blob: c9ca23fcb6fa39cfdfd9588b8101b06e7d60b578 [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 Barchard3b800452020-11-22 12:12:35 -0800773 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-1x4-acc2.c",
774 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-1x4-acc3.c",
775 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-1x4-acc4.c",
776 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-1x4.c",
777 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-2x4-acc2.c",
778 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-2x4.c",
779 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-3x4.c",
780 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-4x4.c",
781 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-5x4.c",
782 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-6x4.c",
783 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-1x4-acc2.c",
784 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-1x4-acc3.c",
785 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-1x4-acc4.c",
786 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-1x4.c",
787 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-2x4-acc2.c",
788 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-2x4.c",
789 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-3x4.c",
790 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-4x4.c",
791 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-5x4.c",
792 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-6x4.c",
Frank Barchardb88d0112020-12-04 09:07:13 -0800793 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-arm-1x4-acc2.c",
794 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-arm-1x4-acc3.c",
795 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-arm-1x4-acc4.c",
796 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-arm-1x4.c",
797 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-arm-2x4-acc2.c",
798 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-arm-2x4.c",
799 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-arm-3x4.c",
800 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-arm-4x4.c",
801 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-x86-1x4-acc2.c",
802 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-x86-1x4-acc3.c",
803 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-x86-1x4-acc4.c",
804 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-x86-1x4.c",
805 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-x86-2x4-acc2.c",
806 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-x86-2x4.c",
807 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-x86-3x4.c",
808 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-x86-4x4.c",
Frank Barchard20a07412020-11-30 23:30:00 -0800809 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-1x4-acc2.c",
810 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-1x4-acc3.c",
811 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-1x4-acc4.c",
812 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-1x4-acc5.c",
813 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-1x4.c",
814 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-2x4-acc2.c",
815 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-2x4-acc3.c",
816 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-2x4.c",
817 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-3x4-acc2.c",
818 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-3x4.c",
819 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-4x4-acc2.c",
820 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-4x4.c",
821 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-5x4.c",
822 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-1x4-acc2.c",
823 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-1x4-acc3.c",
824 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-1x4-acc4.c",
825 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-1x4-acc5.c",
826 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-1x4.c",
827 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-2x4-acc2.c",
828 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-2x4-acc3.c",
829 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-2x4.c",
830 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-3x4-acc2.c",
831 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-3x4.c",
832 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-4x4-acc2.c",
833 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-4x4.c",
834 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-5x4.c",
Frank Barchard729f07b2020-12-03 16:14:33 -0800835 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-arm-1x4-acc2.c",
836 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-arm-1x4-acc3.c",
837 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-arm-1x4-acc4.c",
838 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-arm-1x4-acc5.c",
839 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-arm-1x4.c",
840 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-arm-2x4-acc2.c",
841 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-arm-2x4-acc3.c",
842 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-arm-2x4.c",
843 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-arm-3x4-acc2.c",
844 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-arm-3x4.c",
845 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-x86-1x4-acc2.c",
846 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-x86-1x4-acc3.c",
847 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-x86-1x4-acc4.c",
848 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-x86-1x4-acc5.c",
849 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-x86-1x4.c",
850 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-x86-2x4-acc2.c",
851 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-x86-2x4-acc3.c",
852 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-x86-2x4.c",
853 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-x86-3x4-acc2.c",
854 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-x86-3x4.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700855 "src/f32-gavgpool-cw/wasmsimd-arm-x4.c",
856 "src/f32-gavgpool-cw/wasmsimd-x86-x4.c",
Marat Dukhanc6016802020-07-16 18:51:28 -0700857 "src/f32-gavgpool/7p7x-minmax-wasmsimd-arm-c4.c",
858 "src/f32-gavgpool/7p7x-minmax-wasmsimd-x86-c4.c",
859 "src/f32-gavgpool/7x-minmax-wasmsimd-arm-c4.c",
860 "src/f32-gavgpool/7x-minmax-wasmsimd-x86-c4.c",
Frank 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 Dukhan102e9ea2020-12-05 22:35:18 -08001336PSIMD_ACCMATH_UKERNELS = [
1337 "src/qs8-requantization/fp32-psimd.c",
1338 "src/qs8-requantization/precise-psimd.c",
1339 "src/qu8-requantization/fp32-psimd.c",
1340 "src/qu8-requantization/precise-psimd.c",
1341]
1342
Marat Dukhan08c4a432019-10-03 09:29:21 -07001343# ISA-specific micro-kernels
1344NEON_UKERNELS = [
Marat Dukhanef25c6d2020-07-24 00:59:40 -07001345 "src/f32-argmaxpool/4x-neon-c4.c",
1346 "src/f32-argmaxpool/9p8x-neon-c4.c",
1347 "src/f32-argmaxpool/9x-neon-c4.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001348 "src/f32-avgpool/9p8x-minmax-neon-c4.c",
1349 "src/f32-avgpool/9x-minmax-neon-c4.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001350 "src/f32-clamp/gen/neon-x4.c",
1351 "src/f32-clamp/gen/neon-x8.c",
Marat Dukhan56b10cd2020-05-18 09:35:49 -07001352 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neon-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001353 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neon-2x2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001354 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neon-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001355 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neon-2x2.c",
Marat Dukhan56b10cd2020-05-18 09:35:49 -07001356 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neon-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001357 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neon-2x2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001358 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neon-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001359 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neon-2x2.c",
Marat Dukhanc7634882020-12-07 15:11:12 -08001360 "src/f32-conv-hwc2chw/3x3s2p1c3x4-neon-2x2.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -07001361 "src/f32-dwconv/gen/up4x4-minmax-neon-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001362 "src/f32-dwconv/gen/up4x4-minmax-neon.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001363 "src/f32-dwconv/gen/up4x9-minmax-neon-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001364 "src/f32-dwconv/gen/up4x9-minmax-neon.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -07001365 "src/f32-dwconv/gen/up4x25-minmax-neon-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001366 "src/f32-dwconv/gen/up4x25-minmax-neon.c",
1367 "src/f32-dwconv/gen/up8x4-minmax-neon-acc2.c",
1368 "src/f32-dwconv/gen/up8x4-minmax-neon.c",
1369 "src/f32-dwconv/gen/up8x9-minmax-neon-acc2.c",
1370 "src/f32-dwconv/gen/up8x9-minmax-neon.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -07001371 "src/f32-dwconv/gen/up8x25-minmax-neon-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001372 "src/f32-dwconv/gen/up8x25-minmax-neon.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001373 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-1x4-acc2.c",
1374 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-1x4-acc3.c",
1375 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-1x4-acc4.c",
Marat Dukhanc581e482020-10-24 01:28:11 -07001376 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-1x4.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001377 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-2x4-acc2.c",
Marat Dukhanc581e482020-10-24 01:28:11 -07001378 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-2x4.c",
1379 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-3x4.c",
1380 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-4x4.c",
1381 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-5x4.c",
1382 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-6x4.c",
Marat Dukhan82f0c322020-10-25 19:17:35 -07001383 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neon-1x4-acc2.c",
1384 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neon-1x4-acc3.c",
1385 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neon-1x4-acc4.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001386 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neon-1x4.c",
Marat Dukhan82f0c322020-10-25 19:17:35 -07001387 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neon-2x4-acc2.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001388 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neon-2x4.c",
1389 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neon-3x4.c",
1390 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neon-4x4.c",
Marat Dukhan149f0ea2020-10-26 12:50:33 -07001391 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-1x4-acc2.c",
1392 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-1x4-acc3.c",
1393 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-1x4-acc4.c",
1394 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-1x4-acc5.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001395 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-1x4.c",
Marat Dukhan149f0ea2020-10-26 12:50:33 -07001396 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-2x4-acc2.c",
1397 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-2x4-acc3.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001398 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-2x4.c",
Marat Dukhan149f0ea2020-10-26 12:50:33 -07001399 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-3x4-acc2.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001400 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-3x4.c",
Marat Dukhan149f0ea2020-10-26 12:50:33 -07001401 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-4x4-acc2.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001402 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-4x4.c",
1403 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-5x4.c",
Marat Dukhan30d4b252020-10-29 16:33:22 -07001404 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neon-1x4-acc2.c",
1405 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neon-1x4-acc3.c",
1406 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neon-1x4-acc4.c",
1407 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neon-1x4-acc5.c",
1408 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neon-1x4.c",
1409 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neon-2x4-acc2.c",
1410 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neon-2x4-acc3.c",
1411 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neon-2x4.c",
Marat Dukhan30d4b252020-10-29 16:33:22 -07001412 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neon-3x4-acc2.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07001413 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neon-3x4.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -07001414 "src/f32-gavgpool-cw/neon-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001415 "src/f32-gavgpool/7p7x-minmax-neon-c4.c",
1416 "src/f32-gavgpool/7x-minmax-neon-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001417 "src/f32-gemm/gen-inc/1x8inc-minmax-neon-dup-ld64.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001418 "src/f32-gemm/gen-inc/1x8inc-minmax-neon-lane-ld64.c",
1419 "src/f32-gemm/gen-inc/1x8s4inc-minmax-neon.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001420 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-dup-ld64.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001421 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-dup-ld128.c",
1422 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-lane-ld64.c",
1423 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-lane-ld128.c",
1424 "src/f32-gemm/gen-inc/4x8s4inc-minmax-neon.c",
1425 "src/f32-gemm/gen-inc/5x8inc-minmax-neon-lane-ld64.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001426 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-dup-ld64.c",
1427 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-dup-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001428 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-lane-ld64.c",
1429 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-lane-ld128.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001430 "src/f32-gemm/gen-inc/6x8s4inc-minmax-neon.c",
1431 "src/f32-gemm/gen-inc/8x8s4inc-minmax-neon.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001432 "src/f32-gemm/gen/1x8-minmax-neon-dup-ld64.c",
1433 "src/f32-gemm/gen/1x8-minmax-neon-lane-ld64.c",
1434 "src/f32-gemm/gen/1x8s4-minmax-neon.c",
1435 "src/f32-gemm/gen/4x2-minmax-neon-lane-ld64.c",
1436 "src/f32-gemm/gen/4x8-minmax-neon-dup-ld64.c",
1437 "src/f32-gemm/gen/4x8-minmax-neon-dup-ld128.c",
1438 "src/f32-gemm/gen/4x8-minmax-neon-lane-ld64.c",
1439 "src/f32-gemm/gen/4x8-minmax-neon-lane-ld128.c",
1440 "src/f32-gemm/gen/4x8s4-minmax-neon.c",
1441 "src/f32-gemm/gen/5x8-minmax-neon-lane-ld64.c",
1442 "src/f32-gemm/gen/6x8-minmax-neon-dup-ld64.c",
1443 "src/f32-gemm/gen/6x8-minmax-neon-dup-ld128.c",
1444 "src/f32-gemm/gen/6x8-minmax-neon-lane-ld64.c",
1445 "src/f32-gemm/gen/6x8-minmax-neon-lane-ld128.c",
1446 "src/f32-gemm/gen/6x8s4-minmax-neon.c",
1447 "src/f32-gemm/gen/8x8s4-minmax-neon.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07001448 "src/f32-hswish/gen/hswish-neon-x4.c",
1449 "src/f32-hswish/gen/hswish-neon-x8.c",
1450 "src/f32-hswish/gen/hswish-neon-x16.c",
Marat Dukhan660fd192020-03-10 04:55:30 -07001451 "src/f32-ibilinear/gen/neon-c4.c",
1452 "src/f32-ibilinear/gen/neon-c8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001453 "src/f32-igemm/gen/1x8-minmax-neon-dup-ld64.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001454 "src/f32-igemm/gen/1x8-minmax-neon-lane-ld64.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001455 "src/f32-igemm/gen/1x8s4-minmax-neon.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001456 "src/f32-igemm/gen/4x2-minmax-neon-lane-ld64.c",
1457 "src/f32-igemm/gen/4x4-minmax-neon-lane-ld64.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001458 "src/f32-igemm/gen/4x8-minmax-neon-dup-ld64.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001459 "src/f32-igemm/gen/4x8-minmax-neon-dup-ld128.c",
1460 "src/f32-igemm/gen/4x8-minmax-neon-lane-ld64.c",
1461 "src/f32-igemm/gen/4x8-minmax-neon-lane-ld128.c",
1462 "src/f32-igemm/gen/4x8s4-minmax-neon.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001463 "src/f32-igemm/gen/6x8-minmax-neon-dup-ld64.c",
1464 "src/f32-igemm/gen/6x8-minmax-neon-dup-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001465 "src/f32-igemm/gen/6x8-minmax-neon-lane-ld64.c",
1466 "src/f32-igemm/gen/6x8-minmax-neon-lane-ld128.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001467 "src/f32-igemm/gen/6x8s4-minmax-neon.c",
1468 "src/f32-igemm/gen/8x8s4-minmax-neon.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001469 "src/f32-maxpool/9p8x-minmax-neon-c4.c",
1470 "src/f32-pavgpool/9p8x-minmax-neon-c4.c",
1471 "src/f32-pavgpool/9x-minmax-neon-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001472 "src/f32-ppmm/gen/4x8-minmax-neon.c",
1473 "src/f32-ppmm/gen/8x8-minmax-neon.c",
Frank Barcharda5316982020-07-23 13:19:28 -07001474 "src/f32-prelu/gen/neon-1x4.c",
1475 "src/f32-prelu/gen/neon-1x8.c",
1476 "src/f32-prelu/gen/neon-1x16.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001477 "src/f32-prelu/gen/neon-2x4.c",
1478 "src/f32-prelu/gen/neon-2x8.c",
Frank Barcharda5316982020-07-23 13:19:28 -07001479 "src/f32-prelu/gen/neon-2x16.c",
1480 "src/f32-prelu/gen/neon-4x4.c",
1481 "src/f32-prelu/gen/neon-4x8.c",
1482 "src/f32-prelu/gen/neon-4x16.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001483 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x4.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001484 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x8-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001485 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x8.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001486 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12-acc2.c",
1487 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001488 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001489 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16-acc2.c",
1490 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001491 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001492 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20-acc2.c",
1493 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001494 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20.c",
1495 "src/f32-raddstoreexpminusmax/gen/neon-p5-x4.c",
1496 "src/f32-raddstoreexpminusmax/gen/neon-p5-x8-acc2.c",
1497 "src/f32-raddstoreexpminusmax/gen/neon-p5-x8.c",
1498 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12-acc2.c",
1499 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12-acc3.c",
1500 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12.c",
1501 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16-acc2.c",
1502 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16-acc4.c",
1503 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16.c",
1504 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20-acc2.c",
1505 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20-acc5.c",
1506 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07001507 "src/f32-relu/gen/neon-x4.c",
1508 "src/f32-relu/gen/neon-x8.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001509 "src/f32-rmax/neon.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -08001510 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x4.c",
1511 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x8.c",
1512 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x12.c",
1513 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x16.c",
1514 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x20.c",
1515 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x24.c",
1516 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x4.c",
1517 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x8.c",
1518 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x12.c",
1519 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x16.c",
1520 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x20.c",
1521 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x24.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001522 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x4.c",
1523 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x8.c",
1524 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x12.c",
1525 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x16.c",
1526 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x20.c",
1527 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x24.c",
Marat Dukhan5b86c432020-12-06 19:15:03 -08001528 "src/f32-spmm/gen/4x1-minmax-neon-pipelined.c",
1529 "src/f32-spmm/gen/4x1-minmax-neon-x2.c",
1530 "src/f32-spmm/gen/4x1-minmax-neon.c",
1531 "src/f32-spmm/gen/8x1-minmax-neon-pipelined.c",
1532 "src/f32-spmm/gen/8x1-minmax-neon-x2.c",
1533 "src/f32-spmm/gen/8x1-minmax-neon.c",
1534 "src/f32-spmm/gen/12x1-minmax-neon.c",
1535 "src/f32-spmm/gen/16x1-minmax-neon-pipelined.c",
1536 "src/f32-spmm/gen/16x1-minmax-neon-x2.c",
1537 "src/f32-spmm/gen/16x1-minmax-neon.c",
1538 "src/f32-spmm/gen/32x1-minmax-neon-pipelined.c",
1539 "src/f32-spmm/gen/32x1-minmax-neon-x2.c",
1540 "src/f32-spmm/gen/32x1-minmax-neon.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001541 "src/f32-vbinary/gen/vadd-minmax-neon-x4.c",
1542 "src/f32-vbinary/gen/vadd-minmax-neon-x8.c",
1543 "src/f32-vbinary/gen/vaddc-minmax-neon-x4.c",
1544 "src/f32-vbinary/gen/vaddc-minmax-neon-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -08001545 "src/f32-vbinary/gen/vmax-neon-x4.c",
1546 "src/f32-vbinary/gen/vmax-neon-x8.c",
1547 "src/f32-vbinary/gen/vmaxc-neon-x4.c",
1548 "src/f32-vbinary/gen/vmaxc-neon-x8.c",
1549 "src/f32-vbinary/gen/vmin-neon-x4.c",
1550 "src/f32-vbinary/gen/vmin-neon-x8.c",
1551 "src/f32-vbinary/gen/vminc-neon-x4.c",
1552 "src/f32-vbinary/gen/vminc-neon-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001553 "src/f32-vbinary/gen/vmul-minmax-neon-x4.c",
1554 "src/f32-vbinary/gen/vmul-minmax-neon-x8.c",
1555 "src/f32-vbinary/gen/vmulc-minmax-neon-x4.c",
1556 "src/f32-vbinary/gen/vmulc-minmax-neon-x8.c",
1557 "src/f32-vbinary/gen/vrsubc-minmax-neon-x4.c",
1558 "src/f32-vbinary/gen/vrsubc-minmax-neon-x8.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07001559 "src/f32-vbinary/gen/vsqrdiff-neon-x4.c",
1560 "src/f32-vbinary/gen/vsqrdiff-neon-x8.c",
1561 "src/f32-vbinary/gen/vsqrdiffc-neon-x4.c",
1562 "src/f32-vbinary/gen/vsqrdiffc-neon-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001563 "src/f32-vbinary/gen/vsub-minmax-neon-x4.c",
1564 "src/f32-vbinary/gen/vsub-minmax-neon-x8.c",
1565 "src/f32-vbinary/gen/vsubc-minmax-neon-x4.c",
1566 "src/f32-vbinary/gen/vsubc-minmax-neon-x8.c",
Marat Dukhaned6baaf2020-12-01 15:07:08 -08001567 "src/f32-velu/gen/velu-neon-rr2-lut16-p3-x4.c",
1568 "src/f32-velu/gen/velu-neon-rr2-lut16-p3-x8.c",
1569 "src/f32-velu/gen/velu-neon-rr2-lut16-p3-x12.c",
1570 "src/f32-velu/gen/velu-neon-rr2-lut16-p3-x16.c",
1571 "src/f32-velu/gen/velu-neon-rr2-lut16-p3-x20.c",
1572 "src/f32-velu/gen/velu-neon-rr2-lut16-p3-x24.c",
1573 "src/f32-velu/gen/velu-neon-rr2-p6-x4.c",
1574 "src/f32-velu/gen/velu-neon-rr2-p6-x8.c",
1575 "src/f32-velu/gen/velu-neon-rr2-p6-x12.c",
1576 "src/f32-velu/gen/velu-neon-rr2-p6-x16.c",
1577 "src/f32-velu/gen/velu-neon-rr2-p6-x20.c",
1578 "src/f32-velu/gen/velu-neon-rr2-p6-x24.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07001579 "src/f32-vlrelu/gen/vlrelu-neon-x4.c",
1580 "src/f32-vlrelu/gen/vlrelu-neon-x8.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001581 "src/f32-vmulcaddc/gen/c4-minmax-neon-2x.c",
1582 "src/f32-vmulcaddc/gen/c8-minmax-neon-2x.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07001583 "src/f32-vrnd/gen/vrndd-neon-x4.c",
1584 "src/f32-vrnd/gen/vrndd-neon-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001585 "src/f32-vrnd/gen/vrndne-neon-x4.c",
1586 "src/f32-vrnd/gen/vrndne-neon-x8.c",
1587 "src/f32-vrnd/gen/vrndu-neon-x4.c",
1588 "src/f32-vrnd/gen/vrndu-neon-x8.c",
1589 "src/f32-vrnd/gen/vrndz-neon-x4.c",
1590 "src/f32-vrnd/gen/vrndz-neon-x8.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07001591 "src/f32-vunary/gen/vabs-neon-x4.c",
1592 "src/f32-vunary/gen/vabs-neon-x8.c",
1593 "src/f32-vunary/gen/vneg-neon-x4.c",
1594 "src/f32-vunary/gen/vneg-neon-x8.c",
1595 "src/f32-vunary/gen/vsqr-neon-x4.c",
1596 "src/f32-vunary/gen/vsqr-neon-x8.c",
Marat Dukhande390d42020-11-29 19:32:18 -08001597 "src/math/expm1minus-neon-rr2-lut16-p3.c",
1598 "src/math/expm1minus-neon-rr2-p6.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001599 "src/math/roundd-neon-addsub.c",
1600 "src/math/roundd-neon-cvt.c",
1601 "src/math/roundne-neon-addsub.c",
1602 "src/math/roundu-neon-addsub.c",
1603 "src/math/roundu-neon-cvt.c",
1604 "src/math/roundz-neon-addsub.c",
1605 "src/math/roundz-neon-cvt.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001606 "src/math/sigmoid-neon-rr2-lut64-p2-nr2recps.c",
1607 "src/math/sigmoid-neon-rr2-lut2048-p1-nr2recps.c",
1608 "src/math/sigmoid-neon-rr2-p5-nr2recps.c",
1609 "src/math/sqrt-neon-nr1rsqrts.c",
1610 "src/math/sqrt-neon-nr2rsqrts.c",
1611 "src/math/sqrt-neon-nr3rsqrts.c",
Marat Dukhan023bcf92020-08-10 12:40:50 -07001612 "src/qs8-dwconv/gen/up8x9-minmax-neon-mul16.c",
1613 "src/qs8-dwconv/gen/up16x9-minmax-neon-mul16.c",
1614 "src/qs8-dwconv/gen/up24x9-minmax-neon-mul16.c",
1615 "src/qs8-dwconv/gen/up32x9-minmax-neon-mul16.c",
Marat Dukhan281262d2020-08-10 13:23:21 -07001616 "src/qs8-gavgpool/gen/7p7x-minmax-neon-c8-acc2.c",
1617 "src/qs8-gavgpool/gen/7p7x-minmax-neon-c16-acc2.c",
1618 "src/qs8-gavgpool/gen/7p7x-minmax-neon-c24-acc2.c",
1619 "src/qs8-gavgpool/gen/7p7x-minmax-neon-c32-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001620 "src/qs8-gavgpool/gen/7x-minmax-neon-c8-acc2.c",
1621 "src/qs8-gavgpool/gen/7x-minmax-neon-c16-acc2.c",
1622 "src/qs8-gavgpool/gen/7x-minmax-neon-c24-acc2.c",
1623 "src/qs8-gavgpool/gen/7x-minmax-neon-c32-acc2.c",
Marat Dukhan40bbafe2020-08-04 02:04:22 -07001624 "src/qs8-gemm/gen/1x8-minmax-neon-mlal-lane.c",
Marat Dukhan40bbafe2020-08-04 02:04:22 -07001625 "src/qs8-gemm/gen/1x16-minmax-neon-mlal-lane.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001626 "src/qs8-gemm/gen/2x8-minmax-neon-mlal-lane.c",
Marat Dukhan40bbafe2020-08-04 02:04:22 -07001627 "src/qs8-gemm/gen/2x16-minmax-neon-mlal-lane.c",
1628 "src/qs8-igemm/gen/1x8-minmax-neon-mlal-lane.c",
Marat Dukhan40bbafe2020-08-04 02:04:22 -07001629 "src/qs8-igemm/gen/1x16-minmax-neon-mlal-lane.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001630 "src/qs8-igemm/gen/2x8-minmax-neon-mlal-lane.c",
Marat Dukhan40bbafe2020-08-04 02:04:22 -07001631 "src/qs8-igemm/gen/2x16-minmax-neon-mlal-lane.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -07001632 "src/qs8-requantization/fp32-neon.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001633 "src/qs8-requantization/precise-neon.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -07001634 "src/qs8-requantization/q31-neon.c",
Marat Dukhanba7b2792020-09-02 14:26:45 -07001635 "src/qs8-vadd/gen/minmax-neon-ld64-x8.c",
1636 "src/qs8-vadd/gen/minmax-neon-ld64-x16.c",
1637 "src/qs8-vadd/gen/minmax-neon-ld64-x24.c",
1638 "src/qs8-vadd/gen/minmax-neon-ld64-x32.c",
1639 "src/qs8-vaddc/gen/minmax-neon-ld64-x8.c",
1640 "src/qs8-vaddc/gen/minmax-neon-ld64-x16.c",
1641 "src/qs8-vaddc/gen/minmax-neon-ld64-x24.c",
1642 "src/qs8-vaddc/gen/minmax-neon-ld64-x32.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -07001643 "src/qu8-avgpool/9p8x-minmax-neon-c8.c",
1644 "src/qu8-avgpool/9x-minmax-neon-c8.c",
1645 "src/qu8-dwconv/up8x9-minmax-neon.c",
1646 "src/qu8-gavgpool/7p7x-minmax-neon-c8.c",
1647 "src/qu8-gavgpool/7x-minmax-neon-c8.c",
1648 "src/qu8-gemm/4x8-minmax-neon.c",
1649 "src/qu8-gemm/8x8-minmax-neon.c",
1650 "src/qu8-igemm/4x8-minmax-neon.c",
1651 "src/qu8-igemm/8x8-minmax-neon.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07001652 "src/qu8-requantization/fp32-neon.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001653 "src/qu8-requantization/precise-neon.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07001654 "src/qu8-requantization/q31-neon.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -07001655 "src/qu8-vadd/minmax-neon.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001656 "src/u8-clamp/neon-x64.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001657 "src/u8-maxpool/9p8x-minmax-neon-c16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001658 "src/u8-rmax/neon.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001659 "src/x8-zip/x2-neon.c",
1660 "src/x8-zip/x3-neon.c",
1661 "src/x8-zip/x4-neon.c",
1662 "src/x8-zip/xm-neon.c",
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -07001663 "src/x32-fill/neon.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001664 "src/x32-packx/x4-neon-st4.c",
Marat Dukhan63523d42020-05-22 17:07:33 -07001665 "src/x32-pad/neon.c",
Marat Dukhan57dccd82020-04-14 00:53:10 -07001666 "src/x32-unpool/neon.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001667 "src/x32-zip/x2-neon.c",
1668 "src/x32-zip/x3-neon.c",
1669 "src/x32-zip/x4-neon.c",
1670 "src/x32-zip/xm-neon.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001671]
1672
1673NEONFMA_UKERNELS = [
Frank Barchard04336c12020-10-22 16:48:55 -07001674 "src/f32-dwconv/gen/up4x4-minmax-neonfma-acc2.c",
1675 "src/f32-dwconv/gen/up4x4-minmax-neonfma.c",
1676 "src/f32-dwconv/gen/up4x9-minmax-neonfma-acc2.c",
1677 "src/f32-dwconv/gen/up4x9-minmax-neonfma.c",
1678 "src/f32-dwconv/gen/up4x25-minmax-neonfma-acc2.c",
1679 "src/f32-dwconv/gen/up4x25-minmax-neonfma.c",
1680 "src/f32-dwconv/gen/up8x4-minmax-neonfma-acc2.c",
1681 "src/f32-dwconv/gen/up8x4-minmax-neonfma.c",
1682 "src/f32-dwconv/gen/up8x9-minmax-neonfma-acc2.c",
1683 "src/f32-dwconv/gen/up8x9-minmax-neonfma.c",
1684 "src/f32-dwconv/gen/up8x25-minmax-neonfma-acc2.c",
1685 "src/f32-dwconv/gen/up8x25-minmax-neonfma.c",
1686 "src/f32-gemm/gen-inc/1x8inc-minmax-neonfma-dup-ld64.c",
1687 "src/f32-gemm/gen-inc/1x8s4inc-minmax-neonfma.c",
1688 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-dup-ld64.c",
1689 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-dup-ld128.c",
1690 "src/f32-gemm/gen-inc/4x8s4inc-minmax-neonfma.c",
1691 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-dup-ld64.c",
1692 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-dup-ld128.c",
1693 "src/f32-gemm/gen-inc/6x8s4inc-minmax-neonfma.c",
1694 "src/f32-gemm/gen-inc/8x8s4inc-minmax-neonfma.c",
1695 "src/f32-gemm/gen/1x8-minmax-neonfma-dup-ld64.c",
1696 "src/f32-gemm/gen/1x8s4-minmax-neonfma.c",
1697 "src/f32-gemm/gen/4x8-minmax-neonfma-dup-ld64.c",
1698 "src/f32-gemm/gen/4x8-minmax-neonfma-dup-ld128.c",
1699 "src/f32-gemm/gen/4x8s4-minmax-neonfma.c",
1700 "src/f32-gemm/gen/6x8-minmax-neonfma-dup-ld64.c",
1701 "src/f32-gemm/gen/6x8-minmax-neonfma-dup-ld128.c",
1702 "src/f32-gemm/gen/6x8s4-minmax-neonfma.c",
1703 "src/f32-gemm/gen/8x8s4-minmax-neonfma.c",
Marat Dukhan660fd192020-03-10 04:55:30 -07001704 "src/f32-ibilinear/gen/neonfma-c4.c",
1705 "src/f32-ibilinear/gen/neonfma-c8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001706 "src/f32-igemm/gen/1x8-minmax-neonfma-dup-ld64.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001707 "src/f32-igemm/gen/1x8s4-minmax-neonfma.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001708 "src/f32-igemm/gen/4x8-minmax-neonfma-dup-ld64.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001709 "src/f32-igemm/gen/4x8-minmax-neonfma-dup-ld128.c",
1710 "src/f32-igemm/gen/4x8s4-minmax-neonfma.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001711 "src/f32-igemm/gen/6x8-minmax-neonfma-dup-ld64.c",
1712 "src/f32-igemm/gen/6x8-minmax-neonfma-dup-ld128.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001713 "src/f32-igemm/gen/6x8s4-minmax-neonfma.c",
1714 "src/f32-igemm/gen/8x8s4-minmax-neonfma.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001715 "src/f32-ppmm/gen/4x8-minmax-neonfma.c",
1716 "src/f32-ppmm/gen/8x8-minmax-neonfma.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001717 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x4.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001718 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x8-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001719 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x8.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001720 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12-acc2.c",
1721 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001722 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001723 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16-acc2.c",
1724 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001725 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001726 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20-acc2.c",
1727 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001728 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20.c",
1729 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x4.c",
1730 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x8-acc2.c",
1731 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x8.c",
1732 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12-acc2.c",
1733 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12-acc3.c",
1734 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12.c",
1735 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16-acc2.c",
1736 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16-acc4.c",
1737 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16.c",
1738 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20-acc2.c",
1739 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20-acc5.c",
1740 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -08001741 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x4.c",
1742 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x8.c",
1743 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x12.c",
1744 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x16.c",
1745 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x20.c",
1746 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x24.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001747 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x4.c",
1748 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x8.c",
1749 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x12.c",
1750 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x16.c",
1751 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x20.c",
1752 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x24.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -08001753 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x4.c",
1754 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x8.c",
1755 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x12.c",
1756 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x16.c",
1757 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x20.c",
1758 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x24.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -08001759 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x4.c",
1760 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x8.c",
1761 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x12.c",
1762 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x16.c",
1763 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x20.c",
1764 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x24.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001765 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x4.c",
1766 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x8.c",
1767 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x12.c",
1768 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x16.c",
1769 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x20.c",
1770 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x24.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -08001771 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x4.c",
1772 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x8.c",
1773 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x12.c",
1774 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x16.c",
1775 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x20.c",
1776 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x24.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001777 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x4.c",
1778 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x8.c",
1779 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x12.c",
1780 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x16.c",
1781 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x20.c",
1782 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x24.c",
1783 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x4.c",
1784 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x8.c",
1785 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x12.c",
1786 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x16.c",
1787 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x20.c",
1788 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x24.c",
1789 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x4.c",
1790 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x8.c",
1791 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x12.c",
1792 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x16.c",
1793 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x20.c",
1794 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x24.c",
Marat Dukhan2fa7a0c2020-12-06 19:09:02 -08001795 "src/f32-spmm/gen/4x1-minmax-neonfma-pipelined.c",
1796 "src/f32-spmm/gen/4x1-minmax-neonfma-x2.c",
1797 "src/f32-spmm/gen/4x1-minmax-neonfma.c",
1798 "src/f32-spmm/gen/8x1-minmax-neonfma-pipelined.c",
1799 "src/f32-spmm/gen/8x1-minmax-neonfma-x2.c",
1800 "src/f32-spmm/gen/8x1-minmax-neonfma.c",
1801 "src/f32-spmm/gen/12x1-minmax-neonfma.c",
1802 "src/f32-spmm/gen/16x1-minmax-neonfma-pipelined.c",
1803 "src/f32-spmm/gen/16x1-minmax-neonfma-x2.c",
1804 "src/f32-spmm/gen/16x1-minmax-neonfma.c",
1805 "src/f32-spmm/gen/32x1-minmax-neonfma-pipelined.c",
1806 "src/f32-spmm/gen/32x1-minmax-neonfma-x2.c",
1807 "src/f32-spmm/gen/32x1-minmax-neonfma.c",
Marat Dukhaned6baaf2020-12-01 15:07:08 -08001808 "src/f32-velu/gen/velu-neonfma-rr1-lut16-p3-x4.c",
1809 "src/f32-velu/gen/velu-neonfma-rr1-lut16-p3-x8.c",
1810 "src/f32-velu/gen/velu-neonfma-rr1-lut16-p3-x12.c",
1811 "src/f32-velu/gen/velu-neonfma-rr1-lut16-p3-x16.c",
1812 "src/f32-velu/gen/velu-neonfma-rr1-lut16-p3-x20.c",
1813 "src/f32-velu/gen/velu-neonfma-rr1-lut16-p3-x24.c",
1814 "src/f32-velu/gen/velu-neonfma-rr1-p6-x4.c",
1815 "src/f32-velu/gen/velu-neonfma-rr1-p6-x8.c",
1816 "src/f32-velu/gen/velu-neonfma-rr1-p6-x12.c",
1817 "src/f32-velu/gen/velu-neonfma-rr1-p6-x16.c",
1818 "src/f32-velu/gen/velu-neonfma-rr1-p6-x20.c",
1819 "src/f32-velu/gen/velu-neonfma-rr1-p6-x24.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001820 "src/f32-vmulcaddc/gen/c4-minmax-neonfma-2x.c",
1821 "src/f32-vmulcaddc/gen/c8-minmax-neonfma-2x.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07001822 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x4.c",
1823 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x8.c",
1824 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x12.c",
1825 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x16.c",
1826 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x20.c",
1827 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x24.c",
1828 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x28.c",
1829 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x32.c",
1830 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x36.c",
1831 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x40.c",
1832 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x4.c",
1833 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x8.c",
1834 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x12.c",
1835 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x16.c",
1836 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x20.c",
1837 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x24.c",
1838 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x28.c",
1839 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x32.c",
1840 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x36.c",
1841 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x40.c",
Marat Dukhanb7633f22020-11-20 16:34:56 -08001842 "src/math/exp-neonfma-rr2-lut64-p2.c",
1843 "src/math/exp-neonfma-rr2-p5.c",
Frank Barcharde7223ee2020-12-04 19:04:01 -08001844 "src/math/expm1minus-neonfma-rr1-lut16-p3.c",
1845 "src/math/expm1minus-neonfma-rr1-p6.c",
Marat Dukhan9dd119a2020-11-20 18:20:04 -08001846 "src/math/expminus-neonfma-rr2-lut64-p2.c",
1847 "src/math/expminus-neonfma-rr2-lut2048-p1.c",
1848 "src/math/expminus-neonfma-rr2-p5.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001849 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr1recps1fma.c",
1850 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr2fma.c",
1851 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr2recps.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001852 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr1recps1fma.c",
1853 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr2fma.c",
1854 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr2recps.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001855 "src/math/sigmoid-neonfma-rr1-p5-nr1recps1fma.c",
1856 "src/math/sigmoid-neonfma-rr1-p5-nr2fma.c",
1857 "src/math/sigmoid-neonfma-rr1-p5-nr2recps.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001858 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr1recps1fma.c",
1859 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr2fma.c",
1860 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr2recps.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001861 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr1recps1fma.c",
1862 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr2fma.c",
1863 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr2recps.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001864 "src/math/sigmoid-neonfma-rr2-p5-nr1recps1fma.c",
1865 "src/math/sigmoid-neonfma-rr2-p5-nr2fma.c",
1866 "src/math/sigmoid-neonfma-rr2-p5-nr2recps.c",
Marat Dukhan84000762020-06-29 18:38:43 -07001867 "src/math/sqrt-neonfma-nr1fma.c",
Marat Dukhan84000762020-06-29 18:38:43 -07001868 "src/math/sqrt-neonfma-nr1rsqrts1fma1adj.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001869 "src/math/sqrt-neonfma-nr2fma.c",
1870 "src/math/sqrt-neonfma-nr2fma1adj.c",
1871 "src/math/sqrt-neonfma-nr3fma.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001872]
1873
1874AARCH64_NEONFMA_UKERNELS = [
Marat Dukhan56b10cd2020-05-18 09:35:49 -07001875 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neonfma-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001876 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neonfma-2x2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001877 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neonfma-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001878 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neonfma-2x2.c",
Marat Dukhan56b10cd2020-05-18 09:35:49 -07001879 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neonfma-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001880 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neonfma-2x2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001881 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neonfma-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001882 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neonfma-2x2.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -07001883 "src/f32-conv-hwc2chw/3x3s2p1c3x4-neonfma-2x2.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001884 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-1x4-acc2.c",
1885 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-1x4-acc3.c",
1886 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-1x4-acc4.c",
Marat Dukhan1268a242020-10-24 00:36:32 -07001887 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-1x4.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001888 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-2x4-acc2.c",
Marat Dukhan1268a242020-10-24 00:36:32 -07001889 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-2x4.c",
1890 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-3x4.c",
1891 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-4x4.c",
1892 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-5x4.c",
1893 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-6x4.c",
Marat Dukhan82f0c322020-10-25 19:17:35 -07001894 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neonfma-1x4-acc2.c",
1895 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neonfma-1x4-acc3.c",
1896 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neonfma-1x4-acc4.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001897 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neonfma-1x4.c",
Marat Dukhan82f0c322020-10-25 19:17:35 -07001898 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neonfma-2x4-acc2.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001899 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neonfma-2x4.c",
1900 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neonfma-3x4.c",
1901 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neonfma-4x4.c",
Marat Dukhan149f0ea2020-10-26 12:50:33 -07001902 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-1x4-acc2.c",
1903 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-1x4-acc3.c",
1904 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-1x4-acc4.c",
1905 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-1x4-acc5.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001906 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-1x4.c",
Marat Dukhan149f0ea2020-10-26 12:50:33 -07001907 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-2x4-acc2.c",
1908 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-2x4-acc3.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001909 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-2x4.c",
Marat Dukhan149f0ea2020-10-26 12:50:33 -07001910 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-3x4-acc2.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001911 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-3x4.c",
Marat Dukhan149f0ea2020-10-26 12:50:33 -07001912 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-4x4-acc2.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001913 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-4x4.c",
1914 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-5x4.c",
Marat Dukhan30d4b252020-10-29 16:33:22 -07001915 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neonfma-1x4-acc2.c",
1916 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neonfma-1x4-acc3.c",
1917 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neonfma-1x4-acc4.c",
1918 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neonfma-1x4-acc5.c",
1919 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neonfma-1x4.c",
1920 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neonfma-2x4-acc2.c",
1921 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neonfma-2x4-acc3.c",
1922 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neonfma-2x4.c",
Marat Dukhan30d4b252020-10-29 16:33:22 -07001923 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neonfma-3x4-acc2.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07001924 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neonfma-3x4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001925 "src/f32-gemm/gen-inc/1x8inc-minmax-neonfma-lane-ld64.c",
1926 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-lane-ld64.c",
1927 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-lane-ld128.c",
1928 "src/f32-gemm/gen-inc/5x8inc-minmax-neonfma-lane-ld64.c",
1929 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-lane-ld64.c",
1930 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-lane-ld128.c",
1931 "src/f32-gemm/gen/1x8-minmax-neonfma-lane-ld64.c",
1932 "src/f32-gemm/gen/4x2-minmax-neonfma-lane-ld64.c",
1933 "src/f32-gemm/gen/4x8-minmax-neonfma-lane-ld64.c",
1934 "src/f32-gemm/gen/4x8-minmax-neonfma-lane-ld128.c",
1935 "src/f32-gemm/gen/5x8-minmax-neonfma-lane-ld64.c",
1936 "src/f32-gemm/gen/6x8-minmax-neonfma-lane-ld64.c",
1937 "src/f32-gemm/gen/6x8-minmax-neonfma-lane-ld128.c",
1938 "src/f32-igemm/gen/1x8-minmax-neonfma-lane-ld64.c",
1939 "src/f32-igemm/gen/4x2-minmax-neonfma-lane-ld64.c",
1940 "src/f32-igemm/gen/4x4-minmax-neonfma-lane-ld64.c",
1941 "src/f32-igemm/gen/4x8-minmax-neonfma-lane-ld64.c",
1942 "src/f32-igemm/gen/4x8-minmax-neonfma-lane-ld128.c",
1943 "src/f32-igemm/gen/6x8-minmax-neonfma-lane-ld64.c",
1944 "src/f32-igemm/gen/6x8-minmax-neonfma-lane-ld128.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -08001945 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x4.c",
1946 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x8.c",
1947 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x12.c",
1948 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x16.c",
1949 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x20.c",
1950 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x24.c",
1951 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x4.c",
1952 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x8.c",
1953 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x12.c",
1954 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x16.c",
1955 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x20.c",
1956 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x24.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001957 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x4.c",
1958 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x8.c",
1959 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x12.c",
1960 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x16.c",
1961 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x20.c",
1962 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x24.c",
Marat Dukhan355ab432020-04-09 19:01:52 -07001963 "src/f32-spmm/gen/4x2-minmax-neonfma.c",
1964 "src/f32-spmm/gen/4x4-minmax-neonfma.c",
Marat Dukhan355ab432020-04-09 19:01:52 -07001965 "src/f32-spmm/gen/8x2-minmax-neonfma.c",
1966 "src/f32-spmm/gen/8x4-minmax-neonfma.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001967 "src/f32-spmm/gen/12x2-minmax-neonfma.c",
1968 "src/f32-spmm/gen/12x4-minmax-neonfma.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001969 "src/f32-spmm/gen/16x2-minmax-neonfma.c",
1970 "src/f32-spmm/gen/16x4-minmax-neonfma.c",
Frank Barchard846c0c62020-10-26 15:01:39 -07001971 "src/f32-spmm/gen/32x2-minmax-neonfma.c",
1972 "src/f32-spmm/gen/32x4-minmax-neonfma.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001973 "src/f32-vbinary/gen/vdiv-minmax-neon-x4.c",
1974 "src/f32-vbinary/gen/vdiv-minmax-neon-x8.c",
1975 "src/f32-vbinary/gen/vdivc-minmax-neon-x4.c",
1976 "src/f32-vbinary/gen/vdivc-minmax-neon-x8.c",
1977 "src/f32-vbinary/gen/vrdivc-minmax-neon-x4.c",
1978 "src/f32-vbinary/gen/vrdivc-minmax-neon-x8.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07001979 "src/f32-vsqrt/gen/neon-sqrt-x4.c",
1980 "src/f32-vsqrt/gen/neon-sqrt-x8.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001981 "src/math/sigmoid-neonfma-rr1-lut64-p2-div.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001982 "src/math/sigmoid-neonfma-rr1-lut2048-p1-div.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001983 "src/math/sigmoid-neonfma-rr1-p5-div.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001984 "src/math/sigmoid-neonfma-rr2-lut64-p2-div.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001985 "src/math/sigmoid-neonfma-rr2-lut2048-p1-div.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001986 "src/math/sigmoid-neonfma-rr2-p5-div.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001987]
1988
Marat Dukhan8853b822020-05-07 12:19:01 -07001989NEONV8_UKERNELS = [
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07001990 "src/f32-vrnd/gen/vrndd-neonv8-x4.c",
1991 "src/f32-vrnd/gen/vrndd-neonv8-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001992 "src/f32-vrnd/gen/vrndne-neonv8-x4.c",
1993 "src/f32-vrnd/gen/vrndne-neonv8-x8.c",
1994 "src/f32-vrnd/gen/vrndu-neonv8-x4.c",
1995 "src/f32-vrnd/gen/vrndu-neonv8-x8.c",
1996 "src/f32-vrnd/gen/vrndz-neonv8-x4.c",
1997 "src/f32-vrnd/gen/vrndz-neonv8-x8.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001998 "src/math/roundd-neonv8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001999 "src/math/roundne-neonv8.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07002000 "src/math/roundu-neonv8.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07002001 "src/math/roundz-neonv8.c",
Marat Dukhan8853b822020-05-07 12:19:01 -07002002]
2003
Marat Dukhan08c4a432019-10-03 09:29:21 -07002004AARCH64_NEONFP16ARITH_UKERNELS = [
Frank Barchardb1966592020-05-12 13:47:06 -07002005 "src/f16-clamp/gen/neonfp16arith-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002006 "src/f16-clamp/gen/neonfp16arith-x16.c",
Frank Barchard5a599a62020-06-04 20:12:44 -07002007 "src/f16-dwconv/gen/up8x4-minmax-neonfp16arith-acc2.c",
2008 "src/f16-dwconv/gen/up8x4-minmax-neonfp16arith.c",
2009 "src/f16-dwconv/gen/up8x9-minmax-neonfp16arith-acc2.c",
2010 "src/f16-dwconv/gen/up8x9-minmax-neonfp16arith.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002011 "src/f16-dwconv/gen/up8x25-minmax-neonfp16arith-acc2.c",
2012 "src/f16-dwconv/gen/up8x25-minmax-neonfp16arith.c",
2013 "src/f16-dwconv/gen/up16x4-minmax-neonfp16arith-acc2.c",
2014 "src/f16-dwconv/gen/up16x4-minmax-neonfp16arith.c",
2015 "src/f16-dwconv/gen/up16x9-minmax-neonfp16arith-acc2.c",
2016 "src/f16-dwconv/gen/up16x9-minmax-neonfp16arith.c",
2017 "src/f16-dwconv/gen/up16x25-minmax-neonfp16arith-acc2.c",
2018 "src/f16-dwconv/gen/up16x25-minmax-neonfp16arith.c",
Frank Barchard0bb49a72020-06-04 11:35:11 -07002019 "src/f16-gavgpool/7p7x-minmax-neonfp16arith-c8.c",
2020 "src/f16-gavgpool/7x-minmax-neonfp16arith-c8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002021 "src/f16-gemm/gen-inc/1x8inc-minmax-neonfp16arith-ld64.c",
2022 "src/f16-gemm/gen-inc/1x16inc-minmax-neonfp16arith-ld64.c",
2023 "src/f16-gemm/gen-inc/4x8inc-minmax-neonfp16arith-ld64.c",
2024 "src/f16-gemm/gen-inc/4x16inc-minmax-neonfp16arith-ld64.c",
2025 "src/f16-gemm/gen-inc/6x8inc-minmax-neonfp16arith-ld64.c",
2026 "src/f16-gemm/gen-inc/6x16inc-minmax-neonfp16arith-ld64.c",
2027 "src/f16-gemm/gen-inc/8x8inc-minmax-neonfp16arith-ld64.c",
2028 "src/f16-gemm/gen-inc/8x16inc-minmax-neonfp16arith-ld64.c",
2029 "src/f16-gemm/gen/1x8-minmax-neonfp16arith-ld64.c",
2030 "src/f16-gemm/gen/1x16-minmax-neonfp16arith-ld64.c",
2031 "src/f16-gemm/gen/4x8-minmax-neonfp16arith-ld64.c",
2032 "src/f16-gemm/gen/4x16-minmax-neonfp16arith-ld64.c",
2033 "src/f16-gemm/gen/6x8-minmax-neonfp16arith-ld64.c",
2034 "src/f16-gemm/gen/6x16-minmax-neonfp16arith-ld64.c",
2035 "src/f16-gemm/gen/8x8-minmax-neonfp16arith-ld64.c",
2036 "src/f16-gemm/gen/8x16-minmax-neonfp16arith-ld64.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07002037 "src/f16-hswish/gen/hswish-neonfp16arith-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002038 "src/f16-hswish/gen/hswish-neonfp16arith-x16.c",
2039 "src/f16-igemm/gen/1x8-minmax-neonfp16arith-ld64.c",
2040 "src/f16-igemm/gen/1x16-minmax-neonfp16arith-ld64.c",
2041 "src/f16-igemm/gen/4x8-minmax-neonfp16arith-ld64.c",
2042 "src/f16-igemm/gen/4x16-minmax-neonfp16arith-ld64.c",
2043 "src/f16-igemm/gen/6x8-minmax-neonfp16arith-ld64.c",
2044 "src/f16-igemm/gen/6x16-minmax-neonfp16arith-ld64.c",
2045 "src/f16-igemm/gen/8x8-minmax-neonfp16arith-ld64.c",
2046 "src/f16-igemm/gen/8x16-minmax-neonfp16arith-ld64.c",
Frank Barchardb1966592020-05-12 13:47:06 -07002047 "src/f16-prelu/gen/neonfp16arith-2x8.c",
Frank Barcharda5316982020-07-23 13:19:28 -07002048 "src/f16-prelu/gen/neonfp16arith-2x16.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07002049 "src/f16-relu/gen/neonfp16arith-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002050 "src/f16-relu/gen/neonfp16arith-x16.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07002051 "src/f16-spmm/gen/8x1-minmax-neonfp16arith-x2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002052 "src/f16-spmm/gen/8x1-minmax-neonfp16arith.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07002053 "src/f16-spmm/gen/16x1-minmax-neonfp16arith-x2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002054 "src/f16-spmm/gen/16x1-minmax-neonfp16arith.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07002055 "src/f16-spmm/gen/24x1-minmax-neonfp16arith-x2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002056 "src/f16-spmm/gen/24x1-minmax-neonfp16arith.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07002057 "src/f16-spmm/gen/32x1-minmax-neonfp16arith-x2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002058 "src/f16-spmm/gen/32x1-minmax-neonfp16arith.c",
2059 "src/f16-vbinary/gen/vadd-minmax-neonfp16arith-x8.c",
2060 "src/f16-vbinary/gen/vadd-minmax-neonfp16arith-x16.c",
2061 "src/f16-vbinary/gen/vaddc-minmax-neonfp16arith-x8.c",
2062 "src/f16-vbinary/gen/vaddc-minmax-neonfp16arith-x16.c",
2063 "src/f16-vbinary/gen/vdiv-minmax-neonfp16arith-x8.c",
2064 "src/f16-vbinary/gen/vdiv-minmax-neonfp16arith-x16.c",
2065 "src/f16-vbinary/gen/vdivc-minmax-neonfp16arith-x8.c",
2066 "src/f16-vbinary/gen/vdivc-minmax-neonfp16arith-x16.c",
2067 "src/f16-vbinary/gen/vmax-neonfp16arith-x8.c",
2068 "src/f16-vbinary/gen/vmax-neonfp16arith-x16.c",
2069 "src/f16-vbinary/gen/vmaxc-neonfp16arith-x8.c",
2070 "src/f16-vbinary/gen/vmaxc-neonfp16arith-x16.c",
2071 "src/f16-vbinary/gen/vmin-neonfp16arith-x8.c",
2072 "src/f16-vbinary/gen/vmin-neonfp16arith-x16.c",
2073 "src/f16-vbinary/gen/vminc-neonfp16arith-x8.c",
2074 "src/f16-vbinary/gen/vminc-neonfp16arith-x16.c",
2075 "src/f16-vbinary/gen/vmul-minmax-neonfp16arith-x8.c",
2076 "src/f16-vbinary/gen/vmul-minmax-neonfp16arith-x16.c",
2077 "src/f16-vbinary/gen/vmulc-minmax-neonfp16arith-x8.c",
2078 "src/f16-vbinary/gen/vmulc-minmax-neonfp16arith-x16.c",
2079 "src/f16-vbinary/gen/vrdivc-minmax-neonfp16arith-x8.c",
2080 "src/f16-vbinary/gen/vrdivc-minmax-neonfp16arith-x16.c",
2081 "src/f16-vbinary/gen/vrsubc-minmax-neonfp16arith-x8.c",
2082 "src/f16-vbinary/gen/vrsubc-minmax-neonfp16arith-x16.c",
2083 "src/f16-vbinary/gen/vsub-minmax-neonfp16arith-x8.c",
2084 "src/f16-vbinary/gen/vsub-minmax-neonfp16arith-x16.c",
2085 "src/f16-vbinary/gen/vsubc-minmax-neonfp16arith-x8.c",
2086 "src/f16-vbinary/gen/vsubc-minmax-neonfp16arith-x16.c",
2087 "src/f16-vmulcaddc/gen/c8-minmax-neonfp16arith-2x.c",
2088 "src/f16-vmulcaddc/gen/c16-minmax-neonfp16arith-2x.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002089]
2090
Benoit Jacoba9644732020-08-13 12:48:55 -07002091NEONDOT_UKERNELS = [
Marat Dukhand4c83032020-09-09 23:20:12 -07002092 "src/qs8-gemm/gen/1x8c4-minmax-neondot.c",
Marat Dukhand4c83032020-09-09 23:20:12 -07002093 "src/qs8-gemm/gen/1x16c4-minmax-neondot.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002094 "src/qs8-gemm/gen/4x8c4-minmax-neondot.c",
Benoit Jacoba9644732020-08-13 12:48:55 -07002095 "src/qs8-gemm/gen/4x16c4-minmax-neondot.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002096 "src/qs8-gemm/gen/6x8c4-minmax-neondot.c",
Frank Barchard2fa17452020-09-22 09:51:46 -07002097 "src/qs8-gemm/gen/6x16c4-minmax-neondot.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002098 "src/qs8-gemm/gen/8x8c4-minmax-neondot.c",
Frank Barcharda48848f2020-09-24 00:59:24 -07002099 "src/qs8-gemm/gen/8x16c4-minmax-neondot.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002100 "src/qs8-gemm/gen/12x8c4-minmax-neondot.c",
Marat Dukhand4c83032020-09-09 23:20:12 -07002101 "src/qs8-igemm/gen/1x8c4-minmax-neondot.c",
Marat Dukhand4c83032020-09-09 23:20:12 -07002102 "src/qs8-igemm/gen/1x16c4-minmax-neondot.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002103 "src/qs8-igemm/gen/4x8c4-minmax-neondot.c",
Benoit Jacoba05487f2020-08-14 13:37:21 -07002104 "src/qs8-igemm/gen/4x16c4-minmax-neondot.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002105 "src/qs8-igemm/gen/6x8c4-minmax-neondot.c",
Frank Barchard2fa17452020-09-22 09:51:46 -07002106 "src/qs8-igemm/gen/6x16c4-minmax-neondot.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002107 "src/qs8-igemm/gen/8x8c4-minmax-neondot.c",
Frank Barcharda48848f2020-09-24 00:59:24 -07002108 "src/qs8-igemm/gen/8x16c4-minmax-neondot.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002109 "src/qs8-igemm/gen/12x8c4-minmax-neondot.c",
Benoit Jacoba9644732020-08-13 12:48:55 -07002110]
2111
Marat Dukhan08c4a432019-10-03 09:29:21 -07002112SSE_UKERNELS = [
Marat Dukhan99936602020-04-11 16:47:01 -07002113 "src/f32-avgpool/9p8x-minmax-sse-c4.c",
2114 "src/f32-avgpool/9x-minmax-sse-c4.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -07002115 "src/f32-clamp/gen/sse-x4.c",
2116 "src/f32-clamp/gen/sse-x8.c",
Erich Elsenb1233402020-06-08 15:53:15 -07002117 "src/f32-conv-hwc2chw/3x3s2p1c3x4-sse-1x1.c",
2118 "src/f32-conv-hwc2chw/3x3s2p1c3x4-sse-2x2.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002119 "src/f32-dwconv/gen/up4x4-minmax-sse-acc2.c",
2120 "src/f32-dwconv/gen/up4x4-minmax-sse.c",
2121 "src/f32-dwconv/gen/up4x9-minmax-sse-acc2.c",
2122 "src/f32-dwconv/gen/up4x9-minmax-sse.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002123 "src/f32-dwconv/gen/up4x25-minmax-sse-acc2.c",
2124 "src/f32-dwconv/gen/up4x25-minmax-sse.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002125 "src/f32-dwconv/gen/up8x4-minmax-sse-acc2.c",
2126 "src/f32-dwconv/gen/up8x4-minmax-sse.c",
2127 "src/f32-dwconv/gen/up8x9-minmax-sse-acc2.c",
2128 "src/f32-dwconv/gen/up8x9-minmax-sse.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002129 "src/f32-dwconv/gen/up8x25-minmax-sse-acc2.c",
2130 "src/f32-dwconv/gen/up8x25-minmax-sse.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07002131 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-1x4-acc2.c",
2132 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-1x4-acc3.c",
2133 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-1x4-acc4.c",
Marat Dukhan470078a2020-10-23 22:36:52 -07002134 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-1x4.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07002135 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-2x4-acc2.c",
Marat Dukhan470078a2020-10-23 22:36:52 -07002136 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-2x4.c",
2137 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-3x4.c",
2138 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-4x4.c",
2139 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-5x4.c",
2140 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-6x4.c",
Marat Dukhan0ff97182020-10-25 19:14:03 -07002141 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-sse-1x4-acc2.c",
2142 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-sse-1x4-acc3.c",
2143 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-sse-1x4-acc4.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07002144 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-sse-1x4.c",
Marat Dukhan0ff97182020-10-25 19:14:03 -07002145 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-sse-2x4-acc2.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07002146 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-sse-2x4.c",
2147 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-sse-3x4.c",
2148 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-sse-4x4.c",
Marat Dukhand0503892020-10-30 08:22:04 -07002149 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-1x4-acc2.c",
2150 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-1x4-acc3.c",
2151 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-1x4-acc4.c",
2152 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-1x4-acc5.c",
2153 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-1x4.c",
2154 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-2x4-acc2.c",
2155 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-2x4-acc3.c",
2156 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-2x4.c",
2157 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-3x4-acc2.c",
2158 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-3x4.c",
2159 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-4x4-acc2.c",
2160 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-4x4.c",
2161 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-5x4.c",
Marat Dukhanccca2142020-10-30 17:32:45 -07002162 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-sse-1x4-acc2.c",
2163 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-sse-1x4-acc3.c",
2164 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-sse-1x4-acc4.c",
2165 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-sse-1x4-acc5.c",
2166 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-sse-1x4.c",
2167 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-sse-2x4-acc2.c",
2168 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-sse-2x4-acc3.c",
2169 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-sse-2x4.c",
Marat Dukhanccca2142020-10-30 17:32:45 -07002170 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-sse-3x4-acc2.c",
Frank Barchard8ef44cd2020-11-03 12:30:23 -08002171 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-sse-3x4.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -07002172 "src/f32-gavgpool-cw/sse-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -07002173 "src/f32-gavgpool/7p7x-minmax-sse-c4.c",
2174 "src/f32-gavgpool/7x-minmax-sse-c4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002175 "src/f32-gemm/gen-inc/1x8inc-minmax-sse-dup.c",
2176 "src/f32-gemm/gen-inc/1x8inc-minmax-sse-load1.c",
2177 "src/f32-gemm/gen-inc/1x8s4inc-minmax-sse.c",
2178 "src/f32-gemm/gen-inc/4x8inc-minmax-sse-dup.c",
2179 "src/f32-gemm/gen-inc/4x8inc-minmax-sse-load1.c",
2180 "src/f32-gemm/gen-inc/4x8s4inc-minmax-sse.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002181 "src/f32-gemm/gen/1x8-minmax-sse-dup.c",
2182 "src/f32-gemm/gen/1x8-minmax-sse-load1.c",
2183 "src/f32-gemm/gen/1x8s4-minmax-sse.c",
2184 "src/f32-gemm/gen/4x2c4-minmax-sse.c",
2185 "src/f32-gemm/gen/4x8-minmax-sse-dup.c",
2186 "src/f32-gemm/gen/4x8-minmax-sse-load1.c",
2187 "src/f32-gemm/gen/4x8s4-minmax-sse.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07002188 "src/f32-hswish/gen/hswish-sse-x4.c",
2189 "src/f32-hswish/gen/hswish-sse-x8.c",
Marat Dukhan660fd192020-03-10 04:55:30 -07002190 "src/f32-ibilinear/gen/sse-c4.c",
2191 "src/f32-ibilinear/gen/sse-c8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002192 "src/f32-igemm/gen/1x8-minmax-sse-dup.c",
2193 "src/f32-igemm/gen/1x8-minmax-sse-load1.c",
2194 "src/f32-igemm/gen/1x8s4-minmax-sse.c",
2195 "src/f32-igemm/gen/4x2c4-minmax-sse.c",
2196 "src/f32-igemm/gen/4x8-minmax-sse-dup.c",
2197 "src/f32-igemm/gen/4x8-minmax-sse-load1.c",
2198 "src/f32-igemm/gen/4x8s4-minmax-sse.c",
Marat Dukhan99936602020-04-11 16:47:01 -07002199 "src/f32-maxpool/9p8x-minmax-sse-c4.c",
2200 "src/f32-pavgpool/9p8x-minmax-sse-c4.c",
2201 "src/f32-pavgpool/9x-minmax-sse-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002202 "src/f32-ppmm/gen/4x8-minmax-sse.c",
Marat Dukhan39b5e942020-06-24 15:03:48 -07002203 "src/f32-prelu/gen/sse-2x4.c",
2204 "src/f32-prelu/gen/sse-2x8.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07002205 "src/f32-relu/gen/sse-x4.c",
2206 "src/f32-relu/gen/sse-x8.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002207 "src/f32-rmax/sse.c",
Marat Dukhan355ab432020-04-09 19:01:52 -07002208 "src/f32-spmm/gen/4x1-minmax-sse.c",
2209 "src/f32-spmm/gen/8x1-minmax-sse.c",
Erich Elsen6e80fdc2020-06-09 15:35:37 -07002210 "src/f32-spmm/gen/16x1-minmax-sse.c",
Frank Barchard846c0c62020-10-26 15:01:39 -07002211 "src/f32-spmm/gen/32x1-minmax-sse.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002212 "src/f32-vbinary/gen/vadd-minmax-sse-x4.c",
2213 "src/f32-vbinary/gen/vadd-minmax-sse-x8.c",
2214 "src/f32-vbinary/gen/vaddc-minmax-sse-x4.c",
2215 "src/f32-vbinary/gen/vaddc-minmax-sse-x8.c",
2216 "src/f32-vbinary/gen/vdiv-minmax-sse-x4.c",
2217 "src/f32-vbinary/gen/vdiv-minmax-sse-x8.c",
2218 "src/f32-vbinary/gen/vdivc-minmax-sse-x4.c",
2219 "src/f32-vbinary/gen/vdivc-minmax-sse-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -08002220 "src/f32-vbinary/gen/vmax-sse-x4.c",
2221 "src/f32-vbinary/gen/vmax-sse-x8.c",
2222 "src/f32-vbinary/gen/vmaxc-sse-x4.c",
2223 "src/f32-vbinary/gen/vmaxc-sse-x8.c",
2224 "src/f32-vbinary/gen/vmin-sse-x4.c",
2225 "src/f32-vbinary/gen/vmin-sse-x8.c",
2226 "src/f32-vbinary/gen/vminc-sse-x4.c",
2227 "src/f32-vbinary/gen/vminc-sse-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002228 "src/f32-vbinary/gen/vmul-minmax-sse-x4.c",
2229 "src/f32-vbinary/gen/vmul-minmax-sse-x8.c",
2230 "src/f32-vbinary/gen/vmulc-minmax-sse-x4.c",
2231 "src/f32-vbinary/gen/vmulc-minmax-sse-x8.c",
2232 "src/f32-vbinary/gen/vrdivc-minmax-sse-x4.c",
2233 "src/f32-vbinary/gen/vrdivc-minmax-sse-x8.c",
2234 "src/f32-vbinary/gen/vrsubc-minmax-sse-x4.c",
2235 "src/f32-vbinary/gen/vrsubc-minmax-sse-x8.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07002236 "src/f32-vbinary/gen/vsqrdiff-sse-x4.c",
2237 "src/f32-vbinary/gen/vsqrdiff-sse-x8.c",
2238 "src/f32-vbinary/gen/vsqrdiffc-sse-x4.c",
2239 "src/f32-vbinary/gen/vsqrdiffc-sse-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002240 "src/f32-vbinary/gen/vsub-minmax-sse-x4.c",
2241 "src/f32-vbinary/gen/vsub-minmax-sse-x8.c",
2242 "src/f32-vbinary/gen/vsubc-minmax-sse-x4.c",
2243 "src/f32-vbinary/gen/vsubc-minmax-sse-x8.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07002244 "src/f32-vlrelu/gen/vlrelu-sse-x4.c",
2245 "src/f32-vlrelu/gen/vlrelu-sse-x8.c",
Marat Dukhan99936602020-04-11 16:47:01 -07002246 "src/f32-vmulcaddc/gen/c4-minmax-sse-2x.c",
2247 "src/f32-vmulcaddc/gen/c8-minmax-sse-2x.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07002248 "src/f32-vsqrt/gen/sse-sqrt-x4.c",
2249 "src/f32-vsqrt/gen/sse-sqrt-x8.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07002250 "src/f32-vunary/gen/vabs-sse-x4.c",
2251 "src/f32-vunary/gen/vabs-sse-x8.c",
2252 "src/f32-vunary/gen/vneg-sse-x4.c",
2253 "src/f32-vunary/gen/vneg-sse-x8.c",
2254 "src/f32-vunary/gen/vsqr-sse-x4.c",
2255 "src/f32-vunary/gen/vsqr-sse-x8.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07002256 "src/math/roundd-sse-addsub.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002257 "src/math/roundne-sse-addsub.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07002258 "src/math/roundu-sse-addsub.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07002259 "src/math/roundz-sse-addsub.c",
Marat Dukhan84000762020-06-29 18:38:43 -07002260 "src/math/sqrt-sse-hh1mac.c",
2261 "src/math/sqrt-sse-nr1mac.c",
2262 "src/math/sqrt-sse-nr2mac.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002263 "src/x32-fill/sse.c",
2264 "src/x32-packx/x4-sse.c",
2265 "src/x32-pad/sse.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002266]
2267
2268SSE2_UKERNELS = [
Marat Dukhan329da642019-11-19 21:44:39 -08002269 "src/f32-argmaxpool/4x-sse2-c4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002270 "src/f32-argmaxpool/9p8x-sse2-c4.c",
Marat Dukhan329da642019-11-19 21:44:39 -08002271 "src/f32-argmaxpool/9x-sse2-c4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08002272 "src/f32-prelu/gen/sse2-2x4.c",
2273 "src/f32-prelu/gen/sse2-2x8.c",
Marat Dukhanb39689d2020-01-24 13:32:20 -08002274 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x4.c",
Marat Dukhanb39689d2020-01-24 13:32:20 -08002275 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x8-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002276 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x8.c",
Marat Dukhanb39689d2020-01-24 13:32:20 -08002277 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12-acc2.c",
2278 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002279 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12.c",
Marat Dukhanb39689d2020-01-24 13:32:20 -08002280 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16-acc2.c",
2281 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002282 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16.c",
Marat Dukhanb39689d2020-01-24 13:32:20 -08002283 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20-acc2.c",
2284 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002285 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20.c",
Marat Dukhand243c1a2020-09-17 18:48:10 -07002286 "src/f32-sigmoid/gen/sse2-lut64-p2-div-x4.c",
2287 "src/f32-sigmoid/gen/sse2-lut64-p2-div-x8.c",
2288 "src/f32-sigmoid/gen/sse2-lut64-p2-div-x12.c",
2289 "src/f32-sigmoid/gen/sse2-lut64-p2-div-x16.c",
2290 "src/f32-sigmoid/gen/sse2-lut64-p2-div-x20.c",
2291 "src/f32-sigmoid/gen/sse2-lut64-p2-div-x24.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002292 "src/f32-sigmoid/gen/sse2-p5-div-x4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08002293 "src/f32-sigmoid/gen/sse2-p5-div-x8.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002294 "src/f32-sigmoid/gen/sse2-p5-div-x12.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08002295 "src/f32-sigmoid/gen/sse2-p5-div-x16.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002296 "src/f32-sigmoid/gen/sse2-p5-div-x20.c",
2297 "src/f32-sigmoid/gen/sse2-p5-div-x24.c",
Marat Dukhaned6baaf2020-12-01 15:07:08 -08002298 "src/f32-velu/gen/velu-sse2-rr2-lut16-p3-x4.c",
2299 "src/f32-velu/gen/velu-sse2-rr2-lut16-p3-x8.c",
2300 "src/f32-velu/gen/velu-sse2-rr2-lut16-p3-x12.c",
2301 "src/f32-velu/gen/velu-sse2-rr2-lut16-p3-x16.c",
2302 "src/f32-velu/gen/velu-sse2-rr2-lut16-p3-x20.c",
2303 "src/f32-velu/gen/velu-sse2-rr2-lut16-p3-x24.c",
2304 "src/f32-velu/gen/velu-sse2-rr2-p6-x4.c",
2305 "src/f32-velu/gen/velu-sse2-rr2-p6-x8.c",
2306 "src/f32-velu/gen/velu-sse2-rr2-p6-x12.c",
2307 "src/f32-velu/gen/velu-sse2-rr2-p6-x16.c",
2308 "src/f32-velu/gen/velu-sse2-rr2-p6-x20.c",
2309 "src/f32-velu/gen/velu-sse2-rr2-p6-x24.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07002310 "src/f32-vlrelu/gen/vlrelu-sse2-x4.c",
2311 "src/f32-vlrelu/gen/vlrelu-sse2-x8.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07002312 "src/f32-vrnd/gen/vrndd-sse2-x4.c",
2313 "src/f32-vrnd/gen/vrndd-sse2-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002314 "src/f32-vrnd/gen/vrndne-sse2-x4.c",
2315 "src/f32-vrnd/gen/vrndne-sse2-x8.c",
2316 "src/f32-vrnd/gen/vrndu-sse2-x4.c",
2317 "src/f32-vrnd/gen/vrndu-sse2-x8.c",
2318 "src/f32-vrnd/gen/vrndz-sse2-x4.c",
2319 "src/f32-vrnd/gen/vrndz-sse2-x8.c",
Marat Dukhanb7633f22020-11-20 16:34:56 -08002320 "src/math/exp-sse2-rr2-lut64-p2.c",
2321 "src/math/exp-sse2-rr2-p5.c",
Marat Dukhande390d42020-11-29 19:32:18 -08002322 "src/math/expm1minus-sse2-rr2-lut16-p3.c",
Marat Dukhana438aca2020-11-20 15:45:01 -08002323 "src/math/expm1minus-sse2-rr2-p6.c",
Frank Barchard3b800452020-11-22 12:12:35 -08002324 "src/math/expminus-sse2-rr2-p5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002325 "src/math/roundd-sse2-cvt.c",
2326 "src/math/roundne-sse2-cvt.c",
2327 "src/math/roundu-sse2-cvt.c",
2328 "src/math/roundz-sse2-cvt.c",
2329 "src/math/sigmoid-sse2-rr2-lut64-p2-div.c",
2330 "src/math/sigmoid-sse2-rr2-lut64-p2-nr1.c",
2331 "src/math/sigmoid-sse2-rr2-lut64-p2-nr2.c",
2332 "src/math/sigmoid-sse2-rr2-p5-div.c",
2333 "src/math/sigmoid-sse2-rr2-p5-nr1.c",
2334 "src/math/sigmoid-sse2-rr2-p5-nr2.c",
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07002335 "src/qs8-dwconv/gen/up8x9-minmax-sse2-mul16.c",
2336 "src/qs8-dwconv/gen/up16x9-minmax-sse2-mul16.c",
2337 "src/qs8-dwconv/gen/up24x9-minmax-sse2-mul16.c",
Marat Dukhan159688f2020-08-06 10:34:29 -07002338 "src/qs8-gavgpool/gen/7p7x-minmax-sse2-c8-acc2.c",
2339 "src/qs8-gavgpool/gen/7p7x-minmax-sse2-c16-acc2.c",
2340 "src/qs8-gavgpool/gen/7p7x-minmax-sse2-c24-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002341 "src/qs8-gavgpool/gen/7x-minmax-sse2-c8-acc2.c",
2342 "src/qs8-gavgpool/gen/7x-minmax-sse2-c16-acc2.c",
2343 "src/qs8-gavgpool/gen/7x-minmax-sse2-c24-acc2.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002344 "src/qs8-gemm/gen/1x4c2-minmax-sse2-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002345 "src/qs8-gemm/gen/1x4c2-minmax-sse2-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002346 "src/qs8-gemm/gen/1x4c2-xw-minmax-sse2.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002347 "src/qs8-gemm/gen/1x4c8-minmax-sse2-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002348 "src/qs8-gemm/gen/1x4c8-minmax-sse2-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002349 "src/qs8-gemm/gen/1x4c8-xw-minmax-sse2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002350 "src/qs8-gemm/gen/2x4c8-minmax-sse2-ld64.c",
2351 "src/qs8-gemm/gen/2x4c8-minmax-sse2-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002352 "src/qs8-gemm/gen/2x4c8-xw-minmax-sse2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002353 "src/qs8-gemm/gen/3x4c8-minmax-sse2-ld64.c",
2354 "src/qs8-gemm/gen/3x4c8-minmax-sse2-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002355 "src/qs8-gemm/gen/3x4c8-xw-minmax-sse2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002356 "src/qs8-gemm/gen/4x4c2-minmax-sse2-ld64.c",
2357 "src/qs8-gemm/gen/4x4c2-minmax-sse2-ld128.c",
2358 "src/qs8-gemm/gen/4x4c2-xw-minmax-sse2.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002359 "src/qs8-igemm/gen/1x4c2-minmax-sse2-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002360 "src/qs8-igemm/gen/1x4c2-minmax-sse2-ld128.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002361 "src/qs8-igemm/gen/1x4c8-minmax-sse2-ld64.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002362 "src/qs8-igemm/gen/1x4c8-minmax-sse2-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002363 "src/qs8-igemm/gen/2x4c8-minmax-sse2-ld64.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002364 "src/qs8-igemm/gen/2x4c8-minmax-sse2-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002365 "src/qs8-igemm/gen/3x4c8-minmax-sse2-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002366 "src/qs8-igemm/gen/3x4c8-minmax-sse2-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002367 "src/qs8-igemm/gen/4x4c2-minmax-sse2-ld64.c",
2368 "src/qs8-igemm/gen/4x4c2-minmax-sse2-ld128.c",
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07002369 "src/qs8-requantization/fp32-sse2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002370 "src/qs8-requantization/precise-sse2.c",
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07002371 "src/qs8-requantization/q31-sse2.c",
Marat Dukhand9f3ad42020-08-10 12:30:58 -07002372 "src/qs8-vadd/gen/minmax-sse2-mul16-ld64-x8.c",
2373 "src/qs8-vadd/gen/minmax-sse2-mul16-ld64-x16.c",
2374 "src/qs8-vadd/gen/minmax-sse2-mul16-ld64-x24.c",
2375 "src/qs8-vadd/gen/minmax-sse2-mul16-ld64-x32.c",
Marat Dukhan0270d9f2020-08-11 00:56:46 -07002376 "src/qs8-vaddc/gen/minmax-sse2-mul16-ld64-x8.c",
2377 "src/qs8-vaddc/gen/minmax-sse2-mul16-ld64-x16.c",
2378 "src/qs8-vaddc/gen/minmax-sse2-mul16-ld64-x24.c",
2379 "src/qs8-vaddc/gen/minmax-sse2-mul16-ld64-x32.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -07002380 "src/qu8-avgpool/9p8x-minmax-sse2-c8.c",
2381 "src/qu8-avgpool/9x-minmax-sse2-c8.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -07002382 "src/qu8-dwconv/up8x9-minmax-sse2.c",
2383 "src/qu8-gavgpool/7p7x-minmax-sse2-c8.c",
2384 "src/qu8-gavgpool/7x-minmax-sse2-c8.c",
2385 "src/qu8-gemm/2x4c8-minmax-sse2.c",
2386 "src/qu8-gemm/4x4c2-minmax-sse2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002387 "src/qu8-igemm/4x4c2-minmax-sse2.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07002388 "src/qu8-requantization/fp32-sse2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002389 "src/qu8-requantization/precise-sse2.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07002390 "src/qu8-requantization/q31-sse2.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -07002391 "src/qu8-vadd/minmax-sse2.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -07002392 "src/u8-clamp/sse2-x64.c",
Marat Dukhan99936602020-04-11 16:47:01 -07002393 "src/u8-maxpool/9p8x-minmax-sse2-c16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002394 "src/u8-rmax/sse2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002395 "src/x8-zip/x2-sse2.c",
2396 "src/x8-zip/x3-sse2.c",
2397 "src/x8-zip/x4-sse2.c",
2398 "src/x8-zip/xm-sse2.c",
Marat Dukhan57dccd82020-04-14 00:53:10 -07002399 "src/x32-unpool/sse2.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002400 "src/x32-zip/x2-sse2.c",
2401 "src/x32-zip/x3-sse2.c",
2402 "src/x32-zip/x4-sse2.c",
2403 "src/x32-zip/xm-sse2.c",
Marat Dukhanfe7acb62020-03-09 19:30:05 -07002404]
2405
2406SSSE3_UKERNELS = [
Frank Barchard35db7d02020-10-26 13:37:34 -07002407 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-1x4-acc2.c",
2408 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-1x4-acc3.c",
2409 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-1x4-acc4.c",
Marat Dukhan98f2eeb2020-10-23 23:13:41 -07002410 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-1x4.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07002411 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-2x4-acc2.c",
Marat Dukhan98f2eeb2020-10-23 23:13:41 -07002412 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-2x4.c",
2413 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-3x4.c",
2414 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-4x4.c",
2415 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-5x4.c",
2416 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-6x4.c",
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07002417 "src/qs8-dwconv/gen/up8x9-minmax-ssse3-mul16.c",
2418 "src/qs8-dwconv/gen/up16x9-minmax-ssse3-mul16.c",
2419 "src/qs8-dwconv/gen/up24x9-minmax-ssse3-mul16.c",
Marat Dukhan159688f2020-08-06 10:34:29 -07002420 "src/qs8-gavgpool/gen/7p7x-minmax-ssse3-c8-acc2.c",
2421 "src/qs8-gavgpool/gen/7p7x-minmax-ssse3-c16-acc2.c",
2422 "src/qs8-gavgpool/gen/7p7x-minmax-ssse3-c24-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002423 "src/qs8-gavgpool/gen/7x-minmax-ssse3-c8-acc2.c",
2424 "src/qs8-gavgpool/gen/7x-minmax-ssse3-c16-acc2.c",
2425 "src/qs8-gavgpool/gen/7x-minmax-ssse3-c24-acc2.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002426 "src/qs8-gemm/gen/1x4c2-minmax-ssse3-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002427 "src/qs8-gemm/gen/1x4c2-minmax-ssse3-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002428 "src/qs8-gemm/gen/1x4c2-xw-minmax-ssse3.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002429 "src/qs8-gemm/gen/1x4c8-minmax-ssse3-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002430 "src/qs8-gemm/gen/1x4c8-minmax-ssse3-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002431 "src/qs8-gemm/gen/1x4c8-xw-minmax-ssse3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002432 "src/qs8-gemm/gen/2x4c8-minmax-ssse3-ld64.c",
2433 "src/qs8-gemm/gen/2x4c8-minmax-ssse3-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002434 "src/qs8-gemm/gen/2x4c8-xw-minmax-ssse3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002435 "src/qs8-gemm/gen/3x4c8-minmax-ssse3-ld64.c",
2436 "src/qs8-gemm/gen/3x4c8-minmax-ssse3-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002437 "src/qs8-gemm/gen/3x4c8-xw-minmax-ssse3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002438 "src/qs8-gemm/gen/4x4c2-minmax-ssse3-ld64.c",
2439 "src/qs8-gemm/gen/4x4c2-minmax-ssse3-ld128.c",
2440 "src/qs8-gemm/gen/4x4c2-xw-minmax-ssse3.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002441 "src/qs8-igemm/gen/1x4c2-minmax-ssse3-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002442 "src/qs8-igemm/gen/1x4c2-minmax-ssse3-ld128.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002443 "src/qs8-igemm/gen/1x4c8-minmax-ssse3-ld64.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002444 "src/qs8-igemm/gen/1x4c8-minmax-ssse3-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002445 "src/qs8-igemm/gen/2x4c8-minmax-ssse3-ld64.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002446 "src/qs8-igemm/gen/2x4c8-minmax-ssse3-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002447 "src/qs8-igemm/gen/3x4c8-minmax-ssse3-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002448 "src/qs8-igemm/gen/3x4c8-minmax-ssse3-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002449 "src/qs8-igemm/gen/4x4c2-minmax-ssse3-ld64.c",
2450 "src/qs8-igemm/gen/4x4c2-minmax-ssse3-ld128.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -07002451 "src/qs8-requantization/precise-ssse3.c",
2452 "src/qs8-requantization/q31-ssse3.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07002453 "src/qu8-requantization/precise-ssse3.c",
2454 "src/qu8-requantization/q31-ssse3.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002455]
2456
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08002457SSE41_UKERNELS = [
Marat Dukhan40a672f2019-11-25 03:08:22 -08002458 "src/f32-prelu/gen/sse41-2x4.c",
2459 "src/f32-prelu/gen/sse41-2x8.c",
Marat Dukhand243c1a2020-09-17 18:48:10 -07002460 "src/f32-sigmoid/gen/sse41-lut64-p2-div-x4.c",
2461 "src/f32-sigmoid/gen/sse41-lut64-p2-div-x8.c",
2462 "src/f32-sigmoid/gen/sse41-lut64-p2-div-x12.c",
2463 "src/f32-sigmoid/gen/sse41-lut64-p2-div-x16.c",
2464 "src/f32-sigmoid/gen/sse41-lut64-p2-div-x20.c",
2465 "src/f32-sigmoid/gen/sse41-lut64-p2-div-x24.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002466 "src/f32-sigmoid/gen/sse41-p5-div-x4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08002467 "src/f32-sigmoid/gen/sse41-p5-div-x8.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002468 "src/f32-sigmoid/gen/sse41-p5-div-x12.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08002469 "src/f32-sigmoid/gen/sse41-p5-div-x16.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002470 "src/f32-sigmoid/gen/sse41-p5-div-x20.c",
2471 "src/f32-sigmoid/gen/sse41-p5-div-x24.c",
Marat Dukhaned6baaf2020-12-01 15:07:08 -08002472 "src/f32-velu/gen/velu-sse41-rr2-lut16-p3-x4.c",
2473 "src/f32-velu/gen/velu-sse41-rr2-lut16-p3-x8.c",
2474 "src/f32-velu/gen/velu-sse41-rr2-lut16-p3-x12.c",
2475 "src/f32-velu/gen/velu-sse41-rr2-lut16-p3-x16.c",
2476 "src/f32-velu/gen/velu-sse41-rr2-lut16-p3-x20.c",
2477 "src/f32-velu/gen/velu-sse41-rr2-lut16-p3-x24.c",
2478 "src/f32-velu/gen/velu-sse41-rr2-p6-x4.c",
2479 "src/f32-velu/gen/velu-sse41-rr2-p6-x8.c",
2480 "src/f32-velu/gen/velu-sse41-rr2-p6-x12.c",
2481 "src/f32-velu/gen/velu-sse41-rr2-p6-x16.c",
2482 "src/f32-velu/gen/velu-sse41-rr2-p6-x20.c",
2483 "src/f32-velu/gen/velu-sse41-rr2-p6-x24.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07002484 "src/f32-vlrelu/gen/vlrelu-sse41-x4.c",
2485 "src/f32-vlrelu/gen/vlrelu-sse41-x8.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07002486 "src/f32-vrnd/gen/vrndd-sse41-x4.c",
2487 "src/f32-vrnd/gen/vrndd-sse41-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002488 "src/f32-vrnd/gen/vrndne-sse41-x4.c",
2489 "src/f32-vrnd/gen/vrndne-sse41-x8.c",
2490 "src/f32-vrnd/gen/vrndu-sse41-x4.c",
2491 "src/f32-vrnd/gen/vrndu-sse41-x8.c",
2492 "src/f32-vrnd/gen/vrndz-sse41-x4.c",
2493 "src/f32-vrnd/gen/vrndz-sse41-x8.c",
2494 "src/math/roundd-sse41.c",
2495 "src/math/roundne-sse41.c",
2496 "src/math/roundu-sse41.c",
2497 "src/math/roundz-sse41.c",
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07002498 "src/qs8-dwconv/gen/up8x9-minmax-sse41-mul16.c",
2499 "src/qs8-dwconv/gen/up16x9-minmax-sse41-mul16.c",
2500 "src/qs8-dwconv/gen/up24x9-minmax-sse41-mul16.c",
Marat Dukhan159688f2020-08-06 10:34:29 -07002501 "src/qs8-gavgpool/gen/7p7x-minmax-sse41-c8-acc2.c",
2502 "src/qs8-gavgpool/gen/7p7x-minmax-sse41-c16-acc2.c",
2503 "src/qs8-gavgpool/gen/7p7x-minmax-sse41-c24-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002504 "src/qs8-gavgpool/gen/7x-minmax-sse41-c8-acc2.c",
2505 "src/qs8-gavgpool/gen/7x-minmax-sse41-c16-acc2.c",
2506 "src/qs8-gavgpool/gen/7x-minmax-sse41-c24-acc2.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002507 "src/qs8-gemm/gen/1x4c2-minmax-sse41-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002508 "src/qs8-gemm/gen/1x4c2-minmax-sse41-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002509 "src/qs8-gemm/gen/1x4c2-xw-minmax-sse41.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002510 "src/qs8-gemm/gen/1x4c8-minmax-sse41-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002511 "src/qs8-gemm/gen/1x4c8-minmax-sse41-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002512 "src/qs8-gemm/gen/1x4c8-xw-minmax-sse41.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002513 "src/qs8-gemm/gen/2x4c8-minmax-sse41-ld64.c",
2514 "src/qs8-gemm/gen/2x4c8-minmax-sse41-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002515 "src/qs8-gemm/gen/2x4c8-xw-minmax-sse41.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002516 "src/qs8-gemm/gen/3x4c8-minmax-sse41-ld64.c",
2517 "src/qs8-gemm/gen/3x4c8-minmax-sse41-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002518 "src/qs8-gemm/gen/3x4c8-xw-minmax-sse41.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002519 "src/qs8-gemm/gen/4x4c2-minmax-sse41-ld64.c",
2520 "src/qs8-gemm/gen/4x4c2-minmax-sse41-ld128.c",
2521 "src/qs8-gemm/gen/4x4c2-xw-minmax-sse41.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002522 "src/qs8-igemm/gen/1x4c2-minmax-sse41-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002523 "src/qs8-igemm/gen/1x4c2-minmax-sse41-ld128.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002524 "src/qs8-igemm/gen/1x4c8-minmax-sse41-ld64.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002525 "src/qs8-igemm/gen/1x4c8-minmax-sse41-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002526 "src/qs8-igemm/gen/2x4c8-minmax-sse41-ld64.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002527 "src/qs8-igemm/gen/2x4c8-minmax-sse41-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002528 "src/qs8-igemm/gen/3x4c8-minmax-sse41-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002529 "src/qs8-igemm/gen/3x4c8-minmax-sse41-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002530 "src/qs8-igemm/gen/4x4c2-minmax-sse41-ld64.c",
2531 "src/qs8-igemm/gen/4x4c2-minmax-sse41-ld128.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -07002532 "src/qs8-requantization/fp32-sse4.c",
2533 "src/qs8-requantization/precise-sse4.c",
2534 "src/qs8-requantization/q31-sse4.c",
Marat Dukhand9f3ad42020-08-10 12:30:58 -07002535 "src/qs8-vadd/gen/minmax-sse41-mul16-ld64-x8.c",
2536 "src/qs8-vadd/gen/minmax-sse41-mul16-ld64-x16.c",
2537 "src/qs8-vadd/gen/minmax-sse41-mul16-ld64-x24.c",
2538 "src/qs8-vadd/gen/minmax-sse41-mul16-ld64-x32.c",
Marat Dukhanbb9225e2020-09-06 22:40:56 -07002539 "src/qs8-vadd/gen/minmax-sse41-mul32-ld32-x8.c",
2540 "src/qs8-vadd/gen/minmax-sse41-mul32-ld32-x16.c",
2541 "src/qs8-vadd/gen/minmax-sse41-mul32-ld32-x24.c",
2542 "src/qs8-vadd/gen/minmax-sse41-mul32-ld32-x32.c",
Marat Dukhan0270d9f2020-08-11 00:56:46 -07002543 "src/qs8-vaddc/gen/minmax-sse41-mul16-ld64-x8.c",
2544 "src/qs8-vaddc/gen/minmax-sse41-mul16-ld64-x16.c",
2545 "src/qs8-vaddc/gen/minmax-sse41-mul16-ld64-x24.c",
2546 "src/qs8-vaddc/gen/minmax-sse41-mul16-ld64-x32.c",
Marat Dukhanbb9225e2020-09-06 22:40:56 -07002547 "src/qs8-vaddc/gen/minmax-sse41-mul32-ld32-x8.c",
2548 "src/qs8-vaddc/gen/minmax-sse41-mul32-ld32-x16.c",
2549 "src/qs8-vaddc/gen/minmax-sse41-mul32-ld32-x24.c",
2550 "src/qs8-vaddc/gen/minmax-sse41-mul32-ld32-x32.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07002551 "src/qu8-requantization/precise-sse4.c",
2552 "src/qu8-requantization/q31-sse4.c",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08002553]
2554
Marat Dukhan08c4a432019-10-03 09:29:21 -07002555AVX_UKERNELS = [
Marat Dukhan5c5fa962020-03-10 18:38:33 -07002556 "src/f32-clamp/gen/avx-x8.c",
2557 "src/f32-clamp/gen/avx-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002558 "src/f32-dwconv/gen/up8x4-minmax-avx-acc2.c",
2559 "src/f32-dwconv/gen/up8x4-minmax-avx.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002560 "src/f32-dwconv/gen/up8x9-minmax-avx-acc2.c",
2561 "src/f32-dwconv/gen/up8x9-minmax-avx.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002562 "src/f32-dwconv/gen/up8x25-minmax-avx-acc2.c",
2563 "src/f32-dwconv/gen/up8x25-minmax-avx.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002564 "src/f32-dwconv/gen/up16x4-minmax-avx-acc2.c",
2565 "src/f32-dwconv/gen/up16x4-minmax-avx.c",
2566 "src/f32-dwconv/gen/up16x9-minmax-avx-acc2.c",
2567 "src/f32-dwconv/gen/up16x9-minmax-avx.c",
2568 "src/f32-dwconv/gen/up16x25-minmax-avx-acc2.c",
2569 "src/f32-dwconv/gen/up16x25-minmax-avx.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002570 "src/f32-gemm/gen-inc/1x8inc-minmax-avx-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002571 "src/f32-gemm/gen-inc/1x16inc-minmax-avx-broadcast.c",
2572 "src/f32-gemm/gen-inc/3x16inc-minmax-avx-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002573 "src/f32-gemm/gen-inc/4x8inc-minmax-avx-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002574 "src/f32-gemm/gen-inc/4x16inc-minmax-avx-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002575 "src/f32-gemm/gen-inc/5x8inc-minmax-avx-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002576 "src/f32-gemm/gen-inc/5x16inc-minmax-avx-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002577 "src/f32-gemm/gen-inc/6x8inc-minmax-avx-broadcast.c",
2578 "src/f32-gemm/gen-inc/7x8inc-minmax-avx-broadcast.c",
2579 "src/f32-gemm/gen/1x8-minmax-avx-broadcast.c",
2580 "src/f32-gemm/gen/1x16-minmax-avx-broadcast.c",
2581 "src/f32-gemm/gen/3x16-minmax-avx-broadcast.c",
2582 "src/f32-gemm/gen/4x8-minmax-avx-broadcast.c",
2583 "src/f32-gemm/gen/4x16-minmax-avx-broadcast.c",
2584 "src/f32-gemm/gen/5x8-minmax-avx-broadcast.c",
2585 "src/f32-gemm/gen/5x16-minmax-avx-broadcast.c",
2586 "src/f32-gemm/gen/6x8-minmax-avx-broadcast.c",
2587 "src/f32-gemm/gen/7x8-minmax-avx-broadcast.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07002588 "src/f32-hswish/gen/hswish-avx-x8.c",
2589 "src/f32-hswish/gen/hswish-avx-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002590 "src/f32-igemm/gen/1x8-minmax-avx-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002591 "src/f32-igemm/gen/1x16-minmax-avx-broadcast.c",
2592 "src/f32-igemm/gen/3x16-minmax-avx-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002593 "src/f32-igemm/gen/4x8-minmax-avx-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002594 "src/f32-igemm/gen/4x16-minmax-avx-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002595 "src/f32-igemm/gen/5x8-minmax-avx-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002596 "src/f32-igemm/gen/5x16-minmax-avx-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002597 "src/f32-igemm/gen/6x8-minmax-avx-broadcast.c",
2598 "src/f32-igemm/gen/7x8-minmax-avx-broadcast.c",
Marat Dukhan90eca0a2020-03-11 00:52:23 -07002599 "src/f32-prelu/gen/avx-2x8.c",
2600 "src/f32-prelu/gen/avx-2x16.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07002601 "src/f32-relu/gen/avx-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002602 "src/f32-relu/gen/avx-x16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002603 "src/f32-rmax/avx.c",
T.J. Alumbaughdc2b29c2020-10-14 13:56:08 -07002604 "src/f32-sigmoid/gen/avx-rr2-p5-div-x8.c",
2605 "src/f32-sigmoid/gen/avx-rr2-p5-div-x16.c",
2606 "src/f32-sigmoid/gen/avx-rr2-p5-div-x24.c",
2607 "src/f32-sigmoid/gen/avx-rr2-p5-div-x32.c",
2608 "src/f32-sigmoid/gen/avx-rr2-p5-div-x40.c",
2609 "src/f32-sigmoid/gen/avx-rr2-p5-div-x48.c",
2610 "src/f32-sigmoid/gen/avx-rr2-p5-div-x56.c",
2611 "src/f32-sigmoid/gen/avx-rr2-p5-div-x64.c",
2612 "src/f32-sigmoid/gen/avx-rr2-p5-div-x72.c",
2613 "src/f32-sigmoid/gen/avx-rr2-p5-div-x80.c",
2614 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x8.c",
2615 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x16.c",
2616 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x24.c",
2617 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x32.c",
2618 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x40.c",
2619 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x48.c",
2620 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x56.c",
2621 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x64.c",
2622 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x72.c",
2623 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x80.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002624 "src/f32-vbinary/gen/vadd-minmax-avx-x8.c",
2625 "src/f32-vbinary/gen/vadd-minmax-avx-x16.c",
2626 "src/f32-vbinary/gen/vaddc-minmax-avx-x8.c",
2627 "src/f32-vbinary/gen/vaddc-minmax-avx-x16.c",
2628 "src/f32-vbinary/gen/vdiv-minmax-avx-x8.c",
2629 "src/f32-vbinary/gen/vdiv-minmax-avx-x16.c",
2630 "src/f32-vbinary/gen/vdivc-minmax-avx-x8.c",
2631 "src/f32-vbinary/gen/vdivc-minmax-avx-x16.c",
Marat Dukhan9a88efe2019-12-10 15:54:24 -08002632 "src/f32-vbinary/gen/vmax-avx-x8.c",
2633 "src/f32-vbinary/gen/vmax-avx-x16.c",
2634 "src/f32-vbinary/gen/vmaxc-avx-x8.c",
2635 "src/f32-vbinary/gen/vmaxc-avx-x16.c",
2636 "src/f32-vbinary/gen/vmin-avx-x8.c",
2637 "src/f32-vbinary/gen/vmin-avx-x16.c",
2638 "src/f32-vbinary/gen/vminc-avx-x8.c",
2639 "src/f32-vbinary/gen/vminc-avx-x16.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002640 "src/f32-vbinary/gen/vmul-minmax-avx-x8.c",
2641 "src/f32-vbinary/gen/vmul-minmax-avx-x16.c",
2642 "src/f32-vbinary/gen/vmulc-minmax-avx-x8.c",
2643 "src/f32-vbinary/gen/vmulc-minmax-avx-x16.c",
2644 "src/f32-vbinary/gen/vrdivc-minmax-avx-x8.c",
2645 "src/f32-vbinary/gen/vrdivc-minmax-avx-x16.c",
2646 "src/f32-vbinary/gen/vrsubc-minmax-avx-x8.c",
2647 "src/f32-vbinary/gen/vrsubc-minmax-avx-x16.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07002648 "src/f32-vbinary/gen/vsqrdiff-avx-x8.c",
2649 "src/f32-vbinary/gen/vsqrdiff-avx-x16.c",
2650 "src/f32-vbinary/gen/vsqrdiffc-avx-x8.c",
2651 "src/f32-vbinary/gen/vsqrdiffc-avx-x16.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002652 "src/f32-vbinary/gen/vsub-minmax-avx-x8.c",
2653 "src/f32-vbinary/gen/vsub-minmax-avx-x16.c",
2654 "src/f32-vbinary/gen/vsubc-minmax-avx-x8.c",
2655 "src/f32-vbinary/gen/vsubc-minmax-avx-x16.c",
Marat Dukhaned6baaf2020-12-01 15:07:08 -08002656 "src/f32-velu/gen/velu-avx-rr2-lut4-p4-perm-x8.c",
2657 "src/f32-velu/gen/velu-avx-rr2-lut4-p4-perm-x16.c",
2658 "src/f32-velu/gen/velu-avx-rr2-lut4-p4-perm-x24.c",
2659 "src/f32-velu/gen/velu-avx-rr2-lut4-p4-perm-x32.c",
2660 "src/f32-velu/gen/velu-avx-rr2-lut4-p4-perm-x40.c",
2661 "src/f32-velu/gen/velu-avx-rr2-lut4-p4-perm-x48.c",
2662 "src/f32-velu/gen/velu-avx-rr2-lut16-p3-x8.c",
2663 "src/f32-velu/gen/velu-avx-rr2-lut16-p3-x16.c",
2664 "src/f32-velu/gen/velu-avx-rr2-lut16-p3-x24.c",
2665 "src/f32-velu/gen/velu-avx-rr2-lut16-p3-x32.c",
2666 "src/f32-velu/gen/velu-avx-rr2-lut16-p3-x40.c",
2667 "src/f32-velu/gen/velu-avx-rr2-lut16-p3-x48.c",
2668 "src/f32-velu/gen/velu-avx-rr2-p6-x8.c",
2669 "src/f32-velu/gen/velu-avx-rr2-p6-x16.c",
2670 "src/f32-velu/gen/velu-avx-rr2-p6-x24.c",
2671 "src/f32-velu/gen/velu-avx-rr2-p6-x32.c",
2672 "src/f32-velu/gen/velu-avx-rr2-p6-x40.c",
2673 "src/f32-velu/gen/velu-avx-rr2-p6-x48.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07002674 "src/f32-vlrelu/gen/vlrelu-avx-x8.c",
2675 "src/f32-vlrelu/gen/vlrelu-avx-x16.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07002676 "src/f32-vrnd/gen/vrndd-avx-x8.c",
2677 "src/f32-vrnd/gen/vrndd-avx-x16.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002678 "src/f32-vrnd/gen/vrndne-avx-x8.c",
2679 "src/f32-vrnd/gen/vrndne-avx-x16.c",
2680 "src/f32-vrnd/gen/vrndu-avx-x8.c",
2681 "src/f32-vrnd/gen/vrndu-avx-x16.c",
2682 "src/f32-vrnd/gen/vrndz-avx-x8.c",
2683 "src/f32-vrnd/gen/vrndz-avx-x16.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07002684 "src/f32-vscale/avx-x32.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07002685 "src/f32-vsqrt/gen/avx-sqrt-x8.c",
2686 "src/f32-vsqrt/gen/avx-sqrt-x16.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07002687 "src/f32-vunary/gen/vabs-avx-x8.c",
2688 "src/f32-vunary/gen/vabs-avx-x16.c",
2689 "src/f32-vunary/gen/vneg-avx-x8.c",
2690 "src/f32-vunary/gen/vneg-avx-x16.c",
2691 "src/f32-vunary/gen/vsqr-avx-x8.c",
2692 "src/f32-vunary/gen/vsqr-avx-x16.c",
Marat Dukhanb7633f22020-11-20 16:34:56 -08002693 "src/math/exp-avx-rr2-p5.c",
Marat Dukhande390d42020-11-29 19:32:18 -08002694 "src/math/expm1minus-avx-rr2-lut4-p4-perm.c",
2695 "src/math/expm1minus-avx-rr2-lut16-p3.c",
2696 "src/math/expm1minus-avx-rr2-p6.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002697 "src/math/sigmoid-avx-rr2-lut64-p2-div.c",
Marat Dukhan2e4727b2020-09-24 00:48:39 -07002698 "src/math/sigmoid-avx-rr2-p5-div.c",
Marat Dukhan39ab16f2020-10-16 14:30:35 -07002699 "src/math/sigmoid-avx-rr2-p5-nr1.c",
2700 "src/math/sigmoid-avx-rr2-p5-nr2.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002701]
2702
Marat Dukhan1566fee2020-08-02 21:55:41 -07002703XOP_UKERNELS = [
2704 "src/qs8-gemm/gen/1x4c2-minmax-xop-ld64.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002705 "src/qs8-gemm/gen/1x4c2-minmax-xop-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002706 "src/qs8-gemm/gen/1x4c2-xw-minmax-xop.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002707 "src/qs8-gemm/gen/1x4c8-minmax-xop-ld64.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002708 "src/qs8-gemm/gen/1x4c8-minmax-xop-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002709 "src/qs8-gemm/gen/1x4c8-xw-minmax-xop.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002710 "src/qs8-gemm/gen/2x4c8-minmax-xop-ld64.c",
2711 "src/qs8-gemm/gen/2x4c8-minmax-xop-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002712 "src/qs8-gemm/gen/2x4c8-xw-minmax-xop.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002713 "src/qs8-gemm/gen/3x4c8-minmax-xop-ld64.c",
2714 "src/qs8-gemm/gen/3x4c8-minmax-xop-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002715 "src/qs8-gemm/gen/3x4c8-xw-minmax-xop.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002716 "src/qs8-gemm/gen/4x4c2-minmax-xop-ld64.c",
2717 "src/qs8-gemm/gen/4x4c2-minmax-xop-ld128.c",
2718 "src/qs8-gemm/gen/4x4c2-xw-minmax-xop.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002719 "src/qs8-igemm/gen/1x4c2-minmax-xop-ld64.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002720 "src/qs8-igemm/gen/1x4c2-minmax-xop-ld128.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002721 "src/qs8-igemm/gen/1x4c8-minmax-xop-ld64.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002722 "src/qs8-igemm/gen/1x4c8-minmax-xop-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002723 "src/qs8-igemm/gen/2x4c8-minmax-xop-ld64.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002724 "src/qs8-igemm/gen/2x4c8-minmax-xop-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002725 "src/qs8-igemm/gen/3x4c8-minmax-xop-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002726 "src/qs8-igemm/gen/3x4c8-minmax-xop-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002727 "src/qs8-igemm/gen/4x4c2-minmax-xop-ld64.c",
2728 "src/qs8-igemm/gen/4x4c2-minmax-xop-ld128.c",
Marat Dukhanbb9225e2020-09-06 22:40:56 -07002729 "src/qs8-vadd/gen/minmax-xop-mul32-ld32-x8.c",
2730 "src/qs8-vadd/gen/minmax-xop-mul32-ld32-x16.c",
2731 "src/qs8-vadd/gen/minmax-xop-mul32-ld32-x24.c",
2732 "src/qs8-vadd/gen/minmax-xop-mul32-ld32-x32.c",
2733 "src/qs8-vaddc/gen/minmax-xop-mul32-ld32-x8.c",
2734 "src/qs8-vaddc/gen/minmax-xop-mul32-ld32-x16.c",
2735 "src/qs8-vaddc/gen/minmax-xop-mul32-ld32-x24.c",
2736 "src/qs8-vaddc/gen/minmax-xop-mul32-ld32-x32.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002737]
2738
Marat Dukhanfda12b82019-11-21 12:27:59 -08002739FMA3_UKERNELS = [
Marat Dukhan1c587112020-04-08 20:04:28 -07002740 "src/f32-dwconv/gen/up8x4-minmax-fma3-acc2.c",
2741 "src/f32-dwconv/gen/up8x4-minmax-fma3.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002742 "src/f32-dwconv/gen/up8x9-minmax-fma3-acc2.c",
2743 "src/f32-dwconv/gen/up8x9-minmax-fma3.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002744 "src/f32-dwconv/gen/up8x25-minmax-fma3-acc2.c",
2745 "src/f32-dwconv/gen/up8x25-minmax-fma3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002746 "src/f32-dwconv/gen/up16x4-minmax-fma3-acc2.c",
2747 "src/f32-dwconv/gen/up16x4-minmax-fma3.c",
2748 "src/f32-dwconv/gen/up16x9-minmax-fma3-acc2.c",
2749 "src/f32-dwconv/gen/up16x9-minmax-fma3.c",
2750 "src/f32-dwconv/gen/up16x25-minmax-fma3-acc2.c",
2751 "src/f32-dwconv/gen/up16x25-minmax-fma3.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002752 "src/f32-gemm/gen-inc/1x8inc-minmax-fma3-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002753 "src/f32-gemm/gen-inc/1x16inc-minmax-fma3-broadcast.c",
2754 "src/f32-gemm/gen-inc/1x16s4inc-minmax-fma3-broadcast.c",
2755 "src/f32-gemm/gen-inc/3x16inc-minmax-fma3-broadcast.c",
2756 "src/f32-gemm/gen-inc/3x16s4inc-minmax-fma3-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002757 "src/f32-gemm/gen-inc/4x8inc-minmax-fma3-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002758 "src/f32-gemm/gen-inc/4x16inc-minmax-fma3-broadcast.c",
2759 "src/f32-gemm/gen-inc/4x16s4inc-minmax-fma3-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002760 "src/f32-gemm/gen-inc/5x8inc-minmax-fma3-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002761 "src/f32-gemm/gen-inc/5x16inc-minmax-fma3-broadcast.c",
2762 "src/f32-gemm/gen-inc/5x16s4inc-minmax-fma3-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002763 "src/f32-gemm/gen-inc/6x8inc-minmax-fma3-broadcast.c",
2764 "src/f32-gemm/gen-inc/7x8inc-minmax-fma3-broadcast.c",
2765 "src/f32-gemm/gen-inc/8x8inc-minmax-fma3-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002766 "src/f32-gemm/gen/1x8-minmax-fma3-broadcast.c",
2767 "src/f32-gemm/gen/1x16-minmax-fma3-broadcast.c",
2768 "src/f32-gemm/gen/1x16s4-minmax-fma3-broadcast.c",
2769 "src/f32-gemm/gen/3x16-minmax-fma3-broadcast.c",
2770 "src/f32-gemm/gen/3x16s4-minmax-fma3-broadcast.c",
2771 "src/f32-gemm/gen/4x8-minmax-fma3-broadcast.c",
2772 "src/f32-gemm/gen/4x16-minmax-fma3-broadcast.c",
2773 "src/f32-gemm/gen/4x16s4-minmax-fma3-broadcast.c",
2774 "src/f32-gemm/gen/5x8-minmax-fma3-broadcast.c",
2775 "src/f32-gemm/gen/5x16-minmax-fma3-broadcast.c",
2776 "src/f32-gemm/gen/5x16s4-minmax-fma3-broadcast.c",
2777 "src/f32-gemm/gen/6x8-minmax-fma3-broadcast.c",
2778 "src/f32-gemm/gen/7x8-minmax-fma3-broadcast.c",
2779 "src/f32-gemm/gen/8x8-minmax-fma3-broadcast.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07002780 "src/f32-hswish/gen/hswish-fma3-x8.c",
2781 "src/f32-hswish/gen/hswish-fma3-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002782 "src/f32-igemm/gen/1x8-minmax-fma3-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002783 "src/f32-igemm/gen/1x16-minmax-fma3-broadcast.c",
2784 "src/f32-igemm/gen/1x16s4-minmax-fma3-broadcast.c",
2785 "src/f32-igemm/gen/3x16-minmax-fma3-broadcast.c",
2786 "src/f32-igemm/gen/3x16s4-minmax-fma3-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002787 "src/f32-igemm/gen/4x8-minmax-fma3-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002788 "src/f32-igemm/gen/4x16-minmax-fma3-broadcast.c",
2789 "src/f32-igemm/gen/4x16s4-minmax-fma3-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002790 "src/f32-igemm/gen/5x8-minmax-fma3-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002791 "src/f32-igemm/gen/5x16-minmax-fma3-broadcast.c",
2792 "src/f32-igemm/gen/5x16s4-minmax-fma3-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002793 "src/f32-igemm/gen/6x8-minmax-fma3-broadcast.c",
2794 "src/f32-igemm/gen/7x8-minmax-fma3-broadcast.c",
2795 "src/f32-igemm/gen/8x8-minmax-fma3-broadcast.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07002796 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x8.c",
2797 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x16.c",
2798 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x24.c",
2799 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x32.c",
2800 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x40.c",
2801 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x48.c",
2802 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x56.c",
2803 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x64.c",
Marat Dukhan84000762020-06-29 18:38:43 -07002804 "src/math/sqrt-fma3-nr1fma.c",
Marat Dukhan84000762020-06-29 18:38:43 -07002805 "src/math/sqrt-fma3-nr1fma1adj.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002806 "src/math/sqrt-fma3-nr2fma.c",
Marat Dukhanfda12b82019-11-21 12:27:59 -08002807]
2808
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002809AVX2_UKERNELS = [
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002810 "src/f32-raddexpminusmax/gen/avx2-p5-x64-acc2.c",
2811 "src/f32-raddexpminusmax/gen/avx2-p5-x64-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002812 "src/f32-raddexpminusmax/gen/avx2-p5-x64.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002813 "src/f32-raddexpminusmax/gen/avx2-p5-x72-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002814 "src/f32-raddexpminusmax/gen/avx2-p5-x72.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002815 "src/f32-raddexpminusmax/gen/avx2-p5-x80-acc2.c",
2816 "src/f32-raddexpminusmax/gen/avx2-p5-x80-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002817 "src/f32-raddexpminusmax/gen/avx2-p5-x80.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002818 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc2.c",
2819 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc3.c",
2820 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc6.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002821 "src/f32-raddexpminusmax/gen/avx2-p5-x96.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002822 "src/f32-raddextexp/gen/avx2-p5-x64-acc2.c",
2823 "src/f32-raddextexp/gen/avx2-p5-x64-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002824 "src/f32-raddextexp/gen/avx2-p5-x64.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002825 "src/f32-raddextexp/gen/avx2-p5-x72-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002826 "src/f32-raddextexp/gen/avx2-p5-x72.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002827 "src/f32-raddextexp/gen/avx2-p5-x80-acc2.c",
2828 "src/f32-raddextexp/gen/avx2-p5-x80-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002829 "src/f32-raddextexp/gen/avx2-p5-x80.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002830 "src/f32-raddextexp/gen/avx2-p5-x96-acc2.c",
2831 "src/f32-raddextexp/gen/avx2-p5-x96-acc3.c",
2832 "src/f32-raddextexp/gen/avx2-p5-x96-acc6.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002833 "src/f32-raddextexp/gen/avx2-p5-x96.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002834 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64-acc2.c",
2835 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002836 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002837 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x72-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002838 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x72.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002839 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80-acc2.c",
2840 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002841 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002842 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc2.c",
2843 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc3.c",
2844 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc6.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002845 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96.c",
Marat Dukhanfa0a4322020-01-06 16:14:29 -08002846 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x8.c",
2847 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x16.c",
2848 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x24.c",
2849 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x32.c",
2850 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x40.c",
2851 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x48.c",
2852 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x56.c",
2853 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x64.c",
2854 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x72.c",
2855 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x80.c",
2856 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x8.c",
2857 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x16.c",
2858 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x24.c",
2859 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x32.c",
2860 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x40.c",
2861 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x48.c",
2862 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x56.c",
2863 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x64.c",
2864 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x72.c",
2865 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x80.c",
2866 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x8.c",
2867 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x16.c",
2868 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x24.c",
2869 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x32.c",
2870 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x40.c",
2871 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x48.c",
2872 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x56.c",
2873 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x64.c",
2874 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x72.c",
2875 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x80.c",
Marat Dukhaned6baaf2020-12-01 15:07:08 -08002876 "src/f32-velu/gen/velu-avx2-rr1-lut4-p4-perm-x8.c",
2877 "src/f32-velu/gen/velu-avx2-rr1-lut4-p4-perm-x16.c",
2878 "src/f32-velu/gen/velu-avx2-rr1-lut4-p4-perm-x24.c",
2879 "src/f32-velu/gen/velu-avx2-rr1-lut4-p4-perm-x32.c",
2880 "src/f32-velu/gen/velu-avx2-rr1-lut4-p4-perm-x40.c",
2881 "src/f32-velu/gen/velu-avx2-rr1-lut4-p4-perm-x48.c",
2882 "src/f32-velu/gen/velu-avx2-rr1-lut4-p4-perm-x56.c",
2883 "src/f32-velu/gen/velu-avx2-rr1-lut4-p4-perm-x64.c",
2884 "src/f32-velu/gen/velu-avx2-rr1-lut4-p4-perm-x72.c",
2885 "src/f32-velu/gen/velu-avx2-rr1-lut4-p4-perm-x80.c",
2886 "src/f32-velu/gen/velu-avx2-rr1-lut8-p4-perm-x8.c",
2887 "src/f32-velu/gen/velu-avx2-rr1-lut8-p4-perm-x16.c",
2888 "src/f32-velu/gen/velu-avx2-rr1-lut8-p4-perm-x24.c",
2889 "src/f32-velu/gen/velu-avx2-rr1-lut8-p4-perm-x32.c",
2890 "src/f32-velu/gen/velu-avx2-rr1-lut8-p4-perm-x40.c",
2891 "src/f32-velu/gen/velu-avx2-rr1-lut8-p4-perm-x48.c",
2892 "src/f32-velu/gen/velu-avx2-rr1-lut8-p4-perm-x56.c",
2893 "src/f32-velu/gen/velu-avx2-rr1-lut8-p4-perm-x64.c",
2894 "src/f32-velu/gen/velu-avx2-rr1-lut8-p4-perm-x72.c",
2895 "src/f32-velu/gen/velu-avx2-rr1-lut8-p4-perm-x80.c",
2896 "src/f32-velu/gen/velu-avx2-rr1-lut16-p3-gather-x8.c",
2897 "src/f32-velu/gen/velu-avx2-rr1-lut16-p3-gather-x16.c",
2898 "src/f32-velu/gen/velu-avx2-rr1-lut16-p3-gather-x24.c",
2899 "src/f32-velu/gen/velu-avx2-rr1-lut16-p3-gather-x32.c",
2900 "src/f32-velu/gen/velu-avx2-rr1-lut16-p3-gather-x40.c",
2901 "src/f32-velu/gen/velu-avx2-rr1-lut16-p3-gather-x48.c",
2902 "src/f32-velu/gen/velu-avx2-rr1-lut16-p3-gather-x56.c",
2903 "src/f32-velu/gen/velu-avx2-rr1-lut16-p3-gather-x64.c",
2904 "src/f32-velu/gen/velu-avx2-rr1-lut16-p3-gather-x72.c",
2905 "src/f32-velu/gen/velu-avx2-rr1-lut16-p3-gather-x80.c",
2906 "src/f32-velu/gen/velu-avx2-rr1-p6-x8.c",
2907 "src/f32-velu/gen/velu-avx2-rr1-p6-x16.c",
2908 "src/f32-velu/gen/velu-avx2-rr1-p6-x24.c",
2909 "src/f32-velu/gen/velu-avx2-rr1-p6-x32.c",
2910 "src/f32-velu/gen/velu-avx2-rr1-p6-x40.c",
2911 "src/f32-velu/gen/velu-avx2-rr1-p6-x48.c",
2912 "src/f32-velu/gen/velu-avx2-rr1-p6-x56.c",
2913 "src/f32-velu/gen/velu-avx2-rr1-p6-x64.c",
2914 "src/f32-velu/gen/velu-avx2-rr1-p6-x72.c",
2915 "src/f32-velu/gen/velu-avx2-rr1-p6-x80.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002916 "src/f32-vscaleexpminusmax/gen/avx2-p5-x8.c",
2917 "src/f32-vscaleexpminusmax/gen/avx2-p5-x16.c",
2918 "src/f32-vscaleexpminusmax/gen/avx2-p5-x24.c",
2919 "src/f32-vscaleexpminusmax/gen/avx2-p5-x32.c",
2920 "src/f32-vscaleexpminusmax/gen/avx2-p5-x40.c",
2921 "src/f32-vscaleexpminusmax/gen/avx2-p5-x48.c",
2922 "src/f32-vscaleexpminusmax/gen/avx2-p5-x56.c",
2923 "src/f32-vscaleexpminusmax/gen/avx2-p5-x64.c",
2924 "src/f32-vscaleexpminusmax/gen/avx2-p5-x72.c",
2925 "src/f32-vscaleexpminusmax/gen/avx2-p5-x80.c",
2926 "src/f32-vscaleexpminusmax/gen/avx2-p5-x88.c",
2927 "src/f32-vscaleexpminusmax/gen/avx2-p5-x96.c",
2928 "src/f32-vscaleextexp/gen/avx2-p5-x8.c",
2929 "src/f32-vscaleextexp/gen/avx2-p5-x16.c",
2930 "src/f32-vscaleextexp/gen/avx2-p5-x24.c",
2931 "src/f32-vscaleextexp/gen/avx2-p5-x32.c",
2932 "src/f32-vscaleextexp/gen/avx2-p5-x40.c",
2933 "src/f32-vscaleextexp/gen/avx2-p5-x48.c",
2934 "src/f32-vscaleextexp/gen/avx2-p5-x56.c",
2935 "src/f32-vscaleextexp/gen/avx2-p5-x64.c",
2936 "src/f32-vscaleextexp/gen/avx2-p5-x72.c",
2937 "src/f32-vscaleextexp/gen/avx2-p5-x80.c",
2938 "src/f32-vscaleextexp/gen/avx2-p5-x88.c",
2939 "src/f32-vscaleextexp/gen/avx2-p5-x96.c",
Marat Dukhanb7633f22020-11-20 16:34:56 -08002940 "src/math/exp-avx2-rr2-lut8-p3-perm.c",
2941 "src/math/exp-avx2-rr2-lut8-p4-perm.c",
2942 "src/math/exp-avx2-rr2-p5.c",
Marat Dukhande390d42020-11-29 19:32:18 -08002943 "src/math/expm1minus-avx2-rr1-lut4-p4-perm.c",
2944 "src/math/expm1minus-avx2-rr1-lut8-p4-perm.c",
2945 "src/math/expm1minus-avx2-rr1-lut16-p3-gather.c",
2946 "src/math/expm1minus-avx2-rr1-p6.c",
Frank Barcharde7223ee2020-12-04 19:04:01 -08002947 "src/math/expminus-avx2-rr2-p5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002948 "src/math/extexp-avx2-p5.c",
2949 "src/math/sigmoid-avx2-rr1-lut64-p2-gather-div.c",
2950 "src/math/sigmoid-avx2-rr1-lut64-p2-gather-nr1fma.c",
2951 "src/math/sigmoid-avx2-rr1-lut64-p2-gather-nr2fma.c",
2952 "src/math/sigmoid-avx2-rr1-lut64-p2-gather-nr2fma1adj.c",
2953 "src/math/sigmoid-avx2-rr1-p5-div.c",
2954 "src/math/sigmoid-avx2-rr1-p5-nr1fma.c",
2955 "src/math/sigmoid-avx2-rr1-p5-nr2fma.c",
2956 "src/math/sigmoid-avx2-rr2-lut64-p2-gather-div.c",
2957 "src/math/sigmoid-avx2-rr2-lut64-p2-gather-nr1fma.c",
2958 "src/math/sigmoid-avx2-rr2-lut64-p2-gather-nr2fma.c",
2959 "src/math/sigmoid-avx2-rr2-lut64-p2-gather-nr2fma1adj.c",
2960 "src/math/sigmoid-avx2-rr2-p5-div.c",
2961 "src/math/sigmoid-avx2-rr2-p5-nr1fma.c",
2962 "src/math/sigmoid-avx2-rr2-p5-nr2fma.c",
Marat Dukhand65a1522020-08-04 19:28:18 -07002963 "src/qs8-dwconv/gen/up8x9-minmax-avx2-mul32.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002964 "src/qs8-dwconv/gen/up16x9-minmax-avx2-mul16.c",
Marat Dukhand65a1522020-08-04 19:28:18 -07002965 "src/qs8-dwconv/gen/up16x9-minmax-avx2-mul32.c",
2966 "src/qs8-dwconv/gen/up24x9-minmax-avx2-mul32.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002967 "src/qs8-dwconv/gen/up32x9-minmax-avx2-mul16.c",
Marat Dukhand65a1522020-08-04 19:28:18 -07002968 "src/qs8-dwconv/gen/up32x9-minmax-avx2-mul32.c",
Marat Dukhan12809522020-08-02 22:23:51 -07002969 "src/qs8-gemm/gen/1x8c8-minmax-avx2.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002970 "src/qs8-gemm/gen/1x8c8-xw-minmax-avx2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002971 "src/qs8-gemm/gen/2x8c8-minmax-avx2.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002972 "src/qs8-gemm/gen/2x8c8-xw-minmax-avx2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002973 "src/qs8-gemm/gen/3x8c8-minmax-avx2.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002974 "src/qs8-gemm/gen/3x8c8-xw-minmax-avx2.c",
Marat Dukhan12809522020-08-02 22:23:51 -07002975 "src/qs8-igemm/gen/1x8c8-minmax-avx2.c",
2976 "src/qs8-igemm/gen/2x8c8-minmax-avx2.c",
2977 "src/qs8-igemm/gen/3x8c8-minmax-avx2.c",
Marat Dukhane6dc0b62020-09-08 23:57:14 -07002978 "src/qs8-vadd/gen/minmax-avx2-mul32-ld64-x8.c",
2979 "src/qs8-vadd/gen/minmax-avx2-mul32-ld64-x16.c",
2980 "src/qs8-vadd/gen/minmax-avx2-mul32-ld64-x24.c",
2981 "src/qs8-vadd/gen/minmax-avx2-mul32-ld64-x32.c",
2982 "src/qs8-vaddc/gen/minmax-avx2-mul32-ld64-x8.c",
2983 "src/qs8-vaddc/gen/minmax-avx2-mul32-ld64-x16.c",
2984 "src/qs8-vaddc/gen/minmax-avx2-mul32-ld64-x24.c",
2985 "src/qs8-vaddc/gen/minmax-avx2-mul32-ld64-x32.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002986]
2987
Marat Dukhan08c4a432019-10-03 09:29:21 -07002988AVX512F_UKERNELS = [
Marat Dukhan5c5fa962020-03-10 18:38:33 -07002989 "src/f32-clamp/gen/avx512f-x16.c",
2990 "src/f32-clamp/gen/avx512f-x32.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002991 "src/f32-dwconv/gen/up16x4-minmax-avx512f-acc2.c",
2992 "src/f32-dwconv/gen/up16x4-minmax-avx512f.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002993 "src/f32-dwconv/gen/up16x9-minmax-avx512f-acc2.c",
2994 "src/f32-dwconv/gen/up16x9-minmax-avx512f.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002995 "src/f32-dwconv/gen/up16x25-minmax-avx512f-acc2.c",
2996 "src/f32-dwconv/gen/up16x25-minmax-avx512f.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002997 "src/f32-dwconv/gen/up32x4-minmax-avx512f-acc2.c",
2998 "src/f32-dwconv/gen/up32x4-minmax-avx512f.c",
2999 "src/f32-dwconv/gen/up32x9-minmax-avx512f-acc2.c",
3000 "src/f32-dwconv/gen/up32x9-minmax-avx512f.c",
3001 "src/f32-dwconv/gen/up32x25-minmax-avx512f-acc2.c",
3002 "src/f32-dwconv/gen/up32x25-minmax-avx512f.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07003003 "src/f32-gemm/gen-inc/1x16inc-minmax-avx512f-broadcast.c",
3004 "src/f32-gemm/gen-inc/4x16inc-minmax-avx512f-broadcast.c",
3005 "src/f32-gemm/gen-inc/5x16inc-minmax-avx512f-broadcast.c",
3006 "src/f32-gemm/gen-inc/6x16inc-minmax-avx512f-broadcast.c",
3007 "src/f32-gemm/gen-inc/7x16inc-minmax-avx512f-broadcast.c",
3008 "src/f32-gemm/gen-inc/8x16inc-minmax-avx512f-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003009 "src/f32-gemm/gen/1x16-minmax-avx512f-broadcast.c",
3010 "src/f32-gemm/gen/4x16-minmax-avx512f-broadcast.c",
3011 "src/f32-gemm/gen/5x16-minmax-avx512f-broadcast.c",
3012 "src/f32-gemm/gen/6x16-minmax-avx512f-broadcast.c",
3013 "src/f32-gemm/gen/7x16-minmax-avx512f-broadcast.c",
3014 "src/f32-gemm/gen/8x16-minmax-avx512f-broadcast.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07003015 "src/f32-hswish/gen/hswish-avx512f-x16.c",
3016 "src/f32-hswish/gen/hswish-avx512f-x32.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07003017 "src/f32-igemm/gen/1x16-minmax-avx512f-broadcast.c",
3018 "src/f32-igemm/gen/4x16-minmax-avx512f-broadcast.c",
3019 "src/f32-igemm/gen/5x16-minmax-avx512f-broadcast.c",
3020 "src/f32-igemm/gen/6x16-minmax-avx512f-broadcast.c",
3021 "src/f32-igemm/gen/7x16-minmax-avx512f-broadcast.c",
3022 "src/f32-igemm/gen/8x16-minmax-avx512f-broadcast.c",
Marat Dukhan90eca0a2020-03-11 00:52:23 -07003023 "src/f32-prelu/gen/avx512f-2x16.c",
3024 "src/f32-prelu/gen/avx512f-2x32.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003025 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128-acc2.c",
3026 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003027 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003028 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x144-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003029 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x144.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003030 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160-acc2.c",
3031 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003032 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003033 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc2.c",
3034 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc3.c",
3035 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc6.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003036 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003037 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128-acc2.c",
3038 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003039 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003040 "src/f32-raddextexp/gen/avx512f-p5-scalef-x144-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003041 "src/f32-raddextexp/gen/avx512f-p5-scalef-x144.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003042 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160-acc2.c",
3043 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003044 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003045 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc2.c",
3046 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc3.c",
3047 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc6.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003048 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003049 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128-acc2.c",
3050 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003051 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003052 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x144-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003053 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x144.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003054 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160-acc2.c",
3055 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003056 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003057 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc2.c",
3058 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc3.c",
3059 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc6.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003060 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07003061 "src/f32-relu/gen/avx512f-x16.c",
3062 "src/f32-relu/gen/avx512f-x32.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003063 "src/f32-rmax/avx512f.c",
Marat Dukhand9ca7e62020-09-23 23:45:29 -07003064 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x16.c",
3065 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x32.c",
3066 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x48.c",
3067 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x64.c",
3068 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x80.c",
3069 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x96.c",
3070 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x112.c",
3071 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x128.c",
3072 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x16.c",
3073 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x32.c",
3074 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x48.c",
3075 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x64.c",
3076 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x80.c",
3077 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x96.c",
3078 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x112.c",
3079 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003080 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x16.c",
3081 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x32.c",
3082 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x48.c",
3083 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x64.c",
3084 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x80.c",
3085 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x96.c",
3086 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x112.c",
3087 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x128.c",
3088 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x16.c",
3089 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x32.c",
3090 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x48.c",
3091 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x64.c",
3092 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x80.c",
3093 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x96.c",
3094 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x112.c",
3095 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x128.c",
Marat Dukhand9ca7e62020-09-23 23:45:29 -07003096 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x16.c",
3097 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x32.c",
3098 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x48.c",
3099 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x64.c",
3100 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x80.c",
3101 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x96.c",
3102 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x112.c",
3103 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x128.c",
3104 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x16.c",
3105 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x32.c",
3106 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x48.c",
3107 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x64.c",
3108 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x80.c",
3109 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x96.c",
3110 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x112.c",
3111 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x128.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07003112 "src/f32-vbinary/gen/vadd-minmax-avx512f-x16.c",
3113 "src/f32-vbinary/gen/vadd-minmax-avx512f-x32.c",
3114 "src/f32-vbinary/gen/vaddc-minmax-avx512f-x16.c",
3115 "src/f32-vbinary/gen/vaddc-minmax-avx512f-x32.c",
3116 "src/f32-vbinary/gen/vdiv-minmax-avx512f-x16.c",
3117 "src/f32-vbinary/gen/vdiv-minmax-avx512f-x32.c",
3118 "src/f32-vbinary/gen/vdivc-minmax-avx512f-x16.c",
3119 "src/f32-vbinary/gen/vdivc-minmax-avx512f-x32.c",
Marat Dukhan9a88efe2019-12-10 15:54:24 -08003120 "src/f32-vbinary/gen/vmax-avx512f-x16.c",
3121 "src/f32-vbinary/gen/vmax-avx512f-x32.c",
3122 "src/f32-vbinary/gen/vmaxc-avx512f-x16.c",
3123 "src/f32-vbinary/gen/vmaxc-avx512f-x32.c",
3124 "src/f32-vbinary/gen/vmin-avx512f-x16.c",
3125 "src/f32-vbinary/gen/vmin-avx512f-x32.c",
3126 "src/f32-vbinary/gen/vminc-avx512f-x16.c",
3127 "src/f32-vbinary/gen/vminc-avx512f-x32.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07003128 "src/f32-vbinary/gen/vmul-minmax-avx512f-x16.c",
3129 "src/f32-vbinary/gen/vmul-minmax-avx512f-x32.c",
3130 "src/f32-vbinary/gen/vmulc-minmax-avx512f-x16.c",
3131 "src/f32-vbinary/gen/vmulc-minmax-avx512f-x32.c",
3132 "src/f32-vbinary/gen/vrdivc-minmax-avx512f-x16.c",
3133 "src/f32-vbinary/gen/vrdivc-minmax-avx512f-x32.c",
3134 "src/f32-vbinary/gen/vrsubc-minmax-avx512f-x16.c",
3135 "src/f32-vbinary/gen/vrsubc-minmax-avx512f-x32.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07003136 "src/f32-vbinary/gen/vsqrdiff-avx512f-x16.c",
3137 "src/f32-vbinary/gen/vsqrdiff-avx512f-x32.c",
3138 "src/f32-vbinary/gen/vsqrdiffc-avx512f-x16.c",
3139 "src/f32-vbinary/gen/vsqrdiffc-avx512f-x32.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07003140 "src/f32-vbinary/gen/vsub-minmax-avx512f-x16.c",
3141 "src/f32-vbinary/gen/vsub-minmax-avx512f-x32.c",
3142 "src/f32-vbinary/gen/vsubc-minmax-avx512f-x16.c",
3143 "src/f32-vbinary/gen/vsubc-minmax-avx512f-x32.c",
Marat Dukhaned6baaf2020-12-01 15:07:08 -08003144 "src/f32-velu/gen/velu-avx512f-rr1-lut16-p3-perm-x16.c",
3145 "src/f32-velu/gen/velu-avx512f-rr1-lut16-p3-perm-x32.c",
3146 "src/f32-velu/gen/velu-avx512f-rr1-lut16-p3-perm-x48.c",
3147 "src/f32-velu/gen/velu-avx512f-rr1-lut16-p3-perm-x64.c",
3148 "src/f32-velu/gen/velu-avx512f-rr1-lut16-p3-perm-x80.c",
3149 "src/f32-velu/gen/velu-avx512f-rr1-lut16-p3-perm-x96.c",
3150 "src/f32-velu/gen/velu-avx512f-rr1-lut16-p3-perm-x112.c",
3151 "src/f32-velu/gen/velu-avx512f-rr1-lut16-p3-perm-x128.c",
3152 "src/f32-velu/gen/velu-avx512f-rr1-p6-x16.c",
3153 "src/f32-velu/gen/velu-avx512f-rr1-p6-x32.c",
3154 "src/f32-velu/gen/velu-avx512f-rr1-p6-x48.c",
3155 "src/f32-velu/gen/velu-avx512f-rr1-p6-x64.c",
3156 "src/f32-velu/gen/velu-avx512f-rr1-p6-x80.c",
3157 "src/f32-velu/gen/velu-avx512f-rr1-p6-x96.c",
3158 "src/f32-velu/gen/velu-avx512f-rr1-p6-x112.c",
3159 "src/f32-velu/gen/velu-avx512f-rr1-p6-x128.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07003160 "src/f32-vlrelu/gen/vlrelu-avx512f-x16.c",
3161 "src/f32-vlrelu/gen/vlrelu-avx512f-x32.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003162 "src/f32-vrnd/gen/vrndd-avx512f-x16.c",
3163 "src/f32-vrnd/gen/vrndd-avx512f-x32.c",
3164 "src/f32-vrnd/gen/vrndne-avx512f-x16.c",
3165 "src/f32-vrnd/gen/vrndne-avx512f-x32.c",
3166 "src/f32-vrnd/gen/vrndu-avx512f-x16.c",
3167 "src/f32-vrnd/gen/vrndu-avx512f-x32.c",
3168 "src/f32-vrnd/gen/vrndz-avx512f-x16.c",
3169 "src/f32-vrnd/gen/vrndz-avx512f-x32.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07003170 "src/f32-vscale/avx512f-x64.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003171 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x16.c",
3172 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x32.c",
3173 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x48.c",
3174 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x64.c",
3175 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x80.c",
3176 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x96.c",
3177 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x112.c",
3178 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x128.c",
3179 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x144.c",
3180 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x160.c",
3181 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x176.c",
3182 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x192.c",
3183 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x16.c",
3184 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x32.c",
3185 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x48.c",
3186 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x64.c",
3187 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x80.c",
3188 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x96.c",
3189 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x112.c",
3190 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x128.c",
3191 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x144.c",
3192 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x160.c",
3193 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x176.c",
3194 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x192.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07003195 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x16.c",
3196 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x32.c",
3197 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x48.c",
3198 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x64.c",
3199 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x80.c",
3200 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x96.c",
3201 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x112.c",
3202 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x128.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07003203 "src/f32-vunary/gen/vabs-avx512f-x16.c",
3204 "src/f32-vunary/gen/vabs-avx512f-x32.c",
3205 "src/f32-vunary/gen/vneg-avx512f-x16.c",
3206 "src/f32-vunary/gen/vneg-avx512f-x32.c",
3207 "src/f32-vunary/gen/vsqr-avx512f-x16.c",
3208 "src/f32-vunary/gen/vsqr-avx512f-x32.c",
Marat Dukhanb7633f22020-11-20 16:34:56 -08003209 "src/math/exp-avx512f-rr2-lut16-p3-perm-scalef.c",
3210 "src/math/exp-avx512f-rr2-lut16-p3-perm.c",
3211 "src/math/exp-avx512f-rr2-lut32-p2-perm2-scalef.c",
3212 "src/math/exp-avx512f-rr2-lut32-p2-perm2.c",
3213 "src/math/exp-avx512f-rr2-p5-scalef.c",
3214 "src/math/exp-avx512f-rr2-p5.c",
Marat Dukhande390d42020-11-29 19:32:18 -08003215 "src/math/expm1minus-avx512f-rr1-lut16-p3-perm.c",
3216 "src/math/expm1minus-avx512f-rr1-p6.c",
Marat Dukhan98ba4412019-10-23 02:14:28 -07003217 "src/math/extexp-avx512f-p5.c",
Marat Dukhan6a34c5f2020-09-22 21:44:15 -07003218 "src/math/sigmoid-avx512f-rr1-lut16-p3-perm-scalef-div.c",
Marat Dukhan6a34c5f2020-09-22 21:44:15 -07003219 "src/math/sigmoid-avx512f-rr1-lut16-p3-perm-scalef-nr1fma.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003220 "src/math/sigmoid-avx512f-rr1-lut16-p3-perm-scalef-nr1fma1adj.c",
Marat Dukhan6a34c5f2020-09-22 21:44:15 -07003221 "src/math/sigmoid-avx512f-rr1-lut32-p2-perm2-scalef-div.c",
Marat Dukhan6a34c5f2020-09-22 21:44:15 -07003222 "src/math/sigmoid-avx512f-rr1-lut32-p2-perm2-scalef-nr1fma.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003223 "src/math/sigmoid-avx512f-rr1-lut32-p2-perm2-scalef-nr1fma1adj.c",
Marat Dukhan36173d22020-10-15 17:14:26 -07003224 "src/math/sigmoid-avx512f-rr1-lut64-p2-gather-scalef-div.c",
Marat Dukhan36173d22020-10-15 17:14:26 -07003225 "src/math/sigmoid-avx512f-rr1-lut64-p2-gather-scalef-nr1fma.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003226 "src/math/sigmoid-avx512f-rr1-lut64-p2-gather-scalef-nr1fma1adj.c",
3227 "src/math/sigmoid-avx512f-rr1-p5-scalef-div.c",
3228 "src/math/sigmoid-avx512f-rr1-p5-scalef-nr1fma.c",
3229 "src/math/sigmoid-avx512f-rr1-p5-scalef-nr1fma1adj.c",
3230 "src/math/sigmoid-avx512f-rr2-lut16-p3-perm-scalef-div.c",
3231 "src/math/sigmoid-avx512f-rr2-lut16-p3-perm-scalef-nr1fma.c",
3232 "src/math/sigmoid-avx512f-rr2-lut16-p3-perm-scalef-nr1fma1adj.c",
3233 "src/math/sigmoid-avx512f-rr2-lut32-p2-perm2-scalef-div.c",
3234 "src/math/sigmoid-avx512f-rr2-lut32-p2-perm2-scalef-nr1fma.c",
3235 "src/math/sigmoid-avx512f-rr2-lut32-p2-perm2-scalef-nr1fma1adj.c",
Marat Dukhan36173d22020-10-15 17:14:26 -07003236 "src/math/sigmoid-avx512f-rr2-lut64-p2-gather-scalef-div.c",
Marat Dukhan36173d22020-10-15 17:14:26 -07003237 "src/math/sigmoid-avx512f-rr2-lut64-p2-gather-scalef-nr1fma.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003238 "src/math/sigmoid-avx512f-rr2-lut64-p2-gather-scalef-nr1fma1adj.c",
3239 "src/math/sigmoid-avx512f-rr2-p5-scalef-div.c",
3240 "src/math/sigmoid-avx512f-rr2-p5-scalef-nr1fma.c",
3241 "src/math/sigmoid-avx512f-rr2-p5-scalef-nr1fma1adj.c",
Marat Dukhan84000762020-06-29 18:38:43 -07003242 "src/math/sqrt-avx512f-nr1fma.c",
Marat Dukhan84000762020-06-29 18:38:43 -07003243 "src/math/sqrt-avx512f-nr1fma1adj.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003244 "src/math/sqrt-avx512f-nr2fma.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003245]
3246
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07003247AVX512SKX_UKERNELS = [
Marat Dukhan2ffc5e62020-09-06 22:33:38 -07003248 "src/qs8-dwconv/gen/up16x9-minmax-avx512skx-mul32.c",
3249 "src/qs8-dwconv/gen/up32x9-minmax-avx512skx-mul32.c",
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07003250 "src/qs8-gemm/gen/1x16c8-minmax-avx512skx.c",
3251 "src/qs8-gemm/gen/2x16c8-minmax-avx512skx.c",
3252 "src/qs8-gemm/gen/3x16c8-minmax-avx512skx.c",
3253 "src/qs8-gemm/gen/4x16c8-minmax-avx512skx.c",
3254 "src/qs8-igemm/gen/1x16c8-minmax-avx512skx.c",
3255 "src/qs8-igemm/gen/2x16c8-minmax-avx512skx.c",
3256 "src/qs8-igemm/gen/3x16c8-minmax-avx512skx.c",
3257 "src/qs8-igemm/gen/4x16c8-minmax-avx512skx.c",
3258]
3259
Frank Barchardbcedc082020-08-17 18:00:51 -07003260WASM32_ASM_UKERNELS = [
3261 "src/f32-relu/wasm_shr_x1.S",
3262 "src/f32-relu/wasm_shr_x2.S",
3263 "src/f32-relu/wasm_shr_x4.S",
3264]
3265
Marat Dukhan08c4a432019-10-03 09:29:21 -07003266AARCH32_ASM_UKERNELS = [
Marat Dukhan32f93812020-05-17 20:31:21 -07003267 "src/f32-gemm/4x4-aarch32-vfp-ld64.S",
Marat Dukhan3b98f6b2020-05-17 10:09:22 -07003268 "src/f32-gemm/4x4-minmax-aarch32-vfp-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003269 "src/f32-gemm/4x8-minmax-aarch32-neon-cortex-a53.S",
3270 "src/f32-gemm/4x8-minmax-aarch32-neon-cortex-a55.S",
Frank Barchard490febe2020-07-16 18:42:17 -07003271 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-cortex-a7.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003272 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-cortex-a75.S",
Frank Barchard569561d2020-06-17 13:11:12 -07003273 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-ld64.S",
Frank Barchard490febe2020-07-16 18:42:17 -07003274 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-pld-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003275 "src/f32-igemm/4x8-minmax-aarch32-neon-cortex-a53.S",
3276 "src/f32-igemm/4x8-minmax-aarch32-neon-cortex-a55.S",
Frank Barchard490febe2020-07-16 18:42:17 -07003277 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-cortex-a7.S",
3278 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-cortex-a75.S",
3279 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-ld64.S",
3280 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-pld-cortex-a75.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003281]
3282
3283AARCH64_ASM_UKERNELS = [
Frank Barchardbddfbcd2020-04-15 12:32:41 -07003284 "src/f16-gemm/gen-inc/1x8inc-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchard04336c12020-10-22 16:48:55 -07003285 "src/f16-gemm/gen-inc/1x16inc-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchardbddfbcd2020-04-15 12:32:41 -07003286 "src/f16-gemm/gen-inc/4x8inc-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchard04336c12020-10-22 16:48:55 -07003287 "src/f16-gemm/gen-inc/4x16inc-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchardbddfbcd2020-04-15 12:32:41 -07003288 "src/f16-gemm/gen-inc/6x8inc-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchard04336c12020-10-22 16:48:55 -07003289 "src/f16-gemm/gen-inc/6x16inc-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchard3b8e5662020-04-20 12:12:53 -07003290 "src/f16-gemm/gen-inc/8x8inc-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchard04336c12020-10-22 16:48:55 -07003291 "src/f16-gemm/gen/1x8-minmax-aarch64-neonfp16arith-ld64.S",
3292 "src/f16-gemm/gen/1x16-minmax-aarch64-neonfp16arith-ld32.S",
3293 "src/f16-gemm/gen/4x8-minmax-aarch64-neonfp16arith-ld64.S",
3294 "src/f16-gemm/gen/4x16-minmax-aarch64-neonfp16arith-ld32.S",
3295 "src/f16-gemm/gen/6x8-minmax-aarch64-neonfp16arith-ld64.S",
3296 "src/f16-gemm/gen/6x16-minmax-aarch64-neonfp16arith-ld32.S",
3297 "src/f16-gemm/gen/8x8-minmax-aarch64-neonfp16arith-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003298 "src/f32-dwconv/up4x9-minmax-aarch64-neonfma-cortex-a55.S",
3299 "src/f32-dwconv/up4x9-minmax-aarch64-neonfma.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003300 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a53.S",
3301 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a57.S",
3302 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a75.S",
Frank Barchard04336c12020-10-22 16:48:55 -07003303 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-ld64.S",
3304 "src/f32-gemm/gen-inc/1x12inc-minmax-aarch64-neonfma-cortex-a53.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003305 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a53.S",
3306 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a55.S",
3307 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a57.S",
3308 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003309 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-ld64.S",
Frank Barchard04336c12020-10-22 16:48:55 -07003310 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-ld128.S",
3311 "src/f32-gemm/gen-inc/4x12inc-minmax-aarch64-neonfma-cortex-a53.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003312 "src/f32-gemm/gen-inc/5x8inc-minmax-aarch64-neonfma-cortex-a57.S",
3313 "src/f32-gemm/gen-inc/5x8inc-minmax-aarch64-neonfma-cortex-a75.S",
3314 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a53.S",
3315 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a55.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003316 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a57.S",
Frank Barchard04336c12020-10-22 16:48:55 -07003317 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a73.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003318 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003319 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-ld64.S",
Frank Barchard04336c12020-10-22 16:48:55 -07003320 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-ld128.S",
3321 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a53.S",
3322 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a57.S",
3323 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a75.S",
3324 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-ld64.S",
3325 "src/f32-gemm/gen/1x12-minmax-aarch64-neonfma-cortex-a53.S",
3326 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a53.S",
3327 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a55.S",
3328 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a57.S",
3329 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a75.S",
3330 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-ld64.S",
3331 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-ld128.S",
3332 "src/f32-gemm/gen/4x12-minmax-aarch64-neonfma-cortex-a53.S",
3333 "src/f32-gemm/gen/5x8-minmax-aarch64-neonfma-cortex-a57.S",
3334 "src/f32-gemm/gen/5x8-minmax-aarch64-neonfma-cortex-a75.S",
3335 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a53.S",
3336 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a55.S",
3337 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a57.S",
3338 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a73.S",
3339 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a75.S",
3340 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-ld64.S",
3341 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-ld128.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003342 "src/f32-igemm/1x8-minmax-aarch64-neonfma-cortex-a53.S",
Frank Barchard04336c12020-10-22 16:48:55 -07003343 "src/f32-igemm/1x12-minmax-aarch64-neonfma-cortex-a53.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003344 "src/f32-igemm/4x8-minmax-aarch64-neonfma-cortex-a53.S",
3345 "src/f32-igemm/4x8-minmax-aarch64-neonfma-cortex-a55.S",
Frank Barchard04336c12020-10-22 16:48:55 -07003346 "src/f32-igemm/4x12-minmax-aarch64-neonfma-cortex-a53.S",
3347 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a53.S",
3348 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a55.S",
3349 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a73.S",
3350 "src/f32-igemm/gen/1x8-minmax-aarch64-neonfma-cortex-a57.S",
3351 "src/f32-igemm/gen/1x8-minmax-aarch64-neonfma-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003352 "src/f32-igemm/gen/4x8-minmax-aarch64-neonfma-cortex-a57.S",
3353 "src/f32-igemm/gen/4x8-minmax-aarch64-neonfma-cortex-a75.S",
3354 "src/f32-igemm/gen/5x8-minmax-aarch64-neonfma-cortex-a57.S",
3355 "src/f32-igemm/gen/5x8-minmax-aarch64-neonfma-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003356 "src/f32-igemm/gen/6x8-minmax-aarch64-neonfma-cortex-a57.S",
3357 "src/f32-igemm/gen/6x8-minmax-aarch64-neonfma-cortex-a75.S",
Frank Barchardf1fd89e2020-10-06 23:27:11 -07003358 "src/qs8-gemm/1x16c4-aarch64-neondot-ld64.S",
Frank Barchardf2742c42020-10-12 21:16:46 -07003359 "src/qs8-gemm/4x16c4-aarch64-neondot-cortex-a55.S",
Frank Barcharda5237a52020-10-06 16:18:46 -07003360 "src/qs8-gemm/4x16c4-aarch64-neondot-ld64.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003361]
3362
Marat Dukhan1b354632020-03-23 12:50:22 -07003363INTERNAL_MICROKERNEL_HDRS = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003364 "src/xnnpack/argmaxpool.h",
3365 "src/xnnpack/avgpool.h",
3366 "src/xnnpack/clamp.h",
3367 "src/xnnpack/common.h",
3368 "src/xnnpack/conv.h",
Yury Kartynnike7841862020-11-04 18:22:18 -08003369 "src/xnnpack/depthtospace.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003370 "src/xnnpack/dwconv.h",
Frank Barchard04336c12020-10-22 16:48:55 -07003371 "src/xnnpack/fill.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003372 "src/xnnpack/gavgpool.h",
3373 "src/xnnpack/gemm.h",
3374 "src/xnnpack/hswish.h",
Marat Dukhan660fd192020-03-10 04:55:30 -07003375 "src/xnnpack/ibilinear.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003376 "src/xnnpack/igemm.h",
Marat Dukhancfb31342019-12-05 10:42:57 -08003377 "src/xnnpack/intrinsics-polyfill.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003378 "src/xnnpack/lut.h",
3379 "src/xnnpack/math.h",
3380 "src/xnnpack/maxpool.h",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003381 "src/xnnpack/memory.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003382 "src/xnnpack/packx.h",
3383 "src/xnnpack/pad.h",
3384 "src/xnnpack/params.h",
3385 "src/xnnpack/pavgpool.h",
3386 "src/xnnpack/ppmm.h",
3387 "src/xnnpack/prelu.h",
Marat Dukhan97579532019-10-18 16:40:39 -07003388 "src/xnnpack/raddexpminusmax.h",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07003389 "src/xnnpack/raddextexp.h",
Marat Dukhan97579532019-10-18 16:40:39 -07003390 "src/xnnpack/raddstoreexpminusmax.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003391 "src/xnnpack/rmax.h",
3392 "src/xnnpack/scalar-utils.h",
3393 "src/xnnpack/spmm.h",
3394 "src/xnnpack/unpool.h",
3395 "src/xnnpack/vadd.h",
Marat Dukhan1e782c42019-11-21 17:02:40 -08003396 "src/xnnpack/vbinary.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003397 "src/xnnpack/vmulcaddc.h",
Marat Dukhan05ac8e32019-10-21 15:39:33 -07003398 "src/xnnpack/vscale.h",
Marat Dukhan97579532019-10-18 16:40:39 -07003399 "src/xnnpack/vscaleexpminusmax.h",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07003400 "src/xnnpack/vscaleextexp.h",
Marat Dukhan1e782c42019-11-21 17:02:40 -08003401 "src/xnnpack/vunary.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003402 "src/xnnpack/zip.h",
Marat Dukhan1b354632020-03-23 12:50:22 -07003403]
3404
3405INTERNAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003406 "include/xnnpack.h",
3407 "src/xnnpack/allocator.h",
3408 "src/xnnpack/compute.h",
3409 "src/xnnpack/im2col.h",
3410 "src/xnnpack/indirection.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003411 "src/xnnpack/math-stubs.h",
Chao Mei6ddfc602020-05-13 22:29:36 -07003412 "src/xnnpack/memory-planner.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003413 "src/xnnpack/operator.h",
3414 "src/xnnpack/pack.h",
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07003415 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003416 "src/xnnpack/requantization-stubs.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003417 "src/xnnpack/requantization.h",
Marat Dukhan1d75a542020-02-03 12:23:01 -08003418 "src/xnnpack/subgraph.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003419]
3420
Marat Dukhan1b354632020-03-23 12:50:22 -07003421ACCURACY_EVAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003422 "src/xnnpack/math-stubs.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003423]
3424
Marat Dukhan1b354632020-03-23 12:50:22 -07003425MICROKERNEL_BENCHMARK_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003426 "include/xnnpack.h",
Frank Barchard35db7d02020-10-26 13:37:34 -07003427 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003428]
3429
Marat Dukhan1b354632020-03-23 12:50:22 -07003430MICROKERNEL_TEST_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Frank Barchard35db7d02020-10-26 13:37:34 -07003431 "include/xnnpack.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003432 "src/xnnpack/isa-checks.h",
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07003433 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003434 "src/xnnpack/requantization.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003435]
3436
3437OPERATOR_TEST_PARAMS_HDRS = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003438 "src/xnnpack/common.h",
Frank Barchard04336c12020-10-22 16:48:55 -07003439 "src/xnnpack/params.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003440]
3441
3442WEIGHTS_PACK_HDRS = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003443 "src/xnnpack/compute.h",
Frank Barchard04336c12020-10-22 16:48:55 -07003444 "src/xnnpack/operator.h",
3445 "src/xnnpack/pack.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003446]
3447
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07003448LOGGING_COPTS = select({
3449 # No logging in optimized mode
3450 ":optimized_build": ["-DXNN_LOG_LEVEL=0"],
3451 # Full logging in debug mode
3452 ":debug_build": ["-DXNN_LOG_LEVEL=5"],
3453 # Error-only logging in default (fastbuild) mode
3454 "//conditions:default": ["-DXNN_LOG_LEVEL=2"],
3455})
3456
Marat Dukhan3b59de22020-06-03 20:15:19 -07003457LOGGING_SRCS = select({
3458 # No logging in optimized mode
3459 ":optimized_build": [],
3460 "//conditions:default": [
3461 "src/operator-strings.c",
3462 "src/subgraph-strings.c",
3463 ],
3464})
3465
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07003466LOGGING_HDRS = [
3467 "src/xnnpack/log.h",
3468]
3469
Marat Dukhan08c4a432019-10-03 09:29:21 -07003470xnnpack_cc_library(
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003471 name = "tables",
3472 srcs = TABLE_SRCS,
3473 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003474 gcc_copts = xnnpack_gcc_std_copts(),
3475 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003476)
3477
3478xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003479 name = "scalar_ukernels",
3480 srcs = SCALAR_UKERNELS,
3481 hdrs = INTERNAL_HDRS,
3482 aarch32_copts = ["-marm"],
Marat Dukhan10a38082020-04-17 03:58:35 -07003483 gcc_copts = xnnpack_gcc_std_copts(),
3484 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003485 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003486 ":tables",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003487 "@FP16",
3488 "@FXdiv",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003489 "@pthreadpool",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003490 ],
3491)
3492
3493xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003494 name = "scalar_ukernels_test_mode",
3495 srcs = SCALAR_UKERNELS,
3496 hdrs = INTERNAL_HDRS,
3497 aarch32_copts = ["-marm"],
3498 copts = [
3499 "-UNDEBUG",
3500 "-DXNN_TEST_MODE=1",
3501 ],
3502 gcc_copts = xnnpack_gcc_std_copts(),
3503 msvc_copts = xnnpack_msvc_std_copts(),
3504 deps = [
3505 ":tables",
3506 "@FP16",
3507 "@FXdiv",
3508 "@pthreadpool",
3509 ],
3510)
3511
3512xnnpack_cc_library(
Marat Dukhan436ebe62019-12-04 15:10:12 -08003513 name = "wasm_ukernels",
3514 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003515 gcc_copts = xnnpack_gcc_std_copts(),
3516 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan436ebe62019-12-04 15:10:12 -08003517 wasm_srcs = WASM_UKERNELS,
Marat Dukhan1c6e3892020-06-25 23:56:10 -07003518 wasmsimd_srcs = WASM_UKERNELS + WASMSIMD_UKERNELS,
Marat Dukhan436ebe62019-12-04 15:10:12 -08003519 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003520 ":tables",
Marat Dukhan436ebe62019-12-04 15:10:12 -08003521 "@FP16",
3522 "@FXdiv",
3523 "@pthreadpool",
3524 ],
3525)
3526
3527xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003528 name = "wasm_ukernels_test_mode",
3529 hdrs = INTERNAL_HDRS,
3530 copts = [
3531 "-UNDEBUG",
3532 "-DXNN_TEST_MODE=1",
3533 ],
3534 gcc_copts = xnnpack_gcc_std_copts(),
3535 msvc_copts = xnnpack_msvc_std_copts(),
3536 wasm_srcs = WASM_UKERNELS,
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07003537 wasmsimd_srcs = WASM_UKERNELS + WASMSIMD_UKERNELS,
Marat Dukhan33fcf782020-05-24 14:27:15 -07003538 deps = [
3539 ":tables",
3540 "@FP16",
3541 "@FXdiv",
3542 "@pthreadpool",
3543 ],
3544)
3545
3546xnnpack_cc_library(
Marat Dukhan102e9ea2020-12-05 22:35:18 -08003547 name = "psimd_accmath_ukernels",
3548 hdrs = INTERNAL_HDRS,
3549 aarch32_copts = [
3550 "-marm",
3551 "-mfpu=neon",
3552 ],
3553 gcc_copts = xnnpack_gcc_std_copts(),
3554 gcc_x86_copts = ["-msse2"],
3555 msvc_copts = xnnpack_msvc_std_copts(),
3556 optimized_copts = [
3557 "-O3",
3558 ],
3559 psimd_srcs = PSIMD_ACCMATH_UKERNELS,
3560 deps = [
3561 ":tables",
3562 "@FP16",
3563 "@psimd",
3564 "@pthreadpool",
3565 ],
3566)
3567
3568xnnpack_cc_library(
3569 name = "psimd_accmath_ukernels_test_mode",
3570 hdrs = INTERNAL_HDRS,
3571 aarch32_copts = [
3572 "-marm",
3573 "-mfpu=neon",
3574 ],
3575 copts = [
3576 "-UNDEBUG",
3577 "-DXNN_TEST_MODE=1",
3578 ],
3579 gcc_copts = xnnpack_gcc_std_copts(),
3580 gcc_x86_copts = ["-msse2"],
3581 msvc_copts = xnnpack_msvc_std_copts(),
3582 optimized_copts = [
3583 "-O3",
3584 ],
3585 psimd_srcs = PSIMD_ACCMATH_UKERNELS,
3586 deps = [
3587 ":tables",
3588 "@FP16",
3589 "@psimd",
3590 "@pthreadpool",
3591 ],
3592)
3593
3594xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003595 name = "neon_ukernels",
3596 hdrs = INTERNAL_HDRS,
3597 aarch32_copts = [
3598 "-marm",
Marat Dukhan8853b822020-05-07 12:19:01 -07003599 "-march=armv7-a",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003600 "-mfpu=neon",
3601 ],
3602 aarch32_srcs = NEON_UKERNELS,
3603 aarch64_srcs = NEON_UKERNELS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003604 gcc_copts = xnnpack_gcc_std_copts(),
3605 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08003606 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003607 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003608 "@FP16",
3609 "@pthreadpool",
3610 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003611)
3612
3613xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003614 name = "neon_ukernels_test_mode",
3615 hdrs = INTERNAL_HDRS,
3616 aarch32_copts = [
3617 "-marm",
3618 "-march=armv7-a",
3619 "-mfpu=neon",
3620 ],
3621 aarch32_srcs = NEON_UKERNELS,
3622 aarch64_srcs = NEON_UKERNELS,
3623 copts = [
3624 "-UNDEBUG",
3625 "-DXNN_TEST_MODE=1",
3626 ],
3627 gcc_copts = xnnpack_gcc_std_copts(),
3628 msvc_copts = xnnpack_msvc_std_copts(),
3629 deps = [
3630 ":tables",
3631 "@FP16",
3632 "@pthreadpool",
3633 ],
3634)
3635
3636xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003637 name = "neonfma_ukernels",
3638 hdrs = INTERNAL_HDRS,
3639 aarch32_copts = [
3640 "-marm",
Marat Dukhan8853b822020-05-07 12:19:01 -07003641 "-march=armv7-a",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003642 "-mfpu=neon-vfpv4",
3643 ],
3644 aarch32_srcs = NEONFMA_UKERNELS,
3645 aarch64_srcs = NEONFMA_UKERNELS + AARCH64_NEONFMA_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07003646 apple_aarch32_copts = [
3647 "-mcpu=swift",
3648 "-mtune=generic",
3649 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07003650 gcc_copts = xnnpack_gcc_std_copts(),
3651 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08003652 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003653 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003654 "@FP16",
3655 "@pthreadpool",
3656 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003657)
3658
3659xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003660 name = "neonfma_ukernels_test_mode",
3661 hdrs = INTERNAL_HDRS,
3662 aarch32_copts = [
3663 "-marm",
3664 "-march=armv7-a",
3665 "-mfpu=neon-vfpv4",
3666 ],
3667 aarch32_srcs = NEONFMA_UKERNELS,
3668 aarch64_srcs = NEONFMA_UKERNELS + AARCH64_NEONFMA_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07003669 apple_aarch32_copts = [
3670 "-mcpu=swift",
3671 "-mtune=generic",
3672 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07003673 copts = [
3674 "-UNDEBUG",
3675 "-DXNN_TEST_MODE=1",
3676 ],
3677 gcc_copts = xnnpack_gcc_std_copts(),
3678 msvc_copts = xnnpack_msvc_std_copts(),
3679 deps = [
3680 ":tables",
3681 "@FP16",
3682 "@pthreadpool",
3683 ],
3684)
3685
3686xnnpack_cc_library(
Marat Dukhan8853b822020-05-07 12:19:01 -07003687 name = "neonv8_ukernels",
3688 hdrs = INTERNAL_HDRS,
3689 aarch32_copts = [
3690 "-marm",
3691 "-march=armv8-a",
3692 "-mfpu=neon-fp-armv8",
3693 ],
3694 aarch32_srcs = NEONV8_UKERNELS,
3695 aarch64_srcs = NEONV8_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07003696 apple_aarch32_copts = [
3697 "-mcpu=cyclone",
3698 "-mtune=generic",
3699 ],
Marat Dukhan8853b822020-05-07 12:19:01 -07003700 gcc_copts = xnnpack_gcc_std_copts(),
3701 msvc_copts = xnnpack_msvc_std_copts(),
3702 deps = [
3703 ":tables",
3704 "@FP16",
3705 "@pthreadpool",
3706 ],
3707)
3708
3709xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003710 name = "neonv8_ukernels_test_mode",
3711 hdrs = INTERNAL_HDRS,
3712 aarch32_copts = [
3713 "-marm",
3714 "-march=armv8-a",
3715 "-mfpu=neon-fp-armv8",
3716 ],
3717 aarch32_srcs = NEONV8_UKERNELS,
3718 aarch64_srcs = NEONV8_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07003719 apple_aarch32_copts = [
3720 "-mcpu=cyclone",
3721 "-mtune=generic",
3722 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07003723 copts = [
3724 "-UNDEBUG",
3725 "-DXNN_TEST_MODE=1",
3726 ],
3727 gcc_copts = xnnpack_gcc_std_copts(),
3728 msvc_copts = xnnpack_msvc_std_copts(),
3729 deps = [
3730 ":tables",
3731 "@FP16",
3732 "@pthreadpool",
3733 ],
3734)
3735
3736xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003737 name = "neonfp16arith_ukernels",
3738 hdrs = INTERNAL_HDRS,
3739 aarch64_copts = ["-march=armv8.2-a+fp16"],
3740 aarch64_srcs = AARCH64_NEONFP16ARITH_UKERNELS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003741 gcc_copts = xnnpack_gcc_std_copts(),
3742 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08003743 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003744 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003745 "@FP16",
3746 "@pthreadpool",
3747 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003748)
3749
3750xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003751 name = "neonfp16arith_ukernels_test_mode",
3752 hdrs = INTERNAL_HDRS,
3753 aarch64_copts = ["-march=armv8.2-a+fp16"],
3754 aarch64_srcs = AARCH64_NEONFP16ARITH_UKERNELS,
3755 copts = [
3756 "-UNDEBUG",
3757 "-DXNN_TEST_MODE=1",
3758 ],
3759 gcc_copts = xnnpack_gcc_std_copts(),
3760 msvc_copts = xnnpack_msvc_std_copts(),
3761 deps = [
3762 ":tables",
3763 "@FP16",
3764 "@pthreadpool",
3765 ],
3766)
3767
3768xnnpack_cc_library(
Benoit Jacoba9644732020-08-13 12:48:55 -07003769 name = "neondot_ukernels",
3770 hdrs = INTERNAL_HDRS,
Marat Dukhan799ac752020-08-13 16:08:03 -07003771 aarch32_copts = [
3772 "-marm",
3773 "-march=armv8.2-a+dotprod",
3774 "-mfpu=neon-fp-armv8",
3775 ],
Benoit Jacoba9644732020-08-13 12:48:55 -07003776 aarch32_srcs = NEONDOT_UKERNELS,
3777 aarch64_copts = ["-march=armv8.2-a+dotprod"],
3778 aarch64_srcs = NEONDOT_UKERNELS,
3779 gcc_copts = xnnpack_gcc_std_copts(),
3780 msvc_copts = xnnpack_msvc_std_copts(),
3781 deps = [
3782 ":tables",
3783 "@FP16",
3784 "@pthreadpool",
3785 ],
3786)
3787
3788xnnpack_cc_library(
3789 name = "neondot_ukernels_test_mode",
3790 hdrs = INTERNAL_HDRS,
Marat Dukhan799ac752020-08-13 16:08:03 -07003791 aarch32_copts = [
3792 "-marm",
3793 "-march=armv8.2-a+dotprod",
3794 "-mfpu=neon-fp-armv8",
3795 ],
Benoit Jacoba9644732020-08-13 12:48:55 -07003796 aarch32_srcs = NEONDOT_UKERNELS,
3797 aarch64_copts = ["-march=armv8.2-a+dotprod"],
3798 aarch64_srcs = NEONDOT_UKERNELS,
3799 copts = [
3800 "-UNDEBUG",
3801 "-DXNN_TEST_MODE=1",
3802 ],
3803 gcc_copts = xnnpack_gcc_std_copts(),
3804 msvc_copts = xnnpack_msvc_std_copts(),
3805 deps = [
3806 ":tables",
3807 "@FP16",
3808 "@pthreadpool",
3809 ],
3810)
3811
3812xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003813 name = "sse2_ukernels",
3814 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003815 gcc_copts = xnnpack_gcc_std_copts(),
3816 gcc_x86_copts = ["-msse2"],
3817 msvc_copts = xnnpack_msvc_std_copts(),
3818 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003819 x86_srcs = SSE_UKERNELS + SSE2_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08003820 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003821 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003822 "@FP16",
3823 "@pthreadpool",
3824 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003825)
3826
3827xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003828 name = "sse2_ukernels_test_mode",
3829 hdrs = INTERNAL_HDRS,
3830 copts = [
3831 "-UNDEBUG",
3832 "-DXNN_TEST_MODE=1",
3833 ],
3834 gcc_copts = xnnpack_gcc_std_copts(),
3835 gcc_x86_copts = ["-msse2"],
3836 msvc_copts = xnnpack_msvc_std_copts(),
3837 msvc_x86_32_copts = ["/arch:SSE2"],
3838 x86_srcs = SSE_UKERNELS + SSE2_UKERNELS,
3839 deps = [
3840 ":tables",
3841 "@FP16",
3842 "@pthreadpool",
3843 ],
3844)
3845
3846xnnpack_cc_library(
Marat Dukhanfe7acb62020-03-09 19:30:05 -07003847 name = "ssse3_ukernels",
3848 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003849 gcc_copts = xnnpack_gcc_std_copts(),
3850 gcc_x86_copts = ["-mssse3"],
3851 msvc_copts = xnnpack_msvc_std_copts(),
3852 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhanfe7acb62020-03-09 19:30:05 -07003853 x86_srcs = SSSE3_UKERNELS,
3854 deps = [
3855 ":tables",
3856 "@FP16",
3857 "@pthreadpool",
3858 ],
3859)
3860
3861xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003862 name = "ssse3_ukernels_test_mode",
3863 hdrs = INTERNAL_HDRS,
3864 copts = [
3865 "-UNDEBUG",
3866 "-DXNN_TEST_MODE=1",
3867 ],
3868 gcc_copts = xnnpack_gcc_std_copts(),
3869 gcc_x86_copts = ["-mssse3"],
3870 msvc_copts = xnnpack_msvc_std_copts(),
3871 msvc_x86_32_copts = ["/arch:SSE2"],
3872 x86_srcs = SSSE3_UKERNELS,
3873 deps = [
3874 ":tables",
3875 "@FP16",
3876 "@pthreadpool",
3877 ],
3878)
3879
3880xnnpack_cc_library(
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08003881 name = "sse41_ukernels",
3882 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003883 gcc_copts = xnnpack_gcc_std_copts(),
3884 gcc_x86_copts = ["-msse4.1"],
3885 msvc_copts = xnnpack_msvc_std_copts(),
3886 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08003887 x86_srcs = SSE41_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08003888 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003889 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003890 "@FP16",
3891 "@pthreadpool",
3892 ],
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08003893)
3894
3895xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003896 name = "sse41_ukernels_test_mode",
3897 hdrs = INTERNAL_HDRS,
3898 copts = [
3899 "-UNDEBUG",
3900 "-DXNN_TEST_MODE=1",
3901 ],
3902 gcc_copts = xnnpack_gcc_std_copts(),
3903 gcc_x86_copts = ["-msse4.1"],
3904 msvc_copts = xnnpack_msvc_std_copts(),
3905 msvc_x86_32_copts = ["/arch:SSE2"],
3906 x86_srcs = SSE41_UKERNELS,
3907 deps = [
3908 ":tables",
3909 "@FP16",
3910 "@pthreadpool",
3911 ],
3912)
3913
3914xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003915 name = "avx_ukernels",
3916 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003917 gcc_copts = xnnpack_gcc_std_copts(),
3918 gcc_x86_copts = ["-mavx"],
3919 msvc_copts = xnnpack_msvc_std_copts(),
3920 msvc_x86_32_copts = ["/arch:AVX"],
3921 msvc_x86_64_copts = ["/arch:AVX"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003922 x86_srcs = AVX_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08003923 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003924 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003925 "@FP16",
3926 "@pthreadpool",
3927 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003928)
3929
3930xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003931 name = "avx_ukernels_test_mode",
3932 hdrs = INTERNAL_HDRS,
3933 copts = [
3934 "-UNDEBUG",
3935 "-DXNN_TEST_MODE=1",
3936 ],
3937 gcc_copts = xnnpack_gcc_std_copts(),
3938 gcc_x86_copts = ["-mavx"],
3939 msvc_copts = xnnpack_msvc_std_copts(),
3940 msvc_x86_32_copts = ["/arch:AVX"],
3941 msvc_x86_64_copts = ["/arch:AVX"],
3942 x86_srcs = AVX_UKERNELS,
3943 deps = [
3944 ":tables",
3945 "@FP16",
3946 "@pthreadpool",
3947 ],
3948)
3949
3950xnnpack_cc_library(
Marat Dukhan1566fee2020-08-02 21:55:41 -07003951 name = "xop_ukernels",
3952 hdrs = INTERNAL_HDRS,
3953 gcc_copts = xnnpack_gcc_std_copts(),
3954 gcc_x86_copts = ["-mxop"],
3955 msvc_copts = xnnpack_msvc_std_copts(),
3956 msvc_x86_32_copts = ["/arch:AVX"],
3957 msvc_x86_64_copts = ["/arch:AVX"],
3958 x86_srcs = XOP_UKERNELS,
3959 deps = [
3960 ":tables",
3961 "@FP16",
3962 "@pthreadpool",
3963 ],
3964)
3965
3966xnnpack_cc_library(
3967 name = "xop_ukernels_test_mode",
3968 hdrs = INTERNAL_HDRS,
3969 copts = [
3970 "-UNDEBUG",
3971 "-DXNN_TEST_MODE=1",
3972 ],
3973 gcc_copts = xnnpack_gcc_std_copts(),
3974 gcc_x86_copts = ["-mxop"],
3975 msvc_copts = xnnpack_msvc_std_copts(),
3976 msvc_x86_32_copts = ["/arch:AVX"],
3977 msvc_x86_64_copts = ["/arch:AVX"],
3978 x86_srcs = XOP_UKERNELS,
3979 deps = [
3980 ":tables",
3981 "@FP16",
3982 "@pthreadpool",
3983 ],
3984)
3985
3986xnnpack_cc_library(
Marat Dukhanfda12b82019-11-21 12:27:59 -08003987 name = "fma3_ukernels",
3988 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003989 gcc_copts = xnnpack_gcc_std_copts(),
3990 gcc_x86_copts = ["-mfma"],
3991 msvc_copts = xnnpack_msvc_std_copts(),
3992 msvc_x86_32_copts = ["/arch:AVX"],
3993 msvc_x86_64_copts = ["/arch:AVX"],
Marat Dukhanfda12b82019-11-21 12:27:59 -08003994 x86_srcs = FMA3_UKERNELS,
3995 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003996 ":tables",
Marat Dukhanfda12b82019-11-21 12:27:59 -08003997 "@FP16",
3998 "@pthreadpool",
3999 ],
4000)
4001
4002xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07004003 name = "fma3_ukernels_test_mode",
4004 hdrs = INTERNAL_HDRS,
4005 copts = [
4006 "-UNDEBUG",
4007 "-DXNN_TEST_MODE=1",
4008 ],
4009 gcc_copts = xnnpack_gcc_std_copts(),
4010 gcc_x86_copts = ["-mfma"],
4011 msvc_copts = xnnpack_msvc_std_copts(),
4012 msvc_x86_32_copts = ["/arch:AVX"],
4013 msvc_x86_64_copts = ["/arch:AVX"],
4014 x86_srcs = FMA3_UKERNELS,
4015 deps = [
4016 ":tables",
4017 "@FP16",
4018 "@pthreadpool",
4019 ],
4020)
4021
4022xnnpack_cc_library(
Marat Dukhan6adff4e2019-10-14 18:32:07 -07004023 name = "avx2_ukernels",
4024 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004025 gcc_copts = xnnpack_gcc_std_copts(),
4026 gcc_x86_copts = [
Marat Dukhan6adff4e2019-10-14 18:32:07 -07004027 "-mfma",
4028 "-mavx2",
4029 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07004030 msvc_copts = xnnpack_msvc_std_copts(),
4031 msvc_x86_32_copts = ["/arch:AVX2"],
4032 msvc_x86_64_copts = ["/arch:AVX2"],
Marat Dukhan6adff4e2019-10-14 18:32:07 -07004033 x86_srcs = AVX2_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08004034 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08004035 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08004036 "@FP16",
4037 "@pthreadpool",
4038 ],
Marat Dukhan6adff4e2019-10-14 18:32:07 -07004039)
4040
4041xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07004042 name = "avx2_ukernels_test_mode",
4043 hdrs = INTERNAL_HDRS,
4044 copts = [
4045 "-UNDEBUG",
4046 "-DXNN_TEST_MODE=1",
4047 ],
4048 gcc_copts = xnnpack_gcc_std_copts(),
4049 gcc_x86_copts = [
4050 "-mfma",
4051 "-mavx2",
4052 ],
4053 msvc_copts = xnnpack_msvc_std_copts(),
4054 msvc_x86_32_copts = ["/arch:AVX2"],
4055 msvc_x86_64_copts = ["/arch:AVX2"],
4056 x86_srcs = AVX2_UKERNELS,
4057 deps = [
4058 ":tables",
4059 "@FP16",
4060 "@pthreadpool",
4061 ],
4062)
4063
4064xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004065 name = "avx512f_ukernels",
4066 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004067 gcc_copts = xnnpack_gcc_std_copts(),
4068 gcc_x86_copts = ["-mavx512f"],
4069 mingw_copts = ["-fno-asynchronous-unwind-tables"],
4070 msvc_copts = xnnpack_msvc_std_copts(),
4071 msvc_x86_32_copts = ["/arch:AVX512"],
4072 msvc_x86_64_copts = ["/arch:AVX512"],
4073 msys_copts = ["-fno-asynchronous-unwind-tables"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004074 x86_srcs = AVX512F_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08004075 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08004076 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08004077 "@FP16",
4078 "@pthreadpool",
4079 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004080)
4081
4082xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07004083 name = "avx512f_ukernels_test_mode",
4084 hdrs = INTERNAL_HDRS,
4085 copts = [
4086 "-UNDEBUG",
4087 "-DXNN_TEST_MODE=1",
4088 ],
4089 gcc_copts = xnnpack_gcc_std_copts(),
4090 gcc_x86_copts = ["-mavx512f"],
4091 mingw_copts = ["-fno-asynchronous-unwind-tables"],
4092 msvc_copts = xnnpack_msvc_std_copts(),
4093 msvc_x86_32_copts = ["/arch:AVX512"],
4094 msvc_x86_64_copts = ["/arch:AVX512"],
4095 msys_copts = ["-fno-asynchronous-unwind-tables"],
4096 x86_srcs = AVX512F_UKERNELS,
4097 deps = [
4098 ":tables",
4099 "@FP16",
4100 "@pthreadpool",
4101 ],
4102)
4103
4104xnnpack_cc_library(
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07004105 name = "avx512skx_ukernels",
4106 hdrs = INTERNAL_HDRS,
4107 gcc_copts = xnnpack_gcc_std_copts(),
4108 gcc_x86_copts = [
4109 "-mavx512f",
4110 "-mavx512cd",
4111 "-mavx512bw",
4112 "-mavx512dq",
4113 "-mavx512vl",
4114 ],
4115 mingw_copts = ["-fno-asynchronous-unwind-tables"],
4116 msvc_copts = xnnpack_msvc_std_copts(),
4117 msvc_x86_32_copts = ["/arch:AVX512"],
4118 msvc_x86_64_copts = ["/arch:AVX512"],
4119 msys_copts = ["-fno-asynchronous-unwind-tables"],
4120 x86_srcs = AVX512SKX_UKERNELS,
4121 deps = [
4122 ":tables",
4123 "@FP16",
4124 "@pthreadpool",
4125 ],
4126)
4127
4128xnnpack_cc_library(
4129 name = "avx512skx_ukernels_test_mode",
4130 hdrs = INTERNAL_HDRS,
4131 copts = [
4132 "-UNDEBUG",
4133 "-DXNN_TEST_MODE=1",
4134 ],
4135 gcc_copts = xnnpack_gcc_std_copts(),
4136 gcc_x86_copts = [
4137 "-mavx512f",
4138 "-mavx512cd",
4139 "-mavx512bw",
4140 "-mavx512dq",
4141 "-mavx512vl",
4142 ],
4143 mingw_copts = ["-fno-asynchronous-unwind-tables"],
4144 msvc_copts = xnnpack_msvc_std_copts(),
4145 msvc_x86_32_copts = ["/arch:AVX512"],
4146 msvc_x86_64_copts = ["/arch:AVX512"],
4147 msys_copts = ["-fno-asynchronous-unwind-tables"],
4148 x86_srcs = AVX512SKX_UKERNELS,
4149 deps = [
4150 ":tables",
4151 "@FP16",
4152 "@pthreadpool",
4153 ],
4154)
4155
4156xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004157 name = "asm_ukernels",
4158 hdrs = ["src/xnnpack/assembly.h"],
4159 aarch32_srcs = AARCH32_ASM_UKERNELS,
Frank Barchard31bb45b2020-10-06 00:26:33 -07004160 aarch64_copts = ["-march=armv8.2-a+fp16+dotprod"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004161 aarch64_srcs = AARCH64_ASM_UKERNELS,
Frank Barchardbcedc082020-08-17 18:00:51 -07004162 wasm_srcs = WASM32_ASM_UKERNELS,
4163 wasmsimd_srcs = WASM32_ASM_UKERNELS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07004164)
4165
Marat Dukhan3b59de22020-06-03 20:15:19 -07004166xnnpack_cc_library(
4167 name = "logging_utils",
4168 srcs = LOGGING_SRCS,
4169 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
4170 copts = LOGGING_COPTS + [
4171 "-Isrc",
4172 "-Iinclude",
4173 ] + select({
4174 ":debug_build": [],
4175 "//conditions:default": xnnpack_min_size_copts(),
4176 }),
4177 gcc_copts = xnnpack_gcc_std_copts(),
4178 msvc_copts = xnnpack_msvc_std_copts(),
4179 visibility = xnnpack_visibility(),
4180 deps = [
4181 "@FP16",
4182 "@clog",
4183 "@pthreadpool",
4184 ],
4185)
4186
Marat Dukhan08c4a432019-10-03 09:29:21 -07004187xnnpack_aggregate_library(
4188 name = "ukernels",
4189 aarch32_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07004190 ":neon_ukernels",
4191 ":neonfma_ukernels",
Marat Dukhan8853b822020-05-07 12:19:01 -07004192 ":neonv8_ukernels",
Benoit Jacoba9644732020-08-13 12:48:55 -07004193 ":neondot_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004194 ":asm_ukernels",
4195 ],
4196 aarch64_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07004197 ":neon_ukernels",
4198 ":neonfma_ukernels",
Marat Dukhan8853b822020-05-07 12:19:01 -07004199 ":neonv8_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004200 ":neonfp16arith_ukernels",
Benoit Jacoba9644732020-08-13 12:48:55 -07004201 ":neondot_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004202 ":asm_ukernels",
4203 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07004204 generic_deps = [
4205 ":scalar_ukernels",
4206 ],
Marat Dukhan102e9ea2020-12-05 22:35:18 -08004207 psimd_deps = [
4208 ":psimd_accmath_ukernels",
4209 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07004210 wasm_deps = [
4211 ":wasm_ukernels",
Frank Barchardbcedc082020-08-17 18:00:51 -07004212 ":asm_ukernels",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004213 ],
4214 wasmsimd_deps = [
4215 ":wasm_ukernels",
Frank Barchardbcedc082020-08-17 18:00:51 -07004216 ":asm_ukernels",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004217 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004218 x86_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07004219 ":sse2_ukernels",
Marat Dukhanfe7acb62020-03-09 19:30:05 -07004220 ":ssse3_ukernels",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08004221 ":sse41_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004222 ":avx_ukernels",
Marat Dukhan1566fee2020-08-02 21:55:41 -07004223 ":xop_ukernels",
Marat Dukhanfda12b82019-11-21 12:27:59 -08004224 ":fma3_ukernels",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07004225 ":avx2_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004226 ":avx512f_ukernels",
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07004227 ":avx512skx_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004228 ],
4229)
4230
Marat Dukhan33fcf782020-05-24 14:27:15 -07004231xnnpack_aggregate_library(
4232 name = "ukernels_test_mode",
4233 aarch32_deps = [
4234 ":neon_ukernels_test_mode",
4235 ":neonfma_ukernels_test_mode",
4236 ":neonv8_ukernels_test_mode",
Benoit Jacoba9644732020-08-13 12:48:55 -07004237 ":neondot_ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004238 ":asm_ukernels",
4239 ],
4240 aarch64_deps = [
4241 ":neon_ukernels_test_mode",
4242 ":neonfma_ukernels_test_mode",
4243 ":neonv8_ukernels_test_mode",
4244 ":neonfp16arith_ukernels_test_mode",
Benoit Jacoba9644732020-08-13 12:48:55 -07004245 ":neondot_ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004246 ":asm_ukernels",
4247 ],
4248 generic_deps = [
4249 ":scalar_ukernels_test_mode",
4250 ],
Marat Dukhan102e9ea2020-12-05 22:35:18 -08004251 psimd_deps = [
4252 ":psimd_accmath_ukernels_test_mode",
4253 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07004254 wasm_deps = [
4255 ":wasm_ukernels_test_mode",
Frank Barchardbcedc082020-08-17 18:00:51 -07004256 ":asm_ukernels",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004257 ],
4258 wasmsimd_deps = [
4259 ":wasm_ukernels_test_mode",
Frank Barchardbcedc082020-08-17 18:00:51 -07004260 ":asm_ukernels",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004261 ],
4262 x86_deps = [
4263 ":sse2_ukernels_test_mode",
4264 ":ssse3_ukernels_test_mode",
4265 ":sse41_ukernels_test_mode",
4266 ":avx_ukernels_test_mode",
Marat Dukhan1566fee2020-08-02 21:55:41 -07004267 ":xop_ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004268 ":fma3_ukernels_test_mode",
4269 ":avx2_ukernels_test_mode",
4270 ":avx512f_ukernels_test_mode",
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07004271 ":avx512skx_ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004272 ],
4273)
4274
Marat Dukhan86df3842020-06-17 14:03:31 -07004275# This target must be compatible with Android NDK r17 and compile with Clang 6.
4276# It is similar to :ukernels target, but excludes microkernels which require
4277# NEON FP16 arithmetics, NEON Dot Product, and newer intrinsics.
Marat Dukhanae046f52020-06-15 13:16:14 -07004278xnnpack_aggregate_library(
Marat Dukhan86df3842020-06-17 14:03:31 -07004279 name = "ukernels_max_compatibility",
Marat Dukhanae046f52020-06-15 13:16:14 -07004280 aarch32_deps = [
4281 ":neon_ukernels",
4282 ":neonfma_ukernels",
4283 ":neonv8_ukernels",
4284 ":asm_ukernels",
4285 ],
4286 aarch64_deps = [
4287 ":neon_ukernels",
4288 ":neonfma_ukernels",
4289 ":neonv8_ukernels",
4290 ":asm_ukernels",
4291 ],
4292 generic_deps = [
4293 ":scalar_ukernels",
4294 ],
Marat Dukhan102e9ea2020-12-05 22:35:18 -08004295 psimd_deps = [
4296 ":psimd_accmath_ukernels",
4297 ],
Marat Dukhanae046f52020-06-15 13:16:14 -07004298 wasm_deps = [
4299 ":wasm_ukernels",
4300 ],
4301 wasmsimd_deps = [
4302 ":wasm_ukernels",
4303 ],
4304 x86_deps = [
4305 ":sse2_ukernels",
4306 ":ssse3_ukernels",
4307 ":sse41_ukernels",
4308 ":avx_ukernels",
Marat Dukhan1566fee2020-08-02 21:55:41 -07004309 ":xop_ukernels",
Marat Dukhanae046f52020-06-15 13:16:14 -07004310 ":fma3_ukernels",
4311 ":avx2_ukernels",
4312 ":avx512f_ukernels",
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07004313 ":avx512skx_ukernels",
Marat Dukhanae046f52020-06-15 13:16:14 -07004314 ],
4315)
4316
Marat Dukhan08c4a432019-10-03 09:29:21 -07004317xnnpack_cc_library(
4318 name = "im2col",
4319 srcs = ["src/im2col.c"],
4320 hdrs = [
4321 "src/xnnpack/common.h",
4322 "src/xnnpack/im2col.h",
4323 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07004324 gcc_copts = xnnpack_gcc_std_copts(),
4325 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004326)
4327
4328xnnpack_cc_library(
4329 name = "indirection",
4330 srcs = ["src/indirection.c"],
4331 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004332 gcc_copts = xnnpack_gcc_std_copts(),
4333 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004334 deps = [
4335 "@FP16",
4336 "@FXdiv",
4337 "@pthreadpool",
4338 ],
4339)
4340
4341xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07004342 name = "indirection_test_mode",
4343 srcs = ["src/indirection.c"],
4344 hdrs = INTERNAL_HDRS,
4345 copts = [
4346 "-UNDEBUG",
4347 "-DXNN_TEST_MODE=1",
4348 ],
4349 gcc_copts = xnnpack_gcc_std_copts(),
4350 msvc_copts = xnnpack_msvc_std_copts(),
4351 deps = [
4352 "@FP16",
4353 "@FXdiv",
4354 "@pthreadpool",
4355 ],
4356)
4357
4358xnnpack_cc_library(
Marat Dukhanab582382020-07-06 13:32:08 -07004359 name = "packing",
4360 srcs = ["src/packing.c"],
4361 hdrs = INTERNAL_HDRS,
4362 gcc_copts = xnnpack_gcc_std_copts(),
4363 msvc_copts = xnnpack_msvc_std_copts(),
4364 deps = [
4365 "@FP16",
4366 "@FXdiv",
4367 "@pthreadpool",
4368 ],
4369)
4370
4371xnnpack_cc_library(
4372 name = "packing_test_mode",
4373 srcs = ["src/packing.c"],
4374 hdrs = INTERNAL_HDRS,
4375 copts = [
4376 "-UNDEBUG",
4377 "-DXNN_TEST_MODE=1",
4378 ],
4379 gcc_copts = xnnpack_gcc_std_copts(),
4380 msvc_copts = xnnpack_msvc_std_copts(),
4381 deps = [
4382 "@FP16",
4383 "@FXdiv",
4384 "@pthreadpool",
4385 ],
4386)
4387
4388xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004389 name = "operator_run",
4390 srcs = ["src/operator-run.c"],
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07004391 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004392 copts = LOGGING_COPTS + select({
Marat Dukhan05702cf2020-03-26 15:41:33 -07004393 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4394 "//conditions:default": [],
4395 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07004396 gcc_copts = xnnpack_gcc_std_copts(),
4397 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004398 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07004399 ":logging_utils",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004400 "@FP16",
4401 "@FXdiv",
4402 "@clog",
4403 "@pthreadpool",
4404 ],
4405)
4406
Chao Mei6ddfc602020-05-13 22:29:36 -07004407xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07004408 name = "operator_run_test_mode",
4409 srcs = ["src/operator-run.c"],
4410 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
4411 copts = LOGGING_COPTS + [
4412 "-UNDEBUG",
4413 "-DXNN_TEST_MODE=1",
4414 ] + select({
4415 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4416 "//conditions:default": [],
4417 }),
4418 gcc_copts = xnnpack_gcc_std_copts(),
4419 msvc_copts = xnnpack_msvc_std_copts(),
4420 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07004421 ":logging_utils",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004422 "@FP16",
4423 "@FXdiv",
4424 "@clog",
4425 "@pthreadpool",
4426 ],
4427)
4428
4429xnnpack_cc_library(
Chao Mei6ddfc602020-05-13 22:29:36 -07004430 name = "memory_planner",
4431 srcs = ["src/memory-planner.c"],
4432 hdrs = INTERNAL_HDRS,
4433 defines = select({
4434 ":xnn_enable_memopt_explicit_true": ["XNN_ENABLE_MEMOPT=1"],
4435 ":xnn_enable_memopt_explicit_false": ["XNN_ENABLE_MEMOPT=0"],
4436 "//conditions:default": ["XNN_ENABLE_MEMOPT=1"],
4437 }),
4438 gcc_copts = xnnpack_gcc_std_copts(),
4439 msvc_copts = xnnpack_msvc_std_copts(),
4440 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07004441 ":logging_utils",
Chao Mei6ddfc602020-05-13 22:29:36 -07004442 "@pthreadpool",
4443 ],
4444)
4445
Marat Dukhan33fcf782020-05-24 14:27:15 -07004446xnnpack_cc_library(
4447 name = "memory_planner_test_mode",
4448 srcs = ["src/memory-planner.c"],
4449 hdrs = INTERNAL_HDRS,
4450 copts = [
4451 "-UNDEBUG",
4452 "-DXNN_TEST_MODE=1",
4453 ],
4454 defines = select({
4455 ":xnn_enable_memopt_explicit_true": ["XNN_ENABLE_MEMOPT=1"],
4456 ":xnn_enable_memopt_explicit_false": ["XNN_ENABLE_MEMOPT=0"],
4457 "//conditions:default": ["XNN_ENABLE_MEMOPT=1"],
4458 }),
4459 gcc_copts = xnnpack_gcc_std_copts(),
4460 msvc_copts = xnnpack_msvc_std_copts(),
4461 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07004462 ":logging_utils",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004463 "@pthreadpool",
4464 ],
4465)
4466
Marat Dukhan08c4a432019-10-03 09:29:21 -07004467cc_library(
4468 name = "enable_assembly",
4469 defines = select({
4470 ":xnn_enable_assembly_explicit_true": ["XNN_ENABLE_ASSEMBLY=1"],
4471 ":xnn_enable_assembly_explicit_false": ["XNN_ENABLE_ASSEMBLY=0"],
Frank Barchard810171d2019-10-10 10:34:51 -07004472 "//conditions:default": ["XNN_ENABLE_ASSEMBLY=1"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004473 }),
4474)
4475
Marat Dukhan9de90e02020-06-18 16:04:12 -07004476cc_library(
4477 name = "enable_sparse",
4478 defines = select({
4479 ":xnn_enable_sparse_explicit_true": ["XNN_ENABLE_SPARSE=1"],
4480 ":xnn_enable_sparse_explicit_false": ["XNN_ENABLE_SPARSE=0"],
Marat Dukhan8b283aa2020-06-18 16:48:57 -07004481 "//conditions:default": ["XNN_ENABLE_SPARSE=0"],
Marat Dukhan9de90e02020-06-18 16:04:12 -07004482 }),
4483)
4484
Marat Dukhancf056b22019-10-07 10:26:29 -07004485xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004486 name = "operators",
4487 srcs = OPERATOR_SRCS + [
Marat Dukhan04f03be2019-11-19 12:36:47 -08004488 "src/memory.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004489 "src/operator-delete.c",
Marat Dukhancf056b22019-10-07 10:26:29 -07004490 ],
4491 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004492 copts = LOGGING_COPTS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07004493 "-Isrc",
4494 "-Iinclude",
4495 ] + select({
4496 ":debug_build": [],
4497 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07004498 }) + select({
4499 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4500 "//conditions:default": [],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004501 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07004502 gcc_copts = xnnpack_gcc_std_copts(),
4503 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004504 deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07004505 ":indirection",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004506 ":logging_utils",
Marat Dukhanab582382020-07-06 13:32:08 -07004507 ":packing",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004508 "@FP16",
4509 "@FXdiv",
4510 "@clog",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004511 "@pthreadpool",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004512 ],
4513)
4514
Marat Dukhan10a38082020-04-17 03:58:35 -07004515xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07004516 name = "operators_test_mode",
4517 srcs = OPERATOR_SRCS + [
4518 "src/memory.c",
4519 "src/operator-delete.c",
4520 ],
4521 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
4522 copts = LOGGING_COPTS + [
4523 "-Isrc",
4524 "-Iinclude",
4525 "-UNDEBUG",
4526 "-DXNN_TEST_MODE=1",
4527 ] + select({
4528 ":debug_build": [],
4529 "//conditions:default": xnnpack_min_size_copts(),
4530 }) + select({
4531 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4532 "//conditions:default": [],
4533 }),
4534 gcc_copts = xnnpack_gcc_std_copts(),
4535 msvc_copts = xnnpack_msvc_std_copts(),
4536 deps = [
4537 ":indirection_test_mode",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004538 ":logging_utils",
Marat Dukhanab582382020-07-06 13:32:08 -07004539 ":packing_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004540 "@FP16",
4541 "@FXdiv",
4542 "@clog",
4543 "@pthreadpool",
4544 ],
4545)
4546
4547xnnpack_cc_library(
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004548 name = "XNNPACK",
4549 srcs = [
4550 "src/init.c",
Marat Dukhanccfdbd12020-02-03 14:27:45 -08004551 "src/runtime.c",
4552 "src/subgraph.c",
4553 "src/tensor.c",
Marat Dukhanf03da0d2020-06-10 16:00:20 -07004554 ] + SUBGRAPH_SRCS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004555 hdrs = ["include/xnnpack.h"],
4556 copts = LOGGING_COPTS + [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004557 "-Isrc",
4558 "-Iinclude",
4559 ] + select({
4560 ":debug_build": [],
4561 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07004562 }) + select({
4563 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4564 "//conditions:default": [],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004565 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07004566 gcc_copts = xnnpack_gcc_std_copts(),
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004567 includes = ["include"],
Marat Dukhan10a38082020-04-17 03:58:35 -07004568 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004569 visibility = xnnpack_visibility(),
4570 deps = [
4571 ":enable_assembly",
Marat Dukhan9de90e02020-06-18 16:04:12 -07004572 ":enable_sparse",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004573 ":logging_utils",
Chao Mei6ddfc602020-05-13 22:29:36 -07004574 ":memory_planner",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004575 ":operator_run",
4576 ":operators",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004577 ":ukernels",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004578 "@clog",
Marat Dukhanab2946c2020-05-21 20:04:13 -07004579 "@FP16",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004580 "@pthreadpool",
Marat Dukhand343c222019-10-07 09:22:14 -07004581 ] + select({
4582 ":emscripten": [],
4583 "//conditions:default": ["@cpuinfo"],
4584 }),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004585)
4586
Marat Dukhan10a38082020-04-17 03:58:35 -07004587xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07004588 name = "XNNPACK_test_mode",
4589 srcs = [
4590 "src/init.c",
4591 "src/runtime.c",
4592 "src/subgraph.c",
4593 "src/tensor.c",
Marat Dukhanf03da0d2020-06-10 16:00:20 -07004594 ] + SUBGRAPH_SRCS,
Marat Dukhan33fcf782020-05-24 14:27:15 -07004595 hdrs = ["include/xnnpack.h"],
4596 copts = LOGGING_COPTS + [
4597 "-Isrc",
4598 "-Iinclude",
4599 "-UNDEBUG",
4600 "-DXNN_TEST_MODE=1",
4601 ] + select({
4602 ":debug_build": [],
4603 "//conditions:default": xnnpack_min_size_copts(),
4604 }) + select({
4605 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4606 "//conditions:default": [],
4607 }),
4608 gcc_copts = xnnpack_gcc_std_copts(),
4609 includes = ["include"],
4610 msvc_copts = xnnpack_msvc_std_copts(),
4611 visibility = xnnpack_visibility(),
4612 deps = [
4613 ":enable_assembly",
Marat Dukhan9de90e02020-06-18 16:04:12 -07004614 ":enable_sparse",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004615 ":logging_utils",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004616 ":memory_planner_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004617 ":operator_run_test_mode",
4618 ":operators_test_mode",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004619 ":ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004620 "@clog",
4621 "@FP16",
4622 "@pthreadpool",
4623 ] + select({
4624 ":emscripten": [],
4625 "//conditions:default": ["@cpuinfo"],
4626 }),
4627)
4628
4629xnnpack_cc_library(
Marat Dukhanae046f52020-06-15 13:16:14 -07004630 name = "xnnpack_f32",
4631 srcs = [
4632 "src/init.c",
4633 "src/runtime.c",
4634 "src/subgraph.c",
4635 "src/tensor.c",
4636 ] + SUBGRAPH_SRCS,
4637 hdrs = ["include/xnnpack.h"],
4638 copts = LOGGING_COPTS + [
4639 "-Isrc",
4640 "-Iinclude",
4641 ] + select({
4642 ":debug_build": [],
4643 "//conditions:default": xnnpack_min_size_copts(),
4644 }) + select({
4645 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4646 "//conditions:default": [],
4647 }),
4648 defines = [
Marat Dukhan16f1e1a2020-08-04 16:38:22 -07004649 "XNN_NO_QS8_OPERATORS",
Marat Dukhan08b7a972020-07-14 18:17:29 -07004650 "XNN_NO_QU8_OPERATORS",
Marat Dukhanae046f52020-06-15 13:16:14 -07004651 "XNN_NO_U8_OPERATORS",
4652 "XNN_NO_X8_OPERATORS",
4653 "XNN_NO_F16_OPERATORS",
4654 "XNN_NO_X16_OPERATORS",
4655 ],
4656 gcc_copts = xnnpack_gcc_std_copts(),
4657 includes = ["include"],
4658 msvc_copts = xnnpack_msvc_std_copts(),
4659 visibility = xnnpack_visibility(),
4660 deps = [
4661 ":enable_assembly",
Marat Dukhan9de90e02020-06-18 16:04:12 -07004662 ":enable_sparse",
Marat Dukhanae046f52020-06-15 13:16:14 -07004663 ":logging_utils",
4664 ":memory_planner",
4665 ":operator_run",
4666 ":operators",
Marat Dukhan86df3842020-06-17 14:03:31 -07004667 ":ukernels_max_compatibility",
Marat Dukhanae046f52020-06-15 13:16:14 -07004668 "@clog",
4669 "@FP16",
4670 "@pthreadpool",
4671 ] + select({
4672 ":emscripten": [],
4673 "//conditions:default": ["@cpuinfo"],
4674 }),
4675)
4676
4677xnnpack_cc_library(
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004678 name = "xnnpack_operators_nhwc_f32",
4679 srcs = [
4680 "src/init.c",
4681 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07004682 hdrs = ["include/xnnpack.h"],
4683 copts = LOGGING_COPTS + [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004684 "-Isrc",
4685 "-Iinclude",
4686 ] + select({
4687 ":debug_build": [],
4688 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07004689 }) + select({
4690 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4691 "//conditions:default": [],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004692 }),
4693 defines = [
Marat Dukhan16f1e1a2020-08-04 16:38:22 -07004694 "XNN_NO_QS8_OPERATORS",
Marat Dukhan08b7a972020-07-14 18:17:29 -07004695 "XNN_NO_QU8_OPERATORS",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004696 "XNN_NO_U8_OPERATORS",
4697 "XNN_NO_X8_OPERATORS",
Marat Dukhanefc47b82019-11-18 09:25:38 -08004698 "XNN_NO_NCHW_OPERATORS",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004699 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07004700 gcc_copts = xnnpack_gcc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004701 includes = ["include"],
Marat Dukhan10a38082020-04-17 03:58:35 -07004702 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004703 visibility = xnnpack_visibility(),
4704 deps = [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004705 ":enable_assembly",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004706 ":logging_utils",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004707 ":operator_run",
4708 ":operators",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004709 ":ukernels",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004710 "@clog",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004711 "@pthreadpool",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004712 ] + select({
4713 ":emscripten": [],
4714 "//conditions:default": ["@cpuinfo"],
4715 }),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004716)
4717
Marat Dukhancf056b22019-10-07 10:26:29 -07004718xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004719 name = "bench_utils",
4720 srcs = ["bench/utils.cc"],
4721 hdrs = ["bench/utils.h"],
Marat Dukhanbad48fe2019-11-04 10:35:22 -08004722 deps = [
4723 "@com_google_benchmark//:benchmark",
4724 "@cpuinfo",
4725 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004726)
4727
Frank Barchard7e955972019-10-11 10:34:25 -07004728######################### Benchmarks for micro-kernels #########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07004729
4730xnnpack_benchmark(
Marat Dukhan595e1702020-07-31 10:12:52 -07004731 name = "qs8_gemm_bench",
4732 srcs = [
4733 "bench/gemm.h",
4734 "bench/qs8-gemm.cc",
4735 "src/xnnpack/AlignedAllocator.h",
4736 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Frank Barchard31328cb2020-10-12 11:55:18 -07004737 copts = xnnpack_optional_ruy_copts() + xnnpack_optional_gemmlowp_copts(),
4738 deps = MICROKERNEL_BENCHMARK_DEPS + [":packing"] + xnnpack_optional_ruy_deps() + xnnpack_optional_gemmlowp_deps(),
Marat Dukhan595e1702020-07-31 10:12:52 -07004739)
4740
4741xnnpack_benchmark(
Marat Dukhan56bdd4a2020-08-03 19:47:04 -07004742 name = "qs8_requantization_bench",
4743 srcs = [
4744 "bench/qs8-requantization.cc",
Marat Dukhan56bdd4a2020-08-03 19:47:04 -07004745 "src/xnnpack/AlignedAllocator.h",
Frank Barchard04336c12020-10-22 16:48:55 -07004746 "src/xnnpack/requantization-stubs.h",
Marat Dukhan56bdd4a2020-08-03 19:47:04 -07004747 ] + MICROKERNEL_BENCHMARK_HDRS,
4748 deps = MICROKERNEL_BENCHMARK_DEPS,
4749)
4750
4751xnnpack_benchmark(
Marat Dukhan08b7a972020-07-14 18:17:29 -07004752 name = "qu8_gemm_bench",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004753 srcs = [
4754 "bench/gemm.h",
Marat Dukhan08b7a972020-07-14 18:17:29 -07004755 "bench/qu8-gemm.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004756 "src/xnnpack/AlignedAllocator.h",
4757 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004758 copts = xnnpack_optional_ruy_copts() + xnnpack_optional_gemmlowp_copts(),
Marat Dukhanab582382020-07-06 13:32:08 -07004759 deps = MICROKERNEL_BENCHMARK_DEPS + [":packing"] + xnnpack_optional_ruy_deps() + xnnpack_optional_gemmlowp_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004760)
4761
4762xnnpack_benchmark(
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07004763 name = "qu8_requantization_bench",
4764 srcs = [
4765 "bench/qu8-requantization.cc",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07004766 "src/xnnpack/AlignedAllocator.h",
Frank Barchard04336c12020-10-22 16:48:55 -07004767 "src/xnnpack/requantization-stubs.h",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07004768 ] + MICROKERNEL_BENCHMARK_HDRS,
4769 deps = MICROKERNEL_BENCHMARK_DEPS,
4770)
4771
4772xnnpack_benchmark(
Frank Barchard40d20fe2020-05-05 00:37:45 -07004773 name = "f16_igemm_bench",
4774 srcs = [
4775 "bench/f16-igemm.cc",
4776 "bench/conv.h",
4777 "bench/google/conv.h",
4778 "src/xnnpack/AlignedAllocator.h",
4779 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004780 deps = MICROKERNEL_BENCHMARK_DEPS + [
4781 ":indirection",
4782 ":packing",
4783 ],
Frank Barchard40d20fe2020-05-05 00:37:45 -07004784)
4785
4786xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004787 name = "f16_gemm_bench",
4788 srcs = [
4789 "bench/f16-gemm.cc",
4790 "bench/gemm.h",
4791 "src/xnnpack/AlignedAllocator.h",
4792 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004793 deps = MICROKERNEL_BENCHMARK_DEPS + [
4794 ":packing",
4795 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004796)
4797
4798xnnpack_benchmark(
Marat Dukhanbdb56f52020-02-05 21:42:49 -08004799 name = "f16_spmm_bench",
4800 srcs = [
4801 "bench/f16-spmm.cc",
Marat Dukhan1631e3e2020-12-06 19:29:31 -08004802 "bench/spmm.h",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08004803 "src/xnnpack/AlignedAllocator.h",
4804 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanbdb56f52020-02-05 21:42:49 -08004805 deps = MICROKERNEL_BENCHMARK_DEPS,
4806)
4807
4808xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004809 name = "f32_igemm_bench",
4810 srcs = [
4811 "bench/f32-igemm.cc",
4812 "bench/conv.h",
4813 "src/xnnpack/AlignedAllocator.h",
4814 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004815 deps = MICROKERNEL_BENCHMARK_DEPS + [
4816 ":indirection",
4817 ":packing",
4818 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004819)
4820
4821xnnpack_benchmark(
Frank Barchardc2010062020-07-19 10:04:12 -07004822 name = "f16_relu_bench",
4823 srcs = [
4824 "bench/f16-relu.cc",
4825 "src/xnnpack/AlignedAllocator.h",
4826 ] + MICROKERNEL_BENCHMARK_HDRS,
4827 deps = MICROKERNEL_BENCHMARK_DEPS,
4828)
4829
4830xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004831 name = "f32_conv_hwc_bench",
4832 srcs = [
4833 "bench/f32-conv-hwc.cc",
4834 "bench/dconv.h",
4835 "src/xnnpack/AlignedAllocator.h",
4836 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004837 deps = MICROKERNEL_BENCHMARK_DEPS + [
4838 ":packing",
4839 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004840)
4841
4842xnnpack_benchmark(
Marat Dukhan1f29b802020-05-15 23:46:39 -07004843 name = "f32_conv_hwc2chw_bench",
Erich Elsen563df5f2019-10-23 08:02:21 -07004844 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07004845 "bench/f32-conv-hwc2chw.cc",
Erich Elsen563df5f2019-10-23 08:02:21 -07004846 "bench/dconv.h",
4847 "src/xnnpack/AlignedAllocator.h",
4848 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004849 deps = MICROKERNEL_BENCHMARK_DEPS + [
4850 ":packing",
4851 ],
Erich Elsen563df5f2019-10-23 08:02:21 -07004852)
4853
4854xnnpack_benchmark(
Frank Barchard5a599a62020-06-04 20:12:44 -07004855 name = "f16_dwconv_bench",
4856 srcs = [
4857 "bench/f16-dwconv.cc",
4858 "bench/dwconv.h",
4859 "bench/google/dwconv.h",
4860 "src/xnnpack/AlignedAllocator.h",
4861 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004862 deps = MICROKERNEL_BENCHMARK_DEPS + [
4863 ":indirection",
4864 ":packing",
4865 ],
Frank Barchard5a599a62020-06-04 20:12:44 -07004866)
4867
4868xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004869 name = "f32_dwconv_bench",
4870 srcs = [
4871 "bench/f32-dwconv.cc",
4872 "bench/dwconv.h",
4873 "src/xnnpack/AlignedAllocator.h",
4874 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004875 deps = MICROKERNEL_BENCHMARK_DEPS + [
4876 ":indirection",
4877 ":packing",
4878 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004879)
4880
4881xnnpack_benchmark(
Marat Dukhanbf715f92020-10-23 20:17:00 -07004882 name = "f32_dwconv2d_chw_bench",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004883 srcs = [
Marat Dukhanbf715f92020-10-23 20:17:00 -07004884 "bench/f32-dwconv2d-chw.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004885 "bench/dwconv.h",
4886 "src/xnnpack/AlignedAllocator.h",
4887 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004888 deps = MICROKERNEL_BENCHMARK_DEPS + [
4889 ":indirection",
4890 ":packing",
4891 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004892)
4893
4894xnnpack_benchmark(
4895 name = "f32_gemm_bench",
4896 srcs = [
4897 "bench/f32-gemm.cc",
4898 "bench/gemm.h",
4899 "src/xnnpack/AlignedAllocator.h",
4900 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004901 copts = xnnpack_optional_ruy_copts(),
Marat Dukhanab582382020-07-06 13:32:08 -07004902 deps = MICROKERNEL_BENCHMARK_DEPS + [":packing"] + xnnpack_optional_ruy_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004903)
4904
4905xnnpack_benchmark(
Marat Dukhana11ca342020-06-25 23:45:07 -07004906 name = "f32_hswish_bench",
4907 srcs = [
4908 "bench/f32-hswish.cc",
4909 "src/xnnpack/AlignedAllocator.h",
4910 ] + MICROKERNEL_BENCHMARK_HDRS,
4911 deps = MICROKERNEL_BENCHMARK_DEPS,
4912)
4913
4914xnnpack_benchmark(
Marat Dukhan4c4eb002019-12-08 21:27:49 -08004915 name = "f32_raddexpminusmax_bench",
4916 srcs = [
4917 "bench/f32-raddexpminusmax.cc",
4918 "src/xnnpack/AlignedAllocator.h",
4919 ] + MICROKERNEL_BENCHMARK_HDRS,
4920 deps = MICROKERNEL_BENCHMARK_DEPS,
4921)
4922
4923xnnpack_benchmark(
4924 name = "f32_raddextexp_bench",
4925 srcs = [
4926 "bench/f32-raddextexp.cc",
4927 "src/xnnpack/AlignedAllocator.h",
4928 ] + MICROKERNEL_BENCHMARK_HDRS,
4929 deps = MICROKERNEL_BENCHMARK_DEPS,
4930)
4931
4932xnnpack_benchmark(
4933 name = "f32_raddstoreexpminusmax_bench",
4934 srcs = [
4935 "bench/f32-raddstoreexpminusmax.cc",
4936 "src/xnnpack/AlignedAllocator.h",
4937 ] + MICROKERNEL_BENCHMARK_HDRS,
4938 deps = MICROKERNEL_BENCHMARK_DEPS,
4939)
4940
4941xnnpack_benchmark(
Frank Barchard3c4a9522020-07-18 16:11:22 -07004942 name = "f32_relu_bench",
4943 srcs = [
4944 "bench/f32-relu.cc",
4945 "src/xnnpack/AlignedAllocator.h",
4946 ] + MICROKERNEL_BENCHMARK_HDRS,
4947 deps = MICROKERNEL_BENCHMARK_DEPS,
4948)
4949
4950xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004951 name = "f32_rmax_bench",
4952 srcs = [
4953 "bench/f32-rmax.cc",
4954 "src/xnnpack/AlignedAllocator.h",
4955 ] + MICROKERNEL_BENCHMARK_HDRS,
4956 deps = MICROKERNEL_BENCHMARK_DEPS,
4957)
4958
4959xnnpack_benchmark(
Marat Dukhan14bec502019-11-18 11:35:31 -08004960 name = "f32_sigmoid_bench",
4961 srcs = [
4962 "bench/f32-sigmoid.cc",
4963 "src/xnnpack/AlignedAllocator.h",
4964 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan14bec502019-11-18 11:35:31 -08004965 deps = MICROKERNEL_BENCHMARK_DEPS,
4966)
4967
4968xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004969 name = "f32_spmm_bench",
4970 srcs = [
4971 "bench/f32-spmm.cc",
Marat Dukhan1631e3e2020-12-06 19:29:31 -08004972 "bench/spmm.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004973 "src/xnnpack/AlignedAllocator.h",
4974 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07004975 deps = MICROKERNEL_BENCHMARK_DEPS,
4976)
4977
4978xnnpack_benchmark(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08004979 name = "f32_softmax_bench",
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07004980 srcs = [
Marat Dukhanfd8e6892020-01-27 15:25:25 -08004981 "bench/f32-softmax.cc",
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07004982 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004983 copts = xnnpack_optional_dnnl_copts(),
Marat Dukhan8d3c6932020-03-06 20:27:27 -08004984 deps = MICROKERNEL_BENCHMARK_DEPS + xnnpack_optional_dnnl_deps(),
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07004985)
4986
4987xnnpack_benchmark(
Marat Dukhaned6baaf2020-12-01 15:07:08 -08004988 name = "f32_velu_bench",
4989 srcs = [
4990 "bench/f32-velu.cc",
4991 "src/xnnpack/AlignedAllocator.h",
4992 ] + MICROKERNEL_BENCHMARK_HDRS,
4993 deps = MICROKERNEL_BENCHMARK_DEPS,
4994)
4995
4996xnnpack_benchmark(
Marat Dukhan4c4eb002019-12-08 21:27:49 -08004997 name = "f32_vscaleexpminusmax_bench",
4998 srcs = [
4999 "bench/f32-vscaleexpminusmax.cc",
5000 "src/xnnpack/AlignedAllocator.h",
5001 ] + MICROKERNEL_BENCHMARK_HDRS,
5002 deps = MICROKERNEL_BENCHMARK_DEPS,
5003)
5004
5005xnnpack_benchmark(
5006 name = "f32_vscaleextexp_bench",
5007 srcs = [
5008 "bench/f32-vscaleextexp.cc",
5009 "src/xnnpack/AlignedAllocator.h",
5010 ] + MICROKERNEL_BENCHMARK_HDRS,
5011 deps = MICROKERNEL_BENCHMARK_DEPS,
5012)
5013
5014xnnpack_benchmark(
Marat Dukhanf4db2f32020-06-30 10:55:30 -07005015 name = "f32_vsqrt_bench",
5016 srcs = [
5017 "bench/f32-vsqrt.cc",
5018 "src/xnnpack/AlignedAllocator.h",
5019 ] + MICROKERNEL_BENCHMARK_HDRS,
5020 deps = MICROKERNEL_BENCHMARK_DEPS,
5021)
5022
5023xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005024 name = "f32_im2col_gemm_bench",
5025 srcs = [
5026 "bench/f32-im2col-gemm.cc",
5027 "bench/conv.h",
5028 "src/xnnpack/AlignedAllocator.h",
5029 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005030 deps = MICROKERNEL_BENCHMARK_DEPS + [
5031 ":im2col",
5032 ":packing",
5033 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005034)
5035
Marat Dukhanfe7acb62020-03-09 19:30:05 -07005036xnnpack_benchmark(
Marat Dukhanffbf96a2020-05-14 02:59:08 -07005037 name = "rounding_bench",
5038 srcs = [
5039 "bench/rounding.cc",
Marat Dukhanffbf96a2020-05-14 02:59:08 -07005040 "src/xnnpack/AlignedAllocator.h",
Frank Barchard04336c12020-10-22 16:48:55 -07005041 "src/xnnpack/math-stubs.h",
Marat Dukhanffbf96a2020-05-14 02:59:08 -07005042 ] + MICROKERNEL_BENCHMARK_HDRS,
5043 deps = MICROKERNEL_BENCHMARK_DEPS,
5044)
5045
Marat Dukhan08c4a432019-10-03 09:29:21 -07005046########################### Benchmarks for operators ###########################
5047
5048xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005049 name = "average_pooling_bench",
5050 srcs = ["bench/average-pooling.cc"],
Marat Dukhan7a16d8b2020-03-11 04:22:44 -07005051 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07005052 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07005053 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07005054)
5055
5056xnnpack_benchmark(
Marat Dukhan87727142020-06-24 15:24:10 -07005057 name = "bankers_rounding_bench",
5058 srcs = ["bench/bankers-rounding.cc"],
5059 copts = xnnpack_optional_tflite_copts(),
5060 tags = ["nowin32"],
5061 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
5062)
5063
5064xnnpack_benchmark(
5065 name = "ceiling_bench",
5066 srcs = ["bench/ceiling.cc"],
5067 copts = xnnpack_optional_tflite_copts(),
5068 tags = ["nowin32"],
5069 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
5070)
5071
5072xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005073 name = "channel_shuffle_bench",
5074 srcs = ["bench/channel-shuffle.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07005075 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005076)
5077
5078xnnpack_benchmark(
5079 name = "convolution_bench",
5080 srcs = ["bench/convolution.cc"],
5081 copts = xnnpack_optional_tflite_copts() + xnnpack_optional_armcl_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07005082 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07005083 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps() + xnnpack_optional_armcl_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07005084)
5085
5086xnnpack_benchmark(
5087 name = "deconvolution_bench",
5088 srcs = ["bench/deconvolution.cc"],
5089 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07005090 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07005091 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07005092)
5093
5094xnnpack_benchmark(
Marat Dukhanb6bd4bc2020-12-01 17:01:40 -08005095 name = "elu_bench",
5096 srcs = ["bench/elu.cc"],
5097 copts = xnnpack_optional_tflite_copts(),
5098 tags = ["nowin32"],
5099 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
5100)
5101
5102xnnpack_benchmark(
Marat Dukhan87727142020-06-24 15:24:10 -07005103 name = "floor_bench",
5104 srcs = ["bench/floor.cc"],
5105 copts = xnnpack_optional_tflite_copts(),
5106 tags = ["nowin32"],
5107 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
5108)
5109
5110xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005111 name = "global_average_pooling_bench",
5112 srcs = ["bench/global-average-pooling.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07005113 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005114)
5115
5116xnnpack_benchmark(
Marat Dukhanad352602020-06-25 21:50:54 -07005117 name = "hardswish_bench",
5118 srcs = ["bench/hardswish.cc"],
5119 copts = xnnpack_optional_tflite_copts(),
5120 tags = ["nowin32"],
5121 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
5122)
5123
5124xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005125 name = "max_pooling_bench",
5126 srcs = ["bench/max-pooling.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07005127 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005128)
5129
5130xnnpack_benchmark(
5131 name = "sigmoid_bench",
5132 srcs = ["bench/sigmoid.cc"],
Marat Dukhanc3b9e862019-11-17 13:18:54 -08005133 copts = xnnpack_optional_tflite_copts(),
Marat Dukhanca2ba702020-04-24 01:31:47 -07005134 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07005135 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07005136)
5137
5138xnnpack_benchmark(
Marat Dukhan95b22432019-10-30 16:30:14 -07005139 name = "prelu_bench",
5140 srcs = ["bench/prelu.cc"],
5141 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07005142 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07005143 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan95b22432019-10-30 16:30:14 -07005144)
5145
5146xnnpack_benchmark(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08005147 name = "softmax_bench",
5148 srcs = ["bench/softmax.cc"],
Marat Dukhan9c0db962020-01-28 12:30:14 -08005149 copts = xnnpack_optional_tflite_copts(),
Marat Dukhanca2ba702020-04-24 01:31:47 -07005150 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07005151 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07005152)
5153
Marat Dukhan87727142020-06-24 15:24:10 -07005154xnnpack_benchmark(
Marat Dukhan6804bbd2020-06-30 19:26:11 -07005155 name = "square_root_bench",
5156 srcs = ["bench/square-root.cc"],
5157 copts = xnnpack_optional_tflite_copts(),
5158 tags = ["nowin32"],
5159 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
5160)
5161
5162xnnpack_benchmark(
Marat Dukhan87727142020-06-24 15:24:10 -07005163 name = "truncation_bench",
5164 srcs = ["bench/truncation.cc"],
5165 deps = OPERATOR_BENCHMARK_DEPS,
5166)
5167
Marat Dukhanc068bb62019-10-04 13:24:39 -07005168############################# End-to-end benchmarks ############################
5169
5170cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07005171 name = "fp32_mobilenet_v1",
5172 srcs = ["models/fp32-mobilenet-v1.cc"],
Marat Dukhanc068bb62019-10-04 13:24:39 -07005173 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08005174 copts = xnnpack_std_cxxopts(),
Marat Dukhanc068bb62019-10-04 13:24:39 -07005175 linkstatic = True,
5176 deps = [
5177 ":XNNPACK",
5178 "@pthreadpool",
5179 ],
5180)
5181
5182cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07005183 name = "fp16_mobilenet_v1",
5184 srcs = ["models/fp16-mobilenet-v1.cc"],
5185 hdrs = ["models/models.h"],
5186 copts = xnnpack_std_cxxopts(),
5187 linkstatic = True,
5188 deps = [
5189 ":XNNPACK",
5190 "@FP16",
5191 "@pthreadpool",
5192 ],
5193)
5194
5195cc_library(
Marat Dukhan0743cdf2020-08-04 18:52:07 -07005196 name = "qs8_mobilenet_v1",
5197 srcs = ["models/qs8-mobilenet-v1.cc"],
5198 hdrs = ["models/models.h"],
5199 copts = xnnpack_std_cxxopts(),
5200 linkstatic = True,
5201 deps = [
5202 ":XNNPACK",
5203 "@pthreadpool",
5204 ],
5205)
5206
5207cc_library(
Marat Dukhan70a96182020-09-03 17:13:58 -07005208 name = "qs8_mobilenet_v2",
5209 srcs = ["models/qs8-mobilenet-v2.cc"],
5210 hdrs = ["models/models.h"],
5211 copts = xnnpack_std_cxxopts(),
5212 linkstatic = True,
5213 deps = [
5214 ":XNNPACK",
5215 "@pthreadpool",
5216 ],
5217)
5218
5219cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07005220 name = "fp32_mobilenet_v2",
5221 srcs = ["models/fp32-mobilenet-v2.cc"],
Marat Dukhanc068bb62019-10-04 13:24:39 -07005222 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08005223 copts = xnnpack_std_cxxopts(),
Marat Dukhanc068bb62019-10-04 13:24:39 -07005224 linkstatic = True,
5225 deps = [
5226 ":XNNPACK",
5227 "@pthreadpool",
5228 ],
5229)
5230
Marat Dukhanc08cdf52019-12-09 09:17:51 -08005231cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07005232 name = "fp16_mobilenet_v2",
5233 srcs = ["models/fp16-mobilenet-v2.cc"],
5234 hdrs = ["models/models.h"],
5235 copts = xnnpack_std_cxxopts(),
5236 linkstatic = True,
5237 deps = [
5238 ":XNNPACK",
5239 "@FP16",
5240 "@pthreadpool",
5241 ],
5242)
5243
5244cc_library(
5245 name = "fp32_mobilenet_v3_large",
5246 srcs = ["models/fp32-mobilenet-v3-large.cc"],
Marat Dukhanc08cdf52019-12-09 09:17:51 -08005247 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08005248 copts = xnnpack_std_cxxopts(),
Marat Dukhanc08cdf52019-12-09 09:17:51 -08005249 linkstatic = True,
5250 deps = [
5251 ":XNNPACK",
5252 "@pthreadpool",
5253 ],
5254)
5255
5256cc_library(
Marat Dukhan66f3ccd2020-09-14 16:09:38 -07005257 name = "fp16_mobilenet_v3_large",
5258 srcs = ["models/fp16-mobilenet-v3-large.cc"],
5259 hdrs = ["models/models.h"],
5260 copts = xnnpack_std_cxxopts(),
5261 linkstatic = True,
5262 deps = [
5263 ":XNNPACK",
5264 "@FP16",
5265 "@pthreadpool",
5266 ],
5267)
5268
5269cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07005270 name = "fp32_mobilenet_v3_small",
5271 srcs = ["models/fp32-mobilenet-v3-small.cc"],
Marat Dukhanc08cdf52019-12-09 09:17:51 -08005272 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08005273 copts = xnnpack_std_cxxopts(),
Marat Dukhanc08cdf52019-12-09 09:17:51 -08005274 linkstatic = True,
5275 deps = [
5276 ":XNNPACK",
5277 "@pthreadpool",
5278 ],
5279)
5280
Marat Dukhan66f3ccd2020-09-14 16:09:38 -07005281cc_library(
5282 name = "fp16_mobilenet_v3_small",
5283 srcs = ["models/fp16-mobilenet-v3-small.cc"],
5284 hdrs = ["models/models.h"],
5285 copts = xnnpack_std_cxxopts(),
5286 linkstatic = True,
5287 deps = [
5288 ":XNNPACK",
5289 "@FP16",
5290 "@pthreadpool",
5291 ],
5292)
5293
Marat Dukhanc068bb62019-10-04 13:24:39 -07005294xnnpack_benchmark(
Marat Dukhanef4416e2019-10-31 13:44:40 -07005295 name = "f32_dwconv_e2e_bench",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08005296 srcs = [
5297 "bench/f32-dwconv-e2e.cc",
5298 "bench/end2end.h",
5299 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanef4416e2019-10-31 13:44:40 -07005300 deps = MICROKERNEL_BENCHMARK_DEPS + [
5301 ":XNNPACK",
Marat Dukhan270a2c42020-06-26 16:45:52 -07005302 ":fp32_mobilenet_v1",
5303 ":fp32_mobilenet_v2",
5304 ":fp32_mobilenet_v3_large",
5305 ":fp32_mobilenet_v3_small",
Marat Dukhanef4416e2019-10-31 13:44:40 -07005306 ],
5307)
5308
5309xnnpack_benchmark(
Marat Dukhan5f18d262019-10-31 10:24:14 -07005310 name = "f32_gemm_e2e_bench",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08005311 srcs = [
5312 "bench/f32-gemm-e2e.cc",
5313 "bench/end2end.h",
5314 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan5f18d262019-10-31 10:24:14 -07005315 deps = MICROKERNEL_BENCHMARK_DEPS + [
5316 ":XNNPACK",
Marat Dukhan270a2c42020-06-26 16:45:52 -07005317 ":fp32_mobilenet_v1",
5318 ":fp32_mobilenet_v2",
5319 ":fp32_mobilenet_v3_large",
5320 ":fp32_mobilenet_v3_small",
Marat Dukhan5f18d262019-10-31 10:24:14 -07005321 ],
5322)
5323
5324xnnpack_benchmark(
Marat Dukhanc068bb62019-10-04 13:24:39 -07005325 name = "end2end_bench",
5326 srcs = ["bench/end2end.cc"],
5327 deps = [
5328 ":XNNPACK",
Frank Barchardc712fa42019-10-31 14:00:21 -07005329 ":bench_utils",
Marat Dukhan270a2c42020-06-26 16:45:52 -07005330 ":fp16_mobilenet_v1",
5331 ":fp16_mobilenet_v2",
Marat Dukhan66f3ccd2020-09-14 16:09:38 -07005332 ":fp16_mobilenet_v3_large",
5333 ":fp16_mobilenet_v3_small",
Marat Dukhan270a2c42020-06-26 16:45:52 -07005334 ":fp32_mobilenet_v1",
5335 ":fp32_mobilenet_v2",
5336 ":fp32_mobilenet_v3_large",
5337 ":fp32_mobilenet_v3_small",
Marat Dukhan0743cdf2020-08-04 18:52:07 -07005338 ":qs8_mobilenet_v1",
Marat Dukhan70a96182020-09-03 17:13:58 -07005339 ":qs8_mobilenet_v2",
Marat Dukhanc068bb62019-10-04 13:24:39 -07005340 "@pthreadpool",
5341 ],
5342)
5343
Marat Dukhan6adff4e2019-10-14 18:32:07 -07005344#################### Accuracy evaluation for math functions ####################
5345
5346xnnpack_benchmark(
5347 name = "f32_exp_eval",
5348 srcs = [
5349 "eval/f32-exp.cc",
5350 "src/xnnpack/AlignedAllocator.h",
5351 ] + ACCURACY_EVAL_HDRS,
Marat Dukhan3a305212020-12-06 19:24:27 -08005352 deps = ACCURACY_EVAL_DEPS + [
5353 ":bench_utils",
5354 "@cpuinfo",
5355 ],
Marat Dukhan6adff4e2019-10-14 18:32:07 -07005356)
5357
Marat Dukhan515c9772019-10-17 18:07:57 -07005358xnnpack_benchmark(
5359 name = "f32_expminus_eval",
5360 srcs = [
5361 "eval/f32-expminus.cc",
5362 "src/xnnpack/AlignedAllocator.h",
5363 ] + ACCURACY_EVAL_HDRS,
Marat Dukhan3a305212020-12-06 19:24:27 -08005364 deps = ACCURACY_EVAL_DEPS + [
5365 ":bench_utils",
5366 "@cpuinfo",
5367 ],
Marat Dukhan515c9772019-10-17 18:07:57 -07005368)
5369
Marat Dukhan98ba4412019-10-23 02:14:28 -07005370xnnpack_benchmark(
Marat Dukhana438aca2020-11-20 15:45:01 -08005371 name = "f32_expm1minus_eval",
5372 srcs = [
5373 "eval/f32-expm1minus.cc",
5374 "src/xnnpack/AlignedAllocator.h",
5375 ] + ACCURACY_EVAL_HDRS,
Marat Dukhande390d42020-11-29 19:32:18 -08005376 deps = ACCURACY_EVAL_DEPS + [
5377 ":bench_utils",
Marat Dukhan3a305212020-12-06 19:24:27 -08005378 "@cpuinfo",
Marat Dukhande390d42020-11-29 19:32:18 -08005379 ],
Marat Dukhana438aca2020-11-20 15:45:01 -08005380)
5381
5382xnnpack_benchmark(
Marat Dukhan98ba4412019-10-23 02:14:28 -07005383 name = "f32_extexp_eval",
5384 srcs = [
5385 "eval/f32-extexp.cc",
5386 "src/xnnpack/AlignedAllocator.h",
5387 ] + ACCURACY_EVAL_HDRS,
Marat Dukhan3a305212020-12-06 19:24:27 -08005388 deps = ACCURACY_EVAL_DEPS + [
5389 ":bench_utils",
5390 "@cpuinfo",
5391 ],
Marat Dukhan98ba4412019-10-23 02:14:28 -07005392)
5393
Marat Dukhan8853b822020-05-07 12:19:01 -07005394xnnpack_unit_test(
5395 name = "f32_roundne_eval",
5396 srcs = [
5397 "eval/f32-roundne.cc",
5398 "src/xnnpack/AlignedAllocator.h",
5399 "src/xnnpack/math-stubs.h",
5400 ] + MICROKERNEL_TEST_HDRS,
Marat Dukhan22eed3d2020-05-11 20:13:37 -07005401 automatic = False,
Marat Dukhan8853b822020-05-07 12:19:01 -07005402 deps = MICROKERNEL_TEST_DEPS,
5403)
5404
Marat Dukhan2dbb9442020-05-12 20:43:43 -07005405xnnpack_unit_test(
Marat Dukhanc9852ba2020-05-13 17:21:29 -07005406 name = "f32_roundd_eval",
5407 srcs = [
5408 "eval/f32-roundd.cc",
5409 "src/xnnpack/AlignedAllocator.h",
5410 "src/xnnpack/math-stubs.h",
5411 ] + MICROKERNEL_TEST_HDRS,
5412 automatic = False,
5413 deps = MICROKERNEL_TEST_DEPS,
5414)
5415
5416xnnpack_unit_test(
5417 name = "f32_roundu_eval",
5418 srcs = [
5419 "eval/f32-roundu.cc",
5420 "src/xnnpack/AlignedAllocator.h",
5421 "src/xnnpack/math-stubs.h",
5422 ] + MICROKERNEL_TEST_HDRS,
5423 automatic = False,
5424 deps = MICROKERNEL_TEST_DEPS,
5425)
5426
5427xnnpack_unit_test(
Marat Dukhan2dbb9442020-05-12 20:43:43 -07005428 name = "f32_roundz_eval",
5429 srcs = [
5430 "eval/f32-roundz.cc",
5431 "src/xnnpack/AlignedAllocator.h",
5432 "src/xnnpack/math-stubs.h",
5433 ] + MICROKERNEL_TEST_HDRS,
Marat Dukhanc9852ba2020-05-13 17:21:29 -07005434 automatic = False,
Marat Dukhan2dbb9442020-05-12 20:43:43 -07005435 deps = MICROKERNEL_TEST_DEPS,
5436)
5437
Marat Dukhan346a9e52019-11-15 09:06:30 -08005438xnnpack_benchmark(
5439 name = "f32_sigmoid_eval",
5440 srcs = [
5441 "eval/f32-sigmoid.cc",
5442 "src/xnnpack/AlignedAllocator.h",
5443 ] + ACCURACY_EVAL_HDRS,
Marat Dukhan3a305212020-12-06 19:24:27 -08005444 deps = ACCURACY_EVAL_DEPS + [
5445 ":bench_utils",
5446 "@cpuinfo",
5447 ],
Marat Dukhan346a9e52019-11-15 09:06:30 -08005448)
5449
Marat Dukhan84000762020-06-29 18:38:43 -07005450xnnpack_benchmark(
5451 name = "f32_sqrt_eval",
5452 srcs = [
5453 "eval/f32-sqrt.cc",
5454 "src/xnnpack/AlignedAllocator.h",
5455 ] + ACCURACY_EVAL_HDRS,
5456 deps = ACCURACY_EVAL_DEPS + [
5457 ":bench_utils",
Marat Dukhan3a305212020-12-06 19:24:27 -08005458 "@cpuinfo",
Marat Dukhan84000762020-06-29 18:38:43 -07005459 ],
5460)
5461
Marat Dukhan08c4a432019-10-03 09:29:21 -07005462######################### Unit tests for micro-kernels #########################
5463
5464xnnpack_unit_test(
Marat Dukhande06f492020-04-09 00:19:31 -07005465 name = "f16_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005466 srcs = [
Marat Dukhande06f492020-04-09 00:19:31 -07005467 "test/f16-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005468 "test/gemm-microkernel-tester.h",
5469 "src/xnnpack/AlignedAllocator.h",
5470 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005471 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005472)
5473
5474xnnpack_unit_test(
Marat Dukhan355ab432020-04-09 19:01:52 -07005475 name = "f16_spmm_minmax_test",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08005476 srcs = [
Marat Dukhan355ab432020-04-09 19:01:52 -07005477 "test/f16-spmm-minmax.cc",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08005478 "test/spmm-microkernel-tester.h",
5479 "src/xnnpack/AlignedAllocator.h",
5480 ] + MICROKERNEL_TEST_HDRS,
5481 deps = MICROKERNEL_TEST_DEPS,
5482)
5483
5484xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005485 name = "f32_argmaxpool_test",
5486 srcs = [
5487 "test/f32-argmaxpool.cc",
5488 "test/argmaxpool-microkernel-tester.h",
5489 "src/xnnpack/AlignedAllocator.h",
5490 ] + MICROKERNEL_TEST_HDRS,
5491 deps = MICROKERNEL_TEST_DEPS,
5492)
5493
5494xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07005495 name = "f32_avgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005496 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07005497 "test/f32-avgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005498 "test/avgpool-microkernel-tester.h",
5499 "src/xnnpack/AlignedAllocator.h",
5500 ] + MICROKERNEL_TEST_HDRS,
5501 deps = MICROKERNEL_TEST_DEPS,
5502)
5503
5504xnnpack_unit_test(
Marat Dukhan660fd192020-03-10 04:55:30 -07005505 name = "f32_ibilinear_test",
Marat Dukhan35dacfb2019-11-07 19:18:16 -08005506 srcs = [
Marat Dukhan660fd192020-03-10 04:55:30 -07005507 "test/f32-ibilinear.cc",
5508 "test/ibilinear-microkernel-tester.h",
Marat Dukhan35dacfb2019-11-07 19:18:16 -08005509 "src/xnnpack/AlignedAllocator.h",
5510 ] + MICROKERNEL_TEST_HDRS,
5511 deps = MICROKERNEL_TEST_DEPS,
5512)
5513
5514xnnpack_unit_test(
XNNPACK Team21432672020-10-19 19:58:48 -07005515 name = "f32_ibilinear_chw_test",
5516 srcs = [
5517 "test/f32-ibilinear-chw.cc",
XNNPACK Team6be46b22020-10-22 23:34:54 -07005518 "test/ibilinear-microkernel-tester.h",
XNNPACK Team21432672020-10-19 19:58:48 -07005519 "src/xnnpack/AlignedAllocator.h",
5520 ] + MICROKERNEL_TEST_HDRS,
5521 deps = MICROKERNEL_TEST_DEPS,
5522)
5523
5524xnnpack_unit_test(
Frank Barchardb1966592020-05-12 13:47:06 -07005525 name = "f16_clamp_test",
5526 srcs = [
5527 "test/f16-clamp.cc",
5528 "test/clamp-microkernel-tester.h",
5529 ] + MICROKERNEL_TEST_HDRS,
5530 deps = MICROKERNEL_TEST_DEPS,
5531)
5532
5533xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005534 name = "f32_clamp_test",
5535 srcs = [
5536 "test/f32-clamp.cc",
5537 "test/clamp-microkernel-tester.h",
5538 ] + MICROKERNEL_TEST_HDRS,
5539 deps = MICROKERNEL_TEST_DEPS,
5540)
5541
5542xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07005543 name = "f32_igemm_test",
5544 srcs = [
5545 "test/f32-igemm.cc",
5546 "test/gemm-microkernel-tester.h",
5547 "src/xnnpack/AlignedAllocator.h",
5548 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005549 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan163a7e62020-04-09 04:19:26 -07005550)
5551
5552xnnpack_unit_test(
Marat Dukhan467f6362020-05-22 23:21:55 -07005553 name = "f32_igemm_relu_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005554 srcs = [
Marat Dukhan467f6362020-05-22 23:21:55 -07005555 "test/f32-igemm-relu.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005556 "test/gemm-microkernel-tester.h",
5557 "src/xnnpack/AlignedAllocator.h",
5558 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005559 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005560)
5561
5562xnnpack_unit_test(
Marat Dukhane207b7b2020-05-28 16:27:42 -07005563 name = "f32_igemm_minmax_test",
5564 srcs = [
5565 "test/f32-igemm-minmax.cc",
5566 "test/gemm-microkernel-tester.h",
5567 "src/xnnpack/AlignedAllocator.h",
5568 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005569 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhane207b7b2020-05-28 16:27:42 -07005570)
5571
5572xnnpack_unit_test(
Frank Barchardb0e4fae2020-05-04 15:27:51 -07005573 name = "f16_igemm_minmax_test",
5574 srcs = [
5575 "test/f16-igemm-minmax.cc",
5576 "test/gemm-microkernel-tester.h",
5577 "src/xnnpack/AlignedAllocator.h",
5578 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005579 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Frank Barchardb0e4fae2020-05-04 15:27:51 -07005580)
5581
5582xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005583 name = "f32_conv_hwc_test",
5584 srcs = [
5585 "test/f32-conv-hwc.cc",
5586 "test/conv-hwc-microkernel-tester.h",
5587 "src/xnnpack/AlignedAllocator.h",
5588 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005589 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005590)
5591
5592xnnpack_unit_test(
Marat Dukhan1f29b802020-05-15 23:46:39 -07005593 name = "f32_conv_hwc2chw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005594 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07005595 "test/f32-conv-hwc2chw.cc",
5596 "test/conv-hwc2chw-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005597 "src/xnnpack/AlignedAllocator.h",
5598 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005599 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005600)
5601
5602xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07005603 name = "f32_dwconv_test",
5604 srcs = [
5605 "test/f32-dwconv.cc",
5606 "test/dwconv-microkernel-tester.h",
5607 "src/xnnpack/AlignedAllocator.h",
5608 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005609 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan163a7e62020-04-09 04:19:26 -07005610)
5611
5612xnnpack_unit_test(
Frank Barchard5a599a62020-06-04 20:12:44 -07005613 name = "f16_dwconv_minmax_test",
5614 srcs = [
5615 "test/f16-dwconv-minmax.cc",
5616 "test/dwconv-microkernel-tester.h",
5617 "src/xnnpack/AlignedAllocator.h",
5618 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005619 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Frank Barchard5a599a62020-06-04 20:12:44 -07005620)
5621
5622xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07005623 name = "f32_dwconv_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005624 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07005625 "test/f32-dwconv-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005626 "test/dwconv-microkernel-tester.h",
5627 "src/xnnpack/AlignedAllocator.h",
5628 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005629 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005630)
5631
5632xnnpack_unit_test(
Marat Dukhanbf715f92020-10-23 20:17:00 -07005633 name = "f32_dwconv2d_chw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005634 srcs = [
Marat Dukhanbf715f92020-10-23 20:17:00 -07005635 "test/f32-dwconv2d-chw.cc",
5636 "test/dwconv2d-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005637 "src/xnnpack/AlignedAllocator.h",
5638 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005639 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005640)
5641
5642xnnpack_unit_test(
Frank Barchard0bb49a72020-06-04 11:35:11 -07005643 name = "f16_gavgpool_minmax_test",
5644 srcs = [
5645 "test/f16-gavgpool-minmax.cc",
5646 "test/gavgpool-microkernel-tester.h",
5647 "src/xnnpack/AlignedAllocator.h",
5648 ] + MICROKERNEL_TEST_HDRS,
5649 deps = MICROKERNEL_TEST_DEPS,
5650)
5651
5652xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07005653 name = "f32_gavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005654 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07005655 "test/f32-gavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005656 "test/gavgpool-microkernel-tester.h",
5657 "src/xnnpack/AlignedAllocator.h",
5658 ] + MICROKERNEL_TEST_HDRS,
5659 deps = MICROKERNEL_TEST_DEPS,
5660)
5661
5662xnnpack_unit_test(
Marat Dukhan1f29b802020-05-15 23:46:39 -07005663 name = "f32_gavgpool_cw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005664 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07005665 "test/f32-gavgpool-cw.cc",
5666 "test/gavgpool-cw-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005667 "src/xnnpack/AlignedAllocator.h",
5668 ] + MICROKERNEL_TEST_HDRS,
5669 deps = MICROKERNEL_TEST_DEPS,
5670)
5671
5672xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07005673 name = "f32_gemm_test",
5674 srcs = [
5675 "test/f32-gemm.cc",
5676 "test/gemm-microkernel-tester.h",
5677 "src/xnnpack/AlignedAllocator.h",
5678 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005679 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan163a7e62020-04-09 04:19:26 -07005680)
5681
5682xnnpack_unit_test(
Marat Dukhan467f6362020-05-22 23:21:55 -07005683 name = "f32_gemm_relu_test",
5684 srcs = [
5685 "test/f32-gemm-relu.cc",
5686 "test/gemm-microkernel-tester.h",
5687 "src/xnnpack/AlignedAllocator.h",
5688 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005689 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan467f6362020-05-22 23:21:55 -07005690)
5691
5692xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07005693 name = "f32_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005694 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07005695 "test/f32-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005696 "test/gemm-microkernel-tester.h",
5697 "src/xnnpack/AlignedAllocator.h",
5698 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005699 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005700)
5701
5702xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07005703 name = "f32_gemminc_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005704 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07005705 "test/f32-gemminc-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005706 "test/gemm-microkernel-tester.h",
5707 "src/xnnpack/AlignedAllocator.h",
5708 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005709 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005710)
5711
5712xnnpack_unit_test(
Frank Barchardb1966592020-05-12 13:47:06 -07005713 name = "f16_hswish_test",
5714 srcs = [
5715 "test/f16-hswish.cc",
5716 "test/hswish-microkernel-tester.h",
5717 ] + MICROKERNEL_TEST_HDRS,
5718 deps = MICROKERNEL_TEST_DEPS,
5719)
5720
5721xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005722 name = "f32_hswish_test",
5723 srcs = [
5724 "test/f32-hswish.cc",
5725 "test/hswish-microkernel-tester.h",
5726 ] + MICROKERNEL_TEST_HDRS,
5727 deps = MICROKERNEL_TEST_DEPS,
5728)
5729
5730xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07005731 name = "f32_maxpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005732 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07005733 "test/f32-maxpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005734 "test/maxpool-microkernel-tester.h",
5735 ] + MICROKERNEL_TEST_HDRS,
5736 deps = MICROKERNEL_TEST_DEPS,
5737)
5738
5739xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07005740 name = "f32_pavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005741 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07005742 "test/f32-pavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005743 "test/avgpool-microkernel-tester.h",
5744 "src/xnnpack/AlignedAllocator.h",
5745 ] + MICROKERNEL_TEST_HDRS,
5746 deps = MICROKERNEL_TEST_DEPS,
5747)
5748
5749xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07005750 name = "f32_ppmm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005751 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07005752 "test/f32-ppmm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005753 "test/gemm-microkernel-tester.h",
5754 "src/xnnpack/AlignedAllocator.h",
5755 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005756 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005757)
5758
5759xnnpack_unit_test(
Frank Barchardb1966592020-05-12 13:47:06 -07005760 name = "f16_prelu_test",
5761 srcs = [
5762 "test/f16-prelu.cc",
5763 "test/prelu-microkernel-tester.h",
5764 "src/xnnpack/AlignedAllocator.h",
5765 ] + MICROKERNEL_TEST_HDRS,
5766 deps = MICROKERNEL_TEST_DEPS,
5767)
5768
5769xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005770 name = "f32_prelu_test",
5771 srcs = [
5772 "test/f32-prelu.cc",
5773 "test/prelu-microkernel-tester.h",
5774 "src/xnnpack/AlignedAllocator.h",
5775 ] + MICROKERNEL_TEST_HDRS,
5776 deps = MICROKERNEL_TEST_DEPS,
5777)
5778
5779xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07005780 name = "f32_raddexpminusmax_test",
5781 srcs = [
5782 "test/f32-raddexpminusmax.cc",
5783 "test/raddexpminusmax-microkernel-tester.h",
5784 ] + MICROKERNEL_TEST_HDRS,
5785 deps = MICROKERNEL_TEST_DEPS,
5786)
5787
5788xnnpack_unit_test(
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07005789 name = "f32_raddextexp_test",
5790 srcs = [
5791 "test/f32-raddextexp.cc",
5792 "test/raddextexp-microkernel-tester.h",
5793 ] + MICROKERNEL_TEST_HDRS,
5794 deps = MICROKERNEL_TEST_DEPS,
5795)
5796
5797xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07005798 name = "f32_raddstoreexpminusmax_test",
5799 srcs = [
5800 "test/f32-raddstoreexpminusmax.cc",
5801 "test/raddstoreexpminusmax-microkernel-tester.h",
5802 ] + MICROKERNEL_TEST_HDRS,
5803 deps = MICROKERNEL_TEST_DEPS,
5804)
5805
5806xnnpack_unit_test(
Frank Barchardfb158e22020-07-15 16:10:10 -07005807 name = "f32_relu_test",
5808 srcs = [
5809 "test/f32-relu.cc",
5810 "test/vunary-microkernel-tester.h",
5811 ] + MICROKERNEL_TEST_HDRS,
5812 deps = MICROKERNEL_TEST_DEPS,
5813)
5814
5815xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005816 name = "f32_rmax_test",
5817 srcs = [
5818 "test/f32-rmax.cc",
5819 "test/rmax-microkernel-tester.h",
5820 ] + MICROKERNEL_TEST_HDRS,
5821 deps = MICROKERNEL_TEST_DEPS,
5822)
5823
5824xnnpack_unit_test(
Marat Dukhan346a9e52019-11-15 09:06:30 -08005825 name = "f32_sigmoid_test",
5826 srcs = [
5827 "test/f32-sigmoid.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005828 "test/vunary-microkernel-tester.h",
Marat Dukhan346a9e52019-11-15 09:06:30 -08005829 ] + MICROKERNEL_TEST_HDRS,
5830 deps = MICROKERNEL_TEST_DEPS,
5831)
5832
5833xnnpack_unit_test(
Marat Dukhan355ab432020-04-09 19:01:52 -07005834 name = "f32_spmm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005835 srcs = [
Marat Dukhan355ab432020-04-09 19:01:52 -07005836 "test/f32-spmm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005837 "test/spmm-microkernel-tester.h",
5838 "src/xnnpack/AlignedAllocator.h",
5839 ] + MICROKERNEL_TEST_HDRS,
5840 deps = MICROKERNEL_TEST_DEPS,
5841)
5842
5843xnnpack_unit_test(
Frank Barchardd793f6c2020-05-08 13:37:43 -07005844 name = "f16_vadd_minmax_test",
5845 srcs = [
5846 "test/f16-vadd-minmax.cc",
5847 "test/vbinary-microkernel-tester.h",
5848 ] + MICROKERNEL_TEST_HDRS,
5849 deps = MICROKERNEL_TEST_DEPS,
5850)
5851
5852xnnpack_unit_test(
5853 name = "f16_vaddc_minmax_test",
5854 srcs = [
5855 "test/f16-vaddc-minmax.cc",
5856 "test/vbinaryc-microkernel-tester.h",
5857 ] + MICROKERNEL_TEST_HDRS,
5858 deps = MICROKERNEL_TEST_DEPS,
5859)
5860
5861xnnpack_unit_test(
5862 name = "f16_vdiv_minmax_test",
5863 srcs = [
5864 "test/f16-vdiv-minmax.cc",
5865 "test/vbinary-microkernel-tester.h",
5866 ] + MICROKERNEL_TEST_HDRS,
5867 deps = MICROKERNEL_TEST_DEPS,
5868)
5869
5870xnnpack_unit_test(
5871 name = "f16_vdivc_minmax_test",
5872 srcs = [
5873 "test/f16-vdivc-minmax.cc",
5874 "test/vbinaryc-microkernel-tester.h",
5875 ] + MICROKERNEL_TEST_HDRS,
5876 deps = MICROKERNEL_TEST_DEPS,
5877)
5878
5879xnnpack_unit_test(
5880 name = "f16_vrdivc_minmax_test",
5881 srcs = [
5882 "test/f16-vrdivc-minmax.cc",
5883 "test/vbinaryc-microkernel-tester.h",
5884 ] + MICROKERNEL_TEST_HDRS,
5885 deps = MICROKERNEL_TEST_DEPS,
5886)
5887
5888xnnpack_unit_test(
5889 name = "f16_vmax_test",
5890 srcs = [
5891 "test/f16-vmax.cc",
5892 "test/vbinary-microkernel-tester.h",
5893 ] + MICROKERNEL_TEST_HDRS,
5894 deps = MICROKERNEL_TEST_DEPS,
5895)
5896
5897xnnpack_unit_test(
5898 name = "f16_vmaxc_test",
5899 srcs = [
5900 "test/f16-vmaxc.cc",
5901 "test/vbinaryc-microkernel-tester.h",
5902 ] + MICROKERNEL_TEST_HDRS,
5903 deps = MICROKERNEL_TEST_DEPS,
5904)
5905
5906xnnpack_unit_test(
5907 name = "f16_vmin_test",
5908 srcs = [
5909 "test/f16-vmin.cc",
5910 "test/vbinary-microkernel-tester.h",
5911 ] + MICROKERNEL_TEST_HDRS,
5912 deps = MICROKERNEL_TEST_DEPS,
5913)
5914
5915xnnpack_unit_test(
5916 name = "f16_vminc_test",
5917 srcs = [
5918 "test/f16-vminc.cc",
5919 "test/vbinaryc-microkernel-tester.h",
5920 ] + MICROKERNEL_TEST_HDRS,
5921 deps = MICROKERNEL_TEST_DEPS,
5922)
5923
5924xnnpack_unit_test(
5925 name = "f16_vmul_minmax_test",
5926 srcs = [
5927 "test/f16-vmul-minmax.cc",
5928 "test/vbinary-microkernel-tester.h",
5929 ] + MICROKERNEL_TEST_HDRS,
5930 deps = MICROKERNEL_TEST_DEPS,
5931)
5932
5933xnnpack_unit_test(
5934 name = "f16_vmulc_minmax_test",
5935 srcs = [
5936 "test/f16-vmulc-minmax.cc",
5937 "test/vbinaryc-microkernel-tester.h",
5938 ] + MICROKERNEL_TEST_HDRS,
5939 deps = MICROKERNEL_TEST_DEPS,
5940)
5941
5942xnnpack_unit_test(
5943 name = "f16_vsub_minmax_test",
5944 srcs = [
5945 "test/f16-vsub-minmax.cc",
5946 "test/vbinary-microkernel-tester.h",
5947 ] + MICROKERNEL_TEST_HDRS,
5948 deps = MICROKERNEL_TEST_DEPS,
5949)
5950
5951xnnpack_unit_test(
5952 name = "f16_vsubc_minmax_test",
5953 srcs = [
5954 "test/f16-vsubc-minmax.cc",
5955 "test/vbinaryc-microkernel-tester.h",
5956 ] + MICROKERNEL_TEST_HDRS,
5957 deps = MICROKERNEL_TEST_DEPS,
5958)
5959
5960xnnpack_unit_test(
5961 name = "f16_vrsubc_minmax_test",
5962 srcs = [
5963 "test/f16-vrsubc-minmax.cc",
5964 "test/vbinaryc-microkernel-tester.h",
5965 ] + MICROKERNEL_TEST_HDRS,
5966 deps = MICROKERNEL_TEST_DEPS,
5967)
5968
5969xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07005970 name = "f32_vabs_test",
5971 srcs = [
5972 "test/f32-vabs.cc",
5973 "test/vunary-microkernel-tester.h",
5974 ] + MICROKERNEL_TEST_HDRS,
5975 deps = MICROKERNEL_TEST_DEPS,
5976)
5977
5978xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005979 name = "f32_vadd_test",
5980 srcs = [
5981 "test/f32-vadd.cc",
5982 "test/vbinary-microkernel-tester.h",
5983 ] + MICROKERNEL_TEST_HDRS,
5984 deps = MICROKERNEL_TEST_DEPS,
5985)
5986
5987xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005988 name = "f32_vadd_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005989 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005990 "test/f32-vadd-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005991 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005992 ] + MICROKERNEL_TEST_HDRS,
5993 deps = MICROKERNEL_TEST_DEPS,
5994)
5995
5996xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005997 name = "f32_vadd_relu_test",
5998 srcs = [
5999 "test/f32-vadd-relu.cc",
6000 "test/vbinary-microkernel-tester.h",
6001 ] + MICROKERNEL_TEST_HDRS,
6002 deps = MICROKERNEL_TEST_DEPS,
6003)
6004
6005xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07006006 name = "f32_vaddc_test",
6007 srcs = [
6008 "test/f32-vaddc.cc",
6009 "test/vbinaryc-microkernel-tester.h",
6010 ] + MICROKERNEL_TEST_HDRS,
6011 deps = MICROKERNEL_TEST_DEPS,
6012)
6013
6014xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006015 name = "f32_vaddc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08006016 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006017 "test/f32-vaddc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08006018 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006019 ] + MICROKERNEL_TEST_HDRS,
6020 deps = MICROKERNEL_TEST_DEPS,
6021)
6022
6023xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07006024 name = "f32_vaddc_relu_test",
6025 srcs = [
6026 "test/f32-vaddc-relu.cc",
6027 "test/vbinaryc-microkernel-tester.h",
6028 ] + MICROKERNEL_TEST_HDRS,
6029 deps = MICROKERNEL_TEST_DEPS,
6030)
6031
6032xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07006033 name = "f32_vdiv_test",
6034 srcs = [
6035 "test/f32-vdiv.cc",
6036 "test/vbinary-microkernel-tester.h",
6037 ] + MICROKERNEL_TEST_HDRS,
6038 deps = MICROKERNEL_TEST_DEPS,
6039)
6040
6041xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006042 name = "f32_vdiv_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08006043 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006044 "test/f32-vdiv-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08006045 "test/vbinary-microkernel-tester.h",
6046 ] + MICROKERNEL_TEST_HDRS,
6047 deps = MICROKERNEL_TEST_DEPS,
6048)
6049
6050xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07006051 name = "f32_vdiv_relu_test",
6052 srcs = [
6053 "test/f32-vdiv-relu.cc",
6054 "test/vbinary-microkernel-tester.h",
6055 ] + MICROKERNEL_TEST_HDRS,
6056 deps = MICROKERNEL_TEST_DEPS,
6057)
6058
6059xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07006060 name = "f32_vdivc_test",
6061 srcs = [
6062 "test/f32-vdivc.cc",
6063 "test/vbinaryc-microkernel-tester.h",
6064 ] + MICROKERNEL_TEST_HDRS,
6065 deps = MICROKERNEL_TEST_DEPS,
6066)
6067
6068xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006069 name = "f32_vdivc_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08006070 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006071 "test/f32-vdivc-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08006072 "test/vbinaryc-microkernel-tester.h",
6073 ] + MICROKERNEL_TEST_HDRS,
6074 deps = MICROKERNEL_TEST_DEPS,
6075)
6076
6077xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07006078 name = "f32_vdivc_relu_test",
6079 srcs = [
6080 "test/f32-vdivc-relu.cc",
6081 "test/vbinaryc-microkernel-tester.h",
6082 ] + MICROKERNEL_TEST_HDRS,
6083 deps = MICROKERNEL_TEST_DEPS,
6084)
6085
6086xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07006087 name = "f32_vrdivc_test",
6088 srcs = [
6089 "test/f32-vrdivc.cc",
6090 "test/vbinaryc-microkernel-tester.h",
6091 ] + MICROKERNEL_TEST_HDRS,
6092 deps = MICROKERNEL_TEST_DEPS,
6093)
6094
6095xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006096 name = "f32_vrdivc_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08006097 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006098 "test/f32-vrdivc-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08006099 "test/vbinaryc-microkernel-tester.h",
6100 ] + MICROKERNEL_TEST_HDRS,
6101 deps = MICROKERNEL_TEST_DEPS,
6102)
6103
6104xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07006105 name = "f32_vrdivc_relu_test",
6106 srcs = [
6107 "test/f32-vrdivc-relu.cc",
6108 "test/vbinaryc-microkernel-tester.h",
6109 ] + MICROKERNEL_TEST_HDRS,
6110 deps = MICROKERNEL_TEST_DEPS,
6111)
6112
6113xnnpack_unit_test(
Marat Dukhaned6baaf2020-12-01 15:07:08 -08006114 name = "f32_velu_test",
6115 srcs = [
6116 "test/f32-velu.cc",
6117 "test/vunary-microkernel-tester.h",
6118 ] + MICROKERNEL_TEST_HDRS,
6119 deps = MICROKERNEL_TEST_DEPS,
6120)
6121
6122xnnpack_unit_test(
Marat Dukhan403b7d42019-12-05 12:49:11 -08006123 name = "f32_vmax_test",
6124 srcs = [
6125 "test/f32-vmax.cc",
6126 "test/vbinary-microkernel-tester.h",
6127 ] + MICROKERNEL_TEST_HDRS,
6128 deps = MICROKERNEL_TEST_DEPS,
6129)
6130
6131xnnpack_unit_test(
6132 name = "f32_vmaxc_test",
6133 srcs = [
6134 "test/f32-vmaxc.cc",
6135 "test/vbinaryc-microkernel-tester.h",
6136 ] + MICROKERNEL_TEST_HDRS,
6137 deps = MICROKERNEL_TEST_DEPS,
6138)
6139
6140xnnpack_unit_test(
6141 name = "f32_vmin_test",
6142 srcs = [
6143 "test/f32-vmin.cc",
6144 "test/vbinary-microkernel-tester.h",
6145 ] + MICROKERNEL_TEST_HDRS,
6146 deps = MICROKERNEL_TEST_DEPS,
6147)
6148
6149xnnpack_unit_test(
6150 name = "f32_vminc_test",
6151 srcs = [
6152 "test/f32-vminc.cc",
6153 "test/vbinaryc-microkernel-tester.h",
6154 ] + MICROKERNEL_TEST_HDRS,
6155 deps = MICROKERNEL_TEST_DEPS,
6156)
6157
6158xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07006159 name = "f32_vmul_test",
6160 srcs = [
6161 "test/f32-vmul.cc",
6162 "test/vbinary-microkernel-tester.h",
6163 ] + MICROKERNEL_TEST_HDRS,
6164 deps = MICROKERNEL_TEST_DEPS,
6165)
6166
6167xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006168 name = "f32_vmul_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006169 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006170 "test/f32-vmul-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08006171 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08006172 ] + MICROKERNEL_TEST_HDRS,
6173 deps = MICROKERNEL_TEST_DEPS,
6174)
6175
6176xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07006177 name = "f32_vmul_relu_test",
6178 srcs = [
6179 "test/f32-vmul-relu.cc",
6180 "test/vbinary-microkernel-tester.h",
6181 ] + MICROKERNEL_TEST_HDRS,
6182 deps = MICROKERNEL_TEST_DEPS,
6183)
6184
6185xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07006186 name = "f32_vmulc_test",
6187 srcs = [
6188 "test/f32-vmulc.cc",
6189 "test/vbinaryc-microkernel-tester.h",
6190 ] + MICROKERNEL_TEST_HDRS,
6191 deps = MICROKERNEL_TEST_DEPS,
6192)
6193
6194xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006195 name = "f32_vmulc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08006196 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006197 "test/f32-vmulc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08006198 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006199 ] + MICROKERNEL_TEST_HDRS,
6200 deps = MICROKERNEL_TEST_DEPS,
6201)
6202
6203xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07006204 name = "f32_vmulc_relu_test",
6205 srcs = [
6206 "test/f32-vmulc-relu.cc",
6207 "test/vbinaryc-microkernel-tester.h",
6208 ] + MICROKERNEL_TEST_HDRS,
6209 deps = MICROKERNEL_TEST_DEPS,
6210)
6211
6212xnnpack_unit_test(
Frank Barchard2a1049c2020-06-03 02:31:27 -07006213 name = "f16_vmulcaddc_minmax_test",
6214 srcs = [
6215 "test/f16-vmulcaddc-minmax.cc",
6216 "test/vmulcaddc-microkernel-tester.h",
6217 "src/xnnpack/AlignedAllocator.h",
6218 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07006219 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Frank Barchard2a1049c2020-06-03 02:31:27 -07006220)
6221
6222xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07006223 name = "f32_vmulcaddc_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006224 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07006225 "test/f32-vmulcaddc-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006226 "test/vmulcaddc-microkernel-tester.h",
6227 "src/xnnpack/AlignedAllocator.h",
6228 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07006229 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07006230)
6231
6232xnnpack_unit_test(
Marat Dukhan8cc7efe2020-06-10 16:24:27 -07006233 name = "f32_vlrelu_test",
6234 srcs = [
6235 "test/f32-vlrelu.cc",
6236 "test/vunary-microkernel-tester.h",
6237 ] + MICROKERNEL_TEST_HDRS,
6238 deps = MICROKERNEL_TEST_DEPS,
6239)
6240
6241xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07006242 name = "f32_vneg_test",
6243 srcs = [
6244 "test/f32-vneg.cc",
6245 "test/vunary-microkernel-tester.h",
6246 ] + MICROKERNEL_TEST_HDRS,
6247 deps = MICROKERNEL_TEST_DEPS,
6248)
6249
6250xnnpack_unit_test(
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07006251 name = "f32_vrndne_test",
6252 srcs = [
6253 "test/f32-vrndne.cc",
6254 "test/vunary-microkernel-tester.h",
6255 ] + MICROKERNEL_TEST_HDRS,
6256 deps = MICROKERNEL_TEST_DEPS,
6257)
6258
6259xnnpack_unit_test(
6260 name = "f32_vrndz_test",
6261 srcs = [
6262 "test/f32-vrndz.cc",
6263 "test/vunary-microkernel-tester.h",
6264 ] + MICROKERNEL_TEST_HDRS,
6265 deps = MICROKERNEL_TEST_DEPS,
6266)
6267
6268xnnpack_unit_test(
6269 name = "f32_vrndu_test",
6270 srcs = [
6271 "test/f32-vrndu.cc",
6272 "test/vunary-microkernel-tester.h",
6273 ] + MICROKERNEL_TEST_HDRS,
6274 deps = MICROKERNEL_TEST_DEPS,
6275)
6276
6277xnnpack_unit_test(
6278 name = "f32_vrndd_test",
6279 srcs = [
6280 "test/f32-vrndd.cc",
6281 "test/vunary-microkernel-tester.h",
6282 ] + MICROKERNEL_TEST_HDRS,
6283 deps = MICROKERNEL_TEST_DEPS,
6284)
6285
6286xnnpack_unit_test(
Marat Dukhan05ac8e32019-10-21 15:39:33 -07006287 name = "f32_vscale_test",
6288 srcs = [
6289 "test/f32-vscale.cc",
6290 "test/vscale-microkernel-tester.h",
6291 ] + MICROKERNEL_TEST_HDRS,
6292 deps = MICROKERNEL_TEST_DEPS,
6293)
6294
6295xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07006296 name = "f32_vscaleexpminusmax_test",
6297 srcs = [
6298 "test/f32-vscaleexpminusmax.cc",
6299 "test/vscaleexpminusmax-microkernel-tester.h",
6300 ] + MICROKERNEL_TEST_HDRS,
6301 deps = MICROKERNEL_TEST_DEPS,
6302)
6303
6304xnnpack_unit_test(
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07006305 name = "f32_vscaleextexp_test",
6306 srcs = [
6307 "test/f32-vscaleextexp.cc",
6308 "test/vscaleextexp-microkernel-tester.h",
6309 ] + MICROKERNEL_TEST_HDRS,
6310 deps = MICROKERNEL_TEST_DEPS,
6311)
6312
6313xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07006314 name = "f32_vsqr_test",
6315 srcs = [
6316 "test/f32-vsqr.cc",
6317 "test/vunary-microkernel-tester.h",
6318 ] + MICROKERNEL_TEST_HDRS,
6319 deps = MICROKERNEL_TEST_DEPS,
6320)
6321
6322xnnpack_unit_test(
Marat Dukhan13bafb02020-06-05 00:43:11 -07006323 name = "f32_vsqrdiff_test",
6324 srcs = [
6325 "test/f32-vsqrdiff.cc",
6326 "test/vbinary-microkernel-tester.h",
6327 ] + MICROKERNEL_TEST_HDRS,
6328 deps = MICROKERNEL_TEST_DEPS,
6329)
6330
6331xnnpack_unit_test(
6332 name = "f32_vsqrdiffc_test",
6333 srcs = [
6334 "test/f32-vsqrdiffc.cc",
6335 "test/vbinaryc-microkernel-tester.h",
6336 ] + MICROKERNEL_TEST_HDRS,
6337 deps = MICROKERNEL_TEST_DEPS,
6338)
6339
6340xnnpack_unit_test(
Marat Dukhanf4db2f32020-06-30 10:55:30 -07006341 name = "f32_vsqrt_test",
6342 srcs = [
6343 "test/f32-vsqrt.cc",
6344 "test/vunary-microkernel-tester.h",
6345 ] + MICROKERNEL_TEST_HDRS,
6346 deps = MICROKERNEL_TEST_DEPS,
6347)
6348
6349xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07006350 name = "f32_vsub_test",
6351 srcs = [
6352 "test/f32-vsub.cc",
6353 "test/vbinary-microkernel-tester.h",
6354 ] + MICROKERNEL_TEST_HDRS,
6355 deps = MICROKERNEL_TEST_DEPS,
6356)
6357
6358xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006359 name = "f32_vsub_minmax_test",
Marat Dukhan97579532019-10-18 16:40:39 -07006360 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006361 "test/f32-vsub-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08006362 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08006363 ] + MICROKERNEL_TEST_HDRS,
6364 deps = MICROKERNEL_TEST_DEPS,
6365)
6366
6367xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07006368 name = "f32_vsub_relu_test",
6369 srcs = [
6370 "test/f32-vsub-relu.cc",
6371 "test/vbinary-microkernel-tester.h",
6372 ] + MICROKERNEL_TEST_HDRS,
6373 deps = MICROKERNEL_TEST_DEPS,
6374)
6375
6376xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07006377 name = "f32_vsubc_test",
6378 srcs = [
6379 "test/f32-vsubc.cc",
6380 "test/vbinaryc-microkernel-tester.h",
6381 ] + MICROKERNEL_TEST_HDRS,
6382 deps = MICROKERNEL_TEST_DEPS,
6383)
6384
6385xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006386 name = "f32_vsubc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08006387 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006388 "test/f32-vsubc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08006389 "test/vbinaryc-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08006390 ] + MICROKERNEL_TEST_HDRS,
6391 deps = MICROKERNEL_TEST_DEPS,
6392)
6393
6394xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07006395 name = "f32_vsubc_relu_test",
6396 srcs = [
6397 "test/f32-vsubc-relu.cc",
6398 "test/vbinaryc-microkernel-tester.h",
6399 ] + MICROKERNEL_TEST_HDRS,
6400 deps = MICROKERNEL_TEST_DEPS,
6401)
6402
6403xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07006404 name = "f32_vrsubc_test",
6405 srcs = [
6406 "test/f32-vrsubc.cc",
6407 "test/vbinaryc-microkernel-tester.h",
6408 ] + MICROKERNEL_TEST_HDRS,
6409 deps = MICROKERNEL_TEST_DEPS,
6410)
6411
6412xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006413 name = "f32_vrsubc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08006414 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006415 "test/f32-vrsubc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08006416 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan97579532019-10-18 16:40:39 -07006417 ] + MICROKERNEL_TEST_HDRS,
6418 deps = MICROKERNEL_TEST_DEPS,
6419)
6420
6421xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07006422 name = "f32_vrsubc_relu_test",
6423 srcs = [
6424 "test/f32-vrsubc-relu.cc",
6425 "test/vbinaryc-microkernel-tester.h",
6426 ] + MICROKERNEL_TEST_HDRS,
6427 deps = MICROKERNEL_TEST_DEPS,
6428)
6429
6430xnnpack_unit_test(
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07006431 name = "qs8_dwconv_minmax_test",
6432 srcs = [
6433 "test/qs8-dwconv-minmax.cc",
6434 "test/dwconv-microkernel-tester.h",
6435 "src/xnnpack/AlignedAllocator.h",
6436 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
6437 deps = MICROKERNEL_TEST_DEPS + [":packing"],
6438)
6439
6440xnnpack_unit_test(
Marat Dukhan4ed53f42020-08-06 01:12:55 -07006441 name = "qs8_gavgpool_minmax_test",
6442 srcs = [
6443 "test/qs8-gavgpool-minmax.cc",
6444 "test/gavgpool-microkernel-tester.h",
6445 "src/xnnpack/AlignedAllocator.h",
6446 ] + MICROKERNEL_TEST_HDRS,
6447 deps = MICROKERNEL_TEST_DEPS,
6448)
6449
6450xnnpack_unit_test(
Marat Dukhan595e1702020-07-31 10:12:52 -07006451 name = "qs8_gemm_minmax_test",
6452 srcs = [
6453 "test/qs8-gemm-minmax.cc",
6454 "test/gemm-microkernel-tester.h",
6455 "src/xnnpack/AlignedAllocator.h",
6456 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
6457 deps = MICROKERNEL_TEST_DEPS + [":packing"],
6458)
6459
6460xnnpack_unit_test(
Marat Dukhanf9480682020-07-31 14:50:24 -07006461 name = "qs8_igemm_minmax_test",
6462 srcs = [
6463 "test/qs8-igemm-minmax.cc",
6464 "test/gemm-microkernel-tester.h",
6465 "src/xnnpack/AlignedAllocator.h",
6466 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
6467 deps = MICROKERNEL_TEST_DEPS + [":packing"],
6468)
6469
6470xnnpack_unit_test(
6471 name = "qs8_requantization_test",
6472 srcs = [
6473 "src/xnnpack/requantization-stubs.h",
6474 "test/qs8-requantization.cc",
6475 "test/requantization-tester.h",
6476 ] + MICROKERNEL_TEST_HDRS,
6477 deps = MICROKERNEL_TEST_DEPS,
6478)
6479
6480xnnpack_unit_test(
Marat Dukhand9f3ad42020-08-10 12:30:58 -07006481 name = "qs8_vadd_minmax_test",
6482 srcs = [
6483 "test/qs8-vadd-minmax.cc",
6484 "test/vadd-microkernel-tester.h",
6485 ] + MICROKERNEL_TEST_HDRS,
6486 deps = MICROKERNEL_TEST_DEPS,
6487)
6488
6489xnnpack_unit_test(
Marat Dukhan0270d9f2020-08-11 00:56:46 -07006490 name = "qs8_vaddc_minmax_test",
6491 srcs = [
6492 "test/qs8-vaddc-minmax.cc",
6493 "test/vaddc-microkernel-tester.h",
6494 ] + MICROKERNEL_TEST_HDRS,
6495 deps = MICROKERNEL_TEST_DEPS,
6496)
6497
6498xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07006499 name = "qu8_avgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006500 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07006501 "test/qu8-avgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006502 "test/avgpool-microkernel-tester.h",
6503 "src/xnnpack/AlignedAllocator.h",
6504 ] + MICROKERNEL_TEST_HDRS,
6505 deps = MICROKERNEL_TEST_DEPS,
6506)
6507
6508xnnpack_unit_test(
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07006509 name = "qu8_dwconv_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006510 srcs = [
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07006511 "test/qu8-dwconv-minmax.cc",
6512 "test/dwconv-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006513 "src/xnnpack/AlignedAllocator.h",
6514 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07006515 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07006516)
6517
6518xnnpack_unit_test(
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07006519 name = "qu8_igemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006520 srcs = [
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07006521 "test/qu8-igemm-minmax.cc",
6522 "test/gemm-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006523 "src/xnnpack/AlignedAllocator.h",
6524 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07006525 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07006526)
6527
6528xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07006529 name = "qu8_gavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006530 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07006531 "test/qu8-gavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006532 "test/gavgpool-microkernel-tester.h",
6533 "src/xnnpack/AlignedAllocator.h",
6534 ] + MICROKERNEL_TEST_HDRS,
6535 deps = MICROKERNEL_TEST_DEPS,
6536)
6537
6538xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07006539 name = "qu8_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006540 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07006541 "test/qu8-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006542 "test/gemm-microkernel-tester.h",
6543 "src/xnnpack/AlignedAllocator.h",
6544 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07006545 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07006546)
6547
6548xnnpack_unit_test(
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07006549 name = "qu8_requantization_test",
6550 srcs = [
6551 "src/xnnpack/requantization-stubs.h",
6552 "test/qu8-requantization.cc",
6553 "test/requantization-tester.h",
6554 ] + MICROKERNEL_TEST_HDRS,
6555 deps = MICROKERNEL_TEST_DEPS,
6556)
6557
6558xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07006559 name = "qu8_vadd_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006560 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07006561 "test/qu8-vadd-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006562 "test/vadd-microkernel-tester.h",
6563 ] + MICROKERNEL_TEST_HDRS,
6564 deps = MICROKERNEL_TEST_DEPS,
6565)
6566
6567xnnpack_unit_test(
6568 name = "u8_clamp_test",
6569 srcs = [
6570 "test/u8-clamp.cc",
6571 "test/clamp-microkernel-tester.h",
6572 ] + MICROKERNEL_TEST_HDRS,
6573 deps = MICROKERNEL_TEST_DEPS,
6574)
6575
6576xnnpack_unit_test(
6577 name = "u8_lut32norm_test",
6578 srcs = [
6579 "test/u8-lut32norm.cc",
6580 "test/lut-norm-microkernel-tester.h",
6581 ] + MICROKERNEL_TEST_HDRS,
6582 deps = MICROKERNEL_TEST_DEPS,
6583)
6584
6585xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07006586 name = "u8_maxpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006587 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07006588 "test/u8-maxpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006589 "test/maxpool-microkernel-tester.h",
6590 ] + MICROKERNEL_TEST_HDRS,
6591 deps = MICROKERNEL_TEST_DEPS,
6592)
6593
6594xnnpack_unit_test(
6595 name = "u8_rmax_test",
6596 srcs = [
6597 "test/u8-rmax.cc",
6598 "test/rmax-microkernel-tester.h",
6599 ] + MICROKERNEL_TEST_HDRS,
6600 deps = MICROKERNEL_TEST_DEPS,
6601)
6602
6603xnnpack_unit_test(
Marat Dukhanad71b9a2020-11-20 00:01:51 -08006604 name = "x32_depthtospace2d_chw2hwc_test",
Yury Kartynnike7841862020-11-04 18:22:18 -08006605 srcs = [
Marat Dukhanad71b9a2020-11-20 00:01:51 -08006606 "test/x32-depthtospace2d-chw2hwc.cc",
6607 "test/depthtospace-microkernel-tester.h",
Yury Kartynnike7841862020-11-04 18:22:18 -08006608 ] + MICROKERNEL_TEST_HDRS,
6609 deps = MICROKERNEL_TEST_DEPS,
6610)
6611
6612xnnpack_unit_test(
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -07006613 name = "x32_fill_test",
6614 srcs = [
6615 "test/x32-fill.cc",
6616 "test/fill-microkernel-tester.h",
6617 ] + MICROKERNEL_TEST_HDRS,
6618 deps = MICROKERNEL_TEST_DEPS,
6619)
6620
6621xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07006622 name = "x32_packx_test",
6623 srcs = [
6624 "test/x32-packx.cc",
6625 "test/pack-microkernel-tester.h",
6626 "src/xnnpack/AlignedAllocator.h",
6627 ] + MICROKERNEL_TEST_HDRS,
6628 deps = MICROKERNEL_TEST_DEPS,
6629)
6630
6631xnnpack_unit_test(
6632 name = "x32_pad_test",
6633 srcs = [
6634 "test/x32-pad.cc",
6635 "test/pad-microkernel-tester.h",
6636 ] + MICROKERNEL_TEST_HDRS,
6637 deps = MICROKERNEL_TEST_DEPS,
6638)
6639
6640xnnpack_unit_test(
6641 name = "x32_unpool_test",
6642 srcs = [
6643 "test/x32-unpool.cc",
6644 "test/unpool-microkernel-tester.h",
6645 ] + MICROKERNEL_TEST_HDRS,
6646 deps = MICROKERNEL_TEST_DEPS,
6647)
6648
6649xnnpack_unit_test(
6650 name = "x32_zip_test",
6651 srcs = [
6652 "test/x32-zip.cc",
6653 "test/zip-microkernel-tester.h",
6654 ] + MICROKERNEL_TEST_HDRS,
6655 deps = MICROKERNEL_TEST_DEPS,
6656)
6657
6658xnnpack_unit_test(
6659 name = "x8_lut_test",
6660 srcs = [
6661 "test/x8-lut.cc",
6662 "test/lut-microkernel-tester.h",
6663 ] + MICROKERNEL_TEST_HDRS,
6664 deps = MICROKERNEL_TEST_DEPS,
6665)
6666
6667xnnpack_unit_test(
6668 name = "x8_zip_test",
6669 srcs = [
6670 "test/x8-zip.cc",
6671 "test/zip-microkernel-tester.h",
6672 ] + MICROKERNEL_TEST_HDRS,
6673 deps = MICROKERNEL_TEST_DEPS,
6674)
6675
Marat Dukhan20c3b922020-03-10 03:45:06 -07006676########################## Size tests for the library #########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07006677
6678xnnpack_binary(
Marat Dukhan20c3b922020-03-10 03:45:06 -07006679 name = "operator_size_test",
6680 srcs = ["test/operator-size.c"],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07006681 deps = [":xnnpack_operators_nhwc_f32"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07006682)
6683
Marat Dukhan20c3b922020-03-10 03:45:06 -07006684xnnpack_binary(
6685 name = "subgraph_size_test",
6686 srcs = ["test/subgraph-size.c"],
6687 deps = [":XNNPACK"],
6688)
6689
6690########################### Unit tests for operators ##########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07006691
6692xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07006693 name = "abs_nc_test",
6694 srcs = [
6695 "test/abs-nc.cc",
6696 "test/abs-operator-tester.h",
6697 ],
6698 deps = OPERATOR_TEST_DEPS,
6699)
6700
6701xnnpack_unit_test(
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08006702 name = "add_nd_test",
6703 srcs = [
6704 "test/add-nd.cc",
6705 "test/binary-elementwise-operator-tester.h",
6706 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006707 deps = OPERATOR_TEST_DEPS,
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08006708)
6709
6710xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006711 name = "argmax_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006712 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006713 "test/argmax-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006714 "test/argmax-pooling-operator-tester.h",
6715 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07006716 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006717)
6718
6719xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006720 name = "average_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006721 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006722 "test/average-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006723 "test/average-pooling-operator-tester.h",
6724 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07006725 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006726)
6727
6728xnnpack_unit_test(
Marat Dukhan64e52512020-06-09 13:41:16 -07006729 name = "bankers_rounding_nc_test",
6730 srcs = [
6731 "test/bankers-rounding-nc.cc",
6732 "test/bankers-rounding-operator-tester.h",
6733 ],
6734 deps = OPERATOR_TEST_DEPS,
6735)
6736
6737xnnpack_unit_test(
6738 name = "ceiling_nc_test",
6739 srcs = [
6740 "test/ceiling-nc.cc",
6741 "test/ceiling-operator-tester.h",
6742 ],
6743 deps = OPERATOR_TEST_DEPS,
6744)
6745
6746xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006747 name = "channel_shuffle_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006748 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006749 "test/channel-shuffle-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006750 "test/channel-shuffle-operator-tester.h",
6751 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006752 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006753)
6754
6755xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006756 name = "clamp_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006757 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006758 "test/clamp-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006759 "test/clamp-operator-tester.h",
6760 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006761 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006762)
6763
6764xnnpack_unit_test(
Marat Dukhan065b11e2020-05-22 09:49:41 -07006765 name = "constant_pad_nd_test",
6766 srcs = [
6767 "test/constant-pad-nd.cc",
6768 "test/constant-pad-operator-tester.h",
6769 ],
6770 deps = OPERATOR_TEST_DEPS,
6771)
6772
6773xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006774 name = "convolution_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006775 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006776 "test/convolution-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006777 "test/convolution-operator-tester.h",
6778 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006779 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006780)
6781
6782xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006783 name = "convolution_nchw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006784 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006785 "test/convolution-nchw.cc",
6786 "test/convolution-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006787 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006788 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006789)
6790
6791xnnpack_unit_test(
Marat Dukhan4e21b272020-06-04 18:45:01 -07006792 name = "copy_nc_test",
6793 srcs = [
6794 "test/copy-nc.cc",
6795 "test/copy-operator-tester.h",
6796 ],
6797 deps = OPERATOR_TEST_DEPS,
6798)
6799
6800xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006801 name = "deconvolution_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006802 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006803 "test/deconvolution-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006804 "test/deconvolution-operator-tester.h",
6805 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07006806 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006807)
6808
6809xnnpack_unit_test(
Marat Dukhan188d1042020-11-24 23:39:40 -08006810 name = "depth_to_space_nchw2nhwc_test",
Artsiom Ablavatski0f1dc182020-11-05 19:21:50 -08006811 srcs = [
Marat Dukhan188d1042020-11-24 23:39:40 -08006812 "test/depth-to-space-nchw2nhwc.cc",
Artsiom Ablavatski0f1dc182020-11-05 19:21:50 -08006813 "test/depth-to-space-operator-tester.h",
6814 ] + OPERATOR_TEST_PARAMS_HDRS,
6815 deps = OPERATOR_TEST_DEPS,
6816)
6817
6818xnnpack_unit_test(
Marat Dukhan0e521172020-11-25 13:10:04 -08006819 name = "depth_to_space_nhwc_test",
6820 srcs = [
6821 "test/depth-to-space-nhwc.cc",
6822 "test/depth-to-space-operator-tester.h",
6823 ] + OPERATOR_TEST_PARAMS_HDRS,
6824 deps = OPERATOR_TEST_DEPS,
6825)
6826
6827xnnpack_unit_test(
Marat Dukhan69180502019-12-06 15:00:31 -08006828 name = "divide_nd_test",
6829 srcs = [
6830 "test/binary-elementwise-operator-tester.h",
6831 "test/divide-nd.cc",
6832 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006833 deps = OPERATOR_TEST_DEPS,
Marat Dukhan69180502019-12-06 15:00:31 -08006834)
6835
6836xnnpack_unit_test(
Marat Dukhanb6bd4bc2020-12-01 17:01:40 -08006837 name = "elu_nc_test",
6838 srcs = [
6839 "test/elu-nc.cc",
6840 "test/elu-operator-tester.h",
6841 ],
6842 deps = OPERATOR_TEST_DEPS,
6843)
6844
6845xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006846 name = "fully_connected_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006847 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006848 "test/fully-connected-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006849 "test/fully-connected-operator-tester.h",
6850 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006851 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006852)
6853
6854xnnpack_unit_test(
Marat Dukhan64e52512020-06-09 13:41:16 -07006855 name = "floor_nc_test",
6856 srcs = [
6857 "test/floor-nc.cc",
6858 "test/floor-operator-tester.h",
6859 ],
6860 deps = OPERATOR_TEST_DEPS,
6861)
6862
6863xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006864 name = "global_average_pooling_nwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006865 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006866 "test/global-average-pooling-nwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006867 "test/global-average-pooling-operator-tester.h",
Marat Dukhanef61d022020-06-19 13:54:49 -07006868 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07006869 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006870)
6871
6872xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006873 name = "global_average_pooling_ncw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006874 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006875 "test/global-average-pooling-ncw.cc",
6876 "test/global-average-pooling-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006877 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006878 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006879)
6880
6881xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006882 name = "hardswish_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006883 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006884 "test/hardswish-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006885 "test/hardswish-operator-tester.h",
6886 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006887 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006888)
6889
6890xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006891 name = "leaky_relu_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006892 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006893 "test/leaky-relu-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006894 "test/leaky-relu-operator-tester.h",
6895 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006896 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006897)
6898
6899xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006900 name = "max_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006901 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006902 "test/max-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006903 "test/max-pooling-operator-tester.h",
6904 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07006905 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006906)
6907
6908xnnpack_unit_test(
Marat Dukhan79e7f842019-12-05 14:35:50 -08006909 name = "maximum_nd_test",
6910 srcs = [
6911 "test/binary-elementwise-operator-tester.h",
6912 "test/maximum-nd.cc",
6913 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006914 deps = OPERATOR_TEST_DEPS,
Marat Dukhan79e7f842019-12-05 14:35:50 -08006915)
6916
6917xnnpack_unit_test(
6918 name = "minimum_nd_test",
6919 srcs = [
6920 "test/binary-elementwise-operator-tester.h",
6921 "test/minimum-nd.cc",
6922 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006923 deps = OPERATOR_TEST_DEPS,
Marat Dukhan79e7f842019-12-05 14:35:50 -08006924)
6925
6926xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006927 name = "multiply_nd_test",
Marat Dukhanca2733c2019-11-15 23:21:17 -08006928 srcs = [
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08006929 "test/binary-elementwise-operator-tester.h",
Marat Dukhanefc47b82019-11-18 09:25:38 -08006930 "test/multiply-nd.cc",
Marat Dukhanca2733c2019-11-15 23:21:17 -08006931 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006932 deps = OPERATOR_TEST_DEPS,
Marat Dukhanca2733c2019-11-15 23:21:17 -08006933)
6934
6935xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07006936 name = "negate_nc_test",
6937 srcs = [
6938 "test/negate-nc.cc",
6939 "test/negate-operator-tester.h",
6940 ],
6941 deps = OPERATOR_TEST_DEPS,
6942)
6943
6944xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006945 name = "prelu_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006946 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006947 "test/prelu-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006948 "test/prelu-operator-tester.h",
6949 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07006950 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006951)
6952
6953xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006954 name = "resize_bilinear_nhwc_test",
Marat Dukhan69722492019-11-11 19:55:50 -08006955 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006956 "test/resize-bilinear-nhwc.cc",
Marat Dukhan69722492019-11-11 19:55:50 -08006957 "test/resize-bilinear-operator-tester.h",
6958 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07006959 deps = OPERATOR_TEST_DEPS,
Marat Dukhan69722492019-11-11 19:55:50 -08006960)
6961
6962xnnpack_unit_test(
Artsiom Ablavatski97918102020-10-27 15:52:59 -07006963 name = "resize_bilinear_nchw_test",
6964 srcs = [
6965 "test/resize-bilinear-nchw.cc",
6966 "test/resize-bilinear-operator-tester.h",
6967 ] + OPERATOR_TEST_PARAMS_HDRS,
6968 deps = OPERATOR_TEST_DEPS,
6969)
6970
6971xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006972 name = "sigmoid_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006973 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006974 "test/sigmoid-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006975 "test/sigmoid-operator-tester.h",
6976 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006977 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006978)
6979
6980xnnpack_unit_test(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08006981 name = "softmax_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006982 srcs = [
Marat Dukhanfd8e6892020-01-27 15:25:25 -08006983 "test/softmax-nc.cc",
6984 "test/softmax-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006985 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006986 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006987)
6988
6989xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07006990 name = "square_nc_test",
6991 srcs = [
6992 "test/square-nc.cc",
6993 "test/square-operator-tester.h",
6994 ],
6995 deps = OPERATOR_TEST_DEPS,
6996)
6997
6998xnnpack_unit_test(
Marat Dukhan6804bbd2020-06-30 19:26:11 -07006999 name = "square_root_nc_test",
7000 srcs = [
7001 "test/square-root-nc.cc",
7002 "test/square-root-operator-tester.h",
7003 ],
7004 deps = OPERATOR_TEST_DEPS,
7005)
7006
7007xnnpack_unit_test(
Marat Dukhanf7399262020-06-05 10:58:44 -07007008 name = "squared_difference_nd_test",
7009 srcs = [
7010 "test/binary-elementwise-operator-tester.h",
7011 "test/squared-difference-nd.cc",
7012 ],
7013 deps = OPERATOR_TEST_DEPS,
7014)
7015
7016xnnpack_unit_test(
Marat Dukhan05f3f6d2019-12-03 15:13:53 -08007017 name = "subtract_nd_test",
7018 srcs = [
7019 "test/binary-elementwise-operator-tester.h",
7020 "test/subtract-nd.cc",
7021 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07007022 deps = OPERATOR_TEST_DEPS,
Marat Dukhan05f3f6d2019-12-03 15:13:53 -08007023)
7024
7025xnnpack_unit_test(
Marat Dukhan64e52512020-06-09 13:41:16 -07007026 name = "truncation_nc_test",
7027 srcs = [
7028 "test/truncation-nc.cc",
7029 "test/truncation-operator-tester.h",
7030 ],
7031 deps = OPERATOR_TEST_DEPS,
7032)
7033
7034xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08007035 name = "unpooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007036 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08007037 "test/unpooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007038 "test/unpooling-operator-tester.h",
7039 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07007040 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07007041)
7042
Chao Mei6ddfc602020-05-13 22:29:36 -07007043############################### Misc unit tests ###############################
7044
7045xnnpack_unit_test(
7046 name = "memory_planner_test",
7047 srcs = [
7048 "test/memory-planner-test.cc",
7049 ],
7050 deps = [
7051 ":XNNPACK",
7052 ":memory_planner",
7053 ],
7054)
7055
XNNPACK Teamab8c4c82020-10-09 08:05:51 -07007056xnnpack_unit_test(
7057 name = "subgraph_nchw_test",
7058 srcs = [
7059 "src/xnnpack/subgraph.h",
7060 "test/subgraph-nchw.cc",
7061 "test/subgraph-tester.h",
7062 ],
7063 deps = [
7064 ":XNNPACK",
7065 ],
7066)
7067
Marat Dukhan08c4a432019-10-03 09:29:21 -07007068############################# Build configurations #############################
7069
Marat Dukhanb8642352019-10-30 15:43:02 -07007070# Enables usage of assembly kernels.
Marat Dukhan08c4a432019-10-03 09:29:21 -07007071config_setting(
Marat Dukhanb8642352019-10-30 15:43:02 -07007072 name = "xnn_enable_assembly_explicit_true",
7073 define_values = {"xnn_enable_assembly": "true"},
7074)
7075
7076# Disables usage of assembly kernels.
7077config_setting(
7078 name = "xnn_enable_assembly_explicit_false",
7079 define_values = {"xnn_enable_assembly": "false"},
7080)
7081
Marat Dukhan9de90e02020-06-18 16:04:12 -07007082# Enables usage of sparse inference.
7083config_setting(
7084 name = "xnn_enable_sparse_explicit_true",
7085 define_values = {"xnn_enable_sparse": "true"},
7086)
7087
7088# Disables usage of sparse inference.
7089config_setting(
7090 name = "xnn_enable_sparse_explicit_false",
7091 define_values = {"xnn_enable_sparse": "false"},
7092)
7093
Marat Dukhan05702cf2020-03-26 15:41:33 -07007094# Disables usage of HMP-aware optimizations.
7095config_setting(
7096 name = "xnn_enable_hmp_explicit_false",
7097 define_values = {"xnn_enable_hmp": "false"},
7098)
7099
Chao Mei6ddfc602020-05-13 22:29:36 -07007100# Enable usage of optimized memory allocation
7101config_setting(
7102 name = "xnn_enable_memopt_explicit_true",
7103 define_values = {"xnnpack_enable_memopt": "true"},
7104)
7105
7106# Disable usage of optimized memory allocation
7107config_setting(
7108 name = "xnn_enable_memopt_explicit_false",
7109 define_values = {"xnnpack_enable_memopt": "false"},
7110)
7111
Marat Dukhanb8642352019-10-30 15:43:02 -07007112# Builds with -c dbg
7113config_setting(
7114 name = "debug_build",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007115 values = {
Marat Dukhanb8642352019-10-30 15:43:02 -07007116 "compilation_mode": "dbg",
7117 },
7118)
7119
7120# Builds with -c opt
7121config_setting(
7122 name = "optimized_build",
7123 values = {
7124 "compilation_mode": "opt",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007125 },
7126)
7127
7128config_setting(
Marat Dukhanb8642352019-10-30 15:43:02 -07007129 name = "linux_k8",
7130 values = {"cpu": "k8"},
7131)
7132
7133config_setting(
Marat Dukhan582094e2020-04-30 17:21:25 -07007134 name = "linux_arm",
7135 values = {"cpu": "arm"},
Marat Dukhan4e45e662019-10-03 15:40:24 -07007136)
7137
7138config_setting(
Marat Dukhanf0bd4de2020-06-15 15:53:19 -07007139 name = "linux_armeabi",
7140 values = {"cpu": "armeabi"},
7141)
7142
7143config_setting(
Terry Heo68eef3f2020-04-13 22:53:52 -07007144 name = "linux_armhf",
7145 values = {"cpu": "armhf"},
7146)
7147
7148config_setting(
Marat Dukhana720e932020-06-10 13:01:11 -07007149 name = "linux_armv7a",
7150 values = {"cpu": "armv7a"},
7151)
7152
7153config_setting(
Marat Dukhan582094e2020-04-30 17:21:25 -07007154 name = "linux_aarch64",
7155 values = {"cpu": "aarch64"},
7156)
7157
7158config_setting(
Marat Dukhan08c4a432019-10-03 09:29:21 -07007159 name = "android",
7160 values = {"crosstool_top": "//external:android/crosstool"},
7161)
7162
7163config_setting(
7164 name = "android_armv7",
7165 values = {
7166 "crosstool_top": "//external:android/crosstool",
7167 "cpu": "armeabi-v7a",
7168 },
7169)
7170
7171config_setting(
7172 name = "android_arm64",
7173 values = {
7174 "crosstool_top": "//external:android/crosstool",
7175 "cpu": "arm64-v8a",
7176 },
7177)
7178
7179config_setting(
7180 name = "android_x86",
7181 values = {
7182 "crosstool_top": "//external:android/crosstool",
7183 "cpu": "x86",
7184 },
7185)
7186
7187config_setting(
7188 name = "android_x86_64",
7189 values = {
7190 "crosstool_top": "//external:android/crosstool",
7191 "cpu": "x86_64",
7192 },
7193)
7194
7195config_setting(
Marat Dukhan10a38082020-04-17 03:58:35 -07007196 name = "windows_x86_64",
7197 values = {"cpu": "x64_windows"},
Marat Dukhan9fe932e2020-04-11 17:14:15 -07007198)
7199
7200config_setting(
Marat Dukhan10a38082020-04-17 03:58:35 -07007201 name = "windows_x86_64_clang",
7202 values = {
7203 "compiler": "clang-cl",
7204 "cpu": "x64_windows",
7205 },
7206)
7207
7208config_setting(
7209 name = "windows_x86_64_mingw",
7210 values = {
7211 "compiler": "mingw-gcc",
7212 "cpu": "x64_windows",
7213 },
7214)
7215
7216config_setting(
7217 name = "windows_x86_64_msys",
7218 values = {
7219 "compiler": "msys-gcc",
7220 "cpu": "x64_windows",
7221 },
Marat Dukhan9fe932e2020-04-11 17:14:15 -07007222)
7223
7224config_setting(
Marat Dukhan885ca242019-10-07 09:17:32 -07007225 name = "macos_x86_64",
7226 values = {
7227 "apple_platform_type": "macos",
7228 "cpu": "darwin",
7229 },
7230)
7231
7232config_setting(
Marat Dukhan08c4a432019-10-03 09:29:21 -07007233 name = "emscripten",
Marat Dukhan1a729ec2019-10-07 09:31:44 -07007234 values = {"crosstool_top": "//toolchain:emscripten"},
Marat Dukhan08c4a432019-10-03 09:29:21 -07007235)
7236
7237config_setting(
7238 name = "emscripten_wasm",
7239 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07007240 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007241 "cpu": "wasm",
7242 },
7243)
7244
7245config_setting(
7246 name = "emscripten_wasmsimd",
7247 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07007248 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007249 "cpu": "wasm",
Marat Dukhan81c62602020-05-29 13:22:49 -07007250 "copt": "-msimd128",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007251 },
7252)
7253
7254config_setting(
Marat Dukhan1498d1d2020-02-11 20:00:05 -08007255 name = "ios_armv7",
7256 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08007257 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08007258 "cpu": "ios_armv7",
7259 },
7260)
7261
7262config_setting(
7263 name = "ios_arm64",
7264 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08007265 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08007266 "cpu": "ios_arm64",
7267 },
7268)
7269
7270config_setting(
7271 name = "ios_arm64e",
7272 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08007273 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08007274 "cpu": "ios_arm64e",
7275 },
7276)
7277
7278config_setting(
7279 name = "ios_x86",
7280 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08007281 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08007282 "cpu": "ios_i386",
7283 },
7284)
7285
7286config_setting(
7287 name = "ios_x86_64",
7288 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08007289 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08007290 "cpu": "ios_x86_64",
7291 },
7292)
7293
7294config_setting(
7295 name = "watchos_armv7k",
7296 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08007297 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08007298 "cpu": "watchos_armv7k",
7299 },
7300)
7301
7302config_setting(
7303 name = "watchos_arm64_32",
7304 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08007305 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08007306 "cpu": "watchos_arm64_32",
7307 },
7308)
7309
7310config_setting(
7311 name = "watchos_x86",
7312 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08007313 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08007314 "cpu": "watchos_i386",
7315 },
7316)
7317
7318config_setting(
7319 name = "watchos_x86_64",
7320 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08007321 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08007322 "cpu": "watchos_x86_64",
7323 },
7324)
7325
7326config_setting(
7327 name = "tvos_arm64",
7328 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08007329 "apple_platform_type": "tvos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08007330 "cpu": "tvos_arm64",
7331 },
7332)
7333
7334config_setting(
7335 name = "tvos_x86_64",
7336 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08007337 "apple_platform_type": "tvos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08007338 "cpu": "tvos_x86_64",
7339 },
7340)