blob: 4dfff90d596b2097ea7e6e5ed0c22da24de06251 [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",
Frank Barchard0725b8d2020-12-07 11:07:35 -0800743 "src/f32-dwconv/gen/up4x4-minmax-wasmsimd-arm-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700744 "src/f32-dwconv/gen/up4x4-minmax-wasmsimd-arm.c",
Frank Barchard0725b8d2020-12-07 11:07:35 -0800745 "src/f32-dwconv/gen/up4x4-minmax-wasmsimd-x86-acc2.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",
Frank Barchard0725b8d2020-12-07 11:07:35 -0800748 "src/f32-dwconv/gen/up4x9-minmax-wasmsimd-arm-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700749 "src/f32-dwconv/gen/up4x9-minmax-wasmsimd-arm.c",
Frank Barchard0725b8d2020-12-07 11:07:35 -0800750 "src/f32-dwconv/gen/up4x9-minmax-wasmsimd-x86-acc2.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",
Frank Barchard0725b8d2020-12-07 11:07:35 -0800753 "src/f32-dwconv/gen/up4x25-minmax-wasmsimd-arm-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700754 "src/f32-dwconv/gen/up4x25-minmax-wasmsimd-arm.c",
Frank Barchard0725b8d2020-12-07 11:07:35 -0800755 "src/f32-dwconv/gen/up4x25-minmax-wasmsimd-x86-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700756 "src/f32-dwconv/gen/up4x25-minmax-wasmsimd-x86.c",
757 "src/f32-dwconv/gen/up4x25-wasmsimd.c",
Frank Barchard0725b8d2020-12-07 11:07:35 -0800758 "src/f32-dwconv/gen/up8x4-minmax-wasmsimd-arm-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700759 "src/f32-dwconv/gen/up8x4-minmax-wasmsimd-arm.c",
Frank Barchard0725b8d2020-12-07 11:07:35 -0800760 "src/f32-dwconv/gen/up8x4-minmax-wasmsimd-x86-acc2.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",
Frank Barchard0725b8d2020-12-07 11:07:35 -0800763 "src/f32-dwconv/gen/up8x9-minmax-wasmsimd-arm-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700764 "src/f32-dwconv/gen/up8x9-minmax-wasmsimd-arm.c",
Frank Barchard0725b8d2020-12-07 11:07:35 -0800765 "src/f32-dwconv/gen/up8x9-minmax-wasmsimd-x86-acc2.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",
Frank Barchard0725b8d2020-12-07 11:07:35 -0800768 "src/f32-dwconv/gen/up8x25-minmax-wasmsimd-arm-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700769 "src/f32-dwconv/gen/up8x25-minmax-wasmsimd-arm.c",
Frank Barchard0725b8d2020-12-07 11:07:35 -0800770 "src/f32-dwconv/gen/up8x25-minmax-wasmsimd-x86-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700771 "src/f32-dwconv/gen/up8x25-minmax-wasmsimd-x86.c",
772 "src/f32-dwconv/gen/up8x25-wasmsimd.c",
Frank Barchard412e2f42020-12-11 11:40:50 -0800773 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-loadsplat-1x4-acc2.c",
774 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-loadsplat-1x4-acc3.c",
775 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-loadsplat-1x4-acc4.c",
776 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-loadsplat-1x4.c",
777 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-loadsplat-2x4-acc2.c",
778 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-loadsplat-2x4.c",
779 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-loadsplat-3x4.c",
780 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-loadsplat-4x4.c",
781 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-loadsplat-5x4.c",
782 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-loadsplat-6x4.c",
783 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-loadsplat-1x4-acc2.c",
784 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-loadsplat-1x4-acc3.c",
785 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-loadsplat-1x4-acc4.c",
786 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-loadsplat-1x4.c",
787 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-loadsplat-2x4-acc2.c",
788 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-loadsplat-2x4.c",
789 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-loadsplat-3x4.c",
790 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-loadsplat-4x4.c",
791 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-loadsplat-5x4.c",
792 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-loadsplat-6x4.c",
793 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-arm-splat-1x4-acc2.c",
794 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-arm-splat-1x4-acc3.c",
795 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-arm-splat-1x4-acc4.c",
796 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-arm-splat-1x4.c",
797 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-arm-splat-2x4-acc2.c",
798 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-arm-splat-2x4.c",
799 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-arm-splat-3x4.c",
800 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-arm-splat-4x4.c",
801 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-x86-splat-1x4-acc2.c",
802 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-x86-splat-1x4-acc3.c",
803 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-x86-splat-1x4-acc4.c",
804 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-x86-splat-1x4.c",
805 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-x86-splat-2x4-acc2.c",
806 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-x86-splat-2x4.c",
807 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-x86-splat-3x4.c",
808 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-x86-splat-4x4.c",
809 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-splat-1x4-acc2.c",
810 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-splat-1x4-acc3.c",
811 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-splat-1x4-acc4.c",
812 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-splat-1x4-acc5.c",
813 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-splat-1x4.c",
814 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-splat-2x4-acc2.c",
815 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-splat-2x4-acc3.c",
816 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-splat-2x4.c",
817 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-splat-3x4-acc2.c",
818 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-splat-3x4.c",
819 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-splat-4x4-acc2.c",
820 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-splat-4x4.c",
821 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-splat-5x4.c",
822 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-splat-1x4-acc2.c",
823 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-splat-1x4-acc3.c",
824 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-splat-1x4-acc4.c",
825 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-splat-1x4-acc5.c",
826 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-splat-1x4.c",
827 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-splat-2x4-acc2.c",
828 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-splat-2x4-acc3.c",
829 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-splat-2x4.c",
830 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-splat-3x4-acc2.c",
831 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-splat-3x4.c",
832 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-splat-4x4-acc2.c",
833 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-splat-4x4.c",
834 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-splat-5x4.c",
835 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-arm-splat-1x4-acc2.c",
836 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-arm-splat-1x4-acc3.c",
837 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-arm-splat-1x4-acc4.c",
838 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-arm-splat-1x4-acc5.c",
839 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-arm-splat-1x4.c",
840 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-arm-splat-2x4-acc2.c",
841 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-arm-splat-2x4-acc3.c",
842 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-arm-splat-2x4.c",
843 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-arm-splat-3x4-acc2.c",
844 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-arm-splat-3x4.c",
845 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-x86-splat-1x4-acc2.c",
846 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-x86-splat-1x4-acc3.c",
847 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-x86-splat-1x4-acc4.c",
848 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-x86-splat-1x4-acc5.c",
849 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-x86-splat-1x4.c",
850 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-x86-splat-2x4-acc2.c",
851 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-x86-splat-2x4-acc3.c",
852 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-x86-splat-2x4.c",
853 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-x86-splat-3x4-acc2.c",
854 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-x86-splat-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 Barchard0725b8d2020-12-07 11:07:35 -0800861 "src/f32-gemm/gen-inc/1x8inc-minmax-wasmsimd-arm-loadsplat.c",
862 "src/f32-gemm/gen-inc/1x8inc-minmax-wasmsimd-arm-splat.c",
863 "src/f32-gemm/gen-inc/1x8inc-minmax-wasmsimd-x86-loadsplat.c",
864 "src/f32-gemm/gen-inc/1x8inc-minmax-wasmsimd-x86-splat.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700865 "src/f32-gemm/gen-inc/1x8s4inc-minmax-wasmsimd-arm.c",
866 "src/f32-gemm/gen-inc/1x8s4inc-minmax-wasmsimd-x86.c",
Frank Barchard0725b8d2020-12-07 11:07:35 -0800867 "src/f32-gemm/gen-inc/3x8inc-minmax-wasmsimd-arm-loadsplat.c",
868 "src/f32-gemm/gen-inc/3x8inc-minmax-wasmsimd-arm-splat.c",
869 "src/f32-gemm/gen-inc/3x8inc-minmax-wasmsimd-x86-loadsplat.c",
870 "src/f32-gemm/gen-inc/3x8inc-minmax-wasmsimd-x86-splat.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700871 "src/f32-gemm/gen-inc/3x8s4inc-minmax-wasmsimd-arm.c",
872 "src/f32-gemm/gen-inc/3x8s4inc-minmax-wasmsimd-x86.c",
Frank Barchard0725b8d2020-12-07 11:07:35 -0800873 "src/f32-gemm/gen-inc/4x8inc-minmax-wasmsimd-arm-loadsplat.c",
874 "src/f32-gemm/gen-inc/4x8inc-minmax-wasmsimd-arm-splat.c",
875 "src/f32-gemm/gen-inc/4x8inc-minmax-wasmsimd-x86-loadsplat.c",
876 "src/f32-gemm/gen-inc/4x8inc-minmax-wasmsimd-x86-splat.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700877 "src/f32-gemm/gen-inc/4x8s4inc-minmax-wasmsimd-arm.c",
878 "src/f32-gemm/gen-inc/4x8s4inc-minmax-wasmsimd-x86.c",
Frank Barchard0725b8d2020-12-07 11:07:35 -0800879 "src/f32-gemm/gen-inc/5x8inc-minmax-wasmsimd-arm-loadsplat.c",
880 "src/f32-gemm/gen-inc/5x8inc-minmax-wasmsimd-arm-splat.c",
881 "src/f32-gemm/gen-inc/5x8inc-minmax-wasmsimd-x86-loadsplat.c",
882 "src/f32-gemm/gen-inc/5x8inc-minmax-wasmsimd-x86-splat.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700883 "src/f32-gemm/gen-inc/5x8s4inc-minmax-wasmsimd-arm.c",
884 "src/f32-gemm/gen-inc/5x8s4inc-minmax-wasmsimd-x86.c",
Frank Barchard0725b8d2020-12-07 11:07:35 -0800885 "src/f32-gemm/gen-inc/6x8inc-minmax-wasmsimd-arm-loadsplat.c",
886 "src/f32-gemm/gen-inc/6x8inc-minmax-wasmsimd-arm-splat.c",
887 "src/f32-gemm/gen-inc/6x8inc-minmax-wasmsimd-x86-loadsplat.c",
888 "src/f32-gemm/gen-inc/6x8inc-minmax-wasmsimd-x86-splat.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700889 "src/f32-gemm/gen-inc/6x8s4inc-minmax-wasmsimd-arm.c",
890 "src/f32-gemm/gen-inc/6x8s4inc-minmax-wasmsimd-x86.c",
Frank Barchard0725b8d2020-12-07 11:07:35 -0800891 "src/f32-gemm/gen/1x8-minmax-wasmsimd-arm-loadsplat.c",
892 "src/f32-gemm/gen/1x8-minmax-wasmsimd-arm-splat.c",
893 "src/f32-gemm/gen/1x8-minmax-wasmsimd-x86-loadsplat.c",
894 "src/f32-gemm/gen/1x8-minmax-wasmsimd-x86-splat.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700895 "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",
Frank Barchard0725b8d2020-12-07 11:07:35 -0800899 "src/f32-gemm/gen/3x8-minmax-wasmsimd-arm-loadsplat.c",
900 "src/f32-gemm/gen/3x8-minmax-wasmsimd-arm-splat.c",
901 "src/f32-gemm/gen/3x8-minmax-wasmsimd-x86-loadsplat.c",
902 "src/f32-gemm/gen/3x8-minmax-wasmsimd-x86-splat.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700903 "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",
Frank Barchard0725b8d2020-12-07 11:07:35 -0800909 "src/f32-gemm/gen/4x8-minmax-wasmsimd-arm-loadsplat.c",
910 "src/f32-gemm/gen/4x8-minmax-wasmsimd-arm-splat.c",
911 "src/f32-gemm/gen/4x8-minmax-wasmsimd-x86-loadsplat.c",
912 "src/f32-gemm/gen/4x8-minmax-wasmsimd-x86-splat.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700913 "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",
Frank Barchard0725b8d2020-12-07 11:07:35 -0800917 "src/f32-gemm/gen/5x8-minmax-wasmsimd-arm-loadsplat.c",
918 "src/f32-gemm/gen/5x8-minmax-wasmsimd-arm-splat.c",
919 "src/f32-gemm/gen/5x8-minmax-wasmsimd-x86-loadsplat.c",
920 "src/f32-gemm/gen/5x8-minmax-wasmsimd-x86-splat.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700921 "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",
Frank Barchard0725b8d2020-12-07 11:07:35 -0800925 "src/f32-gemm/gen/6x8-minmax-wasmsimd-arm-loadsplat.c",
926 "src/f32-gemm/gen/6x8-minmax-wasmsimd-arm-splat.c",
927 "src/f32-gemm/gen/6x8-minmax-wasmsimd-x86-loadsplat.c",
928 "src/f32-gemm/gen/6x8-minmax-wasmsimd-x86-splat.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700929 "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 Barchard0725b8d2020-12-07 11:07:35 -0800938 "src/f32-igemm/gen/1x8-minmax-wasmsimd-arm-loadsplat.c",
939 "src/f32-igemm/gen/1x8-minmax-wasmsimd-arm-splat.c",
940 "src/f32-igemm/gen/1x8-minmax-wasmsimd-x86-loadsplat.c",
941 "src/f32-igemm/gen/1x8-minmax-wasmsimd-x86-splat.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700942 "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",
Frank Barchard0725b8d2020-12-07 11:07:35 -0800946 "src/f32-igemm/gen/3x8-minmax-wasmsimd-arm-loadsplat.c",
947 "src/f32-igemm/gen/3x8-minmax-wasmsimd-arm-splat.c",
948 "src/f32-igemm/gen/3x8-minmax-wasmsimd-x86-loadsplat.c",
949 "src/f32-igemm/gen/3x8-minmax-wasmsimd-x86-splat.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700950 "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",
Frank Barchard0725b8d2020-12-07 11:07:35 -0800956 "src/f32-igemm/gen/4x8-minmax-wasmsimd-arm-loadsplat.c",
957 "src/f32-igemm/gen/4x8-minmax-wasmsimd-arm-splat.c",
958 "src/f32-igemm/gen/4x8-minmax-wasmsimd-x86-loadsplat.c",
959 "src/f32-igemm/gen/4x8-minmax-wasmsimd-x86-splat.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700960 "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",
Frank Barchard0725b8d2020-12-07 11:07:35 -0800964 "src/f32-igemm/gen/5x8-minmax-wasmsimd-arm-loadsplat.c",
965 "src/f32-igemm/gen/5x8-minmax-wasmsimd-arm-splat.c",
966 "src/f32-igemm/gen/5x8-minmax-wasmsimd-x86-loadsplat.c",
967 "src/f32-igemm/gen/5x8-minmax-wasmsimd-x86-splat.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700968 "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",
Frank Barchard0725b8d2020-12-07 11:07:35 -0800972 "src/f32-igemm/gen/6x8-minmax-wasmsimd-arm-loadsplat.c",
973 "src/f32-igemm/gen/6x8-minmax-wasmsimd-arm-splat.c",
974 "src/f32-igemm/gen/6x8-minmax-wasmsimd-x86-loadsplat.c",
975 "src/f32-igemm/gen/6x8-minmax-wasmsimd-x86-splat.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700976 "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",
Frank Barchard0725b8d2020-12-07 11:07:35 -0800984 "src/f32-ppmm/gen/4x8-minmax-wasmsimd-arm-splat.c",
985 "src/f32-ppmm/gen/4x8-minmax-wasmsimd-x86-splat.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 Dukhan08c4a432019-10-03 09:29:21 -07001336# ISA-specific micro-kernels
1337NEON_UKERNELS = [
Marat Dukhanef25c6d2020-07-24 00:59:40 -07001338 "src/f32-argmaxpool/4x-neon-c4.c",
1339 "src/f32-argmaxpool/9p8x-neon-c4.c",
1340 "src/f32-argmaxpool/9x-neon-c4.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001341 "src/f32-avgpool/9p8x-minmax-neon-c4.c",
1342 "src/f32-avgpool/9x-minmax-neon-c4.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001343 "src/f32-clamp/gen/neon-x4.c",
1344 "src/f32-clamp/gen/neon-x8.c",
Marat Dukhan56b10cd2020-05-18 09:35:49 -07001345 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neon-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001346 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neon-2x2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001347 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neon-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001348 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neon-2x2.c",
Marat Dukhan56b10cd2020-05-18 09:35:49 -07001349 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neon-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001350 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neon-2x2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001351 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neon-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001352 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neon-2x2.c",
Marat Dukhanc7634882020-12-07 15:11:12 -08001353 "src/f32-conv-hwc2chw/3x3s2p1c3x4-neon-2x2.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -07001354 "src/f32-dwconv/gen/up4x4-minmax-neon-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001355 "src/f32-dwconv/gen/up4x4-minmax-neon.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001356 "src/f32-dwconv/gen/up4x9-minmax-neon-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001357 "src/f32-dwconv/gen/up4x9-minmax-neon.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -07001358 "src/f32-dwconv/gen/up4x25-minmax-neon-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001359 "src/f32-dwconv/gen/up4x25-minmax-neon.c",
1360 "src/f32-dwconv/gen/up8x4-minmax-neon-acc2.c",
1361 "src/f32-dwconv/gen/up8x4-minmax-neon.c",
1362 "src/f32-dwconv/gen/up8x9-minmax-neon-acc2.c",
1363 "src/f32-dwconv/gen/up8x9-minmax-neon.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -07001364 "src/f32-dwconv/gen/up8x25-minmax-neon-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001365 "src/f32-dwconv/gen/up8x25-minmax-neon.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001366 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-1x4-acc2.c",
1367 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-1x4-acc3.c",
1368 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-1x4-acc4.c",
Marat Dukhanc581e482020-10-24 01:28:11 -07001369 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-1x4.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001370 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-2x4-acc2.c",
Marat Dukhanc581e482020-10-24 01:28:11 -07001371 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-2x4.c",
1372 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-3x4.c",
1373 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-4x4.c",
1374 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-5x4.c",
1375 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-6x4.c",
Marat Dukhan82f0c322020-10-25 19:17:35 -07001376 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neon-1x4-acc2.c",
1377 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neon-1x4-acc3.c",
1378 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neon-1x4-acc4.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001379 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neon-1x4.c",
Marat Dukhan82f0c322020-10-25 19:17:35 -07001380 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neon-2x4-acc2.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001381 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neon-2x4.c",
1382 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neon-3x4.c",
1383 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neon-4x4.c",
Marat Dukhan149f0ea2020-10-26 12:50:33 -07001384 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-1x4-acc2.c",
1385 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-1x4-acc3.c",
1386 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-1x4-acc4.c",
1387 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-1x4-acc5.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001388 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-1x4.c",
Marat Dukhan149f0ea2020-10-26 12:50:33 -07001389 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-2x4-acc2.c",
1390 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-2x4-acc3.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001391 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-2x4.c",
Marat Dukhan149f0ea2020-10-26 12:50:33 -07001392 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-3x4-acc2.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001393 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-3x4.c",
Marat Dukhan149f0ea2020-10-26 12:50:33 -07001394 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-4x4-acc2.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001395 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-4x4.c",
1396 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-5x4.c",
Marat Dukhan30d4b252020-10-29 16:33:22 -07001397 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neon-1x4-acc2.c",
1398 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neon-1x4-acc3.c",
1399 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neon-1x4-acc4.c",
1400 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neon-1x4-acc5.c",
1401 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neon-1x4.c",
1402 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neon-2x4-acc2.c",
1403 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neon-2x4-acc3.c",
1404 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neon-2x4.c",
Marat Dukhan30d4b252020-10-29 16:33:22 -07001405 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neon-3x4-acc2.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07001406 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neon-3x4.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -07001407 "src/f32-gavgpool-cw/neon-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001408 "src/f32-gavgpool/7p7x-minmax-neon-c4.c",
1409 "src/f32-gavgpool/7x-minmax-neon-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001410 "src/f32-gemm/gen-inc/1x8inc-minmax-neon-dup-ld64.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001411 "src/f32-gemm/gen-inc/1x8inc-minmax-neon-lane-ld64.c",
1412 "src/f32-gemm/gen-inc/1x8s4inc-minmax-neon.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001413 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-dup-ld64.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001414 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-dup-ld128.c",
1415 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-lane-ld64.c",
1416 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-lane-ld128.c",
1417 "src/f32-gemm/gen-inc/4x8s4inc-minmax-neon.c",
1418 "src/f32-gemm/gen-inc/5x8inc-minmax-neon-lane-ld64.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001419 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-dup-ld64.c",
1420 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-dup-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001421 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-lane-ld64.c",
1422 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-lane-ld128.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001423 "src/f32-gemm/gen-inc/6x8s4inc-minmax-neon.c",
1424 "src/f32-gemm/gen-inc/8x8s4inc-minmax-neon.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001425 "src/f32-gemm/gen/1x8-minmax-neon-dup-ld64.c",
1426 "src/f32-gemm/gen/1x8-minmax-neon-lane-ld64.c",
1427 "src/f32-gemm/gen/1x8s4-minmax-neon.c",
1428 "src/f32-gemm/gen/4x2-minmax-neon-lane-ld64.c",
1429 "src/f32-gemm/gen/4x8-minmax-neon-dup-ld64.c",
1430 "src/f32-gemm/gen/4x8-minmax-neon-dup-ld128.c",
1431 "src/f32-gemm/gen/4x8-minmax-neon-lane-ld64.c",
1432 "src/f32-gemm/gen/4x8-minmax-neon-lane-ld128.c",
1433 "src/f32-gemm/gen/4x8s4-minmax-neon.c",
1434 "src/f32-gemm/gen/5x8-minmax-neon-lane-ld64.c",
1435 "src/f32-gemm/gen/6x8-minmax-neon-dup-ld64.c",
1436 "src/f32-gemm/gen/6x8-minmax-neon-dup-ld128.c",
1437 "src/f32-gemm/gen/6x8-minmax-neon-lane-ld64.c",
1438 "src/f32-gemm/gen/6x8-minmax-neon-lane-ld128.c",
1439 "src/f32-gemm/gen/6x8s4-minmax-neon.c",
1440 "src/f32-gemm/gen/8x8s4-minmax-neon.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07001441 "src/f32-hswish/gen/hswish-neon-x4.c",
1442 "src/f32-hswish/gen/hswish-neon-x8.c",
1443 "src/f32-hswish/gen/hswish-neon-x16.c",
Marat Dukhan660fd192020-03-10 04:55:30 -07001444 "src/f32-ibilinear/gen/neon-c4.c",
1445 "src/f32-ibilinear/gen/neon-c8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001446 "src/f32-igemm/gen/1x8-minmax-neon-dup-ld64.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001447 "src/f32-igemm/gen/1x8-minmax-neon-lane-ld64.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001448 "src/f32-igemm/gen/1x8s4-minmax-neon.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001449 "src/f32-igemm/gen/4x2-minmax-neon-lane-ld64.c",
1450 "src/f32-igemm/gen/4x4-minmax-neon-lane-ld64.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001451 "src/f32-igemm/gen/4x8-minmax-neon-dup-ld64.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001452 "src/f32-igemm/gen/4x8-minmax-neon-dup-ld128.c",
1453 "src/f32-igemm/gen/4x8-minmax-neon-lane-ld64.c",
1454 "src/f32-igemm/gen/4x8-minmax-neon-lane-ld128.c",
1455 "src/f32-igemm/gen/4x8s4-minmax-neon.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001456 "src/f32-igemm/gen/6x8-minmax-neon-dup-ld64.c",
1457 "src/f32-igemm/gen/6x8-minmax-neon-dup-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001458 "src/f32-igemm/gen/6x8-minmax-neon-lane-ld64.c",
1459 "src/f32-igemm/gen/6x8-minmax-neon-lane-ld128.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001460 "src/f32-igemm/gen/6x8s4-minmax-neon.c",
1461 "src/f32-igemm/gen/8x8s4-minmax-neon.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001462 "src/f32-maxpool/9p8x-minmax-neon-c4.c",
1463 "src/f32-pavgpool/9p8x-minmax-neon-c4.c",
1464 "src/f32-pavgpool/9x-minmax-neon-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001465 "src/f32-ppmm/gen/4x8-minmax-neon.c",
1466 "src/f32-ppmm/gen/8x8-minmax-neon.c",
Frank Barcharda5316982020-07-23 13:19:28 -07001467 "src/f32-prelu/gen/neon-1x4.c",
1468 "src/f32-prelu/gen/neon-1x8.c",
1469 "src/f32-prelu/gen/neon-1x16.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001470 "src/f32-prelu/gen/neon-2x4.c",
1471 "src/f32-prelu/gen/neon-2x8.c",
Frank Barcharda5316982020-07-23 13:19:28 -07001472 "src/f32-prelu/gen/neon-2x16.c",
1473 "src/f32-prelu/gen/neon-4x4.c",
1474 "src/f32-prelu/gen/neon-4x8.c",
1475 "src/f32-prelu/gen/neon-4x16.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001476 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x4.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001477 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x8-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001478 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x8.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001479 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12-acc2.c",
1480 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001481 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001482 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16-acc2.c",
1483 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001484 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001485 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20-acc2.c",
1486 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001487 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20.c",
1488 "src/f32-raddstoreexpminusmax/gen/neon-p5-x4.c",
1489 "src/f32-raddstoreexpminusmax/gen/neon-p5-x8-acc2.c",
1490 "src/f32-raddstoreexpminusmax/gen/neon-p5-x8.c",
1491 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12-acc2.c",
1492 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12-acc3.c",
1493 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12.c",
1494 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16-acc2.c",
1495 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16-acc4.c",
1496 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16.c",
1497 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20-acc2.c",
1498 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20-acc5.c",
1499 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07001500 "src/f32-relu/gen/neon-x4.c",
1501 "src/f32-relu/gen/neon-x8.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001502 "src/f32-rmax/neon.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -08001503 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x4.c",
1504 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x8.c",
1505 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x12.c",
1506 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x16.c",
1507 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x20.c",
1508 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x24.c",
1509 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x4.c",
1510 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x8.c",
1511 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x12.c",
1512 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x16.c",
1513 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x20.c",
1514 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x24.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001515 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x4.c",
1516 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x8.c",
1517 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x12.c",
1518 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x16.c",
1519 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x20.c",
1520 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x24.c",
Marat Dukhan5b86c432020-12-06 19:15:03 -08001521 "src/f32-spmm/gen/4x1-minmax-neon-pipelined.c",
1522 "src/f32-spmm/gen/4x1-minmax-neon-x2.c",
1523 "src/f32-spmm/gen/4x1-minmax-neon.c",
1524 "src/f32-spmm/gen/8x1-minmax-neon-pipelined.c",
1525 "src/f32-spmm/gen/8x1-minmax-neon-x2.c",
1526 "src/f32-spmm/gen/8x1-minmax-neon.c",
1527 "src/f32-spmm/gen/12x1-minmax-neon.c",
1528 "src/f32-spmm/gen/16x1-minmax-neon-pipelined.c",
1529 "src/f32-spmm/gen/16x1-minmax-neon-x2.c",
1530 "src/f32-spmm/gen/16x1-minmax-neon.c",
1531 "src/f32-spmm/gen/32x1-minmax-neon-pipelined.c",
1532 "src/f32-spmm/gen/32x1-minmax-neon-x2.c",
1533 "src/f32-spmm/gen/32x1-minmax-neon.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001534 "src/f32-vbinary/gen/vadd-minmax-neon-x4.c",
1535 "src/f32-vbinary/gen/vadd-minmax-neon-x8.c",
1536 "src/f32-vbinary/gen/vaddc-minmax-neon-x4.c",
1537 "src/f32-vbinary/gen/vaddc-minmax-neon-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -08001538 "src/f32-vbinary/gen/vmax-neon-x4.c",
1539 "src/f32-vbinary/gen/vmax-neon-x8.c",
1540 "src/f32-vbinary/gen/vmaxc-neon-x4.c",
1541 "src/f32-vbinary/gen/vmaxc-neon-x8.c",
1542 "src/f32-vbinary/gen/vmin-neon-x4.c",
1543 "src/f32-vbinary/gen/vmin-neon-x8.c",
1544 "src/f32-vbinary/gen/vminc-neon-x4.c",
1545 "src/f32-vbinary/gen/vminc-neon-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001546 "src/f32-vbinary/gen/vmul-minmax-neon-x4.c",
1547 "src/f32-vbinary/gen/vmul-minmax-neon-x8.c",
1548 "src/f32-vbinary/gen/vmulc-minmax-neon-x4.c",
1549 "src/f32-vbinary/gen/vmulc-minmax-neon-x8.c",
1550 "src/f32-vbinary/gen/vrsubc-minmax-neon-x4.c",
1551 "src/f32-vbinary/gen/vrsubc-minmax-neon-x8.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07001552 "src/f32-vbinary/gen/vsqrdiff-neon-x4.c",
1553 "src/f32-vbinary/gen/vsqrdiff-neon-x8.c",
1554 "src/f32-vbinary/gen/vsqrdiffc-neon-x4.c",
1555 "src/f32-vbinary/gen/vsqrdiffc-neon-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001556 "src/f32-vbinary/gen/vsub-minmax-neon-x4.c",
1557 "src/f32-vbinary/gen/vsub-minmax-neon-x8.c",
1558 "src/f32-vbinary/gen/vsubc-minmax-neon-x4.c",
1559 "src/f32-vbinary/gen/vsubc-minmax-neon-x8.c",
Marat Dukhaned6baaf2020-12-01 15:07:08 -08001560 "src/f32-velu/gen/velu-neon-rr2-lut16-p3-x4.c",
1561 "src/f32-velu/gen/velu-neon-rr2-lut16-p3-x8.c",
1562 "src/f32-velu/gen/velu-neon-rr2-lut16-p3-x12.c",
1563 "src/f32-velu/gen/velu-neon-rr2-lut16-p3-x16.c",
1564 "src/f32-velu/gen/velu-neon-rr2-lut16-p3-x20.c",
1565 "src/f32-velu/gen/velu-neon-rr2-lut16-p3-x24.c",
1566 "src/f32-velu/gen/velu-neon-rr2-p6-x4.c",
1567 "src/f32-velu/gen/velu-neon-rr2-p6-x8.c",
1568 "src/f32-velu/gen/velu-neon-rr2-p6-x12.c",
1569 "src/f32-velu/gen/velu-neon-rr2-p6-x16.c",
1570 "src/f32-velu/gen/velu-neon-rr2-p6-x20.c",
1571 "src/f32-velu/gen/velu-neon-rr2-p6-x24.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07001572 "src/f32-vlrelu/gen/vlrelu-neon-x4.c",
1573 "src/f32-vlrelu/gen/vlrelu-neon-x8.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001574 "src/f32-vmulcaddc/gen/c4-minmax-neon-2x.c",
1575 "src/f32-vmulcaddc/gen/c8-minmax-neon-2x.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07001576 "src/f32-vrnd/gen/vrndd-neon-x4.c",
1577 "src/f32-vrnd/gen/vrndd-neon-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001578 "src/f32-vrnd/gen/vrndne-neon-x4.c",
1579 "src/f32-vrnd/gen/vrndne-neon-x8.c",
1580 "src/f32-vrnd/gen/vrndu-neon-x4.c",
1581 "src/f32-vrnd/gen/vrndu-neon-x8.c",
1582 "src/f32-vrnd/gen/vrndz-neon-x4.c",
1583 "src/f32-vrnd/gen/vrndz-neon-x8.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07001584 "src/f32-vunary/gen/vabs-neon-x4.c",
1585 "src/f32-vunary/gen/vabs-neon-x8.c",
1586 "src/f32-vunary/gen/vneg-neon-x4.c",
1587 "src/f32-vunary/gen/vneg-neon-x8.c",
1588 "src/f32-vunary/gen/vsqr-neon-x4.c",
1589 "src/f32-vunary/gen/vsqr-neon-x8.c",
Marat Dukhande390d42020-11-29 19:32:18 -08001590 "src/math/expm1minus-neon-rr2-lut16-p3.c",
1591 "src/math/expm1minus-neon-rr2-p6.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001592 "src/math/roundd-neon-addsub.c",
1593 "src/math/roundd-neon-cvt.c",
1594 "src/math/roundne-neon-addsub.c",
1595 "src/math/roundu-neon-addsub.c",
1596 "src/math/roundu-neon-cvt.c",
1597 "src/math/roundz-neon-addsub.c",
1598 "src/math/roundz-neon-cvt.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001599 "src/math/sigmoid-neon-rr2-lut64-p2-nr2recps.c",
1600 "src/math/sigmoid-neon-rr2-lut2048-p1-nr2recps.c",
1601 "src/math/sigmoid-neon-rr2-p5-nr2recps.c",
1602 "src/math/sqrt-neon-nr1rsqrts.c",
1603 "src/math/sqrt-neon-nr2rsqrts.c",
1604 "src/math/sqrt-neon-nr3rsqrts.c",
Marat Dukhan023bcf92020-08-10 12:40:50 -07001605 "src/qs8-dwconv/gen/up8x9-minmax-neon-mul16.c",
1606 "src/qs8-dwconv/gen/up16x9-minmax-neon-mul16.c",
1607 "src/qs8-dwconv/gen/up24x9-minmax-neon-mul16.c",
1608 "src/qs8-dwconv/gen/up32x9-minmax-neon-mul16.c",
Marat Dukhan281262d2020-08-10 13:23:21 -07001609 "src/qs8-gavgpool/gen/7p7x-minmax-neon-c8-acc2.c",
1610 "src/qs8-gavgpool/gen/7p7x-minmax-neon-c16-acc2.c",
1611 "src/qs8-gavgpool/gen/7p7x-minmax-neon-c24-acc2.c",
1612 "src/qs8-gavgpool/gen/7p7x-minmax-neon-c32-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001613 "src/qs8-gavgpool/gen/7x-minmax-neon-c8-acc2.c",
1614 "src/qs8-gavgpool/gen/7x-minmax-neon-c16-acc2.c",
1615 "src/qs8-gavgpool/gen/7x-minmax-neon-c24-acc2.c",
1616 "src/qs8-gavgpool/gen/7x-minmax-neon-c32-acc2.c",
Marat Dukhan40bbafe2020-08-04 02:04:22 -07001617 "src/qs8-gemm/gen/1x8-minmax-neon-mlal-lane.c",
Marat Dukhan40bbafe2020-08-04 02:04:22 -07001618 "src/qs8-gemm/gen/1x16-minmax-neon-mlal-lane.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001619 "src/qs8-gemm/gen/2x8-minmax-neon-mlal-lane.c",
Marat Dukhan40bbafe2020-08-04 02:04:22 -07001620 "src/qs8-gemm/gen/2x16-minmax-neon-mlal-lane.c",
1621 "src/qs8-igemm/gen/1x8-minmax-neon-mlal-lane.c",
Marat Dukhan40bbafe2020-08-04 02:04:22 -07001622 "src/qs8-igemm/gen/1x16-minmax-neon-mlal-lane.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001623 "src/qs8-igemm/gen/2x8-minmax-neon-mlal-lane.c",
Marat Dukhan40bbafe2020-08-04 02:04:22 -07001624 "src/qs8-igemm/gen/2x16-minmax-neon-mlal-lane.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -07001625 "src/qs8-requantization/fp32-neon.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001626 "src/qs8-requantization/precise-neon.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -07001627 "src/qs8-requantization/q31-neon.c",
Marat Dukhanba7b2792020-09-02 14:26:45 -07001628 "src/qs8-vadd/gen/minmax-neon-ld64-x8.c",
1629 "src/qs8-vadd/gen/minmax-neon-ld64-x16.c",
1630 "src/qs8-vadd/gen/minmax-neon-ld64-x24.c",
1631 "src/qs8-vadd/gen/minmax-neon-ld64-x32.c",
1632 "src/qs8-vaddc/gen/minmax-neon-ld64-x8.c",
1633 "src/qs8-vaddc/gen/minmax-neon-ld64-x16.c",
1634 "src/qs8-vaddc/gen/minmax-neon-ld64-x24.c",
1635 "src/qs8-vaddc/gen/minmax-neon-ld64-x32.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -07001636 "src/qu8-avgpool/9p8x-minmax-neon-c8.c",
1637 "src/qu8-avgpool/9x-minmax-neon-c8.c",
1638 "src/qu8-dwconv/up8x9-minmax-neon.c",
1639 "src/qu8-gavgpool/7p7x-minmax-neon-c8.c",
1640 "src/qu8-gavgpool/7x-minmax-neon-c8.c",
1641 "src/qu8-gemm/4x8-minmax-neon.c",
1642 "src/qu8-gemm/8x8-minmax-neon.c",
1643 "src/qu8-igemm/4x8-minmax-neon.c",
1644 "src/qu8-igemm/8x8-minmax-neon.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07001645 "src/qu8-requantization/fp32-neon.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001646 "src/qu8-requantization/precise-neon.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07001647 "src/qu8-requantization/q31-neon.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -07001648 "src/qu8-vadd/minmax-neon.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001649 "src/u8-clamp/neon-x64.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001650 "src/u8-maxpool/9p8x-minmax-neon-c16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001651 "src/u8-rmax/neon.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001652 "src/x8-zip/x2-neon.c",
1653 "src/x8-zip/x3-neon.c",
1654 "src/x8-zip/x4-neon.c",
1655 "src/x8-zip/xm-neon.c",
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -07001656 "src/x32-fill/neon.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001657 "src/x32-packx/x4-neon-st4.c",
Marat Dukhan63523d42020-05-22 17:07:33 -07001658 "src/x32-pad/neon.c",
Marat Dukhan57dccd82020-04-14 00:53:10 -07001659 "src/x32-unpool/neon.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001660 "src/x32-zip/x2-neon.c",
1661 "src/x32-zip/x3-neon.c",
1662 "src/x32-zip/x4-neon.c",
1663 "src/x32-zip/xm-neon.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001664]
1665
1666NEONFMA_UKERNELS = [
Frank Barchard04336c12020-10-22 16:48:55 -07001667 "src/f32-dwconv/gen/up4x4-minmax-neonfma-acc2.c",
1668 "src/f32-dwconv/gen/up4x4-minmax-neonfma.c",
1669 "src/f32-dwconv/gen/up4x9-minmax-neonfma-acc2.c",
1670 "src/f32-dwconv/gen/up4x9-minmax-neonfma.c",
1671 "src/f32-dwconv/gen/up4x25-minmax-neonfma-acc2.c",
1672 "src/f32-dwconv/gen/up4x25-minmax-neonfma.c",
1673 "src/f32-dwconv/gen/up8x4-minmax-neonfma-acc2.c",
1674 "src/f32-dwconv/gen/up8x4-minmax-neonfma.c",
1675 "src/f32-dwconv/gen/up8x9-minmax-neonfma-acc2.c",
1676 "src/f32-dwconv/gen/up8x9-minmax-neonfma.c",
1677 "src/f32-dwconv/gen/up8x25-minmax-neonfma-acc2.c",
1678 "src/f32-dwconv/gen/up8x25-minmax-neonfma.c",
1679 "src/f32-gemm/gen-inc/1x8inc-minmax-neonfma-dup-ld64.c",
1680 "src/f32-gemm/gen-inc/1x8s4inc-minmax-neonfma.c",
1681 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-dup-ld64.c",
1682 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-dup-ld128.c",
1683 "src/f32-gemm/gen-inc/4x8s4inc-minmax-neonfma.c",
1684 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-dup-ld64.c",
1685 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-dup-ld128.c",
1686 "src/f32-gemm/gen-inc/6x8s4inc-minmax-neonfma.c",
1687 "src/f32-gemm/gen-inc/8x8s4inc-minmax-neonfma.c",
1688 "src/f32-gemm/gen/1x8-minmax-neonfma-dup-ld64.c",
1689 "src/f32-gemm/gen/1x8s4-minmax-neonfma.c",
1690 "src/f32-gemm/gen/4x8-minmax-neonfma-dup-ld64.c",
1691 "src/f32-gemm/gen/4x8-minmax-neonfma-dup-ld128.c",
1692 "src/f32-gemm/gen/4x8s4-minmax-neonfma.c",
1693 "src/f32-gemm/gen/6x8-minmax-neonfma-dup-ld64.c",
1694 "src/f32-gemm/gen/6x8-minmax-neonfma-dup-ld128.c",
1695 "src/f32-gemm/gen/6x8s4-minmax-neonfma.c",
1696 "src/f32-gemm/gen/8x8s4-minmax-neonfma.c",
Marat Dukhan660fd192020-03-10 04:55:30 -07001697 "src/f32-ibilinear/gen/neonfma-c4.c",
1698 "src/f32-ibilinear/gen/neonfma-c8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001699 "src/f32-igemm/gen/1x8-minmax-neonfma-dup-ld64.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001700 "src/f32-igemm/gen/1x8s4-minmax-neonfma.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001701 "src/f32-igemm/gen/4x8-minmax-neonfma-dup-ld64.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001702 "src/f32-igemm/gen/4x8-minmax-neonfma-dup-ld128.c",
1703 "src/f32-igemm/gen/4x8s4-minmax-neonfma.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001704 "src/f32-igemm/gen/6x8-minmax-neonfma-dup-ld64.c",
1705 "src/f32-igemm/gen/6x8-minmax-neonfma-dup-ld128.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001706 "src/f32-igemm/gen/6x8s4-minmax-neonfma.c",
1707 "src/f32-igemm/gen/8x8s4-minmax-neonfma.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001708 "src/f32-ppmm/gen/4x8-minmax-neonfma.c",
1709 "src/f32-ppmm/gen/8x8-minmax-neonfma.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001710 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x4.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001711 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x8-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001712 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x8.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001713 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12-acc2.c",
1714 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001715 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001716 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16-acc2.c",
1717 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001718 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001719 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20-acc2.c",
1720 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001721 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20.c",
1722 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x4.c",
1723 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x8-acc2.c",
1724 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x8.c",
1725 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12-acc2.c",
1726 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12-acc3.c",
1727 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12.c",
1728 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16-acc2.c",
1729 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16-acc4.c",
1730 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16.c",
1731 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20-acc2.c",
1732 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20-acc5.c",
1733 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -08001734 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x4.c",
1735 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x8.c",
1736 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x12.c",
1737 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x16.c",
1738 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x20.c",
1739 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x24.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001740 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x4.c",
1741 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x8.c",
1742 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x12.c",
1743 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x16.c",
1744 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x20.c",
1745 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x24.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -08001746 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x4.c",
1747 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x8.c",
1748 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x12.c",
1749 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x16.c",
1750 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x20.c",
1751 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x24.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -08001752 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x4.c",
1753 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x8.c",
1754 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x12.c",
1755 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x16.c",
1756 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x20.c",
1757 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x24.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001758 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x4.c",
1759 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x8.c",
1760 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x12.c",
1761 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x16.c",
1762 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x20.c",
1763 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x24.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -08001764 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x4.c",
1765 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x8.c",
1766 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x12.c",
1767 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x16.c",
1768 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x20.c",
1769 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x24.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001770 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x4.c",
1771 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x8.c",
1772 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x12.c",
1773 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x16.c",
1774 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x20.c",
1775 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x24.c",
1776 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x4.c",
1777 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x8.c",
1778 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x12.c",
1779 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x16.c",
1780 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x20.c",
1781 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x24.c",
1782 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x4.c",
1783 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x8.c",
1784 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x12.c",
1785 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x16.c",
1786 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x20.c",
1787 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x24.c",
Marat Dukhan2fa7a0c2020-12-06 19:09:02 -08001788 "src/f32-spmm/gen/4x1-minmax-neonfma-pipelined.c",
1789 "src/f32-spmm/gen/4x1-minmax-neonfma-x2.c",
1790 "src/f32-spmm/gen/4x1-minmax-neonfma.c",
1791 "src/f32-spmm/gen/8x1-minmax-neonfma-pipelined.c",
1792 "src/f32-spmm/gen/8x1-minmax-neonfma-x2.c",
1793 "src/f32-spmm/gen/8x1-minmax-neonfma.c",
1794 "src/f32-spmm/gen/12x1-minmax-neonfma.c",
1795 "src/f32-spmm/gen/16x1-minmax-neonfma-pipelined.c",
1796 "src/f32-spmm/gen/16x1-minmax-neonfma-x2.c",
1797 "src/f32-spmm/gen/16x1-minmax-neonfma.c",
1798 "src/f32-spmm/gen/32x1-minmax-neonfma-pipelined.c",
1799 "src/f32-spmm/gen/32x1-minmax-neonfma-x2.c",
1800 "src/f32-spmm/gen/32x1-minmax-neonfma.c",
Marat Dukhaned6baaf2020-12-01 15:07:08 -08001801 "src/f32-velu/gen/velu-neonfma-rr1-lut16-p3-x4.c",
1802 "src/f32-velu/gen/velu-neonfma-rr1-lut16-p3-x8.c",
1803 "src/f32-velu/gen/velu-neonfma-rr1-lut16-p3-x12.c",
1804 "src/f32-velu/gen/velu-neonfma-rr1-lut16-p3-x16.c",
1805 "src/f32-velu/gen/velu-neonfma-rr1-lut16-p3-x20.c",
1806 "src/f32-velu/gen/velu-neonfma-rr1-lut16-p3-x24.c",
1807 "src/f32-velu/gen/velu-neonfma-rr1-p6-x4.c",
1808 "src/f32-velu/gen/velu-neonfma-rr1-p6-x8.c",
1809 "src/f32-velu/gen/velu-neonfma-rr1-p6-x12.c",
1810 "src/f32-velu/gen/velu-neonfma-rr1-p6-x16.c",
1811 "src/f32-velu/gen/velu-neonfma-rr1-p6-x20.c",
1812 "src/f32-velu/gen/velu-neonfma-rr1-p6-x24.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001813 "src/f32-vmulcaddc/gen/c4-minmax-neonfma-2x.c",
1814 "src/f32-vmulcaddc/gen/c8-minmax-neonfma-2x.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07001815 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x4.c",
1816 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x8.c",
1817 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x12.c",
1818 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x16.c",
1819 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x20.c",
1820 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x24.c",
1821 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x28.c",
1822 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x32.c",
1823 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x36.c",
1824 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x40.c",
1825 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x4.c",
1826 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x8.c",
1827 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x12.c",
1828 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x16.c",
1829 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x20.c",
1830 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x24.c",
1831 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x28.c",
1832 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x32.c",
1833 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x36.c",
1834 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x40.c",
Marat Dukhanb7633f22020-11-20 16:34:56 -08001835 "src/math/exp-neonfma-rr2-lut64-p2.c",
1836 "src/math/exp-neonfma-rr2-p5.c",
Frank Barcharde7223ee2020-12-04 19:04:01 -08001837 "src/math/expm1minus-neonfma-rr1-lut16-p3.c",
1838 "src/math/expm1minus-neonfma-rr1-p6.c",
Marat Dukhan9dd119a2020-11-20 18:20:04 -08001839 "src/math/expminus-neonfma-rr2-lut64-p2.c",
1840 "src/math/expminus-neonfma-rr2-lut2048-p1.c",
1841 "src/math/expminus-neonfma-rr2-p5.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001842 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr1recps1fma.c",
1843 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr2fma.c",
1844 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr2recps.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001845 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr1recps1fma.c",
1846 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr2fma.c",
1847 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr2recps.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001848 "src/math/sigmoid-neonfma-rr1-p5-nr1recps1fma.c",
1849 "src/math/sigmoid-neonfma-rr1-p5-nr2fma.c",
1850 "src/math/sigmoid-neonfma-rr1-p5-nr2recps.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001851 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr1recps1fma.c",
1852 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr2fma.c",
1853 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr2recps.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001854 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr1recps1fma.c",
1855 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr2fma.c",
1856 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr2recps.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001857 "src/math/sigmoid-neonfma-rr2-p5-nr1recps1fma.c",
1858 "src/math/sigmoid-neonfma-rr2-p5-nr2fma.c",
1859 "src/math/sigmoid-neonfma-rr2-p5-nr2recps.c",
Marat Dukhan84000762020-06-29 18:38:43 -07001860 "src/math/sqrt-neonfma-nr1fma.c",
Marat Dukhan84000762020-06-29 18:38:43 -07001861 "src/math/sqrt-neonfma-nr1rsqrts1fma1adj.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001862 "src/math/sqrt-neonfma-nr2fma.c",
1863 "src/math/sqrt-neonfma-nr2fma1adj.c",
1864 "src/math/sqrt-neonfma-nr3fma.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001865]
1866
1867AARCH64_NEONFMA_UKERNELS = [
Marat Dukhan56b10cd2020-05-18 09:35:49 -07001868 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neonfma-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001869 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neonfma-2x2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001870 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neonfma-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001871 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neonfma-2x2.c",
Marat Dukhan56b10cd2020-05-18 09:35:49 -07001872 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neonfma-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001873 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neonfma-2x2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001874 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neonfma-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001875 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neonfma-2x2.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -07001876 "src/f32-conv-hwc2chw/3x3s2p1c3x4-neonfma-2x2.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001877 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-1x4-acc2.c",
1878 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-1x4-acc3.c",
1879 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-1x4-acc4.c",
Marat Dukhan1268a242020-10-24 00:36:32 -07001880 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-1x4.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001881 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-2x4-acc2.c",
Marat Dukhan1268a242020-10-24 00:36:32 -07001882 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-2x4.c",
1883 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-3x4.c",
1884 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-4x4.c",
1885 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-5x4.c",
1886 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-6x4.c",
Marat Dukhan82f0c322020-10-25 19:17:35 -07001887 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neonfma-1x4-acc2.c",
1888 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neonfma-1x4-acc3.c",
1889 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neonfma-1x4-acc4.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001890 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neonfma-1x4.c",
Marat Dukhan82f0c322020-10-25 19:17:35 -07001891 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neonfma-2x4-acc2.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001892 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neonfma-2x4.c",
1893 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neonfma-3x4.c",
1894 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neonfma-4x4.c",
Marat Dukhan149f0ea2020-10-26 12:50:33 -07001895 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-1x4-acc2.c",
1896 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-1x4-acc3.c",
1897 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-1x4-acc4.c",
1898 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-1x4-acc5.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001899 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-1x4.c",
Marat Dukhan149f0ea2020-10-26 12:50:33 -07001900 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-2x4-acc2.c",
1901 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-2x4-acc3.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001902 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-2x4.c",
Marat Dukhan149f0ea2020-10-26 12:50:33 -07001903 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-3x4-acc2.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001904 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-3x4.c",
Marat Dukhan149f0ea2020-10-26 12:50:33 -07001905 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-4x4-acc2.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001906 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-4x4.c",
1907 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-5x4.c",
Marat Dukhan30d4b252020-10-29 16:33:22 -07001908 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neonfma-1x4-acc2.c",
1909 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neonfma-1x4-acc3.c",
1910 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neonfma-1x4-acc4.c",
1911 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neonfma-1x4-acc5.c",
1912 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neonfma-1x4.c",
1913 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neonfma-2x4-acc2.c",
1914 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neonfma-2x4-acc3.c",
1915 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neonfma-2x4.c",
Marat Dukhan30d4b252020-10-29 16:33:22 -07001916 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neonfma-3x4-acc2.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07001917 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neonfma-3x4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001918 "src/f32-gemm/gen-inc/1x8inc-minmax-neonfma-lane-ld64.c",
1919 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-lane-ld64.c",
1920 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-lane-ld128.c",
1921 "src/f32-gemm/gen-inc/5x8inc-minmax-neonfma-lane-ld64.c",
1922 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-lane-ld64.c",
1923 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-lane-ld128.c",
1924 "src/f32-gemm/gen/1x8-minmax-neonfma-lane-ld64.c",
1925 "src/f32-gemm/gen/4x2-minmax-neonfma-lane-ld64.c",
1926 "src/f32-gemm/gen/4x8-minmax-neonfma-lane-ld64.c",
1927 "src/f32-gemm/gen/4x8-minmax-neonfma-lane-ld128.c",
1928 "src/f32-gemm/gen/5x8-minmax-neonfma-lane-ld64.c",
1929 "src/f32-gemm/gen/6x8-minmax-neonfma-lane-ld64.c",
1930 "src/f32-gemm/gen/6x8-minmax-neonfma-lane-ld128.c",
1931 "src/f32-igemm/gen/1x8-minmax-neonfma-lane-ld64.c",
1932 "src/f32-igemm/gen/4x2-minmax-neonfma-lane-ld64.c",
1933 "src/f32-igemm/gen/4x4-minmax-neonfma-lane-ld64.c",
1934 "src/f32-igemm/gen/4x8-minmax-neonfma-lane-ld64.c",
1935 "src/f32-igemm/gen/4x8-minmax-neonfma-lane-ld128.c",
1936 "src/f32-igemm/gen/6x8-minmax-neonfma-lane-ld64.c",
1937 "src/f32-igemm/gen/6x8-minmax-neonfma-lane-ld128.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -08001938 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x4.c",
1939 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x8.c",
1940 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x12.c",
1941 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x16.c",
1942 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x20.c",
1943 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x24.c",
1944 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x4.c",
1945 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x8.c",
1946 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x12.c",
1947 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x16.c",
1948 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x20.c",
1949 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x24.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001950 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x4.c",
1951 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x8.c",
1952 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x12.c",
1953 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x16.c",
1954 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x20.c",
1955 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x24.c",
Marat Dukhan355ab432020-04-09 19:01:52 -07001956 "src/f32-spmm/gen/4x2-minmax-neonfma.c",
1957 "src/f32-spmm/gen/4x4-minmax-neonfma.c",
Marat Dukhan355ab432020-04-09 19:01:52 -07001958 "src/f32-spmm/gen/8x2-minmax-neonfma.c",
1959 "src/f32-spmm/gen/8x4-minmax-neonfma.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001960 "src/f32-spmm/gen/12x2-minmax-neonfma.c",
1961 "src/f32-spmm/gen/12x4-minmax-neonfma.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001962 "src/f32-spmm/gen/16x2-minmax-neonfma.c",
1963 "src/f32-spmm/gen/16x4-minmax-neonfma.c",
Frank Barchard846c0c62020-10-26 15:01:39 -07001964 "src/f32-spmm/gen/32x2-minmax-neonfma.c",
1965 "src/f32-spmm/gen/32x4-minmax-neonfma.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001966 "src/f32-vbinary/gen/vdiv-minmax-neon-x4.c",
1967 "src/f32-vbinary/gen/vdiv-minmax-neon-x8.c",
1968 "src/f32-vbinary/gen/vdivc-minmax-neon-x4.c",
1969 "src/f32-vbinary/gen/vdivc-minmax-neon-x8.c",
1970 "src/f32-vbinary/gen/vrdivc-minmax-neon-x4.c",
1971 "src/f32-vbinary/gen/vrdivc-minmax-neon-x8.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07001972 "src/f32-vsqrt/gen/neon-sqrt-x4.c",
1973 "src/f32-vsqrt/gen/neon-sqrt-x8.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001974 "src/math/sigmoid-neonfma-rr1-lut64-p2-div.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001975 "src/math/sigmoid-neonfma-rr1-lut2048-p1-div.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001976 "src/math/sigmoid-neonfma-rr1-p5-div.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001977 "src/math/sigmoid-neonfma-rr2-lut64-p2-div.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001978 "src/math/sigmoid-neonfma-rr2-lut2048-p1-div.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001979 "src/math/sigmoid-neonfma-rr2-p5-div.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001980]
1981
Marat Dukhan8853b822020-05-07 12:19:01 -07001982NEONV8_UKERNELS = [
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07001983 "src/f32-vrnd/gen/vrndd-neonv8-x4.c",
1984 "src/f32-vrnd/gen/vrndd-neonv8-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001985 "src/f32-vrnd/gen/vrndne-neonv8-x4.c",
1986 "src/f32-vrnd/gen/vrndne-neonv8-x8.c",
1987 "src/f32-vrnd/gen/vrndu-neonv8-x4.c",
1988 "src/f32-vrnd/gen/vrndu-neonv8-x8.c",
1989 "src/f32-vrnd/gen/vrndz-neonv8-x4.c",
1990 "src/f32-vrnd/gen/vrndz-neonv8-x8.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001991 "src/math/roundd-neonv8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001992 "src/math/roundne-neonv8.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001993 "src/math/roundu-neonv8.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07001994 "src/math/roundz-neonv8.c",
Marat Dukhan8853b822020-05-07 12:19:01 -07001995]
1996
Marat Dukhan08c4a432019-10-03 09:29:21 -07001997AARCH64_NEONFP16ARITH_UKERNELS = [
Frank Barchardb1966592020-05-12 13:47:06 -07001998 "src/f16-clamp/gen/neonfp16arith-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001999 "src/f16-clamp/gen/neonfp16arith-x16.c",
Frank Barchard5a599a62020-06-04 20:12:44 -07002000 "src/f16-dwconv/gen/up8x4-minmax-neonfp16arith-acc2.c",
2001 "src/f16-dwconv/gen/up8x4-minmax-neonfp16arith.c",
2002 "src/f16-dwconv/gen/up8x9-minmax-neonfp16arith-acc2.c",
2003 "src/f16-dwconv/gen/up8x9-minmax-neonfp16arith.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002004 "src/f16-dwconv/gen/up8x25-minmax-neonfp16arith-acc2.c",
2005 "src/f16-dwconv/gen/up8x25-minmax-neonfp16arith.c",
2006 "src/f16-dwconv/gen/up16x4-minmax-neonfp16arith-acc2.c",
2007 "src/f16-dwconv/gen/up16x4-minmax-neonfp16arith.c",
2008 "src/f16-dwconv/gen/up16x9-minmax-neonfp16arith-acc2.c",
2009 "src/f16-dwconv/gen/up16x9-minmax-neonfp16arith.c",
2010 "src/f16-dwconv/gen/up16x25-minmax-neonfp16arith-acc2.c",
2011 "src/f16-dwconv/gen/up16x25-minmax-neonfp16arith.c",
Frank Barchard0bb49a72020-06-04 11:35:11 -07002012 "src/f16-gavgpool/7p7x-minmax-neonfp16arith-c8.c",
2013 "src/f16-gavgpool/7x-minmax-neonfp16arith-c8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002014 "src/f16-gemm/gen-inc/1x8inc-minmax-neonfp16arith-ld64.c",
2015 "src/f16-gemm/gen-inc/1x16inc-minmax-neonfp16arith-ld64.c",
2016 "src/f16-gemm/gen-inc/4x8inc-minmax-neonfp16arith-ld64.c",
2017 "src/f16-gemm/gen-inc/4x16inc-minmax-neonfp16arith-ld64.c",
2018 "src/f16-gemm/gen-inc/6x8inc-minmax-neonfp16arith-ld64.c",
2019 "src/f16-gemm/gen-inc/6x16inc-minmax-neonfp16arith-ld64.c",
2020 "src/f16-gemm/gen-inc/8x8inc-minmax-neonfp16arith-ld64.c",
2021 "src/f16-gemm/gen-inc/8x16inc-minmax-neonfp16arith-ld64.c",
2022 "src/f16-gemm/gen/1x8-minmax-neonfp16arith-ld64.c",
2023 "src/f16-gemm/gen/1x16-minmax-neonfp16arith-ld64.c",
2024 "src/f16-gemm/gen/4x8-minmax-neonfp16arith-ld64.c",
2025 "src/f16-gemm/gen/4x16-minmax-neonfp16arith-ld64.c",
2026 "src/f16-gemm/gen/6x8-minmax-neonfp16arith-ld64.c",
2027 "src/f16-gemm/gen/6x16-minmax-neonfp16arith-ld64.c",
2028 "src/f16-gemm/gen/8x8-minmax-neonfp16arith-ld64.c",
2029 "src/f16-gemm/gen/8x16-minmax-neonfp16arith-ld64.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07002030 "src/f16-hswish/gen/hswish-neonfp16arith-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002031 "src/f16-hswish/gen/hswish-neonfp16arith-x16.c",
2032 "src/f16-igemm/gen/1x8-minmax-neonfp16arith-ld64.c",
2033 "src/f16-igemm/gen/1x16-minmax-neonfp16arith-ld64.c",
2034 "src/f16-igemm/gen/4x8-minmax-neonfp16arith-ld64.c",
2035 "src/f16-igemm/gen/4x16-minmax-neonfp16arith-ld64.c",
2036 "src/f16-igemm/gen/6x8-minmax-neonfp16arith-ld64.c",
2037 "src/f16-igemm/gen/6x16-minmax-neonfp16arith-ld64.c",
2038 "src/f16-igemm/gen/8x8-minmax-neonfp16arith-ld64.c",
2039 "src/f16-igemm/gen/8x16-minmax-neonfp16arith-ld64.c",
Frank Barchardb1966592020-05-12 13:47:06 -07002040 "src/f16-prelu/gen/neonfp16arith-2x8.c",
Frank Barcharda5316982020-07-23 13:19:28 -07002041 "src/f16-prelu/gen/neonfp16arith-2x16.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07002042 "src/f16-relu/gen/neonfp16arith-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002043 "src/f16-relu/gen/neonfp16arith-x16.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07002044 "src/f16-spmm/gen/8x1-minmax-neonfp16arith-x2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002045 "src/f16-spmm/gen/8x1-minmax-neonfp16arith.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07002046 "src/f16-spmm/gen/16x1-minmax-neonfp16arith-x2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002047 "src/f16-spmm/gen/16x1-minmax-neonfp16arith.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07002048 "src/f16-spmm/gen/24x1-minmax-neonfp16arith-x2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002049 "src/f16-spmm/gen/24x1-minmax-neonfp16arith.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07002050 "src/f16-spmm/gen/32x1-minmax-neonfp16arith-x2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002051 "src/f16-spmm/gen/32x1-minmax-neonfp16arith.c",
2052 "src/f16-vbinary/gen/vadd-minmax-neonfp16arith-x8.c",
2053 "src/f16-vbinary/gen/vadd-minmax-neonfp16arith-x16.c",
2054 "src/f16-vbinary/gen/vaddc-minmax-neonfp16arith-x8.c",
2055 "src/f16-vbinary/gen/vaddc-minmax-neonfp16arith-x16.c",
2056 "src/f16-vbinary/gen/vdiv-minmax-neonfp16arith-x8.c",
2057 "src/f16-vbinary/gen/vdiv-minmax-neonfp16arith-x16.c",
2058 "src/f16-vbinary/gen/vdivc-minmax-neonfp16arith-x8.c",
2059 "src/f16-vbinary/gen/vdivc-minmax-neonfp16arith-x16.c",
2060 "src/f16-vbinary/gen/vmax-neonfp16arith-x8.c",
2061 "src/f16-vbinary/gen/vmax-neonfp16arith-x16.c",
2062 "src/f16-vbinary/gen/vmaxc-neonfp16arith-x8.c",
2063 "src/f16-vbinary/gen/vmaxc-neonfp16arith-x16.c",
2064 "src/f16-vbinary/gen/vmin-neonfp16arith-x8.c",
2065 "src/f16-vbinary/gen/vmin-neonfp16arith-x16.c",
2066 "src/f16-vbinary/gen/vminc-neonfp16arith-x8.c",
2067 "src/f16-vbinary/gen/vminc-neonfp16arith-x16.c",
2068 "src/f16-vbinary/gen/vmul-minmax-neonfp16arith-x8.c",
2069 "src/f16-vbinary/gen/vmul-minmax-neonfp16arith-x16.c",
2070 "src/f16-vbinary/gen/vmulc-minmax-neonfp16arith-x8.c",
2071 "src/f16-vbinary/gen/vmulc-minmax-neonfp16arith-x16.c",
2072 "src/f16-vbinary/gen/vrdivc-minmax-neonfp16arith-x8.c",
2073 "src/f16-vbinary/gen/vrdivc-minmax-neonfp16arith-x16.c",
2074 "src/f16-vbinary/gen/vrsubc-minmax-neonfp16arith-x8.c",
2075 "src/f16-vbinary/gen/vrsubc-minmax-neonfp16arith-x16.c",
2076 "src/f16-vbinary/gen/vsub-minmax-neonfp16arith-x8.c",
2077 "src/f16-vbinary/gen/vsub-minmax-neonfp16arith-x16.c",
2078 "src/f16-vbinary/gen/vsubc-minmax-neonfp16arith-x8.c",
2079 "src/f16-vbinary/gen/vsubc-minmax-neonfp16arith-x16.c",
2080 "src/f16-vmulcaddc/gen/c8-minmax-neonfp16arith-2x.c",
2081 "src/f16-vmulcaddc/gen/c16-minmax-neonfp16arith-2x.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002082]
2083
Benoit Jacoba9644732020-08-13 12:48:55 -07002084NEONDOT_UKERNELS = [
Marat Dukhand4c83032020-09-09 23:20:12 -07002085 "src/qs8-gemm/gen/1x8c4-minmax-neondot.c",
Marat Dukhand4c83032020-09-09 23:20:12 -07002086 "src/qs8-gemm/gen/1x16c4-minmax-neondot.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002087 "src/qs8-gemm/gen/4x8c4-minmax-neondot.c",
Benoit Jacoba9644732020-08-13 12:48:55 -07002088 "src/qs8-gemm/gen/4x16c4-minmax-neondot.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002089 "src/qs8-gemm/gen/6x8c4-minmax-neondot.c",
Frank Barchard2fa17452020-09-22 09:51:46 -07002090 "src/qs8-gemm/gen/6x16c4-minmax-neondot.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002091 "src/qs8-gemm/gen/8x8c4-minmax-neondot.c",
Frank Barcharda48848f2020-09-24 00:59:24 -07002092 "src/qs8-gemm/gen/8x16c4-minmax-neondot.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002093 "src/qs8-gemm/gen/12x8c4-minmax-neondot.c",
Marat Dukhand4c83032020-09-09 23:20:12 -07002094 "src/qs8-igemm/gen/1x8c4-minmax-neondot.c",
Marat Dukhand4c83032020-09-09 23:20:12 -07002095 "src/qs8-igemm/gen/1x16c4-minmax-neondot.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002096 "src/qs8-igemm/gen/4x8c4-minmax-neondot.c",
Benoit Jacoba05487f2020-08-14 13:37:21 -07002097 "src/qs8-igemm/gen/4x16c4-minmax-neondot.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002098 "src/qs8-igemm/gen/6x8c4-minmax-neondot.c",
Frank Barchard2fa17452020-09-22 09:51:46 -07002099 "src/qs8-igemm/gen/6x16c4-minmax-neondot.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002100 "src/qs8-igemm/gen/8x8c4-minmax-neondot.c",
Frank Barcharda48848f2020-09-24 00:59:24 -07002101 "src/qs8-igemm/gen/8x16c4-minmax-neondot.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002102 "src/qs8-igemm/gen/12x8c4-minmax-neondot.c",
Benoit Jacoba9644732020-08-13 12:48:55 -07002103]
2104
Marat Dukhan08c4a432019-10-03 09:29:21 -07002105SSE_UKERNELS = [
Marat Dukhan99936602020-04-11 16:47:01 -07002106 "src/f32-avgpool/9p8x-minmax-sse-c4.c",
2107 "src/f32-avgpool/9x-minmax-sse-c4.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -07002108 "src/f32-clamp/gen/sse-x4.c",
2109 "src/f32-clamp/gen/sse-x8.c",
Erich Elsenb1233402020-06-08 15:53:15 -07002110 "src/f32-conv-hwc2chw/3x3s2p1c3x4-sse-1x1.c",
2111 "src/f32-conv-hwc2chw/3x3s2p1c3x4-sse-2x2.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002112 "src/f32-dwconv/gen/up4x4-minmax-sse-acc2.c",
2113 "src/f32-dwconv/gen/up4x4-minmax-sse.c",
2114 "src/f32-dwconv/gen/up4x9-minmax-sse-acc2.c",
2115 "src/f32-dwconv/gen/up4x9-minmax-sse.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002116 "src/f32-dwconv/gen/up4x25-minmax-sse-acc2.c",
2117 "src/f32-dwconv/gen/up4x25-minmax-sse.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002118 "src/f32-dwconv/gen/up8x4-minmax-sse-acc2.c",
2119 "src/f32-dwconv/gen/up8x4-minmax-sse.c",
2120 "src/f32-dwconv/gen/up8x9-minmax-sse-acc2.c",
2121 "src/f32-dwconv/gen/up8x9-minmax-sse.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002122 "src/f32-dwconv/gen/up8x25-minmax-sse-acc2.c",
2123 "src/f32-dwconv/gen/up8x25-minmax-sse.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07002124 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-1x4-acc2.c",
2125 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-1x4-acc3.c",
2126 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-1x4-acc4.c",
Marat Dukhan470078a2020-10-23 22:36:52 -07002127 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-1x4.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07002128 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-2x4-acc2.c",
Marat Dukhan470078a2020-10-23 22:36:52 -07002129 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-2x4.c",
2130 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-3x4.c",
2131 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-4x4.c",
2132 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-5x4.c",
2133 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-6x4.c",
Marat Dukhan0ff97182020-10-25 19:14:03 -07002134 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-sse-1x4-acc2.c",
2135 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-sse-1x4-acc3.c",
2136 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-sse-1x4-acc4.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07002137 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-sse-1x4.c",
Marat Dukhan0ff97182020-10-25 19:14:03 -07002138 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-sse-2x4-acc2.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07002139 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-sse-2x4.c",
2140 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-sse-3x4.c",
2141 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-sse-4x4.c",
Marat Dukhand0503892020-10-30 08:22:04 -07002142 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-1x4-acc2.c",
2143 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-1x4-acc3.c",
2144 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-1x4-acc4.c",
2145 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-1x4-acc5.c",
2146 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-1x4.c",
2147 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-2x4-acc2.c",
2148 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-2x4-acc3.c",
2149 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-2x4.c",
2150 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-3x4-acc2.c",
2151 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-3x4.c",
2152 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-4x4-acc2.c",
2153 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-4x4.c",
2154 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-5x4.c",
Marat Dukhanccca2142020-10-30 17:32:45 -07002155 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-sse-1x4-acc2.c",
2156 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-sse-1x4-acc3.c",
2157 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-sse-1x4-acc4.c",
2158 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-sse-1x4-acc5.c",
2159 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-sse-1x4.c",
2160 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-sse-2x4-acc2.c",
2161 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-sse-2x4-acc3.c",
2162 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-sse-2x4.c",
Marat Dukhanccca2142020-10-30 17:32:45 -07002163 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-sse-3x4-acc2.c",
Frank Barchard8ef44cd2020-11-03 12:30:23 -08002164 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-sse-3x4.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -07002165 "src/f32-gavgpool-cw/sse-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -07002166 "src/f32-gavgpool/7p7x-minmax-sse-c4.c",
2167 "src/f32-gavgpool/7x-minmax-sse-c4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002168 "src/f32-gemm/gen-inc/1x8inc-minmax-sse-dup.c",
2169 "src/f32-gemm/gen-inc/1x8inc-minmax-sse-load1.c",
2170 "src/f32-gemm/gen-inc/1x8s4inc-minmax-sse.c",
Marat Dukhan802fcae2020-12-11 14:37:25 -08002171 "src/f32-gemm/gen-inc/3x8inc-minmax-sse-dup.c",
2172 "src/f32-gemm/gen-inc/3x8inc-minmax-sse-load1.c",
2173 "src/f32-gemm/gen-inc/3x8s4inc-minmax-sse.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002174 "src/f32-gemm/gen-inc/4x8inc-minmax-sse-dup.c",
2175 "src/f32-gemm/gen-inc/4x8inc-minmax-sse-load1.c",
2176 "src/f32-gemm/gen-inc/4x8s4inc-minmax-sse.c",
Marat Dukhan802fcae2020-12-11 14:37:25 -08002177 "src/f32-gemm/gen-inc/5x8inc-minmax-sse-dup.c",
2178 "src/f32-gemm/gen-inc/5x8inc-minmax-sse-load1.c",
2179 "src/f32-gemm/gen-inc/5x8s4inc-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",
Marat Dukhan802fcae2020-12-11 14:37:25 -08002183 "src/f32-gemm/gen/3x8-minmax-sse-dup.c",
2184 "src/f32-gemm/gen/3x8-minmax-sse-load1.c",
2185 "src/f32-gemm/gen/3x8s4-minmax-sse.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002186 "src/f32-gemm/gen/4x2c4-minmax-sse.c",
2187 "src/f32-gemm/gen/4x8-minmax-sse-dup.c",
2188 "src/f32-gemm/gen/4x8-minmax-sse-load1.c",
2189 "src/f32-gemm/gen/4x8s4-minmax-sse.c",
Marat Dukhan802fcae2020-12-11 14:37:25 -08002190 "src/f32-gemm/gen/5x8-minmax-sse-dup.c",
2191 "src/f32-gemm/gen/5x8-minmax-sse-load1.c",
2192 "src/f32-gemm/gen/5x8s4-minmax-sse.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07002193 "src/f32-hswish/gen/hswish-sse-x4.c",
2194 "src/f32-hswish/gen/hswish-sse-x8.c",
Marat Dukhan660fd192020-03-10 04:55:30 -07002195 "src/f32-ibilinear/gen/sse-c4.c",
2196 "src/f32-ibilinear/gen/sse-c8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002197 "src/f32-igemm/gen/1x8-minmax-sse-dup.c",
2198 "src/f32-igemm/gen/1x8-minmax-sse-load1.c",
2199 "src/f32-igemm/gen/1x8s4-minmax-sse.c",
Marat Dukhan802fcae2020-12-11 14:37:25 -08002200 "src/f32-igemm/gen/3x8-minmax-sse-dup.c",
2201 "src/f32-igemm/gen/3x8-minmax-sse-load1.c",
2202 "src/f32-igemm/gen/3x8s4-minmax-sse.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002203 "src/f32-igemm/gen/4x2c4-minmax-sse.c",
2204 "src/f32-igemm/gen/4x8-minmax-sse-dup.c",
2205 "src/f32-igemm/gen/4x8-minmax-sse-load1.c",
2206 "src/f32-igemm/gen/4x8s4-minmax-sse.c",
Marat Dukhan802fcae2020-12-11 14:37:25 -08002207 "src/f32-igemm/gen/5x8-minmax-sse-dup.c",
2208 "src/f32-igemm/gen/5x8-minmax-sse-load1.c",
2209 "src/f32-igemm/gen/5x8s4-minmax-sse.c",
Marat Dukhan99936602020-04-11 16:47:01 -07002210 "src/f32-maxpool/9p8x-minmax-sse-c4.c",
2211 "src/f32-pavgpool/9p8x-minmax-sse-c4.c",
2212 "src/f32-pavgpool/9x-minmax-sse-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002213 "src/f32-ppmm/gen/4x8-minmax-sse.c",
Marat Dukhan39b5e942020-06-24 15:03:48 -07002214 "src/f32-prelu/gen/sse-2x4.c",
2215 "src/f32-prelu/gen/sse-2x8.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07002216 "src/f32-relu/gen/sse-x4.c",
2217 "src/f32-relu/gen/sse-x8.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002218 "src/f32-rmax/sse.c",
Marat Dukhan355ab432020-04-09 19:01:52 -07002219 "src/f32-spmm/gen/4x1-minmax-sse.c",
2220 "src/f32-spmm/gen/8x1-minmax-sse.c",
Erich Elsen6e80fdc2020-06-09 15:35:37 -07002221 "src/f32-spmm/gen/16x1-minmax-sse.c",
Frank Barchard846c0c62020-10-26 15:01:39 -07002222 "src/f32-spmm/gen/32x1-minmax-sse.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002223 "src/f32-vbinary/gen/vadd-minmax-sse-x4.c",
2224 "src/f32-vbinary/gen/vadd-minmax-sse-x8.c",
2225 "src/f32-vbinary/gen/vaddc-minmax-sse-x4.c",
2226 "src/f32-vbinary/gen/vaddc-minmax-sse-x8.c",
2227 "src/f32-vbinary/gen/vdiv-minmax-sse-x4.c",
2228 "src/f32-vbinary/gen/vdiv-minmax-sse-x8.c",
2229 "src/f32-vbinary/gen/vdivc-minmax-sse-x4.c",
2230 "src/f32-vbinary/gen/vdivc-minmax-sse-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -08002231 "src/f32-vbinary/gen/vmax-sse-x4.c",
2232 "src/f32-vbinary/gen/vmax-sse-x8.c",
2233 "src/f32-vbinary/gen/vmaxc-sse-x4.c",
2234 "src/f32-vbinary/gen/vmaxc-sse-x8.c",
2235 "src/f32-vbinary/gen/vmin-sse-x4.c",
2236 "src/f32-vbinary/gen/vmin-sse-x8.c",
2237 "src/f32-vbinary/gen/vminc-sse-x4.c",
2238 "src/f32-vbinary/gen/vminc-sse-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002239 "src/f32-vbinary/gen/vmul-minmax-sse-x4.c",
2240 "src/f32-vbinary/gen/vmul-minmax-sse-x8.c",
2241 "src/f32-vbinary/gen/vmulc-minmax-sse-x4.c",
2242 "src/f32-vbinary/gen/vmulc-minmax-sse-x8.c",
2243 "src/f32-vbinary/gen/vrdivc-minmax-sse-x4.c",
2244 "src/f32-vbinary/gen/vrdivc-minmax-sse-x8.c",
2245 "src/f32-vbinary/gen/vrsubc-minmax-sse-x4.c",
2246 "src/f32-vbinary/gen/vrsubc-minmax-sse-x8.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07002247 "src/f32-vbinary/gen/vsqrdiff-sse-x4.c",
2248 "src/f32-vbinary/gen/vsqrdiff-sse-x8.c",
2249 "src/f32-vbinary/gen/vsqrdiffc-sse-x4.c",
2250 "src/f32-vbinary/gen/vsqrdiffc-sse-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002251 "src/f32-vbinary/gen/vsub-minmax-sse-x4.c",
2252 "src/f32-vbinary/gen/vsub-minmax-sse-x8.c",
2253 "src/f32-vbinary/gen/vsubc-minmax-sse-x4.c",
2254 "src/f32-vbinary/gen/vsubc-minmax-sse-x8.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07002255 "src/f32-vlrelu/gen/vlrelu-sse-x4.c",
2256 "src/f32-vlrelu/gen/vlrelu-sse-x8.c",
Marat Dukhan99936602020-04-11 16:47:01 -07002257 "src/f32-vmulcaddc/gen/c4-minmax-sse-2x.c",
2258 "src/f32-vmulcaddc/gen/c8-minmax-sse-2x.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07002259 "src/f32-vsqrt/gen/sse-sqrt-x4.c",
2260 "src/f32-vsqrt/gen/sse-sqrt-x8.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07002261 "src/f32-vunary/gen/vabs-sse-x4.c",
2262 "src/f32-vunary/gen/vabs-sse-x8.c",
2263 "src/f32-vunary/gen/vneg-sse-x4.c",
2264 "src/f32-vunary/gen/vneg-sse-x8.c",
2265 "src/f32-vunary/gen/vsqr-sse-x4.c",
2266 "src/f32-vunary/gen/vsqr-sse-x8.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07002267 "src/math/roundd-sse-addsub.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002268 "src/math/roundne-sse-addsub.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07002269 "src/math/roundu-sse-addsub.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07002270 "src/math/roundz-sse-addsub.c",
Marat Dukhan84000762020-06-29 18:38:43 -07002271 "src/math/sqrt-sse-hh1mac.c",
2272 "src/math/sqrt-sse-nr1mac.c",
2273 "src/math/sqrt-sse-nr2mac.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002274 "src/x32-fill/sse.c",
2275 "src/x32-packx/x4-sse.c",
2276 "src/x32-pad/sse.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002277]
2278
2279SSE2_UKERNELS = [
Marat Dukhan329da642019-11-19 21:44:39 -08002280 "src/f32-argmaxpool/4x-sse2-c4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002281 "src/f32-argmaxpool/9p8x-sse2-c4.c",
Marat Dukhan329da642019-11-19 21:44:39 -08002282 "src/f32-argmaxpool/9x-sse2-c4.c",
Marat Dukhan802fcae2020-12-11 14:37:25 -08002283 "src/f32-gemm/gen-inc/1x8inc-minmax-sse2-dup.c",
2284 "src/f32-gemm/gen-inc/3x8inc-minmax-sse2-dup.c",
2285 "src/f32-gemm/gen-inc/4x8inc-minmax-sse2-dup.c",
2286 "src/f32-gemm/gen-inc/5x8inc-minmax-sse2-dup.c",
2287 "src/f32-gemm/gen/1x8-minmax-sse2-dup.c",
2288 "src/f32-gemm/gen/3x8-minmax-sse2-dup.c",
2289 "src/f32-gemm/gen/4x8-minmax-sse2-dup.c",
2290 "src/f32-gemm/gen/5x8-minmax-sse2-dup.c",
2291 "src/f32-igemm/gen/1x8-minmax-sse2-dup.c",
2292 "src/f32-igemm/gen/3x8-minmax-sse2-dup.c",
2293 "src/f32-igemm/gen/4x8-minmax-sse2-dup.c",
2294 "src/f32-igemm/gen/5x8-minmax-sse2-dup.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08002295 "src/f32-prelu/gen/sse2-2x4.c",
2296 "src/f32-prelu/gen/sse2-2x8.c",
Marat Dukhanb39689d2020-01-24 13:32:20 -08002297 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x4.c",
Marat Dukhanb39689d2020-01-24 13:32:20 -08002298 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x8-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002299 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x8.c",
Marat Dukhanb39689d2020-01-24 13:32:20 -08002300 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12-acc2.c",
2301 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002302 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12.c",
Marat Dukhanb39689d2020-01-24 13:32:20 -08002303 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16-acc2.c",
2304 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002305 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16.c",
Marat Dukhanb39689d2020-01-24 13:32:20 -08002306 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20-acc2.c",
2307 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002308 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20.c",
Marat Dukhand243c1a2020-09-17 18:48:10 -07002309 "src/f32-sigmoid/gen/sse2-lut64-p2-div-x4.c",
2310 "src/f32-sigmoid/gen/sse2-lut64-p2-div-x8.c",
2311 "src/f32-sigmoid/gen/sse2-lut64-p2-div-x12.c",
2312 "src/f32-sigmoid/gen/sse2-lut64-p2-div-x16.c",
2313 "src/f32-sigmoid/gen/sse2-lut64-p2-div-x20.c",
2314 "src/f32-sigmoid/gen/sse2-lut64-p2-div-x24.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002315 "src/f32-sigmoid/gen/sse2-p5-div-x4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08002316 "src/f32-sigmoid/gen/sse2-p5-div-x8.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002317 "src/f32-sigmoid/gen/sse2-p5-div-x12.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08002318 "src/f32-sigmoid/gen/sse2-p5-div-x16.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002319 "src/f32-sigmoid/gen/sse2-p5-div-x20.c",
2320 "src/f32-sigmoid/gen/sse2-p5-div-x24.c",
Marat Dukhaned6baaf2020-12-01 15:07:08 -08002321 "src/f32-velu/gen/velu-sse2-rr2-lut16-p3-x4.c",
2322 "src/f32-velu/gen/velu-sse2-rr2-lut16-p3-x8.c",
2323 "src/f32-velu/gen/velu-sse2-rr2-lut16-p3-x12.c",
2324 "src/f32-velu/gen/velu-sse2-rr2-lut16-p3-x16.c",
2325 "src/f32-velu/gen/velu-sse2-rr2-lut16-p3-x20.c",
2326 "src/f32-velu/gen/velu-sse2-rr2-lut16-p3-x24.c",
2327 "src/f32-velu/gen/velu-sse2-rr2-p6-x4.c",
2328 "src/f32-velu/gen/velu-sse2-rr2-p6-x8.c",
2329 "src/f32-velu/gen/velu-sse2-rr2-p6-x12.c",
2330 "src/f32-velu/gen/velu-sse2-rr2-p6-x16.c",
2331 "src/f32-velu/gen/velu-sse2-rr2-p6-x20.c",
2332 "src/f32-velu/gen/velu-sse2-rr2-p6-x24.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07002333 "src/f32-vlrelu/gen/vlrelu-sse2-x4.c",
2334 "src/f32-vlrelu/gen/vlrelu-sse2-x8.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07002335 "src/f32-vrnd/gen/vrndd-sse2-x4.c",
2336 "src/f32-vrnd/gen/vrndd-sse2-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002337 "src/f32-vrnd/gen/vrndne-sse2-x4.c",
2338 "src/f32-vrnd/gen/vrndne-sse2-x8.c",
2339 "src/f32-vrnd/gen/vrndu-sse2-x4.c",
2340 "src/f32-vrnd/gen/vrndu-sse2-x8.c",
2341 "src/f32-vrnd/gen/vrndz-sse2-x4.c",
2342 "src/f32-vrnd/gen/vrndz-sse2-x8.c",
Marat Dukhanb7633f22020-11-20 16:34:56 -08002343 "src/math/exp-sse2-rr2-lut64-p2.c",
2344 "src/math/exp-sse2-rr2-p5.c",
Marat Dukhande390d42020-11-29 19:32:18 -08002345 "src/math/expm1minus-sse2-rr2-lut16-p3.c",
Marat Dukhana438aca2020-11-20 15:45:01 -08002346 "src/math/expm1minus-sse2-rr2-p6.c",
Frank Barchard3b800452020-11-22 12:12:35 -08002347 "src/math/expminus-sse2-rr2-p5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002348 "src/math/roundd-sse2-cvt.c",
2349 "src/math/roundne-sse2-cvt.c",
2350 "src/math/roundu-sse2-cvt.c",
2351 "src/math/roundz-sse2-cvt.c",
2352 "src/math/sigmoid-sse2-rr2-lut64-p2-div.c",
2353 "src/math/sigmoid-sse2-rr2-lut64-p2-nr1.c",
2354 "src/math/sigmoid-sse2-rr2-lut64-p2-nr2.c",
2355 "src/math/sigmoid-sse2-rr2-p5-div.c",
2356 "src/math/sigmoid-sse2-rr2-p5-nr1.c",
2357 "src/math/sigmoid-sse2-rr2-p5-nr2.c",
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07002358 "src/qs8-dwconv/gen/up8x9-minmax-sse2-mul16.c",
2359 "src/qs8-dwconv/gen/up16x9-minmax-sse2-mul16.c",
2360 "src/qs8-dwconv/gen/up24x9-minmax-sse2-mul16.c",
Marat Dukhan159688f2020-08-06 10:34:29 -07002361 "src/qs8-gavgpool/gen/7p7x-minmax-sse2-c8-acc2.c",
2362 "src/qs8-gavgpool/gen/7p7x-minmax-sse2-c16-acc2.c",
2363 "src/qs8-gavgpool/gen/7p7x-minmax-sse2-c24-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002364 "src/qs8-gavgpool/gen/7x-minmax-sse2-c8-acc2.c",
2365 "src/qs8-gavgpool/gen/7x-minmax-sse2-c16-acc2.c",
2366 "src/qs8-gavgpool/gen/7x-minmax-sse2-c24-acc2.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002367 "src/qs8-gemm/gen/1x4c2-minmax-sse2-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002368 "src/qs8-gemm/gen/1x4c2-minmax-sse2-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002369 "src/qs8-gemm/gen/1x4c2-xw-minmax-sse2.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002370 "src/qs8-gemm/gen/1x4c8-minmax-sse2-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002371 "src/qs8-gemm/gen/1x4c8-minmax-sse2-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002372 "src/qs8-gemm/gen/1x4c8-xw-minmax-sse2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002373 "src/qs8-gemm/gen/2x4c8-minmax-sse2-ld64.c",
2374 "src/qs8-gemm/gen/2x4c8-minmax-sse2-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002375 "src/qs8-gemm/gen/2x4c8-xw-minmax-sse2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002376 "src/qs8-gemm/gen/3x4c8-minmax-sse2-ld64.c",
2377 "src/qs8-gemm/gen/3x4c8-minmax-sse2-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002378 "src/qs8-gemm/gen/3x4c8-xw-minmax-sse2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002379 "src/qs8-gemm/gen/4x4c2-minmax-sse2-ld64.c",
2380 "src/qs8-gemm/gen/4x4c2-minmax-sse2-ld128.c",
2381 "src/qs8-gemm/gen/4x4c2-xw-minmax-sse2.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002382 "src/qs8-igemm/gen/1x4c2-minmax-sse2-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002383 "src/qs8-igemm/gen/1x4c2-minmax-sse2-ld128.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002384 "src/qs8-igemm/gen/1x4c8-minmax-sse2-ld64.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002385 "src/qs8-igemm/gen/1x4c8-minmax-sse2-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002386 "src/qs8-igemm/gen/2x4c8-minmax-sse2-ld64.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002387 "src/qs8-igemm/gen/2x4c8-minmax-sse2-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002388 "src/qs8-igemm/gen/3x4c8-minmax-sse2-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002389 "src/qs8-igemm/gen/3x4c8-minmax-sse2-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002390 "src/qs8-igemm/gen/4x4c2-minmax-sse2-ld64.c",
2391 "src/qs8-igemm/gen/4x4c2-minmax-sse2-ld128.c",
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07002392 "src/qs8-requantization/fp32-sse2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002393 "src/qs8-requantization/precise-sse2.c",
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07002394 "src/qs8-requantization/q31-sse2.c",
Marat Dukhand9f3ad42020-08-10 12:30:58 -07002395 "src/qs8-vadd/gen/minmax-sse2-mul16-ld64-x8.c",
2396 "src/qs8-vadd/gen/minmax-sse2-mul16-ld64-x16.c",
2397 "src/qs8-vadd/gen/minmax-sse2-mul16-ld64-x24.c",
2398 "src/qs8-vadd/gen/minmax-sse2-mul16-ld64-x32.c",
Marat Dukhan0270d9f2020-08-11 00:56:46 -07002399 "src/qs8-vaddc/gen/minmax-sse2-mul16-ld64-x8.c",
2400 "src/qs8-vaddc/gen/minmax-sse2-mul16-ld64-x16.c",
2401 "src/qs8-vaddc/gen/minmax-sse2-mul16-ld64-x24.c",
2402 "src/qs8-vaddc/gen/minmax-sse2-mul16-ld64-x32.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -07002403 "src/qu8-avgpool/9p8x-minmax-sse2-c8.c",
2404 "src/qu8-avgpool/9x-minmax-sse2-c8.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -07002405 "src/qu8-dwconv/up8x9-minmax-sse2.c",
2406 "src/qu8-gavgpool/7p7x-minmax-sse2-c8.c",
2407 "src/qu8-gavgpool/7x-minmax-sse2-c8.c",
2408 "src/qu8-gemm/2x4c8-minmax-sse2.c",
2409 "src/qu8-gemm/4x4c2-minmax-sse2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002410 "src/qu8-igemm/4x4c2-minmax-sse2.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07002411 "src/qu8-requantization/fp32-sse2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002412 "src/qu8-requantization/precise-sse2.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07002413 "src/qu8-requantization/q31-sse2.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -07002414 "src/qu8-vadd/minmax-sse2.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -07002415 "src/u8-clamp/sse2-x64.c",
Marat Dukhan99936602020-04-11 16:47:01 -07002416 "src/u8-maxpool/9p8x-minmax-sse2-c16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002417 "src/u8-rmax/sse2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002418 "src/x8-zip/x2-sse2.c",
2419 "src/x8-zip/x3-sse2.c",
2420 "src/x8-zip/x4-sse2.c",
2421 "src/x8-zip/xm-sse2.c",
Marat Dukhan57dccd82020-04-14 00:53:10 -07002422 "src/x32-unpool/sse2.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002423 "src/x32-zip/x2-sse2.c",
2424 "src/x32-zip/x3-sse2.c",
2425 "src/x32-zip/x4-sse2.c",
2426 "src/x32-zip/xm-sse2.c",
Marat Dukhanfe7acb62020-03-09 19:30:05 -07002427]
2428
2429SSSE3_UKERNELS = [
Frank Barchard35db7d02020-10-26 13:37:34 -07002430 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-1x4-acc2.c",
2431 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-1x4-acc3.c",
2432 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-1x4-acc4.c",
Marat Dukhan98f2eeb2020-10-23 23:13:41 -07002433 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-1x4.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07002434 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-2x4-acc2.c",
Marat Dukhan98f2eeb2020-10-23 23:13:41 -07002435 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-2x4.c",
2436 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-3x4.c",
2437 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-4x4.c",
2438 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-5x4.c",
2439 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-6x4.c",
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07002440 "src/qs8-dwconv/gen/up8x9-minmax-ssse3-mul16.c",
2441 "src/qs8-dwconv/gen/up16x9-minmax-ssse3-mul16.c",
2442 "src/qs8-dwconv/gen/up24x9-minmax-ssse3-mul16.c",
Marat Dukhan159688f2020-08-06 10:34:29 -07002443 "src/qs8-gavgpool/gen/7p7x-minmax-ssse3-c8-acc2.c",
2444 "src/qs8-gavgpool/gen/7p7x-minmax-ssse3-c16-acc2.c",
2445 "src/qs8-gavgpool/gen/7p7x-minmax-ssse3-c24-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002446 "src/qs8-gavgpool/gen/7x-minmax-ssse3-c8-acc2.c",
2447 "src/qs8-gavgpool/gen/7x-minmax-ssse3-c16-acc2.c",
2448 "src/qs8-gavgpool/gen/7x-minmax-ssse3-c24-acc2.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002449 "src/qs8-gemm/gen/1x4c2-minmax-ssse3-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002450 "src/qs8-gemm/gen/1x4c2-minmax-ssse3-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002451 "src/qs8-gemm/gen/1x4c2-xw-minmax-ssse3.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002452 "src/qs8-gemm/gen/1x4c8-minmax-ssse3-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002453 "src/qs8-gemm/gen/1x4c8-minmax-ssse3-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002454 "src/qs8-gemm/gen/1x4c8-xw-minmax-ssse3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002455 "src/qs8-gemm/gen/2x4c8-minmax-ssse3-ld64.c",
2456 "src/qs8-gemm/gen/2x4c8-minmax-ssse3-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002457 "src/qs8-gemm/gen/2x4c8-xw-minmax-ssse3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002458 "src/qs8-gemm/gen/3x4c8-minmax-ssse3-ld64.c",
2459 "src/qs8-gemm/gen/3x4c8-minmax-ssse3-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002460 "src/qs8-gemm/gen/3x4c8-xw-minmax-ssse3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002461 "src/qs8-gemm/gen/4x4c2-minmax-ssse3-ld64.c",
2462 "src/qs8-gemm/gen/4x4c2-minmax-ssse3-ld128.c",
2463 "src/qs8-gemm/gen/4x4c2-xw-minmax-ssse3.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002464 "src/qs8-igemm/gen/1x4c2-minmax-ssse3-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002465 "src/qs8-igemm/gen/1x4c2-minmax-ssse3-ld128.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002466 "src/qs8-igemm/gen/1x4c8-minmax-ssse3-ld64.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002467 "src/qs8-igemm/gen/1x4c8-minmax-ssse3-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002468 "src/qs8-igemm/gen/2x4c8-minmax-ssse3-ld64.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002469 "src/qs8-igemm/gen/2x4c8-minmax-ssse3-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002470 "src/qs8-igemm/gen/3x4c8-minmax-ssse3-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002471 "src/qs8-igemm/gen/3x4c8-minmax-ssse3-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002472 "src/qs8-igemm/gen/4x4c2-minmax-ssse3-ld64.c",
2473 "src/qs8-igemm/gen/4x4c2-minmax-ssse3-ld128.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -07002474 "src/qs8-requantization/precise-ssse3.c",
2475 "src/qs8-requantization/q31-ssse3.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07002476 "src/qu8-requantization/precise-ssse3.c",
2477 "src/qu8-requantization/q31-ssse3.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002478]
2479
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08002480SSE41_UKERNELS = [
Marat Dukhan40a672f2019-11-25 03:08:22 -08002481 "src/f32-prelu/gen/sse41-2x4.c",
2482 "src/f32-prelu/gen/sse41-2x8.c",
Marat Dukhand243c1a2020-09-17 18:48:10 -07002483 "src/f32-sigmoid/gen/sse41-lut64-p2-div-x4.c",
2484 "src/f32-sigmoid/gen/sse41-lut64-p2-div-x8.c",
2485 "src/f32-sigmoid/gen/sse41-lut64-p2-div-x12.c",
2486 "src/f32-sigmoid/gen/sse41-lut64-p2-div-x16.c",
2487 "src/f32-sigmoid/gen/sse41-lut64-p2-div-x20.c",
2488 "src/f32-sigmoid/gen/sse41-lut64-p2-div-x24.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002489 "src/f32-sigmoid/gen/sse41-p5-div-x4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08002490 "src/f32-sigmoid/gen/sse41-p5-div-x8.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002491 "src/f32-sigmoid/gen/sse41-p5-div-x12.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08002492 "src/f32-sigmoid/gen/sse41-p5-div-x16.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002493 "src/f32-sigmoid/gen/sse41-p5-div-x20.c",
2494 "src/f32-sigmoid/gen/sse41-p5-div-x24.c",
Marat Dukhaned6baaf2020-12-01 15:07:08 -08002495 "src/f32-velu/gen/velu-sse41-rr2-lut16-p3-x4.c",
2496 "src/f32-velu/gen/velu-sse41-rr2-lut16-p3-x8.c",
2497 "src/f32-velu/gen/velu-sse41-rr2-lut16-p3-x12.c",
2498 "src/f32-velu/gen/velu-sse41-rr2-lut16-p3-x16.c",
2499 "src/f32-velu/gen/velu-sse41-rr2-lut16-p3-x20.c",
2500 "src/f32-velu/gen/velu-sse41-rr2-lut16-p3-x24.c",
2501 "src/f32-velu/gen/velu-sse41-rr2-p6-x4.c",
2502 "src/f32-velu/gen/velu-sse41-rr2-p6-x8.c",
2503 "src/f32-velu/gen/velu-sse41-rr2-p6-x12.c",
2504 "src/f32-velu/gen/velu-sse41-rr2-p6-x16.c",
2505 "src/f32-velu/gen/velu-sse41-rr2-p6-x20.c",
2506 "src/f32-velu/gen/velu-sse41-rr2-p6-x24.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07002507 "src/f32-vlrelu/gen/vlrelu-sse41-x4.c",
2508 "src/f32-vlrelu/gen/vlrelu-sse41-x8.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07002509 "src/f32-vrnd/gen/vrndd-sse41-x4.c",
2510 "src/f32-vrnd/gen/vrndd-sse41-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002511 "src/f32-vrnd/gen/vrndne-sse41-x4.c",
2512 "src/f32-vrnd/gen/vrndne-sse41-x8.c",
2513 "src/f32-vrnd/gen/vrndu-sse41-x4.c",
2514 "src/f32-vrnd/gen/vrndu-sse41-x8.c",
2515 "src/f32-vrnd/gen/vrndz-sse41-x4.c",
2516 "src/f32-vrnd/gen/vrndz-sse41-x8.c",
2517 "src/math/roundd-sse41.c",
2518 "src/math/roundne-sse41.c",
2519 "src/math/roundu-sse41.c",
2520 "src/math/roundz-sse41.c",
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07002521 "src/qs8-dwconv/gen/up8x9-minmax-sse41-mul16.c",
2522 "src/qs8-dwconv/gen/up16x9-minmax-sse41-mul16.c",
2523 "src/qs8-dwconv/gen/up24x9-minmax-sse41-mul16.c",
Marat Dukhan159688f2020-08-06 10:34:29 -07002524 "src/qs8-gavgpool/gen/7p7x-minmax-sse41-c8-acc2.c",
2525 "src/qs8-gavgpool/gen/7p7x-minmax-sse41-c16-acc2.c",
2526 "src/qs8-gavgpool/gen/7p7x-minmax-sse41-c24-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002527 "src/qs8-gavgpool/gen/7x-minmax-sse41-c8-acc2.c",
2528 "src/qs8-gavgpool/gen/7x-minmax-sse41-c16-acc2.c",
2529 "src/qs8-gavgpool/gen/7x-minmax-sse41-c24-acc2.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002530 "src/qs8-gemm/gen/1x4c2-minmax-sse41-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002531 "src/qs8-gemm/gen/1x4c2-minmax-sse41-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002532 "src/qs8-gemm/gen/1x4c2-xw-minmax-sse41.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002533 "src/qs8-gemm/gen/1x4c8-minmax-sse41-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002534 "src/qs8-gemm/gen/1x4c8-minmax-sse41-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002535 "src/qs8-gemm/gen/1x4c8-xw-minmax-sse41.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002536 "src/qs8-gemm/gen/2x4c8-minmax-sse41-ld64.c",
2537 "src/qs8-gemm/gen/2x4c8-minmax-sse41-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002538 "src/qs8-gemm/gen/2x4c8-xw-minmax-sse41.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002539 "src/qs8-gemm/gen/3x4c8-minmax-sse41-ld64.c",
2540 "src/qs8-gemm/gen/3x4c8-minmax-sse41-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002541 "src/qs8-gemm/gen/3x4c8-xw-minmax-sse41.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002542 "src/qs8-gemm/gen/4x4c2-minmax-sse41-ld64.c",
2543 "src/qs8-gemm/gen/4x4c2-minmax-sse41-ld128.c",
2544 "src/qs8-gemm/gen/4x4c2-xw-minmax-sse41.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002545 "src/qs8-igemm/gen/1x4c2-minmax-sse41-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002546 "src/qs8-igemm/gen/1x4c2-minmax-sse41-ld128.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002547 "src/qs8-igemm/gen/1x4c8-minmax-sse41-ld64.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002548 "src/qs8-igemm/gen/1x4c8-minmax-sse41-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002549 "src/qs8-igemm/gen/2x4c8-minmax-sse41-ld64.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002550 "src/qs8-igemm/gen/2x4c8-minmax-sse41-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002551 "src/qs8-igemm/gen/3x4c8-minmax-sse41-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002552 "src/qs8-igemm/gen/3x4c8-minmax-sse41-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002553 "src/qs8-igemm/gen/4x4c2-minmax-sse41-ld64.c",
2554 "src/qs8-igemm/gen/4x4c2-minmax-sse41-ld128.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -07002555 "src/qs8-requantization/fp32-sse4.c",
2556 "src/qs8-requantization/precise-sse4.c",
2557 "src/qs8-requantization/q31-sse4.c",
Marat Dukhand9f3ad42020-08-10 12:30:58 -07002558 "src/qs8-vadd/gen/minmax-sse41-mul16-ld64-x8.c",
2559 "src/qs8-vadd/gen/minmax-sse41-mul16-ld64-x16.c",
2560 "src/qs8-vadd/gen/minmax-sse41-mul16-ld64-x24.c",
2561 "src/qs8-vadd/gen/minmax-sse41-mul16-ld64-x32.c",
Marat Dukhanbb9225e2020-09-06 22:40:56 -07002562 "src/qs8-vadd/gen/minmax-sse41-mul32-ld32-x8.c",
2563 "src/qs8-vadd/gen/minmax-sse41-mul32-ld32-x16.c",
2564 "src/qs8-vadd/gen/minmax-sse41-mul32-ld32-x24.c",
2565 "src/qs8-vadd/gen/minmax-sse41-mul32-ld32-x32.c",
Marat Dukhan0270d9f2020-08-11 00:56:46 -07002566 "src/qs8-vaddc/gen/minmax-sse41-mul16-ld64-x8.c",
2567 "src/qs8-vaddc/gen/minmax-sse41-mul16-ld64-x16.c",
2568 "src/qs8-vaddc/gen/minmax-sse41-mul16-ld64-x24.c",
2569 "src/qs8-vaddc/gen/minmax-sse41-mul16-ld64-x32.c",
Marat Dukhanbb9225e2020-09-06 22:40:56 -07002570 "src/qs8-vaddc/gen/minmax-sse41-mul32-ld32-x8.c",
2571 "src/qs8-vaddc/gen/minmax-sse41-mul32-ld32-x16.c",
2572 "src/qs8-vaddc/gen/minmax-sse41-mul32-ld32-x24.c",
2573 "src/qs8-vaddc/gen/minmax-sse41-mul32-ld32-x32.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07002574 "src/qu8-requantization/precise-sse4.c",
2575 "src/qu8-requantization/q31-sse4.c",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08002576]
2577
Marat Dukhan08c4a432019-10-03 09:29:21 -07002578AVX_UKERNELS = [
Marat Dukhan5c5fa962020-03-10 18:38:33 -07002579 "src/f32-clamp/gen/avx-x8.c",
2580 "src/f32-clamp/gen/avx-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002581 "src/f32-dwconv/gen/up8x4-minmax-avx-acc2.c",
2582 "src/f32-dwconv/gen/up8x4-minmax-avx.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002583 "src/f32-dwconv/gen/up8x9-minmax-avx-acc2.c",
2584 "src/f32-dwconv/gen/up8x9-minmax-avx.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002585 "src/f32-dwconv/gen/up8x25-minmax-avx-acc2.c",
2586 "src/f32-dwconv/gen/up8x25-minmax-avx.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002587 "src/f32-dwconv/gen/up16x4-minmax-avx-acc2.c",
2588 "src/f32-dwconv/gen/up16x4-minmax-avx.c",
2589 "src/f32-dwconv/gen/up16x9-minmax-avx-acc2.c",
2590 "src/f32-dwconv/gen/up16x9-minmax-avx.c",
2591 "src/f32-dwconv/gen/up16x25-minmax-avx-acc2.c",
2592 "src/f32-dwconv/gen/up16x25-minmax-avx.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002593 "src/f32-gemm/gen-inc/1x8inc-minmax-avx-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002594 "src/f32-gemm/gen-inc/1x16inc-minmax-avx-broadcast.c",
2595 "src/f32-gemm/gen-inc/3x16inc-minmax-avx-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002596 "src/f32-gemm/gen-inc/4x8inc-minmax-avx-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002597 "src/f32-gemm/gen-inc/4x16inc-minmax-avx-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002598 "src/f32-gemm/gen-inc/5x8inc-minmax-avx-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002599 "src/f32-gemm/gen-inc/5x16inc-minmax-avx-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002600 "src/f32-gemm/gen-inc/6x8inc-minmax-avx-broadcast.c",
2601 "src/f32-gemm/gen-inc/7x8inc-minmax-avx-broadcast.c",
2602 "src/f32-gemm/gen/1x8-minmax-avx-broadcast.c",
2603 "src/f32-gemm/gen/1x16-minmax-avx-broadcast.c",
2604 "src/f32-gemm/gen/3x16-minmax-avx-broadcast.c",
2605 "src/f32-gemm/gen/4x8-minmax-avx-broadcast.c",
2606 "src/f32-gemm/gen/4x16-minmax-avx-broadcast.c",
2607 "src/f32-gemm/gen/5x8-minmax-avx-broadcast.c",
2608 "src/f32-gemm/gen/5x16-minmax-avx-broadcast.c",
2609 "src/f32-gemm/gen/6x8-minmax-avx-broadcast.c",
2610 "src/f32-gemm/gen/7x8-minmax-avx-broadcast.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07002611 "src/f32-hswish/gen/hswish-avx-x8.c",
2612 "src/f32-hswish/gen/hswish-avx-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002613 "src/f32-igemm/gen/1x8-minmax-avx-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002614 "src/f32-igemm/gen/1x16-minmax-avx-broadcast.c",
2615 "src/f32-igemm/gen/3x16-minmax-avx-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002616 "src/f32-igemm/gen/4x8-minmax-avx-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002617 "src/f32-igemm/gen/4x16-minmax-avx-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002618 "src/f32-igemm/gen/5x8-minmax-avx-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002619 "src/f32-igemm/gen/5x16-minmax-avx-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002620 "src/f32-igemm/gen/6x8-minmax-avx-broadcast.c",
2621 "src/f32-igemm/gen/7x8-minmax-avx-broadcast.c",
Marat Dukhan90eca0a2020-03-11 00:52:23 -07002622 "src/f32-prelu/gen/avx-2x8.c",
2623 "src/f32-prelu/gen/avx-2x16.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07002624 "src/f32-relu/gen/avx-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002625 "src/f32-relu/gen/avx-x16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002626 "src/f32-rmax/avx.c",
T.J. Alumbaughdc2b29c2020-10-14 13:56:08 -07002627 "src/f32-sigmoid/gen/avx-rr2-p5-div-x8.c",
2628 "src/f32-sigmoid/gen/avx-rr2-p5-div-x16.c",
2629 "src/f32-sigmoid/gen/avx-rr2-p5-div-x24.c",
2630 "src/f32-sigmoid/gen/avx-rr2-p5-div-x32.c",
2631 "src/f32-sigmoid/gen/avx-rr2-p5-div-x40.c",
2632 "src/f32-sigmoid/gen/avx-rr2-p5-div-x48.c",
2633 "src/f32-sigmoid/gen/avx-rr2-p5-div-x56.c",
2634 "src/f32-sigmoid/gen/avx-rr2-p5-div-x64.c",
2635 "src/f32-sigmoid/gen/avx-rr2-p5-div-x72.c",
2636 "src/f32-sigmoid/gen/avx-rr2-p5-div-x80.c",
2637 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x8.c",
2638 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x16.c",
2639 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x24.c",
2640 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x32.c",
2641 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x40.c",
2642 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x48.c",
2643 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x56.c",
2644 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x64.c",
2645 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x72.c",
2646 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x80.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002647 "src/f32-vbinary/gen/vadd-minmax-avx-x8.c",
2648 "src/f32-vbinary/gen/vadd-minmax-avx-x16.c",
2649 "src/f32-vbinary/gen/vaddc-minmax-avx-x8.c",
2650 "src/f32-vbinary/gen/vaddc-minmax-avx-x16.c",
2651 "src/f32-vbinary/gen/vdiv-minmax-avx-x8.c",
2652 "src/f32-vbinary/gen/vdiv-minmax-avx-x16.c",
2653 "src/f32-vbinary/gen/vdivc-minmax-avx-x8.c",
2654 "src/f32-vbinary/gen/vdivc-minmax-avx-x16.c",
Marat Dukhan9a88efe2019-12-10 15:54:24 -08002655 "src/f32-vbinary/gen/vmax-avx-x8.c",
2656 "src/f32-vbinary/gen/vmax-avx-x16.c",
2657 "src/f32-vbinary/gen/vmaxc-avx-x8.c",
2658 "src/f32-vbinary/gen/vmaxc-avx-x16.c",
2659 "src/f32-vbinary/gen/vmin-avx-x8.c",
2660 "src/f32-vbinary/gen/vmin-avx-x16.c",
2661 "src/f32-vbinary/gen/vminc-avx-x8.c",
2662 "src/f32-vbinary/gen/vminc-avx-x16.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002663 "src/f32-vbinary/gen/vmul-minmax-avx-x8.c",
2664 "src/f32-vbinary/gen/vmul-minmax-avx-x16.c",
2665 "src/f32-vbinary/gen/vmulc-minmax-avx-x8.c",
2666 "src/f32-vbinary/gen/vmulc-minmax-avx-x16.c",
2667 "src/f32-vbinary/gen/vrdivc-minmax-avx-x8.c",
2668 "src/f32-vbinary/gen/vrdivc-minmax-avx-x16.c",
2669 "src/f32-vbinary/gen/vrsubc-minmax-avx-x8.c",
2670 "src/f32-vbinary/gen/vrsubc-minmax-avx-x16.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07002671 "src/f32-vbinary/gen/vsqrdiff-avx-x8.c",
2672 "src/f32-vbinary/gen/vsqrdiff-avx-x16.c",
2673 "src/f32-vbinary/gen/vsqrdiffc-avx-x8.c",
2674 "src/f32-vbinary/gen/vsqrdiffc-avx-x16.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002675 "src/f32-vbinary/gen/vsub-minmax-avx-x8.c",
2676 "src/f32-vbinary/gen/vsub-minmax-avx-x16.c",
2677 "src/f32-vbinary/gen/vsubc-minmax-avx-x8.c",
2678 "src/f32-vbinary/gen/vsubc-minmax-avx-x16.c",
Marat Dukhaned6baaf2020-12-01 15:07:08 -08002679 "src/f32-velu/gen/velu-avx-rr2-lut4-p4-perm-x8.c",
2680 "src/f32-velu/gen/velu-avx-rr2-lut4-p4-perm-x16.c",
2681 "src/f32-velu/gen/velu-avx-rr2-lut4-p4-perm-x24.c",
2682 "src/f32-velu/gen/velu-avx-rr2-lut4-p4-perm-x32.c",
2683 "src/f32-velu/gen/velu-avx-rr2-lut4-p4-perm-x40.c",
2684 "src/f32-velu/gen/velu-avx-rr2-lut4-p4-perm-x48.c",
2685 "src/f32-velu/gen/velu-avx-rr2-lut16-p3-x8.c",
2686 "src/f32-velu/gen/velu-avx-rr2-lut16-p3-x16.c",
2687 "src/f32-velu/gen/velu-avx-rr2-lut16-p3-x24.c",
2688 "src/f32-velu/gen/velu-avx-rr2-lut16-p3-x32.c",
2689 "src/f32-velu/gen/velu-avx-rr2-lut16-p3-x40.c",
2690 "src/f32-velu/gen/velu-avx-rr2-lut16-p3-x48.c",
2691 "src/f32-velu/gen/velu-avx-rr2-p6-x8.c",
2692 "src/f32-velu/gen/velu-avx-rr2-p6-x16.c",
2693 "src/f32-velu/gen/velu-avx-rr2-p6-x24.c",
2694 "src/f32-velu/gen/velu-avx-rr2-p6-x32.c",
2695 "src/f32-velu/gen/velu-avx-rr2-p6-x40.c",
2696 "src/f32-velu/gen/velu-avx-rr2-p6-x48.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07002697 "src/f32-vlrelu/gen/vlrelu-avx-x8.c",
2698 "src/f32-vlrelu/gen/vlrelu-avx-x16.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07002699 "src/f32-vrnd/gen/vrndd-avx-x8.c",
2700 "src/f32-vrnd/gen/vrndd-avx-x16.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002701 "src/f32-vrnd/gen/vrndne-avx-x8.c",
2702 "src/f32-vrnd/gen/vrndne-avx-x16.c",
2703 "src/f32-vrnd/gen/vrndu-avx-x8.c",
2704 "src/f32-vrnd/gen/vrndu-avx-x16.c",
2705 "src/f32-vrnd/gen/vrndz-avx-x8.c",
2706 "src/f32-vrnd/gen/vrndz-avx-x16.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07002707 "src/f32-vscale/avx-x32.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07002708 "src/f32-vsqrt/gen/avx-sqrt-x8.c",
2709 "src/f32-vsqrt/gen/avx-sqrt-x16.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07002710 "src/f32-vunary/gen/vabs-avx-x8.c",
2711 "src/f32-vunary/gen/vabs-avx-x16.c",
2712 "src/f32-vunary/gen/vneg-avx-x8.c",
2713 "src/f32-vunary/gen/vneg-avx-x16.c",
2714 "src/f32-vunary/gen/vsqr-avx-x8.c",
2715 "src/f32-vunary/gen/vsqr-avx-x16.c",
Marat Dukhanb7633f22020-11-20 16:34:56 -08002716 "src/math/exp-avx-rr2-p5.c",
Marat Dukhande390d42020-11-29 19:32:18 -08002717 "src/math/expm1minus-avx-rr2-lut4-p4-perm.c",
2718 "src/math/expm1minus-avx-rr2-lut16-p3.c",
2719 "src/math/expm1minus-avx-rr2-p6.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002720 "src/math/sigmoid-avx-rr2-lut64-p2-div.c",
Marat Dukhan2e4727b2020-09-24 00:48:39 -07002721 "src/math/sigmoid-avx-rr2-p5-div.c",
Marat Dukhan39ab16f2020-10-16 14:30:35 -07002722 "src/math/sigmoid-avx-rr2-p5-nr1.c",
2723 "src/math/sigmoid-avx-rr2-p5-nr2.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002724]
2725
Marat Dukhan1566fee2020-08-02 21:55:41 -07002726XOP_UKERNELS = [
2727 "src/qs8-gemm/gen/1x4c2-minmax-xop-ld64.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002728 "src/qs8-gemm/gen/1x4c2-minmax-xop-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002729 "src/qs8-gemm/gen/1x4c2-xw-minmax-xop.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002730 "src/qs8-gemm/gen/1x4c8-minmax-xop-ld64.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002731 "src/qs8-gemm/gen/1x4c8-minmax-xop-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002732 "src/qs8-gemm/gen/1x4c8-xw-minmax-xop.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002733 "src/qs8-gemm/gen/2x4c8-minmax-xop-ld64.c",
2734 "src/qs8-gemm/gen/2x4c8-minmax-xop-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002735 "src/qs8-gemm/gen/2x4c8-xw-minmax-xop.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002736 "src/qs8-gemm/gen/3x4c8-minmax-xop-ld64.c",
2737 "src/qs8-gemm/gen/3x4c8-minmax-xop-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002738 "src/qs8-gemm/gen/3x4c8-xw-minmax-xop.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002739 "src/qs8-gemm/gen/4x4c2-minmax-xop-ld64.c",
2740 "src/qs8-gemm/gen/4x4c2-minmax-xop-ld128.c",
2741 "src/qs8-gemm/gen/4x4c2-xw-minmax-xop.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002742 "src/qs8-igemm/gen/1x4c2-minmax-xop-ld64.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002743 "src/qs8-igemm/gen/1x4c2-minmax-xop-ld128.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002744 "src/qs8-igemm/gen/1x4c8-minmax-xop-ld64.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002745 "src/qs8-igemm/gen/1x4c8-minmax-xop-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002746 "src/qs8-igemm/gen/2x4c8-minmax-xop-ld64.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002747 "src/qs8-igemm/gen/2x4c8-minmax-xop-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002748 "src/qs8-igemm/gen/3x4c8-minmax-xop-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002749 "src/qs8-igemm/gen/3x4c8-minmax-xop-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002750 "src/qs8-igemm/gen/4x4c2-minmax-xop-ld64.c",
2751 "src/qs8-igemm/gen/4x4c2-minmax-xop-ld128.c",
Marat Dukhanbb9225e2020-09-06 22:40:56 -07002752 "src/qs8-vadd/gen/minmax-xop-mul32-ld32-x8.c",
2753 "src/qs8-vadd/gen/minmax-xop-mul32-ld32-x16.c",
2754 "src/qs8-vadd/gen/minmax-xop-mul32-ld32-x24.c",
2755 "src/qs8-vadd/gen/minmax-xop-mul32-ld32-x32.c",
2756 "src/qs8-vaddc/gen/minmax-xop-mul32-ld32-x8.c",
2757 "src/qs8-vaddc/gen/minmax-xop-mul32-ld32-x16.c",
2758 "src/qs8-vaddc/gen/minmax-xop-mul32-ld32-x24.c",
2759 "src/qs8-vaddc/gen/minmax-xop-mul32-ld32-x32.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002760]
2761
Marat Dukhanfda12b82019-11-21 12:27:59 -08002762FMA3_UKERNELS = [
Marat Dukhan1c587112020-04-08 20:04:28 -07002763 "src/f32-dwconv/gen/up8x4-minmax-fma3-acc2.c",
2764 "src/f32-dwconv/gen/up8x4-minmax-fma3.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002765 "src/f32-dwconv/gen/up8x9-minmax-fma3-acc2.c",
2766 "src/f32-dwconv/gen/up8x9-minmax-fma3.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002767 "src/f32-dwconv/gen/up8x25-minmax-fma3-acc2.c",
2768 "src/f32-dwconv/gen/up8x25-minmax-fma3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002769 "src/f32-dwconv/gen/up16x4-minmax-fma3-acc2.c",
2770 "src/f32-dwconv/gen/up16x4-minmax-fma3.c",
2771 "src/f32-dwconv/gen/up16x9-minmax-fma3-acc2.c",
2772 "src/f32-dwconv/gen/up16x9-minmax-fma3.c",
2773 "src/f32-dwconv/gen/up16x25-minmax-fma3-acc2.c",
2774 "src/f32-dwconv/gen/up16x25-minmax-fma3.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002775 "src/f32-gemm/gen-inc/1x8inc-minmax-fma3-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002776 "src/f32-gemm/gen-inc/1x16inc-minmax-fma3-broadcast.c",
2777 "src/f32-gemm/gen-inc/1x16s4inc-minmax-fma3-broadcast.c",
2778 "src/f32-gemm/gen-inc/3x16inc-minmax-fma3-broadcast.c",
2779 "src/f32-gemm/gen-inc/3x16s4inc-minmax-fma3-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002780 "src/f32-gemm/gen-inc/4x8inc-minmax-fma3-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002781 "src/f32-gemm/gen-inc/4x16inc-minmax-fma3-broadcast.c",
2782 "src/f32-gemm/gen-inc/4x16s4inc-minmax-fma3-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002783 "src/f32-gemm/gen-inc/5x8inc-minmax-fma3-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002784 "src/f32-gemm/gen-inc/5x16inc-minmax-fma3-broadcast.c",
2785 "src/f32-gemm/gen-inc/5x16s4inc-minmax-fma3-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002786 "src/f32-gemm/gen-inc/6x8inc-minmax-fma3-broadcast.c",
2787 "src/f32-gemm/gen-inc/7x8inc-minmax-fma3-broadcast.c",
2788 "src/f32-gemm/gen-inc/8x8inc-minmax-fma3-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002789 "src/f32-gemm/gen/1x8-minmax-fma3-broadcast.c",
2790 "src/f32-gemm/gen/1x16-minmax-fma3-broadcast.c",
2791 "src/f32-gemm/gen/1x16s4-minmax-fma3-broadcast.c",
2792 "src/f32-gemm/gen/3x16-minmax-fma3-broadcast.c",
2793 "src/f32-gemm/gen/3x16s4-minmax-fma3-broadcast.c",
2794 "src/f32-gemm/gen/4x8-minmax-fma3-broadcast.c",
2795 "src/f32-gemm/gen/4x16-minmax-fma3-broadcast.c",
2796 "src/f32-gemm/gen/4x16s4-minmax-fma3-broadcast.c",
2797 "src/f32-gemm/gen/5x8-minmax-fma3-broadcast.c",
2798 "src/f32-gemm/gen/5x16-minmax-fma3-broadcast.c",
2799 "src/f32-gemm/gen/5x16s4-minmax-fma3-broadcast.c",
2800 "src/f32-gemm/gen/6x8-minmax-fma3-broadcast.c",
2801 "src/f32-gemm/gen/7x8-minmax-fma3-broadcast.c",
2802 "src/f32-gemm/gen/8x8-minmax-fma3-broadcast.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07002803 "src/f32-hswish/gen/hswish-fma3-x8.c",
2804 "src/f32-hswish/gen/hswish-fma3-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002805 "src/f32-igemm/gen/1x8-minmax-fma3-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002806 "src/f32-igemm/gen/1x16-minmax-fma3-broadcast.c",
2807 "src/f32-igemm/gen/1x16s4-minmax-fma3-broadcast.c",
2808 "src/f32-igemm/gen/3x16-minmax-fma3-broadcast.c",
2809 "src/f32-igemm/gen/3x16s4-minmax-fma3-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002810 "src/f32-igemm/gen/4x8-minmax-fma3-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002811 "src/f32-igemm/gen/4x16-minmax-fma3-broadcast.c",
2812 "src/f32-igemm/gen/4x16s4-minmax-fma3-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002813 "src/f32-igemm/gen/5x8-minmax-fma3-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002814 "src/f32-igemm/gen/5x16-minmax-fma3-broadcast.c",
2815 "src/f32-igemm/gen/5x16s4-minmax-fma3-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002816 "src/f32-igemm/gen/6x8-minmax-fma3-broadcast.c",
2817 "src/f32-igemm/gen/7x8-minmax-fma3-broadcast.c",
2818 "src/f32-igemm/gen/8x8-minmax-fma3-broadcast.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07002819 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x8.c",
2820 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x16.c",
2821 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x24.c",
2822 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x32.c",
2823 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x40.c",
2824 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x48.c",
2825 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x56.c",
2826 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x64.c",
Marat Dukhan84000762020-06-29 18:38:43 -07002827 "src/math/sqrt-fma3-nr1fma.c",
Marat Dukhan84000762020-06-29 18:38:43 -07002828 "src/math/sqrt-fma3-nr1fma1adj.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002829 "src/math/sqrt-fma3-nr2fma.c",
Marat Dukhanfda12b82019-11-21 12:27:59 -08002830]
2831
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002832AVX2_UKERNELS = [
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002833 "src/f32-raddexpminusmax/gen/avx2-p5-x64-acc2.c",
2834 "src/f32-raddexpminusmax/gen/avx2-p5-x64-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002835 "src/f32-raddexpminusmax/gen/avx2-p5-x64.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002836 "src/f32-raddexpminusmax/gen/avx2-p5-x72-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002837 "src/f32-raddexpminusmax/gen/avx2-p5-x72.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002838 "src/f32-raddexpminusmax/gen/avx2-p5-x80-acc2.c",
2839 "src/f32-raddexpminusmax/gen/avx2-p5-x80-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002840 "src/f32-raddexpminusmax/gen/avx2-p5-x80.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002841 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc2.c",
2842 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc3.c",
2843 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc6.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002844 "src/f32-raddexpminusmax/gen/avx2-p5-x96.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002845 "src/f32-raddextexp/gen/avx2-p5-x64-acc2.c",
2846 "src/f32-raddextexp/gen/avx2-p5-x64-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002847 "src/f32-raddextexp/gen/avx2-p5-x64.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002848 "src/f32-raddextexp/gen/avx2-p5-x72-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002849 "src/f32-raddextexp/gen/avx2-p5-x72.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002850 "src/f32-raddextexp/gen/avx2-p5-x80-acc2.c",
2851 "src/f32-raddextexp/gen/avx2-p5-x80-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002852 "src/f32-raddextexp/gen/avx2-p5-x80.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002853 "src/f32-raddextexp/gen/avx2-p5-x96-acc2.c",
2854 "src/f32-raddextexp/gen/avx2-p5-x96-acc3.c",
2855 "src/f32-raddextexp/gen/avx2-p5-x96-acc6.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002856 "src/f32-raddextexp/gen/avx2-p5-x96.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002857 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64-acc2.c",
2858 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002859 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002860 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x72-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002861 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x72.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002862 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80-acc2.c",
2863 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002864 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002865 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc2.c",
2866 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc3.c",
2867 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc6.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002868 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96.c",
Marat Dukhanfa0a4322020-01-06 16:14:29 -08002869 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x8.c",
2870 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x16.c",
2871 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x24.c",
2872 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x32.c",
2873 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x40.c",
2874 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x48.c",
2875 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x56.c",
2876 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x64.c",
2877 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x72.c",
2878 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x80.c",
2879 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x8.c",
2880 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x16.c",
2881 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x24.c",
2882 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x32.c",
2883 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x40.c",
2884 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x48.c",
2885 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x56.c",
2886 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x64.c",
2887 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x72.c",
2888 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x80.c",
2889 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x8.c",
2890 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x16.c",
2891 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x24.c",
2892 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x32.c",
2893 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x40.c",
2894 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x48.c",
2895 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x56.c",
2896 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x64.c",
2897 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x72.c",
2898 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x80.c",
Marat Dukhaned6baaf2020-12-01 15:07:08 -08002899 "src/f32-velu/gen/velu-avx2-rr1-lut4-p4-perm-x8.c",
2900 "src/f32-velu/gen/velu-avx2-rr1-lut4-p4-perm-x16.c",
2901 "src/f32-velu/gen/velu-avx2-rr1-lut4-p4-perm-x24.c",
2902 "src/f32-velu/gen/velu-avx2-rr1-lut4-p4-perm-x32.c",
2903 "src/f32-velu/gen/velu-avx2-rr1-lut4-p4-perm-x40.c",
2904 "src/f32-velu/gen/velu-avx2-rr1-lut4-p4-perm-x48.c",
2905 "src/f32-velu/gen/velu-avx2-rr1-lut4-p4-perm-x56.c",
2906 "src/f32-velu/gen/velu-avx2-rr1-lut4-p4-perm-x64.c",
2907 "src/f32-velu/gen/velu-avx2-rr1-lut4-p4-perm-x72.c",
2908 "src/f32-velu/gen/velu-avx2-rr1-lut4-p4-perm-x80.c",
2909 "src/f32-velu/gen/velu-avx2-rr1-lut8-p4-perm-x8.c",
2910 "src/f32-velu/gen/velu-avx2-rr1-lut8-p4-perm-x16.c",
2911 "src/f32-velu/gen/velu-avx2-rr1-lut8-p4-perm-x24.c",
2912 "src/f32-velu/gen/velu-avx2-rr1-lut8-p4-perm-x32.c",
2913 "src/f32-velu/gen/velu-avx2-rr1-lut8-p4-perm-x40.c",
2914 "src/f32-velu/gen/velu-avx2-rr1-lut8-p4-perm-x48.c",
2915 "src/f32-velu/gen/velu-avx2-rr1-lut8-p4-perm-x56.c",
2916 "src/f32-velu/gen/velu-avx2-rr1-lut8-p4-perm-x64.c",
2917 "src/f32-velu/gen/velu-avx2-rr1-lut8-p4-perm-x72.c",
2918 "src/f32-velu/gen/velu-avx2-rr1-lut8-p4-perm-x80.c",
2919 "src/f32-velu/gen/velu-avx2-rr1-lut16-p3-gather-x8.c",
2920 "src/f32-velu/gen/velu-avx2-rr1-lut16-p3-gather-x16.c",
2921 "src/f32-velu/gen/velu-avx2-rr1-lut16-p3-gather-x24.c",
2922 "src/f32-velu/gen/velu-avx2-rr1-lut16-p3-gather-x32.c",
2923 "src/f32-velu/gen/velu-avx2-rr1-lut16-p3-gather-x40.c",
2924 "src/f32-velu/gen/velu-avx2-rr1-lut16-p3-gather-x48.c",
2925 "src/f32-velu/gen/velu-avx2-rr1-lut16-p3-gather-x56.c",
2926 "src/f32-velu/gen/velu-avx2-rr1-lut16-p3-gather-x64.c",
2927 "src/f32-velu/gen/velu-avx2-rr1-lut16-p3-gather-x72.c",
2928 "src/f32-velu/gen/velu-avx2-rr1-lut16-p3-gather-x80.c",
2929 "src/f32-velu/gen/velu-avx2-rr1-p6-x8.c",
2930 "src/f32-velu/gen/velu-avx2-rr1-p6-x16.c",
2931 "src/f32-velu/gen/velu-avx2-rr1-p6-x24.c",
2932 "src/f32-velu/gen/velu-avx2-rr1-p6-x32.c",
2933 "src/f32-velu/gen/velu-avx2-rr1-p6-x40.c",
2934 "src/f32-velu/gen/velu-avx2-rr1-p6-x48.c",
2935 "src/f32-velu/gen/velu-avx2-rr1-p6-x56.c",
2936 "src/f32-velu/gen/velu-avx2-rr1-p6-x64.c",
2937 "src/f32-velu/gen/velu-avx2-rr1-p6-x72.c",
2938 "src/f32-velu/gen/velu-avx2-rr1-p6-x80.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002939 "src/f32-vscaleexpminusmax/gen/avx2-p5-x8.c",
2940 "src/f32-vscaleexpminusmax/gen/avx2-p5-x16.c",
2941 "src/f32-vscaleexpminusmax/gen/avx2-p5-x24.c",
2942 "src/f32-vscaleexpminusmax/gen/avx2-p5-x32.c",
2943 "src/f32-vscaleexpminusmax/gen/avx2-p5-x40.c",
2944 "src/f32-vscaleexpminusmax/gen/avx2-p5-x48.c",
2945 "src/f32-vscaleexpminusmax/gen/avx2-p5-x56.c",
2946 "src/f32-vscaleexpminusmax/gen/avx2-p5-x64.c",
2947 "src/f32-vscaleexpminusmax/gen/avx2-p5-x72.c",
2948 "src/f32-vscaleexpminusmax/gen/avx2-p5-x80.c",
2949 "src/f32-vscaleexpminusmax/gen/avx2-p5-x88.c",
2950 "src/f32-vscaleexpminusmax/gen/avx2-p5-x96.c",
2951 "src/f32-vscaleextexp/gen/avx2-p5-x8.c",
2952 "src/f32-vscaleextexp/gen/avx2-p5-x16.c",
2953 "src/f32-vscaleextexp/gen/avx2-p5-x24.c",
2954 "src/f32-vscaleextexp/gen/avx2-p5-x32.c",
2955 "src/f32-vscaleextexp/gen/avx2-p5-x40.c",
2956 "src/f32-vscaleextexp/gen/avx2-p5-x48.c",
2957 "src/f32-vscaleextexp/gen/avx2-p5-x56.c",
2958 "src/f32-vscaleextexp/gen/avx2-p5-x64.c",
2959 "src/f32-vscaleextexp/gen/avx2-p5-x72.c",
2960 "src/f32-vscaleextexp/gen/avx2-p5-x80.c",
2961 "src/f32-vscaleextexp/gen/avx2-p5-x88.c",
2962 "src/f32-vscaleextexp/gen/avx2-p5-x96.c",
Marat Dukhanb7633f22020-11-20 16:34:56 -08002963 "src/math/exp-avx2-rr2-lut8-p3-perm.c",
2964 "src/math/exp-avx2-rr2-lut8-p4-perm.c",
2965 "src/math/exp-avx2-rr2-p5.c",
Marat Dukhande390d42020-11-29 19:32:18 -08002966 "src/math/expm1minus-avx2-rr1-lut4-p4-perm.c",
2967 "src/math/expm1minus-avx2-rr1-lut8-p4-perm.c",
2968 "src/math/expm1minus-avx2-rr1-lut16-p3-gather.c",
2969 "src/math/expm1minus-avx2-rr1-p6.c",
Frank Barcharde7223ee2020-12-04 19:04:01 -08002970 "src/math/expminus-avx2-rr2-p5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002971 "src/math/extexp-avx2-p5.c",
2972 "src/math/sigmoid-avx2-rr1-lut64-p2-gather-div.c",
2973 "src/math/sigmoid-avx2-rr1-lut64-p2-gather-nr1fma.c",
2974 "src/math/sigmoid-avx2-rr1-lut64-p2-gather-nr2fma.c",
2975 "src/math/sigmoid-avx2-rr1-lut64-p2-gather-nr2fma1adj.c",
2976 "src/math/sigmoid-avx2-rr1-p5-div.c",
2977 "src/math/sigmoid-avx2-rr1-p5-nr1fma.c",
2978 "src/math/sigmoid-avx2-rr1-p5-nr2fma.c",
2979 "src/math/sigmoid-avx2-rr2-lut64-p2-gather-div.c",
2980 "src/math/sigmoid-avx2-rr2-lut64-p2-gather-nr1fma.c",
2981 "src/math/sigmoid-avx2-rr2-lut64-p2-gather-nr2fma.c",
2982 "src/math/sigmoid-avx2-rr2-lut64-p2-gather-nr2fma1adj.c",
2983 "src/math/sigmoid-avx2-rr2-p5-div.c",
2984 "src/math/sigmoid-avx2-rr2-p5-nr1fma.c",
2985 "src/math/sigmoid-avx2-rr2-p5-nr2fma.c",
Marat Dukhand65a1522020-08-04 19:28:18 -07002986 "src/qs8-dwconv/gen/up8x9-minmax-avx2-mul32.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002987 "src/qs8-dwconv/gen/up16x9-minmax-avx2-mul16.c",
Marat Dukhand65a1522020-08-04 19:28:18 -07002988 "src/qs8-dwconv/gen/up16x9-minmax-avx2-mul32.c",
2989 "src/qs8-dwconv/gen/up24x9-minmax-avx2-mul32.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002990 "src/qs8-dwconv/gen/up32x9-minmax-avx2-mul16.c",
Marat Dukhand65a1522020-08-04 19:28:18 -07002991 "src/qs8-dwconv/gen/up32x9-minmax-avx2-mul32.c",
Marat Dukhan12809522020-08-02 22:23:51 -07002992 "src/qs8-gemm/gen/1x8c8-minmax-avx2.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002993 "src/qs8-gemm/gen/1x8c8-xw-minmax-avx2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002994 "src/qs8-gemm/gen/2x8c8-minmax-avx2.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002995 "src/qs8-gemm/gen/2x8c8-xw-minmax-avx2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002996 "src/qs8-gemm/gen/3x8c8-minmax-avx2.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002997 "src/qs8-gemm/gen/3x8c8-xw-minmax-avx2.c",
Marat Dukhan12809522020-08-02 22:23:51 -07002998 "src/qs8-igemm/gen/1x8c8-minmax-avx2.c",
2999 "src/qs8-igemm/gen/2x8c8-minmax-avx2.c",
3000 "src/qs8-igemm/gen/3x8c8-minmax-avx2.c",
Marat Dukhane6dc0b62020-09-08 23:57:14 -07003001 "src/qs8-vadd/gen/minmax-avx2-mul32-ld64-x8.c",
3002 "src/qs8-vadd/gen/minmax-avx2-mul32-ld64-x16.c",
3003 "src/qs8-vadd/gen/minmax-avx2-mul32-ld64-x24.c",
3004 "src/qs8-vadd/gen/minmax-avx2-mul32-ld64-x32.c",
3005 "src/qs8-vaddc/gen/minmax-avx2-mul32-ld64-x8.c",
3006 "src/qs8-vaddc/gen/minmax-avx2-mul32-ld64-x16.c",
3007 "src/qs8-vaddc/gen/minmax-avx2-mul32-ld64-x24.c",
3008 "src/qs8-vaddc/gen/minmax-avx2-mul32-ld64-x32.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003009]
3010
Marat Dukhan08c4a432019-10-03 09:29:21 -07003011AVX512F_UKERNELS = [
Marat Dukhan5c5fa962020-03-10 18:38:33 -07003012 "src/f32-clamp/gen/avx512f-x16.c",
3013 "src/f32-clamp/gen/avx512f-x32.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07003014 "src/f32-dwconv/gen/up16x4-minmax-avx512f-acc2.c",
3015 "src/f32-dwconv/gen/up16x4-minmax-avx512f.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07003016 "src/f32-dwconv/gen/up16x9-minmax-avx512f-acc2.c",
3017 "src/f32-dwconv/gen/up16x9-minmax-avx512f.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07003018 "src/f32-dwconv/gen/up16x25-minmax-avx512f-acc2.c",
3019 "src/f32-dwconv/gen/up16x25-minmax-avx512f.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003020 "src/f32-dwconv/gen/up32x4-minmax-avx512f-acc2.c",
3021 "src/f32-dwconv/gen/up32x4-minmax-avx512f.c",
3022 "src/f32-dwconv/gen/up32x9-minmax-avx512f-acc2.c",
3023 "src/f32-dwconv/gen/up32x9-minmax-avx512f.c",
3024 "src/f32-dwconv/gen/up32x25-minmax-avx512f-acc2.c",
3025 "src/f32-dwconv/gen/up32x25-minmax-avx512f.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07003026 "src/f32-gemm/gen-inc/1x16inc-minmax-avx512f-broadcast.c",
3027 "src/f32-gemm/gen-inc/4x16inc-minmax-avx512f-broadcast.c",
3028 "src/f32-gemm/gen-inc/5x16inc-minmax-avx512f-broadcast.c",
3029 "src/f32-gemm/gen-inc/6x16inc-minmax-avx512f-broadcast.c",
3030 "src/f32-gemm/gen-inc/7x16inc-minmax-avx512f-broadcast.c",
3031 "src/f32-gemm/gen-inc/8x16inc-minmax-avx512f-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003032 "src/f32-gemm/gen/1x16-minmax-avx512f-broadcast.c",
3033 "src/f32-gemm/gen/4x16-minmax-avx512f-broadcast.c",
3034 "src/f32-gemm/gen/5x16-minmax-avx512f-broadcast.c",
3035 "src/f32-gemm/gen/6x16-minmax-avx512f-broadcast.c",
3036 "src/f32-gemm/gen/7x16-minmax-avx512f-broadcast.c",
3037 "src/f32-gemm/gen/8x16-minmax-avx512f-broadcast.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07003038 "src/f32-hswish/gen/hswish-avx512f-x16.c",
3039 "src/f32-hswish/gen/hswish-avx512f-x32.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07003040 "src/f32-igemm/gen/1x16-minmax-avx512f-broadcast.c",
3041 "src/f32-igemm/gen/4x16-minmax-avx512f-broadcast.c",
3042 "src/f32-igemm/gen/5x16-minmax-avx512f-broadcast.c",
3043 "src/f32-igemm/gen/6x16-minmax-avx512f-broadcast.c",
3044 "src/f32-igemm/gen/7x16-minmax-avx512f-broadcast.c",
3045 "src/f32-igemm/gen/8x16-minmax-avx512f-broadcast.c",
Marat Dukhan90eca0a2020-03-11 00:52:23 -07003046 "src/f32-prelu/gen/avx512f-2x16.c",
3047 "src/f32-prelu/gen/avx512f-2x32.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003048 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128-acc2.c",
3049 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003050 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003051 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x144-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003052 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x144.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003053 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160-acc2.c",
3054 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003055 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003056 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc2.c",
3057 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc3.c",
3058 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc6.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003059 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003060 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128-acc2.c",
3061 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003062 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003063 "src/f32-raddextexp/gen/avx512f-p5-scalef-x144-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003064 "src/f32-raddextexp/gen/avx512f-p5-scalef-x144.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003065 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160-acc2.c",
3066 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003067 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003068 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc2.c",
3069 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc3.c",
3070 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc6.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003071 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003072 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128-acc2.c",
3073 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003074 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003075 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x144-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003076 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x144.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003077 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160-acc2.c",
3078 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003079 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003080 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc2.c",
3081 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc3.c",
3082 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc6.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003083 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07003084 "src/f32-relu/gen/avx512f-x16.c",
3085 "src/f32-relu/gen/avx512f-x32.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003086 "src/f32-rmax/avx512f.c",
Marat Dukhand9ca7e62020-09-23 23:45:29 -07003087 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x16.c",
3088 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x32.c",
3089 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x48.c",
3090 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x64.c",
3091 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x80.c",
3092 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x96.c",
3093 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x112.c",
3094 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x128.c",
3095 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x16.c",
3096 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x32.c",
3097 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x48.c",
3098 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x64.c",
3099 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x80.c",
3100 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x96.c",
3101 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x112.c",
3102 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003103 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x16.c",
3104 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x32.c",
3105 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x48.c",
3106 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x64.c",
3107 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x80.c",
3108 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x96.c",
3109 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x112.c",
3110 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x128.c",
3111 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x16.c",
3112 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x32.c",
3113 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x48.c",
3114 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x64.c",
3115 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x80.c",
3116 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x96.c",
3117 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x112.c",
3118 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x128.c",
Marat Dukhand9ca7e62020-09-23 23:45:29 -07003119 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x16.c",
3120 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x32.c",
3121 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x48.c",
3122 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x64.c",
3123 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x80.c",
3124 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x96.c",
3125 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x112.c",
3126 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x128.c",
3127 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x16.c",
3128 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x32.c",
3129 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x48.c",
3130 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x64.c",
3131 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x80.c",
3132 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x96.c",
3133 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x112.c",
3134 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x128.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07003135 "src/f32-vbinary/gen/vadd-minmax-avx512f-x16.c",
3136 "src/f32-vbinary/gen/vadd-minmax-avx512f-x32.c",
3137 "src/f32-vbinary/gen/vaddc-minmax-avx512f-x16.c",
3138 "src/f32-vbinary/gen/vaddc-minmax-avx512f-x32.c",
3139 "src/f32-vbinary/gen/vdiv-minmax-avx512f-x16.c",
3140 "src/f32-vbinary/gen/vdiv-minmax-avx512f-x32.c",
3141 "src/f32-vbinary/gen/vdivc-minmax-avx512f-x16.c",
3142 "src/f32-vbinary/gen/vdivc-minmax-avx512f-x32.c",
Marat Dukhan9a88efe2019-12-10 15:54:24 -08003143 "src/f32-vbinary/gen/vmax-avx512f-x16.c",
3144 "src/f32-vbinary/gen/vmax-avx512f-x32.c",
3145 "src/f32-vbinary/gen/vmaxc-avx512f-x16.c",
3146 "src/f32-vbinary/gen/vmaxc-avx512f-x32.c",
3147 "src/f32-vbinary/gen/vmin-avx512f-x16.c",
3148 "src/f32-vbinary/gen/vmin-avx512f-x32.c",
3149 "src/f32-vbinary/gen/vminc-avx512f-x16.c",
3150 "src/f32-vbinary/gen/vminc-avx512f-x32.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07003151 "src/f32-vbinary/gen/vmul-minmax-avx512f-x16.c",
3152 "src/f32-vbinary/gen/vmul-minmax-avx512f-x32.c",
3153 "src/f32-vbinary/gen/vmulc-minmax-avx512f-x16.c",
3154 "src/f32-vbinary/gen/vmulc-minmax-avx512f-x32.c",
3155 "src/f32-vbinary/gen/vrdivc-minmax-avx512f-x16.c",
3156 "src/f32-vbinary/gen/vrdivc-minmax-avx512f-x32.c",
3157 "src/f32-vbinary/gen/vrsubc-minmax-avx512f-x16.c",
3158 "src/f32-vbinary/gen/vrsubc-minmax-avx512f-x32.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07003159 "src/f32-vbinary/gen/vsqrdiff-avx512f-x16.c",
3160 "src/f32-vbinary/gen/vsqrdiff-avx512f-x32.c",
3161 "src/f32-vbinary/gen/vsqrdiffc-avx512f-x16.c",
3162 "src/f32-vbinary/gen/vsqrdiffc-avx512f-x32.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07003163 "src/f32-vbinary/gen/vsub-minmax-avx512f-x16.c",
3164 "src/f32-vbinary/gen/vsub-minmax-avx512f-x32.c",
3165 "src/f32-vbinary/gen/vsubc-minmax-avx512f-x16.c",
3166 "src/f32-vbinary/gen/vsubc-minmax-avx512f-x32.c",
Marat Dukhaned6baaf2020-12-01 15:07:08 -08003167 "src/f32-velu/gen/velu-avx512f-rr1-lut16-p3-perm-x16.c",
3168 "src/f32-velu/gen/velu-avx512f-rr1-lut16-p3-perm-x32.c",
3169 "src/f32-velu/gen/velu-avx512f-rr1-lut16-p3-perm-x48.c",
3170 "src/f32-velu/gen/velu-avx512f-rr1-lut16-p3-perm-x64.c",
3171 "src/f32-velu/gen/velu-avx512f-rr1-lut16-p3-perm-x80.c",
3172 "src/f32-velu/gen/velu-avx512f-rr1-lut16-p3-perm-x96.c",
3173 "src/f32-velu/gen/velu-avx512f-rr1-lut16-p3-perm-x112.c",
3174 "src/f32-velu/gen/velu-avx512f-rr1-lut16-p3-perm-x128.c",
3175 "src/f32-velu/gen/velu-avx512f-rr1-p6-x16.c",
3176 "src/f32-velu/gen/velu-avx512f-rr1-p6-x32.c",
3177 "src/f32-velu/gen/velu-avx512f-rr1-p6-x48.c",
3178 "src/f32-velu/gen/velu-avx512f-rr1-p6-x64.c",
3179 "src/f32-velu/gen/velu-avx512f-rr1-p6-x80.c",
3180 "src/f32-velu/gen/velu-avx512f-rr1-p6-x96.c",
3181 "src/f32-velu/gen/velu-avx512f-rr1-p6-x112.c",
3182 "src/f32-velu/gen/velu-avx512f-rr1-p6-x128.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07003183 "src/f32-vlrelu/gen/vlrelu-avx512f-x16.c",
3184 "src/f32-vlrelu/gen/vlrelu-avx512f-x32.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003185 "src/f32-vrnd/gen/vrndd-avx512f-x16.c",
3186 "src/f32-vrnd/gen/vrndd-avx512f-x32.c",
3187 "src/f32-vrnd/gen/vrndne-avx512f-x16.c",
3188 "src/f32-vrnd/gen/vrndne-avx512f-x32.c",
3189 "src/f32-vrnd/gen/vrndu-avx512f-x16.c",
3190 "src/f32-vrnd/gen/vrndu-avx512f-x32.c",
3191 "src/f32-vrnd/gen/vrndz-avx512f-x16.c",
3192 "src/f32-vrnd/gen/vrndz-avx512f-x32.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07003193 "src/f32-vscale/avx512f-x64.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003194 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x16.c",
3195 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x32.c",
3196 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x48.c",
3197 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x64.c",
3198 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x80.c",
3199 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x96.c",
3200 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x112.c",
3201 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x128.c",
3202 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x144.c",
3203 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x160.c",
3204 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x176.c",
3205 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x192.c",
3206 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x16.c",
3207 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x32.c",
3208 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x48.c",
3209 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x64.c",
3210 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x80.c",
3211 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x96.c",
3212 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x112.c",
3213 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x128.c",
3214 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x144.c",
3215 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x160.c",
3216 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x176.c",
3217 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x192.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07003218 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x16.c",
3219 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x32.c",
3220 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x48.c",
3221 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x64.c",
3222 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x80.c",
3223 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x96.c",
3224 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x112.c",
3225 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x128.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07003226 "src/f32-vunary/gen/vabs-avx512f-x16.c",
3227 "src/f32-vunary/gen/vabs-avx512f-x32.c",
3228 "src/f32-vunary/gen/vneg-avx512f-x16.c",
3229 "src/f32-vunary/gen/vneg-avx512f-x32.c",
3230 "src/f32-vunary/gen/vsqr-avx512f-x16.c",
3231 "src/f32-vunary/gen/vsqr-avx512f-x32.c",
Marat Dukhanb7633f22020-11-20 16:34:56 -08003232 "src/math/exp-avx512f-rr2-lut16-p3-perm-scalef.c",
3233 "src/math/exp-avx512f-rr2-lut16-p3-perm.c",
3234 "src/math/exp-avx512f-rr2-lut32-p2-perm2-scalef.c",
3235 "src/math/exp-avx512f-rr2-lut32-p2-perm2.c",
3236 "src/math/exp-avx512f-rr2-p5-scalef.c",
3237 "src/math/exp-avx512f-rr2-p5.c",
Marat Dukhande390d42020-11-29 19:32:18 -08003238 "src/math/expm1minus-avx512f-rr1-lut16-p3-perm.c",
3239 "src/math/expm1minus-avx512f-rr1-p6.c",
Marat Dukhan98ba4412019-10-23 02:14:28 -07003240 "src/math/extexp-avx512f-p5.c",
Marat Dukhan6a34c5f2020-09-22 21:44:15 -07003241 "src/math/sigmoid-avx512f-rr1-lut16-p3-perm-scalef-div.c",
Marat Dukhan6a34c5f2020-09-22 21:44:15 -07003242 "src/math/sigmoid-avx512f-rr1-lut16-p3-perm-scalef-nr1fma.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003243 "src/math/sigmoid-avx512f-rr1-lut16-p3-perm-scalef-nr1fma1adj.c",
Marat Dukhan6a34c5f2020-09-22 21:44:15 -07003244 "src/math/sigmoid-avx512f-rr1-lut32-p2-perm2-scalef-div.c",
Marat Dukhan6a34c5f2020-09-22 21:44:15 -07003245 "src/math/sigmoid-avx512f-rr1-lut32-p2-perm2-scalef-nr1fma.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003246 "src/math/sigmoid-avx512f-rr1-lut32-p2-perm2-scalef-nr1fma1adj.c",
Marat Dukhan36173d22020-10-15 17:14:26 -07003247 "src/math/sigmoid-avx512f-rr1-lut64-p2-gather-scalef-div.c",
Marat Dukhan36173d22020-10-15 17:14:26 -07003248 "src/math/sigmoid-avx512f-rr1-lut64-p2-gather-scalef-nr1fma.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003249 "src/math/sigmoid-avx512f-rr1-lut64-p2-gather-scalef-nr1fma1adj.c",
3250 "src/math/sigmoid-avx512f-rr1-p5-scalef-div.c",
3251 "src/math/sigmoid-avx512f-rr1-p5-scalef-nr1fma.c",
3252 "src/math/sigmoid-avx512f-rr1-p5-scalef-nr1fma1adj.c",
3253 "src/math/sigmoid-avx512f-rr2-lut16-p3-perm-scalef-div.c",
3254 "src/math/sigmoid-avx512f-rr2-lut16-p3-perm-scalef-nr1fma.c",
3255 "src/math/sigmoid-avx512f-rr2-lut16-p3-perm-scalef-nr1fma1adj.c",
3256 "src/math/sigmoid-avx512f-rr2-lut32-p2-perm2-scalef-div.c",
3257 "src/math/sigmoid-avx512f-rr2-lut32-p2-perm2-scalef-nr1fma.c",
3258 "src/math/sigmoid-avx512f-rr2-lut32-p2-perm2-scalef-nr1fma1adj.c",
Marat Dukhan36173d22020-10-15 17:14:26 -07003259 "src/math/sigmoid-avx512f-rr2-lut64-p2-gather-scalef-div.c",
Marat Dukhan36173d22020-10-15 17:14:26 -07003260 "src/math/sigmoid-avx512f-rr2-lut64-p2-gather-scalef-nr1fma.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003261 "src/math/sigmoid-avx512f-rr2-lut64-p2-gather-scalef-nr1fma1adj.c",
3262 "src/math/sigmoid-avx512f-rr2-p5-scalef-div.c",
3263 "src/math/sigmoid-avx512f-rr2-p5-scalef-nr1fma.c",
3264 "src/math/sigmoid-avx512f-rr2-p5-scalef-nr1fma1adj.c",
Marat Dukhan84000762020-06-29 18:38:43 -07003265 "src/math/sqrt-avx512f-nr1fma.c",
Marat Dukhan84000762020-06-29 18:38:43 -07003266 "src/math/sqrt-avx512f-nr1fma1adj.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003267 "src/math/sqrt-avx512f-nr2fma.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003268]
3269
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07003270AVX512SKX_UKERNELS = [
Marat Dukhan2ffc5e62020-09-06 22:33:38 -07003271 "src/qs8-dwconv/gen/up16x9-minmax-avx512skx-mul32.c",
3272 "src/qs8-dwconv/gen/up32x9-minmax-avx512skx-mul32.c",
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07003273 "src/qs8-gemm/gen/1x16c8-minmax-avx512skx.c",
3274 "src/qs8-gemm/gen/2x16c8-minmax-avx512skx.c",
3275 "src/qs8-gemm/gen/3x16c8-minmax-avx512skx.c",
3276 "src/qs8-gemm/gen/4x16c8-minmax-avx512skx.c",
3277 "src/qs8-igemm/gen/1x16c8-minmax-avx512skx.c",
3278 "src/qs8-igemm/gen/2x16c8-minmax-avx512skx.c",
3279 "src/qs8-igemm/gen/3x16c8-minmax-avx512skx.c",
3280 "src/qs8-igemm/gen/4x16c8-minmax-avx512skx.c",
3281]
3282
Frank Barchardbcedc082020-08-17 18:00:51 -07003283WASM32_ASM_UKERNELS = [
3284 "src/f32-relu/wasm_shr_x1.S",
3285 "src/f32-relu/wasm_shr_x2.S",
3286 "src/f32-relu/wasm_shr_x4.S",
3287]
3288
Marat Dukhan08c4a432019-10-03 09:29:21 -07003289AARCH32_ASM_UKERNELS = [
Marat Dukhan32f93812020-05-17 20:31:21 -07003290 "src/f32-gemm/4x4-aarch32-vfp-ld64.S",
Marat Dukhan3b98f6b2020-05-17 10:09:22 -07003291 "src/f32-gemm/4x4-minmax-aarch32-vfp-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003292 "src/f32-gemm/4x8-minmax-aarch32-neon-cortex-a53.S",
3293 "src/f32-gemm/4x8-minmax-aarch32-neon-cortex-a55.S",
Frank Barchard490febe2020-07-16 18:42:17 -07003294 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-cortex-a7.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003295 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-cortex-a75.S",
Frank Barchard569561d2020-06-17 13:11:12 -07003296 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-ld64.S",
Frank Barchard490febe2020-07-16 18:42:17 -07003297 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-pld-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003298 "src/f32-igemm/4x8-minmax-aarch32-neon-cortex-a53.S",
3299 "src/f32-igemm/4x8-minmax-aarch32-neon-cortex-a55.S",
Frank Barchard490febe2020-07-16 18:42:17 -07003300 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-cortex-a7.S",
3301 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-cortex-a75.S",
3302 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-ld64.S",
3303 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-pld-cortex-a75.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003304]
3305
3306AARCH64_ASM_UKERNELS = [
Frank Barchardbddfbcd2020-04-15 12:32:41 -07003307 "src/f16-gemm/gen-inc/1x8inc-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchard04336c12020-10-22 16:48:55 -07003308 "src/f16-gemm/gen-inc/1x16inc-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchardbddfbcd2020-04-15 12:32:41 -07003309 "src/f16-gemm/gen-inc/4x8inc-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchard04336c12020-10-22 16:48:55 -07003310 "src/f16-gemm/gen-inc/4x16inc-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchardbddfbcd2020-04-15 12:32:41 -07003311 "src/f16-gemm/gen-inc/6x8inc-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchard04336c12020-10-22 16:48:55 -07003312 "src/f16-gemm/gen-inc/6x16inc-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchard3b8e5662020-04-20 12:12:53 -07003313 "src/f16-gemm/gen-inc/8x8inc-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchard04336c12020-10-22 16:48:55 -07003314 "src/f16-gemm/gen/1x8-minmax-aarch64-neonfp16arith-ld64.S",
3315 "src/f16-gemm/gen/1x16-minmax-aarch64-neonfp16arith-ld32.S",
3316 "src/f16-gemm/gen/4x8-minmax-aarch64-neonfp16arith-ld64.S",
3317 "src/f16-gemm/gen/4x16-minmax-aarch64-neonfp16arith-ld32.S",
3318 "src/f16-gemm/gen/6x8-minmax-aarch64-neonfp16arith-ld64.S",
3319 "src/f16-gemm/gen/6x16-minmax-aarch64-neonfp16arith-ld32.S",
3320 "src/f16-gemm/gen/8x8-minmax-aarch64-neonfp16arith-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003321 "src/f32-dwconv/up4x9-minmax-aarch64-neonfma-cortex-a55.S",
3322 "src/f32-dwconv/up4x9-minmax-aarch64-neonfma.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003323 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a53.S",
3324 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a57.S",
3325 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a75.S",
Frank Barchard04336c12020-10-22 16:48:55 -07003326 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-ld64.S",
3327 "src/f32-gemm/gen-inc/1x12inc-minmax-aarch64-neonfma-cortex-a53.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003328 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a53.S",
3329 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a55.S",
3330 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a57.S",
3331 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003332 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-ld64.S",
Frank Barchard04336c12020-10-22 16:48:55 -07003333 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-ld128.S",
3334 "src/f32-gemm/gen-inc/4x12inc-minmax-aarch64-neonfma-cortex-a53.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003335 "src/f32-gemm/gen-inc/5x8inc-minmax-aarch64-neonfma-cortex-a57.S",
3336 "src/f32-gemm/gen-inc/5x8inc-minmax-aarch64-neonfma-cortex-a75.S",
3337 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a53.S",
3338 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a55.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003339 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a57.S",
Frank Barchard04336c12020-10-22 16:48:55 -07003340 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a73.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003341 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003342 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-ld64.S",
Frank Barchard04336c12020-10-22 16:48:55 -07003343 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-ld128.S",
3344 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a53.S",
3345 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a57.S",
3346 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a75.S",
3347 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-ld64.S",
3348 "src/f32-gemm/gen/1x12-minmax-aarch64-neonfma-cortex-a53.S",
3349 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a53.S",
3350 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a55.S",
3351 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a57.S",
3352 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a75.S",
3353 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-ld64.S",
3354 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-ld128.S",
3355 "src/f32-gemm/gen/4x12-minmax-aarch64-neonfma-cortex-a53.S",
3356 "src/f32-gemm/gen/5x8-minmax-aarch64-neonfma-cortex-a57.S",
3357 "src/f32-gemm/gen/5x8-minmax-aarch64-neonfma-cortex-a75.S",
3358 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a53.S",
3359 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a55.S",
3360 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a57.S",
3361 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a73.S",
3362 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a75.S",
3363 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-ld64.S",
3364 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-ld128.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003365 "src/f32-igemm/1x8-minmax-aarch64-neonfma-cortex-a53.S",
Frank Barchard04336c12020-10-22 16:48:55 -07003366 "src/f32-igemm/1x12-minmax-aarch64-neonfma-cortex-a53.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003367 "src/f32-igemm/4x8-minmax-aarch64-neonfma-cortex-a53.S",
3368 "src/f32-igemm/4x8-minmax-aarch64-neonfma-cortex-a55.S",
Frank Barchard04336c12020-10-22 16:48:55 -07003369 "src/f32-igemm/4x12-minmax-aarch64-neonfma-cortex-a53.S",
3370 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a53.S",
3371 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a55.S",
3372 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a73.S",
3373 "src/f32-igemm/gen/1x8-minmax-aarch64-neonfma-cortex-a57.S",
3374 "src/f32-igemm/gen/1x8-minmax-aarch64-neonfma-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003375 "src/f32-igemm/gen/4x8-minmax-aarch64-neonfma-cortex-a57.S",
3376 "src/f32-igemm/gen/4x8-minmax-aarch64-neonfma-cortex-a75.S",
3377 "src/f32-igemm/gen/5x8-minmax-aarch64-neonfma-cortex-a57.S",
3378 "src/f32-igemm/gen/5x8-minmax-aarch64-neonfma-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003379 "src/f32-igemm/gen/6x8-minmax-aarch64-neonfma-cortex-a57.S",
3380 "src/f32-igemm/gen/6x8-minmax-aarch64-neonfma-cortex-a75.S",
Frank Barchardf1fd89e2020-10-06 23:27:11 -07003381 "src/qs8-gemm/1x16c4-aarch64-neondot-ld64.S",
Frank Barchardf2742c42020-10-12 21:16:46 -07003382 "src/qs8-gemm/4x16c4-aarch64-neondot-cortex-a55.S",
Frank Barcharda5237a52020-10-06 16:18:46 -07003383 "src/qs8-gemm/4x16c4-aarch64-neondot-ld64.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003384]
3385
Marat Dukhan1b354632020-03-23 12:50:22 -07003386INTERNAL_MICROKERNEL_HDRS = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003387 "src/xnnpack/argmaxpool.h",
3388 "src/xnnpack/avgpool.h",
3389 "src/xnnpack/clamp.h",
3390 "src/xnnpack/common.h",
3391 "src/xnnpack/conv.h",
Yury Kartynnike7841862020-11-04 18:22:18 -08003392 "src/xnnpack/depthtospace.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003393 "src/xnnpack/dwconv.h",
Frank Barchard04336c12020-10-22 16:48:55 -07003394 "src/xnnpack/fill.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003395 "src/xnnpack/gavgpool.h",
3396 "src/xnnpack/gemm.h",
3397 "src/xnnpack/hswish.h",
Marat Dukhan660fd192020-03-10 04:55:30 -07003398 "src/xnnpack/ibilinear.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003399 "src/xnnpack/igemm.h",
Marat Dukhancfb31342019-12-05 10:42:57 -08003400 "src/xnnpack/intrinsics-polyfill.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003401 "src/xnnpack/lut.h",
3402 "src/xnnpack/math.h",
3403 "src/xnnpack/maxpool.h",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003404 "src/xnnpack/memory.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003405 "src/xnnpack/packx.h",
3406 "src/xnnpack/pad.h",
3407 "src/xnnpack/params.h",
3408 "src/xnnpack/pavgpool.h",
3409 "src/xnnpack/ppmm.h",
3410 "src/xnnpack/prelu.h",
Marat Dukhan97579532019-10-18 16:40:39 -07003411 "src/xnnpack/raddexpminusmax.h",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07003412 "src/xnnpack/raddextexp.h",
Marat Dukhan97579532019-10-18 16:40:39 -07003413 "src/xnnpack/raddstoreexpminusmax.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003414 "src/xnnpack/rmax.h",
3415 "src/xnnpack/scalar-utils.h",
3416 "src/xnnpack/spmm.h",
3417 "src/xnnpack/unpool.h",
3418 "src/xnnpack/vadd.h",
Marat Dukhan1e782c42019-11-21 17:02:40 -08003419 "src/xnnpack/vbinary.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003420 "src/xnnpack/vmulcaddc.h",
Marat Dukhan05ac8e32019-10-21 15:39:33 -07003421 "src/xnnpack/vscale.h",
Marat Dukhan97579532019-10-18 16:40:39 -07003422 "src/xnnpack/vscaleexpminusmax.h",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07003423 "src/xnnpack/vscaleextexp.h",
Marat Dukhan1e782c42019-11-21 17:02:40 -08003424 "src/xnnpack/vunary.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003425 "src/xnnpack/zip.h",
Marat Dukhan1b354632020-03-23 12:50:22 -07003426]
3427
3428INTERNAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003429 "include/xnnpack.h",
3430 "src/xnnpack/allocator.h",
3431 "src/xnnpack/compute.h",
3432 "src/xnnpack/im2col.h",
3433 "src/xnnpack/indirection.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003434 "src/xnnpack/math-stubs.h",
Chao Mei6ddfc602020-05-13 22:29:36 -07003435 "src/xnnpack/memory-planner.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003436 "src/xnnpack/operator.h",
3437 "src/xnnpack/pack.h",
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07003438 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003439 "src/xnnpack/requantization-stubs.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003440 "src/xnnpack/requantization.h",
Marat Dukhan1d75a542020-02-03 12:23:01 -08003441 "src/xnnpack/subgraph.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003442]
3443
Marat Dukhan1b354632020-03-23 12:50:22 -07003444ACCURACY_EVAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003445 "src/xnnpack/math-stubs.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003446]
3447
Marat Dukhan1b354632020-03-23 12:50:22 -07003448MICROKERNEL_BENCHMARK_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003449 "include/xnnpack.h",
Frank Barchard35db7d02020-10-26 13:37:34 -07003450 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003451]
3452
Marat Dukhan1b354632020-03-23 12:50:22 -07003453MICROKERNEL_TEST_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Frank Barchard35db7d02020-10-26 13:37:34 -07003454 "include/xnnpack.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003455 "src/xnnpack/isa-checks.h",
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07003456 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003457 "src/xnnpack/requantization.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003458]
3459
3460OPERATOR_TEST_PARAMS_HDRS = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003461 "src/xnnpack/common.h",
Frank Barchard04336c12020-10-22 16:48:55 -07003462 "src/xnnpack/params.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003463]
3464
3465WEIGHTS_PACK_HDRS = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003466 "src/xnnpack/compute.h",
Frank Barchard04336c12020-10-22 16:48:55 -07003467 "src/xnnpack/operator.h",
3468 "src/xnnpack/pack.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003469]
3470
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07003471LOGGING_COPTS = select({
3472 # No logging in optimized mode
3473 ":optimized_build": ["-DXNN_LOG_LEVEL=0"],
3474 # Full logging in debug mode
3475 ":debug_build": ["-DXNN_LOG_LEVEL=5"],
3476 # Error-only logging in default (fastbuild) mode
3477 "//conditions:default": ["-DXNN_LOG_LEVEL=2"],
3478})
3479
Marat Dukhan3b59de22020-06-03 20:15:19 -07003480LOGGING_SRCS = select({
3481 # No logging in optimized mode
3482 ":optimized_build": [],
3483 "//conditions:default": [
3484 "src/operator-strings.c",
3485 "src/subgraph-strings.c",
3486 ],
3487})
3488
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07003489LOGGING_HDRS = [
3490 "src/xnnpack/log.h",
3491]
3492
Marat Dukhan08c4a432019-10-03 09:29:21 -07003493xnnpack_cc_library(
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003494 name = "tables",
3495 srcs = TABLE_SRCS,
3496 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003497 gcc_copts = xnnpack_gcc_std_copts(),
3498 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003499)
3500
3501xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003502 name = "scalar_ukernels",
3503 srcs = SCALAR_UKERNELS,
3504 hdrs = INTERNAL_HDRS,
3505 aarch32_copts = ["-marm"],
Marat Dukhan10a38082020-04-17 03:58:35 -07003506 gcc_copts = xnnpack_gcc_std_copts(),
3507 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003508 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003509 ":tables",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003510 "@FP16",
3511 "@FXdiv",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003512 "@pthreadpool",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003513 ],
3514)
3515
3516xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003517 name = "scalar_ukernels_test_mode",
3518 srcs = SCALAR_UKERNELS,
3519 hdrs = INTERNAL_HDRS,
3520 aarch32_copts = ["-marm"],
3521 copts = [
3522 "-UNDEBUG",
3523 "-DXNN_TEST_MODE=1",
3524 ],
3525 gcc_copts = xnnpack_gcc_std_copts(),
3526 msvc_copts = xnnpack_msvc_std_copts(),
3527 deps = [
3528 ":tables",
3529 "@FP16",
3530 "@FXdiv",
3531 "@pthreadpool",
3532 ],
3533)
3534
3535xnnpack_cc_library(
Marat Dukhan436ebe62019-12-04 15:10:12 -08003536 name = "wasm_ukernels",
3537 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003538 gcc_copts = xnnpack_gcc_std_copts(),
3539 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan436ebe62019-12-04 15:10:12 -08003540 wasm_srcs = WASM_UKERNELS,
Marat Dukhan1c6e3892020-06-25 23:56:10 -07003541 wasmsimd_srcs = WASM_UKERNELS + WASMSIMD_UKERNELS,
Marat Dukhan436ebe62019-12-04 15:10:12 -08003542 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003543 ":tables",
Marat Dukhan436ebe62019-12-04 15:10:12 -08003544 "@FP16",
3545 "@FXdiv",
3546 "@pthreadpool",
3547 ],
3548)
3549
3550xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003551 name = "wasm_ukernels_test_mode",
3552 hdrs = INTERNAL_HDRS,
3553 copts = [
3554 "-UNDEBUG",
3555 "-DXNN_TEST_MODE=1",
3556 ],
3557 gcc_copts = xnnpack_gcc_std_copts(),
3558 msvc_copts = xnnpack_msvc_std_copts(),
3559 wasm_srcs = WASM_UKERNELS,
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07003560 wasmsimd_srcs = WASM_UKERNELS + WASMSIMD_UKERNELS,
Marat Dukhan33fcf782020-05-24 14:27:15 -07003561 deps = [
3562 ":tables",
3563 "@FP16",
3564 "@FXdiv",
3565 "@pthreadpool",
3566 ],
3567)
3568
3569xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003570 name = "neon_ukernels",
3571 hdrs = INTERNAL_HDRS,
3572 aarch32_copts = [
3573 "-marm",
Marat Dukhan8853b822020-05-07 12:19:01 -07003574 "-march=armv7-a",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003575 "-mfpu=neon",
3576 ],
3577 aarch32_srcs = NEON_UKERNELS,
3578 aarch64_srcs = NEON_UKERNELS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003579 gcc_copts = xnnpack_gcc_std_copts(),
3580 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08003581 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003582 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003583 "@FP16",
3584 "@pthreadpool",
3585 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003586)
3587
3588xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003589 name = "neon_ukernels_test_mode",
3590 hdrs = INTERNAL_HDRS,
3591 aarch32_copts = [
3592 "-marm",
3593 "-march=armv7-a",
3594 "-mfpu=neon",
3595 ],
3596 aarch32_srcs = NEON_UKERNELS,
3597 aarch64_srcs = NEON_UKERNELS,
3598 copts = [
3599 "-UNDEBUG",
3600 "-DXNN_TEST_MODE=1",
3601 ],
3602 gcc_copts = xnnpack_gcc_std_copts(),
3603 msvc_copts = xnnpack_msvc_std_copts(),
3604 deps = [
3605 ":tables",
3606 "@FP16",
3607 "@pthreadpool",
3608 ],
3609)
3610
3611xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003612 name = "neonfma_ukernels",
3613 hdrs = INTERNAL_HDRS,
3614 aarch32_copts = [
3615 "-marm",
Marat Dukhan8853b822020-05-07 12:19:01 -07003616 "-march=armv7-a",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003617 "-mfpu=neon-vfpv4",
3618 ],
3619 aarch32_srcs = NEONFMA_UKERNELS,
3620 aarch64_srcs = NEONFMA_UKERNELS + AARCH64_NEONFMA_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07003621 apple_aarch32_copts = [
3622 "-mcpu=swift",
3623 "-mtune=generic",
3624 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07003625 gcc_copts = xnnpack_gcc_std_copts(),
3626 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08003627 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003628 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003629 "@FP16",
3630 "@pthreadpool",
3631 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003632)
3633
3634xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003635 name = "neonfma_ukernels_test_mode",
3636 hdrs = INTERNAL_HDRS,
3637 aarch32_copts = [
3638 "-marm",
3639 "-march=armv7-a",
3640 "-mfpu=neon-vfpv4",
3641 ],
3642 aarch32_srcs = NEONFMA_UKERNELS,
3643 aarch64_srcs = NEONFMA_UKERNELS + AARCH64_NEONFMA_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07003644 apple_aarch32_copts = [
3645 "-mcpu=swift",
3646 "-mtune=generic",
3647 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07003648 copts = [
3649 "-UNDEBUG",
3650 "-DXNN_TEST_MODE=1",
3651 ],
3652 gcc_copts = xnnpack_gcc_std_copts(),
3653 msvc_copts = xnnpack_msvc_std_copts(),
3654 deps = [
3655 ":tables",
3656 "@FP16",
3657 "@pthreadpool",
3658 ],
3659)
3660
3661xnnpack_cc_library(
Marat Dukhan8853b822020-05-07 12:19:01 -07003662 name = "neonv8_ukernels",
3663 hdrs = INTERNAL_HDRS,
3664 aarch32_copts = [
3665 "-marm",
3666 "-march=armv8-a",
3667 "-mfpu=neon-fp-armv8",
3668 ],
3669 aarch32_srcs = NEONV8_UKERNELS,
3670 aarch64_srcs = NEONV8_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07003671 apple_aarch32_copts = [
3672 "-mcpu=cyclone",
3673 "-mtune=generic",
3674 ],
Marat Dukhan8853b822020-05-07 12:19:01 -07003675 gcc_copts = xnnpack_gcc_std_copts(),
3676 msvc_copts = xnnpack_msvc_std_copts(),
3677 deps = [
3678 ":tables",
3679 "@FP16",
3680 "@pthreadpool",
3681 ],
3682)
3683
3684xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003685 name = "neonv8_ukernels_test_mode",
3686 hdrs = INTERNAL_HDRS,
3687 aarch32_copts = [
3688 "-marm",
3689 "-march=armv8-a",
3690 "-mfpu=neon-fp-armv8",
3691 ],
3692 aarch32_srcs = NEONV8_UKERNELS,
3693 aarch64_srcs = NEONV8_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07003694 apple_aarch32_copts = [
3695 "-mcpu=cyclone",
3696 "-mtune=generic",
3697 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07003698 copts = [
3699 "-UNDEBUG",
3700 "-DXNN_TEST_MODE=1",
3701 ],
3702 gcc_copts = xnnpack_gcc_std_copts(),
3703 msvc_copts = xnnpack_msvc_std_copts(),
3704 deps = [
3705 ":tables",
3706 "@FP16",
3707 "@pthreadpool",
3708 ],
3709)
3710
3711xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003712 name = "neonfp16arith_ukernels",
3713 hdrs = INTERNAL_HDRS,
3714 aarch64_copts = ["-march=armv8.2-a+fp16"],
3715 aarch64_srcs = AARCH64_NEONFP16ARITH_UKERNELS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003716 gcc_copts = xnnpack_gcc_std_copts(),
3717 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08003718 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003719 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003720 "@FP16",
3721 "@pthreadpool",
3722 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003723)
3724
3725xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003726 name = "neonfp16arith_ukernels_test_mode",
3727 hdrs = INTERNAL_HDRS,
3728 aarch64_copts = ["-march=armv8.2-a+fp16"],
3729 aarch64_srcs = AARCH64_NEONFP16ARITH_UKERNELS,
3730 copts = [
3731 "-UNDEBUG",
3732 "-DXNN_TEST_MODE=1",
3733 ],
3734 gcc_copts = xnnpack_gcc_std_copts(),
3735 msvc_copts = xnnpack_msvc_std_copts(),
3736 deps = [
3737 ":tables",
3738 "@FP16",
3739 "@pthreadpool",
3740 ],
3741)
3742
3743xnnpack_cc_library(
Benoit Jacoba9644732020-08-13 12:48:55 -07003744 name = "neondot_ukernels",
3745 hdrs = INTERNAL_HDRS,
Marat Dukhan799ac752020-08-13 16:08:03 -07003746 aarch32_copts = [
3747 "-marm",
3748 "-march=armv8.2-a+dotprod",
3749 "-mfpu=neon-fp-armv8",
3750 ],
Benoit Jacoba9644732020-08-13 12:48:55 -07003751 aarch32_srcs = NEONDOT_UKERNELS,
3752 aarch64_copts = ["-march=armv8.2-a+dotprod"],
3753 aarch64_srcs = NEONDOT_UKERNELS,
3754 gcc_copts = xnnpack_gcc_std_copts(),
3755 msvc_copts = xnnpack_msvc_std_copts(),
3756 deps = [
3757 ":tables",
3758 "@FP16",
3759 "@pthreadpool",
3760 ],
3761)
3762
3763xnnpack_cc_library(
3764 name = "neondot_ukernels_test_mode",
3765 hdrs = INTERNAL_HDRS,
Marat Dukhan799ac752020-08-13 16:08:03 -07003766 aarch32_copts = [
3767 "-marm",
3768 "-march=armv8.2-a+dotprod",
3769 "-mfpu=neon-fp-armv8",
3770 ],
Benoit Jacoba9644732020-08-13 12:48:55 -07003771 aarch32_srcs = NEONDOT_UKERNELS,
3772 aarch64_copts = ["-march=armv8.2-a+dotprod"],
3773 aarch64_srcs = NEONDOT_UKERNELS,
3774 copts = [
3775 "-UNDEBUG",
3776 "-DXNN_TEST_MODE=1",
3777 ],
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(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003788 name = "sse2_ukernels",
3789 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003790 gcc_copts = xnnpack_gcc_std_copts(),
3791 gcc_x86_copts = ["-msse2"],
3792 msvc_copts = xnnpack_msvc_std_copts(),
3793 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003794 x86_srcs = SSE_UKERNELS + SSE2_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08003795 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003796 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003797 "@FP16",
3798 "@pthreadpool",
3799 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003800)
3801
3802xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003803 name = "sse2_ukernels_test_mode",
3804 hdrs = INTERNAL_HDRS,
3805 copts = [
3806 "-UNDEBUG",
3807 "-DXNN_TEST_MODE=1",
3808 ],
3809 gcc_copts = xnnpack_gcc_std_copts(),
3810 gcc_x86_copts = ["-msse2"],
3811 msvc_copts = xnnpack_msvc_std_copts(),
3812 msvc_x86_32_copts = ["/arch:SSE2"],
3813 x86_srcs = SSE_UKERNELS + SSE2_UKERNELS,
3814 deps = [
3815 ":tables",
3816 "@FP16",
3817 "@pthreadpool",
3818 ],
3819)
3820
3821xnnpack_cc_library(
Marat Dukhanfe7acb62020-03-09 19:30:05 -07003822 name = "ssse3_ukernels",
3823 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003824 gcc_copts = xnnpack_gcc_std_copts(),
3825 gcc_x86_copts = ["-mssse3"],
3826 msvc_copts = xnnpack_msvc_std_copts(),
3827 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhanfe7acb62020-03-09 19:30:05 -07003828 x86_srcs = SSSE3_UKERNELS,
3829 deps = [
3830 ":tables",
3831 "@FP16",
3832 "@pthreadpool",
3833 ],
3834)
3835
3836xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003837 name = "ssse3_ukernels_test_mode",
3838 hdrs = INTERNAL_HDRS,
3839 copts = [
3840 "-UNDEBUG",
3841 "-DXNN_TEST_MODE=1",
3842 ],
3843 gcc_copts = xnnpack_gcc_std_copts(),
3844 gcc_x86_copts = ["-mssse3"],
3845 msvc_copts = xnnpack_msvc_std_copts(),
3846 msvc_x86_32_copts = ["/arch:SSE2"],
3847 x86_srcs = SSSE3_UKERNELS,
3848 deps = [
3849 ":tables",
3850 "@FP16",
3851 "@pthreadpool",
3852 ],
3853)
3854
3855xnnpack_cc_library(
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08003856 name = "sse41_ukernels",
3857 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003858 gcc_copts = xnnpack_gcc_std_copts(),
3859 gcc_x86_copts = ["-msse4.1"],
3860 msvc_copts = xnnpack_msvc_std_copts(),
3861 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08003862 x86_srcs = SSE41_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08003863 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003864 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003865 "@FP16",
3866 "@pthreadpool",
3867 ],
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08003868)
3869
3870xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003871 name = "sse41_ukernels_test_mode",
3872 hdrs = INTERNAL_HDRS,
3873 copts = [
3874 "-UNDEBUG",
3875 "-DXNN_TEST_MODE=1",
3876 ],
3877 gcc_copts = xnnpack_gcc_std_copts(),
3878 gcc_x86_copts = ["-msse4.1"],
3879 msvc_copts = xnnpack_msvc_std_copts(),
3880 msvc_x86_32_copts = ["/arch:SSE2"],
3881 x86_srcs = SSE41_UKERNELS,
3882 deps = [
3883 ":tables",
3884 "@FP16",
3885 "@pthreadpool",
3886 ],
3887)
3888
3889xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003890 name = "avx_ukernels",
3891 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003892 gcc_copts = xnnpack_gcc_std_copts(),
3893 gcc_x86_copts = ["-mavx"],
3894 msvc_copts = xnnpack_msvc_std_copts(),
3895 msvc_x86_32_copts = ["/arch:AVX"],
3896 msvc_x86_64_copts = ["/arch:AVX"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003897 x86_srcs = AVX_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08003898 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003899 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003900 "@FP16",
3901 "@pthreadpool",
3902 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003903)
3904
3905xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003906 name = "avx_ukernels_test_mode",
3907 hdrs = INTERNAL_HDRS,
3908 copts = [
3909 "-UNDEBUG",
3910 "-DXNN_TEST_MODE=1",
3911 ],
3912 gcc_copts = xnnpack_gcc_std_copts(),
3913 gcc_x86_copts = ["-mavx"],
3914 msvc_copts = xnnpack_msvc_std_copts(),
3915 msvc_x86_32_copts = ["/arch:AVX"],
3916 msvc_x86_64_copts = ["/arch:AVX"],
3917 x86_srcs = AVX_UKERNELS,
3918 deps = [
3919 ":tables",
3920 "@FP16",
3921 "@pthreadpool",
3922 ],
3923)
3924
3925xnnpack_cc_library(
Marat Dukhan1566fee2020-08-02 21:55:41 -07003926 name = "xop_ukernels",
3927 hdrs = INTERNAL_HDRS,
3928 gcc_copts = xnnpack_gcc_std_copts(),
3929 gcc_x86_copts = ["-mxop"],
3930 msvc_copts = xnnpack_msvc_std_copts(),
3931 msvc_x86_32_copts = ["/arch:AVX"],
3932 msvc_x86_64_copts = ["/arch:AVX"],
3933 x86_srcs = XOP_UKERNELS,
3934 deps = [
3935 ":tables",
3936 "@FP16",
3937 "@pthreadpool",
3938 ],
3939)
3940
3941xnnpack_cc_library(
3942 name = "xop_ukernels_test_mode",
3943 hdrs = INTERNAL_HDRS,
3944 copts = [
3945 "-UNDEBUG",
3946 "-DXNN_TEST_MODE=1",
3947 ],
3948 gcc_copts = xnnpack_gcc_std_copts(),
3949 gcc_x86_copts = ["-mxop"],
3950 msvc_copts = xnnpack_msvc_std_copts(),
3951 msvc_x86_32_copts = ["/arch:AVX"],
3952 msvc_x86_64_copts = ["/arch:AVX"],
3953 x86_srcs = XOP_UKERNELS,
3954 deps = [
3955 ":tables",
3956 "@FP16",
3957 "@pthreadpool",
3958 ],
3959)
3960
3961xnnpack_cc_library(
Marat Dukhanfda12b82019-11-21 12:27:59 -08003962 name = "fma3_ukernels",
3963 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003964 gcc_copts = xnnpack_gcc_std_copts(),
3965 gcc_x86_copts = ["-mfma"],
3966 msvc_copts = xnnpack_msvc_std_copts(),
3967 msvc_x86_32_copts = ["/arch:AVX"],
3968 msvc_x86_64_copts = ["/arch:AVX"],
Marat Dukhanfda12b82019-11-21 12:27:59 -08003969 x86_srcs = FMA3_UKERNELS,
3970 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003971 ":tables",
Marat Dukhanfda12b82019-11-21 12:27:59 -08003972 "@FP16",
3973 "@pthreadpool",
3974 ],
3975)
3976
3977xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003978 name = "fma3_ukernels_test_mode",
3979 hdrs = INTERNAL_HDRS,
3980 copts = [
3981 "-UNDEBUG",
3982 "-DXNN_TEST_MODE=1",
3983 ],
3984 gcc_copts = xnnpack_gcc_std_copts(),
3985 gcc_x86_copts = ["-mfma"],
3986 msvc_copts = xnnpack_msvc_std_copts(),
3987 msvc_x86_32_copts = ["/arch:AVX"],
3988 msvc_x86_64_copts = ["/arch:AVX"],
3989 x86_srcs = FMA3_UKERNELS,
3990 deps = [
3991 ":tables",
3992 "@FP16",
3993 "@pthreadpool",
3994 ],
3995)
3996
3997xnnpack_cc_library(
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003998 name = "avx2_ukernels",
3999 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004000 gcc_copts = xnnpack_gcc_std_copts(),
4001 gcc_x86_copts = [
Marat Dukhan6adff4e2019-10-14 18:32:07 -07004002 "-mfma",
4003 "-mavx2",
4004 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07004005 msvc_copts = xnnpack_msvc_std_copts(),
4006 msvc_x86_32_copts = ["/arch:AVX2"],
4007 msvc_x86_64_copts = ["/arch:AVX2"],
Marat Dukhan6adff4e2019-10-14 18:32:07 -07004008 x86_srcs = AVX2_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08004009 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08004010 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08004011 "@FP16",
4012 "@pthreadpool",
4013 ],
Marat Dukhan6adff4e2019-10-14 18:32:07 -07004014)
4015
4016xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07004017 name = "avx2_ukernels_test_mode",
4018 hdrs = INTERNAL_HDRS,
4019 copts = [
4020 "-UNDEBUG",
4021 "-DXNN_TEST_MODE=1",
4022 ],
4023 gcc_copts = xnnpack_gcc_std_copts(),
4024 gcc_x86_copts = [
4025 "-mfma",
4026 "-mavx2",
4027 ],
4028 msvc_copts = xnnpack_msvc_std_copts(),
4029 msvc_x86_32_copts = ["/arch:AVX2"],
4030 msvc_x86_64_copts = ["/arch:AVX2"],
4031 x86_srcs = AVX2_UKERNELS,
4032 deps = [
4033 ":tables",
4034 "@FP16",
4035 "@pthreadpool",
4036 ],
4037)
4038
4039xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004040 name = "avx512f_ukernels",
4041 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004042 gcc_copts = xnnpack_gcc_std_copts(),
4043 gcc_x86_copts = ["-mavx512f"],
4044 mingw_copts = ["-fno-asynchronous-unwind-tables"],
4045 msvc_copts = xnnpack_msvc_std_copts(),
4046 msvc_x86_32_copts = ["/arch:AVX512"],
4047 msvc_x86_64_copts = ["/arch:AVX512"],
4048 msys_copts = ["-fno-asynchronous-unwind-tables"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004049 x86_srcs = AVX512F_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08004050 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08004051 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08004052 "@FP16",
4053 "@pthreadpool",
4054 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004055)
4056
4057xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07004058 name = "avx512f_ukernels_test_mode",
4059 hdrs = INTERNAL_HDRS,
4060 copts = [
4061 "-UNDEBUG",
4062 "-DXNN_TEST_MODE=1",
4063 ],
4064 gcc_copts = xnnpack_gcc_std_copts(),
4065 gcc_x86_copts = ["-mavx512f"],
4066 mingw_copts = ["-fno-asynchronous-unwind-tables"],
4067 msvc_copts = xnnpack_msvc_std_copts(),
4068 msvc_x86_32_copts = ["/arch:AVX512"],
4069 msvc_x86_64_copts = ["/arch:AVX512"],
4070 msys_copts = ["-fno-asynchronous-unwind-tables"],
4071 x86_srcs = AVX512F_UKERNELS,
4072 deps = [
4073 ":tables",
4074 "@FP16",
4075 "@pthreadpool",
4076 ],
4077)
4078
4079xnnpack_cc_library(
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07004080 name = "avx512skx_ukernels",
4081 hdrs = INTERNAL_HDRS,
4082 gcc_copts = xnnpack_gcc_std_copts(),
4083 gcc_x86_copts = [
4084 "-mavx512f",
4085 "-mavx512cd",
4086 "-mavx512bw",
4087 "-mavx512dq",
4088 "-mavx512vl",
4089 ],
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 = AVX512SKX_UKERNELS,
4096 deps = [
4097 ":tables",
4098 "@FP16",
4099 "@pthreadpool",
4100 ],
4101)
4102
4103xnnpack_cc_library(
4104 name = "avx512skx_ukernels_test_mode",
4105 hdrs = INTERNAL_HDRS,
4106 copts = [
4107 "-UNDEBUG",
4108 "-DXNN_TEST_MODE=1",
4109 ],
4110 gcc_copts = xnnpack_gcc_std_copts(),
4111 gcc_x86_copts = [
4112 "-mavx512f",
4113 "-mavx512cd",
4114 "-mavx512bw",
4115 "-mavx512dq",
4116 "-mavx512vl",
4117 ],
4118 mingw_copts = ["-fno-asynchronous-unwind-tables"],
4119 msvc_copts = xnnpack_msvc_std_copts(),
4120 msvc_x86_32_copts = ["/arch:AVX512"],
4121 msvc_x86_64_copts = ["/arch:AVX512"],
4122 msys_copts = ["-fno-asynchronous-unwind-tables"],
4123 x86_srcs = AVX512SKX_UKERNELS,
4124 deps = [
4125 ":tables",
4126 "@FP16",
4127 "@pthreadpool",
4128 ],
4129)
4130
4131xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004132 name = "asm_ukernels",
4133 hdrs = ["src/xnnpack/assembly.h"],
4134 aarch32_srcs = AARCH32_ASM_UKERNELS,
Frank Barchard31bb45b2020-10-06 00:26:33 -07004135 aarch64_copts = ["-march=armv8.2-a+fp16+dotprod"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004136 aarch64_srcs = AARCH64_ASM_UKERNELS,
Frank Barchardbcedc082020-08-17 18:00:51 -07004137 wasm_srcs = WASM32_ASM_UKERNELS,
4138 wasmsimd_srcs = WASM32_ASM_UKERNELS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07004139)
4140
Marat Dukhan3b59de22020-06-03 20:15:19 -07004141xnnpack_cc_library(
4142 name = "logging_utils",
4143 srcs = LOGGING_SRCS,
4144 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
4145 copts = LOGGING_COPTS + [
4146 "-Isrc",
4147 "-Iinclude",
4148 ] + select({
4149 ":debug_build": [],
4150 "//conditions:default": xnnpack_min_size_copts(),
4151 }),
4152 gcc_copts = xnnpack_gcc_std_copts(),
4153 msvc_copts = xnnpack_msvc_std_copts(),
4154 visibility = xnnpack_visibility(),
4155 deps = [
4156 "@FP16",
4157 "@clog",
4158 "@pthreadpool",
4159 ],
4160)
4161
Marat Dukhan08c4a432019-10-03 09:29:21 -07004162xnnpack_aggregate_library(
4163 name = "ukernels",
4164 aarch32_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07004165 ":neon_ukernels",
4166 ":neonfma_ukernels",
Marat Dukhan8853b822020-05-07 12:19:01 -07004167 ":neonv8_ukernels",
Benoit Jacoba9644732020-08-13 12:48:55 -07004168 ":neondot_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004169 ":asm_ukernels",
4170 ],
4171 aarch64_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07004172 ":neon_ukernels",
4173 ":neonfma_ukernels",
Marat Dukhan8853b822020-05-07 12:19:01 -07004174 ":neonv8_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004175 ":neonfp16arith_ukernels",
Benoit Jacoba9644732020-08-13 12:48:55 -07004176 ":neondot_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004177 ":asm_ukernels",
4178 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07004179 generic_deps = [
4180 ":scalar_ukernels",
4181 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07004182 wasm_deps = [
4183 ":wasm_ukernels",
Frank Barchardbcedc082020-08-17 18:00:51 -07004184 ":asm_ukernels",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004185 ],
4186 wasmsimd_deps = [
4187 ":wasm_ukernels",
Frank Barchardbcedc082020-08-17 18:00:51 -07004188 ":asm_ukernels",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004189 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004190 x86_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07004191 ":sse2_ukernels",
Marat Dukhanfe7acb62020-03-09 19:30:05 -07004192 ":ssse3_ukernels",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08004193 ":sse41_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004194 ":avx_ukernels",
Marat Dukhan1566fee2020-08-02 21:55:41 -07004195 ":xop_ukernels",
Marat Dukhanfda12b82019-11-21 12:27:59 -08004196 ":fma3_ukernels",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07004197 ":avx2_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004198 ":avx512f_ukernels",
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07004199 ":avx512skx_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004200 ],
4201)
4202
Marat Dukhan33fcf782020-05-24 14:27:15 -07004203xnnpack_aggregate_library(
4204 name = "ukernels_test_mode",
4205 aarch32_deps = [
4206 ":neon_ukernels_test_mode",
4207 ":neonfma_ukernels_test_mode",
4208 ":neonv8_ukernels_test_mode",
Benoit Jacoba9644732020-08-13 12:48:55 -07004209 ":neondot_ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004210 ":asm_ukernels",
4211 ],
4212 aarch64_deps = [
4213 ":neon_ukernels_test_mode",
4214 ":neonfma_ukernels_test_mode",
4215 ":neonv8_ukernels_test_mode",
4216 ":neonfp16arith_ukernels_test_mode",
Benoit Jacoba9644732020-08-13 12:48:55 -07004217 ":neondot_ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004218 ":asm_ukernels",
4219 ],
4220 generic_deps = [
4221 ":scalar_ukernels_test_mode",
4222 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07004223 wasm_deps = [
4224 ":wasm_ukernels_test_mode",
Frank Barchardbcedc082020-08-17 18:00:51 -07004225 ":asm_ukernels",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004226 ],
4227 wasmsimd_deps = [
4228 ":wasm_ukernels_test_mode",
Frank Barchardbcedc082020-08-17 18:00:51 -07004229 ":asm_ukernels",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004230 ],
4231 x86_deps = [
4232 ":sse2_ukernels_test_mode",
4233 ":ssse3_ukernels_test_mode",
4234 ":sse41_ukernels_test_mode",
4235 ":avx_ukernels_test_mode",
Marat Dukhan1566fee2020-08-02 21:55:41 -07004236 ":xop_ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004237 ":fma3_ukernels_test_mode",
4238 ":avx2_ukernels_test_mode",
4239 ":avx512f_ukernels_test_mode",
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07004240 ":avx512skx_ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004241 ],
4242)
4243
Marat Dukhan86df3842020-06-17 14:03:31 -07004244# This target must be compatible with Android NDK r17 and compile with Clang 6.
4245# It is similar to :ukernels target, but excludes microkernels which require
4246# NEON FP16 arithmetics, NEON Dot Product, and newer intrinsics.
Marat Dukhanae046f52020-06-15 13:16:14 -07004247xnnpack_aggregate_library(
Marat Dukhan86df3842020-06-17 14:03:31 -07004248 name = "ukernels_max_compatibility",
Marat Dukhanae046f52020-06-15 13:16:14 -07004249 aarch32_deps = [
4250 ":neon_ukernels",
4251 ":neonfma_ukernels",
4252 ":neonv8_ukernels",
4253 ":asm_ukernels",
4254 ],
4255 aarch64_deps = [
4256 ":neon_ukernels",
4257 ":neonfma_ukernels",
4258 ":neonv8_ukernels",
4259 ":asm_ukernels",
4260 ],
4261 generic_deps = [
4262 ":scalar_ukernels",
4263 ],
Marat Dukhanae046f52020-06-15 13:16:14 -07004264 wasm_deps = [
4265 ":wasm_ukernels",
4266 ],
4267 wasmsimd_deps = [
4268 ":wasm_ukernels",
4269 ],
4270 x86_deps = [
4271 ":sse2_ukernels",
4272 ":ssse3_ukernels",
4273 ":sse41_ukernels",
4274 ":avx_ukernels",
Marat Dukhan1566fee2020-08-02 21:55:41 -07004275 ":xop_ukernels",
Marat Dukhanae046f52020-06-15 13:16:14 -07004276 ":fma3_ukernels",
4277 ":avx2_ukernels",
4278 ":avx512f_ukernels",
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07004279 ":avx512skx_ukernels",
Marat Dukhanae046f52020-06-15 13:16:14 -07004280 ],
4281)
4282
Marat Dukhan08c4a432019-10-03 09:29:21 -07004283xnnpack_cc_library(
4284 name = "im2col",
4285 srcs = ["src/im2col.c"],
4286 hdrs = [
4287 "src/xnnpack/common.h",
4288 "src/xnnpack/im2col.h",
4289 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07004290 gcc_copts = xnnpack_gcc_std_copts(),
4291 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004292)
4293
4294xnnpack_cc_library(
4295 name = "indirection",
4296 srcs = ["src/indirection.c"],
4297 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004298 gcc_copts = xnnpack_gcc_std_copts(),
4299 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004300 deps = [
4301 "@FP16",
4302 "@FXdiv",
4303 "@pthreadpool",
4304 ],
4305)
4306
4307xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07004308 name = "indirection_test_mode",
4309 srcs = ["src/indirection.c"],
4310 hdrs = INTERNAL_HDRS,
4311 copts = [
4312 "-UNDEBUG",
4313 "-DXNN_TEST_MODE=1",
4314 ],
4315 gcc_copts = xnnpack_gcc_std_copts(),
4316 msvc_copts = xnnpack_msvc_std_copts(),
4317 deps = [
4318 "@FP16",
4319 "@FXdiv",
4320 "@pthreadpool",
4321 ],
4322)
4323
4324xnnpack_cc_library(
Marat Dukhanab582382020-07-06 13:32:08 -07004325 name = "packing",
4326 srcs = ["src/packing.c"],
4327 hdrs = INTERNAL_HDRS,
4328 gcc_copts = xnnpack_gcc_std_copts(),
4329 msvc_copts = xnnpack_msvc_std_copts(),
4330 deps = [
4331 "@FP16",
4332 "@FXdiv",
4333 "@pthreadpool",
4334 ],
4335)
4336
4337xnnpack_cc_library(
4338 name = "packing_test_mode",
4339 srcs = ["src/packing.c"],
4340 hdrs = INTERNAL_HDRS,
4341 copts = [
4342 "-UNDEBUG",
4343 "-DXNN_TEST_MODE=1",
4344 ],
4345 gcc_copts = xnnpack_gcc_std_copts(),
4346 msvc_copts = xnnpack_msvc_std_copts(),
4347 deps = [
4348 "@FP16",
4349 "@FXdiv",
4350 "@pthreadpool",
4351 ],
4352)
4353
4354xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004355 name = "operator_run",
4356 srcs = ["src/operator-run.c"],
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07004357 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004358 copts = LOGGING_COPTS + select({
Marat Dukhan05702cf2020-03-26 15:41:33 -07004359 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4360 "//conditions:default": [],
4361 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07004362 gcc_copts = xnnpack_gcc_std_copts(),
4363 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004364 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07004365 ":logging_utils",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004366 "@FP16",
4367 "@FXdiv",
4368 "@clog",
4369 "@pthreadpool",
4370 ],
4371)
4372
Chao Mei6ddfc602020-05-13 22:29:36 -07004373xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07004374 name = "operator_run_test_mode",
4375 srcs = ["src/operator-run.c"],
4376 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
4377 copts = LOGGING_COPTS + [
4378 "-UNDEBUG",
4379 "-DXNN_TEST_MODE=1",
4380 ] + select({
4381 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4382 "//conditions:default": [],
4383 }),
4384 gcc_copts = xnnpack_gcc_std_copts(),
4385 msvc_copts = xnnpack_msvc_std_copts(),
4386 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07004387 ":logging_utils",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004388 "@FP16",
4389 "@FXdiv",
4390 "@clog",
4391 "@pthreadpool",
4392 ],
4393)
4394
4395xnnpack_cc_library(
Chao Mei6ddfc602020-05-13 22:29:36 -07004396 name = "memory_planner",
4397 srcs = ["src/memory-planner.c"],
4398 hdrs = INTERNAL_HDRS,
4399 defines = select({
4400 ":xnn_enable_memopt_explicit_true": ["XNN_ENABLE_MEMOPT=1"],
4401 ":xnn_enable_memopt_explicit_false": ["XNN_ENABLE_MEMOPT=0"],
4402 "//conditions:default": ["XNN_ENABLE_MEMOPT=1"],
4403 }),
4404 gcc_copts = xnnpack_gcc_std_copts(),
4405 msvc_copts = xnnpack_msvc_std_copts(),
4406 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07004407 ":logging_utils",
Chao Mei6ddfc602020-05-13 22:29:36 -07004408 "@pthreadpool",
4409 ],
4410)
4411
Marat Dukhan33fcf782020-05-24 14:27:15 -07004412xnnpack_cc_library(
4413 name = "memory_planner_test_mode",
4414 srcs = ["src/memory-planner.c"],
4415 hdrs = INTERNAL_HDRS,
4416 copts = [
4417 "-UNDEBUG",
4418 "-DXNN_TEST_MODE=1",
4419 ],
4420 defines = select({
4421 ":xnn_enable_memopt_explicit_true": ["XNN_ENABLE_MEMOPT=1"],
4422 ":xnn_enable_memopt_explicit_false": ["XNN_ENABLE_MEMOPT=0"],
4423 "//conditions:default": ["XNN_ENABLE_MEMOPT=1"],
4424 }),
4425 gcc_copts = xnnpack_gcc_std_copts(),
4426 msvc_copts = xnnpack_msvc_std_copts(),
4427 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07004428 ":logging_utils",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004429 "@pthreadpool",
4430 ],
4431)
4432
Marat Dukhan08c4a432019-10-03 09:29:21 -07004433cc_library(
4434 name = "enable_assembly",
4435 defines = select({
4436 ":xnn_enable_assembly_explicit_true": ["XNN_ENABLE_ASSEMBLY=1"],
4437 ":xnn_enable_assembly_explicit_false": ["XNN_ENABLE_ASSEMBLY=0"],
Frank Barchard810171d2019-10-10 10:34:51 -07004438 "//conditions:default": ["XNN_ENABLE_ASSEMBLY=1"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004439 }),
4440)
4441
Marat Dukhan9de90e02020-06-18 16:04:12 -07004442cc_library(
4443 name = "enable_sparse",
4444 defines = select({
4445 ":xnn_enable_sparse_explicit_true": ["XNN_ENABLE_SPARSE=1"],
4446 ":xnn_enable_sparse_explicit_false": ["XNN_ENABLE_SPARSE=0"],
Marat Dukhanb36582b2020-12-08 11:16:28 -08004447 "//conditions:default": ["XNN_ENABLE_SPARSE=1"],
Marat Dukhan9de90e02020-06-18 16:04:12 -07004448 }),
4449)
4450
Marat Dukhancf056b22019-10-07 10:26:29 -07004451xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004452 name = "operators",
4453 srcs = OPERATOR_SRCS + [
Marat Dukhan04f03be2019-11-19 12:36:47 -08004454 "src/memory.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004455 "src/operator-delete.c",
Marat Dukhancf056b22019-10-07 10:26:29 -07004456 ],
4457 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004458 copts = LOGGING_COPTS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07004459 "-Isrc",
4460 "-Iinclude",
4461 ] + select({
4462 ":debug_build": [],
4463 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07004464 }) + select({
4465 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4466 "//conditions:default": [],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004467 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07004468 gcc_copts = xnnpack_gcc_std_copts(),
4469 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004470 deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07004471 ":indirection",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004472 ":logging_utils",
Marat Dukhanab582382020-07-06 13:32:08 -07004473 ":packing",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004474 "@FP16",
4475 "@FXdiv",
4476 "@clog",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004477 "@pthreadpool",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004478 ],
4479)
4480
Marat Dukhan10a38082020-04-17 03:58:35 -07004481xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07004482 name = "operators_test_mode",
4483 srcs = OPERATOR_SRCS + [
4484 "src/memory.c",
4485 "src/operator-delete.c",
4486 ],
4487 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
4488 copts = LOGGING_COPTS + [
4489 "-Isrc",
4490 "-Iinclude",
4491 "-UNDEBUG",
4492 "-DXNN_TEST_MODE=1",
4493 ] + select({
4494 ":debug_build": [],
4495 "//conditions:default": xnnpack_min_size_copts(),
4496 }) + select({
4497 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4498 "//conditions:default": [],
4499 }),
4500 gcc_copts = xnnpack_gcc_std_copts(),
4501 msvc_copts = xnnpack_msvc_std_copts(),
4502 deps = [
4503 ":indirection_test_mode",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004504 ":logging_utils",
Marat Dukhanab582382020-07-06 13:32:08 -07004505 ":packing_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004506 "@FP16",
4507 "@FXdiv",
4508 "@clog",
4509 "@pthreadpool",
4510 ],
4511)
4512
4513xnnpack_cc_library(
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004514 name = "XNNPACK",
4515 srcs = [
4516 "src/init.c",
Marat Dukhanccfdbd12020-02-03 14:27:45 -08004517 "src/runtime.c",
4518 "src/subgraph.c",
4519 "src/tensor.c",
Marat Dukhanf03da0d2020-06-10 16:00:20 -07004520 ] + SUBGRAPH_SRCS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004521 hdrs = ["include/xnnpack.h"],
4522 copts = LOGGING_COPTS + [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004523 "-Isrc",
4524 "-Iinclude",
4525 ] + select({
4526 ":debug_build": [],
4527 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07004528 }) + select({
4529 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4530 "//conditions:default": [],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004531 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07004532 gcc_copts = xnnpack_gcc_std_copts(),
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004533 includes = ["include"],
Marat Dukhan10a38082020-04-17 03:58:35 -07004534 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004535 visibility = xnnpack_visibility(),
4536 deps = [
4537 ":enable_assembly",
Marat Dukhan9de90e02020-06-18 16:04:12 -07004538 ":enable_sparse",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004539 ":logging_utils",
Chao Mei6ddfc602020-05-13 22:29:36 -07004540 ":memory_planner",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004541 ":operator_run",
4542 ":operators",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004543 ":ukernels",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004544 "@clog",
Marat Dukhanab2946c2020-05-21 20:04:13 -07004545 "@FP16",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004546 "@pthreadpool",
Marat Dukhand343c222019-10-07 09:22:14 -07004547 ] + select({
4548 ":emscripten": [],
4549 "//conditions:default": ["@cpuinfo"],
4550 }),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004551)
4552
Marat Dukhan10a38082020-04-17 03:58:35 -07004553xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07004554 name = "XNNPACK_test_mode",
4555 srcs = [
4556 "src/init.c",
4557 "src/runtime.c",
4558 "src/subgraph.c",
4559 "src/tensor.c",
Marat Dukhanf03da0d2020-06-10 16:00:20 -07004560 ] + SUBGRAPH_SRCS,
Marat Dukhan33fcf782020-05-24 14:27:15 -07004561 hdrs = ["include/xnnpack.h"],
4562 copts = LOGGING_COPTS + [
4563 "-Isrc",
4564 "-Iinclude",
4565 "-UNDEBUG",
4566 "-DXNN_TEST_MODE=1",
4567 ] + select({
4568 ":debug_build": [],
4569 "//conditions:default": xnnpack_min_size_copts(),
4570 }) + select({
4571 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4572 "//conditions:default": [],
4573 }),
4574 gcc_copts = xnnpack_gcc_std_copts(),
4575 includes = ["include"],
4576 msvc_copts = xnnpack_msvc_std_copts(),
4577 visibility = xnnpack_visibility(),
4578 deps = [
4579 ":enable_assembly",
Marat Dukhan9de90e02020-06-18 16:04:12 -07004580 ":enable_sparse",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004581 ":logging_utils",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004582 ":memory_planner_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004583 ":operator_run_test_mode",
4584 ":operators_test_mode",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004585 ":ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004586 "@clog",
4587 "@FP16",
4588 "@pthreadpool",
4589 ] + select({
4590 ":emscripten": [],
4591 "//conditions:default": ["@cpuinfo"],
4592 }),
4593)
4594
4595xnnpack_cc_library(
Marat Dukhanae046f52020-06-15 13:16:14 -07004596 name = "xnnpack_f32",
4597 srcs = [
4598 "src/init.c",
4599 "src/runtime.c",
4600 "src/subgraph.c",
4601 "src/tensor.c",
4602 ] + SUBGRAPH_SRCS,
4603 hdrs = ["include/xnnpack.h"],
4604 copts = LOGGING_COPTS + [
4605 "-Isrc",
4606 "-Iinclude",
4607 ] + select({
4608 ":debug_build": [],
4609 "//conditions:default": xnnpack_min_size_copts(),
4610 }) + select({
4611 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4612 "//conditions:default": [],
4613 }),
4614 defines = [
Marat Dukhan16f1e1a2020-08-04 16:38:22 -07004615 "XNN_NO_QS8_OPERATORS",
Marat Dukhan08b7a972020-07-14 18:17:29 -07004616 "XNN_NO_QU8_OPERATORS",
Marat Dukhanae046f52020-06-15 13:16:14 -07004617 "XNN_NO_U8_OPERATORS",
4618 "XNN_NO_X8_OPERATORS",
4619 "XNN_NO_F16_OPERATORS",
4620 "XNN_NO_X16_OPERATORS",
4621 ],
4622 gcc_copts = xnnpack_gcc_std_copts(),
4623 includes = ["include"],
4624 msvc_copts = xnnpack_msvc_std_copts(),
4625 visibility = xnnpack_visibility(),
4626 deps = [
4627 ":enable_assembly",
Marat Dukhan9de90e02020-06-18 16:04:12 -07004628 ":enable_sparse",
Marat Dukhanae046f52020-06-15 13:16:14 -07004629 ":logging_utils",
4630 ":memory_planner",
4631 ":operator_run",
4632 ":operators",
Marat Dukhan86df3842020-06-17 14:03:31 -07004633 ":ukernels_max_compatibility",
Marat Dukhanae046f52020-06-15 13:16:14 -07004634 "@clog",
4635 "@FP16",
4636 "@pthreadpool",
4637 ] + select({
4638 ":emscripten": [],
4639 "//conditions:default": ["@cpuinfo"],
4640 }),
4641)
4642
4643xnnpack_cc_library(
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004644 name = "xnnpack_operators_nhwc_f32",
4645 srcs = [
4646 "src/init.c",
4647 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07004648 hdrs = ["include/xnnpack.h"],
4649 copts = LOGGING_COPTS + [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004650 "-Isrc",
4651 "-Iinclude",
4652 ] + select({
4653 ":debug_build": [],
4654 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07004655 }) + select({
4656 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4657 "//conditions:default": [],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004658 }),
4659 defines = [
Marat Dukhan16f1e1a2020-08-04 16:38:22 -07004660 "XNN_NO_QS8_OPERATORS",
Marat Dukhan08b7a972020-07-14 18:17:29 -07004661 "XNN_NO_QU8_OPERATORS",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004662 "XNN_NO_U8_OPERATORS",
4663 "XNN_NO_X8_OPERATORS",
Marat Dukhanefc47b82019-11-18 09:25:38 -08004664 "XNN_NO_NCHW_OPERATORS",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004665 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07004666 gcc_copts = xnnpack_gcc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004667 includes = ["include"],
Marat Dukhan10a38082020-04-17 03:58:35 -07004668 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004669 visibility = xnnpack_visibility(),
4670 deps = [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004671 ":enable_assembly",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004672 ":logging_utils",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004673 ":operator_run",
4674 ":operators",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004675 ":ukernels",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004676 "@clog",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004677 "@pthreadpool",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004678 ] + select({
4679 ":emscripten": [],
4680 "//conditions:default": ["@cpuinfo"],
4681 }),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004682)
4683
Marat Dukhancf056b22019-10-07 10:26:29 -07004684xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004685 name = "bench_utils",
4686 srcs = ["bench/utils.cc"],
4687 hdrs = ["bench/utils.h"],
Marat Dukhanbad48fe2019-11-04 10:35:22 -08004688 deps = [
4689 "@com_google_benchmark//:benchmark",
4690 "@cpuinfo",
4691 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004692)
4693
Frank Barchard7e955972019-10-11 10:34:25 -07004694######################### Benchmarks for micro-kernels #########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07004695
4696xnnpack_benchmark(
Marat Dukhan595e1702020-07-31 10:12:52 -07004697 name = "qs8_gemm_bench",
4698 srcs = [
4699 "bench/gemm.h",
4700 "bench/qs8-gemm.cc",
4701 "src/xnnpack/AlignedAllocator.h",
4702 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Frank Barchard31328cb2020-10-12 11:55:18 -07004703 copts = xnnpack_optional_ruy_copts() + xnnpack_optional_gemmlowp_copts(),
4704 deps = MICROKERNEL_BENCHMARK_DEPS + [":packing"] + xnnpack_optional_ruy_deps() + xnnpack_optional_gemmlowp_deps(),
Marat Dukhan595e1702020-07-31 10:12:52 -07004705)
4706
4707xnnpack_benchmark(
Marat Dukhan56bdd4a2020-08-03 19:47:04 -07004708 name = "qs8_requantization_bench",
4709 srcs = [
4710 "bench/qs8-requantization.cc",
Marat Dukhan56bdd4a2020-08-03 19:47:04 -07004711 "src/xnnpack/AlignedAllocator.h",
Frank Barchard04336c12020-10-22 16:48:55 -07004712 "src/xnnpack/requantization-stubs.h",
Marat Dukhan56bdd4a2020-08-03 19:47:04 -07004713 ] + MICROKERNEL_BENCHMARK_HDRS,
4714 deps = MICROKERNEL_BENCHMARK_DEPS,
4715)
4716
4717xnnpack_benchmark(
Marat Dukhan08b7a972020-07-14 18:17:29 -07004718 name = "qu8_gemm_bench",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004719 srcs = [
4720 "bench/gemm.h",
Marat Dukhan08b7a972020-07-14 18:17:29 -07004721 "bench/qu8-gemm.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004722 "src/xnnpack/AlignedAllocator.h",
4723 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004724 copts = xnnpack_optional_ruy_copts() + xnnpack_optional_gemmlowp_copts(),
Marat Dukhanab582382020-07-06 13:32:08 -07004725 deps = MICROKERNEL_BENCHMARK_DEPS + [":packing"] + xnnpack_optional_ruy_deps() + xnnpack_optional_gemmlowp_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004726)
4727
4728xnnpack_benchmark(
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07004729 name = "qu8_requantization_bench",
4730 srcs = [
4731 "bench/qu8-requantization.cc",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07004732 "src/xnnpack/AlignedAllocator.h",
Frank Barchard04336c12020-10-22 16:48:55 -07004733 "src/xnnpack/requantization-stubs.h",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07004734 ] + MICROKERNEL_BENCHMARK_HDRS,
4735 deps = MICROKERNEL_BENCHMARK_DEPS,
4736)
4737
4738xnnpack_benchmark(
Frank Barchard40d20fe2020-05-05 00:37:45 -07004739 name = "f16_igemm_bench",
4740 srcs = [
4741 "bench/f16-igemm.cc",
4742 "bench/conv.h",
4743 "bench/google/conv.h",
4744 "src/xnnpack/AlignedAllocator.h",
4745 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004746 deps = MICROKERNEL_BENCHMARK_DEPS + [
4747 ":indirection",
4748 ":packing",
4749 ],
Frank Barchard40d20fe2020-05-05 00:37:45 -07004750)
4751
4752xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004753 name = "f16_gemm_bench",
4754 srcs = [
4755 "bench/f16-gemm.cc",
4756 "bench/gemm.h",
4757 "src/xnnpack/AlignedAllocator.h",
4758 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004759 deps = MICROKERNEL_BENCHMARK_DEPS + [
4760 ":packing",
4761 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004762)
4763
4764xnnpack_benchmark(
Marat Dukhanbdb56f52020-02-05 21:42:49 -08004765 name = "f16_spmm_bench",
4766 srcs = [
4767 "bench/f16-spmm.cc",
Marat Dukhan1631e3e2020-12-06 19:29:31 -08004768 "bench/spmm.h",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08004769 "src/xnnpack/AlignedAllocator.h",
4770 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanbdb56f52020-02-05 21:42:49 -08004771 deps = MICROKERNEL_BENCHMARK_DEPS,
4772)
4773
4774xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004775 name = "f32_igemm_bench",
4776 srcs = [
4777 "bench/f32-igemm.cc",
4778 "bench/conv.h",
4779 "src/xnnpack/AlignedAllocator.h",
4780 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004781 deps = MICROKERNEL_BENCHMARK_DEPS + [
4782 ":indirection",
4783 ":packing",
4784 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004785)
4786
4787xnnpack_benchmark(
Frank Barchardc2010062020-07-19 10:04:12 -07004788 name = "f16_relu_bench",
4789 srcs = [
4790 "bench/f16-relu.cc",
4791 "src/xnnpack/AlignedAllocator.h",
4792 ] + MICROKERNEL_BENCHMARK_HDRS,
4793 deps = MICROKERNEL_BENCHMARK_DEPS,
4794)
4795
4796xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004797 name = "f32_conv_hwc_bench",
4798 srcs = [
4799 "bench/f32-conv-hwc.cc",
4800 "bench/dconv.h",
4801 "src/xnnpack/AlignedAllocator.h",
4802 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004803 deps = MICROKERNEL_BENCHMARK_DEPS + [
4804 ":packing",
4805 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004806)
4807
4808xnnpack_benchmark(
Marat Dukhan1f29b802020-05-15 23:46:39 -07004809 name = "f32_conv_hwc2chw_bench",
Erich Elsen563df5f2019-10-23 08:02:21 -07004810 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07004811 "bench/f32-conv-hwc2chw.cc",
Erich Elsen563df5f2019-10-23 08:02:21 -07004812 "bench/dconv.h",
4813 "src/xnnpack/AlignedAllocator.h",
4814 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004815 deps = MICROKERNEL_BENCHMARK_DEPS + [
4816 ":packing",
4817 ],
Erich Elsen563df5f2019-10-23 08:02:21 -07004818)
4819
4820xnnpack_benchmark(
Frank Barchard5a599a62020-06-04 20:12:44 -07004821 name = "f16_dwconv_bench",
4822 srcs = [
4823 "bench/f16-dwconv.cc",
4824 "bench/dwconv.h",
4825 "bench/google/dwconv.h",
4826 "src/xnnpack/AlignedAllocator.h",
4827 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004828 deps = MICROKERNEL_BENCHMARK_DEPS + [
4829 ":indirection",
4830 ":packing",
4831 ],
Frank Barchard5a599a62020-06-04 20:12:44 -07004832)
4833
4834xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004835 name = "f32_dwconv_bench",
4836 srcs = [
4837 "bench/f32-dwconv.cc",
4838 "bench/dwconv.h",
4839 "src/xnnpack/AlignedAllocator.h",
4840 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004841 deps = MICROKERNEL_BENCHMARK_DEPS + [
4842 ":indirection",
4843 ":packing",
4844 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004845)
4846
4847xnnpack_benchmark(
Marat Dukhanbf715f92020-10-23 20:17:00 -07004848 name = "f32_dwconv2d_chw_bench",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004849 srcs = [
Marat Dukhanbf715f92020-10-23 20:17:00 -07004850 "bench/f32-dwconv2d-chw.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004851 "bench/dwconv.h",
4852 "src/xnnpack/AlignedAllocator.h",
4853 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004854 deps = MICROKERNEL_BENCHMARK_DEPS + [
4855 ":indirection",
4856 ":packing",
4857 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004858)
4859
4860xnnpack_benchmark(
4861 name = "f32_gemm_bench",
4862 srcs = [
4863 "bench/f32-gemm.cc",
4864 "bench/gemm.h",
4865 "src/xnnpack/AlignedAllocator.h",
4866 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004867 copts = xnnpack_optional_ruy_copts(),
Marat Dukhanab582382020-07-06 13:32:08 -07004868 deps = MICROKERNEL_BENCHMARK_DEPS + [":packing"] + xnnpack_optional_ruy_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004869)
4870
4871xnnpack_benchmark(
Marat Dukhana11ca342020-06-25 23:45:07 -07004872 name = "f32_hswish_bench",
4873 srcs = [
4874 "bench/f32-hswish.cc",
4875 "src/xnnpack/AlignedAllocator.h",
4876 ] + MICROKERNEL_BENCHMARK_HDRS,
4877 deps = MICROKERNEL_BENCHMARK_DEPS,
4878)
4879
4880xnnpack_benchmark(
Marat Dukhan4c4eb002019-12-08 21:27:49 -08004881 name = "f32_raddexpminusmax_bench",
4882 srcs = [
4883 "bench/f32-raddexpminusmax.cc",
4884 "src/xnnpack/AlignedAllocator.h",
4885 ] + MICROKERNEL_BENCHMARK_HDRS,
4886 deps = MICROKERNEL_BENCHMARK_DEPS,
4887)
4888
4889xnnpack_benchmark(
4890 name = "f32_raddextexp_bench",
4891 srcs = [
4892 "bench/f32-raddextexp.cc",
4893 "src/xnnpack/AlignedAllocator.h",
4894 ] + MICROKERNEL_BENCHMARK_HDRS,
4895 deps = MICROKERNEL_BENCHMARK_DEPS,
4896)
4897
4898xnnpack_benchmark(
4899 name = "f32_raddstoreexpminusmax_bench",
4900 srcs = [
4901 "bench/f32-raddstoreexpminusmax.cc",
4902 "src/xnnpack/AlignedAllocator.h",
4903 ] + MICROKERNEL_BENCHMARK_HDRS,
4904 deps = MICROKERNEL_BENCHMARK_DEPS,
4905)
4906
4907xnnpack_benchmark(
Frank Barchard3c4a9522020-07-18 16:11:22 -07004908 name = "f32_relu_bench",
4909 srcs = [
4910 "bench/f32-relu.cc",
4911 "src/xnnpack/AlignedAllocator.h",
4912 ] + MICROKERNEL_BENCHMARK_HDRS,
4913 deps = MICROKERNEL_BENCHMARK_DEPS,
4914)
4915
4916xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004917 name = "f32_rmax_bench",
4918 srcs = [
4919 "bench/f32-rmax.cc",
4920 "src/xnnpack/AlignedAllocator.h",
4921 ] + MICROKERNEL_BENCHMARK_HDRS,
4922 deps = MICROKERNEL_BENCHMARK_DEPS,
4923)
4924
4925xnnpack_benchmark(
Marat Dukhan14bec502019-11-18 11:35:31 -08004926 name = "f32_sigmoid_bench",
4927 srcs = [
4928 "bench/f32-sigmoid.cc",
4929 "src/xnnpack/AlignedAllocator.h",
4930 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan14bec502019-11-18 11:35:31 -08004931 deps = MICROKERNEL_BENCHMARK_DEPS,
4932)
4933
4934xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004935 name = "f32_spmm_bench",
4936 srcs = [
4937 "bench/f32-spmm.cc",
Marat Dukhan1631e3e2020-12-06 19:29:31 -08004938 "bench/spmm.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004939 "src/xnnpack/AlignedAllocator.h",
4940 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07004941 deps = MICROKERNEL_BENCHMARK_DEPS,
4942)
4943
4944xnnpack_benchmark(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08004945 name = "f32_softmax_bench",
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07004946 srcs = [
Marat Dukhanfd8e6892020-01-27 15:25:25 -08004947 "bench/f32-softmax.cc",
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07004948 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004949 copts = xnnpack_optional_dnnl_copts(),
Marat Dukhan8d3c6932020-03-06 20:27:27 -08004950 deps = MICROKERNEL_BENCHMARK_DEPS + xnnpack_optional_dnnl_deps(),
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07004951)
4952
4953xnnpack_benchmark(
Marat Dukhaned6baaf2020-12-01 15:07:08 -08004954 name = "f32_velu_bench",
4955 srcs = [
4956 "bench/f32-velu.cc",
4957 "src/xnnpack/AlignedAllocator.h",
4958 ] + MICROKERNEL_BENCHMARK_HDRS,
4959 deps = MICROKERNEL_BENCHMARK_DEPS,
4960)
4961
4962xnnpack_benchmark(
Marat Dukhan4c4eb002019-12-08 21:27:49 -08004963 name = "f32_vscaleexpminusmax_bench",
4964 srcs = [
4965 "bench/f32-vscaleexpminusmax.cc",
4966 "src/xnnpack/AlignedAllocator.h",
4967 ] + MICROKERNEL_BENCHMARK_HDRS,
4968 deps = MICROKERNEL_BENCHMARK_DEPS,
4969)
4970
4971xnnpack_benchmark(
4972 name = "f32_vscaleextexp_bench",
4973 srcs = [
4974 "bench/f32-vscaleextexp.cc",
4975 "src/xnnpack/AlignedAllocator.h",
4976 ] + MICROKERNEL_BENCHMARK_HDRS,
4977 deps = MICROKERNEL_BENCHMARK_DEPS,
4978)
4979
4980xnnpack_benchmark(
Marat Dukhanf4db2f32020-06-30 10:55:30 -07004981 name = "f32_vsqrt_bench",
4982 srcs = [
4983 "bench/f32-vsqrt.cc",
4984 "src/xnnpack/AlignedAllocator.h",
4985 ] + MICROKERNEL_BENCHMARK_HDRS,
4986 deps = MICROKERNEL_BENCHMARK_DEPS,
4987)
4988
4989xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004990 name = "f32_im2col_gemm_bench",
4991 srcs = [
4992 "bench/f32-im2col-gemm.cc",
4993 "bench/conv.h",
4994 "src/xnnpack/AlignedAllocator.h",
4995 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004996 deps = MICROKERNEL_BENCHMARK_DEPS + [
4997 ":im2col",
4998 ":packing",
4999 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005000)
5001
Marat Dukhanfe7acb62020-03-09 19:30:05 -07005002xnnpack_benchmark(
Marat Dukhanffbf96a2020-05-14 02:59:08 -07005003 name = "rounding_bench",
5004 srcs = [
5005 "bench/rounding.cc",
Marat Dukhanffbf96a2020-05-14 02:59:08 -07005006 "src/xnnpack/AlignedAllocator.h",
Frank Barchard04336c12020-10-22 16:48:55 -07005007 "src/xnnpack/math-stubs.h",
Marat Dukhanffbf96a2020-05-14 02:59:08 -07005008 ] + MICROKERNEL_BENCHMARK_HDRS,
5009 deps = MICROKERNEL_BENCHMARK_DEPS,
5010)
5011
Marat Dukhan08c4a432019-10-03 09:29:21 -07005012########################### Benchmarks for operators ###########################
5013
5014xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005015 name = "average_pooling_bench",
5016 srcs = ["bench/average-pooling.cc"],
Marat Dukhan7a16d8b2020-03-11 04:22:44 -07005017 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07005018 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07005019 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07005020)
5021
5022xnnpack_benchmark(
Marat Dukhan87727142020-06-24 15:24:10 -07005023 name = "bankers_rounding_bench",
5024 srcs = ["bench/bankers-rounding.cc"],
5025 copts = xnnpack_optional_tflite_copts(),
5026 tags = ["nowin32"],
5027 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
5028)
5029
5030xnnpack_benchmark(
5031 name = "ceiling_bench",
5032 srcs = ["bench/ceiling.cc"],
5033 copts = xnnpack_optional_tflite_copts(),
5034 tags = ["nowin32"],
5035 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
5036)
5037
5038xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005039 name = "channel_shuffle_bench",
5040 srcs = ["bench/channel-shuffle.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07005041 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005042)
5043
5044xnnpack_benchmark(
5045 name = "convolution_bench",
5046 srcs = ["bench/convolution.cc"],
5047 copts = xnnpack_optional_tflite_copts() + xnnpack_optional_armcl_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07005048 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07005049 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps() + xnnpack_optional_armcl_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07005050)
5051
5052xnnpack_benchmark(
5053 name = "deconvolution_bench",
5054 srcs = ["bench/deconvolution.cc"],
5055 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07005056 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07005057 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07005058)
5059
5060xnnpack_benchmark(
Marat Dukhanb6bd4bc2020-12-01 17:01:40 -08005061 name = "elu_bench",
5062 srcs = ["bench/elu.cc"],
5063 copts = xnnpack_optional_tflite_copts(),
5064 tags = ["nowin32"],
5065 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
5066)
5067
5068xnnpack_benchmark(
Marat Dukhan87727142020-06-24 15:24:10 -07005069 name = "floor_bench",
5070 srcs = ["bench/floor.cc"],
5071 copts = xnnpack_optional_tflite_copts(),
5072 tags = ["nowin32"],
5073 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
5074)
5075
5076xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005077 name = "global_average_pooling_bench",
5078 srcs = ["bench/global-average-pooling.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07005079 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005080)
5081
5082xnnpack_benchmark(
Marat Dukhanad352602020-06-25 21:50:54 -07005083 name = "hardswish_bench",
5084 srcs = ["bench/hardswish.cc"],
5085 copts = xnnpack_optional_tflite_copts(),
5086 tags = ["nowin32"],
5087 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
5088)
5089
5090xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005091 name = "max_pooling_bench",
5092 srcs = ["bench/max-pooling.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07005093 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005094)
5095
5096xnnpack_benchmark(
5097 name = "sigmoid_bench",
5098 srcs = ["bench/sigmoid.cc"],
Marat Dukhanc3b9e862019-11-17 13:18:54 -08005099 copts = xnnpack_optional_tflite_copts(),
Marat Dukhanca2ba702020-04-24 01:31:47 -07005100 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07005101 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07005102)
5103
5104xnnpack_benchmark(
Marat Dukhan95b22432019-10-30 16:30:14 -07005105 name = "prelu_bench",
5106 srcs = ["bench/prelu.cc"],
5107 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07005108 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07005109 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan95b22432019-10-30 16:30:14 -07005110)
5111
5112xnnpack_benchmark(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08005113 name = "softmax_bench",
5114 srcs = ["bench/softmax.cc"],
Marat Dukhan9c0db962020-01-28 12:30:14 -08005115 copts = xnnpack_optional_tflite_copts(),
Marat Dukhanca2ba702020-04-24 01:31:47 -07005116 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07005117 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07005118)
5119
Marat Dukhan87727142020-06-24 15:24:10 -07005120xnnpack_benchmark(
Marat Dukhan6804bbd2020-06-30 19:26:11 -07005121 name = "square_root_bench",
5122 srcs = ["bench/square-root.cc"],
5123 copts = xnnpack_optional_tflite_copts(),
5124 tags = ["nowin32"],
5125 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
5126)
5127
5128xnnpack_benchmark(
Marat Dukhan87727142020-06-24 15:24:10 -07005129 name = "truncation_bench",
5130 srcs = ["bench/truncation.cc"],
5131 deps = OPERATOR_BENCHMARK_DEPS,
5132)
5133
Marat Dukhanc068bb62019-10-04 13:24:39 -07005134############################# End-to-end benchmarks ############################
5135
5136cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07005137 name = "fp32_mobilenet_v1",
5138 srcs = ["models/fp32-mobilenet-v1.cc"],
Marat Dukhanc068bb62019-10-04 13:24:39 -07005139 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08005140 copts = xnnpack_std_cxxopts(),
Marat Dukhanc068bb62019-10-04 13:24:39 -07005141 linkstatic = True,
5142 deps = [
5143 ":XNNPACK",
5144 "@pthreadpool",
5145 ],
5146)
5147
5148cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07005149 name = "fp16_mobilenet_v1",
5150 srcs = ["models/fp16-mobilenet-v1.cc"],
5151 hdrs = ["models/models.h"],
5152 copts = xnnpack_std_cxxopts(),
5153 linkstatic = True,
5154 deps = [
5155 ":XNNPACK",
5156 "@FP16",
5157 "@pthreadpool",
5158 ],
5159)
5160
5161cc_library(
Marat Dukhan0743cdf2020-08-04 18:52:07 -07005162 name = "qs8_mobilenet_v1",
5163 srcs = ["models/qs8-mobilenet-v1.cc"],
5164 hdrs = ["models/models.h"],
5165 copts = xnnpack_std_cxxopts(),
5166 linkstatic = True,
5167 deps = [
5168 ":XNNPACK",
5169 "@pthreadpool",
5170 ],
5171)
5172
5173cc_library(
Marat Dukhan70a96182020-09-03 17:13:58 -07005174 name = "qs8_mobilenet_v2",
5175 srcs = ["models/qs8-mobilenet-v2.cc"],
5176 hdrs = ["models/models.h"],
5177 copts = xnnpack_std_cxxopts(),
5178 linkstatic = True,
5179 deps = [
5180 ":XNNPACK",
5181 "@pthreadpool",
5182 ],
5183)
5184
5185cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07005186 name = "fp32_mobilenet_v2",
5187 srcs = ["models/fp32-mobilenet-v2.cc"],
Marat Dukhanc068bb62019-10-04 13:24:39 -07005188 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08005189 copts = xnnpack_std_cxxopts(),
Marat Dukhanc068bb62019-10-04 13:24:39 -07005190 linkstatic = True,
5191 deps = [
5192 ":XNNPACK",
5193 "@pthreadpool",
5194 ],
5195)
5196
Marat Dukhanc08cdf52019-12-09 09:17:51 -08005197cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07005198 name = "fp16_mobilenet_v2",
5199 srcs = ["models/fp16-mobilenet-v2.cc"],
5200 hdrs = ["models/models.h"],
5201 copts = xnnpack_std_cxxopts(),
5202 linkstatic = True,
5203 deps = [
5204 ":XNNPACK",
5205 "@FP16",
5206 "@pthreadpool",
5207 ],
5208)
5209
5210cc_library(
5211 name = "fp32_mobilenet_v3_large",
5212 srcs = ["models/fp32-mobilenet-v3-large.cc"],
Marat Dukhanc08cdf52019-12-09 09:17:51 -08005213 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08005214 copts = xnnpack_std_cxxopts(),
Marat Dukhanc08cdf52019-12-09 09:17:51 -08005215 linkstatic = True,
5216 deps = [
5217 ":XNNPACK",
5218 "@pthreadpool",
5219 ],
5220)
5221
5222cc_library(
Marat Dukhan66f3ccd2020-09-14 16:09:38 -07005223 name = "fp16_mobilenet_v3_large",
5224 srcs = ["models/fp16-mobilenet-v3-large.cc"],
5225 hdrs = ["models/models.h"],
5226 copts = xnnpack_std_cxxopts(),
5227 linkstatic = True,
5228 deps = [
5229 ":XNNPACK",
5230 "@FP16",
5231 "@pthreadpool",
5232 ],
5233)
5234
5235cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07005236 name = "fp32_mobilenet_v3_small",
5237 srcs = ["models/fp32-mobilenet-v3-small.cc"],
Marat Dukhanc08cdf52019-12-09 09:17:51 -08005238 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08005239 copts = xnnpack_std_cxxopts(),
Marat Dukhanc08cdf52019-12-09 09:17:51 -08005240 linkstatic = True,
5241 deps = [
5242 ":XNNPACK",
5243 "@pthreadpool",
5244 ],
5245)
5246
Marat Dukhan66f3ccd2020-09-14 16:09:38 -07005247cc_library(
5248 name = "fp16_mobilenet_v3_small",
5249 srcs = ["models/fp16-mobilenet-v3-small.cc"],
5250 hdrs = ["models/models.h"],
5251 copts = xnnpack_std_cxxopts(),
5252 linkstatic = True,
5253 deps = [
5254 ":XNNPACK",
5255 "@FP16",
5256 "@pthreadpool",
5257 ],
5258)
5259
Marat Dukhanc068bb62019-10-04 13:24:39 -07005260xnnpack_benchmark(
Marat Dukhanef4416e2019-10-31 13:44:40 -07005261 name = "f32_dwconv_e2e_bench",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08005262 srcs = [
5263 "bench/f32-dwconv-e2e.cc",
5264 "bench/end2end.h",
5265 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanef4416e2019-10-31 13:44:40 -07005266 deps = MICROKERNEL_BENCHMARK_DEPS + [
5267 ":XNNPACK",
Marat Dukhan270a2c42020-06-26 16:45:52 -07005268 ":fp32_mobilenet_v1",
5269 ":fp32_mobilenet_v2",
5270 ":fp32_mobilenet_v3_large",
5271 ":fp32_mobilenet_v3_small",
Marat Dukhanef4416e2019-10-31 13:44:40 -07005272 ],
5273)
5274
5275xnnpack_benchmark(
Marat Dukhan5f18d262019-10-31 10:24:14 -07005276 name = "f32_gemm_e2e_bench",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08005277 srcs = [
5278 "bench/f32-gemm-e2e.cc",
5279 "bench/end2end.h",
5280 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan5f18d262019-10-31 10:24:14 -07005281 deps = MICROKERNEL_BENCHMARK_DEPS + [
5282 ":XNNPACK",
Marat Dukhan270a2c42020-06-26 16:45:52 -07005283 ":fp32_mobilenet_v1",
5284 ":fp32_mobilenet_v2",
5285 ":fp32_mobilenet_v3_large",
5286 ":fp32_mobilenet_v3_small",
Marat Dukhan5f18d262019-10-31 10:24:14 -07005287 ],
5288)
5289
5290xnnpack_benchmark(
Marat Dukhanc068bb62019-10-04 13:24:39 -07005291 name = "end2end_bench",
5292 srcs = ["bench/end2end.cc"],
5293 deps = [
5294 ":XNNPACK",
Frank Barchardc712fa42019-10-31 14:00:21 -07005295 ":bench_utils",
Marat Dukhan270a2c42020-06-26 16:45:52 -07005296 ":fp16_mobilenet_v1",
5297 ":fp16_mobilenet_v2",
Marat Dukhan66f3ccd2020-09-14 16:09:38 -07005298 ":fp16_mobilenet_v3_large",
5299 ":fp16_mobilenet_v3_small",
Marat Dukhan270a2c42020-06-26 16:45:52 -07005300 ":fp32_mobilenet_v1",
5301 ":fp32_mobilenet_v2",
5302 ":fp32_mobilenet_v3_large",
5303 ":fp32_mobilenet_v3_small",
Marat Dukhan0743cdf2020-08-04 18:52:07 -07005304 ":qs8_mobilenet_v1",
Marat Dukhan70a96182020-09-03 17:13:58 -07005305 ":qs8_mobilenet_v2",
Marat Dukhanc068bb62019-10-04 13:24:39 -07005306 "@pthreadpool",
5307 ],
5308)
5309
Marat Dukhan6adff4e2019-10-14 18:32:07 -07005310#################### Accuracy evaluation for math functions ####################
5311
5312xnnpack_benchmark(
Marat Dukhancf67ec62020-12-14 09:56:37 -08005313 name = "f32_exp_ulp_eval",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07005314 srcs = [
Marat Dukhancf67ec62020-12-14 09:56:37 -08005315 "eval/f32-exp-ulp.cc",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07005316 "src/xnnpack/AlignedAllocator.h",
5317 ] + ACCURACY_EVAL_HDRS,
Marat Dukhan3a305212020-12-06 19:24:27 -08005318 deps = ACCURACY_EVAL_DEPS + [
5319 ":bench_utils",
5320 "@cpuinfo",
5321 ],
Marat Dukhan6adff4e2019-10-14 18:32:07 -07005322)
5323
Marat Dukhan515c9772019-10-17 18:07:57 -07005324xnnpack_benchmark(
Marat Dukhancf67ec62020-12-14 09:56:37 -08005325 name = "f32_expminus_ulp_eval",
Marat Dukhan515c9772019-10-17 18:07:57 -07005326 srcs = [
Marat Dukhancf67ec62020-12-14 09:56:37 -08005327 "eval/f32-expminus-ulp.cc",
Marat Dukhan515c9772019-10-17 18:07:57 -07005328 "src/xnnpack/AlignedAllocator.h",
5329 ] + ACCURACY_EVAL_HDRS,
Marat Dukhan3a305212020-12-06 19:24:27 -08005330 deps = ACCURACY_EVAL_DEPS + [
5331 ":bench_utils",
5332 "@cpuinfo",
5333 ],
Marat Dukhan515c9772019-10-17 18:07:57 -07005334)
5335
Marat Dukhan98ba4412019-10-23 02:14:28 -07005336xnnpack_benchmark(
Marat Dukhancf67ec62020-12-14 09:56:37 -08005337 name = "f32_expm1minus_ulp_eval",
Marat Dukhana438aca2020-11-20 15:45:01 -08005338 srcs = [
Marat Dukhancf67ec62020-12-14 09:56:37 -08005339 "eval/f32-expm1minus-ulp.cc",
Marat Dukhana438aca2020-11-20 15:45:01 -08005340 "src/xnnpack/AlignedAllocator.h",
5341 ] + ACCURACY_EVAL_HDRS,
Marat Dukhande390d42020-11-29 19:32:18 -08005342 deps = ACCURACY_EVAL_DEPS + [
5343 ":bench_utils",
Marat Dukhan3a305212020-12-06 19:24:27 -08005344 "@cpuinfo",
Marat Dukhande390d42020-11-29 19:32:18 -08005345 ],
Marat Dukhana438aca2020-11-20 15:45:01 -08005346)
5347
5348xnnpack_benchmark(
Marat Dukhancf67ec62020-12-14 09:56:37 -08005349 name = "f32_extexp_ulp_eval",
Marat Dukhan98ba4412019-10-23 02:14:28 -07005350 srcs = [
Marat Dukhancf67ec62020-12-14 09:56:37 -08005351 "eval/f32-extexp-ulp.cc",
Marat Dukhan98ba4412019-10-23 02:14:28 -07005352 "src/xnnpack/AlignedAllocator.h",
5353 ] + ACCURACY_EVAL_HDRS,
Marat Dukhan3a305212020-12-06 19:24:27 -08005354 deps = ACCURACY_EVAL_DEPS + [
5355 ":bench_utils",
5356 "@cpuinfo",
5357 ],
Marat Dukhan98ba4412019-10-23 02:14:28 -07005358)
5359
Marat Dukhanf44f0222020-12-14 11:53:27 -08005360xnnpack_benchmark(
5361 name = "f32_sigmoid_ulp_eval",
5362 srcs = [
5363 "eval/f32-sigmoid-ulp.cc",
5364 "src/xnnpack/AlignedAllocator.h",
5365 ] + ACCURACY_EVAL_HDRS,
5366 deps = ACCURACY_EVAL_DEPS + [
5367 ":bench_utils",
5368 "@cpuinfo",
5369 ],
5370)
5371
5372xnnpack_benchmark(
5373 name = "f32_sqrt_ulp_eval",
5374 srcs = [
5375 "eval/f32-sqrt-ulp.cc",
5376 "src/xnnpack/AlignedAllocator.h",
5377 ] + ACCURACY_EVAL_HDRS,
5378 deps = ACCURACY_EVAL_DEPS + [
5379 ":bench_utils",
5380 "@cpuinfo",
5381 ],
5382)
5383
5384################### Accuracy verification for math functions ##################
5385
5386xnnpack_unit_test(
Marat Dukhanf7291fc2020-12-15 11:02:50 -08005387 name = "f32_exp_eval",
5388 srcs = [
5389 "eval/f32-exp.cc",
5390 "src/xnnpack/AlignedAllocator.h",
5391 "src/xnnpack/math-stubs.h",
5392 ] + MICROKERNEL_TEST_HDRS,
5393 automatic = False,
5394 deps = MICROKERNEL_TEST_DEPS,
5395)
5396
5397xnnpack_unit_test(
Marat Dukhanf44f0222020-12-14 11:53:27 -08005398 name = "f32_expm1minus_eval",
5399 srcs = [
5400 "eval/f32-expm1minus.cc",
5401 "src/xnnpack/AlignedAllocator.h",
5402 "src/xnnpack/math-stubs.h",
5403 ] + MICROKERNEL_TEST_HDRS,
5404 automatic = False,
5405 deps = MICROKERNEL_TEST_DEPS,
5406)
5407
Marat Dukhan8853b822020-05-07 12:19:01 -07005408xnnpack_unit_test(
Marat Dukhand28a5a22020-12-14 15:27:22 -08005409 name = "f32_expminus_eval",
5410 srcs = [
5411 "eval/f32-expminus.cc",
5412 "src/xnnpack/AlignedAllocator.h",
5413 "src/xnnpack/math-stubs.h",
5414 ] + MICROKERNEL_TEST_HDRS,
5415 automatic = False,
5416 deps = MICROKERNEL_TEST_DEPS,
5417)
5418
5419xnnpack_unit_test(
Marat Dukhan8853b822020-05-07 12:19:01 -07005420 name = "f32_roundne_eval",
5421 srcs = [
5422 "eval/f32-roundne.cc",
5423 "src/xnnpack/AlignedAllocator.h",
5424 "src/xnnpack/math-stubs.h",
5425 ] + MICROKERNEL_TEST_HDRS,
Marat Dukhan22eed3d2020-05-11 20:13:37 -07005426 automatic = False,
Marat Dukhan8853b822020-05-07 12:19:01 -07005427 deps = MICROKERNEL_TEST_DEPS,
5428)
5429
Marat Dukhan2dbb9442020-05-12 20:43:43 -07005430xnnpack_unit_test(
Marat Dukhanc9852ba2020-05-13 17:21:29 -07005431 name = "f32_roundd_eval",
5432 srcs = [
5433 "eval/f32-roundd.cc",
5434 "src/xnnpack/AlignedAllocator.h",
5435 "src/xnnpack/math-stubs.h",
5436 ] + MICROKERNEL_TEST_HDRS,
5437 automatic = False,
5438 deps = MICROKERNEL_TEST_DEPS,
5439)
5440
5441xnnpack_unit_test(
5442 name = "f32_roundu_eval",
5443 srcs = [
5444 "eval/f32-roundu.cc",
5445 "src/xnnpack/AlignedAllocator.h",
5446 "src/xnnpack/math-stubs.h",
5447 ] + MICROKERNEL_TEST_HDRS,
5448 automatic = False,
5449 deps = MICROKERNEL_TEST_DEPS,
5450)
5451
5452xnnpack_unit_test(
Marat Dukhan2dbb9442020-05-12 20:43:43 -07005453 name = "f32_roundz_eval",
5454 srcs = [
5455 "eval/f32-roundz.cc",
5456 "src/xnnpack/AlignedAllocator.h",
5457 "src/xnnpack/math-stubs.h",
5458 ] + MICROKERNEL_TEST_HDRS,
Marat Dukhanc9852ba2020-05-13 17:21:29 -07005459 automatic = False,
Marat Dukhan2dbb9442020-05-12 20:43:43 -07005460 deps = MICROKERNEL_TEST_DEPS,
5461)
5462
Marat Dukhan08c4a432019-10-03 09:29:21 -07005463######################### Unit tests for micro-kernels #########################
5464
5465xnnpack_unit_test(
Marat Dukhande06f492020-04-09 00:19:31 -07005466 name = "f16_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005467 srcs = [
Marat Dukhande06f492020-04-09 00:19:31 -07005468 "test/f16-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005469 "test/gemm-microkernel-tester.h",
5470 "src/xnnpack/AlignedAllocator.h",
5471 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005472 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005473)
5474
5475xnnpack_unit_test(
Marat Dukhan355ab432020-04-09 19:01:52 -07005476 name = "f16_spmm_minmax_test",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08005477 srcs = [
Marat Dukhan355ab432020-04-09 19:01:52 -07005478 "test/f16-spmm-minmax.cc",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08005479 "test/spmm-microkernel-tester.h",
5480 "src/xnnpack/AlignedAllocator.h",
5481 ] + MICROKERNEL_TEST_HDRS,
5482 deps = MICROKERNEL_TEST_DEPS,
5483)
5484
5485xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005486 name = "f32_argmaxpool_test",
5487 srcs = [
5488 "test/f32-argmaxpool.cc",
5489 "test/argmaxpool-microkernel-tester.h",
5490 "src/xnnpack/AlignedAllocator.h",
5491 ] + MICROKERNEL_TEST_HDRS,
5492 deps = MICROKERNEL_TEST_DEPS,
5493)
5494
5495xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07005496 name = "f32_avgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005497 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07005498 "test/f32-avgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005499 "test/avgpool-microkernel-tester.h",
5500 "src/xnnpack/AlignedAllocator.h",
5501 ] + MICROKERNEL_TEST_HDRS,
5502 deps = MICROKERNEL_TEST_DEPS,
5503)
5504
5505xnnpack_unit_test(
Marat Dukhan660fd192020-03-10 04:55:30 -07005506 name = "f32_ibilinear_test",
Marat Dukhan35dacfb2019-11-07 19:18:16 -08005507 srcs = [
Marat Dukhan660fd192020-03-10 04:55:30 -07005508 "test/f32-ibilinear.cc",
5509 "test/ibilinear-microkernel-tester.h",
Marat Dukhan35dacfb2019-11-07 19:18:16 -08005510 "src/xnnpack/AlignedAllocator.h",
5511 ] + MICROKERNEL_TEST_HDRS,
5512 deps = MICROKERNEL_TEST_DEPS,
5513)
5514
5515xnnpack_unit_test(
XNNPACK Team21432672020-10-19 19:58:48 -07005516 name = "f32_ibilinear_chw_test",
5517 srcs = [
5518 "test/f32-ibilinear-chw.cc",
XNNPACK Team6be46b22020-10-22 23:34:54 -07005519 "test/ibilinear-microkernel-tester.h",
XNNPACK Team21432672020-10-19 19:58:48 -07005520 "src/xnnpack/AlignedAllocator.h",
5521 ] + MICROKERNEL_TEST_HDRS,
5522 deps = MICROKERNEL_TEST_DEPS,
5523)
5524
5525xnnpack_unit_test(
Frank Barchardb1966592020-05-12 13:47:06 -07005526 name = "f16_clamp_test",
5527 srcs = [
5528 "test/f16-clamp.cc",
5529 "test/clamp-microkernel-tester.h",
5530 ] + MICROKERNEL_TEST_HDRS,
5531 deps = MICROKERNEL_TEST_DEPS,
5532)
5533
5534xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005535 name = "f32_clamp_test",
5536 srcs = [
5537 "test/f32-clamp.cc",
5538 "test/clamp-microkernel-tester.h",
5539 ] + MICROKERNEL_TEST_HDRS,
5540 deps = MICROKERNEL_TEST_DEPS,
5541)
5542
5543xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07005544 name = "f32_igemm_test",
5545 srcs = [
5546 "test/f32-igemm.cc",
5547 "test/gemm-microkernel-tester.h",
5548 "src/xnnpack/AlignedAllocator.h",
5549 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005550 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan163a7e62020-04-09 04:19:26 -07005551)
5552
5553xnnpack_unit_test(
Marat Dukhan467f6362020-05-22 23:21:55 -07005554 name = "f32_igemm_relu_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005555 srcs = [
Marat Dukhan467f6362020-05-22 23:21:55 -07005556 "test/f32-igemm-relu.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005557 "test/gemm-microkernel-tester.h",
5558 "src/xnnpack/AlignedAllocator.h",
5559 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005560 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005561)
5562
5563xnnpack_unit_test(
Marat Dukhane207b7b2020-05-28 16:27:42 -07005564 name = "f32_igemm_minmax_test",
5565 srcs = [
5566 "test/f32-igemm-minmax.cc",
5567 "test/gemm-microkernel-tester.h",
5568 "src/xnnpack/AlignedAllocator.h",
5569 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005570 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhane207b7b2020-05-28 16:27:42 -07005571)
5572
5573xnnpack_unit_test(
Frank Barchardb0e4fae2020-05-04 15:27:51 -07005574 name = "f16_igemm_minmax_test",
5575 srcs = [
5576 "test/f16-igemm-minmax.cc",
5577 "test/gemm-microkernel-tester.h",
5578 "src/xnnpack/AlignedAllocator.h",
5579 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005580 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Frank Barchardb0e4fae2020-05-04 15:27:51 -07005581)
5582
5583xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005584 name = "f32_conv_hwc_test",
5585 srcs = [
5586 "test/f32-conv-hwc.cc",
5587 "test/conv-hwc-microkernel-tester.h",
5588 "src/xnnpack/AlignedAllocator.h",
5589 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005590 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005591)
5592
5593xnnpack_unit_test(
Marat Dukhan1f29b802020-05-15 23:46:39 -07005594 name = "f32_conv_hwc2chw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005595 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07005596 "test/f32-conv-hwc2chw.cc",
5597 "test/conv-hwc2chw-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005598 "src/xnnpack/AlignedAllocator.h",
5599 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005600 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005601)
5602
5603xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07005604 name = "f32_dwconv_test",
5605 srcs = [
5606 "test/f32-dwconv.cc",
5607 "test/dwconv-microkernel-tester.h",
5608 "src/xnnpack/AlignedAllocator.h",
5609 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005610 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan163a7e62020-04-09 04:19:26 -07005611)
5612
5613xnnpack_unit_test(
Frank Barchard5a599a62020-06-04 20:12:44 -07005614 name = "f16_dwconv_minmax_test",
5615 srcs = [
5616 "test/f16-dwconv-minmax.cc",
5617 "test/dwconv-microkernel-tester.h",
5618 "src/xnnpack/AlignedAllocator.h",
5619 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005620 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Frank Barchard5a599a62020-06-04 20:12:44 -07005621)
5622
5623xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07005624 name = "f32_dwconv_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005625 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07005626 "test/f32-dwconv-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005627 "test/dwconv-microkernel-tester.h",
5628 "src/xnnpack/AlignedAllocator.h",
5629 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005630 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005631)
5632
5633xnnpack_unit_test(
Marat Dukhanbf715f92020-10-23 20:17:00 -07005634 name = "f32_dwconv2d_chw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005635 srcs = [
Marat Dukhanbf715f92020-10-23 20:17:00 -07005636 "test/f32-dwconv2d-chw.cc",
5637 "test/dwconv2d-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005638 "src/xnnpack/AlignedAllocator.h",
5639 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005640 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005641)
5642
5643xnnpack_unit_test(
Frank Barchard0bb49a72020-06-04 11:35:11 -07005644 name = "f16_gavgpool_minmax_test",
5645 srcs = [
5646 "test/f16-gavgpool-minmax.cc",
5647 "test/gavgpool-microkernel-tester.h",
5648 "src/xnnpack/AlignedAllocator.h",
5649 ] + MICROKERNEL_TEST_HDRS,
5650 deps = MICROKERNEL_TEST_DEPS,
5651)
5652
5653xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07005654 name = "f32_gavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005655 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07005656 "test/f32-gavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005657 "test/gavgpool-microkernel-tester.h",
5658 "src/xnnpack/AlignedAllocator.h",
5659 ] + MICROKERNEL_TEST_HDRS,
5660 deps = MICROKERNEL_TEST_DEPS,
5661)
5662
5663xnnpack_unit_test(
Marat Dukhan1f29b802020-05-15 23:46:39 -07005664 name = "f32_gavgpool_cw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005665 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07005666 "test/f32-gavgpool-cw.cc",
5667 "test/gavgpool-cw-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005668 "src/xnnpack/AlignedAllocator.h",
5669 ] + MICROKERNEL_TEST_HDRS,
5670 deps = MICROKERNEL_TEST_DEPS,
5671)
5672
5673xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07005674 name = "f32_gemm_test",
5675 srcs = [
5676 "test/f32-gemm.cc",
5677 "test/gemm-microkernel-tester.h",
5678 "src/xnnpack/AlignedAllocator.h",
5679 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005680 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan163a7e62020-04-09 04:19:26 -07005681)
5682
5683xnnpack_unit_test(
Marat Dukhan467f6362020-05-22 23:21:55 -07005684 name = "f32_gemm_relu_test",
5685 srcs = [
5686 "test/f32-gemm-relu.cc",
5687 "test/gemm-microkernel-tester.h",
5688 "src/xnnpack/AlignedAllocator.h",
5689 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005690 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan467f6362020-05-22 23:21:55 -07005691)
5692
5693xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07005694 name = "f32_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005695 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07005696 "test/f32-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005697 "test/gemm-microkernel-tester.h",
5698 "src/xnnpack/AlignedAllocator.h",
5699 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005700 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005701)
5702
5703xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07005704 name = "f32_gemminc_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005705 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07005706 "test/f32-gemminc-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005707 "test/gemm-microkernel-tester.h",
5708 "src/xnnpack/AlignedAllocator.h",
5709 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005710 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005711)
5712
5713xnnpack_unit_test(
Frank Barchardb1966592020-05-12 13:47:06 -07005714 name = "f16_hswish_test",
5715 srcs = [
5716 "test/f16-hswish.cc",
5717 "test/hswish-microkernel-tester.h",
5718 ] + MICROKERNEL_TEST_HDRS,
5719 deps = MICROKERNEL_TEST_DEPS,
5720)
5721
5722xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005723 name = "f32_hswish_test",
5724 srcs = [
5725 "test/f32-hswish.cc",
5726 "test/hswish-microkernel-tester.h",
5727 ] + MICROKERNEL_TEST_HDRS,
5728 deps = MICROKERNEL_TEST_DEPS,
5729)
5730
5731xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07005732 name = "f32_maxpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005733 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07005734 "test/f32-maxpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005735 "test/maxpool-microkernel-tester.h",
5736 ] + MICROKERNEL_TEST_HDRS,
5737 deps = MICROKERNEL_TEST_DEPS,
5738)
5739
5740xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07005741 name = "f32_pavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005742 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07005743 "test/f32-pavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005744 "test/avgpool-microkernel-tester.h",
5745 "src/xnnpack/AlignedAllocator.h",
5746 ] + MICROKERNEL_TEST_HDRS,
5747 deps = MICROKERNEL_TEST_DEPS,
5748)
5749
5750xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07005751 name = "f32_ppmm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005752 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07005753 "test/f32-ppmm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005754 "test/gemm-microkernel-tester.h",
5755 "src/xnnpack/AlignedAllocator.h",
5756 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005757 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005758)
5759
5760xnnpack_unit_test(
Frank Barchardb1966592020-05-12 13:47:06 -07005761 name = "f16_prelu_test",
5762 srcs = [
5763 "test/f16-prelu.cc",
5764 "test/prelu-microkernel-tester.h",
5765 "src/xnnpack/AlignedAllocator.h",
5766 ] + MICROKERNEL_TEST_HDRS,
5767 deps = MICROKERNEL_TEST_DEPS,
5768)
5769
5770xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005771 name = "f32_prelu_test",
5772 srcs = [
5773 "test/f32-prelu.cc",
5774 "test/prelu-microkernel-tester.h",
5775 "src/xnnpack/AlignedAllocator.h",
5776 ] + MICROKERNEL_TEST_HDRS,
5777 deps = MICROKERNEL_TEST_DEPS,
5778)
5779
5780xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07005781 name = "f32_raddexpminusmax_test",
5782 srcs = [
5783 "test/f32-raddexpminusmax.cc",
5784 "test/raddexpminusmax-microkernel-tester.h",
5785 ] + MICROKERNEL_TEST_HDRS,
5786 deps = MICROKERNEL_TEST_DEPS,
5787)
5788
5789xnnpack_unit_test(
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07005790 name = "f32_raddextexp_test",
5791 srcs = [
5792 "test/f32-raddextexp.cc",
5793 "test/raddextexp-microkernel-tester.h",
5794 ] + MICROKERNEL_TEST_HDRS,
5795 deps = MICROKERNEL_TEST_DEPS,
5796)
5797
5798xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07005799 name = "f32_raddstoreexpminusmax_test",
5800 srcs = [
5801 "test/f32-raddstoreexpminusmax.cc",
5802 "test/raddstoreexpminusmax-microkernel-tester.h",
5803 ] + MICROKERNEL_TEST_HDRS,
5804 deps = MICROKERNEL_TEST_DEPS,
5805)
5806
5807xnnpack_unit_test(
Frank Barchardfb158e22020-07-15 16:10:10 -07005808 name = "f32_relu_test",
5809 srcs = [
5810 "test/f32-relu.cc",
5811 "test/vunary-microkernel-tester.h",
5812 ] + MICROKERNEL_TEST_HDRS,
5813 deps = MICROKERNEL_TEST_DEPS,
5814)
5815
5816xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005817 name = "f32_rmax_test",
5818 srcs = [
5819 "test/f32-rmax.cc",
5820 "test/rmax-microkernel-tester.h",
5821 ] + MICROKERNEL_TEST_HDRS,
5822 deps = MICROKERNEL_TEST_DEPS,
5823)
5824
5825xnnpack_unit_test(
Marat Dukhan346a9e52019-11-15 09:06:30 -08005826 name = "f32_sigmoid_test",
5827 srcs = [
5828 "test/f32-sigmoid.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005829 "test/vunary-microkernel-tester.h",
Marat Dukhan346a9e52019-11-15 09:06:30 -08005830 ] + MICROKERNEL_TEST_HDRS,
5831 deps = MICROKERNEL_TEST_DEPS,
5832)
5833
5834xnnpack_unit_test(
Marat Dukhan355ab432020-04-09 19:01:52 -07005835 name = "f32_spmm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005836 srcs = [
Marat Dukhan355ab432020-04-09 19:01:52 -07005837 "test/f32-spmm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005838 "test/spmm-microkernel-tester.h",
5839 "src/xnnpack/AlignedAllocator.h",
5840 ] + MICROKERNEL_TEST_HDRS,
5841 deps = MICROKERNEL_TEST_DEPS,
5842)
5843
5844xnnpack_unit_test(
Frank Barchardd793f6c2020-05-08 13:37:43 -07005845 name = "f16_vadd_minmax_test",
5846 srcs = [
5847 "test/f16-vadd-minmax.cc",
5848 "test/vbinary-microkernel-tester.h",
5849 ] + MICROKERNEL_TEST_HDRS,
5850 deps = MICROKERNEL_TEST_DEPS,
5851)
5852
5853xnnpack_unit_test(
5854 name = "f16_vaddc_minmax_test",
5855 srcs = [
5856 "test/f16-vaddc-minmax.cc",
5857 "test/vbinaryc-microkernel-tester.h",
5858 ] + MICROKERNEL_TEST_HDRS,
5859 deps = MICROKERNEL_TEST_DEPS,
5860)
5861
5862xnnpack_unit_test(
5863 name = "f16_vdiv_minmax_test",
5864 srcs = [
5865 "test/f16-vdiv-minmax.cc",
5866 "test/vbinary-microkernel-tester.h",
5867 ] + MICROKERNEL_TEST_HDRS,
5868 deps = MICROKERNEL_TEST_DEPS,
5869)
5870
5871xnnpack_unit_test(
5872 name = "f16_vdivc_minmax_test",
5873 srcs = [
5874 "test/f16-vdivc-minmax.cc",
5875 "test/vbinaryc-microkernel-tester.h",
5876 ] + MICROKERNEL_TEST_HDRS,
5877 deps = MICROKERNEL_TEST_DEPS,
5878)
5879
5880xnnpack_unit_test(
5881 name = "f16_vrdivc_minmax_test",
5882 srcs = [
5883 "test/f16-vrdivc-minmax.cc",
5884 "test/vbinaryc-microkernel-tester.h",
5885 ] + MICROKERNEL_TEST_HDRS,
5886 deps = MICROKERNEL_TEST_DEPS,
5887)
5888
5889xnnpack_unit_test(
5890 name = "f16_vmax_test",
5891 srcs = [
5892 "test/f16-vmax.cc",
5893 "test/vbinary-microkernel-tester.h",
5894 ] + MICROKERNEL_TEST_HDRS,
5895 deps = MICROKERNEL_TEST_DEPS,
5896)
5897
5898xnnpack_unit_test(
5899 name = "f16_vmaxc_test",
5900 srcs = [
5901 "test/f16-vmaxc.cc",
5902 "test/vbinaryc-microkernel-tester.h",
5903 ] + MICROKERNEL_TEST_HDRS,
5904 deps = MICROKERNEL_TEST_DEPS,
5905)
5906
5907xnnpack_unit_test(
5908 name = "f16_vmin_test",
5909 srcs = [
5910 "test/f16-vmin.cc",
5911 "test/vbinary-microkernel-tester.h",
5912 ] + MICROKERNEL_TEST_HDRS,
5913 deps = MICROKERNEL_TEST_DEPS,
5914)
5915
5916xnnpack_unit_test(
5917 name = "f16_vminc_test",
5918 srcs = [
5919 "test/f16-vminc.cc",
5920 "test/vbinaryc-microkernel-tester.h",
5921 ] + MICROKERNEL_TEST_HDRS,
5922 deps = MICROKERNEL_TEST_DEPS,
5923)
5924
5925xnnpack_unit_test(
5926 name = "f16_vmul_minmax_test",
5927 srcs = [
5928 "test/f16-vmul-minmax.cc",
5929 "test/vbinary-microkernel-tester.h",
5930 ] + MICROKERNEL_TEST_HDRS,
5931 deps = MICROKERNEL_TEST_DEPS,
5932)
5933
5934xnnpack_unit_test(
5935 name = "f16_vmulc_minmax_test",
5936 srcs = [
5937 "test/f16-vmulc-minmax.cc",
5938 "test/vbinaryc-microkernel-tester.h",
5939 ] + MICROKERNEL_TEST_HDRS,
5940 deps = MICROKERNEL_TEST_DEPS,
5941)
5942
5943xnnpack_unit_test(
5944 name = "f16_vsub_minmax_test",
5945 srcs = [
5946 "test/f16-vsub-minmax.cc",
5947 "test/vbinary-microkernel-tester.h",
5948 ] + MICROKERNEL_TEST_HDRS,
5949 deps = MICROKERNEL_TEST_DEPS,
5950)
5951
5952xnnpack_unit_test(
5953 name = "f16_vsubc_minmax_test",
5954 srcs = [
5955 "test/f16-vsubc-minmax.cc",
5956 "test/vbinaryc-microkernel-tester.h",
5957 ] + MICROKERNEL_TEST_HDRS,
5958 deps = MICROKERNEL_TEST_DEPS,
5959)
5960
5961xnnpack_unit_test(
5962 name = "f16_vrsubc_minmax_test",
5963 srcs = [
5964 "test/f16-vrsubc-minmax.cc",
5965 "test/vbinaryc-microkernel-tester.h",
5966 ] + MICROKERNEL_TEST_HDRS,
5967 deps = MICROKERNEL_TEST_DEPS,
5968)
5969
5970xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07005971 name = "f32_vabs_test",
5972 srcs = [
5973 "test/f32-vabs.cc",
5974 "test/vunary-microkernel-tester.h",
5975 ] + MICROKERNEL_TEST_HDRS,
5976 deps = MICROKERNEL_TEST_DEPS,
5977)
5978
5979xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005980 name = "f32_vadd_test",
5981 srcs = [
5982 "test/f32-vadd.cc",
5983 "test/vbinary-microkernel-tester.h",
5984 ] + MICROKERNEL_TEST_HDRS,
5985 deps = MICROKERNEL_TEST_DEPS,
5986)
5987
5988xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005989 name = "f32_vadd_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005990 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005991 "test/f32-vadd-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005992 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005993 ] + MICROKERNEL_TEST_HDRS,
5994 deps = MICROKERNEL_TEST_DEPS,
5995)
5996
5997xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005998 name = "f32_vadd_relu_test",
5999 srcs = [
6000 "test/f32-vadd-relu.cc",
6001 "test/vbinary-microkernel-tester.h",
6002 ] + MICROKERNEL_TEST_HDRS,
6003 deps = MICROKERNEL_TEST_DEPS,
6004)
6005
6006xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07006007 name = "f32_vaddc_test",
6008 srcs = [
6009 "test/f32-vaddc.cc",
6010 "test/vbinaryc-microkernel-tester.h",
6011 ] + MICROKERNEL_TEST_HDRS,
6012 deps = MICROKERNEL_TEST_DEPS,
6013)
6014
6015xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006016 name = "f32_vaddc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08006017 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006018 "test/f32-vaddc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08006019 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006020 ] + MICROKERNEL_TEST_HDRS,
6021 deps = MICROKERNEL_TEST_DEPS,
6022)
6023
6024xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07006025 name = "f32_vaddc_relu_test",
6026 srcs = [
6027 "test/f32-vaddc-relu.cc",
6028 "test/vbinaryc-microkernel-tester.h",
6029 ] + MICROKERNEL_TEST_HDRS,
6030 deps = MICROKERNEL_TEST_DEPS,
6031)
6032
6033xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07006034 name = "f32_vdiv_test",
6035 srcs = [
6036 "test/f32-vdiv.cc",
6037 "test/vbinary-microkernel-tester.h",
6038 ] + MICROKERNEL_TEST_HDRS,
6039 deps = MICROKERNEL_TEST_DEPS,
6040)
6041
6042xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006043 name = "f32_vdiv_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08006044 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006045 "test/f32-vdiv-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08006046 "test/vbinary-microkernel-tester.h",
6047 ] + MICROKERNEL_TEST_HDRS,
6048 deps = MICROKERNEL_TEST_DEPS,
6049)
6050
6051xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07006052 name = "f32_vdiv_relu_test",
6053 srcs = [
6054 "test/f32-vdiv-relu.cc",
6055 "test/vbinary-microkernel-tester.h",
6056 ] + MICROKERNEL_TEST_HDRS,
6057 deps = MICROKERNEL_TEST_DEPS,
6058)
6059
6060xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07006061 name = "f32_vdivc_test",
6062 srcs = [
6063 "test/f32-vdivc.cc",
6064 "test/vbinaryc-microkernel-tester.h",
6065 ] + MICROKERNEL_TEST_HDRS,
6066 deps = MICROKERNEL_TEST_DEPS,
6067)
6068
6069xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006070 name = "f32_vdivc_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08006071 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006072 "test/f32-vdivc-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08006073 "test/vbinaryc-microkernel-tester.h",
6074 ] + MICROKERNEL_TEST_HDRS,
6075 deps = MICROKERNEL_TEST_DEPS,
6076)
6077
6078xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07006079 name = "f32_vdivc_relu_test",
6080 srcs = [
6081 "test/f32-vdivc-relu.cc",
6082 "test/vbinaryc-microkernel-tester.h",
6083 ] + MICROKERNEL_TEST_HDRS,
6084 deps = MICROKERNEL_TEST_DEPS,
6085)
6086
6087xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07006088 name = "f32_vrdivc_test",
6089 srcs = [
6090 "test/f32-vrdivc.cc",
6091 "test/vbinaryc-microkernel-tester.h",
6092 ] + MICROKERNEL_TEST_HDRS,
6093 deps = MICROKERNEL_TEST_DEPS,
6094)
6095
6096xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006097 name = "f32_vrdivc_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08006098 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006099 "test/f32-vrdivc-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08006100 "test/vbinaryc-microkernel-tester.h",
6101 ] + MICROKERNEL_TEST_HDRS,
6102 deps = MICROKERNEL_TEST_DEPS,
6103)
6104
6105xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07006106 name = "f32_vrdivc_relu_test",
6107 srcs = [
6108 "test/f32-vrdivc-relu.cc",
6109 "test/vbinaryc-microkernel-tester.h",
6110 ] + MICROKERNEL_TEST_HDRS,
6111 deps = MICROKERNEL_TEST_DEPS,
6112)
6113
6114xnnpack_unit_test(
Marat Dukhaned6baaf2020-12-01 15:07:08 -08006115 name = "f32_velu_test",
6116 srcs = [
6117 "test/f32-velu.cc",
6118 "test/vunary-microkernel-tester.h",
6119 ] + MICROKERNEL_TEST_HDRS,
6120 deps = MICROKERNEL_TEST_DEPS,
6121)
6122
6123xnnpack_unit_test(
Marat Dukhan403b7d42019-12-05 12:49:11 -08006124 name = "f32_vmax_test",
6125 srcs = [
6126 "test/f32-vmax.cc",
6127 "test/vbinary-microkernel-tester.h",
6128 ] + MICROKERNEL_TEST_HDRS,
6129 deps = MICROKERNEL_TEST_DEPS,
6130)
6131
6132xnnpack_unit_test(
6133 name = "f32_vmaxc_test",
6134 srcs = [
6135 "test/f32-vmaxc.cc",
6136 "test/vbinaryc-microkernel-tester.h",
6137 ] + MICROKERNEL_TEST_HDRS,
6138 deps = MICROKERNEL_TEST_DEPS,
6139)
6140
6141xnnpack_unit_test(
6142 name = "f32_vmin_test",
6143 srcs = [
6144 "test/f32-vmin.cc",
6145 "test/vbinary-microkernel-tester.h",
6146 ] + MICROKERNEL_TEST_HDRS,
6147 deps = MICROKERNEL_TEST_DEPS,
6148)
6149
6150xnnpack_unit_test(
6151 name = "f32_vminc_test",
6152 srcs = [
6153 "test/f32-vminc.cc",
6154 "test/vbinaryc-microkernel-tester.h",
6155 ] + MICROKERNEL_TEST_HDRS,
6156 deps = MICROKERNEL_TEST_DEPS,
6157)
6158
6159xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07006160 name = "f32_vmul_test",
6161 srcs = [
6162 "test/f32-vmul.cc",
6163 "test/vbinary-microkernel-tester.h",
6164 ] + MICROKERNEL_TEST_HDRS,
6165 deps = MICROKERNEL_TEST_DEPS,
6166)
6167
6168xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006169 name = "f32_vmul_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006170 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006171 "test/f32-vmul-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08006172 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08006173 ] + MICROKERNEL_TEST_HDRS,
6174 deps = MICROKERNEL_TEST_DEPS,
6175)
6176
6177xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07006178 name = "f32_vmul_relu_test",
6179 srcs = [
6180 "test/f32-vmul-relu.cc",
6181 "test/vbinary-microkernel-tester.h",
6182 ] + MICROKERNEL_TEST_HDRS,
6183 deps = MICROKERNEL_TEST_DEPS,
6184)
6185
6186xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07006187 name = "f32_vmulc_test",
6188 srcs = [
6189 "test/f32-vmulc.cc",
6190 "test/vbinaryc-microkernel-tester.h",
6191 ] + MICROKERNEL_TEST_HDRS,
6192 deps = MICROKERNEL_TEST_DEPS,
6193)
6194
6195xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006196 name = "f32_vmulc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08006197 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006198 "test/f32-vmulc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08006199 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006200 ] + MICROKERNEL_TEST_HDRS,
6201 deps = MICROKERNEL_TEST_DEPS,
6202)
6203
6204xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07006205 name = "f32_vmulc_relu_test",
6206 srcs = [
6207 "test/f32-vmulc-relu.cc",
6208 "test/vbinaryc-microkernel-tester.h",
6209 ] + MICROKERNEL_TEST_HDRS,
6210 deps = MICROKERNEL_TEST_DEPS,
6211)
6212
6213xnnpack_unit_test(
Frank Barchard2a1049c2020-06-03 02:31:27 -07006214 name = "f16_vmulcaddc_minmax_test",
6215 srcs = [
6216 "test/f16-vmulcaddc-minmax.cc",
6217 "test/vmulcaddc-microkernel-tester.h",
6218 "src/xnnpack/AlignedAllocator.h",
6219 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07006220 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Frank Barchard2a1049c2020-06-03 02:31:27 -07006221)
6222
6223xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07006224 name = "f32_vmulcaddc_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006225 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07006226 "test/f32-vmulcaddc-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006227 "test/vmulcaddc-microkernel-tester.h",
6228 "src/xnnpack/AlignedAllocator.h",
6229 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07006230 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07006231)
6232
6233xnnpack_unit_test(
Marat Dukhan8cc7efe2020-06-10 16:24:27 -07006234 name = "f32_vlrelu_test",
6235 srcs = [
6236 "test/f32-vlrelu.cc",
6237 "test/vunary-microkernel-tester.h",
6238 ] + MICROKERNEL_TEST_HDRS,
6239 deps = MICROKERNEL_TEST_DEPS,
6240)
6241
6242xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07006243 name = "f32_vneg_test",
6244 srcs = [
6245 "test/f32-vneg.cc",
6246 "test/vunary-microkernel-tester.h",
6247 ] + MICROKERNEL_TEST_HDRS,
6248 deps = MICROKERNEL_TEST_DEPS,
6249)
6250
6251xnnpack_unit_test(
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07006252 name = "f32_vrndne_test",
6253 srcs = [
6254 "test/f32-vrndne.cc",
6255 "test/vunary-microkernel-tester.h",
6256 ] + MICROKERNEL_TEST_HDRS,
6257 deps = MICROKERNEL_TEST_DEPS,
6258)
6259
6260xnnpack_unit_test(
6261 name = "f32_vrndz_test",
6262 srcs = [
6263 "test/f32-vrndz.cc",
6264 "test/vunary-microkernel-tester.h",
6265 ] + MICROKERNEL_TEST_HDRS,
6266 deps = MICROKERNEL_TEST_DEPS,
6267)
6268
6269xnnpack_unit_test(
6270 name = "f32_vrndu_test",
6271 srcs = [
6272 "test/f32-vrndu.cc",
6273 "test/vunary-microkernel-tester.h",
6274 ] + MICROKERNEL_TEST_HDRS,
6275 deps = MICROKERNEL_TEST_DEPS,
6276)
6277
6278xnnpack_unit_test(
6279 name = "f32_vrndd_test",
6280 srcs = [
6281 "test/f32-vrndd.cc",
6282 "test/vunary-microkernel-tester.h",
6283 ] + MICROKERNEL_TEST_HDRS,
6284 deps = MICROKERNEL_TEST_DEPS,
6285)
6286
6287xnnpack_unit_test(
Marat Dukhan05ac8e32019-10-21 15:39:33 -07006288 name = "f32_vscale_test",
6289 srcs = [
6290 "test/f32-vscale.cc",
6291 "test/vscale-microkernel-tester.h",
6292 ] + MICROKERNEL_TEST_HDRS,
6293 deps = MICROKERNEL_TEST_DEPS,
6294)
6295
6296xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07006297 name = "f32_vscaleexpminusmax_test",
6298 srcs = [
6299 "test/f32-vscaleexpminusmax.cc",
6300 "test/vscaleexpminusmax-microkernel-tester.h",
6301 ] + MICROKERNEL_TEST_HDRS,
6302 deps = MICROKERNEL_TEST_DEPS,
6303)
6304
6305xnnpack_unit_test(
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07006306 name = "f32_vscaleextexp_test",
6307 srcs = [
6308 "test/f32-vscaleextexp.cc",
6309 "test/vscaleextexp-microkernel-tester.h",
6310 ] + MICROKERNEL_TEST_HDRS,
6311 deps = MICROKERNEL_TEST_DEPS,
6312)
6313
6314xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07006315 name = "f32_vsqr_test",
6316 srcs = [
6317 "test/f32-vsqr.cc",
6318 "test/vunary-microkernel-tester.h",
6319 ] + MICROKERNEL_TEST_HDRS,
6320 deps = MICROKERNEL_TEST_DEPS,
6321)
6322
6323xnnpack_unit_test(
Marat Dukhan13bafb02020-06-05 00:43:11 -07006324 name = "f32_vsqrdiff_test",
6325 srcs = [
6326 "test/f32-vsqrdiff.cc",
6327 "test/vbinary-microkernel-tester.h",
6328 ] + MICROKERNEL_TEST_HDRS,
6329 deps = MICROKERNEL_TEST_DEPS,
6330)
6331
6332xnnpack_unit_test(
6333 name = "f32_vsqrdiffc_test",
6334 srcs = [
6335 "test/f32-vsqrdiffc.cc",
6336 "test/vbinaryc-microkernel-tester.h",
6337 ] + MICROKERNEL_TEST_HDRS,
6338 deps = MICROKERNEL_TEST_DEPS,
6339)
6340
6341xnnpack_unit_test(
Marat Dukhanf4db2f32020-06-30 10:55:30 -07006342 name = "f32_vsqrt_test",
6343 srcs = [
6344 "test/f32-vsqrt.cc",
6345 "test/vunary-microkernel-tester.h",
6346 ] + MICROKERNEL_TEST_HDRS,
6347 deps = MICROKERNEL_TEST_DEPS,
6348)
6349
6350xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07006351 name = "f32_vsub_test",
6352 srcs = [
6353 "test/f32-vsub.cc",
6354 "test/vbinary-microkernel-tester.h",
6355 ] + MICROKERNEL_TEST_HDRS,
6356 deps = MICROKERNEL_TEST_DEPS,
6357)
6358
6359xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006360 name = "f32_vsub_minmax_test",
Marat Dukhan97579532019-10-18 16:40:39 -07006361 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006362 "test/f32-vsub-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08006363 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08006364 ] + MICROKERNEL_TEST_HDRS,
6365 deps = MICROKERNEL_TEST_DEPS,
6366)
6367
6368xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07006369 name = "f32_vsub_relu_test",
6370 srcs = [
6371 "test/f32-vsub-relu.cc",
6372 "test/vbinary-microkernel-tester.h",
6373 ] + MICROKERNEL_TEST_HDRS,
6374 deps = MICROKERNEL_TEST_DEPS,
6375)
6376
6377xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07006378 name = "f32_vsubc_test",
6379 srcs = [
6380 "test/f32-vsubc.cc",
6381 "test/vbinaryc-microkernel-tester.h",
6382 ] + MICROKERNEL_TEST_HDRS,
6383 deps = MICROKERNEL_TEST_DEPS,
6384)
6385
6386xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006387 name = "f32_vsubc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08006388 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006389 "test/f32-vsubc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08006390 "test/vbinaryc-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08006391 ] + MICROKERNEL_TEST_HDRS,
6392 deps = MICROKERNEL_TEST_DEPS,
6393)
6394
6395xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07006396 name = "f32_vsubc_relu_test",
6397 srcs = [
6398 "test/f32-vsubc-relu.cc",
6399 "test/vbinaryc-microkernel-tester.h",
6400 ] + MICROKERNEL_TEST_HDRS,
6401 deps = MICROKERNEL_TEST_DEPS,
6402)
6403
6404xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07006405 name = "f32_vrsubc_test",
6406 srcs = [
6407 "test/f32-vrsubc.cc",
6408 "test/vbinaryc-microkernel-tester.h",
6409 ] + MICROKERNEL_TEST_HDRS,
6410 deps = MICROKERNEL_TEST_DEPS,
6411)
6412
6413xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006414 name = "f32_vrsubc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08006415 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006416 "test/f32-vrsubc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08006417 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan97579532019-10-18 16:40:39 -07006418 ] + MICROKERNEL_TEST_HDRS,
6419 deps = MICROKERNEL_TEST_DEPS,
6420)
6421
6422xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07006423 name = "f32_vrsubc_relu_test",
6424 srcs = [
6425 "test/f32-vrsubc-relu.cc",
6426 "test/vbinaryc-microkernel-tester.h",
6427 ] + MICROKERNEL_TEST_HDRS,
6428 deps = MICROKERNEL_TEST_DEPS,
6429)
6430
6431xnnpack_unit_test(
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07006432 name = "qs8_dwconv_minmax_test",
6433 srcs = [
6434 "test/qs8-dwconv-minmax.cc",
6435 "test/dwconv-microkernel-tester.h",
6436 "src/xnnpack/AlignedAllocator.h",
6437 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
6438 deps = MICROKERNEL_TEST_DEPS + [":packing"],
6439)
6440
6441xnnpack_unit_test(
Marat Dukhan4ed53f42020-08-06 01:12:55 -07006442 name = "qs8_gavgpool_minmax_test",
6443 srcs = [
6444 "test/qs8-gavgpool-minmax.cc",
6445 "test/gavgpool-microkernel-tester.h",
6446 "src/xnnpack/AlignedAllocator.h",
6447 ] + MICROKERNEL_TEST_HDRS,
6448 deps = MICROKERNEL_TEST_DEPS,
6449)
6450
6451xnnpack_unit_test(
Marat Dukhan595e1702020-07-31 10:12:52 -07006452 name = "qs8_gemm_minmax_test",
6453 srcs = [
6454 "test/qs8-gemm-minmax.cc",
6455 "test/gemm-microkernel-tester.h",
6456 "src/xnnpack/AlignedAllocator.h",
6457 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
6458 deps = MICROKERNEL_TEST_DEPS + [":packing"],
6459)
6460
6461xnnpack_unit_test(
Marat Dukhanf9480682020-07-31 14:50:24 -07006462 name = "qs8_igemm_minmax_test",
6463 srcs = [
6464 "test/qs8-igemm-minmax.cc",
6465 "test/gemm-microkernel-tester.h",
6466 "src/xnnpack/AlignedAllocator.h",
6467 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
6468 deps = MICROKERNEL_TEST_DEPS + [":packing"],
6469)
6470
6471xnnpack_unit_test(
6472 name = "qs8_requantization_test",
6473 srcs = [
6474 "src/xnnpack/requantization-stubs.h",
6475 "test/qs8-requantization.cc",
6476 "test/requantization-tester.h",
6477 ] + MICROKERNEL_TEST_HDRS,
6478 deps = MICROKERNEL_TEST_DEPS,
6479)
6480
6481xnnpack_unit_test(
Marat Dukhand9f3ad42020-08-10 12:30:58 -07006482 name = "qs8_vadd_minmax_test",
6483 srcs = [
6484 "test/qs8-vadd-minmax.cc",
6485 "test/vadd-microkernel-tester.h",
6486 ] + MICROKERNEL_TEST_HDRS,
6487 deps = MICROKERNEL_TEST_DEPS,
6488)
6489
6490xnnpack_unit_test(
Marat Dukhan0270d9f2020-08-11 00:56:46 -07006491 name = "qs8_vaddc_minmax_test",
6492 srcs = [
6493 "test/qs8-vaddc-minmax.cc",
6494 "test/vaddc-microkernel-tester.h",
6495 ] + MICROKERNEL_TEST_HDRS,
6496 deps = MICROKERNEL_TEST_DEPS,
6497)
6498
6499xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07006500 name = "qu8_avgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006501 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07006502 "test/qu8-avgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006503 "test/avgpool-microkernel-tester.h",
6504 "src/xnnpack/AlignedAllocator.h",
6505 ] + MICROKERNEL_TEST_HDRS,
6506 deps = MICROKERNEL_TEST_DEPS,
6507)
6508
6509xnnpack_unit_test(
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07006510 name = "qu8_dwconv_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006511 srcs = [
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07006512 "test/qu8-dwconv-minmax.cc",
6513 "test/dwconv-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006514 "src/xnnpack/AlignedAllocator.h",
6515 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07006516 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07006517)
6518
6519xnnpack_unit_test(
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07006520 name = "qu8_igemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006521 srcs = [
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07006522 "test/qu8-igemm-minmax.cc",
6523 "test/gemm-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006524 "src/xnnpack/AlignedAllocator.h",
6525 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07006526 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07006527)
6528
6529xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07006530 name = "qu8_gavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006531 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07006532 "test/qu8-gavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006533 "test/gavgpool-microkernel-tester.h",
6534 "src/xnnpack/AlignedAllocator.h",
6535 ] + MICROKERNEL_TEST_HDRS,
6536 deps = MICROKERNEL_TEST_DEPS,
6537)
6538
6539xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07006540 name = "qu8_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006541 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07006542 "test/qu8-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006543 "test/gemm-microkernel-tester.h",
6544 "src/xnnpack/AlignedAllocator.h",
6545 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07006546 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07006547)
6548
6549xnnpack_unit_test(
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07006550 name = "qu8_requantization_test",
6551 srcs = [
6552 "src/xnnpack/requantization-stubs.h",
6553 "test/qu8-requantization.cc",
6554 "test/requantization-tester.h",
6555 ] + MICROKERNEL_TEST_HDRS,
6556 deps = MICROKERNEL_TEST_DEPS,
6557)
6558
6559xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07006560 name = "qu8_vadd_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006561 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07006562 "test/qu8-vadd-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006563 "test/vadd-microkernel-tester.h",
6564 ] + MICROKERNEL_TEST_HDRS,
6565 deps = MICROKERNEL_TEST_DEPS,
6566)
6567
6568xnnpack_unit_test(
6569 name = "u8_clamp_test",
6570 srcs = [
6571 "test/u8-clamp.cc",
6572 "test/clamp-microkernel-tester.h",
6573 ] + MICROKERNEL_TEST_HDRS,
6574 deps = MICROKERNEL_TEST_DEPS,
6575)
6576
6577xnnpack_unit_test(
6578 name = "u8_lut32norm_test",
6579 srcs = [
6580 "test/u8-lut32norm.cc",
6581 "test/lut-norm-microkernel-tester.h",
6582 ] + MICROKERNEL_TEST_HDRS,
6583 deps = MICROKERNEL_TEST_DEPS,
6584)
6585
6586xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07006587 name = "u8_maxpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006588 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07006589 "test/u8-maxpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006590 "test/maxpool-microkernel-tester.h",
6591 ] + MICROKERNEL_TEST_HDRS,
6592 deps = MICROKERNEL_TEST_DEPS,
6593)
6594
6595xnnpack_unit_test(
6596 name = "u8_rmax_test",
6597 srcs = [
6598 "test/u8-rmax.cc",
6599 "test/rmax-microkernel-tester.h",
6600 ] + MICROKERNEL_TEST_HDRS,
6601 deps = MICROKERNEL_TEST_DEPS,
6602)
6603
6604xnnpack_unit_test(
Marat Dukhanad71b9a2020-11-20 00:01:51 -08006605 name = "x32_depthtospace2d_chw2hwc_test",
Yury Kartynnike7841862020-11-04 18:22:18 -08006606 srcs = [
Marat Dukhanad71b9a2020-11-20 00:01:51 -08006607 "test/x32-depthtospace2d-chw2hwc.cc",
6608 "test/depthtospace-microkernel-tester.h",
Yury Kartynnike7841862020-11-04 18:22:18 -08006609 ] + MICROKERNEL_TEST_HDRS,
6610 deps = MICROKERNEL_TEST_DEPS,
6611)
6612
6613xnnpack_unit_test(
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -07006614 name = "x32_fill_test",
6615 srcs = [
6616 "test/x32-fill.cc",
6617 "test/fill-microkernel-tester.h",
6618 ] + MICROKERNEL_TEST_HDRS,
6619 deps = MICROKERNEL_TEST_DEPS,
6620)
6621
6622xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07006623 name = "x32_packx_test",
6624 srcs = [
6625 "test/x32-packx.cc",
6626 "test/pack-microkernel-tester.h",
6627 "src/xnnpack/AlignedAllocator.h",
6628 ] + MICROKERNEL_TEST_HDRS,
6629 deps = MICROKERNEL_TEST_DEPS,
6630)
6631
6632xnnpack_unit_test(
6633 name = "x32_pad_test",
6634 srcs = [
6635 "test/x32-pad.cc",
6636 "test/pad-microkernel-tester.h",
6637 ] + MICROKERNEL_TEST_HDRS,
6638 deps = MICROKERNEL_TEST_DEPS,
6639)
6640
6641xnnpack_unit_test(
6642 name = "x32_unpool_test",
6643 srcs = [
6644 "test/x32-unpool.cc",
6645 "test/unpool-microkernel-tester.h",
6646 ] + MICROKERNEL_TEST_HDRS,
6647 deps = MICROKERNEL_TEST_DEPS,
6648)
6649
6650xnnpack_unit_test(
6651 name = "x32_zip_test",
6652 srcs = [
6653 "test/x32-zip.cc",
6654 "test/zip-microkernel-tester.h",
6655 ] + MICROKERNEL_TEST_HDRS,
6656 deps = MICROKERNEL_TEST_DEPS,
6657)
6658
6659xnnpack_unit_test(
6660 name = "x8_lut_test",
6661 srcs = [
6662 "test/x8-lut.cc",
6663 "test/lut-microkernel-tester.h",
6664 ] + MICROKERNEL_TEST_HDRS,
6665 deps = MICROKERNEL_TEST_DEPS,
6666)
6667
6668xnnpack_unit_test(
6669 name = "x8_zip_test",
6670 srcs = [
6671 "test/x8-zip.cc",
6672 "test/zip-microkernel-tester.h",
6673 ] + MICROKERNEL_TEST_HDRS,
6674 deps = MICROKERNEL_TEST_DEPS,
6675)
6676
Marat Dukhan20c3b922020-03-10 03:45:06 -07006677########################## Size tests for the library #########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07006678
6679xnnpack_binary(
Marat Dukhan20c3b922020-03-10 03:45:06 -07006680 name = "operator_size_test",
6681 srcs = ["test/operator-size.c"],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07006682 deps = [":xnnpack_operators_nhwc_f32"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07006683)
6684
Marat Dukhan20c3b922020-03-10 03:45:06 -07006685xnnpack_binary(
6686 name = "subgraph_size_test",
6687 srcs = ["test/subgraph-size.c"],
6688 deps = [":XNNPACK"],
6689)
6690
6691########################### Unit tests for operators ##########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07006692
6693xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07006694 name = "abs_nc_test",
6695 srcs = [
6696 "test/abs-nc.cc",
6697 "test/abs-operator-tester.h",
6698 ],
6699 deps = OPERATOR_TEST_DEPS,
6700)
6701
6702xnnpack_unit_test(
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08006703 name = "add_nd_test",
6704 srcs = [
6705 "test/add-nd.cc",
6706 "test/binary-elementwise-operator-tester.h",
6707 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006708 deps = OPERATOR_TEST_DEPS,
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08006709)
6710
6711xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006712 name = "argmax_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006713 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006714 "test/argmax-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006715 "test/argmax-pooling-operator-tester.h",
6716 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07006717 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006718)
6719
6720xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006721 name = "average_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006722 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006723 "test/average-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006724 "test/average-pooling-operator-tester.h",
6725 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07006726 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006727)
6728
6729xnnpack_unit_test(
Marat Dukhan64e52512020-06-09 13:41:16 -07006730 name = "bankers_rounding_nc_test",
6731 srcs = [
6732 "test/bankers-rounding-nc.cc",
6733 "test/bankers-rounding-operator-tester.h",
6734 ],
6735 deps = OPERATOR_TEST_DEPS,
6736)
6737
6738xnnpack_unit_test(
6739 name = "ceiling_nc_test",
6740 srcs = [
6741 "test/ceiling-nc.cc",
6742 "test/ceiling-operator-tester.h",
6743 ],
6744 deps = OPERATOR_TEST_DEPS,
6745)
6746
6747xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006748 name = "channel_shuffle_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006749 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006750 "test/channel-shuffle-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006751 "test/channel-shuffle-operator-tester.h",
6752 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006753 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006754)
6755
6756xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006757 name = "clamp_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006758 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006759 "test/clamp-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006760 "test/clamp-operator-tester.h",
6761 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006762 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006763)
6764
6765xnnpack_unit_test(
Marat Dukhan065b11e2020-05-22 09:49:41 -07006766 name = "constant_pad_nd_test",
6767 srcs = [
6768 "test/constant-pad-nd.cc",
6769 "test/constant-pad-operator-tester.h",
6770 ],
6771 deps = OPERATOR_TEST_DEPS,
6772)
6773
6774xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006775 name = "convolution_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006776 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006777 "test/convolution-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006778 "test/convolution-operator-tester.h",
6779 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006780 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006781)
6782
6783xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006784 name = "convolution_nchw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006785 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006786 "test/convolution-nchw.cc",
6787 "test/convolution-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006788 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006789 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006790)
6791
6792xnnpack_unit_test(
Marat Dukhan4e21b272020-06-04 18:45:01 -07006793 name = "copy_nc_test",
6794 srcs = [
6795 "test/copy-nc.cc",
6796 "test/copy-operator-tester.h",
6797 ],
6798 deps = OPERATOR_TEST_DEPS,
6799)
6800
6801xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006802 name = "deconvolution_nhwc_test",
Artsiom Ablavatskic1aa2972020-12-08 11:23:34 -08006803 timeout = "moderate",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006804 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006805 "test/deconvolution-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006806 "test/deconvolution-operator-tester.h",
6807 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07006808 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006809)
6810
6811xnnpack_unit_test(
Marat Dukhan188d1042020-11-24 23:39:40 -08006812 name = "depth_to_space_nchw2nhwc_test",
Artsiom Ablavatski0f1dc182020-11-05 19:21:50 -08006813 srcs = [
Marat Dukhan188d1042020-11-24 23:39:40 -08006814 "test/depth-to-space-nchw2nhwc.cc",
Artsiom Ablavatski0f1dc182020-11-05 19:21:50 -08006815 "test/depth-to-space-operator-tester.h",
6816 ] + OPERATOR_TEST_PARAMS_HDRS,
6817 deps = OPERATOR_TEST_DEPS,
6818)
6819
6820xnnpack_unit_test(
Marat Dukhan0e521172020-11-25 13:10:04 -08006821 name = "depth_to_space_nhwc_test",
6822 srcs = [
6823 "test/depth-to-space-nhwc.cc",
6824 "test/depth-to-space-operator-tester.h",
6825 ] + OPERATOR_TEST_PARAMS_HDRS,
6826 deps = OPERATOR_TEST_DEPS,
6827)
6828
6829xnnpack_unit_test(
Marat Dukhan69180502019-12-06 15:00:31 -08006830 name = "divide_nd_test",
6831 srcs = [
6832 "test/binary-elementwise-operator-tester.h",
6833 "test/divide-nd.cc",
6834 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006835 deps = OPERATOR_TEST_DEPS,
Marat Dukhan69180502019-12-06 15:00:31 -08006836)
6837
6838xnnpack_unit_test(
Marat Dukhanb6bd4bc2020-12-01 17:01:40 -08006839 name = "elu_nc_test",
6840 srcs = [
6841 "test/elu-nc.cc",
6842 "test/elu-operator-tester.h",
6843 ],
6844 deps = OPERATOR_TEST_DEPS,
6845)
6846
6847xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006848 name = "fully_connected_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006849 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006850 "test/fully-connected-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006851 "test/fully-connected-operator-tester.h",
6852 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006853 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006854)
6855
6856xnnpack_unit_test(
Marat Dukhan64e52512020-06-09 13:41:16 -07006857 name = "floor_nc_test",
6858 srcs = [
6859 "test/floor-nc.cc",
6860 "test/floor-operator-tester.h",
6861 ],
6862 deps = OPERATOR_TEST_DEPS,
6863)
6864
6865xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006866 name = "global_average_pooling_nwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006867 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006868 "test/global-average-pooling-nwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006869 "test/global-average-pooling-operator-tester.h",
Marat Dukhanef61d022020-06-19 13:54:49 -07006870 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07006871 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006872)
6873
6874xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006875 name = "global_average_pooling_ncw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006876 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006877 "test/global-average-pooling-ncw.cc",
6878 "test/global-average-pooling-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006879 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006880 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006881)
6882
6883xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006884 name = "hardswish_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006885 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006886 "test/hardswish-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006887 "test/hardswish-operator-tester.h",
6888 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006889 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006890)
6891
6892xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006893 name = "leaky_relu_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006894 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006895 "test/leaky-relu-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006896 "test/leaky-relu-operator-tester.h",
6897 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006898 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006899)
6900
6901xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006902 name = "max_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006903 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006904 "test/max-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006905 "test/max-pooling-operator-tester.h",
6906 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07006907 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006908)
6909
6910xnnpack_unit_test(
Marat Dukhan79e7f842019-12-05 14:35:50 -08006911 name = "maximum_nd_test",
6912 srcs = [
6913 "test/binary-elementwise-operator-tester.h",
6914 "test/maximum-nd.cc",
6915 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006916 deps = OPERATOR_TEST_DEPS,
Marat Dukhan79e7f842019-12-05 14:35:50 -08006917)
6918
6919xnnpack_unit_test(
6920 name = "minimum_nd_test",
6921 srcs = [
6922 "test/binary-elementwise-operator-tester.h",
6923 "test/minimum-nd.cc",
6924 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006925 deps = OPERATOR_TEST_DEPS,
Marat Dukhan79e7f842019-12-05 14:35:50 -08006926)
6927
6928xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006929 name = "multiply_nd_test",
Marat Dukhanca2733c2019-11-15 23:21:17 -08006930 srcs = [
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08006931 "test/binary-elementwise-operator-tester.h",
Marat Dukhanefc47b82019-11-18 09:25:38 -08006932 "test/multiply-nd.cc",
Marat Dukhanca2733c2019-11-15 23:21:17 -08006933 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006934 deps = OPERATOR_TEST_DEPS,
Marat Dukhanca2733c2019-11-15 23:21:17 -08006935)
6936
6937xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07006938 name = "negate_nc_test",
6939 srcs = [
6940 "test/negate-nc.cc",
6941 "test/negate-operator-tester.h",
6942 ],
6943 deps = OPERATOR_TEST_DEPS,
6944)
6945
6946xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006947 name = "prelu_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006948 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006949 "test/prelu-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006950 "test/prelu-operator-tester.h",
6951 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07006952 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006953)
6954
6955xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006956 name = "resize_bilinear_nhwc_test",
Marat Dukhan69722492019-11-11 19:55:50 -08006957 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006958 "test/resize-bilinear-nhwc.cc",
Marat Dukhan69722492019-11-11 19:55:50 -08006959 "test/resize-bilinear-operator-tester.h",
6960 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07006961 deps = OPERATOR_TEST_DEPS,
Marat Dukhan69722492019-11-11 19:55:50 -08006962)
6963
6964xnnpack_unit_test(
Artsiom Ablavatski97918102020-10-27 15:52:59 -07006965 name = "resize_bilinear_nchw_test",
6966 srcs = [
6967 "test/resize-bilinear-nchw.cc",
6968 "test/resize-bilinear-operator-tester.h",
6969 ] + OPERATOR_TEST_PARAMS_HDRS,
6970 deps = OPERATOR_TEST_DEPS,
6971)
6972
6973xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006974 name = "sigmoid_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006975 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006976 "test/sigmoid-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006977 "test/sigmoid-operator-tester.h",
6978 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006979 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006980)
6981
6982xnnpack_unit_test(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08006983 name = "softmax_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006984 srcs = [
Marat Dukhanfd8e6892020-01-27 15:25:25 -08006985 "test/softmax-nc.cc",
6986 "test/softmax-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006987 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006988 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006989)
6990
6991xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07006992 name = "square_nc_test",
6993 srcs = [
6994 "test/square-nc.cc",
6995 "test/square-operator-tester.h",
6996 ],
6997 deps = OPERATOR_TEST_DEPS,
6998)
6999
7000xnnpack_unit_test(
Marat Dukhan6804bbd2020-06-30 19:26:11 -07007001 name = "square_root_nc_test",
7002 srcs = [
7003 "test/square-root-nc.cc",
7004 "test/square-root-operator-tester.h",
7005 ],
7006 deps = OPERATOR_TEST_DEPS,
7007)
7008
7009xnnpack_unit_test(
Marat Dukhanf7399262020-06-05 10:58:44 -07007010 name = "squared_difference_nd_test",
7011 srcs = [
7012 "test/binary-elementwise-operator-tester.h",
7013 "test/squared-difference-nd.cc",
7014 ],
7015 deps = OPERATOR_TEST_DEPS,
7016)
7017
7018xnnpack_unit_test(
Marat Dukhan05f3f6d2019-12-03 15:13:53 -08007019 name = "subtract_nd_test",
7020 srcs = [
7021 "test/binary-elementwise-operator-tester.h",
7022 "test/subtract-nd.cc",
7023 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07007024 deps = OPERATOR_TEST_DEPS,
Marat Dukhan05f3f6d2019-12-03 15:13:53 -08007025)
7026
7027xnnpack_unit_test(
Marat Dukhan64e52512020-06-09 13:41:16 -07007028 name = "truncation_nc_test",
7029 srcs = [
7030 "test/truncation-nc.cc",
7031 "test/truncation-operator-tester.h",
7032 ],
7033 deps = OPERATOR_TEST_DEPS,
7034)
7035
7036xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08007037 name = "unpooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007038 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08007039 "test/unpooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007040 "test/unpooling-operator-tester.h",
7041 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07007042 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07007043)
7044
Chao Mei6ddfc602020-05-13 22:29:36 -07007045############################### Misc unit tests ###############################
7046
7047xnnpack_unit_test(
7048 name = "memory_planner_test",
7049 srcs = [
7050 "test/memory-planner-test.cc",
7051 ],
7052 deps = [
7053 ":XNNPACK",
7054 ":memory_planner",
7055 ],
7056)
7057
XNNPACK Teamab8c4c82020-10-09 08:05:51 -07007058xnnpack_unit_test(
7059 name = "subgraph_nchw_test",
7060 srcs = [
7061 "src/xnnpack/subgraph.h",
7062 "test/subgraph-nchw.cc",
7063 "test/subgraph-tester.h",
7064 ],
7065 deps = [
7066 ":XNNPACK",
7067 ],
7068)
7069
Marat Dukhan08c4a432019-10-03 09:29:21 -07007070############################# Build configurations #############################
7071
Marat Dukhanb8642352019-10-30 15:43:02 -07007072# Enables usage of assembly kernels.
Marat Dukhan08c4a432019-10-03 09:29:21 -07007073config_setting(
Marat Dukhanb8642352019-10-30 15:43:02 -07007074 name = "xnn_enable_assembly_explicit_true",
7075 define_values = {"xnn_enable_assembly": "true"},
7076)
7077
7078# Disables usage of assembly kernels.
7079config_setting(
7080 name = "xnn_enable_assembly_explicit_false",
7081 define_values = {"xnn_enable_assembly": "false"},
7082)
7083
Marat Dukhan9de90e02020-06-18 16:04:12 -07007084# Enables usage of sparse inference.
7085config_setting(
7086 name = "xnn_enable_sparse_explicit_true",
7087 define_values = {"xnn_enable_sparse": "true"},
7088)
7089
7090# Disables usage of sparse inference.
7091config_setting(
7092 name = "xnn_enable_sparse_explicit_false",
7093 define_values = {"xnn_enable_sparse": "false"},
7094)
7095
Marat Dukhan05702cf2020-03-26 15:41:33 -07007096# Disables usage of HMP-aware optimizations.
7097config_setting(
7098 name = "xnn_enable_hmp_explicit_false",
7099 define_values = {"xnn_enable_hmp": "false"},
7100)
7101
Chao Mei6ddfc602020-05-13 22:29:36 -07007102# Enable usage of optimized memory allocation
7103config_setting(
7104 name = "xnn_enable_memopt_explicit_true",
7105 define_values = {"xnnpack_enable_memopt": "true"},
7106)
7107
7108# Disable usage of optimized memory allocation
7109config_setting(
7110 name = "xnn_enable_memopt_explicit_false",
7111 define_values = {"xnnpack_enable_memopt": "false"},
7112)
7113
Marat Dukhanb8642352019-10-30 15:43:02 -07007114# Builds with -c dbg
7115config_setting(
7116 name = "debug_build",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007117 values = {
Marat Dukhanb8642352019-10-30 15:43:02 -07007118 "compilation_mode": "dbg",
7119 },
7120)
7121
7122# Builds with -c opt
7123config_setting(
7124 name = "optimized_build",
7125 values = {
7126 "compilation_mode": "opt",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007127 },
7128)
7129
7130config_setting(
Marat Dukhanb8642352019-10-30 15:43:02 -07007131 name = "linux_k8",
7132 values = {"cpu": "k8"},
7133)
7134
7135config_setting(
Marat Dukhan582094e2020-04-30 17:21:25 -07007136 name = "linux_arm",
7137 values = {"cpu": "arm"},
Marat Dukhan4e45e662019-10-03 15:40:24 -07007138)
7139
7140config_setting(
Marat Dukhanf0bd4de2020-06-15 15:53:19 -07007141 name = "linux_armeabi",
7142 values = {"cpu": "armeabi"},
7143)
7144
7145config_setting(
Terry Heo68eef3f2020-04-13 22:53:52 -07007146 name = "linux_armhf",
7147 values = {"cpu": "armhf"},
7148)
7149
7150config_setting(
Marat Dukhana720e932020-06-10 13:01:11 -07007151 name = "linux_armv7a",
7152 values = {"cpu": "armv7a"},
7153)
7154
7155config_setting(
Marat Dukhan582094e2020-04-30 17:21:25 -07007156 name = "linux_aarch64",
7157 values = {"cpu": "aarch64"},
7158)
7159
7160config_setting(
Marat Dukhan08c4a432019-10-03 09:29:21 -07007161 name = "android",
7162 values = {"crosstool_top": "//external:android/crosstool"},
7163)
7164
7165config_setting(
7166 name = "android_armv7",
7167 values = {
7168 "crosstool_top": "//external:android/crosstool",
7169 "cpu": "armeabi-v7a",
7170 },
7171)
7172
7173config_setting(
7174 name = "android_arm64",
7175 values = {
7176 "crosstool_top": "//external:android/crosstool",
7177 "cpu": "arm64-v8a",
7178 },
7179)
7180
7181config_setting(
7182 name = "android_x86",
7183 values = {
7184 "crosstool_top": "//external:android/crosstool",
7185 "cpu": "x86",
7186 },
7187)
7188
7189config_setting(
7190 name = "android_x86_64",
7191 values = {
7192 "crosstool_top": "//external:android/crosstool",
7193 "cpu": "x86_64",
7194 },
7195)
7196
7197config_setting(
Marat Dukhan10a38082020-04-17 03:58:35 -07007198 name = "windows_x86_64",
7199 values = {"cpu": "x64_windows"},
Marat Dukhan9fe932e2020-04-11 17:14:15 -07007200)
7201
7202config_setting(
Marat Dukhan10a38082020-04-17 03:58:35 -07007203 name = "windows_x86_64_clang",
7204 values = {
7205 "compiler": "clang-cl",
7206 "cpu": "x64_windows",
7207 },
7208)
7209
7210config_setting(
7211 name = "windows_x86_64_mingw",
7212 values = {
7213 "compiler": "mingw-gcc",
7214 "cpu": "x64_windows",
7215 },
7216)
7217
7218config_setting(
7219 name = "windows_x86_64_msys",
7220 values = {
7221 "compiler": "msys-gcc",
7222 "cpu": "x64_windows",
7223 },
Marat Dukhan9fe932e2020-04-11 17:14:15 -07007224)
7225
7226config_setting(
Marat Dukhan885ca242019-10-07 09:17:32 -07007227 name = "macos_x86_64",
7228 values = {
7229 "apple_platform_type": "macos",
7230 "cpu": "darwin",
7231 },
7232)
7233
7234config_setting(
Marat Dukhan08c4a432019-10-03 09:29:21 -07007235 name = "emscripten",
Marat Dukhan1a729ec2019-10-07 09:31:44 -07007236 values = {"crosstool_top": "//toolchain:emscripten"},
Marat Dukhan08c4a432019-10-03 09:29:21 -07007237)
7238
7239config_setting(
7240 name = "emscripten_wasm",
7241 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07007242 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007243 "cpu": "wasm",
7244 },
7245)
7246
7247config_setting(
7248 name = "emscripten_wasmsimd",
7249 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07007250 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007251 "cpu": "wasm",
Marat Dukhan81c62602020-05-29 13:22:49 -07007252 "copt": "-msimd128",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007253 },
7254)
7255
7256config_setting(
Marat Dukhan1498d1d2020-02-11 20:00:05 -08007257 name = "ios_armv7",
7258 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08007259 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08007260 "cpu": "ios_armv7",
7261 },
7262)
7263
7264config_setting(
7265 name = "ios_arm64",
7266 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08007267 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08007268 "cpu": "ios_arm64",
7269 },
7270)
7271
7272config_setting(
7273 name = "ios_arm64e",
7274 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08007275 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08007276 "cpu": "ios_arm64e",
7277 },
7278)
7279
7280config_setting(
7281 name = "ios_x86",
7282 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08007283 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08007284 "cpu": "ios_i386",
7285 },
7286)
7287
7288config_setting(
7289 name = "ios_x86_64",
7290 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08007291 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08007292 "cpu": "ios_x86_64",
7293 },
7294)
7295
7296config_setting(
7297 name = "watchos_armv7k",
7298 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08007299 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08007300 "cpu": "watchos_armv7k",
7301 },
7302)
7303
7304config_setting(
7305 name = "watchos_arm64_32",
7306 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08007307 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08007308 "cpu": "watchos_arm64_32",
7309 },
7310)
7311
7312config_setting(
7313 name = "watchos_x86",
7314 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08007315 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08007316 "cpu": "watchos_i386",
7317 },
7318)
7319
7320config_setting(
7321 name = "watchos_x86_64",
7322 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08007323 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08007324 "cpu": "watchos_x86_64",
7325 },
7326)
7327
7328config_setting(
7329 name = "tvos_arm64",
7330 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08007331 "apple_platform_type": "tvos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08007332 "cpu": "tvos_arm64",
7333 },
7334)
7335
7336config_setting(
7337 name = "tvos_x86_64",
7338 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08007339 "apple_platform_type": "tvos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08007340 "cpu": "tvos_x86_64",
7341 },
7342)