19 10:04:12 -07004821 name = "f16_relu_bench",
4822 srcs = [
4823 "bench/f16-relu.cc",
4824 "src/xnnpack/AlignedAllocator.h",
4825 ] + MICROKERNEL_BENCHMARK_HDRS,
4826 deps = MICROKERNEL_BENCHMARK_DEPS,
4827)
4828
4829xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004830 name = "f32_conv_hwc_bench",
4831 srcs = [
4832 "bench/f32-conv-hwc.cc",
4833 "bench/dconv.h",
4834 "src/xnnpack/AlignedAllocator.h",
4835 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004836 deps = MICROKERNEL_BENCHMARK_DEPS + [
4837 ":packing",
4838 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004839)
4840
4841xnnpack_benchmark(
Marat Dukhan1f29b802020-05-15 23:46:39 -07004842 name = "f32_conv_hwc2chw_bench",
Erich Elsen563df5f2019-10-23 08:02:21 -07004843 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07004844 "bench/f32-conv-hwc2chw.cc",
Erich Elsen563df5f2019-10-23 08:02:21 -07004845 "bench/dconv.h",
4846 "src/xnnpack/AlignedAllocator.h",
4847 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004848 deps = MICROKERNEL_BENCHMARK_DEPS + [
4849 ":packing",
4850 ],
Erich Elsen563df5f2019-10-23 08:02:21 -07004851)
4852
4853xnnpack_benchmark(
Frank Barchard5a599a62020-06-04 20:12:44 -07004854 name = "f16_dwconv_bench",
4855 srcs = [
4856 "bench/f16-dwconv.cc",
4857 "bench/dwconv.h",
4858 "bench/google/dwconv.h",
4859 "src/xnnpack/AlignedAllocator.h",
4860 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004861 deps = MICROKERNEL_BENCHMARK_DEPS + [
4862 ":indirection",
4863 ":packing",
4864 ],
Frank Barchard5a599a62020-06-04 20:12:44 -07004865)
4866
4867xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004868 name = "f32_dwconv_bench",
4869 srcs = [
4870 "bench/f32-dwconv.cc",
4871 "bench/dwconv.h",
4872 "src/xnnpack/AlignedAllocator.h",
4873 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004874 deps = MICROKERNEL_BENCHMARK_DEPS + [
4875 ":indirection",
4876 ":packing",
4877 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004878)
4879
4880xnnpack_benchmark(
Marat Dukhanbf715f92020-10-23 20:17:00 -07004881 name = "f32_dwconv2d_chw_bench",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004882 srcs = [
Marat Dukhanbf715f92020-10-23 20:17:00 -07004883 "bench/f32-dwconv2d-chw.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004884 "bench/dwconv.h",
4885 "src/xnnpack/AlignedAllocator.h",
4886 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004887 deps = MICROKERNEL_BENCHMARK_DEPS + [
4888 ":indirection",
4889 ":packing",
4890 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004891)
4892
4893xnnpack_benchmark(
4894 name = "f32_gemm_bench",
4895 srcs = [
4896 "bench/f32-gemm.cc",
4897 "bench/gemm.h",
4898 "src/xnnpack/AlignedAllocator.h",
4899 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004900 copts = xnnpack_optional_ruy_copts(),
Marat Dukhanab582382020-07-06 13:32:08 -07004901 deps = MICROKERNEL_BENCHMARK_DEPS + [":packing"] + xnnpack_optional_ruy_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004902)
4903
4904xnnpack_benchmark(
Marat Dukhana11ca342020-06-25 23:45:07 -07004905 name = "f32_hswish_bench",
4906 srcs = [
4907 "bench/f32-hswish.cc",
4908 "src/xnnpack/AlignedAllocator.h",
4909 ] + MICROKERNEL_BENCHMARK_HDRS,
4910 deps = MICROKERNEL_BENCHMARK_DEPS,
4911)
4912
4913xnnpack_benchmark(
Marat Dukhan4c4eb002019-12-08 21:27:49 -08004914 name = "f32_raddexpminusmax_bench",
4915 srcs = [
4916 "bench/f32-raddexpminusmax.cc",
4917 "src/xnnpack/AlignedAllocator.h",
4918 ] + MICROKERNEL_BENCHMARK_HDRS,
4919 deps = MICROKERNEL_BENCHMARK_DEPS,
4920)
4921
4922xnnpack_benchmark(
4923 name = "f32_raddextexp_bench",
4924 srcs = [
4925 "bench/f32-raddextexp.cc",
4926 "src/xnnpack/AlignedAllocator.h",
4927 ] + MICROKERNEL_BENCHMARK_HDRS,
4928 deps = MICROKERNEL_BENCHMARK_DEPS,
4929)
4930
4931xnnpack_benchmark(
4932 name = "f32_raddstoreexpminusmax_bench",
4933 srcs = [
4934 "bench/f32-raddstoreexpminusmax.cc",
4935 "src/xnnpack/AlignedAllocator.h",
4936 ] + MICROKERNEL_BENCHMARK_HDRS,
4937 deps = MICROKERNEL_BENCHMARK_DEPS,
4938)
4939
4940xnnpack_benchmark(
Frank Barchard3c4a9522020-07-18 16:11:22 -07004941 name = "f32_relu_bench",
4942 srcs = [
4943 "bench/f32-relu.cc",
4944 "src/xnnpack/AlignedAllocator.h",
4945 ] + MICROKERNEL_BENCHMARK_HDRS,
4946 deps = MICROKERNEL_BENCHMARK_DEPS,
4947)
4948
4949xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004950 name = "f32_rmax_bench",
4951 srcs = [
4952 "bench/f32-rmax.cc",
4953 "src/xnnpack/AlignedAllocator.h",
4954 ] + MICROKERNEL_BENCHMARK_HDRS,
4955 deps = MICROKERNEL_BENCHMARK_DEPS,
4956)
4957
4958xnnpack_benchmark(
Marat Dukhan14bec502019-11-18 11:35:31 -08004959 name = "f32_sigmoid_bench",
4960 srcs = [
4961 "bench/f32-sigmoid.cc",
4962 "src/xnnpack/AlignedAllocator.h",
4963 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan14bec502019-11-18 11:35:31 -08004964 deps = MICROKERNEL_BENCHMARK_DEPS,
4965)
4966
4967xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004968 name = "f32_spmm_bench",
4969 srcs = [
4970 "bench/f32-spmm.cc",
Marat Dukhan1631e3e2020-12-06 19:29:31 -08004971 "bench/spmm.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004972 "src/xnnpack/AlignedAllocator.h",
4973 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07004974 deps = MICROKERNEL_BENCHMARK_DEPS,
4975)
4976
4977xnnpack_benchmark(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08004978 name = "f32_softmax_bench",
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07004979 srcs = [
Marat Dukhanfd8e6892020-01-27 15:25:25 -08004980 "bench/f32-softmax.cc",
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07004981 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004982 copts = xnnpack_optional_dnnl_copts(),
Marat Dukhan8d3c6932020-03-06 20:27:27 -08004983 deps = MICROKERNEL_BENCHMARK_DEPS + xnnpack_optional_dnnl_deps(),
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07004984)
4985
4986xnnpack_benchmark(
Marat Dukhaned6baaf2020-12-01 15:07:08 -08004987 name = "f32_velu_bench",
4988 srcs = [
4989 "bench/f32-velu.cc",
4990 "src/xnnpack/AlignedAllocator.h",
4991 ] + MICROKERNEL_BENCHMARK_HDRS,
4992 deps = MICROKERNEL_BENCHMARK_DEPS,
4993)
4994
4995xnnpack_benchmark(
Marat Dukhan4c4eb002019-12-08 21:27:49 -08004996 name = "f32_vscaleexpminusmax_bench",
4997 srcs = [
4998 "bench/f32-vscaleexpminusmax.cc",
4999 "src/xnnpack/AlignedAllocator.h",
5000 ] + MICROKERNEL_BENCHMARK_HDRS,
5001 deps = MICROKERNEL_BENCHMARK_DEPS,
5002)
5003
5004xnnpack_benchmark(
5005 name = "f32_vscaleextexp_bench",
5006 srcs = [
5007 "bench/f32-vscaleextexp.cc",
5008 "src/xnnpack/AlignedAllocator.h",
5009 ] + MICROKERNEL_BENCHMARK_HDRS,
5010 deps = MICROKERNEL_BENCHMARK_DEPS,
5011)
5012
5013xnnpack_benchmark(
Marat Dukhanf4db2f32020-06-30 10:55:30 -07005014 name = "f32_vsqrt_bench",
5015 srcs = [
5016 "bench/f32-vsqrt.cc",
5017 "src/xnnpack/AlignedAllocator.h",
5018 ] + MICROKERNEL_BENCHMARK_HDRS,
5019 deps = MICROKERNEL_BENCHMARK_DEPS,
5020)
5021
5022xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005023 name = "f32_im2col_gemm_bench",
5024 srcs = [
5025 "bench/f32-im2col-gemm.cc",
5026 "bench/conv.h",
5027 "src/xnnpack/AlignedAllocator.h",
5028 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005029 deps = MICROKERNEL_BENCHMARK_DEPS + [
5030 ":im2col",
5031 ":packing",
5032 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005033)
5034
Marat Dukhanfe7acb62020-03-09 19:30:05 -07005035xnnpack_benchmark(
Marat Dukhanffbf96a2020-05-14 02:59:08 -07005036 name = "rounding_bench",
5037 srcs = [
5038 "bench/rounding.cc",
Marat Dukhanffbf96a2020-05-14 02:59:08 -07005039 "src/xnnpack/AlignedAllocator.h",
Frank Barchard04336c12020-10-22 16:48:55 -07005040 "src/xnnpack/math-stubs.h",
Marat Dukhanffbf96a2020-05-14 02:59:08 -07005041 ] + MICROKERNEL_BENCHMARK_HDRS,
5042 deps = MICROKERNEL_BENCHMARK_DEPS,
5043)
5044
Marat Dukhan08c4a432019-10-03 09:29:21 -07005045########################### Benchmarks for operators ###########################
5046
5047xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005048 name = "average_pooling_bench",
5049 srcs = ["bench/average-pooling.cc"],
Marat Dukhan7a16d8b2020-03-11 04:22:44 -07005050 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07005051 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07005052 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07005053)
5054
5055xnnpack_benchmark(
Marat Dukhan87727142020-06-24 15:24:10 -07005056 name = "bankers_rounding_bench",
5057 srcs = ["bench/bankers-rounding.cc"],
5058 copts = xnnpack_optional_tflite_copts(),
5059 tags = ["nowin32"],
5060 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
5061)
5062
5063xnnpack_benchmark(
5064 name = "ceiling_bench",
5065 srcs = ["bench/ceiling.cc"],
5066 copts = xnnpack_optional_tflite_copts(),
5067 tags = ["nowin32"],
5068 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
5069)
5070
5071xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005072 name = "channel_shuffle_bench",
5073 srcs = ["bench/channel-shuffle.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07005074 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005075)
5076
5077xnnpack_benchmark(
5078 name = "convolution_bench",
5079 srcs = ["bench/convolution.cc"],
5080 copts = xnnpack_optional_tflite_copts() + xnnpack_optional_armcl_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07005081 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07005082 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps() + xnnpack_optional_armcl_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07005083)
5084
5085xnnpack_benchmark(
5086 name = "deconvolution_bench",
5087 srcs = ["bench/deconvolution.cc"],
5088 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07005089 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07005090 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07005091)
5092
5093xnnpack_benchmark(
Marat Dukhanb6bd4bc2020-12-01 17:01:40 -08005094 name = "elu_bench",
5095 srcs = ["bench/elu.cc"],
5096 copts = xnnpack_optional_tflite_copts(),
5097 tags = ["nowin32"],
5098 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
5099)
5100
5101xnnpack_benchmark(
Marat Dukhan87727142020-06-24 15:24:10 -07005102 name = "floor_bench",
5103 srcs = ["bench/floor.cc"],
5104 copts = xnnpack_optional_tflite_copts(),
5105 tags = ["nowin32"],
5106 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
5107)
5108
5109xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005110 name = "global_average_pooling_bench",
5111 srcs = ["bench/global-average-pooling.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07005112 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005113)
5114
5115xnnpack_benchmark(
Marat Dukhanad352602020-06-25 21:50:54 -07005116 name = "hardswish_bench",
5117 srcs = ["bench/hardswish.cc"],
5118 copts = xnnpack_optional_tflite_copts(),
5119 tags = ["nowin32"],
5120 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
5121)
5122
5123xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005124 name = "max_pooling_bench",
5125 srcs = ["bench/max-pooling.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07005126 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005127)
5128
5129xnnpack_benchmark(
5130 name = "sigmoid_bench",
5131 srcs = ["bench/sigmoid.cc"],
Marat Dukhanc3b9e862019-11-17 13:18:54 -08005132 copts = xnnpack_optional_tflite_copts(),
Marat Dukhanca2ba702020-04-24 01:31:47 -07005133 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07005134 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07005135)
5136
5137xnnpack_benchmark(
Marat Dukhan95b22432019-10-30 16:30:14 -07005138 name = "prelu_bench",
5139 srcs = ["bench/prelu.cc"],
5140 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07005141 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07005142 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan95b22432019-10-30 16:30:14 -07005143)
5144
5145xnnpack_benchmark(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08005146 name = "softmax_bench",
5147 srcs = ["bench/softmax.cc"],
Marat Dukhan9c0db962020-01-28 12:30:14 -08005148 copts = xnnpack_optional_tflite_copts(),
Marat Dukhanca2ba702020-04-24 01:31:47 -07005149 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07005150 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07005151)
5152
Marat Dukhan87727142020-06-24 15:24:10 -07005153xnnpack_benchmark(
Marat Dukhan6804bbd2020-06-30 19:26:11 -07005154 name = "square_root_bench",
5155 srcs = ["bench/square-root.cc"],
5156 copts = xnnpack_optional_tflite_copts(),
5157 tags = ["nowin32"],
5158 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
5159)
5160
5161xnnpack_benchmark(
Marat Dukhan87727142020-06-24 15:24:10 -07005162 name = "truncation_bench",
5163 srcs = ["bench/truncation.cc"],
5164 deps = OPERATOR_BENCHMARK_DEPS,
5165)
5166
Marat Dukhanc068bb62019-10-04 13:24:39 -07005167############################# End-to-end benchmarks ############################
5168
5169cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07005170 name = "fp32_mobilenet_v1",
5171 srcs = ["models/fp32-mobilenet-v1.cc"],
Marat Dukhanc068bb62019-10-04 13:24:39 -07005172 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08005173 copts = xnnpack_std_cxxopts(),
Marat Dukhanc068bb62019-10-04 13:24:39 -07005174 linkstatic = True,
5175 deps = [
5176 ":XNNPACK",
5177 "@pthreadpool",
5178 ],
5179)
5180
5181cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07005182 name = "fp16_mobilenet_v1",
5183 srcs = ["models/fp16-mobilenet-v1.cc"],
5184 hdrs = ["models/models.h"],
5185 copts = xnnpack_std_cxxopts(),
5186 linkstatic = True,
5187 deps = [
5188 ":XNNPACK",
5189 "@FP16",
5190 "@pthreadpool",
5191 ],
5192)
5193
5194cc_library(
Marat Dukhan0743cdf2020-08-04 18:52:07 -07005195 name = "qs8_mobilenet_v1",
5196 srcs = ["models/qs8-mobilenet-v1.cc"],
5197 hdrs = ["models/models.h"],
5198 copts = xnnpack_std_cxxopts(),
5199 linkstatic = True,
5200 deps = [
5201 ":XNNPACK",
5202 "@pthreadpool",
5203 ],
5204)
5205
5206cc_library(
Marat Dukhan70a96182020-09-03 17:13:58 -07005207 name = "qs8_mobilenet_v2",
5208 srcs = ["models/qs8-mobilenet-v2.cc"],
5209 hdrs = ["models/models.h"],
5210 copts = xnnpack_std_cxxopts(),
5211 linkstatic = True,
5212 deps = [
5213 ":XNNPACK",
5214 "@pthreadpool",
5215 ],
5216)
5217
5218cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07005219 name = "fp32_mobilenet_v2",
5220 srcs = ["models/fp32-mobilenet-v2.cc"],
Marat Dukhanc068bb62019-10-04 13:24:39 -07005221 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08005222 copts = xnnpack_std_cxxopts(),
Marat Dukhanc068bb62019-10-04 13:24:39 -07005223 linkstatic = True,
5224 deps = [
5225 ":XNNPACK",
5226 "@pthreadpool",
5227 ],
5228)
5229
Marat Dukhanc08cdf52019-12-09 09:17:51 -08005230cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07005231 name = "fp16_mobilenet_v2",
5232 srcs = ["models/fp16-mobilenet-v2.cc"],
5233 hdrs = ["models/models.h"],
5234 copts = xnnpack_std_cxxopts(),
5235 linkstatic = True,
5236 deps = [
5237 ":XNNPACK",
5238 "@FP16",
5239 "@pthreadpool",
5240 ],
5241)
5242
5243cc_library(
5244 name = "fp32_mobilenet_v3_large",
5245 srcs = ["models/fp32-mobilenet-v3-large.cc"],
Marat Dukhanc08cdf52019-12-09 09:17:51 -08005246 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08005247 copts = xnnpack_std_cxxopts(),
Marat Dukhanc08cdf52019-12-09 09:17:51 -08005248 linkstatic = True,
5249 deps = [
5250 ":XNNPACK",
5251 "@pthreadpool",
5252 ],
5253)
5254
5255cc_library(
Marat Dukhan66f3ccd2020-09-14 16:09:38 -07005256 name = "fp16_mobilenet_v3_large",
5257 srcs = ["models/fp16-mobilenet-v3-large.cc"],
5258 hdrs = ["models/models.h"],
5259 copts = xnnpack_std_cxxopts(),
5260 linkstatic = True,
5261 deps = [
5262 ":XNNPACK",
5263 "@FP16",
5264 "@pthreadpool",
5265 ],
5266)
5267
5268cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07005269 name = "fp32_mobilenet_v3_small",
5270 srcs = ["models/fp32-mobilenet-v3-small.cc"],
Marat Dukhanc08cdf52019-12-09 09:17:51 -08005271 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08005272 copts = xnnpack_std_cxxopts(),
Marat Dukhanc08cdf52019-12-09 09:17:51 -08005273 linkstatic = True,
5274 deps = [
5275 ":XNNPACK",
5276 "@pthreadpool",
5277 ],
5278)
5279
Marat Dukhan66f3ccd2020-09-14 16:09:38 -07005280cc_library(
5281 name = "fp16_mobilenet_v3_small",
5282 srcs = ["models/fp16-mobilenet-v3-small.cc"],
5283 hdrs = ["models/models.h"],
5284 copts = xnnpack_std_cxxopts(),
5285 linkstatic = True,
5286 deps = [
5287 ":XNNPACK",
5288 "@FP16",
5289 "@pthreadpool",
5290 ],
5291)
5292
Marat Dukhanc068bb62019-10-04 13:24:39 -07005293xnnpack_benchmark(
Marat Dukhanef4416e2019-10-31 13:44:40 -07005294 name = "f32_dwconv_e2e_bench",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08005295 srcs = [
5296 "bench/f32-dwconv-e2e.cc",
5297 "bench/end2end.h",
5298 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanef4416e2019-10-31 13:44:40 -07005299 deps = MICROKERNEL_BENCHMARK_DEPS + [
5300 ":XNNPACK",
Marat Dukhan270a2c42020-06-26 16:45:52 -07005301 ":fp32_mobilenet_v1",
5302 ":fp32_mobilenet_v2",
5303 ":fp32_mobilenet_v3_large",
5304 ":fp32_mobilenet_v3_small",
Marat Dukhanef4416e2019-10-31 13:44:40 -07005305 ],
5306)
5307
5308xnnpack_benchmark(
Marat Dukhan5f18d262019-10-31 10:24:14 -07005309 name = "f32_gemm_e2e_bench",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08005310 srcs = [
5311 "bench/f32-gemm-e2e.cc",
5312 "bench/end2end.h",
5313 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan5f18d262019-10-31 10:24:14 -07005314 deps = MICROKERNEL_BENCHMARK_DEPS + [
5315 ":XNNPACK",
Marat Dukhan270a2c42020-06-26 16:45:52 -07005316 ":fp32_mobilenet_v1",
5317 ":fp32_mobilenet_v2",
5318 ":fp32_mobilenet_v3_large",
5319 ":fp32_mobilenet_v3_small",
Marat Dukhan5f18d262019-10-31 10:24:14 -07005320 ],
5321)
5322
5323xnnpack_benchmark(
Marat Dukhanc068bb62019-10-04 13:24:39 -07005324 name = "end2end_bench",
5325 srcs = ["bench/end2end.cc"],
5326 deps = [
5327 ":XNNPACK",
Frank Barchardc712fa42019-10-31 14:00:21 -07005328 ":bench_utils",
Marat Dukhan270a2c42020-06-26 16:45:52 -07005329 ":fp16_mobilenet_v1",
5330 ":fp16_mobilenet_v2",
Marat Dukhan66f3ccd2020-09-14 16:09:38 -07005331 ":fp16_mobilenet_v3_large",
5332 ":fp16_mobilenet_v3_small",
Marat Dukhan270a2c42020-06-26 16:45:52 -07005333 ":fp32_mobilenet_v1",
5334 ":fp32_mobilenet_v2",
5335 ":fp32_mobilenet_v3_large",
5336 ":fp32_mobilenet_v3_small",
Marat Dukhan0743cdf2020-08-04 18:52:07 -07005337 ":qs8_mobilenet_v1",
Marat Dukhan70a96182020-09-03 17:13:58 -07005338 ":qs8_mobilenet_v2",
Marat Dukhanc068bb62019-10-04 13:24:39 -07005339 "@pthreadpool",
5340 ],
5341)
5342
Marat Dukhan6adff4e2019-10-14 18:32:07 -07005343#################### Accuracy evaluation for math functions ####################
5344
5345xnnpack_benchmark(
5346 name = "f32_exp_eval",
5347 srcs = [
5348 "eval/f32-exp.cc",
5349 "src/xnnpack/AlignedAllocator.h",
5350 ] + ACCURACY_EVAL_HDRS,
Marat Dukhan3a305212020-12-06 19:24:27 -08005351 deps = ACCURACY_EVAL_DEPS + [
5352 ":bench_utils",
5353 "@cpuinfo",
5354 ],
Marat Dukhan6adff4e2019-10-14 18:32:07 -07005355)
5356
Marat Dukhan515c9772019-10-17 18:07:57 -07005357xnnpack_benchmark(
5358 name = "f32_expminus_eval",
5359 srcs = [
5360 "eval/f32-expminus.cc",
5361 "src/xnnpack/AlignedAllocator.h",
5362 ] + ACCURACY_EVAL_HDRS,
Marat Dukhan3a305212020-12-06 19:24:27 -08005363 deps = ACCURACY_EVAL_DEPS + [
5364 ":bench_utils",
5365 "@cpuinfo",
5366 ],
Marat Dukhan515c9772019-10-17 18:07:57 -07005367)
5368
Marat Dukhan98ba4412019-10-23 02:14:28 -07005369xnnpack_benchmark(
Marat Dukhana438aca2020-11-20 15:45:01 -08005370 name = "f32_expm1minus_eval",
5371 srcs = [
5372 "eval/f32-expm1minus.cc",
5373 "src/xnnpack/AlignedAllocator.h",
5374 ] + ACCURACY_EVAL_HDRS,
Marat Dukhande390d42020-11-29 19:32:18 -08005375 deps = ACCURACY_EVAL_DEPS + [
5376 ":bench_utils",
Marat Dukhan3a305212020-12-06 19:24:27 -08005377 "@cpuinfo",
Marat Dukhande390d42020-11-29 19:32:18 -08005378 ],
Marat Dukhana438aca2020-11-20 15:45:01 -08005379)
5380
5381xnnpack_benchmark(
Marat Dukhan98ba4412019-10-23 02:14:28 -07005382 name = "f32_extexp_eval",
5383 srcs = [
5384 "eval/f32-extexp.cc",
5385 "src/xnnpack/AlignedAllocator.h",
5386 ] + ACCURACY_EVAL_HDRS,
Marat Dukhan3a305212020-12-06 19:24:27 -08005387 deps = ACCURACY_EVAL_DEPS + [
5388 ":bench_utils",
5389 "@cpuinfo",
5390 ],
Marat Dukhan98ba4412019-10-23 02:14:28 -07005391)
5392
Marat Dukhan8853b822020-05-07 12:19:01 -07005393xnnpack_unit_test(
5394 name = "f32_roundne_eval",
5395 srcs = [
5396 "eval/f32-roundne.cc",
5397 "src/xnnpack/AlignedAllocator.h",
5398 "src/xnnpack/math-stubs.h",
5399 ] + MICROKERNEL_TEST_HDRS,
Marat Dukhan22eed3d2020-05-11 20:13:37 -07005400 automatic = False,
Marat Dukhan8853b822020-05-07 12:19:01 -07005401 deps = MICROKERNEL_TEST_DEPS,
5402)
5403
Marat Dukhan2dbb9442020-05-12 20:43:43 -07005404xnnpack_unit_test(
Marat Dukhanc9852ba2020-05-13 17:21:29 -07005405 name = "f32_roundd_eval",
5406 srcs = [
5407 "eval/f32-roundd.cc",
5408 "src/xnnpack/AlignedAllocator.h",
5409 "src/xnnpack/math-stubs.h",
5410 ] + MICROKERNEL_TEST_HDRS,
5411 automatic = False,
5412 deps = MICROKERNEL_TEST_DEPS,
5413)
5414
5415xnnpack_unit_test(
5416 name = "f32_roundu_eval",
5417 srcs = [
5418 "eval/f32-roundu.cc",
5419 "src/xnnpack/AlignedAllocator.h",
5420 "src/xnnpack/math-stubs.h",
5421 ] + MICROKERNEL_TEST_HDRS,
5422 automatic = False,
5423 deps = MICROKERNEL_TEST_DEPS,
5424)
5425
5426xnnpack_unit_test(
Marat Dukhan2dbb9442020-05-12 20:43:43 -07005427 name = "f32_roundz_eval",
5428 srcs = [
5429 "eval/f32-roundz.cc",
5430 "src/xnnpack/AlignedAllocator.h",
5431 "src/xnnpack/math-stubs.h",
5432 ] + MICROKERNEL_TEST_HDRS,
Marat Dukhanc9852ba2020-05-13 17:21:29 -07005433 automatic = False,
Marat Dukhan2dbb9442020-05-12 20:43:43 -07005434 deps = MICROKERNEL_TEST_DEPS,
5435)
5436
Marat Dukhan346a9e52019-11-15 09:06:30 -08005437xnnpack_benchmark(
5438 name = "f32_sigmoid_eval",
5439 srcs = [
5440 "eval/f32-sigmoid.cc",
5441 "src/xnnpack/AlignedAllocator.h",
5442 ] + ACCURACY_EVAL_HDRS,
Marat Dukhan3a305212020-12-06 19:24:27 -08005443 deps = ACCURACY_EVAL_DEPS + [
5444 ":bench_utils",
5445 "@cpuinfo",
5446 ],
Marat Dukhan346a9e52019-11-15 09:06:30 -08005447)
5448
Marat Dukhan84000762020-06-29 18:38:43 -07005449xnnpack_benchmark(
5450 name = "f32_sqrt_eval",
5451 srcs = [
5452 "eval/f32-sqrt.cc",
5453 "src/xnnpack/AlignedAllocator.h",
5454 ] + ACCURACY_EVAL_HDRS,
5455 deps = ACCURACY_EVAL_DEPS + [
5456 ":bench_utils",
Marat Dukhan3a305212020-12-06 19:24:27 -08005457 "@cpuinfo",
Marat Dukhan84000762020-06-29 18:38:43 -07005458 ],
5459)
5460
Marat Dukhan08c4a432019-10-03 09:29:21 -07005461######################### Unit tests for micro-kernels #########################
5462
5463xnnpack_unit_test(
Marat Dukhande06f492020-04-09 00:19:31 -07005464 name = "f16_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005465 srcs = [
Marat Dukhande06f492020-04-09 00:19:31 -07005466 "test/f16-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005467 "test/gemm-microkernel-tester.h",
5468 "src/xnnpack/AlignedAllocator.h",
5469 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005470 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005471)
5472
5473xnnpack_unit_test(
Marat Dukhan355ab432020-04-09 19:01:52 -07005474 name = "f16_spmm_minmax_test",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08005475 srcs = [
Marat Dukhan355ab432020-04-09 19:01:52 -07005476 "test/f16-spmm-minmax.cc",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08005477 "test/spmm-microkernel-tester.h",
5478 "src/xnnpack/AlignedAllocator.h",
5479 ] + MICROKERNEL_TEST_HDRS,
5480 deps = MICROKERNEL_TEST_DEPS,
5481)
5482
5483xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005484 name = "f32_argmaxpool_test",
5485 srcs = [
5486 "test/f32-argmaxpool.cc",
5487 "test/argmaxpool-microkernel-tester.h",
5488 "src/xnnpack/AlignedAllocator.h",
5489 ] + MICROKERNEL_TEST_HDRS,
5490 deps = MICROKERNEL_TEST_DEPS,
5491)
5492
5493xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07005494 name = "f32_avgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005495 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07005496 "test/f32-avgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005497 "test/avgpool-microkernel-tester.h",
5498 "src/xnnpack/AlignedAllocator.h",
5499 ] + MICROKERNEL_TEST_HDRS,
5500 deps = MICROKERNEL_TEST_DEPS,
5501)
5502
5503xnnpack_unit_test(
Marat Dukhan660fd192020-03-10 04:55:30 -07005504 name = "f32_ibilinear_test",
Marat Dukhan35dacfb2019-11-07 19:18:16 -08005505 srcs = [
Marat Dukhan660fd192020-03-10 04:55:30 -07005506 "test/f32-ibilinear.cc",
5507 "test/ibilinear-microkernel-tester.h",
Marat Dukhan35dacfb2019-11-07 19:18:16 -08005508 "src/xnnpack/AlignedAllocator.h",
5509 ] + MICROKERNEL_TEST_HDRS,
5510 deps = MICROKERNEL_TEST_DEPS,
5511)
5512
5513xnnpack_unit_test(
XNNPACK Team21432672020-10-19 19:58:48 -07005514 name = "f32_ibilinear_chw_test",
5515 srcs = [
5516 "test/f32-ibilinear-chw.cc",
XNNPACK Team6be46b22020-10-22 23:34:54 -07005517 "test/ibilinear-microkernel-tester.h",
XNNPACK Team21432672020-10-19 19:58:48 -07005518 "src/xnnpack/AlignedAllocator.h",
5519 ] + MICROKERNEL_TEST_HDRS,
5520 deps = MICROKERNEL_TEST_DEPS,
5521)
5522
5523xnnpack_unit_test(
Frank Barchardb1966592020-05-12 13:47:06 -07005524 name = "f16_clamp_test",
5525 srcs = [
5526 "test/f16-clamp.cc",
5527 "test/clamp-microkernel-tester.h",
5528 ] + MICROKERNEL_TEST_HDRS,
5529 deps = MICROKERNEL_TEST_DEPS,
5530)
5531
5532xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005533 name = "f32_clamp_test",
5534 srcs = [
5535 "test/f32-clamp.cc",
5536 "test/clamp-microkernel-tester.h",
5537 ] + MICROKERNEL_TEST_HDRS,
5538 deps = MICROKERNEL_TEST_DEPS,
5539)
5540
5541xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07005542 name = "f32_igemm_test",
5543 srcs = [
5544 "test/f32-igemm.cc",
5545 "test/gemm-microkernel-tester.h",
5546 "src/xnnpack/AlignedAllocator.h",
5547 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005548 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan163a7e62020-04-09 04:19:26 -07005549)
5550
5551xnnpack_unit_test(
Marat Dukhan467f6362020-05-22 23:21:55 -07005552 name = "f32_igemm_relu_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005553 srcs = [
Marat Dukhan467f6362020-05-22 23:21:55 -07005554 "test/f32-igemm-relu.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005555 "test/gemm-microkernel-tester.h",
5556 "src/xnnpack/AlignedAllocator.h",
5557 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005558 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005559)
5560
5561xnnpack_unit_test(
Marat Dukhane207b7b2020-05-28 16:27:42 -07005562 name = "f32_igemm_minmax_test",
5563 srcs = [
5564 "test/f32-igemm-minmax.cc",
5565 "test/gemm-microkernel-tester.h",
5566 "src/xnnpack/AlignedAllocator.h",
5567 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005568 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhane207b7b2020-05-28 16:27:42 -07005569)
5570
5571xnnpack_unit_test(
Frank Barchardb0e4fae2020-05-04 15:27:51 -07005572 name = "f16_igemm_minmax_test",
5573 srcs = [
5574 "test/f16-igemm-minmax.cc",
5575 "test/gemm-microkernel-tester.h",
5576 "src/xnnpack/AlignedAllocator.h",
5577 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005578 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Frank Barchardb0e4fae2020-05-04 15:27:51 -07005579)
5580
5581xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005582 name = "f32_conv_hwc_test",
5583 srcs = [
5584 "test/f32-conv-hwc.cc",
5585 "test/conv-hwc-microkernel-tester.h",
5586 "src/xnnpack/AlignedAllocator.h",
5587 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005588 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005589)
5590
5591xnnpack_unit_test(
Marat Dukhan1f29b802020-05-15 23:46:39 -07005592 name = "f32_conv_hwc2chw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005593 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07005594 "test/f32-conv-hwc2chw.cc",
5595 "test/conv-hwc2chw-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005596 "src/xnnpack/AlignedAllocator.h",
5597 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005598 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005599)
5600
5601xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07005602 name = "f32_dwconv_test",
5603 srcs = [
5604 "test/f32-dwconv.cc",
5605 "test/dwconv-microkernel-tester.h",
5606 "src/xnnpack/AlignedAllocator.h",
5607 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005608 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan163a7e62020-04-09 04:19:26 -07005609)
5610
5611xnnpack_unit_test(
Frank Barchard5a599a62020-06-04 20:12:44 -07005612 name = "f16_dwconv_minmax_test",
5613 srcs = [
5614 "test/f16-dwconv-minmax.cc",
5615 "test/dwconv-microkernel-tester.h",
5616 "src/xnnpack/AlignedAllocator.h",
5617 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005618 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Frank Barchard5a599a62020-06-04 20:12:44 -07005619)
5620
5621xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07005622 name = "f32_dwconv_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005623 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07005624 "test/f32-dwconv-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005625 "test/dwconv-microkernel-tester.h",
5626 "src/xnnpack/AlignedAllocator.h",
5627 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005628 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005629)
5630
5631xnnpack_unit_test(
Marat Dukhanbf715f92020-10-23 20:17:00 -07005632 name = "f32_dwconv2d_chw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005633 srcs = [
Marat Dukhanbf715f92020-10-23 20:17:00 -07005634 "test/f32-dwconv2d-chw.cc",
5635 "test/dwconv2d-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005636 "src/xnnpack/AlignedAllocator.h",
5637 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005638 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005639)
5640
5641xnnpack_unit_test(
Frank Barchard0bb49a72020-06-04 11:35:11 -07005642 name = "f16_gavgpool_minmax_test",
5643 srcs = [
5644 "test/f16-gavgpool-minmax.cc",
5645 "test/gavgpool-microkernel-tester.h",
5646 "src/xnnpack/AlignedAllocator.h",
5647 ] + MICROKERNEL_TEST_HDRS,
5648 deps = MICROKERNEL_TEST_DEPS,
5649)
5650
5651xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07005652 name = "f32_gavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005653 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07005654 "test/f32-gavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005655 "test/gavgpool-microkernel-tester.h",
5656 "src/xnnpack/AlignedAllocator.h",
5657 ] + MICROKERNEL_TEST_HDRS,
5658 deps = MICROKERNEL_TEST_DEPS,
5659)
5660
5661xnnpack_unit_test(
Marat Dukhan1f29b802020-05-15 23:46:39 -07005662 name = "f32_gavgpool_cw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005663 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07005664 "test/f32-gavgpool-cw.cc",
5665 "test/gavgpool-cw-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005666 "src/xnnpack/AlignedAllocator.h",
5667 ] + MICROKERNEL_TEST_HDRS,
5668 deps = MICROKERNEL_TEST_DEPS,
5669)
5670
5671xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07005672 name = "f32_gemm_test",
5673 srcs = [
5674 "test/f32-gemm.cc",
5675 "test/gemm-microkernel-tester.h",
5676 "src/xnnpack/AlignedAllocator.h",
5677 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005678 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan163a7e62020-04-09 04:19:26 -07005679)
5680
5681xnnpack_unit_test(
Marat Dukhan467f6362020-05-22 23:21:55 -07005682 name = "f32_gemm_relu_test",
5683 srcs = [
5684 "test/f32-gemm-relu.cc",
5685 "test/gemm-microkernel-tester.h",
5686 "src/xnnpack/AlignedAllocator.h",
5687 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005688 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan467f6362020-05-22 23:21:55 -07005689)
5690
5691xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07005692 name = "f32_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005693 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07005694 "test/f32-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005695 "test/gemm-microkernel-tester.h",
5696 "src/xnnpack/AlignedAllocator.h",
5697 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005698 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005699)
5700
5701xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07005702 name = "f32_gemminc_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005703 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07005704 "test/f32-gemminc-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005705 "test/gemm-microkernel-tester.h",
5706 "src/xnnpack/AlignedAllocator.h",
5707 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005708 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005709)
5710
5711xnnpack_unit_test(
Frank Barchardb1966592020-05-12 13:47:06 -07005712 name = "f16_hswish_test",
5713 srcs = [
5714 "test/f16-hswish.cc",
5715 "test/hswish-microkernel-tester.h",
5716 ] + MICROKERNEL_TEST_HDRS,
5717 deps = MICROKERNEL_TEST_DEPS,
5718)
5719
5720xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005721 name = "f32_hswish_test",
5722 srcs = [
5723 "test/f32-hswish.cc",
5724 "test/hswish-microkernel-tester.h",
5725 ] + MICROKERNEL_TEST_HDRS,
5726 deps = MICROKERNEL_TEST_DEPS,
5727)
5728
5729xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07005730 name = "f32_maxpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005731 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07005732 "test/f32-maxpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005733 "test/maxpool-microkernel-tester.h",
5734 ] + MICROKERNEL_TEST_HDRS,
5735 deps = MICROKERNEL_TEST_DEPS,
5736)
5737
5738xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07005739 name = "f32_pavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005740 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07005741 "test/f32-pavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005742 "test/avgpool-microkernel-tester.h",
5743 "src/xnnpack/AlignedAllocator.h",
5744 ] + MICROKERNEL_TEST_HDRS,
5745 deps = MICROKERNEL_TEST_DEPS,
5746)
5747
5748xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07005749 name = "f32_ppmm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005750 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07005751 "test/f32-ppmm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005752 "test/gemm-microkernel-tester.h",
5753 "src/xnnpack/AlignedAllocator.h",
5754 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005755 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005756)
5757
5758xnnpack_unit_test(
Frank Barchardb1966592020-05-12 13:47:06 -07005759 name = "f16_prelu_test",
5760 srcs = [
5761 "test/f16-prelu.cc",
5762 "test/prelu-microkernel-tester.h",
5763 "src/xnnpack/AlignedAllocator.h",
5764 ] + MICROKERNEL_TEST_HDRS,
5765 deps = MICROKERNEL_TEST_DEPS,
5766)
5767
5768xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005769 name = "f32_prelu_test",
5770 srcs = [
5771 "test/f32-prelu.cc",
5772 "test/prelu-microkernel-tester.h",
5773 "src/xnnpack/AlignedAllocator.h",
5774 ] + MICROKERNEL_TEST_HDRS,
5775 deps = MICROKERNEL_TEST_DEPS,
5776)
5777
5778xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07005779 name = "f32_raddexpminusmax_test",
5780 srcs = [
5781 "test/f32-raddexpminusmax.cc",
5782 "test/raddexpminusmax-microkernel-tester.h",
5783 ] + MICROKERNEL_TEST_HDRS,
5784 deps = MICROKERNEL_TEST_DEPS,
5785)
5786
5787xnnpack_unit_test(
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07005788 name = "f32_raddextexp_test",
5789 srcs = [
5790 "test/f32-raddextexp.cc",
5791 "test/raddextexp-microkernel-tester.h",
5792 ] + MICROKERNEL_TEST_HDRS,
5793 deps = MICROKERNEL_TEST_DEPS,
5794)
5795
5796xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07005797 name = "f32_raddstoreexpminusmax_test",
5798 srcs = [
5799 "test/f32-raddstoreexpminusmax.cc",
5800 "test/raddstoreexpminusmax-microkernel-tester.h",
5801 ] + MICROKERNEL_TEST_HDRS,
5802 deps = MICROKERNEL_TEST_DEPS,
5803)
5804
5805xnnpack_unit_test(
Frank Barchardfb158e22020-07-15 16:10:10 -07005806 name = "f32_relu_test",
5807 srcs = [
5808 "test/f32-relu.cc",
5809 "test/vunary-microkernel-tester.h",
5810 ] + MICROKERNEL_TEST_HDRS,
5811 deps = MICROKERNEL_TEST_DEPS,
5812)
5813
5814xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005815 name = "f32_rmax_test",
5816 srcs = [
5817 "test/f32-rmax.cc",
5818 "test/rmax-microkernel-tester.h",
5819 ] + MICROKERNEL_TEST_HDRS,
5820 deps = MICROKERNEL_TEST_DEPS,
5821)
5822
5823xnnpack_unit_test(
Marat Dukhan346a9e52019-11-15 09:06:30 -08005824 name = "f32_sigmoid_test",
5825 srcs = [
5826 "test/f32-sigmoid.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005827 "test/vunary-microkernel-tester.h",
Marat Dukhan346a9e52019-11-15 09:06:30 -08005828 ] + MICROKERNEL_TEST_HDRS,
5829 deps = MICROKERNEL_TEST_DEPS,
5830)
5831
5832xnnpack_unit_test(
Marat Dukhan355ab432020-04-09 19:01:52 -07005833 name = "f32_spmm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005834 srcs = [
Marat Dukhan355ab432020-04-09 19:01:52 -07005835 "test/f32-spmm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005836 "test/spmm-microkernel-tester.h",
5837 "src/xnnpack/AlignedAllocator.h",
5838 ] + MICROKERNEL_TEST_HDRS,
5839 deps = MICROKERNEL_TEST_DEPS,
5840)
5841
5842xnnpack_unit_test(
Frank Barchardd793f6c2020-05-08 13:37:43 -07005843 name = "f16_vadd_minmax_test",
5844 srcs = [
5845 "test/f16-vadd-minmax.cc",
5846 "test/vbinary-microkernel-tester.h",
5847 ] + MICROKERNEL_TEST_HDRS,
5848 deps = MICROKERNEL_TEST_DEPS,
5849)
5850
5851xnnpack_unit_test(
5852 name = "f16_vaddc_minmax_test",
5853 srcs = [
5854 "test/f16-vaddc-minmax.cc",
5855 "test/vbinaryc-microkernel-tester.h",
5856 ] + MICROKERNEL_TEST_HDRS,
5857 deps = MICROKERNEL_TEST_DEPS,
5858)
5859
5860xnnpack_unit_test(
5861 name = "f16_vdiv_minmax_test",
5862 srcs = [
5863 "test/f16-vdiv-minmax.cc",
5864 "test/vbinary-microkernel-tester.h",
5865 ] + MICROKERNEL_TEST_HDRS,
5866 deps = MICROKERNEL_TEST_DEPS,
5867)
5868
5869xnnpack_unit_test(
5870 name = "f16_vdivc_minmax_test",
5871 srcs = [
5872 "test/f16-vdivc-minmax.cc",
5873 "test/vbinaryc-microkernel-tester.h",
5874 ] + MICROKERNEL_TEST_HDRS,
5875 deps = MICROKERNEL_TEST_DEPS,
5876)
5877
5878xnnpack_unit_test(
5879 name = "f16_vrdivc_minmax_test",
5880 srcs = [
5881 "test/f16-vrdivc-minmax.cc",
5882 "test/vbinaryc-microkernel-tester.h",
5883 ] + MICROKERNEL_TEST_HDRS,
5884 deps = MICROKERNEL_TEST_DEPS,
5885)
5886
5887xnnpack_unit_test(
5888 name = "f16_vmax_test",
5889 srcs = [
5890 "test/f16-vmax.cc",
5891 "test/vbinary-microkernel-tester.h",
5892 ] + MICROKERNEL_TEST_HDRS,
5893 deps = MICROKERNEL_TEST_DEPS,
5894)
5895
5896xnnpack_unit_test(
5897 name = "f16_vmaxc_test",
5898 srcs = [
5899 "test/f16-vmaxc.cc",
5900 "test/vbinaryc-microkernel-tester.h",
5901 ] + MICROKERNEL_TEST_HDRS,
5902 deps = MICROKERNEL_TEST_DEPS,
5903)
5904
5905xnnpack_unit_test(
5906 name = "f16_vmin_test",
5907 srcs = [
5908 "test/f16-vmin.cc",
5909 "test/vbinary-microkernel-tester.h",
5910 ] + MICROKERNEL_TEST_HDRS,
5911 deps = MICROKERNEL_TEST_DEPS,
5912)
5913
5914xnnpack_unit_test(
5915 name = "f16_vminc_test",
5916 srcs = [
5917 "test/f16-vminc.cc",
5918 "test/vbinaryc-microkernel-tester.h",
5919 ] + MICROKERNEL_TEST_HDRS,
5920 deps = MICROKERNEL_TEST_DEPS,
5921)
5922
5923xnnpack_unit_test(
5924 name = "f16_vmul_minmax_test",
5925 srcs = [
5926 "test/f16-vmul-minmax.cc",
5927 "test/vbinary-microkernel-tester.h",
5928 ] + MICROKERNEL_TEST_HDRS,
5929 deps = MICROKERNEL_TEST_DEPS,
5930)
5931
5932xnnpack_unit_test(
5933 name = "f16_vmulc_minmax_test",
5934 srcs = [
5935 "test/f16-vmulc-minmax.cc",
5936 "test/vbinaryc-microkernel-tester.h",
5937 ] + MICROKERNEL_TEST_HDRS,
5938 deps = MICROKERNEL_TEST_DEPS,
5939)
5940
5941xnnpack_unit_test(
5942 name = "f16_vsub_minmax_test",
5943 srcs = [
5944 "test/f16-vsub-minmax.cc",
5945 "test/vbinary-microkernel-tester.h",
5946 ] + MICROKERNEL_TEST_HDRS,
5947 deps = MICROKERNEL_TEST_DEPS,
5948)
5949
5950xnnpack_unit_test(
5951 name = "f16_vsubc_minmax_test",
5952 srcs = [
5953 "test/f16-vsubc-minmax.cc",
5954 "test/vbinaryc-microkernel-tester.h",
5955 ] + MICROKERNEL_TEST_HDRS,
5956 deps = MICROKERNEL_TEST_DEPS,
5957)
5958
5959xnnpack_unit_test(
5960 name = "f16_vrsubc_minmax_test",
5961 srcs = [
5962 "test/f16-vrsubc-minmax.cc",
5963 "test/vbinaryc-microkernel-tester.h",
5964 ] + MICROKERNEL_TEST_HDRS,
5965 deps = MICROKERNEL_TEST_DEPS,
5966)
5967
5968xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07005969 name = "f32_vabs_test",
5970 srcs = [
5971 "test/f32-vabs.cc",
5972 "test/vunary-microkernel-tester.h",
5973 ] + MICROKERNEL_TEST_HDRS,
5974 deps = MICROKERNEL_TEST_DEPS,
5975)
5976
5977xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005978 name = "f32_vadd_test",
5979 srcs = [
5980 "test/f32-vadd.cc",
5981 "test/vbinary-microkernel-tester.h",
5982 ] + MICROKERNEL_TEST_HDRS,
5983 deps = MICROKERNEL_TEST_DEPS,
5984)
5985
5986xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005987 name = "f32_vadd_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005988 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005989 "test/f32-vadd-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005990 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005991 ] + MICROKERNEL_TEST_HDRS,
5992 deps = MICROKERNEL_TEST_DEPS,
5993)
5994
5995xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005996 name = "f32_vadd_relu_test",
5997 srcs = [
5998 "test/f32-vadd-relu.cc",
5999 "test/vbinary-microkernel-tester.h",
6000 ] + MICROKERNEL_TEST_HDRS,
6001 deps = MICROKERNEL_TEST_DEPS,
6002)
6003
6004xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07006005 name = "f32_vaddc_test",
6006 srcs = [
6007 "test/f32-vaddc.cc",
6008 "test/vbinaryc-microkernel-tester.h",
6009 ] + MICROKERNEL_TEST_HDRS,
6010 deps = MICROKERNEL_TEST_DEPS,
6011)
6012
6013xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006014 name = "f32_vaddc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08006015 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006016 "test/f32-vaddc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08006017 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006018 ] + MICROKERNEL_TEST_HDRS,
6019 deps = MICROKERNEL_TEST_DEPS,
6020)
6021
6022xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07006023 name = "f32_vaddc_relu_test",
6024 srcs = [
6025 "test/f32-vaddc-relu.cc",
6026 "test/vbinaryc-microkernel-tester.h",
6027 ] + MICROKERNEL_TEST_HDRS,
6028 deps = MICROKERNEL_TEST_DEPS,
6029)
6030
6031xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07006032 name = "f32_vdiv_test",
6033 srcs = [
6034 "test/f32-vdiv.cc",
6035 "test/vbinary-microkernel-tester.h",
6036 ] + MICROKERNEL_TEST_HDRS,
6037 deps = MICROKERNEL_TEST_DEPS,
6038)
6039
6040xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006041 name = "f32_vdiv_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08006042 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006043 "test/f32-vdiv-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08006044 "test/vbinary-microkernel-tester.h",
6045 ] + MICROKERNEL_TEST_HDRS,
6046 deps = MICROKERNEL_TEST_DEPS,
6047)
6048
6049xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07006050 name = "f32_vdiv_relu_test",
6051 srcs = [
6052 "test/f32-vdiv-relu.cc",
6053 "test/vbinary-microkernel-tester.h",
6054 ] + MICROKERNEL_TEST_HDRS,
6055 deps = MICROKERNEL_TEST_DEPS,
6056)
6057
6058xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07006059 name = "f32_vdivc_test",
6060 srcs = [
6061 "test/f32-vdivc.cc",
6062 "test/vbinaryc-microkernel-tester.h",
6063 ] + MICROKERNEL_TEST_HDRS,
6064 deps = MICROKERNEL_TEST_DEPS,
6065)
6066
6067xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006068 name = "f32_vdivc_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08006069 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006070 "test/f32-vdivc-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08006071 "test/vbinaryc-microkernel-tester.h",
6072 ] + MICROKERNEL_TEST_HDRS,
6073 deps = MICROKERNEL_TEST_DEPS,
6074)
6075
6076xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07006077 name = "f32_vdivc_relu_test",
6078 srcs = [
6079 "test/f32-vdivc-relu.cc",
6080 "test/vbinaryc-microkernel-tester.h",
6081 ] + MICROKERNEL_TEST_HDRS,
6082 deps = MICROKERNEL_TEST_DEPS,
6083)
6084
6085xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07006086 name = "f32_vrdivc_test",
6087 srcs = [
6088 "test/f32-vrdivc.cc",
6089 "test/vbinaryc-microkernel-tester.h",
6090 ] + MICROKERNEL_TEST_HDRS,
6091 deps = MICROKERNEL_TEST_DEPS,
6092)
6093
6094xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006095 name = "f32_vrdivc_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08006096 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006097 "test/f32-vrdivc-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08006098 "test/vbinaryc-microkernel-tester.h",
6099 ] + MICROKERNEL_TEST_HDRS,
6100 deps = MICROKERNEL_TEST_DEPS,
6101)
6102
6103xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07006104 name = "f32_vrdivc_relu_test",
6105 srcs = [
6106 "test/f32-vrdivc-relu.cc",
6107 "test/vbinaryc-microkernel-tester.h",
6108 ] + MICROKERNEL_TEST_HDRS,
6109 deps = MICROKERNEL_TEST_DEPS,
6110)
6111
6112xnnpack_unit_test(
Marat Dukhaned6baaf2020-12-01 15:07:08 -08006113 name = "f32_velu_test",
6114 srcs = [
6115 "test/f32-velu.cc",
6116 "test/vunary-microkernel-tester.h",
6117 ] + MICROKERNEL_TEST_HDRS,
6118 deps = MICROKERNEL_TEST_DEPS,
6119)
6120
6121xnnpack_unit_test(
Marat Dukhan403b7d42019-12-05 12:49:11 -08006122 name = "f32_vmax_test",
6123 srcs = [
6124 "test/f32-vmax.cc",
6125 "test/vbinary-microkernel-tester.h",
6126 ] + MICROKERNEL_TEST_HDRS,
6127 deps = MICROKERNEL_TEST_DEPS,
6128)
6129
6130xnnpack_unit_test(
6131 name = "f32_vmaxc_test",
6132 srcs = [
6133 "test/f32-vmaxc.cc",
6134 "test/vbinaryc-microkernel-tester.h",
6135 ] + MICROKERNEL_TEST_HDRS,
6136 deps = MICROKERNEL_TEST_DEPS,
6137)
6138
6139xnnpack_unit_test(
6140 name = "f32_vmin_test",
6141 srcs = [
6142 "test/f32-vmin.cc",
6143 "test/vbinary-microkernel-tester.h",
6144 ] + MICROKERNEL_TEST_HDRS,
6145 deps = MICROKERNEL_TEST_DEPS,
6146)
6147
6148xnnpack_unit_test(
6149 name = "f32_vminc_test",
6150 srcs = [
6151 "test/f32-vminc.cc",
6152 "test/vbinaryc-microkernel-tester.h",
6153 ] + MICROKERNEL_TEST_HDRS,
6154 deps = MICROKERNEL_TEST_DEPS,
6155)
6156
6157xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07006158 name = "f32_vmul_test",
6159 srcs = [
6160 "test/f32-vmul.cc",
6161 "test/vbinary-microkernel-tester.h",
6162 ] + MICROKERNEL_TEST_HDRS,
6163 deps = MICROKERNEL_TEST_DEPS,
6164)
6165
6166xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006167 name = "f32_vmul_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006168 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006169 "test/f32-vmul-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08006170 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08006171 ] + MICROKERNEL_TEST_HDRS,
6172 deps = MICROKERNEL_TEST_DEPS,
6173)
6174
6175xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07006176 name = "f32_vmul_relu_test",
6177 srcs = [
6178 "test/f32-vmul-relu.cc",
6179 "test/vbinary-microkernel-tester.h",
6180 ] + MICROKERNEL_TEST_HDRS,
6181 deps = MICROKERNEL_TEST_DEPS,
6182)
6183
6184xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07006185 name = "f32_vmulc_test",
6186 srcs = [
6187 "test/f32-vmulc.cc",
6188 "test/vbinaryc-microkernel-tester.h",
6189 ] + MICROKERNEL_TEST_HDRS,
6190 deps = MICROKERNEL_TEST_DEPS,
6191)
6192
6193xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006194 name = "f32_vmulc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08006195 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006196 "test/f32-vmulc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08006197 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006198 ] + MICROKERNEL_TEST_HDRS,
6199 deps = MICROKERNEL_TEST_DEPS,
6200)
6201
6202xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07006203 name = "f32_vmulc_relu_test",
6204 srcs = [
6205 "test/f32-vmulc-relu.cc",
6206 "test/vbinaryc-microkernel-tester.h",
6207 ] + MICROKERNEL_TEST_HDRS,
6208 deps = MICROKERNEL_TEST_DEPS,
6209)
6210
6211xnnpack_unit_test(
Frank Barchard2a1049c2020-06-03 02:31:27 -07006212 name = "f16_vmulcaddc_minmax_test",
6213 srcs = [
6214 "test/f16-vmulcaddc-minmax.cc",
6215 "test/vmulcaddc-microkernel-tester.h",
6216 "src/xnnpack/AlignedAllocator.h",
6217 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07006218 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Frank Barchard2a1049c2020-06-03 02:31:27 -07006219)
6220
6221xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07006222 name = "f32_vmulcaddc_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006223 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07006224 "test/f32-vmulcaddc-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006225 "test/vmulcaddc-microkernel-tester.h",
6226 "src/xnnpack/AlignedAllocator.h",
6227 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07006228 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07006229)
6230
6231xnnpack_unit_test(
Marat Dukhan8cc7efe2020-06-10 16:24:27 -07006232 name = "f32_vlrelu_test",
6233 srcs = [
6234 "test/f32-vlrelu.cc",
6235 "test/vunary-microkernel-tester.h",
6236 ] + MICROKERNEL_TEST_HDRS,
6237 deps = MICROKERNEL_TEST_DEPS,
6238)
6239
6240xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07006241 name = "f32_vneg_test",
6242 srcs = [
6243 "test/f32-vneg.cc",
6244 "test/vunary-microkernel-tester.h",
6245 ] + MICROKERNEL_TEST_HDRS,
6246 deps = MICROKERNEL_TEST_DEPS,
6247)
6248
6249xnnpack_unit_test(
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07006250 name = "f32_vrndne_test",
6251 srcs = [
6252 "test/f32-vrndne.cc",
6253 "test/vunary-microkernel-tester.h",
6254 ] + MICROKERNEL_TEST_HDRS,
6255 deps = MICROKERNEL_TEST_DEPS,
6256)
6257
6258xnnpack_unit_test(
6259 name = "f32_vrndz_test",
6260 srcs = [
6261 "test/f32-vrndz.cc",
6262 "test/vunary-microkernel-tester.h",
6263 ] + MICROKERNEL_TEST_HDRS,
6264 deps = MICROKERNEL_TEST_DEPS,
6265)
6266
6267xnnpack_unit_test(
6268 name = "f32_vrndu_test",
6269 srcs = [
6270 "test/f32-vrndu.cc",
6271 "test/vunary-microkernel-tester.h",
6272 ] + MICROKERNEL_TEST_HDRS,
6273 deps = MICROKERNEL_TEST_DEPS,
6274)
6275
6276xnnpack_unit_test(
6277 name = "f32_vrndd_test",
6278 srcs = [
6279 "test/f32-vrndd.cc",
6280 "test/vunary-microkernel-tester.h",
6281 ] + MICROKERNEL_TEST_HDRS,
6282 deps = MICROKERNEL_TEST_DEPS,
6283)
6284
6285xnnpack_unit_test(
Marat Dukhan05ac8e32019-10-21 15:39:33 -07006286 name = "f32_vscale_test",
6287 srcs = [
6288 "test/f32-vscale.cc",
6289 "test/vscale-microkernel-tester.h",
6290 ] + MICROKERNEL_TEST_HDRS,
6291 deps = MICROKERNEL_TEST_DEPS,
6292)
6293
6294xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07006295 name = "f32_vscaleexpminusmax_test",
6296 srcs = [
6297 "test/f32-vscaleexpminusmax.cc",
6298 "test/vscaleexpminusmax-microkernel-tester.h",
6299 ] + MICROKERNEL_TEST_HDRS,
6300 deps = MICROKERNEL_TEST_DEPS,
6301)
6302
6303xnnpack_unit_test(
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07006304 name = "f32_vscaleextexp_test",
6305 srcs = [
6306 "test/f32-vscaleextexp.cc",
6307 "test/vscaleextexp-microkernel-tester.h",
6308 ] + MICROKERNEL_TEST_HDRS,
6309 deps = MICROKERNEL_TEST_DEPS,
6310)
6311
6312xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07006313 name = "f32_vsqr_test",
6314 srcs = [
6315 "test/f32-vsqr.cc",
6316 "test/vunary-microkernel-tester.h",
6317 ] + MICROKERNEL_TEST_HDRS,
6318 deps = MICROKERNEL_TEST_DEPS,
6319)
6320
6321xnnpack_unit_test(
Marat Dukhan13bafb02020-06-05 00:43:11 -07006322 name = "f32_vsqrdiff_test",
6323 srcs = [
6324 "test/f32-vsqrdiff.cc",
6325 "test/vbinary-microkernel-tester.h",
6326 ] + MICROKERNEL_TEST_HDRS,
6327 deps = MICROKERNEL_TEST_DEPS,
6328)
6329
6330xnnpack_unit_test(
6331 name = "f32_vsqrdiffc_test",
6332 srcs = [
6333 "test/f32-vsqrdiffc.cc",
6334 "test/vbinaryc-microkernel-tester.h",
6335 ] + MICROKERNEL_TEST_HDRS,
6336 deps = MICROKERNEL_TEST_DEPS,
6337)
6338
6339xnnpack_unit_test(
Marat Dukhanf4db2f32020-06-30 10:55:30 -07006340 name = "f32_vsqrt_test",
6341 srcs = [
6342 "test/f32-vsqrt.cc",
6343 "test/vunary-microkernel-tester.h",
6344 ] + MICROKERNEL_TEST_HDRS,
6345 deps = MICROKERNEL_TEST_DEPS,
6346)
6347
6348xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07006349 name = "f32_vsub_test",
6350 srcs = [
6351 "test/f32-vsub.cc",
6352 "test/vbinary-microkernel-tester.h",
6353 ] + MICROKERNEL_TEST_HDRS,
6354 deps = MICROKERNEL_TEST_DEPS,
6355)
6356
6357xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006358 name = "f32_vsub_minmax_test",
Marat Dukhan97579532019-10-18 16:40:39 -07006359 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006360 "test/f32-vsub-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08006361 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08006362 ] + MICROKERNEL_TEST_HDRS,
6363 deps = MICROKERNEL_TEST_DEPS,
6364)
6365
6366xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07006367 name = "f32_vsub_relu_test",
6368 srcs = [
6369 "test/f32-vsub-relu.cc",
6370 "test/vbinary-microkernel-tester.h",
6371 ] + MICROKERNEL_TEST_HDRS,
6372 deps = MICROKERNEL_TEST_DEPS,
6373)
6374
6375xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07006376 name = "f32_vsubc_test",
6377 srcs = [
6378 "test/f32-vsubc.cc",
6379 "test/vbinaryc-microkernel-tester.h",
6380 ] + MICROKERNEL_TEST_HDRS,
6381 deps = MICROKERNEL_TEST_DEPS,
6382)
6383
6384xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006385 name = "f32_vsubc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08006386 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006387 "test/f32-vsubc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08006388 "test/vbinaryc-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08006389 ] + MICROKERNEL_TEST_HDRS,
6390 deps = MICROKERNEL_TEST_DEPS,
6391)
6392
6393xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07006394 name = "f32_vsubc_relu_test",
6395 srcs = [
6396 "test/f32-vsubc-relu.cc",
6397 "test/vbinaryc-microkernel-tester.h",
6398 ] + MICROKERNEL_TEST_HDRS,
6399 deps = MICROKERNEL_TEST_DEPS,
6400)
6401
6402xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07006403 name = "f32_vrsubc_test",
6404 srcs = [
6405 "test/f32-vrsubc.cc",
6406 "test/vbinaryc-microkernel-tester.h",
6407 ] + MICROKERNEL_TEST_HDRS,
6408 deps = MICROKERNEL_TEST_DEPS,
6409)
6410
6411xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006412 name = "f32_vrsubc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08006413 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006414 "test/f32-vrsubc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08006415 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan97579532019-10-18 16:40:39 -07006416 ] + MICROKERNEL_TEST_HDRS,
6417 deps = MICROKERNEL_TEST_DEPS,
6418)
6419
6420xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07006421 name = "f32_vrsubc_relu_test",
6422 srcs = [
6423 "test/f32-vrsubc-relu.cc",
6424 "test/vbinaryc-microkernel-tester.h",
6425 ] + MICROKERNEL_TEST_HDRS,
6426 deps = MICROKERNEL_TEST_DEPS,
6427)
6428
6429xnnpack_unit_test(
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07006430 name = "qs8_dwconv_minmax_test",
6431 srcs = [
6432 "test/qs8-dwconv-minmax.cc",
6433 "test/dwconv-microkernel-tester.h",
6434 "src/xnnpack/AlignedAllocator.h",
6435 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
6436 deps = MICROKERNEL_TEST_DEPS + [":packing"],
6437)
6438
6439xnnpack_unit_test(
Marat Dukhan4ed53f42020-08-06 01:12:55 -07006440 name = "qs8_gavgpool_minmax_test",
6441 srcs = [
6442 "test/qs8-gavgpool-minmax.cc",
6443 "test/gavgpool-microkernel-tester.h",
6444 "src/xnnpack/AlignedAllocator.h",
6445 ] + MICROKERNEL_TEST_HDRS,
6446 deps = MICROKERNEL_TEST_DEPS,
6447)
6448
6449xnnpack_unit_test(
Marat Dukhan595e1702020-07-31 10:12:52 -07006450 name = "qs8_gemm_minmax_test",
6451 srcs = [
6452 "test/qs8-gemm-minmax.cc",
6453 "test/gemm-microkernel-tester.h",
6454 "src/xnnpack/AlignedAllocator.h",
6455 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
6456 deps = MICROKERNEL_TEST_DEPS + [":packing"],
6457)
6458
6459xnnpack_unit_test(
Marat Dukhanf9480682020-07-31 14:50:24 -07006460 name = "qs8_igemm_minmax_test",
6461 srcs = [
6462 "test/qs8-igemm-minmax.cc",
6463 "test/gemm-microkernel-tester.h",
6464 "src/xnnpack/AlignedAllocator.h",
6465 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
6466 deps = MICROKERNEL_TEST_DEPS + [":packing"],
6467)
6468
6469xnnpack_unit_test(
6470 name = "qs8_requantization_test",
6471 srcs = [
6472 "src/xnnpack/requantization-stubs.h",
6473 "test/qs8-requantization.cc",
6474 "test/requantization-tester.h",
6475 ] + MICROKERNEL_TEST_HDRS,
6476 deps = MICROKERNEL_TEST_DEPS,
6477)
6478
6479xnnpack_unit_test(
Marat Dukhand9f3ad42020-08-10 12:30:58 -07006480 name = "qs8_vadd_minmax_test",
6481 srcs = [
6482 "test/qs8-vadd-minmax.cc",
6483 "test/vadd-microkernel-tester.h",
6484 ] + MICROKERNEL_TEST_HDRS,
6485 deps = MICROKERNEL_TEST_DEPS,
6486)
6487
6488xnnpack_unit_test(
Marat Dukhan0270d9f2020-08-11 00:56:46 -07006489 name = "qs8_vaddc_minmax_test",
6490 srcs = [
6491 "test/qs8-vaddc-minmax.cc",
6492 "test/vaddc-microkernel-tester.h",
6493 ] + MICROKERNEL_TEST_HDRS,
6494 deps = MICROKERNEL_TEST_DEPS,
6495)
6496
6497xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07006498 name = "qu8_avgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006499 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07006500 "test/qu8-avgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006501 "test/avgpool-microkernel-tester.h",
6502 "src/xnnpack/AlignedAllocator.h",
6503 ] + MICROKERNEL_TEST_HDRS,
6504 deps = MICROKERNEL_TEST_DEPS,
6505)
6506
6507xnnpack_unit_test(
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07006508 name = "qu8_dwconv_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006509 srcs = [
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07006510 "test/qu8-dwconv-minmax.cc",
6511 "test/dwconv-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006512 "src/xnnpack/AlignedAllocator.h",
6513 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07006514 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07006515)
6516
6517xnnpack_unit_test(
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07006518 name = "qu8_igemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006519 srcs = [
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07006520 "test/qu8-igemm-minmax.cc",
6521 "test/gemm-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006522 "src/xnnpack/AlignedAllocator.h",
6523 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07006524 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07006525)
6526
6527xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07006528 name = "qu8_gavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006529 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07006530 "test/qu8-gavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006531 "test/gavgpool-microkernel-tester.h",
6532 "src/xnnpack/AlignedAllocator.h",
6533 ] + MICROKERNEL_TEST_HDRS,
6534 deps = MICROKERNEL_TEST_DEPS,
6535)
6536
6537xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07006538 name = "qu8_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006539 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07006540 "test/qu8-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006541 "test/gemm-microkernel-tester.h",
6542 "src/xnnpack/AlignedAllocator.h",
6543 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07006544 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07006545)
6546
6547xnnpack_unit_test(
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07006548 name = "qu8_requantization_test",
6549 srcs = [
6550 "src/xnnpack/requantization-stubs.h",
6551 "test/qu8-requantization.cc",
6552 "test/requantization-tester.h",
6553 ] + MICROKERNEL_TEST_HDRS,
6554 deps = MICROKERNEL_TEST_DEPS,
6555)
6556
6557xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07006558 name = "qu8_vadd_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006559 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07006560 "test/qu8-vadd-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006561 "test/vadd-microkernel-tester.h",
6562 ] + MICROKERNEL_TEST_HDRS,
6563 deps = MICROKERNEL_TEST_DEPS,
6564)
6565
6566xnnpack_unit_test(
6567 name = "u8_clamp_test",
6568 srcs = [
6569 "test/u8-clamp.cc",
6570 "test/clamp-microkernel-tester.h",
6571 ] + MICROKERNEL_TEST_HDRS,
6572 deps = MICROKERNEL_TEST_DEPS,
6573)
6574
6575xnnpack_unit_test(
6576 name = "u8_lut32norm_test",
6577 srcs = [
6578 "test/u8-lut32norm.cc",
6579 "test/lut-norm-microkernel-tester.h",
6580 ] + MICROKERNEL_TEST_HDRS,
6581 deps = MICROKERNEL_TEST_DEPS,
6582)
6583
6584xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07006585 name = "u8_maxpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006586 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07006587 "test/u8-maxpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006588 "test/maxpool-microkernel-tester.h",
6589 ] + MICROKERNEL_TEST_HDRS,
6590 deps = MICROKERNEL_TEST_DEPS,
6591)
6592
6593xnnpack_unit_test(
6594 name = "u8_rmax_test",
6595 srcs = [
6596 "test/u8-rmax.cc",
6597 "test/rmax-microkernel-tester.h",
6598 ] + MICROKERNEL_TEST_HDRS,
6599 deps = MICROKERNEL_TEST_DEPS,
6600)
6601
6602xnnpack_unit_test(
Marat Dukhanad71b9a2020-11-20 00:01:51 -08006603 name = "x32_depthtospace2d_chw2hwc_test",
Yury Kartynnike7841862020-11-04 18:22:18 -08006604 srcs = [
Marat Dukhanad71b9a2020-11-20 00:01:51 -08006605 "test/x32-depthtospace2d-chw2hwc.cc",
6606 "test/depthtospace-microkernel-tester.h",
Yury Kartynnike7841862020-11-04 18:22:18 -08006607 ] + MICROKERNEL_TEST_HDRS,
6608 deps = MICROKERNEL_TEST_DEPS,
6609)
6610
6611xnnpack_unit_test(
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -07006612 name = "x32_fill_test",
6613 srcs = [
6614 "test/x32-fill.cc",
6615 "test/fill-microkernel-tester.h",
6616 ] + MICROKERNEL_TEST_HDRS,
6617 deps = MICROKERNEL_TEST_DEPS,
6618)
6619
6620xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07006621 name = "x32_packx_test",
6622 srcs = [
6623 "test/x32-packx.cc",
6624 "test/pack-microkernel-tester.h",
6625 "src/xnnpack/AlignedAllocator.h",
6626 ] + MICROKERNEL_TEST_HDRS,
6627 deps = MICROKERNEL_TEST_DEPS,
6628)
6629
6630xnnpack_unit_test(
6631 name = "x32_pad_test",
6632 srcs = [
6633 "test/x32-pad.cc",
6634 "test/pad-microkernel-tester.h",
6635 ] + MICROKERNEL_TEST_HDRS,
6636 deps = MICROKERNEL_TEST_DEPS,
6637)
6638
6639xnnpack_unit_test(
6640 name = "x32_unpool_test",
6641 srcs = [
6642 "test/x32-unpool.cc",
6643 "test/unpool-microkernel-tester.h",
6644 ] + MICROKERNEL_TEST_HDRS,
6645 deps = MICROKERNEL_TEST_DEPS,
6646)
6647
6648xnnpack_unit_test(
6649 name = "x32_zip_test",
6650 srcs = [
6651 "test/x32-zip.cc",
6652 "test/zip-microkernel-tester.h",
6653 ] + MICROKERNEL_TEST_HDRS,
6654 deps = MICROKERNEL_TEST_DEPS,
6655)
6656
6657xnnpack_unit_test(
6658 name = "x8_lut_test",
6659 srcs = [
6660 "test/x8-lut.cc",
6661 "test/lut-microkernel-tester.h",
6662 ] + MICROKERNEL_TEST_HDRS,
6663 deps = MICROKERNEL_TEST_DEPS,
6664)
6665
6666xnnpack_unit_test(
6667 name = "x8_zip_test",
6668 srcs = [
6669 "test/x8-zip.cc",
6670 "test/zip-microkernel-tester.h",
6671 ] + MICROKERNEL_TEST_HDRS,
6672 deps = MICROKERNEL_TEST_DEPS,
6673)
6674
Marat Dukhan20c3b922020-03-10 03:45:06 -07006675########################## Size tests for the library #########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07006676
6677xnnpack_binary(
Marat Dukhan20c3b922020-03-10 03:45:06 -07006678 name = "operator_size_test",
6679 srcs = ["test/operator-size.c"],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07006680 deps = [":xnnpack_operators_nhwc_f32"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07006681)
6682
Marat Dukhan20c3b922020-03-10 03:45:06 -07006683xnnpack_binary(
6684 name = "subgraph_size_test",
6685 srcs = ["test/subgraph-size.c"],
6686 deps = [":XNNPACK"],
6687)
6688
6689########################### Unit tests for operators ##########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07006690
6691xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07006692 name = "abs_nc_test",
6693 srcs = [
6694 "test/abs-nc.cc",
6695 "test/abs-operator-tester.h",
6696 ],
6697 deps = OPERATOR_TEST_DEPS,
6698)
6699
6700xnnpack_unit_test(
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08006701 name = "add_nd_test",
6702 srcs = [
6703 "test/add-nd.cc",
6704 "test/binary-elementwise-operator-tester.h",
6705 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006706 deps = OPERATOR_TEST_DEPS,
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08006707)
6708
6709xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006710 name = "argmax_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006711 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006712 "test/argmax-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006713 "test/argmax-pooling-operator-tester.h",
6714 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07006715 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006716)
6717
6718xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006719 name = "average_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006720 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006721 "test/average-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006722 "test/average-pooling-operator-tester.h",
6723 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07006724 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006725)
6726
6727xnnpack_unit_test(
Marat Dukhan64e52512020-06-09 13:41:16 -07006728 name = "bankers_rounding_nc_test",
6729 srcs = [
6730 "test/bankers-rounding-nc.cc",
6731 "test/bankers-rounding-operator-tester.h",
6732 ],
6733 deps = OPERATOR_TEST_DEPS,
6734)
6735
6736xnnpack_unit_test(
6737 name = "ceiling_nc_test",
6738 srcs = [
6739 "test/ceiling-nc.cc",
6740 "test/ceiling-operator-tester.h",
6741 ],
6742 deps = OPERATOR_TEST_DEPS,
6743)
6744
6745xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006746 name = "channel_shuffle_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006747 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006748 "test/channel-shuffle-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006749 "test/channel-shuffle-operator-tester.h",
6750 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006751 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006752)
6753
6754xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006755 name = "clamp_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006756 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006757 "test/clamp-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006758 "test/clamp-operator-tester.h",
6759 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006760 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006761)
6762
6763xnnpack_unit_test(
Marat Dukhan065b11e2020-05-22 09:49:41 -07006764 name = "constant_pad_nd_test",
6765 srcs = [
6766 "test/constant-pad-nd.cc",
6767 "test/constant-pad-operator-tester.h",
6768 ],
6769 deps = OPERATOR_TEST_DEPS,
6770)
6771
6772xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006773 name = "convolution_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006774 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006775 "test/convolution-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006776 "test/convolution-operator-tester.h",
6777 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006778 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006779)
6780
6781xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006782 name = "convolution_nchw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006783 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006784 "test/convolution-nchw.cc",
6785 "test/convolution-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006786 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006787 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006788)
6789
6790xnnpack_unit_test(
Marat Dukhan4e21b272020-06-04 18:45:01 -07006791 name = "copy_nc_test",
6792 srcs = [
6793 "test/copy-nc.cc",
6794 "test/copy-operator-tester.h",
6795 ],
6796 deps = OPERATOR_TEST_DEPS,
6797)
6798
6799xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006800 name = "deconvolution_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006801 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006802 "test/deconvolution-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006803 "test/deconvolution-operator-tester.h",
6804 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07006805 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006806)
6807
6808xnnpack_unit_test(
Marat Dukhan188d1042020-11-24 23:39:40 -08006809 name = "depth_to_space_nchw2nhwc_test",
Artsiom Ablavatski0f1dc182020-11-05 19:21:50 -08006810 srcs = [
Marat Dukhan188d1042020-11-24 23:39:40 -08006811 "test/depth-to-space-nchw2nhwc.cc",
Artsiom Ablavatski0f1dc182020-11-05 19:21:50 -08006812 "test/depth-to-space-operator-tester.h",
6813 ] + OPERATOR_TEST_PARAMS_HDRS,
6814 deps = OPERATOR_TEST_DEPS,
6815)
6816
6817xnnpack_unit_test(
Marat Dukhan0e521172020-11-25 13:10:04 -08006818 name = "depth_to_space_nhwc_test",
6819 srcs = [
6820 "test/depth-to-space-nhwc.cc",
6821 "test/depth-to-space-operator-tester.h",
6822 ] + OPERATOR_TEST_PARAMS_HDRS,
6823 deps = OPERATOR_TEST_DEPS,
6824)
6825
6826xnnpack_unit_test(
Marat Dukhan69180502019-12-06 15:00:31 -08006827 name = "divide_nd_test",
6828 srcs = [
6829 "test/binary-elementwise-operator-tester.h",
6830 "test/divide-nd.cc",
6831 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006832 deps = OPERATOR_TEST_DEPS,
Marat Dukhan69180502019-12-06 15:00:31 -08006833)
6834
6835xnnpack_unit_test(
Marat Dukhanb6bd4bc2020-12-01 17:01:40 -08006836 name = "elu_nc_test",
6837 srcs = [
6838 "test/elu-nc.cc",
6839 "test/elu-operator-tester.h",
6840 ],
6841 deps = OPERATOR_TEST_DEPS,
6842)
6843
6844xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006845 name = "fully_connected_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006846 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006847 "test/fully-connected-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006848 "test/fully-connected-operator-tester.h",
6849 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006850 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006851)
6852
6853xnnpack_unit_test(
Marat Dukhan64e52512020-06-09 13:41:16 -07006854 name = "floor_nc_test",
6855 srcs = [
6856 "test/floor-nc.cc",
6857 "test/floor-operator-tester.h",
6858 ],
6859 deps = OPERATOR_TEST_DEPS,
6860)
6861
6862xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006863 name = "global_average_pooling_nwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006864 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006865 "test/global-average-pooling-nwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006866 "test/global-average-pooling-operator-tester.h",
Marat Dukhanef61d022020-06-19 13:54:49 -07006867 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07006868 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006869)
6870
6871xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006872 name = "global_average_pooling_ncw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006873 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006874 "test/global-average-pooling-ncw.cc",
6875 "test/global-average-pooling-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006876 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006877 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006878)
6879
6880xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006881 name = "hardswish_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006882 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006883 "test/hardswish-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006884 "test/hardswish-operator-tester.h",
6885 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006886 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006887)
6888
6889xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006890 name = "leaky_relu_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006891 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006892 "test/leaky-relu-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006893 "test/leaky-relu-operator-tester.h",
6894 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006895 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006896)
6897
6898xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006899 name = "max_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006900 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006901 "test/max-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006902 "test/max-pooling-operator-tester.h",
6903 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07006904 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006905)
6906
6907xnnpack_unit_test(
Marat Dukhan79e7f842019-12-05 14:35:50 -08006908 name = "maximum_nd_test",
6909 srcs = [
6910 "test/binary-elementwise-operator-tester.h",
6911 "test/maximum-nd.cc",
6912 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006913 deps = OPERATOR_TEST_DEPS,
Marat Dukhan79e7f842019-12-05 14:35:50 -08006914)
6915
6916xnnpack_unit_test(
6917 name = "minimum_nd_test",
6918 srcs = [
6919 "test/binary-elementwise-operator-tester.h",
6920 "test/minimum-nd.cc",
6921 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006922 deps = OPERATOR_TEST_DEPS,
Marat Dukhan79e7f842019-12-05 14:35:50 -08006923)
6924
6925xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006926 name = "multiply_nd_test",
Marat Dukhanca2733c2019-11-15 23:21:17 -08006927 srcs = [
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08006928 "test/binary-elementwise-operator-tester.h",
Marat Dukhanefc47b82019-11-18 09:25:38 -08006929 "test/multiply-nd.cc",
Marat Dukhanca2733c2019-11-15 23:21:17 -08006930 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006931 deps = OPERATOR_TEST_DEPS,
Marat Dukhanca2733c2019-11-15 23:21:17 -08006932)
6933
6934xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07006935 name = "negate_nc_test",
6936 srcs = [
6937 "test/negate-nc.cc",
6938 "test/negate-operator-tester.h",
6939 ],
6940 deps = OPERATOR_TEST_DEPS,
6941)
6942
6943xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006944 name = "prelu_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006945 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006946 "test/prelu-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006947 "test/prelu-operator-tester.h",
6948 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07006949 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006950)
6951
6952xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006953 name = "resize_bilinear_nhwc_test",
Marat Dukhan69722492019-11-11 19:55:50 -08006954 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006955 "test/resize-bilinear-nhwc.cc",
Marat Dukhan69722492019-11-11 19:55:50 -08006956 "test/resize-bilinear-operator-tester.h",
6957 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07006958 deps = OPERATOR_TEST_DEPS,
Marat Dukhan69722492019-11-11 19:55:50 -08006959)
6960
6961xnnpack_unit_test(
Artsiom Ablavatski97918102020-10-27 15:52:59 -07006962 name = "resize_bilinear_nchw_test",
6963 srcs = [
6964 "test/resize-bilinear-nchw.cc",
6965 "test/resize-bilinear-operator-tester.h",
6966 ] + OPERATOR_TEST_PARAMS_HDRS,
6967 deps = OPERATOR_TEST_DEPS,
6968)
6969
6970xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006971 name = "sigmoid_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006972 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006973 "test/sigmoid-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006974 "test/sigmoid-operator-tester.h",
6975 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006976 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006977)
6978
6979xnnpack_unit_test(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08006980 name = "softmax_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006981 srcs = [
Marat Dukhanfd8e6892020-01-27 15:25:25 -08006982 "test/softmax-nc.cc",
6983 "test/softmax-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006984 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006985 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006986)
6987
6988xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07006989 name = "square_nc_test",
6990 srcs = [
6991 "test/square-nc.cc",
6992 "test/square-operator-tester.h",
6993 ],
6994 deps = OPERATOR_TEST_DEPS,
6995)
6996
6997xnnpack_unit_test(
Marat Dukhan6804bbd2020-06-30 19:26:11 -07006998 name = "square_root_nc_test",
6999 srcs = [
7000 "test/square-root-nc.cc",
7001 "test/square-root-operator-tester.h",
7002 ],
7003 deps = OPERATOR_TEST_DEPS,
7004)
7005
7006xnnpack_unit_test(
Marat Dukhanf7399262020-06-05 10:58:44 -07007007 name = "squared_difference_nd_test",
7008 srcs = [
7009 "test/binary-elementwise-operator-tester.h",
7010 "test/squared-difference-nd.cc",
7011 ],
7012 deps = OPERATOR_TEST_DEPS,
7013)
7014
7015xnnpack_unit_test(
Marat Dukhan05f3f6d2019-12-03 15:13:53 -08007016 name = "subtract_nd_test",
7017 srcs = [
7018 "test/binary-elementwise-operator-tester.h",
7019 "test/subtract-nd.cc",
7020 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07007021 deps = OPERATOR_TEST_DEPS,
Marat Dukhan05f3f6d2019-12-03 15:13:53 -08007022)
7023
7024xnnpack_unit_test(
Marat Dukhan64e52512020-06-09 13:41:16 -07007025 name = "truncation_nc_test",
7026 srcs = [
7027 "test/truncation-nc.cc",
7028 "test/truncation-operator-tester.h",
7029 ],
7030 deps = OPERATOR_TEST_DEPS,
7031)
7032
7033xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08007034 name = "unpooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007035 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08007036 "test/unpooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007037 "test/unpooling-operator-tester.h",
7038 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07007039 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07007040)
7041
Chao Mei6ddfc602020-05-13 22:29:36 -07007042############################### Misc unit tests ###############################
7043
7044xnnpack_unit_test(
7045 name = "memory_planner_test",
7046 srcs = [
7047 "test/memory-planner-test.cc",
7048 ],
7049 deps = [
7050 ":XNNPACK",
7051 ":memory_planner",
7052 ],
7053)
7054
XNNPACK Teamab8c4c82020-10-09 08:05:51 -07007055xnnpack_unit_test(
7056 name = "subgraph_nchw_test",
7057 srcs = [
7058 "src/xnnpack/subgraph.h",
7059 "test/subgraph-nchw.cc",
7060 "test/subgraph-tester.h",
7061 ],
7062 deps = [
7063 ":XNNPACK",
7064 ],
7065)
7066
Marat Dukhan08c4a432019-10-03 09:29:21 -07007067############################# Build configurations #############################
7068
Marat Dukhanb8642352019-10-30 15:43:02 -07007069# Enables usage of assembly kernels.
Marat Dukhan08c4a432019-10-03 09:29:21 -07007070config_setting(
Marat Dukhanb8642352019-10-30 15:43:02 -07007071 name = "xnn_enable_assembly_explicit_true",
7072 define_values = {"xnn_enable_assembly": "true"},
7073)
7074
7075# Disables usage of assembly kernels.
7076config_setting(
7077 name = "xnn_enable_assembly_explicit_false",
7078 define_values = {"xnn_enable_assembly": "false"},
7079)
7080
Marat Dukhan9de90e02020-06-18 16:04:12 -07007081# Enables usage of sparse inference.
7082config_setting(
7083 name = "xnn_enable_sparse_explicit_true",
7084 define_values = {"xnn_enable_sparse": "true"},
7085)
7086
7087# Disables usage of sparse inference.
7088config_setting(
7089 name = "xnn_enable_sparse_explicit_false",
7090 define_values = {"xnn_enable_sparse": "false"},
7091)
7092
Marat Dukhan05702cf2020-03-26 15:41:33 -07007093# Disables usage of HMP-aware optimizations.
7094config_setting(
7095 name = "xnn_enable_hmp_explicit_false",
7096 define_values = {"xnn_enable_hmp": "false"},
7097)
7098
Chao Mei6ddfc602020-05-13 22:29:36 -07007099# Enable usage of optimized memory allocation
7100config_setting(
7101 name = "xnn_enable_memopt_explicit_true",
7102 define_values = {"xnnpack_enable_memopt": "true"},
7103)
7104
7105# Disable usage of optimized memory allocation
7106config_setting(
7107 name = "xnn_enable_memopt_explicit_false",
7108 define_values = {"xnnpack_enable_memopt": "false"},
7109)
7110
Marat Dukhanb8642352019-10-30 15:43:02 -07007111# Builds with -c dbg
7112config_setting(
7113 name = "debug_build",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007114 values = {
Marat Dukhanb8642352019-10-30 15:43:02 -07007115 "compilation_mode": "dbg",
7116 },
7117)
7118
7119# Builds with -c opt
7120config_setting(
7121 name = "optimized_build",
7122 values = {
7123 "compilation_mode": "opt",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007124 },
7125)
7126
7127config_setting(
Marat Dukhanb8642352019-10-30 15:43:02 -07007128 name = "linux_k8",
7129 values = {"cpu": "k8"},
7130)
7131
7132config_setting(
Marat Dukhan582094e2020-04-30 17:21:25 -07007133 name = "linux_arm",
7134 values = {"cpu": "arm"},
Marat Dukhan4e45e662019-10-03 15:40:24 -07007135)
7136
7137config_setting(
Marat Dukhanf0bd4de2020-06-15 15:53:19 -07007138 name = "linux_armeabi",
7139 values = {"cpu": "armeabi"},
7140)
7141
7142config_setting(
Terry Heo68eef3f2020-04-13 22:53:52 -07007143 name = "linux_armhf",
7144 values = {"cpu": "armhf"},
7145)
7146
7147config_setting(
Marat Dukhana720e932020-06-10 13:01:11 -07007148 name = "linux_armv7a",
7149 values = {"cpu": "armv7a"},
7150)
7151
7152config_setting(
Marat Dukhan582094e2020-04-30 17:21:25 -07007153 name = "linux_aarch64",
7154 values = {"cpu": "aarch64"},
7155)
7156
7157config_setting(
Marat Dukhan08c4a432019-10-03 09:29:21 -07007158 name = "android",
7159 values = {"crosstool_top": "//external:android/crosstool"},
7160)
7161
7162config_setting(
7163 name = "android_armv7",
7164 values = {
7165 "crosstool_top": "//external:android/crosstool",
7166 "cpu": "armeabi-v7a",
7167 },
7168)
7169
7170config_setting(
7171 name = "android_arm64",
7172 values = {
7173 "crosstool_top": "//external:android/crosstool",
7174 "cpu": "arm64-v8a",
7175 },
7176)
7177
7178config_setting(
7179 name = "android_x86",
7180 values = {
7181 "crosstool_top": "//external:android/crosstool",
7182 "cpu": "x86",
7183 },
7184)
7185
7186config_setting(
7187 name = "android_x86_64",
7188 values = {
7189 "crosstool_top": "//external:android/crosstool",
7190 "cpu": "x86_64",
7191 },
7192)
7193
7194config_setting(
Marat Dukhan10a38082020-04-17 03:58:35 -07007195 name = "windows_x86_64",
7196 values = {"cpu": "x64_windows"},
Marat Dukhan9fe932e2020-04-11 17:14:15 -07007197)
7198
7199config_setting(
Marat Dukhan10a38082020-04-17 03:58:35 -07007200 name = "windows_x86_64_clang",
7201 values = {
7202 "compiler": "clang-cl",
7203 "cpu": "x64_windows",
7204 },
7205)
7206
7207config_setting(
7208 name = "windows_x86_64_mingw",
7209 values = {
7210 "compiler": "mingw-gcc",
7211 "cpu": "x64_windows",
7212 },
7213)
7214
7215config_setting(
7216 name = "windows_x86_64_msys",
7217 values = {
7218 "compiler": "msys-gcc",
7219 "cpu": "x64_windows",
7220 },
Marat Dukhan9fe932e2020-04-11 17:14:15 -07007221)
7222
7223config_setting(
Marat Dukhan885ca242019-10-07 09:17:32 -07007224 name = "macos_x86_64",
7225 values = {
7226 "apple_platform_type": "macos",
7227 "cpu": "darwin",
7228 },
7229)
7230
7231config_setting(
Marat Dukhan08c4a432019-10-03 09:29:21 -07007232 name = "emscripten",
Marat Dukhan1a729ec2019-10-07 09:31:44 -07007233 values = {"crosstool_top": "//toolchain:emscripten"},
Marat Dukhan08c4a432019-10-03 09:29:21 -07007234)
7235
7236config_setting(
7237 name = "emscripten_wasm",
7238 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07007239 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007240 "cpu": "wasm",
7241 },
7242)
7243
7244config_setting(
7245 name = "emscripten_wasmsimd",
7246 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07007247 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007248 "cpu": "wasm",
Marat Dukhan81c62602020-05-29 13:22:49 -07007249 "copt": "-msimd128",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007250 },
7251)
7252
7253config_setting(
Marat Dukhan1498d1d2020-02-11 20:00:05 -08007254 name = "ios_armv7",
7255 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08007256 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08007257 "cpu": "ios_armv7",
7258 },
7259)
7260
7261config_setting(
7262 name = "ios_arm64",
7263 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08007264 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08007265 "cpu": "ios_arm64",
7266 },
7267)
7268
7269config_setting(
7270 name = "ios_arm64e",
7271 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08007272 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08007273 "cpu": "ios_arm64e",
7274 },
7275)
7276
7277config_setting(
7278 name = "ios_x86",
7279 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08007280 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08007281 "cpu": "ios_i386",
7282 },
7283)
7284
7285config_setting(
7286 name = "ios_x86_64",
7287 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08007288 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08007289 "cpu": "ios_x86_64",
7290 },
7291)
7292
7293config_setting(
7294 name = "watchos_armv7k",
7295 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08007296 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08007297 "cpu": "watchos_armv7k",
7298 },
7299)
7300
7301config_setting(
7302 name = "watchos_arm64_32",
7303 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08007304 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08007305 "cpu": "watchos_arm64_32",
7306 },
7307)
7308
7309config_setting(
7310 name = "watchos_x86",
7311 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08007312 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08007313 "cpu": "watchos_i386",
7314 },
7315)
7316
7317config_setting(
7318 name = "watchos_x86_64",
7319 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08007320 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08007321 "cpu": "watchos_x86_64",
7322 },
7323)
7324
7325config_setting(
7326 name = "tvos_arm64",
7327 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08007328 "apple_platform_type": "tvos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08007329 "cpu": "tvos_arm64",
7330 },
7331)
7332
7333config_setting(
7334 name = "tvos_x86_64",
7335 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08007336 "apple_platform_type": "tvos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08007337 "cpu": "tvos_x86_64",
7338 },
7339)