blob: 25414a0575cfe05c854d9b5b097854a104f4dede [file] [log] [blame]
Frank Barchard1a953052020-11-16 18:44:58 -08001# Copyright 2020 Google LLC
Marat Dukhan08c4a432019-10-03 09:29:21 -07002#
3# This source code is licensed under the BSD-style license found in the
4# LICENSE file in the root directory of this source tree.
5#
6# Description:
7# XNNPACK - optimized floating-point neural network operators library
8
Marat Dukhan10a38082020-04-17 03:58:35 -07009load(":build_defs.bzl", "xnnpack_aggregate_library", "xnnpack_benchmark", "xnnpack_binary", "xnnpack_cc_library", "xnnpack_gcc_std_copts", "xnnpack_min_size_copts", "xnnpack_msvc_std_copts", "xnnpack_optional_armcl_copts", "xnnpack_optional_armcl_deps", "xnnpack_optional_dnnl_copts", "xnnpack_optional_dnnl_deps", "xnnpack_optional_gemmlowp_copts", "xnnpack_optional_gemmlowp_deps", "xnnpack_optional_ruy_copts", "xnnpack_optional_ruy_deps", "xnnpack_optional_tflite_copts", "xnnpack_optional_tflite_deps", "xnnpack_std_cxxopts", "xnnpack_unit_test", "xnnpack_visibility")
Marat Dukhan69c3f2c2019-11-06 12:30:01 -080010
Marat Dukhan08c4a432019-10-03 09:29:21 -070011licenses(["notice"])
12
13exports_files(["LICENSE"])
14
Marat Dukhan1b354632020-03-23 12:50:22 -070015OPERATOR_BENCHMARK_DEPS = [
16 ":XNNPACK",
17 ":bench_utils",
18 "@cpuinfo",
Frank Barchard0c849732020-06-12 13:31:32 -070019 "@FP16",
Marat Dukhan1b354632020-03-23 12:50:22 -070020 "@pthreadpool",
21]
22
Marat Dukhan08c4a432019-10-03 09:29:21 -070023MICROKERNEL_BENCHMARK_DEPS = [
24 ":ukernels",
25 ":bench_utils",
Frank Barchard7e955972019-10-11 10:34:25 -070026 ":enable_assembly",
Marat Dukhan08c4a432019-10-03 09:29:21 -070027 "@cpuinfo",
28 "@FP16",
29 "@pthreadpool",
30]
31
Marat Dukhan6adff4e2019-10-14 18:32:07 -070032ACCURACY_EVAL_DEPS = [
33 ":XNNPACK",
34 ":ukernels",
35 "@FP16",
36 "@pthreadpool",
37]
38
Marat Dukhan08c4a432019-10-03 09:29:21 -070039MICROKERNEL_TEST_DEPS = [
Marat Dukhan33fcf782020-05-24 14:27:15 -070040 ":ukernels_test_mode",
Frank Barchard7e955972019-10-11 10:34:25 -070041 ":enable_assembly",
Marat Dukhan08c4a432019-10-03 09:29:21 -070042 "@cpuinfo",
43 "@FP16",
44 "@pthreadpool",
45]
46
Marat Dukhan1b354632020-03-23 12:50:22 -070047OPERATOR_TEST_DEPS = [
Marat Dukhan33fcf782020-05-24 14:27:15 -070048 ":XNNPACK_test_mode",
Marat Dukhan1b354632020-03-23 12:50:22 -070049 "@pthreadpool",
50 "@FP16",
51]
52
Marat Dukhan08c4a432019-10-03 09:29:21 -070053OPERATOR_SRCS = [
Marat Dukhane8265432020-04-28 18:42:59 -070054 "src/operators/argmax-pooling-nhwc.c",
55 "src/operators/average-pooling-nhwc.c",
56 "src/operators/binary-elementwise-nd.c",
Marat Dukhane8265432020-04-28 18:42:59 -070057 "src/operators/channel-shuffle-nc.c",
Marat Dukhan065b11e2020-05-22 09:49:41 -070058 "src/operators/constant-pad-nd.c",
Marat Dukhane8265432020-04-28 18:42:59 -070059 "src/operators/convolution-nchw.c",
60 "src/operators/convolution-nhwc.c",
61 "src/operators/deconvolution-nhwc.c",
Marat Dukhan13b68f22020-11-12 11:55:19 -080062 "src/operators/depth-to-space-nchw2nhwc.c",
Marat Dukhan0e521172020-11-25 13:10:04 -080063 "src/operators/depth-to-space-nhwc.c",
Marat Dukhane8265432020-04-28 18:42:59 -070064 "src/operators/fully-connected-nc.c",
65 "src/operators/global-average-pooling-ncw.c",
66 "src/operators/global-average-pooling-nwc.c",
Marat Dukhane8265432020-04-28 18:42:59 -070067 "src/operators/leaky-relu-nc.c",
68 "src/operators/max-pooling-nhwc.c",
69 "src/operators/prelu-nc.c",
Artsiom Ablavatski97918102020-10-27 15:52:59 -070070 "src/operators/resize-bilinear-nchw.c",
Marat Dukhane8265432020-04-28 18:42:59 -070071 "src/operators/resize-bilinear-nhwc.c",
72 "src/operators/sigmoid-nc.c",
73 "src/operators/softmax-nc.c",
Marat Dukhanc3065f52020-06-04 13:33:32 -070074 "src/operators/unary-elementwise-nc.c",
Marat Dukhane8265432020-04-28 18:42:59 -070075 "src/operators/unpooling-nhwc.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -070076]
77
Marat Dukhan95e8b7a2020-06-03 12:46:26 -070078SUBGRAPH_SRCS = [
Marat Dukhan5fab4092020-06-10 01:28:28 -070079 "src/subgraph/abs.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -070080 "src/subgraph/add2.c",
81 "src/subgraph/argmax-pooling-2d.c",
82 "src/subgraph/average-pooling-2d.c",
Marat Dukhan5fab4092020-06-10 01:28:28 -070083 "src/subgraph/bankers-rounding.c",
84 "src/subgraph/ceiling.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -070085 "src/subgraph/clamp.c",
86 "src/subgraph/convolution-2d.c",
87 "src/subgraph/deconvolution-2d.c",
Artsiom Ablavatskibbe85062020-11-05 14:07:37 -080088 "src/subgraph/depth-to-space.c",
Frank Barchard9cef5ea2020-11-18 14:52:08 -080089 "src/subgraph/depthwise-convolution-2d.c",
Marat Dukhan9d3a4592020-06-05 16:52:42 -070090 "src/subgraph/divide.c",
Marat Dukhan5fab4092020-06-10 01:28:28 -070091 "src/subgraph/floor.c",
Frank Barchard04336c12020-10-22 16:48:55 -070092 "src/subgraph/fully-connected.c",
Marat Dukhana059b7d2020-06-11 11:41:27 -070093 "src/subgraph/global-average-pooling-2d.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -070094 "src/subgraph/hardswish.c",
Marat Dukhan5bbebac2020-06-10 19:42:15 -070095 "src/subgraph/leaky-relu.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -070096 "src/subgraph/max-pooling-2d.c",
Marat Dukhan9d3a4592020-06-05 16:52:42 -070097 "src/subgraph/maximum2.c",
98 "src/subgraph/minimum2.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -070099 "src/subgraph/multiply2.c",
Marat Dukhan5fab4092020-06-10 01:28:28 -0700100 "src/subgraph/negate.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -0700101 "src/subgraph/prelu.c",
102 "src/subgraph/sigmoid.c",
103 "src/subgraph/softmax.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700104 "src/subgraph/square-root.c",
105 "src/subgraph/square.c",
106 "src/subgraph/squared-difference.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -0700107 "src/subgraph/static-constant-pad.c",
Marat Dukhand27202d2020-07-09 23:43:40 -0700108 "src/subgraph/static-reshape.c",
Marat Dukhanaff24e22020-07-23 01:43:58 -0700109 "src/subgraph/static-resize-bilinear-2d.c",
Marat Dukhan9d3a4592020-06-05 16:52:42 -0700110 "src/subgraph/subtract.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -0700111 "src/subgraph/unpooling-2d.c",
112]
113
Marat Dukhan3a77ea72019-12-23 12:10:24 -0800114TABLE_SRCS = [
115 "src/tables/exp2-k-over-64.c",
116 "src/tables/exp2-k-over-2048.c",
Marat Dukhande390d42020-11-29 19:32:18 -0800117 "src/tables/exp2minus-k-over-4.c",
118 "src/tables/exp2minus-k-over-8.c",
Marat Dukhanc60742b2020-11-23 12:33:27 -0800119 "src/tables/exp2minus-k-over-16.c",
Marat Dukhan1f256fc2020-09-24 21:27:14 -0700120 "src/tables/exp2minus-k-over-64.c",
121 "src/tables/exp2minus-k-over-2048.c",
Marat Dukhan3a77ea72019-12-23 12:10:24 -0800122]
123
Marat Dukhan08c4a432019-10-03 09:29:21 -0700124SCALAR_UKERNELS = [
Marat Dukhan329da642019-11-19 21:44:39 -0800125 "src/f32-argmaxpool/4x-scalar-c1.c",
Marat Dukhan1e782c42019-11-21 17:02:40 -0800126 "src/f32-argmaxpool/9p8x-scalar-c1.c",
Marat Dukhan329da642019-11-19 21:44:39 -0800127 "src/f32-argmaxpool/9x-scalar-c1.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700128 "src/f32-avgpool/9p8x-minmax-scalar-c1.c",
129 "src/f32-avgpool/9x-minmax-scalar-c1.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -0700130 "src/f32-clamp/gen/scalar-x1.c",
131 "src/f32-clamp/gen/scalar-x2.c",
132 "src/f32-clamp/gen/scalar-x4.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700133 "src/f32-conv-hwc/3x3s2p0p1c3x4-scalar-1x1.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700134 "src/f32-conv-hwc/3x3s2p1c3x4-scalar-1x1.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -0700135 "src/f32-conv-hwc2chw/3x3s2p1c3x4-scalar-1x1.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700136 "src/f32-dwconv/gen/up1x4-minmax-scalar-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700137 "src/f32-dwconv/gen/up1x4-minmax-scalar.c",
138 "src/f32-dwconv/gen/up1x4-scalar-acc2.c",
139 "src/f32-dwconv/gen/up1x4-scalar.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700140 "src/f32-dwconv/gen/up1x9-minmax-scalar-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700141 "src/f32-dwconv/gen/up1x9-minmax-scalar.c",
142 "src/f32-dwconv/gen/up1x9-scalar-acc2.c",
143 "src/f32-dwconv/gen/up1x9-scalar.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700144 "src/f32-dwconv/gen/up1x25-minmax-scalar-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700145 "src/f32-dwconv/gen/up1x25-minmax-scalar.c",
146 "src/f32-dwconv/gen/up1x25-scalar-acc2.c",
147 "src/f32-dwconv/gen/up1x25-scalar.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700148 "src/f32-dwconv/gen/up2x4-minmax-scalar-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700149 "src/f32-dwconv/gen/up2x4-minmax-scalar.c",
150 "src/f32-dwconv/gen/up2x4-scalar-acc2.c",
151 "src/f32-dwconv/gen/up2x4-scalar.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700152 "src/f32-dwconv/gen/up2x9-minmax-scalar-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700153 "src/f32-dwconv/gen/up2x9-minmax-scalar.c",
154 "src/f32-dwconv/gen/up2x9-scalar-acc2.c",
155 "src/f32-dwconv/gen/up2x9-scalar.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700156 "src/f32-dwconv/gen/up2x25-minmax-scalar-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700157 "src/f32-dwconv/gen/up2x25-minmax-scalar.c",
158 "src/f32-dwconv/gen/up2x25-scalar-acc2.c",
159 "src/f32-dwconv/gen/up2x25-scalar.c",
Frank Barchard35db7d02020-10-26 13:37:34 -0700160 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-scalar-1x1-acc2.c",
161 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-scalar-1x1-acc3.c",
162 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-scalar-1x1-acc4.c",
Marat Dukhan91249d22020-10-24 12:02:51 -0700163 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-scalar-1x1.c",
Frank Barchard35db7d02020-10-26 13:37:34 -0700164 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-scalar-2x1-acc2.c",
Marat Dukhan91249d22020-10-24 12:02:51 -0700165 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-scalar-2x1.c",
166 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-scalar-3x1.c",
167 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-scalar-4x1.c",
168 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-scalar-5x1.c",
169 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-scalar-6x1.c",
Marat Dukhancf5b3c32020-10-25 19:21:10 -0700170 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-scalar-1x1-acc2.c",
171 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-scalar-1x1-acc3.c",
172 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-scalar-1x1-acc4.c",
Frank Barchard35db7d02020-10-26 13:37:34 -0700173 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-scalar-1x1.c",
Marat Dukhancf5b3c32020-10-25 19:21:10 -0700174 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-scalar-2x1-acc2.c",
Frank Barchard35db7d02020-10-26 13:37:34 -0700175 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-scalar-2x1.c",
176 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-scalar-3x1.c",
177 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-scalar-4x1.c",
Marat Dukhanc4efb002020-10-25 23:14:47 -0700178 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-scalar-1x1-acc2.c",
179 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-scalar-1x1-acc3.c",
180 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-scalar-1x1-acc4.c",
181 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-scalar-1x1-acc5.c",
Frank Barchard35db7d02020-10-26 13:37:34 -0700182 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-scalar-1x1.c",
Marat Dukhanc4efb002020-10-25 23:14:47 -0700183 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-scalar-2x1-acc2.c",
184 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-scalar-2x1-acc3.c",
Frank Barchard35db7d02020-10-26 13:37:34 -0700185 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-scalar-2x1.c",
Marat Dukhanc4efb002020-10-25 23:14:47 -0700186 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-scalar-3x1-acc2.c",
Frank Barchard35db7d02020-10-26 13:37:34 -0700187 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-scalar-3x1.c",
Marat Dukhan29c0c332020-10-28 22:11:00 -0700188 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-scalar-1x1-acc2.c",
189 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-scalar-1x1-acc3.c",
190 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-scalar-1x1-acc4.c",
191 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-scalar-1x1-acc5.c",
192 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-scalar-1x1.c",
193 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-scalar-2x1-acc2.c",
194 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-scalar-2x1-acc3.c",
195 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-scalar-2x1.c",
196 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-scalar-3x1-acc2.c",
197 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-scalar-3x1.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -0700198 "src/f32-gavgpool-cw/scalar-x1.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700199 "src/f32-gavgpool/7p7x-minmax-scalar-c1.c",
200 "src/f32-gavgpool/7x-minmax-scalar-c1.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700201 "src/f32-gemm/gen-inc/1x4inc-minmax-scalar.c",
202 "src/f32-gemm/gen-inc/2x4inc-minmax-scalar.c",
203 "src/f32-gemm/gen-inc/4x4inc-minmax-scalar.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700204 "src/f32-gemm/gen/1x4-minmax-scalar.c",
205 "src/f32-gemm/gen/1x4-relu-scalar.c",
206 "src/f32-gemm/gen/1x4-scalar.c",
207 "src/f32-gemm/gen/2x4-minmax-scalar.c",
208 "src/f32-gemm/gen/2x4-relu-scalar.c",
209 "src/f32-gemm/gen/2x4-scalar.c",
210 "src/f32-gemm/gen/4x2-minmax-scalar.c",
211 "src/f32-gemm/gen/4x2-relu-scalar.c",
212 "src/f32-gemm/gen/4x2-scalar.c",
213 "src/f32-gemm/gen/4x4-minmax-scalar.c",
214 "src/f32-gemm/gen/4x4-relu-scalar.c",
215 "src/f32-gemm/gen/4x4-scalar.c",
Marat Dukhanefaac272020-07-16 16:58:07 -0700216 "src/f32-hswish/gen/hswish-scalar-x1.c",
217 "src/f32-hswish/gen/hswish-scalar-x2.c",
218 "src/f32-hswish/gen/hswish-scalar-x4.c",
XNNPACK Team21432672020-10-19 19:58:48 -0700219 "src/f32-ibilinear-chw/gen/scalar-p1.c",
220 "src/f32-ibilinear-chw/gen/scalar-p2.c",
221 "src/f32-ibilinear-chw/gen/scalar-p4.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700222 "src/f32-ibilinear/gen/scalar-c1.c",
223 "src/f32-ibilinear/gen/scalar-c2.c",
224 "src/f32-ibilinear/gen/scalar-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700225 "src/f32-igemm/gen/1x4-minmax-scalar.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700226 "src/f32-igemm/gen/1x4-relu-scalar.c",
227 "src/f32-igemm/gen/1x4-scalar.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700228 "src/f32-igemm/gen/2x4-minmax-scalar.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700229 "src/f32-igemm/gen/2x4-relu-scalar.c",
230 "src/f32-igemm/gen/2x4-scalar.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700231 "src/f32-igemm/gen/4x2-minmax-scalar.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700232 "src/f32-igemm/gen/4x2-relu-scalar.c",
233 "src/f32-igemm/gen/4x2-scalar.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700234 "src/f32-igemm/gen/4x4-minmax-scalar.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700235 "src/f32-igemm/gen/4x4-relu-scalar.c",
236 "src/f32-igemm/gen/4x4-scalar.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700237 "src/f32-maxpool/9p8x-minmax-scalar-c1.c",
238 "src/f32-pavgpool/9p8x-minmax-scalar-c1.c",
239 "src/f32-pavgpool/9x-minmax-scalar-c1.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700240 "src/f32-ppmm/gen/2x4-minmax-scalar.c",
241 "src/f32-ppmm/gen/3x3-minmax-scalar.c",
242 "src/f32-ppmm/gen/4x2-minmax-scalar.c",
243 "src/f32-ppmm/gen/4x4-minmax-scalar.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800244 "src/f32-prelu/gen/scalar-2x1.c",
245 "src/f32-prelu/gen/scalar-2x4.c",
Marat Dukhanf46f6752020-01-21 11:03:49 -0800246 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x1.c",
Marat Dukhanf46f6752020-01-21 11:03:49 -0800247 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x2-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700248 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x2.c",
Marat Dukhanf46f6752020-01-21 11:03:49 -0800249 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x4-acc2.c",
250 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x4-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700251 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x4.c",
Marat Dukhanf46f6752020-01-21 11:03:49 -0800252 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x1.c",
Marat Dukhanf46f6752020-01-21 11:03:49 -0800253 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x2-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700254 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x2.c",
Marat Dukhanf46f6752020-01-21 11:03:49 -0800255 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x4-acc2.c",
256 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x4-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700257 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x4.c",
Frank Barchardfb158e22020-07-15 16:10:10 -0700258 "src/f32-relu/gen/scalar-x1.c",
259 "src/f32-relu/gen/scalar-x2.c",
260 "src/f32-relu/gen/scalar-x4.c",
Frank Barchard37297a62020-08-27 12:15:10 -0700261 "src/f32-relu/gen/scalar-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700262 "src/f32-rmax/scalar.c",
Marat Dukhan3a77ea72019-12-23 12:10:24 -0800263 "src/f32-sigmoid/gen/scalar-lut64-p2-div-x1.c",
264 "src/f32-sigmoid/gen/scalar-lut64-p2-div-x2.c",
265 "src/f32-sigmoid/gen/scalar-lut64-p2-div-x4.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700266 "src/f32-sigmoid/gen/scalar-lut2048-p1-div-x1.c",
267 "src/f32-sigmoid/gen/scalar-lut2048-p1-div-x2.c",
268 "src/f32-sigmoid/gen/scalar-lut2048-p1-div-x4.c",
Marat Dukhan3a77ea72019-12-23 12:10:24 -0800269 "src/f32-sigmoid/gen/scalar-p5-div-x1.c",
270 "src/f32-sigmoid/gen/scalar-p5-div-x2.c",
271 "src/f32-sigmoid/gen/scalar-p5-div-x4.c",
Marat Dukhan355ab432020-04-09 19:01:52 -0700272 "src/f32-spmm/gen/1x1-minmax-scalar-pipelined.c",
273 "src/f32-spmm/gen/1x1-minmax-scalar.c",
274 "src/f32-spmm/gen/2x1-minmax-scalar-pipelined.c",
275 "src/f32-spmm/gen/2x1-minmax-scalar.c",
276 "src/f32-spmm/gen/4x1-minmax-scalar-pipelined.c",
277 "src/f32-spmm/gen/4x1-minmax-scalar.c",
278 "src/f32-spmm/gen/8x1-minmax-scalar-pipelined.c",
279 "src/f32-spmm/gen/8x1-minmax-scalar.c",
280 "src/f32-spmm/gen/8x2-minmax-scalar.c",
281 "src/f32-spmm/gen/8x4-minmax-scalar.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700282 "src/f32-vbinary/gen/vadd-minmax-scalar-x1.c",
283 "src/f32-vbinary/gen/vadd-minmax-scalar-x2.c",
284 "src/f32-vbinary/gen/vadd-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700285 "src/f32-vbinary/gen/vadd-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700286 "src/f32-vbinary/gen/vadd-relu-scalar-x1.c",
287 "src/f32-vbinary/gen/vadd-relu-scalar-x2.c",
288 "src/f32-vbinary/gen/vadd-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700289 "src/f32-vbinary/gen/vadd-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700290 "src/f32-vbinary/gen/vadd-scalar-x1.c",
291 "src/f32-vbinary/gen/vadd-scalar-x2.c",
292 "src/f32-vbinary/gen/vadd-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700293 "src/f32-vbinary/gen/vadd-scalar-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700294 "src/f32-vbinary/gen/vaddc-minmax-scalar-x1.c",
295 "src/f32-vbinary/gen/vaddc-minmax-scalar-x2.c",
296 "src/f32-vbinary/gen/vaddc-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700297 "src/f32-vbinary/gen/vaddc-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700298 "src/f32-vbinary/gen/vaddc-relu-scalar-x1.c",
299 "src/f32-vbinary/gen/vaddc-relu-scalar-x2.c",
300 "src/f32-vbinary/gen/vaddc-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700301 "src/f32-vbinary/gen/vaddc-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700302 "src/f32-vbinary/gen/vaddc-scalar-x1.c",
303 "src/f32-vbinary/gen/vaddc-scalar-x2.c",
304 "src/f32-vbinary/gen/vaddc-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700305 "src/f32-vbinary/gen/vaddc-scalar-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700306 "src/f32-vbinary/gen/vdiv-minmax-scalar-x1.c",
307 "src/f32-vbinary/gen/vdiv-minmax-scalar-x2.c",
308 "src/f32-vbinary/gen/vdiv-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700309 "src/f32-vbinary/gen/vdiv-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700310 "src/f32-vbinary/gen/vdiv-relu-scalar-x1.c",
311 "src/f32-vbinary/gen/vdiv-relu-scalar-x2.c",
312 "src/f32-vbinary/gen/vdiv-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700313 "src/f32-vbinary/gen/vdiv-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700314 "src/f32-vbinary/gen/vdiv-scalar-x1.c",
315 "src/f32-vbinary/gen/vdiv-scalar-x2.c",
316 "src/f32-vbinary/gen/vdiv-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700317 "src/f32-vbinary/gen/vdiv-scalar-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700318 "src/f32-vbinary/gen/vdivc-minmax-scalar-x1.c",
319 "src/f32-vbinary/gen/vdivc-minmax-scalar-x2.c",
320 "src/f32-vbinary/gen/vdivc-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700321 "src/f32-vbinary/gen/vdivc-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700322 "src/f32-vbinary/gen/vdivc-relu-scalar-x1.c",
323 "src/f32-vbinary/gen/vdivc-relu-scalar-x2.c",
324 "src/f32-vbinary/gen/vdivc-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700325 "src/f32-vbinary/gen/vdivc-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700326 "src/f32-vbinary/gen/vdivc-scalar-x1.c",
327 "src/f32-vbinary/gen/vdivc-scalar-x2.c",
328 "src/f32-vbinary/gen/vdivc-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700329 "src/f32-vbinary/gen/vdivc-scalar-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800330 "src/f32-vbinary/gen/vmax-scalar-x1.c",
331 "src/f32-vbinary/gen/vmax-scalar-x2.c",
332 "src/f32-vbinary/gen/vmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700333 "src/f32-vbinary/gen/vmax-scalar-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800334 "src/f32-vbinary/gen/vmaxc-scalar-x1.c",
335 "src/f32-vbinary/gen/vmaxc-scalar-x2.c",
336 "src/f32-vbinary/gen/vmaxc-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700337 "src/f32-vbinary/gen/vmaxc-scalar-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800338 "src/f32-vbinary/gen/vmin-scalar-x1.c",
339 "src/f32-vbinary/gen/vmin-scalar-x2.c",
340 "src/f32-vbinary/gen/vmin-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700341 "src/f32-vbinary/gen/vmin-scalar-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800342 "src/f32-vbinary/gen/vminc-scalar-x1.c",
343 "src/f32-vbinary/gen/vminc-scalar-x2.c",
344 "src/f32-vbinary/gen/vminc-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700345 "src/f32-vbinary/gen/vminc-scalar-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700346 "src/f32-vbinary/gen/vmul-minmax-scalar-x1.c",
347 "src/f32-vbinary/gen/vmul-minmax-scalar-x2.c",
348 "src/f32-vbinary/gen/vmul-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700349 "src/f32-vbinary/gen/vmul-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700350 "src/f32-vbinary/gen/vmul-relu-scalar-x1.c",
351 "src/f32-vbinary/gen/vmul-relu-scalar-x2.c",
352 "src/f32-vbinary/gen/vmul-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700353 "src/f32-vbinary/gen/vmul-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700354 "src/f32-vbinary/gen/vmul-scalar-x1.c",
355 "src/f32-vbinary/gen/vmul-scalar-x2.c",
356 "src/f32-vbinary/gen/vmul-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700357 "src/f32-vbinary/gen/vmul-scalar-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700358 "src/f32-vbinary/gen/vmulc-minmax-scalar-x1.c",
359 "src/f32-vbinary/gen/vmulc-minmax-scalar-x2.c",
360 "src/f32-vbinary/gen/vmulc-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700361 "src/f32-vbinary/gen/vmulc-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700362 "src/f32-vbinary/gen/vmulc-relu-scalar-x1.c",
363 "src/f32-vbinary/gen/vmulc-relu-scalar-x2.c",
364 "src/f32-vbinary/gen/vmulc-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700365 "src/f32-vbinary/gen/vmulc-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700366 "src/f32-vbinary/gen/vmulc-scalar-x1.c",
367 "src/f32-vbinary/gen/vmulc-scalar-x2.c",
368 "src/f32-vbinary/gen/vmulc-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700369 "src/f32-vbinary/gen/vmulc-scalar-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700370 "src/f32-vbinary/gen/vrdivc-minmax-scalar-x1.c",
371 "src/f32-vbinary/gen/vrdivc-minmax-scalar-x2.c",
372 "src/f32-vbinary/gen/vrdivc-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700373 "src/f32-vbinary/gen/vrdivc-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700374 "src/f32-vbinary/gen/vrdivc-relu-scalar-x1.c",
375 "src/f32-vbinary/gen/vrdivc-relu-scalar-x2.c",
376 "src/f32-vbinary/gen/vrdivc-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700377 "src/f32-vbinary/gen/vrdivc-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700378 "src/f32-vbinary/gen/vrdivc-scalar-x1.c",
379 "src/f32-vbinary/gen/vrdivc-scalar-x2.c",
380 "src/f32-vbinary/gen/vrdivc-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700381 "src/f32-vbinary/gen/vrdivc-scalar-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700382 "src/f32-vbinary/gen/vrsubc-minmax-scalar-x1.c",
383 "src/f32-vbinary/gen/vrsubc-minmax-scalar-x2.c",
384 "src/f32-vbinary/gen/vrsubc-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700385 "src/f32-vbinary/gen/vrsubc-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700386 "src/f32-vbinary/gen/vrsubc-relu-scalar-x1.c",
387 "src/f32-vbinary/gen/vrsubc-relu-scalar-x2.c",
388 "src/f32-vbinary/gen/vrsubc-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700389 "src/f32-vbinary/gen/vrsubc-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700390 "src/f32-vbinary/gen/vrsubc-scalar-x1.c",
391 "src/f32-vbinary/gen/vrsubc-scalar-x2.c",
392 "src/f32-vbinary/gen/vrsubc-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700393 "src/f32-vbinary/gen/vrsubc-scalar-x8.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -0700394 "src/f32-vbinary/gen/vsqrdiff-scalar-x1.c",
395 "src/f32-vbinary/gen/vsqrdiff-scalar-x2.c",
396 "src/f32-vbinary/gen/vsqrdiff-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700397 "src/f32-vbinary/gen/vsqrdiff-scalar-x8.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -0700398 "src/f32-vbinary/gen/vsqrdiffc-scalar-x1.c",
399 "src/f32-vbinary/gen/vsqrdiffc-scalar-x2.c",
400 "src/f32-vbinary/gen/vsqrdiffc-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700401 "src/f32-vbinary/gen/vsqrdiffc-scalar-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700402 "src/f32-vbinary/gen/vsub-minmax-scalar-x1.c",
403 "src/f32-vbinary/gen/vsub-minmax-scalar-x2.c",
404 "src/f32-vbinary/gen/vsub-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700405 "src/f32-vbinary/gen/vsub-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700406 "src/f32-vbinary/gen/vsub-relu-scalar-x1.c",
407 "src/f32-vbinary/gen/vsub-relu-scalar-x2.c",
408 "src/f32-vbinary/gen/vsub-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700409 "src/f32-vbinary/gen/vsub-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700410 "src/f32-vbinary/gen/vsub-scalar-x1.c",
411 "src/f32-vbinary/gen/vsub-scalar-x2.c",
412 "src/f32-vbinary/gen/vsub-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700413 "src/f32-vbinary/gen/vsub-scalar-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700414 "src/f32-vbinary/gen/vsubc-minmax-scalar-x1.c",
415 "src/f32-vbinary/gen/vsubc-minmax-scalar-x2.c",
416 "src/f32-vbinary/gen/vsubc-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700417 "src/f32-vbinary/gen/vsubc-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700418 "src/f32-vbinary/gen/vsubc-relu-scalar-x1.c",
419 "src/f32-vbinary/gen/vsubc-relu-scalar-x2.c",
420 "src/f32-vbinary/gen/vsubc-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700421 "src/f32-vbinary/gen/vsubc-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700422 "src/f32-vbinary/gen/vsubc-scalar-x1.c",
423 "src/f32-vbinary/gen/vsubc-scalar-x2.c",
424 "src/f32-vbinary/gen/vsubc-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700425 "src/f32-vbinary/gen/vsubc-scalar-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700426 "src/f32-vlrelu/gen/vlrelu-scalar-x1.c",
427 "src/f32-vlrelu/gen/vlrelu-scalar-x2.c",
428 "src/f32-vlrelu/gen/vlrelu-scalar-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700429 "src/f32-vmulcaddc/gen/c1-minmax-scalar-2x.c",
430 "src/f32-vmulcaddc/gen/c2-minmax-scalar-2x.c",
431 "src/f32-vmulcaddc/gen/c4-minmax-scalar-2x.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -0700432 "src/f32-vrnd/gen/vrndd-scalar-libm-x1.c",
433 "src/f32-vrnd/gen/vrndd-scalar-libm-x2.c",
434 "src/f32-vrnd/gen/vrndd-scalar-libm-x4.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700435 "src/f32-vrnd/gen/vrndne-scalar-libm-x1.c",
436 "src/f32-vrnd/gen/vrndne-scalar-libm-x2.c",
437 "src/f32-vrnd/gen/vrndne-scalar-libm-x4.c",
438 "src/f32-vrnd/gen/vrndu-scalar-libm-x1.c",
439 "src/f32-vrnd/gen/vrndu-scalar-libm-x2.c",
440 "src/f32-vrnd/gen/vrndu-scalar-libm-x4.c",
441 "src/f32-vrnd/gen/vrndz-scalar-libm-x1.c",
442 "src/f32-vrnd/gen/vrndz-scalar-libm-x2.c",
443 "src/f32-vrnd/gen/vrndz-scalar-libm-x4.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -0700444 "src/f32-vsqrt/gen/scalar-sqrt-x1.c",
445 "src/f32-vsqrt/gen/scalar-sqrt-x2.c",
446 "src/f32-vsqrt/gen/scalar-sqrt-x4.c",
Marat Dukhan5020b962020-06-08 13:30:10 -0700447 "src/f32-vunary/gen/vabs-scalar-x1.c",
448 "src/f32-vunary/gen/vabs-scalar-x2.c",
449 "src/f32-vunary/gen/vabs-scalar-x4.c",
450 "src/f32-vunary/gen/vneg-scalar-x1.c",
451 "src/f32-vunary/gen/vneg-scalar-x2.c",
452 "src/f32-vunary/gen/vneg-scalar-x4.c",
453 "src/f32-vunary/gen/vsqr-scalar-x1.c",
454 "src/f32-vunary/gen/vsqr-scalar-x2.c",
455 "src/f32-vunary/gen/vsqr-scalar-x4.c",
Marat Dukhande390d42020-11-29 19:32:18 -0800456 "src/math/expm1minus-scalar-rr2-lut4-p4.c",
457 "src/math/expm1minus-scalar-rr2-lut8-p3.c",
458 "src/math/expm1minus-scalar-rr2-lut8-p4.c",
Marat Dukhanc60742b2020-11-23 12:33:27 -0800459 "src/math/expm1minus-scalar-rr2-lut16-p3.c",
460 "src/math/expm1minus-scalar-rr2-lut16-p4.c",
461 "src/math/expm1minus-scalar-rr2-p5.c",
462 "src/math/expm1minus-scalar-rr2-p6.c",
Frank Barchard22136062020-11-24 18:44:46 -0800463 "src/math/expminus-scalar-rr2-lut64-p2.c",
464 "src/math/expminus-scalar-rr2-lut2048-p1.c",
465 "src/math/expminus-scalar-rr2-p5.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -0700466 "src/math/roundd-scalar-addsub.c",
467 "src/math/roundd-scalar-cvt.c",
Marat Dukhanffbf96a2020-05-14 02:59:08 -0700468 "src/math/roundd-scalar-floor.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700469 "src/math/roundne-scalar-addsub.c",
470 "src/math/roundne-scalar-nearbyint.c",
471 "src/math/roundne-scalar-rint.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -0700472 "src/math/roundu-scalar-addsub.c",
Marat Dukhanffbf96a2020-05-14 02:59:08 -0700473 "src/math/roundu-scalar-ceil.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -0700474 "src/math/roundu-scalar-cvt.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -0700475 "src/math/roundz-scalar-addsub.c",
476 "src/math/roundz-scalar-cvt.c",
Marat Dukhanffbf96a2020-05-14 02:59:08 -0700477 "src/math/roundz-scalar-trunc.c",
Marat Dukhanf8475d62020-09-17 15:01:43 -0700478 "src/math/sigmoid-scalar-rr2-lut64-p2-div.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700479 "src/math/sigmoid-scalar-rr2-lut2048-p1-div.c",
Marat Dukhanf8475d62020-09-17 15:01:43 -0700480 "src/math/sigmoid-scalar-rr2-p5-div.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700481 "src/qs8-gemm/gen/8x8c4-minmax-scalar.c",
482 "src/qs8-gemm/gen/12x4c4-minmax-scalar.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -0700483 "src/qs8-requantization/fp32-scalar-lrintf.c",
484 "src/qs8-requantization/fp32-scalar-magic.c",
485 "src/qs8-requantization/precise-scalar-signed64.c",
486 "src/qs8-requantization/precise-scalar-unsigned32.c",
487 "src/qs8-requantization/precise-scalar-unsigned64.c",
488 "src/qs8-requantization/q31-scalar.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -0700489 "src/qu8-avgpool/9p8x-minmax-scalar-c1.c",
490 "src/qu8-avgpool/9x-minmax-scalar-c1.c",
491 "src/qu8-dwconv/up1x9-minmax-scalar.c",
492 "src/qu8-gavgpool/7p7x-minmax-scalar-c1.c",
493 "src/qu8-gavgpool/7x-minmax-scalar-c1.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700494 "src/qu8-gemm/2x2-minmax-scalar.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700495 "src/qu8-gemm/gen/8x8c4-minmax-scalar.c",
496 "src/qu8-gemm/gen/12x4c4-minmax-scalar.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -0700497 "src/qu8-igemm/2x2-minmax-scalar.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -0700498 "src/qu8-requantization/fp32-scalar-lrintf.c",
499 "src/qu8-requantization/fp32-scalar-magic.c",
500 "src/qu8-requantization/precise-scalar-signed64.c",
501 "src/qu8-requantization/precise-scalar-unsigned32.c",
502 "src/qu8-requantization/precise-scalar-unsigned64.c",
503 "src/qu8-requantization/q31-scalar.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -0700504 "src/qu8-vadd/minmax-scalar.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -0700505 "src/u8-clamp/scalar-x4.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700506 "src/u8-lut32norm/scalar.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700507 "src/u8-maxpool/9p8x-minmax-scalar-c1.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700508 "src/u8-rmax/scalar.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700509 "src/x8-lut/scalar.c",
510 "src/x8-zip/x2-scalar.c",
511 "src/x8-zip/x3-scalar.c",
512 "src/x8-zip/x4-scalar.c",
513 "src/x8-zip/xm-scalar.c",
Marat Dukhanad71b9a2020-11-20 00:01:51 -0800514 "src/x32-depthtospace2d-chw2hwc/scalar.c",
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -0700515 "src/x32-fill/scalar-float.c",
516 "src/x32-fill/scalar-int.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700517 "src/x32-packx/x2-scalar.c",
518 "src/x32-packx/x3-scalar.c",
519 "src/x32-packx/x4-scalar.c",
Marat Dukhan63523d42020-05-22 17:07:33 -0700520 "src/x32-pad/scalar-float.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700521 "src/x32-pad/scalar-int.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700522 "src/x32-unpool/scalar.c",
523 "src/x32-zip/x2-scalar.c",
524 "src/x32-zip/x3-scalar.c",
525 "src/x32-zip/x4-scalar.c",
526 "src/x32-zip/xm-scalar.c",
Marat Dukhan048931b2020-11-24 20:53:54 -0800527 "src/xx-copy/memcpy.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700528]
529
Marat Dukhan436ebe62019-12-04 15:10:12 -0800530WASM_UKERNELS = [
Marat Dukhan99936602020-04-11 16:47:01 -0700531 "src/f32-avgpool/9p8x-minmax-wasm-c1.c",
532 "src/f32-avgpool/9x-minmax-wasm-c1.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -0700533 "src/f32-clamp/gen/wasm-x1.c",
534 "src/f32-clamp/gen/wasm-x2.c",
535 "src/f32-clamp/gen/wasm-x4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700536 "src/f32-dwconv/gen/up1x4-minmax-wasm-acc2.c",
537 "src/f32-dwconv/gen/up1x4-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700538 "src/f32-dwconv/gen/up1x4-wasm-acc2.c",
539 "src/f32-dwconv/gen/up1x4-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700540 "src/f32-dwconv/gen/up1x9-minmax-wasm-acc2.c",
541 "src/f32-dwconv/gen/up1x9-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700542 "src/f32-dwconv/gen/up1x9-wasm-acc2.c",
543 "src/f32-dwconv/gen/up1x9-wasm.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700544 "src/f32-dwconv/gen/up1x25-minmax-wasm-acc2.c",
545 "src/f32-dwconv/gen/up1x25-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700546 "src/f32-dwconv/gen/up1x25-wasm-acc2.c",
547 "src/f32-dwconv/gen/up1x25-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700548 "src/f32-dwconv/gen/up2x4-minmax-wasm-acc2.c",
549 "src/f32-dwconv/gen/up2x4-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700550 "src/f32-dwconv/gen/up2x4-wasm-acc2.c",
551 "src/f32-dwconv/gen/up2x4-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700552 "src/f32-dwconv/gen/up2x9-minmax-wasm-acc2.c",
553 "src/f32-dwconv/gen/up2x9-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700554 "src/f32-dwconv/gen/up2x9-wasm-acc2.c",
555 "src/f32-dwconv/gen/up2x9-wasm.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700556 "src/f32-dwconv/gen/up2x25-minmax-wasm-acc2.c",
557 "src/f32-dwconv/gen/up2x25-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700558 "src/f32-dwconv/gen/up2x25-wasm-acc2.c",
559 "src/f32-dwconv/gen/up2x25-wasm.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700560 "src/f32-gavgpool/7p7x-minmax-wasm-c1.c",
561 "src/f32-gavgpool/7x-minmax-wasm-c1.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700562 "src/f32-gemm/gen-inc/1x4inc-minmax-wasm.c",
563 "src/f32-gemm/gen-inc/2x4inc-minmax-wasm.c",
564 "src/f32-gemm/gen-inc/4x4inc-minmax-wasm.c",
565 "src/f32-gemm/gen/1x4-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700566 "src/f32-gemm/gen/1x4-relu-wasm.c",
567 "src/f32-gemm/gen/1x4-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700568 "src/f32-gemm/gen/2x4-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700569 "src/f32-gemm/gen/2x4-relu-wasm.c",
570 "src/f32-gemm/gen/2x4-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700571 "src/f32-gemm/gen/4x2-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700572 "src/f32-gemm/gen/4x2-relu-wasm.c",
573 "src/f32-gemm/gen/4x2-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700574 "src/f32-gemm/gen/4x4-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700575 "src/f32-gemm/gen/4x4-relu-wasm.c",
576 "src/f32-gemm/gen/4x4-wasm.c",
Marat Dukhanefaac272020-07-16 16:58:07 -0700577 "src/f32-hswish/gen/hswish-wasm-x1.c",
578 "src/f32-hswish/gen/hswish-wasm-x2.c",
579 "src/f32-hswish/gen/hswish-wasm-x4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700580 "src/f32-igemm/gen/1x4-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700581 "src/f32-igemm/gen/1x4-relu-wasm.c",
582 "src/f32-igemm/gen/1x4-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700583 "src/f32-igemm/gen/2x4-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700584 "src/f32-igemm/gen/2x4-relu-wasm.c",
585 "src/f32-igemm/gen/2x4-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700586 "src/f32-igemm/gen/4x2-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700587 "src/f32-igemm/gen/4x2-relu-wasm.c",
588 "src/f32-igemm/gen/4x2-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700589 "src/f32-igemm/gen/4x4-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700590 "src/f32-igemm/gen/4x4-relu-wasm.c",
591 "src/f32-igemm/gen/4x4-wasm.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700592 "src/f32-maxpool/9p8x-minmax-wasm-c1.c",
593 "src/f32-pavgpool/9p8x-minmax-wasm-c1.c",
594 "src/f32-pavgpool/9x-minmax-wasm-c1.c",
Marat Dukhan7c1f8082020-06-25 13:26:20 -0700595 "src/f32-prelu/gen/wasm-2x1.c",
596 "src/f32-prelu/gen/wasm-2x4.c",
Frank Barchardfb158e22020-07-15 16:10:10 -0700597 "src/f32-relu/gen/wasm-x1.c",
598 "src/f32-relu/gen/wasm-x2.c",
599 "src/f32-relu/gen/wasm-x4.c",
Frank Barchard37297a62020-08-27 12:15:10 -0700600 "src/f32-relu/gen/wasm-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700601 "src/f32-vbinary/gen/vadd-minmax-wasm-x1.c",
602 "src/f32-vbinary/gen/vadd-minmax-wasm-x2.c",
603 "src/f32-vbinary/gen/vadd-minmax-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700604 "src/f32-vbinary/gen/vadd-minmax-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700605 "src/f32-vbinary/gen/vadd-relu-wasm-x1.c",
606 "src/f32-vbinary/gen/vadd-relu-wasm-x2.c",
607 "src/f32-vbinary/gen/vadd-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700608 "src/f32-vbinary/gen/vadd-relu-wasm-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700609 "src/f32-vbinary/gen/vaddc-minmax-wasm-x1.c",
610 "src/f32-vbinary/gen/vaddc-minmax-wasm-x2.c",
611 "src/f32-vbinary/gen/vaddc-minmax-wasm-x4.c",
612 "src/f32-vbinary/gen/vaddc-minmax-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700613 "src/f32-vbinary/gen/vaddc-relu-wasm-x1.c",
614 "src/f32-vbinary/gen/vaddc-relu-wasm-x2.c",
615 "src/f32-vbinary/gen/vaddc-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700616 "src/f32-vbinary/gen/vaddc-relu-wasm-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700617 "src/f32-vbinary/gen/vdiv-minmax-wasm-x1.c",
618 "src/f32-vbinary/gen/vdiv-minmax-wasm-x2.c",
619 "src/f32-vbinary/gen/vdiv-minmax-wasm-x4.c",
620 "src/f32-vbinary/gen/vdiv-minmax-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700621 "src/f32-vbinary/gen/vdiv-relu-wasm-x1.c",
622 "src/f32-vbinary/gen/vdiv-relu-wasm-x2.c",
623 "src/f32-vbinary/gen/vdiv-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700624 "src/f32-vbinary/gen/vdiv-relu-wasm-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700625 "src/f32-vbinary/gen/vdivc-minmax-wasm-x1.c",
626 "src/f32-vbinary/gen/vdivc-minmax-wasm-x2.c",
627 "src/f32-vbinary/gen/vdivc-minmax-wasm-x4.c",
628 "src/f32-vbinary/gen/vdivc-minmax-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700629 "src/f32-vbinary/gen/vdivc-relu-wasm-x1.c",
630 "src/f32-vbinary/gen/vdivc-relu-wasm-x2.c",
631 "src/f32-vbinary/gen/vdivc-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700632 "src/f32-vbinary/gen/vdivc-relu-wasm-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800633 "src/f32-vbinary/gen/vmax-wasm-x1.c",
634 "src/f32-vbinary/gen/vmax-wasm-x2.c",
635 "src/f32-vbinary/gen/vmax-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700636 "src/f32-vbinary/gen/vmax-wasm-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800637 "src/f32-vbinary/gen/vmaxc-wasm-x1.c",
638 "src/f32-vbinary/gen/vmaxc-wasm-x2.c",
639 "src/f32-vbinary/gen/vmaxc-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700640 "src/f32-vbinary/gen/vmaxc-wasm-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800641 "src/f32-vbinary/gen/vmin-wasm-x1.c",
642 "src/f32-vbinary/gen/vmin-wasm-x2.c",
643 "src/f32-vbinary/gen/vmin-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700644 "src/f32-vbinary/gen/vmin-wasm-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800645 "src/f32-vbinary/gen/vminc-wasm-x1.c",
646 "src/f32-vbinary/gen/vminc-wasm-x2.c",
647 "src/f32-vbinary/gen/vminc-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700648 "src/f32-vbinary/gen/vminc-wasm-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700649 "src/f32-vbinary/gen/vmul-minmax-wasm-x1.c",
650 "src/f32-vbinary/gen/vmul-minmax-wasm-x2.c",
651 "src/f32-vbinary/gen/vmul-minmax-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700652 "src/f32-vbinary/gen/vmul-minmax-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700653 "src/f32-vbinary/gen/vmul-relu-wasm-x1.c",
654 "src/f32-vbinary/gen/vmul-relu-wasm-x2.c",
655 "src/f32-vbinary/gen/vmul-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700656 "src/f32-vbinary/gen/vmul-relu-wasm-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700657 "src/f32-vbinary/gen/vmulc-minmax-wasm-x1.c",
658 "src/f32-vbinary/gen/vmulc-minmax-wasm-x2.c",
659 "src/f32-vbinary/gen/vmulc-minmax-wasm-x4.c",
660 "src/f32-vbinary/gen/vmulc-minmax-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700661 "src/f32-vbinary/gen/vmulc-relu-wasm-x1.c",
662 "src/f32-vbinary/gen/vmulc-relu-wasm-x2.c",
663 "src/f32-vbinary/gen/vmulc-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700664 "src/f32-vbinary/gen/vmulc-relu-wasm-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700665 "src/f32-vbinary/gen/vrdivc-minmax-wasm-x1.c",
666 "src/f32-vbinary/gen/vrdivc-minmax-wasm-x2.c",
667 "src/f32-vbinary/gen/vrdivc-minmax-wasm-x4.c",
668 "src/f32-vbinary/gen/vrdivc-minmax-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700669 "src/f32-vbinary/gen/vrdivc-relu-wasm-x1.c",
670 "src/f32-vbinary/gen/vrdivc-relu-wasm-x2.c",
671 "src/f32-vbinary/gen/vrdivc-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700672 "src/f32-vbinary/gen/vrdivc-relu-wasm-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700673 "src/f32-vbinary/gen/vrsubc-minmax-wasm-x1.c",
674 "src/f32-vbinary/gen/vrsubc-minmax-wasm-x2.c",
675 "src/f32-vbinary/gen/vrsubc-minmax-wasm-x4.c",
676 "src/f32-vbinary/gen/vrsubc-minmax-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700677 "src/f32-vbinary/gen/vrsubc-relu-wasm-x1.c",
678 "src/f32-vbinary/gen/vrsubc-relu-wasm-x2.c",
679 "src/f32-vbinary/gen/vrsubc-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700680 "src/f32-vbinary/gen/vrsubc-relu-wasm-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700681 "src/f32-vbinary/gen/vsub-minmax-wasm-x1.c",
682 "src/f32-vbinary/gen/vsub-minmax-wasm-x2.c",
683 "src/f32-vbinary/gen/vsub-minmax-wasm-x4.c",
684 "src/f32-vbinary/gen/vsub-minmax-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700685 "src/f32-vbinary/gen/vsub-relu-wasm-x1.c",
686 "src/f32-vbinary/gen/vsub-relu-wasm-x2.c",
687 "src/f32-vbinary/gen/vsub-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700688 "src/f32-vbinary/gen/vsub-relu-wasm-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700689 "src/f32-vbinary/gen/vsubc-minmax-wasm-x1.c",
690 "src/f32-vbinary/gen/vsubc-minmax-wasm-x2.c",
691 "src/f32-vbinary/gen/vsubc-minmax-wasm-x4.c",
692 "src/f32-vbinary/gen/vsubc-minmax-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700693 "src/f32-vbinary/gen/vsubc-relu-wasm-x1.c",
694 "src/f32-vbinary/gen/vsubc-relu-wasm-x2.c",
695 "src/f32-vbinary/gen/vsubc-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700696 "src/f32-vbinary/gen/vsubc-relu-wasm-x8.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -0700697 "src/f32-vlrelu/gen/vlrelu-wasm-x1.c",
698 "src/f32-vlrelu/gen/vlrelu-wasm-x2.c",
699 "src/f32-vlrelu/gen/vlrelu-wasm-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700700 "src/f32-vmulcaddc/gen/c1-minmax-wasm-2x.c",
701 "src/f32-vmulcaddc/gen/c2-minmax-wasm-2x.c",
702 "src/f32-vmulcaddc/gen/c4-minmax-wasm-2x.c",
Marat Dukhan436ebe62019-12-04 15:10:12 -0800703]
704
Marat Dukhan290055c2020-06-09 12:24:29 -0700705WASMSIMD_UKERNELS = [
Marat Dukhan40f05522020-07-16 22:33:12 -0700706 "src/f32-argmaxpool/4x-wasmsimd-c4.c",
707 "src/f32-argmaxpool/9p8x-wasmsimd-c4.c",
708 "src/f32-argmaxpool/9x-wasmsimd-c4.c",
Marat Dukhan3b7432d2020-07-16 17:46:32 -0700709 "src/f32-avgpool/9p8x-minmax-wasmsimd-arm-c4.c",
710 "src/f32-avgpool/9p8x-minmax-wasmsimd-x86-c4.c",
711 "src/f32-avgpool/9x-minmax-wasmsimd-arm-c4.c",
712 "src/f32-avgpool/9x-minmax-wasmsimd-x86-c4.c",
Marat Dukhan3fa52c82020-07-08 12:54:33 -0700713 "src/f32-clamp/gen/wasmsimd-arm-x4.c",
714 "src/f32-clamp/gen/wasmsimd-arm-x8.c",
715 "src/f32-clamp/gen/wasmsimd-x86-x4.c",
716 "src/f32-clamp/gen/wasmsimd-x86-x8.c",
Frank Barchard22136062020-11-24 18:44:46 -0800717 "src/f32-conv-hwc2chw/3x3s2p1c3x4-wasmsimd-2x2.c",
Marat Dukhanac014d72020-06-16 08:36:47 -0700718 "src/f32-dwconv/gen/up4x4-minmax-wasmsimd-acc2-arm.c",
Marat Dukhanac014d72020-06-16 08:36:47 -0700719 "src/f32-dwconv/gen/up4x4-minmax-wasmsimd-acc2-x86.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700720 "src/f32-dwconv/gen/up4x4-minmax-wasmsimd-arm.c",
Marat Dukhanac014d72020-06-16 08:36:47 -0700721 "src/f32-dwconv/gen/up4x4-minmax-wasmsimd-x86.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700722 "src/f32-dwconv/gen/up4x4-wasmsimd.c",
723 "src/f32-dwconv/gen/up4x9-minmax-wasmsimd-acc2-arm.c",
Marat Dukhanac014d72020-06-16 08:36:47 -0700724 "src/f32-dwconv/gen/up4x9-minmax-wasmsimd-acc2-x86.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700725 "src/f32-dwconv/gen/up4x9-minmax-wasmsimd-arm.c",
Marat Dukhanac014d72020-06-16 08:36:47 -0700726 "src/f32-dwconv/gen/up4x9-minmax-wasmsimd-x86.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700727 "src/f32-dwconv/gen/up4x9-wasmsimd.c",
728 "src/f32-dwconv/gen/up4x25-minmax-wasmsimd-acc2-arm.c",
729 "src/f32-dwconv/gen/up4x25-minmax-wasmsimd-acc2-x86.c",
730 "src/f32-dwconv/gen/up4x25-minmax-wasmsimd-arm.c",
731 "src/f32-dwconv/gen/up4x25-minmax-wasmsimd-x86.c",
732 "src/f32-dwconv/gen/up4x25-wasmsimd.c",
733 "src/f32-dwconv/gen/up8x4-minmax-wasmsimd-acc2-arm.c",
Marat Dukhanac014d72020-06-16 08:36:47 -0700734 "src/f32-dwconv/gen/up8x4-minmax-wasmsimd-acc2-x86.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700735 "src/f32-dwconv/gen/up8x4-minmax-wasmsimd-arm.c",
Marat Dukhanac014d72020-06-16 08:36:47 -0700736 "src/f32-dwconv/gen/up8x4-minmax-wasmsimd-x86.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700737 "src/f32-dwconv/gen/up8x4-wasmsimd.c",
738 "src/f32-dwconv/gen/up8x9-minmax-wasmsimd-acc2-arm.c",
Marat Dukhanac014d72020-06-16 08:36:47 -0700739 "src/f32-dwconv/gen/up8x9-minmax-wasmsimd-acc2-x86.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700740 "src/f32-dwconv/gen/up8x9-minmax-wasmsimd-arm.c",
Marat Dukhanac014d72020-06-16 08:36:47 -0700741 "src/f32-dwconv/gen/up8x9-minmax-wasmsimd-x86.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700742 "src/f32-dwconv/gen/up8x9-wasmsimd.c",
743 "src/f32-dwconv/gen/up8x25-minmax-wasmsimd-acc2-arm.c",
744 "src/f32-dwconv/gen/up8x25-minmax-wasmsimd-acc2-x86.c",
745 "src/f32-dwconv/gen/up8x25-minmax-wasmsimd-arm.c",
746 "src/f32-dwconv/gen/up8x25-minmax-wasmsimd-x86.c",
747 "src/f32-dwconv/gen/up8x25-wasmsimd.c",
Frank Barchard3b800452020-11-22 12:12:35 -0800748 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-1x4-acc2.c",
749 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-1x4-acc3.c",
750 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-1x4-acc4.c",
751 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-1x4.c",
752 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-2x4-acc2.c",
753 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-2x4.c",
754 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-3x4.c",
755 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-4x4.c",
756 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-5x4.c",
757 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-6x4.c",
758 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-1x4-acc2.c",
759 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-1x4-acc3.c",
760 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-1x4-acc4.c",
761 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-1x4.c",
762 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-2x4-acc2.c",
763 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-2x4.c",
764 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-3x4.c",
765 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-4x4.c",
766 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-5x4.c",
767 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-6x4.c",
Frank Barchard9cef5ea2020-11-18 14:52:08 -0800768 "src/f32-dwconv2d-chw/gen/3x3s2p1-wasmsimd-1x4-acc3-arm.c",
Frank Barcharddb5c32d2020-11-16 23:58:42 -0800769 "src/f32-dwconv2d-chw/gen/3x3s2p1-wasmsimd-1x4-acc3-x86.c",
Frank Barchard9cef5ea2020-11-18 14:52:08 -0800770 "src/f32-dwconv2d-chw/gen/5x5p2-wasmsimd-3x4-arm.c",
Frank Barcharddb5c32d2020-11-16 23:58:42 -0800771 "src/f32-dwconv2d-chw/gen/5x5p2-wasmsimd-3x4-x86.c",
Frank Barchard9cef5ea2020-11-18 14:52:08 -0800772 "src/f32-dwconv2d-chw/gen/5x5s2p2-wasmsimd-1x4-acc2-arm.c",
Frank Barcharddb5c32d2020-11-16 23:58:42 -0800773 "src/f32-dwconv2d-chw/gen/5x5s2p2-wasmsimd-1x4-acc2-x86.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700774 "src/f32-gavgpool-cw/wasmsimd-arm-x4.c",
775 "src/f32-gavgpool-cw/wasmsimd-x86-x4.c",
Marat Dukhanc6016802020-07-16 18:51:28 -0700776 "src/f32-gavgpool/7p7x-minmax-wasmsimd-arm-c4.c",
777 "src/f32-gavgpool/7p7x-minmax-wasmsimd-x86-c4.c",
778 "src/f32-gavgpool/7x-minmax-wasmsimd-arm-c4.c",
779 "src/f32-gavgpool/7x-minmax-wasmsimd-x86-c4.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700780 "src/f32-gemm/gen-inc/1x8inc-minmax-wasmsimd-loadsplat-arm.c",
781 "src/f32-gemm/gen-inc/1x8inc-minmax-wasmsimd-loadsplat-x86.c",
782 "src/f32-gemm/gen-inc/1x8inc-minmax-wasmsimd-splat-arm.c",
783 "src/f32-gemm/gen-inc/1x8inc-minmax-wasmsimd-splat-x86.c",
784 "src/f32-gemm/gen-inc/1x8s4inc-minmax-wasmsimd-arm.c",
785 "src/f32-gemm/gen-inc/1x8s4inc-minmax-wasmsimd-x86.c",
786 "src/f32-gemm/gen-inc/3x8inc-minmax-wasmsimd-loadsplat-arm.c",
787 "src/f32-gemm/gen-inc/3x8inc-minmax-wasmsimd-loadsplat-x86.c",
788 "src/f32-gemm/gen-inc/3x8inc-minmax-wasmsimd-splat-arm.c",
789 "src/f32-gemm/gen-inc/3x8inc-minmax-wasmsimd-splat-x86.c",
790 "src/f32-gemm/gen-inc/3x8s4inc-minmax-wasmsimd-arm.c",
791 "src/f32-gemm/gen-inc/3x8s4inc-minmax-wasmsimd-x86.c",
792 "src/f32-gemm/gen-inc/4x8inc-minmax-wasmsimd-loadsplat-arm.c",
793 "src/f32-gemm/gen-inc/4x8inc-minmax-wasmsimd-loadsplat-x86.c",
794 "src/f32-gemm/gen-inc/4x8inc-minmax-wasmsimd-splat-arm.c",
795 "src/f32-gemm/gen-inc/4x8inc-minmax-wasmsimd-splat-x86.c",
796 "src/f32-gemm/gen-inc/4x8s4inc-minmax-wasmsimd-arm.c",
797 "src/f32-gemm/gen-inc/4x8s4inc-minmax-wasmsimd-x86.c",
798 "src/f32-gemm/gen-inc/5x8inc-minmax-wasmsimd-loadsplat-arm.c",
799 "src/f32-gemm/gen-inc/5x8inc-minmax-wasmsimd-loadsplat-x86.c",
800 "src/f32-gemm/gen-inc/5x8inc-minmax-wasmsimd-splat-arm.c",
801 "src/f32-gemm/gen-inc/5x8inc-minmax-wasmsimd-splat-x86.c",
802 "src/f32-gemm/gen-inc/5x8s4inc-minmax-wasmsimd-arm.c",
803 "src/f32-gemm/gen-inc/5x8s4inc-minmax-wasmsimd-x86.c",
804 "src/f32-gemm/gen-inc/6x8inc-minmax-wasmsimd-loadsplat-arm.c",
805 "src/f32-gemm/gen-inc/6x8inc-minmax-wasmsimd-loadsplat-x86.c",
806 "src/f32-gemm/gen-inc/6x8inc-minmax-wasmsimd-splat-arm.c",
807 "src/f32-gemm/gen-inc/6x8inc-minmax-wasmsimd-splat-x86.c",
808 "src/f32-gemm/gen-inc/6x8s4inc-minmax-wasmsimd-arm.c",
809 "src/f32-gemm/gen-inc/6x8s4inc-minmax-wasmsimd-x86.c",
810 "src/f32-gemm/gen/1x8-minmax-wasmsimd-loadsplat-arm.c",
811 "src/f32-gemm/gen/1x8-minmax-wasmsimd-loadsplat-x86.c",
812 "src/f32-gemm/gen/1x8-minmax-wasmsimd-splat-arm.c",
813 "src/f32-gemm/gen/1x8-minmax-wasmsimd-splat-x86.c",
814 "src/f32-gemm/gen/1x8-relu-wasmsimd-splat.c",
815 "src/f32-gemm/gen/1x8-wasmsimd-splat.c",
816 "src/f32-gemm/gen/1x8s4-minmax-wasmsimd-arm.c",
817 "src/f32-gemm/gen/1x8s4-minmax-wasmsimd-x86.c",
818 "src/f32-gemm/gen/3x8-minmax-wasmsimd-loadsplat-arm.c",
819 "src/f32-gemm/gen/3x8-minmax-wasmsimd-loadsplat-x86.c",
820 "src/f32-gemm/gen/3x8-minmax-wasmsimd-splat-arm.c",
821 "src/f32-gemm/gen/3x8-minmax-wasmsimd-splat-x86.c",
822 "src/f32-gemm/gen/3x8s4-minmax-wasmsimd-arm.c",
823 "src/f32-gemm/gen/3x8s4-minmax-wasmsimd-x86.c",
824 "src/f32-gemm/gen/4x2c4-minmax-wasmsimd-arm.c",
825 "src/f32-gemm/gen/4x2c4-minmax-wasmsimd-x86.c",
826 "src/f32-gemm/gen/4x2c4-relu-wasmsimd.c",
827 "src/f32-gemm/gen/4x2c4-wasmsimd.c",
828 "src/f32-gemm/gen/4x8-minmax-wasmsimd-loadsplat-arm.c",
829 "src/f32-gemm/gen/4x8-minmax-wasmsimd-loadsplat-x86.c",
830 "src/f32-gemm/gen/4x8-minmax-wasmsimd-splat-arm.c",
831 "src/f32-gemm/gen/4x8-minmax-wasmsimd-splat-x86.c",
832 "src/f32-gemm/gen/4x8-relu-wasmsimd-splat.c",
833 "src/f32-gemm/gen/4x8-wasmsimd-splat.c",
834 "src/f32-gemm/gen/4x8s4-minmax-wasmsimd-arm.c",
835 "src/f32-gemm/gen/4x8s4-minmax-wasmsimd-x86.c",
836 "src/f32-gemm/gen/5x8-minmax-wasmsimd-loadsplat-arm.c",
837 "src/f32-gemm/gen/5x8-minmax-wasmsimd-loadsplat-x86.c",
838 "src/f32-gemm/gen/5x8-minmax-wasmsimd-splat-arm.c",
839 "src/f32-gemm/gen/5x8-minmax-wasmsimd-splat-x86.c",
840 "src/f32-gemm/gen/5x8-relu-wasmsimd-splat.c",
841 "src/f32-gemm/gen/5x8-wasmsimd-splat.c",
842 "src/f32-gemm/gen/5x8s4-minmax-wasmsimd-arm.c",
843 "src/f32-gemm/gen/5x8s4-minmax-wasmsimd-x86.c",
844 "src/f32-gemm/gen/6x8-minmax-wasmsimd-loadsplat-arm.c",
845 "src/f32-gemm/gen/6x8-minmax-wasmsimd-loadsplat-x86.c",
846 "src/f32-gemm/gen/6x8-minmax-wasmsimd-splat-arm.c",
847 "src/f32-gemm/gen/6x8-minmax-wasmsimd-splat-x86.c",
848 "src/f32-gemm/gen/6x8s4-minmax-wasmsimd-arm.c",
849 "src/f32-gemm/gen/6x8s4-minmax-wasmsimd-x86.c",
Marat Dukhanefaac272020-07-16 16:58:07 -0700850 "src/f32-hswish/gen/hswish-wasmsimd-x4.c",
851 "src/f32-hswish/gen/hswish-wasmsimd-x8.c",
852 "src/f32-hswish/gen/hswish-wasmsimd-x16.c",
XNNPACK Team965272b2020-10-23 21:10:15 -0700853 "src/f32-ibilinear-chw/gen/wasmsimd-p4.c",
854 "src/f32-ibilinear-chw/gen/wasmsimd-p8.c",
Frank Barchard35db7d02020-10-26 13:37:34 -0700855 "src/f32-ibilinear/gen/wasmsimd-c4.c",
856 "src/f32-ibilinear/gen/wasmsimd-c8.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700857 "src/f32-igemm/gen/1x8-minmax-wasmsimd-loadsplat-arm.c",
858 "src/f32-igemm/gen/1x8-minmax-wasmsimd-loadsplat-x86.c",
859 "src/f32-igemm/gen/1x8-minmax-wasmsimd-splat-arm.c",
860 "src/f32-igemm/gen/1x8-minmax-wasmsimd-splat-x86.c",
861 "src/f32-igemm/gen/1x8-relu-wasmsimd-splat.c",
862 "src/f32-igemm/gen/1x8-wasmsimd-splat.c",
863 "src/f32-igemm/gen/1x8s4-minmax-wasmsimd-arm.c",
864 "src/f32-igemm/gen/1x8s4-minmax-wasmsimd-x86.c",
865 "src/f32-igemm/gen/3x8-minmax-wasmsimd-loadsplat-arm.c",
866 "src/f32-igemm/gen/3x8-minmax-wasmsimd-loadsplat-x86.c",
867 "src/f32-igemm/gen/3x8-minmax-wasmsimd-splat-arm.c",
868 "src/f32-igemm/gen/3x8-minmax-wasmsimd-splat-x86.c",
869 "src/f32-igemm/gen/3x8s4-minmax-wasmsimd-arm.c",
870 "src/f32-igemm/gen/3x8s4-minmax-wasmsimd-x86.c",
871 "src/f32-igemm/gen/4x2c4-minmax-wasmsimd-arm.c",
872 "src/f32-igemm/gen/4x2c4-minmax-wasmsimd-x86.c",
873 "src/f32-igemm/gen/4x2c4-relu-wasmsimd.c",
874 "src/f32-igemm/gen/4x2c4-wasmsimd.c",
875 "src/f32-igemm/gen/4x8-minmax-wasmsimd-loadsplat-arm.c",
876 "src/f32-igemm/gen/4x8-minmax-wasmsimd-loadsplat-x86.c",
877 "src/f32-igemm/gen/4x8-minmax-wasmsimd-splat-arm.c",
878 "src/f32-igemm/gen/4x8-minmax-wasmsimd-splat-x86.c",
879 "src/f32-igemm/gen/4x8-relu-wasmsimd-splat.c",
880 "src/f32-igemm/gen/4x8-wasmsimd-splat.c",
881 "src/f32-igemm/gen/4x8s4-minmax-wasmsimd-arm.c",
882 "src/f32-igemm/gen/4x8s4-minmax-wasmsimd-x86.c",
883 "src/f32-igemm/gen/5x8-minmax-wasmsimd-loadsplat-arm.c",
884 "src/f32-igemm/gen/5x8-minmax-wasmsimd-loadsplat-x86.c",
885 "src/f32-igemm/gen/5x8-minmax-wasmsimd-splat-arm.c",
886 "src/f32-igemm/gen/5x8-minmax-wasmsimd-splat-x86.c",
887 "src/f32-igemm/gen/5x8-relu-wasmsimd-splat.c",
888 "src/f32-igemm/gen/5x8-wasmsimd-splat.c",
889 "src/f32-igemm/gen/5x8s4-minmax-wasmsimd-arm.c",
890 "src/f32-igemm/gen/5x8s4-minmax-wasmsimd-x86.c",
891 "src/f32-igemm/gen/6x8-minmax-wasmsimd-loadsplat-arm.c",
892 "src/f32-igemm/gen/6x8-minmax-wasmsimd-loadsplat-x86.c",
893 "src/f32-igemm/gen/6x8-minmax-wasmsimd-splat-arm.c",
894 "src/f32-igemm/gen/6x8-minmax-wasmsimd-splat-x86.c",
895 "src/f32-igemm/gen/6x8s4-minmax-wasmsimd-arm.c",
896 "src/f32-igemm/gen/6x8s4-minmax-wasmsimd-x86.c",
Marat Dukhanf6e24802020-07-08 22:20:40 -0700897 "src/f32-maxpool/9p8x-minmax-wasmsimd-arm-c4.c",
898 "src/f32-maxpool/9p8x-minmax-wasmsimd-x86-c4.c",
Marat Dukhan1483c532020-07-16 18:08:19 -0700899 "src/f32-pavgpool/9p8x-minmax-wasmsimd-arm-c4.c",
900 "src/f32-pavgpool/9p8x-minmax-wasmsimd-x86-c4.c",
901 "src/f32-pavgpool/9x-minmax-wasmsimd-arm-c4.c",
902 "src/f32-pavgpool/9x-minmax-wasmsimd-x86-c4.c",
Marat Dukhan0d0d8822020-07-23 23:37:56 -0700903 "src/f32-ppmm/gen/4x8-minmax-wasmsimd-splat-arm.c",
904 "src/f32-ppmm/gen/4x8-minmax-wasmsimd-splat-x86.c",
Frank Barcharda5316982020-07-23 13:19:28 -0700905 "src/f32-prelu/gen/wasmsimd-bitselect-1x4.c",
906 "src/f32-prelu/gen/wasmsimd-bitselect-1x8.c",
907 "src/f32-prelu/gen/wasmsimd-bitselect-1x16.c",
Marat Dukhan195f8eb2020-06-25 12:50:57 -0700908 "src/f32-prelu/gen/wasmsimd-bitselect-2x4.c",
909 "src/f32-prelu/gen/wasmsimd-bitselect-2x8.c",
Frank Barcharda5316982020-07-23 13:19:28 -0700910 "src/f32-prelu/gen/wasmsimd-bitselect-2x16.c",
911 "src/f32-prelu/gen/wasmsimd-bitselect-4x4.c",
912 "src/f32-prelu/gen/wasmsimd-bitselect-4x8.c",
913 "src/f32-prelu/gen/wasmsimd-bitselect-4x16.c",
914 "src/f32-prelu/gen/wasmsimd-minmax-1x4.c",
915 "src/f32-prelu/gen/wasmsimd-minmax-1x8.c",
916 "src/f32-prelu/gen/wasmsimd-minmax-1x16.c",
Marat Dukhan195f8eb2020-06-25 12:50:57 -0700917 "src/f32-prelu/gen/wasmsimd-minmax-2x4.c",
918 "src/f32-prelu/gen/wasmsimd-minmax-2x8.c",
Frank Barcharda5316982020-07-23 13:19:28 -0700919 "src/f32-prelu/gen/wasmsimd-minmax-2x16.c",
920 "src/f32-prelu/gen/wasmsimd-minmax-4x4.c",
921 "src/f32-prelu/gen/wasmsimd-minmax-4x8.c",
922 "src/f32-prelu/gen/wasmsimd-minmax-4x16.c",
Marat Dukhan52238f02020-07-16 15:30:28 -0700923 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x4.c",
Marat Dukhan52238f02020-07-16 15:30:28 -0700924 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x8-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700925 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x8.c",
Marat Dukhan52238f02020-07-16 15:30:28 -0700926 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x12-acc2.c",
927 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x12-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700928 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x12.c",
Marat Dukhan52238f02020-07-16 15:30:28 -0700929 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x16-acc2.c",
930 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x16-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700931 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x16.c",
Marat Dukhan52238f02020-07-16 15:30:28 -0700932 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x20-acc2.c",
933 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x20-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700934 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x20.c",
Frank Barchardfb158e22020-07-15 16:10:10 -0700935 "src/f32-relu/gen/wasmsimd-x4.c",
936 "src/f32-relu/gen/wasmsimd-x8.c",
Frank Barchard37297a62020-08-27 12:15:10 -0700937 "src/f32-relu/gen/wasmsimd-x16.c",
Marat Dukhan8c417962020-07-08 12:27:50 -0700938 "src/f32-rmax/wasmsimd-arm.c",
939 "src/f32-rmax/wasmsimd-x86.c",
Marat Dukhand187a5b2020-07-20 01:07:17 -0700940 "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x4.c",
941 "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x8.c",
942 "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x12.c",
943 "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x16.c",
944 "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x20.c",
945 "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x24.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700946 "src/f32-sigmoid/gen/wasmsimd-p5-div-x4.c",
947 "src/f32-sigmoid/gen/wasmsimd-p5-div-x8.c",
948 "src/f32-sigmoid/gen/wasmsimd-p5-div-x12.c",
949 "src/f32-sigmoid/gen/wasmsimd-p5-div-x16.c",
950 "src/f32-sigmoid/gen/wasmsimd-p5-div-x20.c",
951 "src/f32-sigmoid/gen/wasmsimd-p5-div-x24.c",
Frank Barchard8ef44cd2020-11-03 12:30:23 -0800952 "src/f32-spmm/gen/4x1-minmax-wasmsimd-arm-pipelined-x2.c",
953 "src/f32-spmm/gen/4x1-minmax-wasmsimd-arm-pipelined.c",
Frank Barchardbeca6522020-10-30 22:34:35 -0700954 "src/f32-spmm/gen/4x1-minmax-wasmsimd-arm-x2.c",
955 "src/f32-spmm/gen/4x1-minmax-wasmsimd-arm-x4.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700956 "src/f32-spmm/gen/4x1-minmax-wasmsimd-arm.c",
Frank Barchard8ef44cd2020-11-03 12:30:23 -0800957 "src/f32-spmm/gen/4x1-minmax-wasmsimd-x86-pipelined-x2.c",
958 "src/f32-spmm/gen/4x1-minmax-wasmsimd-x86-pipelined.c",
Frank Barchardbeca6522020-10-30 22:34:35 -0700959 "src/f32-spmm/gen/4x1-minmax-wasmsimd-x86-x2.c",
960 "src/f32-spmm/gen/4x1-minmax-wasmsimd-x86-x4.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700961 "src/f32-spmm/gen/4x1-minmax-wasmsimd-x86.c",
Frank Barchard8ef44cd2020-11-03 12:30:23 -0800962 "src/f32-spmm/gen/8x1-minmax-wasmsimd-arm-pipelined-x2.c",
963 "src/f32-spmm/gen/8x1-minmax-wasmsimd-arm-pipelined.c",
Frank Barchardbeca6522020-10-30 22:34:35 -0700964 "src/f32-spmm/gen/8x1-minmax-wasmsimd-arm-x2.c",
965 "src/f32-spmm/gen/8x1-minmax-wasmsimd-arm-x4.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700966 "src/f32-spmm/gen/8x1-minmax-wasmsimd-arm.c",
Frank Barchard8ef44cd2020-11-03 12:30:23 -0800967 "src/f32-spmm/gen/8x1-minmax-wasmsimd-x86-pipelined-x2.c",
968 "src/f32-spmm/gen/8x1-minmax-wasmsimd-x86-pipelined.c",
Frank Barchardbeca6522020-10-30 22:34:35 -0700969 "src/f32-spmm/gen/8x1-minmax-wasmsimd-x86-x2.c",
970 "src/f32-spmm/gen/8x1-minmax-wasmsimd-x86-x4.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700971 "src/f32-spmm/gen/8x1-minmax-wasmsimd-x86.c",
Frank Barchard8ef44cd2020-11-03 12:30:23 -0800972 "src/f32-spmm/gen/16x1-minmax-wasmsimd-arm-pipelined-x2.c",
973 "src/f32-spmm/gen/16x1-minmax-wasmsimd-arm-pipelined.c",
Frank Barchardbeca6522020-10-30 22:34:35 -0700974 "src/f32-spmm/gen/16x1-minmax-wasmsimd-arm-x2.c",
975 "src/f32-spmm/gen/16x1-minmax-wasmsimd-arm-x4.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700976 "src/f32-spmm/gen/16x1-minmax-wasmsimd-arm.c",
Frank Barchard8ef44cd2020-11-03 12:30:23 -0800977 "src/f32-spmm/gen/16x1-minmax-wasmsimd-x86-pipelined-x2.c",
978 "src/f32-spmm/gen/16x1-minmax-wasmsimd-x86-pipelined.c",
Frank Barchardbeca6522020-10-30 22:34:35 -0700979 "src/f32-spmm/gen/16x1-minmax-wasmsimd-x86-x2.c",
980 "src/f32-spmm/gen/16x1-minmax-wasmsimd-x86-x4.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700981 "src/f32-spmm/gen/16x1-minmax-wasmsimd-x86.c",
Frank Barchard8ef44cd2020-11-03 12:30:23 -0800982 "src/f32-spmm/gen/32x1-minmax-wasmsimd-arm-pipelined-x2.c",
983 "src/f32-spmm/gen/32x1-minmax-wasmsimd-arm-pipelined.c",
Frank Barchardbeca6522020-10-30 22:34:35 -0700984 "src/f32-spmm/gen/32x1-minmax-wasmsimd-arm-x2.c",
985 "src/f32-spmm/gen/32x1-minmax-wasmsimd-arm-x4.c",
Frank Barchard846c0c62020-10-26 15:01:39 -0700986 "src/f32-spmm/gen/32x1-minmax-wasmsimd-arm.c",
Frank Barchard8ef44cd2020-11-03 12:30:23 -0800987 "src/f32-spmm/gen/32x1-minmax-wasmsimd-x86-pipelined-x2.c",
988 "src/f32-spmm/gen/32x1-minmax-wasmsimd-x86-pipelined.c",
Frank Barchardbeca6522020-10-30 22:34:35 -0700989 "src/f32-spmm/gen/32x1-minmax-wasmsimd-x86-x2.c",
990 "src/f32-spmm/gen/32x1-minmax-wasmsimd-x86-x4.c",
Frank Barchard846c0c62020-10-26 15:01:39 -0700991 "src/f32-spmm/gen/32x1-minmax-wasmsimd-x86.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700992 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-arm-x4.c",
993 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700994 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-arm-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700995 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-x86-x4.c",
996 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700997 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700998 "src/f32-vbinary/gen/vadd-relu-wasmsimd-x4.c",
999 "src/f32-vbinary/gen/vadd-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001000 "src/f32-vbinary/gen/vadd-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07001001 "src/f32-vbinary/gen/vadd-wasmsimd-x4.c",
1002 "src/f32-vbinary/gen/vadd-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001003 "src/f32-vbinary/gen/vadd-wasmsimd-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001004 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-arm-x4.c",
1005 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001006 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-arm-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001007 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-x86-x4.c",
1008 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001009 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -07001010 "src/f32-vbinary/gen/vaddc-relu-wasmsimd-x4.c",
1011 "src/f32-vbinary/gen/vaddc-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001012 "src/f32-vbinary/gen/vaddc-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07001013 "src/f32-vbinary/gen/vaddc-wasmsimd-x4.c",
1014 "src/f32-vbinary/gen/vaddc-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001015 "src/f32-vbinary/gen/vaddc-wasmsimd-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -07001016 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-arm-x4.c",
1017 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001018 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-arm-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -07001019 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-x86-x4.c",
1020 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001021 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -07001022 "src/f32-vbinary/gen/vdiv-relu-wasmsimd-x4.c",
1023 "src/f32-vbinary/gen/vdiv-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001024 "src/f32-vbinary/gen/vdiv-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07001025 "src/f32-vbinary/gen/vdiv-wasmsimd-x4.c",
1026 "src/f32-vbinary/gen/vdiv-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001027 "src/f32-vbinary/gen/vdiv-wasmsimd-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001028 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-arm-x4.c",
1029 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001030 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-arm-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001031 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-x86-x4.c",
1032 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001033 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -07001034 "src/f32-vbinary/gen/vdivc-relu-wasmsimd-x4.c",
1035 "src/f32-vbinary/gen/vdivc-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001036 "src/f32-vbinary/gen/vdivc-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07001037 "src/f32-vbinary/gen/vdivc-wasmsimd-x4.c",
1038 "src/f32-vbinary/gen/vdivc-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001039 "src/f32-vbinary/gen/vdivc-wasmsimd-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -07001040 "src/f32-vbinary/gen/vmax-wasmsimd-arm-x4.c",
1041 "src/f32-vbinary/gen/vmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001042 "src/f32-vbinary/gen/vmax-wasmsimd-arm-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -07001043 "src/f32-vbinary/gen/vmax-wasmsimd-x86-x4.c",
1044 "src/f32-vbinary/gen/vmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001045 "src/f32-vbinary/gen/vmax-wasmsimd-x86-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001046 "src/f32-vbinary/gen/vmaxc-wasmsimd-arm-x4.c",
1047 "src/f32-vbinary/gen/vmaxc-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001048 "src/f32-vbinary/gen/vmaxc-wasmsimd-arm-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001049 "src/f32-vbinary/gen/vmaxc-wasmsimd-x86-x4.c",
1050 "src/f32-vbinary/gen/vmaxc-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001051 "src/f32-vbinary/gen/vmaxc-wasmsimd-x86-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -07001052 "src/f32-vbinary/gen/vmin-wasmsimd-arm-x4.c",
1053 "src/f32-vbinary/gen/vmin-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001054 "src/f32-vbinary/gen/vmin-wasmsimd-arm-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -07001055 "src/f32-vbinary/gen/vmin-wasmsimd-x86-x4.c",
1056 "src/f32-vbinary/gen/vmin-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001057 "src/f32-vbinary/gen/vmin-wasmsimd-x86-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001058 "src/f32-vbinary/gen/vminc-wasmsimd-arm-x4.c",
1059 "src/f32-vbinary/gen/vminc-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001060 "src/f32-vbinary/gen/vminc-wasmsimd-arm-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001061 "src/f32-vbinary/gen/vminc-wasmsimd-x86-x4.c",
1062 "src/f32-vbinary/gen/vminc-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001063 "src/f32-vbinary/gen/vminc-wasmsimd-x86-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -07001064 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-arm-x4.c",
1065 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001066 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-arm-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -07001067 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-x86-x4.c",
1068 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001069 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -07001070 "src/f32-vbinary/gen/vmul-relu-wasmsimd-x4.c",
1071 "src/f32-vbinary/gen/vmul-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001072 "src/f32-vbinary/gen/vmul-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07001073 "src/f32-vbinary/gen/vmul-wasmsimd-x4.c",
1074 "src/f32-vbinary/gen/vmul-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001075 "src/f32-vbinary/gen/vmul-wasmsimd-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001076 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-arm-x4.c",
1077 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001078 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-arm-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001079 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-x86-x4.c",
1080 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001081 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -07001082 "src/f32-vbinary/gen/vmulc-relu-wasmsimd-x4.c",
1083 "src/f32-vbinary/gen/vmulc-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001084 "src/f32-vbinary/gen/vmulc-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07001085 "src/f32-vbinary/gen/vmulc-wasmsimd-x4.c",
1086 "src/f32-vbinary/gen/vmulc-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001087 "src/f32-vbinary/gen/vmulc-wasmsimd-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001088 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-arm-x4.c",
1089 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001090 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-arm-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001091 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-x86-x4.c",
1092 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001093 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -07001094 "src/f32-vbinary/gen/vrdivc-relu-wasmsimd-x4.c",
1095 "src/f32-vbinary/gen/vrdivc-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001096 "src/f32-vbinary/gen/vrdivc-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07001097 "src/f32-vbinary/gen/vrdivc-wasmsimd-x4.c",
1098 "src/f32-vbinary/gen/vrdivc-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001099 "src/f32-vbinary/gen/vrdivc-wasmsimd-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001100 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-arm-x4.c",
1101 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001102 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-arm-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001103 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-x86-x4.c",
1104 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001105 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -07001106 "src/f32-vbinary/gen/vrsubc-relu-wasmsimd-x4.c",
1107 "src/f32-vbinary/gen/vrsubc-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001108 "src/f32-vbinary/gen/vrsubc-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07001109 "src/f32-vbinary/gen/vrsubc-wasmsimd-x4.c",
1110 "src/f32-vbinary/gen/vrsubc-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001111 "src/f32-vbinary/gen/vrsubc-wasmsimd-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -07001112 "src/f32-vbinary/gen/vsqrdiff-wasmsimd-x4.c",
1113 "src/f32-vbinary/gen/vsqrdiff-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001114 "src/f32-vbinary/gen/vsqrdiff-wasmsimd-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001115 "src/f32-vbinary/gen/vsqrdiffc-wasmsimd-x4.c",
1116 "src/f32-vbinary/gen/vsqrdiffc-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001117 "src/f32-vbinary/gen/vsqrdiffc-wasmsimd-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -07001118 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-arm-x4.c",
1119 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001120 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-arm-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -07001121 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-x86-x4.c",
1122 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001123 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -07001124 "src/f32-vbinary/gen/vsub-relu-wasmsimd-x4.c",
1125 "src/f32-vbinary/gen/vsub-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001126 "src/f32-vbinary/gen/vsub-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07001127 "src/f32-vbinary/gen/vsub-wasmsimd-x4.c",
1128 "src/f32-vbinary/gen/vsub-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001129 "src/f32-vbinary/gen/vsub-wasmsimd-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001130 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-arm-x4.c",
1131 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001132 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-arm-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001133 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-x86-x4.c",
1134 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001135 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -07001136 "src/f32-vbinary/gen/vsubc-relu-wasmsimd-x4.c",
1137 "src/f32-vbinary/gen/vsubc-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001138 "src/f32-vbinary/gen/vsubc-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07001139 "src/f32-vbinary/gen/vsubc-wasmsimd-x4.c",
1140 "src/f32-vbinary/gen/vsubc-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001141 "src/f32-vbinary/gen/vsubc-wasmsimd-x16.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07001142 "src/f32-vlrelu/gen/vlrelu-wasmsimd-bitselect-x4.c",
1143 "src/f32-vlrelu/gen/vlrelu-wasmsimd-bitselect-x8.c",
1144 "src/f32-vlrelu/gen/vlrelu-wasmsimd-minmax-x4.c",
1145 "src/f32-vlrelu/gen/vlrelu-wasmsimd-minmax-x8.c",
Marat Dukhand816f622020-07-15 10:14:39 -07001146 "src/f32-vmulcaddc/gen/c4-minmax-wasmsimd-arm-2x.c",
Marat Dukhand816f622020-07-15 10:14:39 -07001147 "src/f32-vmulcaddc/gen/c4-minmax-wasmsimd-x86-2x.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001148 "src/f32-vmulcaddc/gen/c8-minmax-wasmsimd-arm-2x.c",
Marat Dukhand816f622020-07-15 10:14:39 -07001149 "src/f32-vmulcaddc/gen/c8-minmax-wasmsimd-x86-2x.c",
Marat Dukhanb82b2cd2020-07-16 02:23:42 -07001150 "src/f32-vrnd/gen/vrndd-wasmsimd-addsub-x4.c",
1151 "src/f32-vrnd/gen/vrndd-wasmsimd-addsub-x8.c",
1152 "src/f32-vrnd/gen/vrndd-wasmsimd-cvt-x4.c",
1153 "src/f32-vrnd/gen/vrndd-wasmsimd-cvt-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001154 "src/f32-vrnd/gen/vrndne-wasmsimd-addsub-x4.c",
1155 "src/f32-vrnd/gen/vrndne-wasmsimd-addsub-x8.c",
1156 "src/f32-vrnd/gen/vrndu-wasmsimd-addsub-x4.c",
1157 "src/f32-vrnd/gen/vrndu-wasmsimd-addsub-x8.c",
1158 "src/f32-vrnd/gen/vrndu-wasmsimd-cvt-x4.c",
1159 "src/f32-vrnd/gen/vrndu-wasmsimd-cvt-x8.c",
1160 "src/f32-vrnd/gen/vrndz-wasmsimd-addsub-x4.c",
1161 "src/f32-vrnd/gen/vrndz-wasmsimd-addsub-x8.c",
1162 "src/f32-vrnd/gen/vrndz-wasmsimd-cvt-x4.c",
1163 "src/f32-vrnd/gen/vrndz-wasmsimd-cvt-x8.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07001164 "src/f32-vsqrt/gen/wasmsimd-sqrt-x4.c",
1165 "src/f32-vsqrt/gen/wasmsimd-sqrt-x8.c",
Marat Dukhan37c83512020-06-29 13:25:53 -07001166 "src/f32-vunary/gen/vabs-wasmsimd-x4.c",
1167 "src/f32-vunary/gen/vabs-wasmsimd-x8.c",
1168 "src/f32-vunary/gen/vneg-wasmsimd-x4.c",
1169 "src/f32-vunary/gen/vneg-wasmsimd-x8.c",
1170 "src/f32-vunary/gen/vsqr-wasmsimd-x4.c",
1171 "src/f32-vunary/gen/vsqr-wasmsimd-x8.c",
Marat Dukhande390d42020-11-29 19:32:18 -08001172 "src/math/expm1minus-wasmsimd-rr2-lut16-p3-andnot.c",
1173 "src/math/expm1minus-wasmsimd-rr2-lut16-p3-max.c",
1174 "src/math/expm1minus-wasmsimd-rr2-p6-andnot.c",
1175 "src/math/expm1minus-wasmsimd-rr2-p6-max.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001176 "src/math/roundd-wasmsimd-addsub.c",
1177 "src/math/roundd-wasmsimd-cvt.c",
1178 "src/math/roundne-wasmsimd-addsub.c",
1179 "src/math/roundu-wasmsimd-addsub.c",
1180 "src/math/roundu-wasmsimd-cvt.c",
1181 "src/math/roundz-wasmsimd-addsub.c",
1182 "src/math/roundz-wasmsimd-cvt.c",
1183 "src/math/sigmoid-wasmsimd-rr2-lut64-p2-div.c",
1184 "src/math/sigmoid-wasmsimd-rr2-p5-div.c",
Marat Dukhancc8f34c2020-08-05 16:36:38 -07001185 "src/qs8-dwconv/gen/up8x9-minmax-wasmsimd-mul16.c",
1186 "src/qs8-dwconv/gen/up16x9-minmax-wasmsimd-mul16.c",
1187 "src/qs8-dwconv/gen/up24x9-minmax-wasmsimd-mul16.c",
Marat Dukhanb5e3d172020-08-06 13:29:53 -07001188 "src/qs8-gavgpool/gen/7p7x-minmax-wasmsimd-c8-acc2.c",
1189 "src/qs8-gavgpool/gen/7p7x-minmax-wasmsimd-c16-acc2.c",
1190 "src/qs8-gavgpool/gen/7p7x-minmax-wasmsimd-c24-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001191 "src/qs8-gavgpool/gen/7x-minmax-wasmsimd-c8-acc2.c",
1192 "src/qs8-gavgpool/gen/7x-minmax-wasmsimd-c16-acc2.c",
1193 "src/qs8-gavgpool/gen/7x-minmax-wasmsimd-c24-acc2.c",
Marat Dukhan27203da2020-08-05 15:19:03 -07001194 "src/qs8-gemm/gen/1x4c8-minmax-wasmsimd-ld64.c",
Marat Dukhan27203da2020-08-05 15:19:03 -07001195 "src/qs8-gemm/gen/1x4c8-minmax-wasmsimd-ld128.c",
Marat Dukhan27203da2020-08-05 15:19:03 -07001196 "src/qs8-gemm/gen/1x4c8-xw-minmax-wasmsimd.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001197 "src/qs8-gemm/gen/2x4c8-minmax-wasmsimd-ld64.c",
1198 "src/qs8-gemm/gen/2x4c8-minmax-wasmsimd-ld128.c",
Marat Dukhan27203da2020-08-05 15:19:03 -07001199 "src/qs8-gemm/gen/2x4c8-xw-minmax-wasmsimd.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001200 "src/qs8-gemm/gen/3x4c8-minmax-wasmsimd-ld64.c",
1201 "src/qs8-gemm/gen/3x4c8-minmax-wasmsimd-ld128.c",
Marat Dukhan27203da2020-08-05 15:19:03 -07001202 "src/qs8-gemm/gen/3x4c8-xw-minmax-wasmsimd.c",
1203 "src/qs8-igemm/gen/1x4c8-minmax-wasmsimd-ld64.c",
Marat Dukhan27203da2020-08-05 15:19:03 -07001204 "src/qs8-igemm/gen/1x4c8-minmax-wasmsimd-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001205 "src/qs8-igemm/gen/2x4c8-minmax-wasmsimd-ld64.c",
Marat Dukhan27203da2020-08-05 15:19:03 -07001206 "src/qs8-igemm/gen/2x4c8-minmax-wasmsimd-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001207 "src/qs8-igemm/gen/3x4c8-minmax-wasmsimd-ld64.c",
Marat Dukhan27203da2020-08-05 15:19:03 -07001208 "src/qs8-igemm/gen/3x4c8-minmax-wasmsimd-ld128.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -07001209 "src/qs8-requantization/fp32-wasmsimd.c",
Marat Dukhan138560c2020-08-03 18:57:34 -07001210 "src/qs8-requantization/q31-wasmsimd.c",
Marat Dukhan5df27f82020-09-02 23:59:21 -07001211 "src/qs8-vadd/gen/minmax-wasmsimd-x8.c",
1212 "src/qs8-vadd/gen/minmax-wasmsimd-x16.c",
1213 "src/qs8-vadd/gen/minmax-wasmsimd-x24.c",
1214 "src/qs8-vadd/gen/minmax-wasmsimd-x32.c",
1215 "src/qs8-vaddc/gen/minmax-wasmsimd-x8.c",
1216 "src/qs8-vaddc/gen/minmax-wasmsimd-x16.c",
1217 "src/qs8-vaddc/gen/minmax-wasmsimd-x24.c",
1218 "src/qs8-vaddc/gen/minmax-wasmsimd-x32.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07001219 "src/qu8-requantization/fp32-wasmsimd.c",
Marat Dukhan22de5e72020-08-03 21:53:18 -07001220 "src/qu8-requantization/q31-wasmsimd.c",
Marat Dukhan8ee37012020-07-16 13:17:13 -07001221 "src/x32-fill/wasmsimd.c",
Marat Dukhan66d99e92020-07-16 12:56:21 -07001222 "src/x32-packx/x4-wasmsimd.c",
Marat Dukhan9306ae02020-07-16 15:51:13 -07001223 "src/x32-pad/wasmsimd.c",
Marat Dukhan9d4bfa22020-07-16 19:07:04 -07001224 "src/x32-unpool/wasmsimd.c",
Marat Dukhane3b78762020-07-16 20:02:58 -07001225 "src/x32-zip/x2-wasmsimd.c",
1226 "src/x32-zip/x3-wasmsimd.c",
1227 "src/x32-zip/x4-wasmsimd.c",
1228 "src/x32-zip/xm-wasmsimd.c",
Marat Dukhan290055c2020-06-09 12:24:29 -07001229]
1230
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001231PSIMD_ACCMATH_UKERNELS = [
Marat Dukhan2e23d2b2020-07-29 16:01:37 -07001232 "src/qs8-requantization/fp32-psimd.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001233 "src/qs8-requantization/precise-psimd.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07001234 "src/qu8-requantization/fp32-psimd.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001235 "src/qu8-requantization/precise-psimd.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001236]
1237
Marat Dukhan08c4a432019-10-03 09:29:21 -07001238# ISA-specific micro-kernels
1239NEON_UKERNELS = [
Marat Dukhanef25c6d2020-07-24 00:59:40 -07001240 "src/f32-argmaxpool/4x-neon-c4.c",
1241 "src/f32-argmaxpool/9p8x-neon-c4.c",
1242 "src/f32-argmaxpool/9x-neon-c4.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001243 "src/f32-avgpool/9p8x-minmax-neon-c4.c",
1244 "src/f32-avgpool/9x-minmax-neon-c4.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001245 "src/f32-clamp/gen/neon-x4.c",
1246 "src/f32-clamp/gen/neon-x8.c",
Marat Dukhan56b10cd2020-05-18 09:35:49 -07001247 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neon-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001248 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neon-2x2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001249 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neon-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001250 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neon-2x2.c",
Marat Dukhan56b10cd2020-05-18 09:35:49 -07001251 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neon-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001252 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neon-2x2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001253 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neon-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001254 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neon-2x2.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -07001255 "src/f32-dwconv/gen/up4x4-minmax-neon-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001256 "src/f32-dwconv/gen/up4x4-minmax-neon.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001257 "src/f32-dwconv/gen/up4x9-minmax-neon-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001258 "src/f32-dwconv/gen/up4x9-minmax-neon.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -07001259 "src/f32-dwconv/gen/up4x25-minmax-neon-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001260 "src/f32-dwconv/gen/up4x25-minmax-neon.c",
1261 "src/f32-dwconv/gen/up8x4-minmax-neon-acc2.c",
1262 "src/f32-dwconv/gen/up8x4-minmax-neon.c",
1263 "src/f32-dwconv/gen/up8x9-minmax-neon-acc2.c",
1264 "src/f32-dwconv/gen/up8x9-minmax-neon.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -07001265 "src/f32-dwconv/gen/up8x25-minmax-neon-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001266 "src/f32-dwconv/gen/up8x25-minmax-neon.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001267 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-1x4-acc2.c",
1268 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-1x4-acc3.c",
1269 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-1x4-acc4.c",
Marat Dukhanc581e482020-10-24 01:28:11 -07001270 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-1x4.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001271 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-2x4-acc2.c",
Marat Dukhanc581e482020-10-24 01:28:11 -07001272 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-2x4.c",
1273 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-3x4.c",
1274 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-4x4.c",
1275 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-5x4.c",
1276 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-6x4.c",
Marat Dukhan82f0c322020-10-25 19:17:35 -07001277 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neon-1x4-acc2.c",
1278 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neon-1x4-acc3.c",
1279 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neon-1x4-acc4.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001280 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neon-1x4.c",
Marat Dukhan82f0c322020-10-25 19:17:35 -07001281 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neon-2x4-acc2.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001282 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neon-2x4.c",
1283 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neon-3x4.c",
1284 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neon-4x4.c",
Marat Dukhan149f0ea2020-10-26 12:50:33 -07001285 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-1x4-acc2.c",
1286 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-1x4-acc3.c",
1287 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-1x4-acc4.c",
1288 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-1x4-acc5.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001289 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-1x4.c",
Marat Dukhan149f0ea2020-10-26 12:50:33 -07001290 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-2x4-acc2.c",
1291 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-2x4-acc3.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001292 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-2x4.c",
Marat Dukhan149f0ea2020-10-26 12:50:33 -07001293 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-3x4-acc2.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001294 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-3x4.c",
Marat Dukhan149f0ea2020-10-26 12:50:33 -07001295 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-4x4-acc2.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001296 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-4x4.c",
1297 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-5x4.c",
Marat Dukhan30d4b252020-10-29 16:33:22 -07001298 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neon-1x4-acc2.c",
1299 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neon-1x4-acc3.c",
1300 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neon-1x4-acc4.c",
1301 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neon-1x4-acc5.c",
1302 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neon-1x4.c",
1303 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neon-2x4-acc2.c",
1304 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neon-2x4-acc3.c",
1305 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neon-2x4.c",
Marat Dukhan30d4b252020-10-29 16:33:22 -07001306 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neon-3x4-acc2.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07001307 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neon-3x4.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -07001308 "src/f32-gavgpool-cw/neon-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001309 "src/f32-gavgpool/7p7x-minmax-neon-c4.c",
1310 "src/f32-gavgpool/7x-minmax-neon-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001311 "src/f32-gemm/gen-inc/1x8inc-minmax-neon-dup-ld64.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001312 "src/f32-gemm/gen-inc/1x8inc-minmax-neon-lane-ld64.c",
1313 "src/f32-gemm/gen-inc/1x8s4inc-minmax-neon.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001314 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-dup-ld64.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001315 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-dup-ld128.c",
1316 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-lane-ld64.c",
1317 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-lane-ld128.c",
1318 "src/f32-gemm/gen-inc/4x8s4inc-minmax-neon.c",
1319 "src/f32-gemm/gen-inc/5x8inc-minmax-neon-lane-ld64.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001320 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-dup-ld64.c",
1321 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-dup-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001322 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-lane-ld64.c",
1323 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-lane-ld128.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001324 "src/f32-gemm/gen-inc/6x8s4inc-minmax-neon.c",
1325 "src/f32-gemm/gen-inc/8x8s4inc-minmax-neon.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001326 "src/f32-gemm/gen/1x8-minmax-neon-dup-ld64.c",
1327 "src/f32-gemm/gen/1x8-minmax-neon-lane-ld64.c",
1328 "src/f32-gemm/gen/1x8s4-minmax-neon.c",
1329 "src/f32-gemm/gen/4x2-minmax-neon-lane-ld64.c",
1330 "src/f32-gemm/gen/4x8-minmax-neon-dup-ld64.c",
1331 "src/f32-gemm/gen/4x8-minmax-neon-dup-ld128.c",
1332 "src/f32-gemm/gen/4x8-minmax-neon-lane-ld64.c",
1333 "src/f32-gemm/gen/4x8-minmax-neon-lane-ld128.c",
1334 "src/f32-gemm/gen/4x8s4-minmax-neon.c",
1335 "src/f32-gemm/gen/5x8-minmax-neon-lane-ld64.c",
1336 "src/f32-gemm/gen/6x8-minmax-neon-dup-ld64.c",
1337 "src/f32-gemm/gen/6x8-minmax-neon-dup-ld128.c",
1338 "src/f32-gemm/gen/6x8-minmax-neon-lane-ld64.c",
1339 "src/f32-gemm/gen/6x8-minmax-neon-lane-ld128.c",
1340 "src/f32-gemm/gen/6x8s4-minmax-neon.c",
1341 "src/f32-gemm/gen/8x8s4-minmax-neon.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07001342 "src/f32-hswish/gen/hswish-neon-x4.c",
1343 "src/f32-hswish/gen/hswish-neon-x8.c",
1344 "src/f32-hswish/gen/hswish-neon-x16.c",
Marat Dukhan660fd192020-03-10 04:55:30 -07001345 "src/f32-ibilinear/gen/neon-c4.c",
1346 "src/f32-ibilinear/gen/neon-c8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001347 "src/f32-igemm/gen/1x8-minmax-neon-dup-ld64.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001348 "src/f32-igemm/gen/1x8-minmax-neon-lane-ld64.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001349 "src/f32-igemm/gen/1x8s4-minmax-neon.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001350 "src/f32-igemm/gen/4x2-minmax-neon-lane-ld64.c",
1351 "src/f32-igemm/gen/4x4-minmax-neon-lane-ld64.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001352 "src/f32-igemm/gen/4x8-minmax-neon-dup-ld64.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001353 "src/f32-igemm/gen/4x8-minmax-neon-dup-ld128.c",
1354 "src/f32-igemm/gen/4x8-minmax-neon-lane-ld64.c",
1355 "src/f32-igemm/gen/4x8-minmax-neon-lane-ld128.c",
1356 "src/f32-igemm/gen/4x8s4-minmax-neon.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001357 "src/f32-igemm/gen/6x8-minmax-neon-dup-ld64.c",
1358 "src/f32-igemm/gen/6x8-minmax-neon-dup-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001359 "src/f32-igemm/gen/6x8-minmax-neon-lane-ld64.c",
1360 "src/f32-igemm/gen/6x8-minmax-neon-lane-ld128.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001361 "src/f32-igemm/gen/6x8s4-minmax-neon.c",
1362 "src/f32-igemm/gen/8x8s4-minmax-neon.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001363 "src/f32-maxpool/9p8x-minmax-neon-c4.c",
1364 "src/f32-pavgpool/9p8x-minmax-neon-c4.c",
1365 "src/f32-pavgpool/9x-minmax-neon-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001366 "src/f32-ppmm/gen/4x8-minmax-neon.c",
1367 "src/f32-ppmm/gen/8x8-minmax-neon.c",
Frank Barcharda5316982020-07-23 13:19:28 -07001368 "src/f32-prelu/gen/neon-1x4.c",
1369 "src/f32-prelu/gen/neon-1x8.c",
1370 "src/f32-prelu/gen/neon-1x16.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001371 "src/f32-prelu/gen/neon-2x4.c",
1372 "src/f32-prelu/gen/neon-2x8.c",
Frank Barcharda5316982020-07-23 13:19:28 -07001373 "src/f32-prelu/gen/neon-2x16.c",
1374 "src/f32-prelu/gen/neon-4x4.c",
1375 "src/f32-prelu/gen/neon-4x8.c",
1376 "src/f32-prelu/gen/neon-4x16.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001377 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x4.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001378 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x8-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001379 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x8.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001380 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12-acc2.c",
1381 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001382 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001383 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16-acc2.c",
1384 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001385 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001386 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20-acc2.c",
1387 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001388 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20.c",
1389 "src/f32-raddstoreexpminusmax/gen/neon-p5-x4.c",
1390 "src/f32-raddstoreexpminusmax/gen/neon-p5-x8-acc2.c",
1391 "src/f32-raddstoreexpminusmax/gen/neon-p5-x8.c",
1392 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12-acc2.c",
1393 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12-acc3.c",
1394 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12.c",
1395 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16-acc2.c",
1396 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16-acc4.c",
1397 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16.c",
1398 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20-acc2.c",
1399 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20-acc5.c",
1400 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07001401 "src/f32-relu/gen/neon-x4.c",
1402 "src/f32-relu/gen/neon-x8.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001403 "src/f32-rmax/neon.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001404 "src/f32-sigmoid/gen/neon-frac-p9-p10-nr1recps-x16.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -08001405 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x4.c",
1406 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x8.c",
1407 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x12.c",
1408 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x16.c",
1409 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x20.c",
1410 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x24.c",
1411 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x4.c",
1412 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x8.c",
1413 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x12.c",
1414 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x16.c",
1415 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x20.c",
1416 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x24.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001417 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x4.c",
1418 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x8.c",
1419 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x12.c",
1420 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x16.c",
1421 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x20.c",
1422 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x24.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001423 "src/f32-vbinary/gen/vadd-minmax-neon-x4.c",
1424 "src/f32-vbinary/gen/vadd-minmax-neon-x8.c",
1425 "src/f32-vbinary/gen/vaddc-minmax-neon-x4.c",
1426 "src/f32-vbinary/gen/vaddc-minmax-neon-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -08001427 "src/f32-vbinary/gen/vmax-neon-x4.c",
1428 "src/f32-vbinary/gen/vmax-neon-x8.c",
1429 "src/f32-vbinary/gen/vmaxc-neon-x4.c",
1430 "src/f32-vbinary/gen/vmaxc-neon-x8.c",
1431 "src/f32-vbinary/gen/vmin-neon-x4.c",
1432 "src/f32-vbinary/gen/vmin-neon-x8.c",
1433 "src/f32-vbinary/gen/vminc-neon-x4.c",
1434 "src/f32-vbinary/gen/vminc-neon-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001435 "src/f32-vbinary/gen/vmul-minmax-neon-x4.c",
1436 "src/f32-vbinary/gen/vmul-minmax-neon-x8.c",
1437 "src/f32-vbinary/gen/vmulc-minmax-neon-x4.c",
1438 "src/f32-vbinary/gen/vmulc-minmax-neon-x8.c",
1439 "src/f32-vbinary/gen/vrsubc-minmax-neon-x4.c",
1440 "src/f32-vbinary/gen/vrsubc-minmax-neon-x8.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07001441 "src/f32-vbinary/gen/vsqrdiff-neon-x4.c",
1442 "src/f32-vbinary/gen/vsqrdiff-neon-x8.c",
1443 "src/f32-vbinary/gen/vsqrdiffc-neon-x4.c",
1444 "src/f32-vbinary/gen/vsqrdiffc-neon-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001445 "src/f32-vbinary/gen/vsub-minmax-neon-x4.c",
1446 "src/f32-vbinary/gen/vsub-minmax-neon-x8.c",
1447 "src/f32-vbinary/gen/vsubc-minmax-neon-x4.c",
1448 "src/f32-vbinary/gen/vsubc-minmax-neon-x8.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07001449 "src/f32-vlrelu/gen/vlrelu-neon-x4.c",
1450 "src/f32-vlrelu/gen/vlrelu-neon-x8.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001451 "src/f32-vmulcaddc/gen/c4-minmax-neon-2x.c",
1452 "src/f32-vmulcaddc/gen/c8-minmax-neon-2x.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07001453 "src/f32-vrnd/gen/vrndd-neon-x4.c",
1454 "src/f32-vrnd/gen/vrndd-neon-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001455 "src/f32-vrnd/gen/vrndne-neon-x4.c",
1456 "src/f32-vrnd/gen/vrndne-neon-x8.c",
1457 "src/f32-vrnd/gen/vrndu-neon-x4.c",
1458 "src/f32-vrnd/gen/vrndu-neon-x8.c",
1459 "src/f32-vrnd/gen/vrndz-neon-x4.c",
1460 "src/f32-vrnd/gen/vrndz-neon-x8.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07001461 "src/f32-vunary/gen/vabs-neon-x4.c",
1462 "src/f32-vunary/gen/vabs-neon-x8.c",
1463 "src/f32-vunary/gen/vneg-neon-x4.c",
1464 "src/f32-vunary/gen/vneg-neon-x8.c",
1465 "src/f32-vunary/gen/vsqr-neon-x4.c",
1466 "src/f32-vunary/gen/vsqr-neon-x8.c",
Marat Dukhande390d42020-11-29 19:32:18 -08001467 "src/math/expm1minus-neon-rr2-lut16-p3.c",
1468 "src/math/expm1minus-neon-rr2-p6.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001469 "src/math/roundd-neon-addsub.c",
1470 "src/math/roundd-neon-cvt.c",
1471 "src/math/roundne-neon-addsub.c",
1472 "src/math/roundu-neon-addsub.c",
1473 "src/math/roundu-neon-cvt.c",
1474 "src/math/roundz-neon-addsub.c",
1475 "src/math/roundz-neon-cvt.c",
1476 "src/math/sigmoid-neon-frac-p9-p10-nr1recps.c",
1477 "src/math/sigmoid-neon-rr1-lut64-p2-nr2recps.c",
1478 "src/math/sigmoid-neon-rr1-lut2048-p1-nr2recps.c",
1479 "src/math/sigmoid-neon-rr1-p5-nr2recps.c",
1480 "src/math/sigmoid-neon-rr2-lut64-p2-nr2recps.c",
1481 "src/math/sigmoid-neon-rr2-lut2048-p1-nr2recps.c",
1482 "src/math/sigmoid-neon-rr2-p5-nr2recps.c",
1483 "src/math/sqrt-neon-nr1rsqrts.c",
1484 "src/math/sqrt-neon-nr2rsqrts.c",
1485 "src/math/sqrt-neon-nr3rsqrts.c",
Marat Dukhan023bcf92020-08-10 12:40:50 -07001486 "src/qs8-dwconv/gen/up8x9-minmax-neon-mul16.c",
1487 "src/qs8-dwconv/gen/up16x9-minmax-neon-mul16.c",
1488 "src/qs8-dwconv/gen/up24x9-minmax-neon-mul16.c",
1489 "src/qs8-dwconv/gen/up32x9-minmax-neon-mul16.c",
Marat Dukhan281262d2020-08-10 13:23:21 -07001490 "src/qs8-gavgpool/gen/7p7x-minmax-neon-c8-acc2.c",
1491 "src/qs8-gavgpool/gen/7p7x-minmax-neon-c16-acc2.c",
1492 "src/qs8-gavgpool/gen/7p7x-minmax-neon-c24-acc2.c",
1493 "src/qs8-gavgpool/gen/7p7x-minmax-neon-c32-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001494 "src/qs8-gavgpool/gen/7x-minmax-neon-c8-acc2.c",
1495 "src/qs8-gavgpool/gen/7x-minmax-neon-c16-acc2.c",
1496 "src/qs8-gavgpool/gen/7x-minmax-neon-c24-acc2.c",
1497 "src/qs8-gavgpool/gen/7x-minmax-neon-c32-acc2.c",
Marat Dukhan40bbafe2020-08-04 02:04:22 -07001498 "src/qs8-gemm/gen/1x8-minmax-neon-mlal-lane.c",
Marat Dukhan40bbafe2020-08-04 02:04:22 -07001499 "src/qs8-gemm/gen/1x16-minmax-neon-mlal-lane.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001500 "src/qs8-gemm/gen/2x8-minmax-neon-mlal-lane.c",
Marat Dukhan40bbafe2020-08-04 02:04:22 -07001501 "src/qs8-gemm/gen/2x16-minmax-neon-mlal-lane.c",
1502 "src/qs8-igemm/gen/1x8-minmax-neon-mlal-lane.c",
Marat Dukhan40bbafe2020-08-04 02:04:22 -07001503 "src/qs8-igemm/gen/1x16-minmax-neon-mlal-lane.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001504 "src/qs8-igemm/gen/2x8-minmax-neon-mlal-lane.c",
Marat Dukhan40bbafe2020-08-04 02:04:22 -07001505 "src/qs8-igemm/gen/2x16-minmax-neon-mlal-lane.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -07001506 "src/qs8-requantization/fp32-neon.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001507 "src/qs8-requantization/precise-neon.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -07001508 "src/qs8-requantization/q31-neon.c",
Marat Dukhanba7b2792020-09-02 14:26:45 -07001509 "src/qs8-vadd/gen/minmax-neon-ld64-x8.c",
1510 "src/qs8-vadd/gen/minmax-neon-ld64-x16.c",
1511 "src/qs8-vadd/gen/minmax-neon-ld64-x24.c",
1512 "src/qs8-vadd/gen/minmax-neon-ld64-x32.c",
1513 "src/qs8-vaddc/gen/minmax-neon-ld64-x8.c",
1514 "src/qs8-vaddc/gen/minmax-neon-ld64-x16.c",
1515 "src/qs8-vaddc/gen/minmax-neon-ld64-x24.c",
1516 "src/qs8-vaddc/gen/minmax-neon-ld64-x32.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -07001517 "src/qu8-avgpool/9p8x-minmax-neon-c8.c",
1518 "src/qu8-avgpool/9x-minmax-neon-c8.c",
1519 "src/qu8-dwconv/up8x9-minmax-neon.c",
1520 "src/qu8-gavgpool/7p7x-minmax-neon-c8.c",
1521 "src/qu8-gavgpool/7x-minmax-neon-c8.c",
1522 "src/qu8-gemm/4x8-minmax-neon.c",
1523 "src/qu8-gemm/8x8-minmax-neon.c",
1524 "src/qu8-igemm/4x8-minmax-neon.c",
1525 "src/qu8-igemm/8x8-minmax-neon.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07001526 "src/qu8-requantization/fp32-neon.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001527 "src/qu8-requantization/precise-neon.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07001528 "src/qu8-requantization/q31-neon.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -07001529 "src/qu8-vadd/minmax-neon.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001530 "src/u8-clamp/neon-x64.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001531 "src/u8-maxpool/9p8x-minmax-neon-c16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001532 "src/u8-rmax/neon.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001533 "src/x8-zip/x2-neon.c",
1534 "src/x8-zip/x3-neon.c",
1535 "src/x8-zip/x4-neon.c",
1536 "src/x8-zip/xm-neon.c",
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -07001537 "src/x32-fill/neon.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001538 "src/x32-packx/x4-neon-st4.c",
Marat Dukhan63523d42020-05-22 17:07:33 -07001539 "src/x32-pad/neon.c",
Marat Dukhan57dccd82020-04-14 00:53:10 -07001540 "src/x32-unpool/neon.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001541 "src/x32-zip/x2-neon.c",
1542 "src/x32-zip/x3-neon.c",
1543 "src/x32-zip/x4-neon.c",
1544 "src/x32-zip/xm-neon.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001545]
1546
1547NEONFMA_UKERNELS = [
Frank Barchard04336c12020-10-22 16:48:55 -07001548 "src/f32-dwconv/gen/up4x4-minmax-neonfma-acc2.c",
1549 "src/f32-dwconv/gen/up4x4-minmax-neonfma.c",
1550 "src/f32-dwconv/gen/up4x9-minmax-neonfma-acc2.c",
1551 "src/f32-dwconv/gen/up4x9-minmax-neonfma.c",
1552 "src/f32-dwconv/gen/up4x25-minmax-neonfma-acc2.c",
1553 "src/f32-dwconv/gen/up4x25-minmax-neonfma.c",
1554 "src/f32-dwconv/gen/up8x4-minmax-neonfma-acc2.c",
1555 "src/f32-dwconv/gen/up8x4-minmax-neonfma.c",
1556 "src/f32-dwconv/gen/up8x9-minmax-neonfma-acc2.c",
1557 "src/f32-dwconv/gen/up8x9-minmax-neonfma.c",
1558 "src/f32-dwconv/gen/up8x25-minmax-neonfma-acc2.c",
1559 "src/f32-dwconv/gen/up8x25-minmax-neonfma.c",
1560 "src/f32-gemm/gen-inc/1x8inc-minmax-neonfma-dup-ld64.c",
1561 "src/f32-gemm/gen-inc/1x8s4inc-minmax-neonfma.c",
1562 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-dup-ld64.c",
1563 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-dup-ld128.c",
1564 "src/f32-gemm/gen-inc/4x8s4inc-minmax-neonfma.c",
1565 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-dup-ld64.c",
1566 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-dup-ld128.c",
1567 "src/f32-gemm/gen-inc/6x8s4inc-minmax-neonfma.c",
1568 "src/f32-gemm/gen-inc/8x8s4inc-minmax-neonfma.c",
1569 "src/f32-gemm/gen/1x8-minmax-neonfma-dup-ld64.c",
1570 "src/f32-gemm/gen/1x8s4-minmax-neonfma.c",
1571 "src/f32-gemm/gen/4x8-minmax-neonfma-dup-ld64.c",
1572 "src/f32-gemm/gen/4x8-minmax-neonfma-dup-ld128.c",
1573 "src/f32-gemm/gen/4x8s4-minmax-neonfma.c",
1574 "src/f32-gemm/gen/6x8-minmax-neonfma-dup-ld64.c",
1575 "src/f32-gemm/gen/6x8-minmax-neonfma-dup-ld128.c",
1576 "src/f32-gemm/gen/6x8s4-minmax-neonfma.c",
1577 "src/f32-gemm/gen/8x8s4-minmax-neonfma.c",
Marat Dukhan660fd192020-03-10 04:55:30 -07001578 "src/f32-ibilinear/gen/neonfma-c4.c",
1579 "src/f32-ibilinear/gen/neonfma-c8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001580 "src/f32-igemm/gen/1x8-minmax-neonfma-dup-ld64.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001581 "src/f32-igemm/gen/1x8s4-minmax-neonfma.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001582 "src/f32-igemm/gen/4x8-minmax-neonfma-dup-ld64.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001583 "src/f32-igemm/gen/4x8-minmax-neonfma-dup-ld128.c",
1584 "src/f32-igemm/gen/4x8s4-minmax-neonfma.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001585 "src/f32-igemm/gen/6x8-minmax-neonfma-dup-ld64.c",
1586 "src/f32-igemm/gen/6x8-minmax-neonfma-dup-ld128.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001587 "src/f32-igemm/gen/6x8s4-minmax-neonfma.c",
1588 "src/f32-igemm/gen/8x8s4-minmax-neonfma.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001589 "src/f32-ppmm/gen/4x8-minmax-neonfma.c",
1590 "src/f32-ppmm/gen/8x8-minmax-neonfma.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001591 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x4.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001592 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x8-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001593 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x8.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001594 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12-acc2.c",
1595 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001596 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001597 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16-acc2.c",
1598 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001599 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001600 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20-acc2.c",
1601 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001602 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20.c",
1603 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x4.c",
1604 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x8-acc2.c",
1605 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x8.c",
1606 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12-acc2.c",
1607 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12-acc3.c",
1608 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12.c",
1609 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16-acc2.c",
1610 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16-acc4.c",
1611 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16.c",
1612 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20-acc2.c",
1613 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20-acc5.c",
1614 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -08001615 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x4.c",
1616 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x8.c",
1617 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x12.c",
1618 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x16.c",
1619 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x20.c",
1620 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x24.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001621 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x4.c",
1622 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x8.c",
1623 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x12.c",
1624 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x16.c",
1625 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x20.c",
1626 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x24.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -08001627 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x4.c",
1628 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x8.c",
1629 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x12.c",
1630 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x16.c",
1631 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x20.c",
1632 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x24.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -08001633 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x4.c",
1634 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x8.c",
1635 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x12.c",
1636 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x16.c",
1637 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x20.c",
1638 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x24.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001639 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x4.c",
1640 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x8.c",
1641 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x12.c",
1642 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x16.c",
1643 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x20.c",
1644 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x24.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -08001645 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x4.c",
1646 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x8.c",
1647 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x12.c",
1648 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x16.c",
1649 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x20.c",
1650 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x24.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001651 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x4.c",
1652 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x8.c",
1653 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x12.c",
1654 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x16.c",
1655 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x20.c",
1656 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x24.c",
1657 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x4.c",
1658 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x8.c",
1659 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x12.c",
1660 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x16.c",
1661 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x20.c",
1662 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x24.c",
1663 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x4.c",
1664 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x8.c",
1665 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x12.c",
1666 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x16.c",
1667 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x20.c",
1668 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x24.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001669 "src/f32-vmulcaddc/gen/c4-minmax-neonfma-2x.c",
1670 "src/f32-vmulcaddc/gen/c8-minmax-neonfma-2x.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07001671 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x4.c",
1672 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x8.c",
1673 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x12.c",
1674 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x16.c",
1675 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x20.c",
1676 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x24.c",
1677 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x28.c",
1678 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x32.c",
1679 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x36.c",
1680 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x40.c",
1681 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x4.c",
1682 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x8.c",
1683 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x12.c",
1684 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x16.c",
1685 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x20.c",
1686 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x24.c",
1687 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x28.c",
1688 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x32.c",
1689 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x36.c",
1690 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x40.c",
Marat Dukhanb7633f22020-11-20 16:34:56 -08001691 "src/math/exp-neonfma-rr2-lut64-p2.c",
1692 "src/math/exp-neonfma-rr2-p5.c",
Marat Dukhan9dd119a2020-11-20 18:20:04 -08001693 "src/math/expminus-neonfma-rr2-lut64-p2.c",
1694 "src/math/expminus-neonfma-rr2-lut2048-p1.c",
1695 "src/math/expminus-neonfma-rr2-p5.c",
Marat Dukhande390d42020-11-29 19:32:18 -08001696 "src/math/expm1minus-neonfma-rr1-lut16-p3.c",
1697 "src/math/expm1minus-neonfma-rr1-p6.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001698 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr1recps1fma.c",
1699 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr2fma.c",
1700 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr2recps.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001701 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr1recps1fma.c",
1702 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr2fma.c",
1703 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr2recps.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001704 "src/math/sigmoid-neonfma-rr1-p5-nr1recps1fma.c",
1705 "src/math/sigmoid-neonfma-rr1-p5-nr2fma.c",
1706 "src/math/sigmoid-neonfma-rr1-p5-nr2recps.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001707 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr1recps1fma.c",
1708 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr2fma.c",
1709 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr2recps.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001710 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr1recps1fma.c",
1711 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr2fma.c",
1712 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr2recps.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001713 "src/math/sigmoid-neonfma-rr2-p5-nr1recps1fma.c",
1714 "src/math/sigmoid-neonfma-rr2-p5-nr2fma.c",
1715 "src/math/sigmoid-neonfma-rr2-p5-nr2recps.c",
Marat Dukhan84000762020-06-29 18:38:43 -07001716 "src/math/sqrt-neonfma-nr1fma.c",
Marat Dukhan84000762020-06-29 18:38:43 -07001717 "src/math/sqrt-neonfma-nr1rsqrts1fma1adj.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001718 "src/math/sqrt-neonfma-nr2fma.c",
1719 "src/math/sqrt-neonfma-nr2fma1adj.c",
1720 "src/math/sqrt-neonfma-nr3fma.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001721]
1722
1723AARCH64_NEONFMA_UKERNELS = [
Marat Dukhan56b10cd2020-05-18 09:35:49 -07001724 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neonfma-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001725 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neonfma-2x2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001726 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neonfma-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001727 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neonfma-2x2.c",
Marat Dukhan56b10cd2020-05-18 09:35:49 -07001728 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neonfma-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001729 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neonfma-2x2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001730 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neonfma-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001731 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neonfma-2x2.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -07001732 "src/f32-conv-hwc2chw/3x3s2p1c3x4-neonfma-2x2.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001733 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-1x4-acc2.c",
1734 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-1x4-acc3.c",
1735 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-1x4-acc4.c",
Marat Dukhan1268a242020-10-24 00:36:32 -07001736 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-1x4.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001737 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-2x4-acc2.c",
Marat Dukhan1268a242020-10-24 00:36:32 -07001738 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-2x4.c",
1739 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-3x4.c",
1740 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-4x4.c",
1741 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-5x4.c",
1742 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-6x4.c",
Marat Dukhan82f0c322020-10-25 19:17:35 -07001743 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neonfma-1x4-acc2.c",
1744 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neonfma-1x4-acc3.c",
1745 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neonfma-1x4-acc4.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001746 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neonfma-1x4.c",
Marat Dukhan82f0c322020-10-25 19:17:35 -07001747 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neonfma-2x4-acc2.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001748 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neonfma-2x4.c",
1749 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neonfma-3x4.c",
1750 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neonfma-4x4.c",
Marat Dukhan149f0ea2020-10-26 12:50:33 -07001751 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-1x4-acc2.c",
1752 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-1x4-acc3.c",
1753 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-1x4-acc4.c",
1754 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-1x4-acc5.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001755 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-1x4.c",
Marat Dukhan149f0ea2020-10-26 12:50:33 -07001756 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-2x4-acc2.c",
1757 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-2x4-acc3.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001758 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-2x4.c",
Marat Dukhan149f0ea2020-10-26 12:50:33 -07001759 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-3x4-acc2.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001760 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-3x4.c",
Marat Dukhan149f0ea2020-10-26 12:50:33 -07001761 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-4x4-acc2.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001762 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-4x4.c",
1763 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-5x4.c",
Marat Dukhan30d4b252020-10-29 16:33:22 -07001764 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neonfma-1x4-acc2.c",
1765 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neonfma-1x4-acc3.c",
1766 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neonfma-1x4-acc4.c",
1767 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neonfma-1x4-acc5.c",
1768 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neonfma-1x4.c",
1769 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neonfma-2x4-acc2.c",
1770 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neonfma-2x4-acc3.c",
1771 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neonfma-2x4.c",
Marat Dukhan30d4b252020-10-29 16:33:22 -07001772 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neonfma-3x4-acc2.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07001773 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neonfma-3x4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001774 "src/f32-gemm/gen-inc/1x8inc-minmax-neonfma-lane-ld64.c",
1775 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-lane-ld64.c",
1776 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-lane-ld128.c",
1777 "src/f32-gemm/gen-inc/5x8inc-minmax-neonfma-lane-ld64.c",
1778 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-lane-ld64.c",
1779 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-lane-ld128.c",
1780 "src/f32-gemm/gen/1x8-minmax-neonfma-lane-ld64.c",
1781 "src/f32-gemm/gen/4x2-minmax-neonfma-lane-ld64.c",
1782 "src/f32-gemm/gen/4x8-minmax-neonfma-lane-ld64.c",
1783 "src/f32-gemm/gen/4x8-minmax-neonfma-lane-ld128.c",
1784 "src/f32-gemm/gen/5x8-minmax-neonfma-lane-ld64.c",
1785 "src/f32-gemm/gen/6x8-minmax-neonfma-lane-ld64.c",
1786 "src/f32-gemm/gen/6x8-minmax-neonfma-lane-ld128.c",
1787 "src/f32-igemm/gen/1x8-minmax-neonfma-lane-ld64.c",
1788 "src/f32-igemm/gen/4x2-minmax-neonfma-lane-ld64.c",
1789 "src/f32-igemm/gen/4x4-minmax-neonfma-lane-ld64.c",
1790 "src/f32-igemm/gen/4x8-minmax-neonfma-lane-ld64.c",
1791 "src/f32-igemm/gen/4x8-minmax-neonfma-lane-ld128.c",
1792 "src/f32-igemm/gen/6x8-minmax-neonfma-lane-ld64.c",
1793 "src/f32-igemm/gen/6x8-minmax-neonfma-lane-ld128.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -08001794 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x4.c",
1795 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x8.c",
1796 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x12.c",
1797 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x16.c",
1798 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x20.c",
1799 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x24.c",
1800 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x4.c",
1801 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x8.c",
1802 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x12.c",
1803 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x16.c",
1804 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x20.c",
1805 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x24.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001806 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x4.c",
1807 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x8.c",
1808 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x12.c",
1809 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x16.c",
1810 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x20.c",
1811 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x24.c",
Marat Dukhan355ab432020-04-09 19:01:52 -07001812 "src/f32-spmm/gen/4x1-minmax-neonfma-pipelined.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07001813 "src/f32-spmm/gen/4x1-minmax-neonfma-x2.c",
Marat Dukhan355ab432020-04-09 19:01:52 -07001814 "src/f32-spmm/gen/4x1-minmax-neonfma.c",
1815 "src/f32-spmm/gen/4x2-minmax-neonfma.c",
1816 "src/f32-spmm/gen/4x4-minmax-neonfma.c",
1817 "src/f32-spmm/gen/8x1-minmax-neonfma-pipelined.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07001818 "src/f32-spmm/gen/8x1-minmax-neonfma-x2.c",
Marat Dukhan355ab432020-04-09 19:01:52 -07001819 "src/f32-spmm/gen/8x1-minmax-neonfma.c",
1820 "src/f32-spmm/gen/8x2-minmax-neonfma.c",
1821 "src/f32-spmm/gen/8x4-minmax-neonfma.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001822 "src/f32-spmm/gen/12x1-minmax-neonfma.c",
1823 "src/f32-spmm/gen/12x2-minmax-neonfma.c",
1824 "src/f32-spmm/gen/12x4-minmax-neonfma.c",
1825 "src/f32-spmm/gen/16x1-minmax-neonfma-pipelined.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07001826 "src/f32-spmm/gen/16x1-minmax-neonfma-x2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001827 "src/f32-spmm/gen/16x1-minmax-neonfma.c",
1828 "src/f32-spmm/gen/16x2-minmax-neonfma.c",
1829 "src/f32-spmm/gen/16x4-minmax-neonfma.c",
Frank Barchard846c0c62020-10-26 15:01:39 -07001830 "src/f32-spmm/gen/32x1-minmax-neonfma-pipelined.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07001831 "src/f32-spmm/gen/32x1-minmax-neonfma-x2.c",
Frank Barchard846c0c62020-10-26 15:01:39 -07001832 "src/f32-spmm/gen/32x1-minmax-neonfma.c",
1833 "src/f32-spmm/gen/32x2-minmax-neonfma.c",
1834 "src/f32-spmm/gen/32x4-minmax-neonfma.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001835 "src/f32-vbinary/gen/vdiv-minmax-neon-x4.c",
1836 "src/f32-vbinary/gen/vdiv-minmax-neon-x8.c",
1837 "src/f32-vbinary/gen/vdivc-minmax-neon-x4.c",
1838 "src/f32-vbinary/gen/vdivc-minmax-neon-x8.c",
1839 "src/f32-vbinary/gen/vrdivc-minmax-neon-x4.c",
1840 "src/f32-vbinary/gen/vrdivc-minmax-neon-x8.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07001841 "src/f32-vsqrt/gen/neon-sqrt-x4.c",
1842 "src/f32-vsqrt/gen/neon-sqrt-x8.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001843 "src/math/sigmoid-neonfma-rr1-lut64-p2-div.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001844 "src/math/sigmoid-neonfma-rr1-lut2048-p1-div.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001845 "src/math/sigmoid-neonfma-rr1-p5-div.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001846 "src/math/sigmoid-neonfma-rr2-lut64-p2-div.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001847 "src/math/sigmoid-neonfma-rr2-lut2048-p1-div.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001848 "src/math/sigmoid-neonfma-rr2-p5-div.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001849]
1850
Marat Dukhan8853b822020-05-07 12:19:01 -07001851NEONV8_UKERNELS = [
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07001852 "src/f32-vrnd/gen/vrndd-neonv8-x4.c",
1853 "src/f32-vrnd/gen/vrndd-neonv8-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001854 "src/f32-vrnd/gen/vrndne-neonv8-x4.c",
1855 "src/f32-vrnd/gen/vrndne-neonv8-x8.c",
1856 "src/f32-vrnd/gen/vrndu-neonv8-x4.c",
1857 "src/f32-vrnd/gen/vrndu-neonv8-x8.c",
1858 "src/f32-vrnd/gen/vrndz-neonv8-x4.c",
1859 "src/f32-vrnd/gen/vrndz-neonv8-x8.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001860 "src/math/roundd-neonv8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001861 "src/math/roundne-neonv8.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001862 "src/math/roundu-neonv8.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07001863 "src/math/roundz-neonv8.c",
Marat Dukhan8853b822020-05-07 12:19:01 -07001864]
1865
Marat Dukhan08c4a432019-10-03 09:29:21 -07001866AARCH64_NEONFP16ARITH_UKERNELS = [
Frank Barchardb1966592020-05-12 13:47:06 -07001867 "src/f16-clamp/gen/neonfp16arith-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001868 "src/f16-clamp/gen/neonfp16arith-x16.c",
Frank Barchard5a599a62020-06-04 20:12:44 -07001869 "src/f16-dwconv/gen/up8x4-minmax-neonfp16arith-acc2.c",
1870 "src/f16-dwconv/gen/up8x4-minmax-neonfp16arith.c",
1871 "src/f16-dwconv/gen/up8x9-minmax-neonfp16arith-acc2.c",
1872 "src/f16-dwconv/gen/up8x9-minmax-neonfp16arith.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001873 "src/f16-dwconv/gen/up8x25-minmax-neonfp16arith-acc2.c",
1874 "src/f16-dwconv/gen/up8x25-minmax-neonfp16arith.c",
1875 "src/f16-dwconv/gen/up16x4-minmax-neonfp16arith-acc2.c",
1876 "src/f16-dwconv/gen/up16x4-minmax-neonfp16arith.c",
1877 "src/f16-dwconv/gen/up16x9-minmax-neonfp16arith-acc2.c",
1878 "src/f16-dwconv/gen/up16x9-minmax-neonfp16arith.c",
1879 "src/f16-dwconv/gen/up16x25-minmax-neonfp16arith-acc2.c",
1880 "src/f16-dwconv/gen/up16x25-minmax-neonfp16arith.c",
Frank Barchard0bb49a72020-06-04 11:35:11 -07001881 "src/f16-gavgpool/7p7x-minmax-neonfp16arith-c8.c",
1882 "src/f16-gavgpool/7x-minmax-neonfp16arith-c8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001883 "src/f16-gemm/gen-inc/1x8inc-minmax-neonfp16arith-ld64.c",
1884 "src/f16-gemm/gen-inc/1x16inc-minmax-neonfp16arith-ld64.c",
1885 "src/f16-gemm/gen-inc/4x8inc-minmax-neonfp16arith-ld64.c",
1886 "src/f16-gemm/gen-inc/4x16inc-minmax-neonfp16arith-ld64.c",
1887 "src/f16-gemm/gen-inc/6x8inc-minmax-neonfp16arith-ld64.c",
1888 "src/f16-gemm/gen-inc/6x16inc-minmax-neonfp16arith-ld64.c",
1889 "src/f16-gemm/gen-inc/8x8inc-minmax-neonfp16arith-ld64.c",
1890 "src/f16-gemm/gen-inc/8x16inc-minmax-neonfp16arith-ld64.c",
1891 "src/f16-gemm/gen/1x8-minmax-neonfp16arith-ld64.c",
1892 "src/f16-gemm/gen/1x16-minmax-neonfp16arith-ld64.c",
1893 "src/f16-gemm/gen/4x8-minmax-neonfp16arith-ld64.c",
1894 "src/f16-gemm/gen/4x16-minmax-neonfp16arith-ld64.c",
1895 "src/f16-gemm/gen/6x8-minmax-neonfp16arith-ld64.c",
1896 "src/f16-gemm/gen/6x16-minmax-neonfp16arith-ld64.c",
1897 "src/f16-gemm/gen/8x8-minmax-neonfp16arith-ld64.c",
1898 "src/f16-gemm/gen/8x16-minmax-neonfp16arith-ld64.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07001899 "src/f16-hswish/gen/hswish-neonfp16arith-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001900 "src/f16-hswish/gen/hswish-neonfp16arith-x16.c",
1901 "src/f16-igemm/gen/1x8-minmax-neonfp16arith-ld64.c",
1902 "src/f16-igemm/gen/1x16-minmax-neonfp16arith-ld64.c",
1903 "src/f16-igemm/gen/4x8-minmax-neonfp16arith-ld64.c",
1904 "src/f16-igemm/gen/4x16-minmax-neonfp16arith-ld64.c",
1905 "src/f16-igemm/gen/6x8-minmax-neonfp16arith-ld64.c",
1906 "src/f16-igemm/gen/6x16-minmax-neonfp16arith-ld64.c",
1907 "src/f16-igemm/gen/8x8-minmax-neonfp16arith-ld64.c",
1908 "src/f16-igemm/gen/8x16-minmax-neonfp16arith-ld64.c",
Frank Barchardb1966592020-05-12 13:47:06 -07001909 "src/f16-prelu/gen/neonfp16arith-2x8.c",
Frank Barcharda5316982020-07-23 13:19:28 -07001910 "src/f16-prelu/gen/neonfp16arith-2x16.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07001911 "src/f16-relu/gen/neonfp16arith-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001912 "src/f16-relu/gen/neonfp16arith-x16.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07001913 "src/f16-spmm/gen/8x1-minmax-neonfp16arith-x2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001914 "src/f16-spmm/gen/8x1-minmax-neonfp16arith.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07001915 "src/f16-spmm/gen/16x1-minmax-neonfp16arith-x2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001916 "src/f16-spmm/gen/16x1-minmax-neonfp16arith.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07001917 "src/f16-spmm/gen/24x1-minmax-neonfp16arith-x2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001918 "src/f16-spmm/gen/24x1-minmax-neonfp16arith.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07001919 "src/f16-spmm/gen/32x1-minmax-neonfp16arith-x2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001920 "src/f16-spmm/gen/32x1-minmax-neonfp16arith.c",
1921 "src/f16-vbinary/gen/vadd-minmax-neonfp16arith-x8.c",
1922 "src/f16-vbinary/gen/vadd-minmax-neonfp16arith-x16.c",
1923 "src/f16-vbinary/gen/vaddc-minmax-neonfp16arith-x8.c",
1924 "src/f16-vbinary/gen/vaddc-minmax-neonfp16arith-x16.c",
1925 "src/f16-vbinary/gen/vdiv-minmax-neonfp16arith-x8.c",
1926 "src/f16-vbinary/gen/vdiv-minmax-neonfp16arith-x16.c",
1927 "src/f16-vbinary/gen/vdivc-minmax-neonfp16arith-x8.c",
1928 "src/f16-vbinary/gen/vdivc-minmax-neonfp16arith-x16.c",
1929 "src/f16-vbinary/gen/vmax-neonfp16arith-x8.c",
1930 "src/f16-vbinary/gen/vmax-neonfp16arith-x16.c",
1931 "src/f16-vbinary/gen/vmaxc-neonfp16arith-x8.c",
1932 "src/f16-vbinary/gen/vmaxc-neonfp16arith-x16.c",
1933 "src/f16-vbinary/gen/vmin-neonfp16arith-x8.c",
1934 "src/f16-vbinary/gen/vmin-neonfp16arith-x16.c",
1935 "src/f16-vbinary/gen/vminc-neonfp16arith-x8.c",
1936 "src/f16-vbinary/gen/vminc-neonfp16arith-x16.c",
1937 "src/f16-vbinary/gen/vmul-minmax-neonfp16arith-x8.c",
1938 "src/f16-vbinary/gen/vmul-minmax-neonfp16arith-x16.c",
1939 "src/f16-vbinary/gen/vmulc-minmax-neonfp16arith-x8.c",
1940 "src/f16-vbinary/gen/vmulc-minmax-neonfp16arith-x16.c",
1941 "src/f16-vbinary/gen/vrdivc-minmax-neonfp16arith-x8.c",
1942 "src/f16-vbinary/gen/vrdivc-minmax-neonfp16arith-x16.c",
1943 "src/f16-vbinary/gen/vrsubc-minmax-neonfp16arith-x8.c",
1944 "src/f16-vbinary/gen/vrsubc-minmax-neonfp16arith-x16.c",
1945 "src/f16-vbinary/gen/vsub-minmax-neonfp16arith-x8.c",
1946 "src/f16-vbinary/gen/vsub-minmax-neonfp16arith-x16.c",
1947 "src/f16-vbinary/gen/vsubc-minmax-neonfp16arith-x8.c",
1948 "src/f16-vbinary/gen/vsubc-minmax-neonfp16arith-x16.c",
1949 "src/f16-vmulcaddc/gen/c8-minmax-neonfp16arith-2x.c",
1950 "src/f16-vmulcaddc/gen/c16-minmax-neonfp16arith-2x.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001951]
1952
Benoit Jacoba9644732020-08-13 12:48:55 -07001953NEONDOT_UKERNELS = [
Marat Dukhand4c83032020-09-09 23:20:12 -07001954 "src/qs8-gemm/gen/1x8c4-minmax-neondot.c",
Marat Dukhand4c83032020-09-09 23:20:12 -07001955 "src/qs8-gemm/gen/1x16c4-minmax-neondot.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001956 "src/qs8-gemm/gen/4x8c4-minmax-neondot.c",
Benoit Jacoba9644732020-08-13 12:48:55 -07001957 "src/qs8-gemm/gen/4x16c4-minmax-neondot.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001958 "src/qs8-gemm/gen/6x8c4-minmax-neondot.c",
Frank Barchard2fa17452020-09-22 09:51:46 -07001959 "src/qs8-gemm/gen/6x16c4-minmax-neondot.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001960 "src/qs8-gemm/gen/8x8c4-minmax-neondot.c",
Frank Barcharda48848f2020-09-24 00:59:24 -07001961 "src/qs8-gemm/gen/8x16c4-minmax-neondot.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001962 "src/qs8-gemm/gen/12x8c4-minmax-neondot.c",
Marat Dukhand4c83032020-09-09 23:20:12 -07001963 "src/qs8-igemm/gen/1x8c4-minmax-neondot.c",
Marat Dukhand4c83032020-09-09 23:20:12 -07001964 "src/qs8-igemm/gen/1x16c4-minmax-neondot.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001965 "src/qs8-igemm/gen/4x8c4-minmax-neondot.c",
Benoit Jacoba05487f2020-08-14 13:37:21 -07001966 "src/qs8-igemm/gen/4x16c4-minmax-neondot.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001967 "src/qs8-igemm/gen/6x8c4-minmax-neondot.c",
Frank Barchard2fa17452020-09-22 09:51:46 -07001968 "src/qs8-igemm/gen/6x16c4-minmax-neondot.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001969 "src/qs8-igemm/gen/8x8c4-minmax-neondot.c",
Frank Barcharda48848f2020-09-24 00:59:24 -07001970 "src/qs8-igemm/gen/8x16c4-minmax-neondot.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001971 "src/qs8-igemm/gen/12x8c4-minmax-neondot.c",
Benoit Jacoba9644732020-08-13 12:48:55 -07001972]
1973
Marat Dukhan08c4a432019-10-03 09:29:21 -07001974SSE_UKERNELS = [
Marat Dukhan99936602020-04-11 16:47:01 -07001975 "src/f32-avgpool/9p8x-minmax-sse-c4.c",
1976 "src/f32-avgpool/9x-minmax-sse-c4.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001977 "src/f32-clamp/gen/sse-x4.c",
1978 "src/f32-clamp/gen/sse-x8.c",
Erich Elsenb1233402020-06-08 15:53:15 -07001979 "src/f32-conv-hwc2chw/3x3s2p1c3x4-sse-1x1.c",
1980 "src/f32-conv-hwc2chw/3x3s2p1c3x4-sse-2x2.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001981 "src/f32-dwconv/gen/up4x4-minmax-sse-acc2.c",
1982 "src/f32-dwconv/gen/up4x4-minmax-sse.c",
1983 "src/f32-dwconv/gen/up4x9-minmax-sse-acc2.c",
1984 "src/f32-dwconv/gen/up4x9-minmax-sse.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001985 "src/f32-dwconv/gen/up4x25-minmax-sse-acc2.c",
1986 "src/f32-dwconv/gen/up4x25-minmax-sse.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001987 "src/f32-dwconv/gen/up8x4-minmax-sse-acc2.c",
1988 "src/f32-dwconv/gen/up8x4-minmax-sse.c",
1989 "src/f32-dwconv/gen/up8x9-minmax-sse-acc2.c",
1990 "src/f32-dwconv/gen/up8x9-minmax-sse.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001991 "src/f32-dwconv/gen/up8x25-minmax-sse-acc2.c",
1992 "src/f32-dwconv/gen/up8x25-minmax-sse.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001993 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-1x4-acc2.c",
1994 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-1x4-acc3.c",
1995 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-1x4-acc4.c",
Marat Dukhan470078a2020-10-23 22:36:52 -07001996 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-1x4.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001997 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-2x4-acc2.c",
Marat Dukhan470078a2020-10-23 22:36:52 -07001998 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-2x4.c",
1999 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-3x4.c",
2000 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-4x4.c",
2001 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-5x4.c",
2002 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-6x4.c",
Marat Dukhan0ff97182020-10-25 19:14:03 -07002003 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-sse-1x4-acc2.c",
2004 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-sse-1x4-acc3.c",
2005 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-sse-1x4-acc4.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07002006 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-sse-1x4.c",
Marat Dukhan0ff97182020-10-25 19:14:03 -07002007 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-sse-2x4-acc2.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07002008 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-sse-2x4.c",
2009 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-sse-3x4.c",
2010 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-sse-4x4.c",
Marat Dukhand0503892020-10-30 08:22:04 -07002011 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-1x4-acc2.c",
2012 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-1x4-acc3.c",
2013 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-1x4-acc4.c",
2014 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-1x4-acc5.c",
2015 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-1x4.c",
2016 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-2x4-acc2.c",
2017 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-2x4-acc3.c",
2018 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-2x4.c",
2019 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-3x4-acc2.c",
2020 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-3x4.c",
2021 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-4x4-acc2.c",
2022 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-4x4.c",
2023 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-5x4.c",
Marat Dukhanccca2142020-10-30 17:32:45 -07002024 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-sse-1x4-acc2.c",
2025 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-sse-1x4-acc3.c",
2026 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-sse-1x4-acc4.c",
2027 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-sse-1x4-acc5.c",
2028 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-sse-1x4.c",
2029 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-sse-2x4-acc2.c",
2030 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-sse-2x4-acc3.c",
2031 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-sse-2x4.c",
Marat Dukhanccca2142020-10-30 17:32:45 -07002032 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-sse-3x4-acc2.c",
Frank Barchard8ef44cd2020-11-03 12:30:23 -08002033 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-sse-3x4.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -07002034 "src/f32-gavgpool-cw/sse-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -07002035 "src/f32-gavgpool/7p7x-minmax-sse-c4.c",
2036 "src/f32-gavgpool/7x-minmax-sse-c4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002037 "src/f32-gemm/gen-inc/1x8inc-minmax-sse-dup.c",
2038 "src/f32-gemm/gen-inc/1x8inc-minmax-sse-load1.c",
2039 "src/f32-gemm/gen-inc/1x8s4inc-minmax-sse.c",
2040 "src/f32-gemm/gen-inc/4x8inc-minmax-sse-dup.c",
2041 "src/f32-gemm/gen-inc/4x8inc-minmax-sse-load1.c",
2042 "src/f32-gemm/gen-inc/4x8s4inc-minmax-sse.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002043 "src/f32-gemm/gen/1x8-minmax-sse-dup.c",
2044 "src/f32-gemm/gen/1x8-minmax-sse-load1.c",
2045 "src/f32-gemm/gen/1x8s4-minmax-sse.c",
2046 "src/f32-gemm/gen/4x2c4-minmax-sse.c",
2047 "src/f32-gemm/gen/4x8-minmax-sse-dup.c",
2048 "src/f32-gemm/gen/4x8-minmax-sse-load1.c",
2049 "src/f32-gemm/gen/4x8s4-minmax-sse.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07002050 "src/f32-hswish/gen/hswish-sse-x4.c",
2051 "src/f32-hswish/gen/hswish-sse-x8.c",
Marat Dukhan660fd192020-03-10 04:55:30 -07002052 "src/f32-ibilinear/gen/sse-c4.c",
2053 "src/f32-ibilinear/gen/sse-c8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002054 "src/f32-igemm/gen/1x8-minmax-sse-dup.c",
2055 "src/f32-igemm/gen/1x8-minmax-sse-load1.c",
2056 "src/f32-igemm/gen/1x8s4-minmax-sse.c",
2057 "src/f32-igemm/gen/4x2c4-minmax-sse.c",
2058 "src/f32-igemm/gen/4x8-minmax-sse-dup.c",
2059 "src/f32-igemm/gen/4x8-minmax-sse-load1.c",
2060 "src/f32-igemm/gen/4x8s4-minmax-sse.c",
Marat Dukhan99936602020-04-11 16:47:01 -07002061 "src/f32-maxpool/9p8x-minmax-sse-c4.c",
2062 "src/f32-pavgpool/9p8x-minmax-sse-c4.c",
2063 "src/f32-pavgpool/9x-minmax-sse-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002064 "src/f32-ppmm/gen/4x8-minmax-sse.c",
Marat Dukhan39b5e942020-06-24 15:03:48 -07002065 "src/f32-prelu/gen/sse-2x4.c",
2066 "src/f32-prelu/gen/sse-2x8.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07002067 "src/f32-relu/gen/sse-x4.c",
2068 "src/f32-relu/gen/sse-x8.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002069 "src/f32-rmax/sse.c",
Marat Dukhan355ab432020-04-09 19:01:52 -07002070 "src/f32-spmm/gen/4x1-minmax-sse.c",
2071 "src/f32-spmm/gen/8x1-minmax-sse.c",
Erich Elsen6e80fdc2020-06-09 15:35:37 -07002072 "src/f32-spmm/gen/16x1-minmax-sse.c",
Frank Barchard846c0c62020-10-26 15:01:39 -07002073 "src/f32-spmm/gen/32x1-minmax-sse.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002074 "src/f32-vbinary/gen/vadd-minmax-sse-x4.c",
2075 "src/f32-vbinary/gen/vadd-minmax-sse-x8.c",
2076 "src/f32-vbinary/gen/vaddc-minmax-sse-x4.c",
2077 "src/f32-vbinary/gen/vaddc-minmax-sse-x8.c",
2078 "src/f32-vbinary/gen/vdiv-minmax-sse-x4.c",
2079 "src/f32-vbinary/gen/vdiv-minmax-sse-x8.c",
2080 "src/f32-vbinary/gen/vdivc-minmax-sse-x4.c",
2081 "src/f32-vbinary/gen/vdivc-minmax-sse-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -08002082 "src/f32-vbinary/gen/vmax-sse-x4.c",
2083 "src/f32-vbinary/gen/vmax-sse-x8.c",
2084 "src/f32-vbinary/gen/vmaxc-sse-x4.c",
2085 "src/f32-vbinary/gen/vmaxc-sse-x8.c",
2086 "src/f32-vbinary/gen/vmin-sse-x4.c",
2087 "src/f32-vbinary/gen/vmin-sse-x8.c",
2088 "src/f32-vbinary/gen/vminc-sse-x4.c",
2089 "src/f32-vbinary/gen/vminc-sse-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002090 "src/f32-vbinary/gen/vmul-minmax-sse-x4.c",
2091 "src/f32-vbinary/gen/vmul-minmax-sse-x8.c",
2092 "src/f32-vbinary/gen/vmulc-minmax-sse-x4.c",
2093 "src/f32-vbinary/gen/vmulc-minmax-sse-x8.c",
2094 "src/f32-vbinary/gen/vrdivc-minmax-sse-x4.c",
2095 "src/f32-vbinary/gen/vrdivc-minmax-sse-x8.c",
2096 "src/f32-vbinary/gen/vrsubc-minmax-sse-x4.c",
2097 "src/f32-vbinary/gen/vrsubc-minmax-sse-x8.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07002098 "src/f32-vbinary/gen/vsqrdiff-sse-x4.c",
2099 "src/f32-vbinary/gen/vsqrdiff-sse-x8.c",
2100 "src/f32-vbinary/gen/vsqrdiffc-sse-x4.c",
2101 "src/f32-vbinary/gen/vsqrdiffc-sse-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002102 "src/f32-vbinary/gen/vsub-minmax-sse-x4.c",
2103 "src/f32-vbinary/gen/vsub-minmax-sse-x8.c",
2104 "src/f32-vbinary/gen/vsubc-minmax-sse-x4.c",
2105 "src/f32-vbinary/gen/vsubc-minmax-sse-x8.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07002106 "src/f32-vlrelu/gen/vlrelu-sse-x4.c",
2107 "src/f32-vlrelu/gen/vlrelu-sse-x8.c",
Marat Dukhan99936602020-04-11 16:47:01 -07002108 "src/f32-vmulcaddc/gen/c4-minmax-sse-2x.c",
2109 "src/f32-vmulcaddc/gen/c8-minmax-sse-2x.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07002110 "src/f32-vsqrt/gen/sse-sqrt-x4.c",
2111 "src/f32-vsqrt/gen/sse-sqrt-x8.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07002112 "src/f32-vunary/gen/vabs-sse-x4.c",
2113 "src/f32-vunary/gen/vabs-sse-x8.c",
2114 "src/f32-vunary/gen/vneg-sse-x4.c",
2115 "src/f32-vunary/gen/vneg-sse-x8.c",
2116 "src/f32-vunary/gen/vsqr-sse-x4.c",
2117 "src/f32-vunary/gen/vsqr-sse-x8.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07002118 "src/math/roundd-sse-addsub.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002119 "src/math/roundne-sse-addsub.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07002120 "src/math/roundu-sse-addsub.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07002121 "src/math/roundz-sse-addsub.c",
Marat Dukhan84000762020-06-29 18:38:43 -07002122 "src/math/sqrt-sse-hh1mac.c",
2123 "src/math/sqrt-sse-nr1mac.c",
2124 "src/math/sqrt-sse-nr2mac.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002125 "src/x32-fill/sse.c",
2126 "src/x32-packx/x4-sse.c",
2127 "src/x32-pad/sse.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002128]
2129
2130SSE2_UKERNELS = [
Marat Dukhan329da642019-11-19 21:44:39 -08002131 "src/f32-argmaxpool/4x-sse2-c4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002132 "src/f32-argmaxpool/9p8x-sse2-c4.c",
Marat Dukhan329da642019-11-19 21:44:39 -08002133 "src/f32-argmaxpool/9x-sse2-c4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08002134 "src/f32-prelu/gen/sse2-2x4.c",
2135 "src/f32-prelu/gen/sse2-2x8.c",
Marat Dukhanb39689d2020-01-24 13:32:20 -08002136 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x4.c",
Marat Dukhanb39689d2020-01-24 13:32:20 -08002137 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x8-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002138 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x8.c",
Marat Dukhanb39689d2020-01-24 13:32:20 -08002139 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12-acc2.c",
2140 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002141 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12.c",
Marat Dukhanb39689d2020-01-24 13:32:20 -08002142 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16-acc2.c",
2143 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002144 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16.c",
Marat Dukhanb39689d2020-01-24 13:32:20 -08002145 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20-acc2.c",
2146 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002147 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20.c",
Marat Dukhand243c1a2020-09-17 18:48:10 -07002148 "src/f32-sigmoid/gen/sse2-lut64-p2-div-x4.c",
2149 "src/f32-sigmoid/gen/sse2-lut64-p2-div-x8.c",
2150 "src/f32-sigmoid/gen/sse2-lut64-p2-div-x12.c",
2151 "src/f32-sigmoid/gen/sse2-lut64-p2-div-x16.c",
2152 "src/f32-sigmoid/gen/sse2-lut64-p2-div-x20.c",
2153 "src/f32-sigmoid/gen/sse2-lut64-p2-div-x24.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002154 "src/f32-sigmoid/gen/sse2-p5-div-x4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08002155 "src/f32-sigmoid/gen/sse2-p5-div-x8.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002156 "src/f32-sigmoid/gen/sse2-p5-div-x12.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08002157 "src/f32-sigmoid/gen/sse2-p5-div-x16.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002158 "src/f32-sigmoid/gen/sse2-p5-div-x20.c",
2159 "src/f32-sigmoid/gen/sse2-p5-div-x24.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07002160 "src/f32-vlrelu/gen/vlrelu-sse2-x4.c",
2161 "src/f32-vlrelu/gen/vlrelu-sse2-x8.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07002162 "src/f32-vrnd/gen/vrndd-sse2-x4.c",
2163 "src/f32-vrnd/gen/vrndd-sse2-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002164 "src/f32-vrnd/gen/vrndne-sse2-x4.c",
2165 "src/f32-vrnd/gen/vrndne-sse2-x8.c",
2166 "src/f32-vrnd/gen/vrndu-sse2-x4.c",
2167 "src/f32-vrnd/gen/vrndu-sse2-x8.c",
2168 "src/f32-vrnd/gen/vrndz-sse2-x4.c",
2169 "src/f32-vrnd/gen/vrndz-sse2-x8.c",
Marat Dukhanb7633f22020-11-20 16:34:56 -08002170 "src/math/exp-sse2-rr2-lut64-p2.c",
2171 "src/math/exp-sse2-rr2-p5.c",
Marat Dukhande390d42020-11-29 19:32:18 -08002172 "src/math/expm1minus-sse2-rr2-lut16-p3.c",
Marat Dukhana438aca2020-11-20 15:45:01 -08002173 "src/math/expm1minus-sse2-rr2-p6.c",
Frank Barchard3b800452020-11-22 12:12:35 -08002174 "src/math/expminus-sse2-rr2-p5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002175 "src/math/roundd-sse2-cvt.c",
2176 "src/math/roundne-sse2-cvt.c",
2177 "src/math/roundu-sse2-cvt.c",
2178 "src/math/roundz-sse2-cvt.c",
2179 "src/math/sigmoid-sse2-rr2-lut64-p2-div.c",
2180 "src/math/sigmoid-sse2-rr2-lut64-p2-nr1.c",
2181 "src/math/sigmoid-sse2-rr2-lut64-p2-nr2.c",
2182 "src/math/sigmoid-sse2-rr2-p5-div.c",
2183 "src/math/sigmoid-sse2-rr2-p5-nr1.c",
2184 "src/math/sigmoid-sse2-rr2-p5-nr2.c",
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07002185 "src/qs8-dwconv/gen/up8x9-minmax-sse2-mul16.c",
2186 "src/qs8-dwconv/gen/up16x9-minmax-sse2-mul16.c",
2187 "src/qs8-dwconv/gen/up24x9-minmax-sse2-mul16.c",
Marat Dukhan159688f2020-08-06 10:34:29 -07002188 "src/qs8-gavgpool/gen/7p7x-minmax-sse2-c8-acc2.c",
2189 "src/qs8-gavgpool/gen/7p7x-minmax-sse2-c16-acc2.c",
2190 "src/qs8-gavgpool/gen/7p7x-minmax-sse2-c24-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002191 "src/qs8-gavgpool/gen/7x-minmax-sse2-c8-acc2.c",
2192 "src/qs8-gavgpool/gen/7x-minmax-sse2-c16-acc2.c",
2193 "src/qs8-gavgpool/gen/7x-minmax-sse2-c24-acc2.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002194 "src/qs8-gemm/gen/1x4c2-minmax-sse2-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002195 "src/qs8-gemm/gen/1x4c2-minmax-sse2-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002196 "src/qs8-gemm/gen/1x4c2-xw-minmax-sse2.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002197 "src/qs8-gemm/gen/1x4c8-minmax-sse2-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002198 "src/qs8-gemm/gen/1x4c8-minmax-sse2-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002199 "src/qs8-gemm/gen/1x4c8-xw-minmax-sse2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002200 "src/qs8-gemm/gen/2x4c8-minmax-sse2-ld64.c",
2201 "src/qs8-gemm/gen/2x4c8-minmax-sse2-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002202 "src/qs8-gemm/gen/2x4c8-xw-minmax-sse2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002203 "src/qs8-gemm/gen/3x4c8-minmax-sse2-ld64.c",
2204 "src/qs8-gemm/gen/3x4c8-minmax-sse2-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002205 "src/qs8-gemm/gen/3x4c8-xw-minmax-sse2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002206 "src/qs8-gemm/gen/4x4c2-minmax-sse2-ld64.c",
2207 "src/qs8-gemm/gen/4x4c2-minmax-sse2-ld128.c",
2208 "src/qs8-gemm/gen/4x4c2-xw-minmax-sse2.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002209 "src/qs8-igemm/gen/1x4c2-minmax-sse2-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002210 "src/qs8-igemm/gen/1x4c2-minmax-sse2-ld128.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002211 "src/qs8-igemm/gen/1x4c8-minmax-sse2-ld64.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002212 "src/qs8-igemm/gen/1x4c8-minmax-sse2-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002213 "src/qs8-igemm/gen/2x4c8-minmax-sse2-ld64.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002214 "src/qs8-igemm/gen/2x4c8-minmax-sse2-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002215 "src/qs8-igemm/gen/3x4c8-minmax-sse2-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002216 "src/qs8-igemm/gen/3x4c8-minmax-sse2-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002217 "src/qs8-igemm/gen/4x4c2-minmax-sse2-ld64.c",
2218 "src/qs8-igemm/gen/4x4c2-minmax-sse2-ld128.c",
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07002219 "src/qs8-requantization/fp32-sse2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002220 "src/qs8-requantization/precise-sse2.c",
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07002221 "src/qs8-requantization/q31-sse2.c",
Marat Dukhand9f3ad42020-08-10 12:30:58 -07002222 "src/qs8-vadd/gen/minmax-sse2-mul16-ld64-x8.c",
2223 "src/qs8-vadd/gen/minmax-sse2-mul16-ld64-x16.c",
2224 "src/qs8-vadd/gen/minmax-sse2-mul16-ld64-x24.c",
2225 "src/qs8-vadd/gen/minmax-sse2-mul16-ld64-x32.c",
Marat Dukhan0270d9f2020-08-11 00:56:46 -07002226 "src/qs8-vaddc/gen/minmax-sse2-mul16-ld64-x8.c",
2227 "src/qs8-vaddc/gen/minmax-sse2-mul16-ld64-x16.c",
2228 "src/qs8-vaddc/gen/minmax-sse2-mul16-ld64-x24.c",
2229 "src/qs8-vaddc/gen/minmax-sse2-mul16-ld64-x32.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -07002230 "src/qu8-avgpool/9p8x-minmax-sse2-c8.c",
2231 "src/qu8-avgpool/9x-minmax-sse2-c8.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -07002232 "src/qu8-dwconv/up8x9-minmax-sse2.c",
2233 "src/qu8-gavgpool/7p7x-minmax-sse2-c8.c",
2234 "src/qu8-gavgpool/7x-minmax-sse2-c8.c",
2235 "src/qu8-gemm/2x4c8-minmax-sse2.c",
2236 "src/qu8-gemm/4x4c2-minmax-sse2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002237 "src/qu8-igemm/4x4c2-minmax-sse2.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07002238 "src/qu8-requantization/fp32-sse2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002239 "src/qu8-requantization/precise-sse2.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07002240 "src/qu8-requantization/q31-sse2.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -07002241 "src/qu8-vadd/minmax-sse2.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -07002242 "src/u8-clamp/sse2-x64.c",
Marat Dukhan99936602020-04-11 16:47:01 -07002243 "src/u8-maxpool/9p8x-minmax-sse2-c16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002244 "src/u8-rmax/sse2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002245 "src/x8-zip/x2-sse2.c",
2246 "src/x8-zip/x3-sse2.c",
2247 "src/x8-zip/x4-sse2.c",
2248 "src/x8-zip/xm-sse2.c",
Marat Dukhan57dccd82020-04-14 00:53:10 -07002249 "src/x32-unpool/sse2.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002250 "src/x32-zip/x2-sse2.c",
2251 "src/x32-zip/x3-sse2.c",
2252 "src/x32-zip/x4-sse2.c",
2253 "src/x32-zip/xm-sse2.c",
Marat Dukhanfe7acb62020-03-09 19:30:05 -07002254]
2255
2256SSSE3_UKERNELS = [
Frank Barchard35db7d02020-10-26 13:37:34 -07002257 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-1x4-acc2.c",
2258 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-1x4-acc3.c",
2259 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-1x4-acc4.c",
Marat Dukhan98f2eeb2020-10-23 23:13:41 -07002260 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-1x4.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07002261 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-2x4-acc2.c",
Marat Dukhan98f2eeb2020-10-23 23:13:41 -07002262 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-2x4.c",
2263 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-3x4.c",
2264 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-4x4.c",
2265 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-5x4.c",
2266 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-6x4.c",
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07002267 "src/qs8-dwconv/gen/up8x9-minmax-ssse3-mul16.c",
2268 "src/qs8-dwconv/gen/up16x9-minmax-ssse3-mul16.c",
2269 "src/qs8-dwconv/gen/up24x9-minmax-ssse3-mul16.c",
Marat Dukhan159688f2020-08-06 10:34:29 -07002270 "src/qs8-gavgpool/gen/7p7x-minmax-ssse3-c8-acc2.c",
2271 "src/qs8-gavgpool/gen/7p7x-minmax-ssse3-c16-acc2.c",
2272 "src/qs8-gavgpool/gen/7p7x-minmax-ssse3-c24-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002273 "src/qs8-gavgpool/gen/7x-minmax-ssse3-c8-acc2.c",
2274 "src/qs8-gavgpool/gen/7x-minmax-ssse3-c16-acc2.c",
2275 "src/qs8-gavgpool/gen/7x-minmax-ssse3-c24-acc2.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002276 "src/qs8-gemm/gen/1x4c2-minmax-ssse3-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002277 "src/qs8-gemm/gen/1x4c2-minmax-ssse3-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002278 "src/qs8-gemm/gen/1x4c2-xw-minmax-ssse3.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002279 "src/qs8-gemm/gen/1x4c8-minmax-ssse3-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002280 "src/qs8-gemm/gen/1x4c8-minmax-ssse3-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002281 "src/qs8-gemm/gen/1x4c8-xw-minmax-ssse3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002282 "src/qs8-gemm/gen/2x4c8-minmax-ssse3-ld64.c",
2283 "src/qs8-gemm/gen/2x4c8-minmax-ssse3-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002284 "src/qs8-gemm/gen/2x4c8-xw-minmax-ssse3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002285 "src/qs8-gemm/gen/3x4c8-minmax-ssse3-ld64.c",
2286 "src/qs8-gemm/gen/3x4c8-minmax-ssse3-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002287 "src/qs8-gemm/gen/3x4c8-xw-minmax-ssse3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002288 "src/qs8-gemm/gen/4x4c2-minmax-ssse3-ld64.c",
2289 "src/qs8-gemm/gen/4x4c2-minmax-ssse3-ld128.c",
2290 "src/qs8-gemm/gen/4x4c2-xw-minmax-ssse3.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002291 "src/qs8-igemm/gen/1x4c2-minmax-ssse3-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002292 "src/qs8-igemm/gen/1x4c2-minmax-ssse3-ld128.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002293 "src/qs8-igemm/gen/1x4c8-minmax-ssse3-ld64.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002294 "src/qs8-igemm/gen/1x4c8-minmax-ssse3-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002295 "src/qs8-igemm/gen/2x4c8-minmax-ssse3-ld64.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002296 "src/qs8-igemm/gen/2x4c8-minmax-ssse3-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002297 "src/qs8-igemm/gen/3x4c8-minmax-ssse3-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002298 "src/qs8-igemm/gen/3x4c8-minmax-ssse3-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002299 "src/qs8-igemm/gen/4x4c2-minmax-ssse3-ld64.c",
2300 "src/qs8-igemm/gen/4x4c2-minmax-ssse3-ld128.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -07002301 "src/qs8-requantization/precise-ssse3.c",
2302 "src/qs8-requantization/q31-ssse3.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07002303 "src/qu8-requantization/precise-ssse3.c",
2304 "src/qu8-requantization/q31-ssse3.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002305]
2306
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08002307SSE41_UKERNELS = [
Marat Dukhan40a672f2019-11-25 03:08:22 -08002308 "src/f32-prelu/gen/sse41-2x4.c",
2309 "src/f32-prelu/gen/sse41-2x8.c",
Marat Dukhand243c1a2020-09-17 18:48:10 -07002310 "src/f32-sigmoid/gen/sse41-lut64-p2-div-x4.c",
2311 "src/f32-sigmoid/gen/sse41-lut64-p2-div-x8.c",
2312 "src/f32-sigmoid/gen/sse41-lut64-p2-div-x12.c",
2313 "src/f32-sigmoid/gen/sse41-lut64-p2-div-x16.c",
2314 "src/f32-sigmoid/gen/sse41-lut64-p2-div-x20.c",
2315 "src/f32-sigmoid/gen/sse41-lut64-p2-div-x24.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002316 "src/f32-sigmoid/gen/sse41-p5-div-x4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08002317 "src/f32-sigmoid/gen/sse41-p5-div-x8.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002318 "src/f32-sigmoid/gen/sse41-p5-div-x12.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08002319 "src/f32-sigmoid/gen/sse41-p5-div-x16.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002320 "src/f32-sigmoid/gen/sse41-p5-div-x20.c",
2321 "src/f32-sigmoid/gen/sse41-p5-div-x24.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07002322 "src/f32-vlrelu/gen/vlrelu-sse41-x4.c",
2323 "src/f32-vlrelu/gen/vlrelu-sse41-x8.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07002324 "src/f32-vrnd/gen/vrndd-sse41-x4.c",
2325 "src/f32-vrnd/gen/vrndd-sse41-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002326 "src/f32-vrnd/gen/vrndne-sse41-x4.c",
2327 "src/f32-vrnd/gen/vrndne-sse41-x8.c",
2328 "src/f32-vrnd/gen/vrndu-sse41-x4.c",
2329 "src/f32-vrnd/gen/vrndu-sse41-x8.c",
2330 "src/f32-vrnd/gen/vrndz-sse41-x4.c",
2331 "src/f32-vrnd/gen/vrndz-sse41-x8.c",
2332 "src/math/roundd-sse41.c",
2333 "src/math/roundne-sse41.c",
2334 "src/math/roundu-sse41.c",
2335 "src/math/roundz-sse41.c",
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07002336 "src/qs8-dwconv/gen/up8x9-minmax-sse41-mul16.c",
2337 "src/qs8-dwconv/gen/up16x9-minmax-sse41-mul16.c",
2338 "src/qs8-dwconv/gen/up24x9-minmax-sse41-mul16.c",
Marat Dukhan159688f2020-08-06 10:34:29 -07002339 "src/qs8-gavgpool/gen/7p7x-minmax-sse41-c8-acc2.c",
2340 "src/qs8-gavgpool/gen/7p7x-minmax-sse41-c16-acc2.c",
2341 "src/qs8-gavgpool/gen/7p7x-minmax-sse41-c24-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002342 "src/qs8-gavgpool/gen/7x-minmax-sse41-c8-acc2.c",
2343 "src/qs8-gavgpool/gen/7x-minmax-sse41-c16-acc2.c",
2344 "src/qs8-gavgpool/gen/7x-minmax-sse41-c24-acc2.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002345 "src/qs8-gemm/gen/1x4c2-minmax-sse41-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002346 "src/qs8-gemm/gen/1x4c2-minmax-sse41-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002347 "src/qs8-gemm/gen/1x4c2-xw-minmax-sse41.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002348 "src/qs8-gemm/gen/1x4c8-minmax-sse41-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002349 "src/qs8-gemm/gen/1x4c8-minmax-sse41-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002350 "src/qs8-gemm/gen/1x4c8-xw-minmax-sse41.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002351 "src/qs8-gemm/gen/2x4c8-minmax-sse41-ld64.c",
2352 "src/qs8-gemm/gen/2x4c8-minmax-sse41-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002353 "src/qs8-gemm/gen/2x4c8-xw-minmax-sse41.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002354 "src/qs8-gemm/gen/3x4c8-minmax-sse41-ld64.c",
2355 "src/qs8-gemm/gen/3x4c8-minmax-sse41-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002356 "src/qs8-gemm/gen/3x4c8-xw-minmax-sse41.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002357 "src/qs8-gemm/gen/4x4c2-minmax-sse41-ld64.c",
2358 "src/qs8-gemm/gen/4x4c2-minmax-sse41-ld128.c",
2359 "src/qs8-gemm/gen/4x4c2-xw-minmax-sse41.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002360 "src/qs8-igemm/gen/1x4c2-minmax-sse41-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002361 "src/qs8-igemm/gen/1x4c2-minmax-sse41-ld128.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002362 "src/qs8-igemm/gen/1x4c8-minmax-sse41-ld64.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002363 "src/qs8-igemm/gen/1x4c8-minmax-sse41-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002364 "src/qs8-igemm/gen/2x4c8-minmax-sse41-ld64.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002365 "src/qs8-igemm/gen/2x4c8-minmax-sse41-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002366 "src/qs8-igemm/gen/3x4c8-minmax-sse41-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002367 "src/qs8-igemm/gen/3x4c8-minmax-sse41-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002368 "src/qs8-igemm/gen/4x4c2-minmax-sse41-ld64.c",
2369 "src/qs8-igemm/gen/4x4c2-minmax-sse41-ld128.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -07002370 "src/qs8-requantization/fp32-sse4.c",
2371 "src/qs8-requantization/precise-sse4.c",
2372 "src/qs8-requantization/q31-sse4.c",
Marat Dukhand9f3ad42020-08-10 12:30:58 -07002373 "src/qs8-vadd/gen/minmax-sse41-mul16-ld64-x8.c",
2374 "src/qs8-vadd/gen/minmax-sse41-mul16-ld64-x16.c",
2375 "src/qs8-vadd/gen/minmax-sse41-mul16-ld64-x24.c",
2376 "src/qs8-vadd/gen/minmax-sse41-mul16-ld64-x32.c",
Marat Dukhanbb9225e2020-09-06 22:40:56 -07002377 "src/qs8-vadd/gen/minmax-sse41-mul32-ld32-x8.c",
2378 "src/qs8-vadd/gen/minmax-sse41-mul32-ld32-x16.c",
2379 "src/qs8-vadd/gen/minmax-sse41-mul32-ld32-x24.c",
2380 "src/qs8-vadd/gen/minmax-sse41-mul32-ld32-x32.c",
Marat Dukhan0270d9f2020-08-11 00:56:46 -07002381 "src/qs8-vaddc/gen/minmax-sse41-mul16-ld64-x8.c",
2382 "src/qs8-vaddc/gen/minmax-sse41-mul16-ld64-x16.c",
2383 "src/qs8-vaddc/gen/minmax-sse41-mul16-ld64-x24.c",
2384 "src/qs8-vaddc/gen/minmax-sse41-mul16-ld64-x32.c",
Marat Dukhanbb9225e2020-09-06 22:40:56 -07002385 "src/qs8-vaddc/gen/minmax-sse41-mul32-ld32-x8.c",
2386 "src/qs8-vaddc/gen/minmax-sse41-mul32-ld32-x16.c",
2387 "src/qs8-vaddc/gen/minmax-sse41-mul32-ld32-x24.c",
2388 "src/qs8-vaddc/gen/minmax-sse41-mul32-ld32-x32.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07002389 "src/qu8-requantization/precise-sse4.c",
2390 "src/qu8-requantization/q31-sse4.c",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08002391]
2392
Marat Dukhan08c4a432019-10-03 09:29:21 -07002393AVX_UKERNELS = [
Marat Dukhan5c5fa962020-03-10 18:38:33 -07002394 "src/f32-clamp/gen/avx-x8.c",
2395 "src/f32-clamp/gen/avx-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002396 "src/f32-dwconv/gen/up8x4-minmax-avx-acc2.c",
2397 "src/f32-dwconv/gen/up8x4-minmax-avx.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002398 "src/f32-dwconv/gen/up8x9-minmax-avx-acc2.c",
2399 "src/f32-dwconv/gen/up8x9-minmax-avx.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002400 "src/f32-dwconv/gen/up8x25-minmax-avx-acc2.c",
2401 "src/f32-dwconv/gen/up8x25-minmax-avx.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002402 "src/f32-dwconv/gen/up16x4-minmax-avx-acc2.c",
2403 "src/f32-dwconv/gen/up16x4-minmax-avx.c",
2404 "src/f32-dwconv/gen/up16x9-minmax-avx-acc2.c",
2405 "src/f32-dwconv/gen/up16x9-minmax-avx.c",
2406 "src/f32-dwconv/gen/up16x25-minmax-avx-acc2.c",
2407 "src/f32-dwconv/gen/up16x25-minmax-avx.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002408 "src/f32-gemm/gen-inc/1x8inc-minmax-avx-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002409 "src/f32-gemm/gen-inc/1x16inc-minmax-avx-broadcast.c",
2410 "src/f32-gemm/gen-inc/3x16inc-minmax-avx-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002411 "src/f32-gemm/gen-inc/4x8inc-minmax-avx-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002412 "src/f32-gemm/gen-inc/4x16inc-minmax-avx-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002413 "src/f32-gemm/gen-inc/5x8inc-minmax-avx-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002414 "src/f32-gemm/gen-inc/5x16inc-minmax-avx-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002415 "src/f32-gemm/gen-inc/6x8inc-minmax-avx-broadcast.c",
2416 "src/f32-gemm/gen-inc/7x8inc-minmax-avx-broadcast.c",
2417 "src/f32-gemm/gen/1x8-minmax-avx-broadcast.c",
2418 "src/f32-gemm/gen/1x16-minmax-avx-broadcast.c",
2419 "src/f32-gemm/gen/3x16-minmax-avx-broadcast.c",
2420 "src/f32-gemm/gen/4x8-minmax-avx-broadcast.c",
2421 "src/f32-gemm/gen/4x16-minmax-avx-broadcast.c",
2422 "src/f32-gemm/gen/5x8-minmax-avx-broadcast.c",
2423 "src/f32-gemm/gen/5x16-minmax-avx-broadcast.c",
2424 "src/f32-gemm/gen/6x8-minmax-avx-broadcast.c",
2425 "src/f32-gemm/gen/7x8-minmax-avx-broadcast.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07002426 "src/f32-hswish/gen/hswish-avx-x8.c",
2427 "src/f32-hswish/gen/hswish-avx-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002428 "src/f32-igemm/gen/1x8-minmax-avx-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002429 "src/f32-igemm/gen/1x16-minmax-avx-broadcast.c",
2430 "src/f32-igemm/gen/3x16-minmax-avx-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002431 "src/f32-igemm/gen/4x8-minmax-avx-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002432 "src/f32-igemm/gen/4x16-minmax-avx-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002433 "src/f32-igemm/gen/5x8-minmax-avx-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002434 "src/f32-igemm/gen/5x16-minmax-avx-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002435 "src/f32-igemm/gen/6x8-minmax-avx-broadcast.c",
2436 "src/f32-igemm/gen/7x8-minmax-avx-broadcast.c",
Marat Dukhan90eca0a2020-03-11 00:52:23 -07002437 "src/f32-prelu/gen/avx-2x8.c",
2438 "src/f32-prelu/gen/avx-2x16.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07002439 "src/f32-relu/gen/avx-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002440 "src/f32-relu/gen/avx-x16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002441 "src/f32-rmax/avx.c",
T.J. Alumbaughdc2b29c2020-10-14 13:56:08 -07002442 "src/f32-sigmoid/gen/avx-rr2-p5-div-x8.c",
2443 "src/f32-sigmoid/gen/avx-rr2-p5-div-x16.c",
2444 "src/f32-sigmoid/gen/avx-rr2-p5-div-x24.c",
2445 "src/f32-sigmoid/gen/avx-rr2-p5-div-x32.c",
2446 "src/f32-sigmoid/gen/avx-rr2-p5-div-x40.c",
2447 "src/f32-sigmoid/gen/avx-rr2-p5-div-x48.c",
2448 "src/f32-sigmoid/gen/avx-rr2-p5-div-x56.c",
2449 "src/f32-sigmoid/gen/avx-rr2-p5-div-x64.c",
2450 "src/f32-sigmoid/gen/avx-rr2-p5-div-x72.c",
2451 "src/f32-sigmoid/gen/avx-rr2-p5-div-x80.c",
2452 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x8.c",
2453 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x16.c",
2454 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x24.c",
2455 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x32.c",
2456 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x40.c",
2457 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x48.c",
2458 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x56.c",
2459 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x64.c",
2460 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x72.c",
2461 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x80.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002462 "src/f32-vbinary/gen/vadd-minmax-avx-x8.c",
2463 "src/f32-vbinary/gen/vadd-minmax-avx-x16.c",
2464 "src/f32-vbinary/gen/vaddc-minmax-avx-x8.c",
2465 "src/f32-vbinary/gen/vaddc-minmax-avx-x16.c",
2466 "src/f32-vbinary/gen/vdiv-minmax-avx-x8.c",
2467 "src/f32-vbinary/gen/vdiv-minmax-avx-x16.c",
2468 "src/f32-vbinary/gen/vdivc-minmax-avx-x8.c",
2469 "src/f32-vbinary/gen/vdivc-minmax-avx-x16.c",
Marat Dukhan9a88efe2019-12-10 15:54:24 -08002470 "src/f32-vbinary/gen/vmax-avx-x8.c",
2471 "src/f32-vbinary/gen/vmax-avx-x16.c",
2472 "src/f32-vbinary/gen/vmaxc-avx-x8.c",
2473 "src/f32-vbinary/gen/vmaxc-avx-x16.c",
2474 "src/f32-vbinary/gen/vmin-avx-x8.c",
2475 "src/f32-vbinary/gen/vmin-avx-x16.c",
2476 "src/f32-vbinary/gen/vminc-avx-x8.c",
2477 "src/f32-vbinary/gen/vminc-avx-x16.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002478 "src/f32-vbinary/gen/vmul-minmax-avx-x8.c",
2479 "src/f32-vbinary/gen/vmul-minmax-avx-x16.c",
2480 "src/f32-vbinary/gen/vmulc-minmax-avx-x8.c",
2481 "src/f32-vbinary/gen/vmulc-minmax-avx-x16.c",
2482 "src/f32-vbinary/gen/vrdivc-minmax-avx-x8.c",
2483 "src/f32-vbinary/gen/vrdivc-minmax-avx-x16.c",
2484 "src/f32-vbinary/gen/vrsubc-minmax-avx-x8.c",
2485 "src/f32-vbinary/gen/vrsubc-minmax-avx-x16.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07002486 "src/f32-vbinary/gen/vsqrdiff-avx-x8.c",
2487 "src/f32-vbinary/gen/vsqrdiff-avx-x16.c",
2488 "src/f32-vbinary/gen/vsqrdiffc-avx-x8.c",
2489 "src/f32-vbinary/gen/vsqrdiffc-avx-x16.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002490 "src/f32-vbinary/gen/vsub-minmax-avx-x8.c",
2491 "src/f32-vbinary/gen/vsub-minmax-avx-x16.c",
2492 "src/f32-vbinary/gen/vsubc-minmax-avx-x8.c",
2493 "src/f32-vbinary/gen/vsubc-minmax-avx-x16.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07002494 "src/f32-vlrelu/gen/vlrelu-avx-x8.c",
2495 "src/f32-vlrelu/gen/vlrelu-avx-x16.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07002496 "src/f32-vrnd/gen/vrndd-avx-x8.c",
2497 "src/f32-vrnd/gen/vrndd-avx-x16.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002498 "src/f32-vrnd/gen/vrndne-avx-x8.c",
2499 "src/f32-vrnd/gen/vrndne-avx-x16.c",
2500 "src/f32-vrnd/gen/vrndu-avx-x8.c",
2501 "src/f32-vrnd/gen/vrndu-avx-x16.c",
2502 "src/f32-vrnd/gen/vrndz-avx-x8.c",
2503 "src/f32-vrnd/gen/vrndz-avx-x16.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07002504 "src/f32-vscale/avx-x32.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07002505 "src/f32-vsqrt/gen/avx-sqrt-x8.c",
2506 "src/f32-vsqrt/gen/avx-sqrt-x16.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07002507 "src/f32-vunary/gen/vabs-avx-x8.c",
2508 "src/f32-vunary/gen/vabs-avx-x16.c",
2509 "src/f32-vunary/gen/vneg-avx-x8.c",
2510 "src/f32-vunary/gen/vneg-avx-x16.c",
2511 "src/f32-vunary/gen/vsqr-avx-x8.c",
2512 "src/f32-vunary/gen/vsqr-avx-x16.c",
Marat Dukhanb7633f22020-11-20 16:34:56 -08002513 "src/math/exp-avx-rr2-p5.c",
Marat Dukhande390d42020-11-29 19:32:18 -08002514 "src/math/expm1minus-avx-rr2-lut4-p4-perm.c",
2515 "src/math/expm1minus-avx-rr2-lut16-p3.c",
2516 "src/math/expm1minus-avx-rr2-p6.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002517 "src/math/sigmoid-avx-rr2-lut64-p2-div.c",
Marat Dukhan2e4727b2020-09-24 00:48:39 -07002518 "src/math/sigmoid-avx-rr2-p5-div.c",
Marat Dukhan39ab16f2020-10-16 14:30:35 -07002519 "src/math/sigmoid-avx-rr2-p5-nr1.c",
2520 "src/math/sigmoid-avx-rr2-p5-nr2.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002521]
2522
Marat Dukhan1566fee2020-08-02 21:55:41 -07002523XOP_UKERNELS = [
2524 "src/qs8-gemm/gen/1x4c2-minmax-xop-ld64.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002525 "src/qs8-gemm/gen/1x4c2-minmax-xop-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002526 "src/qs8-gemm/gen/1x4c2-xw-minmax-xop.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002527 "src/qs8-gemm/gen/1x4c8-minmax-xop-ld64.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002528 "src/qs8-gemm/gen/1x4c8-minmax-xop-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002529 "src/qs8-gemm/gen/1x4c8-xw-minmax-xop.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002530 "src/qs8-gemm/gen/2x4c8-minmax-xop-ld64.c",
2531 "src/qs8-gemm/gen/2x4c8-minmax-xop-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002532 "src/qs8-gemm/gen/2x4c8-xw-minmax-xop.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002533 "src/qs8-gemm/gen/3x4c8-minmax-xop-ld64.c",
2534 "src/qs8-gemm/gen/3x4c8-minmax-xop-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002535 "src/qs8-gemm/gen/3x4c8-xw-minmax-xop.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002536 "src/qs8-gemm/gen/4x4c2-minmax-xop-ld64.c",
2537 "src/qs8-gemm/gen/4x4c2-minmax-xop-ld128.c",
2538 "src/qs8-gemm/gen/4x4c2-xw-minmax-xop.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002539 "src/qs8-igemm/gen/1x4c2-minmax-xop-ld64.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002540 "src/qs8-igemm/gen/1x4c2-minmax-xop-ld128.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002541 "src/qs8-igemm/gen/1x4c8-minmax-xop-ld64.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002542 "src/qs8-igemm/gen/1x4c8-minmax-xop-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002543 "src/qs8-igemm/gen/2x4c8-minmax-xop-ld64.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002544 "src/qs8-igemm/gen/2x4c8-minmax-xop-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002545 "src/qs8-igemm/gen/3x4c8-minmax-xop-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002546 "src/qs8-igemm/gen/3x4c8-minmax-xop-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002547 "src/qs8-igemm/gen/4x4c2-minmax-xop-ld64.c",
2548 "src/qs8-igemm/gen/4x4c2-minmax-xop-ld128.c",
Marat Dukhanbb9225e2020-09-06 22:40:56 -07002549 "src/qs8-vadd/gen/minmax-xop-mul32-ld32-x8.c",
2550 "src/qs8-vadd/gen/minmax-xop-mul32-ld32-x16.c",
2551 "src/qs8-vadd/gen/minmax-xop-mul32-ld32-x24.c",
2552 "src/qs8-vadd/gen/minmax-xop-mul32-ld32-x32.c",
2553 "src/qs8-vaddc/gen/minmax-xop-mul32-ld32-x8.c",
2554 "src/qs8-vaddc/gen/minmax-xop-mul32-ld32-x16.c",
2555 "src/qs8-vaddc/gen/minmax-xop-mul32-ld32-x24.c",
2556 "src/qs8-vaddc/gen/minmax-xop-mul32-ld32-x32.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002557]
2558
Marat Dukhanfda12b82019-11-21 12:27:59 -08002559FMA3_UKERNELS = [
Marat Dukhan1c587112020-04-08 20:04:28 -07002560 "src/f32-dwconv/gen/up8x4-minmax-fma3-acc2.c",
2561 "src/f32-dwconv/gen/up8x4-minmax-fma3.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002562 "src/f32-dwconv/gen/up8x9-minmax-fma3-acc2.c",
2563 "src/f32-dwconv/gen/up8x9-minmax-fma3.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002564 "src/f32-dwconv/gen/up8x25-minmax-fma3-acc2.c",
2565 "src/f32-dwconv/gen/up8x25-minmax-fma3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002566 "src/f32-dwconv/gen/up16x4-minmax-fma3-acc2.c",
2567 "src/f32-dwconv/gen/up16x4-minmax-fma3.c",
2568 "src/f32-dwconv/gen/up16x9-minmax-fma3-acc2.c",
2569 "src/f32-dwconv/gen/up16x9-minmax-fma3.c",
2570 "src/f32-dwconv/gen/up16x25-minmax-fma3-acc2.c",
2571 "src/f32-dwconv/gen/up16x25-minmax-fma3.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002572 "src/f32-gemm/gen-inc/1x8inc-minmax-fma3-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002573 "src/f32-gemm/gen-inc/1x16inc-minmax-fma3-broadcast.c",
2574 "src/f32-gemm/gen-inc/1x16s4inc-minmax-fma3-broadcast.c",
2575 "src/f32-gemm/gen-inc/3x16inc-minmax-fma3-broadcast.c",
2576 "src/f32-gemm/gen-inc/3x16s4inc-minmax-fma3-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002577 "src/f32-gemm/gen-inc/4x8inc-minmax-fma3-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002578 "src/f32-gemm/gen-inc/4x16inc-minmax-fma3-broadcast.c",
2579 "src/f32-gemm/gen-inc/4x16s4inc-minmax-fma3-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002580 "src/f32-gemm/gen-inc/5x8inc-minmax-fma3-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002581 "src/f32-gemm/gen-inc/5x16inc-minmax-fma3-broadcast.c",
2582 "src/f32-gemm/gen-inc/5x16s4inc-minmax-fma3-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002583 "src/f32-gemm/gen-inc/6x8inc-minmax-fma3-broadcast.c",
2584 "src/f32-gemm/gen-inc/7x8inc-minmax-fma3-broadcast.c",
2585 "src/f32-gemm/gen-inc/8x8inc-minmax-fma3-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002586 "src/f32-gemm/gen/1x8-minmax-fma3-broadcast.c",
2587 "src/f32-gemm/gen/1x16-minmax-fma3-broadcast.c",
2588 "src/f32-gemm/gen/1x16s4-minmax-fma3-broadcast.c",
2589 "src/f32-gemm/gen/3x16-minmax-fma3-broadcast.c",
2590 "src/f32-gemm/gen/3x16s4-minmax-fma3-broadcast.c",
2591 "src/f32-gemm/gen/4x8-minmax-fma3-broadcast.c",
2592 "src/f32-gemm/gen/4x16-minmax-fma3-broadcast.c",
2593 "src/f32-gemm/gen/4x16s4-minmax-fma3-broadcast.c",
2594 "src/f32-gemm/gen/5x8-minmax-fma3-broadcast.c",
2595 "src/f32-gemm/gen/5x16-minmax-fma3-broadcast.c",
2596 "src/f32-gemm/gen/5x16s4-minmax-fma3-broadcast.c",
2597 "src/f32-gemm/gen/6x8-minmax-fma3-broadcast.c",
2598 "src/f32-gemm/gen/7x8-minmax-fma3-broadcast.c",
2599 "src/f32-gemm/gen/8x8-minmax-fma3-broadcast.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07002600 "src/f32-hswish/gen/hswish-fma3-x8.c",
2601 "src/f32-hswish/gen/hswish-fma3-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002602 "src/f32-igemm/gen/1x8-minmax-fma3-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002603 "src/f32-igemm/gen/1x16-minmax-fma3-broadcast.c",
2604 "src/f32-igemm/gen/1x16s4-minmax-fma3-broadcast.c",
2605 "src/f32-igemm/gen/3x16-minmax-fma3-broadcast.c",
2606 "src/f32-igemm/gen/3x16s4-minmax-fma3-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002607 "src/f32-igemm/gen/4x8-minmax-fma3-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002608 "src/f32-igemm/gen/4x16-minmax-fma3-broadcast.c",
2609 "src/f32-igemm/gen/4x16s4-minmax-fma3-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002610 "src/f32-igemm/gen/5x8-minmax-fma3-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002611 "src/f32-igemm/gen/5x16-minmax-fma3-broadcast.c",
2612 "src/f32-igemm/gen/5x16s4-minmax-fma3-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002613 "src/f32-igemm/gen/6x8-minmax-fma3-broadcast.c",
2614 "src/f32-igemm/gen/7x8-minmax-fma3-broadcast.c",
2615 "src/f32-igemm/gen/8x8-minmax-fma3-broadcast.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07002616 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x8.c",
2617 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x16.c",
2618 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x24.c",
2619 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x32.c",
2620 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x40.c",
2621 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x48.c",
2622 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x56.c",
2623 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x64.c",
Marat Dukhan84000762020-06-29 18:38:43 -07002624 "src/math/sqrt-fma3-nr1fma.c",
Marat Dukhan84000762020-06-29 18:38:43 -07002625 "src/math/sqrt-fma3-nr1fma1adj.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002626 "src/math/sqrt-fma3-nr2fma.c",
Marat Dukhanfda12b82019-11-21 12:27:59 -08002627]
2628
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002629AVX2_UKERNELS = [
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002630 "src/f32-raddexpminusmax/gen/avx2-p5-x64-acc2.c",
2631 "src/f32-raddexpminusmax/gen/avx2-p5-x64-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002632 "src/f32-raddexpminusmax/gen/avx2-p5-x64.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002633 "src/f32-raddexpminusmax/gen/avx2-p5-x72-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002634 "src/f32-raddexpminusmax/gen/avx2-p5-x72.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002635 "src/f32-raddexpminusmax/gen/avx2-p5-x80-acc2.c",
2636 "src/f32-raddexpminusmax/gen/avx2-p5-x80-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002637 "src/f32-raddexpminusmax/gen/avx2-p5-x80.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002638 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc2.c",
2639 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc3.c",
2640 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc6.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002641 "src/f32-raddexpminusmax/gen/avx2-p5-x96.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002642 "src/f32-raddextexp/gen/avx2-p5-x64-acc2.c",
2643 "src/f32-raddextexp/gen/avx2-p5-x64-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002644 "src/f32-raddextexp/gen/avx2-p5-x64.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002645 "src/f32-raddextexp/gen/avx2-p5-x72-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002646 "src/f32-raddextexp/gen/avx2-p5-x72.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002647 "src/f32-raddextexp/gen/avx2-p5-x80-acc2.c",
2648 "src/f32-raddextexp/gen/avx2-p5-x80-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002649 "src/f32-raddextexp/gen/avx2-p5-x80.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002650 "src/f32-raddextexp/gen/avx2-p5-x96-acc2.c",
2651 "src/f32-raddextexp/gen/avx2-p5-x96-acc3.c",
2652 "src/f32-raddextexp/gen/avx2-p5-x96-acc6.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002653 "src/f32-raddextexp/gen/avx2-p5-x96.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002654 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64-acc2.c",
2655 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002656 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002657 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x72-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002658 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x72.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002659 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80-acc2.c",
2660 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002661 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002662 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc2.c",
2663 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc3.c",
2664 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc6.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002665 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96.c",
Marat Dukhanfa0a4322020-01-06 16:14:29 -08002666 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x8.c",
2667 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x16.c",
2668 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x24.c",
2669 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x32.c",
2670 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x40.c",
2671 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x48.c",
2672 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x56.c",
2673 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x64.c",
2674 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x72.c",
2675 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x80.c",
2676 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x8.c",
2677 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x16.c",
2678 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x24.c",
2679 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x32.c",
2680 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x40.c",
2681 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x48.c",
2682 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x56.c",
2683 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x64.c",
2684 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x72.c",
2685 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x80.c",
2686 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x8.c",
2687 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x16.c",
2688 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x24.c",
2689 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x32.c",
2690 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x40.c",
2691 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x48.c",
2692 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x56.c",
2693 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x64.c",
2694 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x72.c",
2695 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x80.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002696 "src/f32-vscaleexpminusmax/gen/avx2-p5-x8.c",
2697 "src/f32-vscaleexpminusmax/gen/avx2-p5-x16.c",
2698 "src/f32-vscaleexpminusmax/gen/avx2-p5-x24.c",
2699 "src/f32-vscaleexpminusmax/gen/avx2-p5-x32.c",
2700 "src/f32-vscaleexpminusmax/gen/avx2-p5-x40.c",
2701 "src/f32-vscaleexpminusmax/gen/avx2-p5-x48.c",
2702 "src/f32-vscaleexpminusmax/gen/avx2-p5-x56.c",
2703 "src/f32-vscaleexpminusmax/gen/avx2-p5-x64.c",
2704 "src/f32-vscaleexpminusmax/gen/avx2-p5-x72.c",
2705 "src/f32-vscaleexpminusmax/gen/avx2-p5-x80.c",
2706 "src/f32-vscaleexpminusmax/gen/avx2-p5-x88.c",
2707 "src/f32-vscaleexpminusmax/gen/avx2-p5-x96.c",
2708 "src/f32-vscaleextexp/gen/avx2-p5-x8.c",
2709 "src/f32-vscaleextexp/gen/avx2-p5-x16.c",
2710 "src/f32-vscaleextexp/gen/avx2-p5-x24.c",
2711 "src/f32-vscaleextexp/gen/avx2-p5-x32.c",
2712 "src/f32-vscaleextexp/gen/avx2-p5-x40.c",
2713 "src/f32-vscaleextexp/gen/avx2-p5-x48.c",
2714 "src/f32-vscaleextexp/gen/avx2-p5-x56.c",
2715 "src/f32-vscaleextexp/gen/avx2-p5-x64.c",
2716 "src/f32-vscaleextexp/gen/avx2-p5-x72.c",
2717 "src/f32-vscaleextexp/gen/avx2-p5-x80.c",
2718 "src/f32-vscaleextexp/gen/avx2-p5-x88.c",
2719 "src/f32-vscaleextexp/gen/avx2-p5-x96.c",
Marat Dukhanb7633f22020-11-20 16:34:56 -08002720 "src/math/exp-avx2-rr2-lut8-p3-perm.c",
2721 "src/math/exp-avx2-rr2-lut8-p4-perm.c",
2722 "src/math/exp-avx2-rr2-p5.c",
Marat Dukhan9dd119a2020-11-20 18:20:04 -08002723 "src/math/expminus-avx2-rr2-p5.c",
Marat Dukhande390d42020-11-29 19:32:18 -08002724 "src/math/expm1minus-avx2-rr1-lut4-p4-perm.c",
2725 "src/math/expm1minus-avx2-rr1-lut8-p4-perm.c",
2726 "src/math/expm1minus-avx2-rr1-lut16-p3-gather.c",
2727 "src/math/expm1minus-avx2-rr1-p6.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002728 "src/math/extexp-avx2-p5.c",
2729 "src/math/sigmoid-avx2-rr1-lut64-p2-gather-div.c",
2730 "src/math/sigmoid-avx2-rr1-lut64-p2-gather-nr1fma.c",
2731 "src/math/sigmoid-avx2-rr1-lut64-p2-gather-nr2fma.c",
2732 "src/math/sigmoid-avx2-rr1-lut64-p2-gather-nr2fma1adj.c",
2733 "src/math/sigmoid-avx2-rr1-p5-div.c",
2734 "src/math/sigmoid-avx2-rr1-p5-nr1fma.c",
2735 "src/math/sigmoid-avx2-rr1-p5-nr2fma.c",
2736 "src/math/sigmoid-avx2-rr2-lut64-p2-gather-div.c",
2737 "src/math/sigmoid-avx2-rr2-lut64-p2-gather-nr1fma.c",
2738 "src/math/sigmoid-avx2-rr2-lut64-p2-gather-nr2fma.c",
2739 "src/math/sigmoid-avx2-rr2-lut64-p2-gather-nr2fma1adj.c",
2740 "src/math/sigmoid-avx2-rr2-p5-div.c",
2741 "src/math/sigmoid-avx2-rr2-p5-nr1fma.c",
2742 "src/math/sigmoid-avx2-rr2-p5-nr2fma.c",
Marat Dukhand65a1522020-08-04 19:28:18 -07002743 "src/qs8-dwconv/gen/up8x9-minmax-avx2-mul32.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002744 "src/qs8-dwconv/gen/up16x9-minmax-avx2-mul16.c",
Marat Dukhand65a1522020-08-04 19:28:18 -07002745 "src/qs8-dwconv/gen/up16x9-minmax-avx2-mul32.c",
2746 "src/qs8-dwconv/gen/up24x9-minmax-avx2-mul32.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002747 "src/qs8-dwconv/gen/up32x9-minmax-avx2-mul16.c",
Marat Dukhand65a1522020-08-04 19:28:18 -07002748 "src/qs8-dwconv/gen/up32x9-minmax-avx2-mul32.c",
Marat Dukhan12809522020-08-02 22:23:51 -07002749 "src/qs8-gemm/gen/1x8c8-minmax-avx2.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002750 "src/qs8-gemm/gen/1x8c8-xw-minmax-avx2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002751 "src/qs8-gemm/gen/2x8c8-minmax-avx2.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002752 "src/qs8-gemm/gen/2x8c8-xw-minmax-avx2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002753 "src/qs8-gemm/gen/3x8c8-minmax-avx2.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002754 "src/qs8-gemm/gen/3x8c8-xw-minmax-avx2.c",
Marat Dukhan12809522020-08-02 22:23:51 -07002755 "src/qs8-igemm/gen/1x8c8-minmax-avx2.c",
2756 "src/qs8-igemm/gen/2x8c8-minmax-avx2.c",
2757 "src/qs8-igemm/gen/3x8c8-minmax-avx2.c",
Marat Dukhane6dc0b62020-09-08 23:57:14 -07002758 "src/qs8-vadd/gen/minmax-avx2-mul32-ld64-x8.c",
2759 "src/qs8-vadd/gen/minmax-avx2-mul32-ld64-x16.c",
2760 "src/qs8-vadd/gen/minmax-avx2-mul32-ld64-x24.c",
2761 "src/qs8-vadd/gen/minmax-avx2-mul32-ld64-x32.c",
2762 "src/qs8-vaddc/gen/minmax-avx2-mul32-ld64-x8.c",
2763 "src/qs8-vaddc/gen/minmax-avx2-mul32-ld64-x16.c",
2764 "src/qs8-vaddc/gen/minmax-avx2-mul32-ld64-x24.c",
2765 "src/qs8-vaddc/gen/minmax-avx2-mul32-ld64-x32.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002766]
2767
Marat Dukhan08c4a432019-10-03 09:29:21 -07002768AVX512F_UKERNELS = [
Marat Dukhan5c5fa962020-03-10 18:38:33 -07002769 "src/f32-clamp/gen/avx512f-x16.c",
2770 "src/f32-clamp/gen/avx512f-x32.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002771 "src/f32-dwconv/gen/up16x4-minmax-avx512f-acc2.c",
2772 "src/f32-dwconv/gen/up16x4-minmax-avx512f.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002773 "src/f32-dwconv/gen/up16x9-minmax-avx512f-acc2.c",
2774 "src/f32-dwconv/gen/up16x9-minmax-avx512f.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002775 "src/f32-dwconv/gen/up16x25-minmax-avx512f-acc2.c",
2776 "src/f32-dwconv/gen/up16x25-minmax-avx512f.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002777 "src/f32-dwconv/gen/up32x4-minmax-avx512f-acc2.c",
2778 "src/f32-dwconv/gen/up32x4-minmax-avx512f.c",
2779 "src/f32-dwconv/gen/up32x9-minmax-avx512f-acc2.c",
2780 "src/f32-dwconv/gen/up32x9-minmax-avx512f.c",
2781 "src/f32-dwconv/gen/up32x25-minmax-avx512f-acc2.c",
2782 "src/f32-dwconv/gen/up32x25-minmax-avx512f.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002783 "src/f32-gemm/gen-inc/1x16inc-minmax-avx512f-broadcast.c",
2784 "src/f32-gemm/gen-inc/4x16inc-minmax-avx512f-broadcast.c",
2785 "src/f32-gemm/gen-inc/5x16inc-minmax-avx512f-broadcast.c",
2786 "src/f32-gemm/gen-inc/6x16inc-minmax-avx512f-broadcast.c",
2787 "src/f32-gemm/gen-inc/7x16inc-minmax-avx512f-broadcast.c",
2788 "src/f32-gemm/gen-inc/8x16inc-minmax-avx512f-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002789 "src/f32-gemm/gen/1x16-minmax-avx512f-broadcast.c",
2790 "src/f32-gemm/gen/4x16-minmax-avx512f-broadcast.c",
2791 "src/f32-gemm/gen/5x16-minmax-avx512f-broadcast.c",
2792 "src/f32-gemm/gen/6x16-minmax-avx512f-broadcast.c",
2793 "src/f32-gemm/gen/7x16-minmax-avx512f-broadcast.c",
2794 "src/f32-gemm/gen/8x16-minmax-avx512f-broadcast.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07002795 "src/f32-hswish/gen/hswish-avx512f-x16.c",
2796 "src/f32-hswish/gen/hswish-avx512f-x32.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002797 "src/f32-igemm/gen/1x16-minmax-avx512f-broadcast.c",
2798 "src/f32-igemm/gen/4x16-minmax-avx512f-broadcast.c",
2799 "src/f32-igemm/gen/5x16-minmax-avx512f-broadcast.c",
2800 "src/f32-igemm/gen/6x16-minmax-avx512f-broadcast.c",
2801 "src/f32-igemm/gen/7x16-minmax-avx512f-broadcast.c",
2802 "src/f32-igemm/gen/8x16-minmax-avx512f-broadcast.c",
Marat Dukhan90eca0a2020-03-11 00:52:23 -07002803 "src/f32-prelu/gen/avx512f-2x16.c",
2804 "src/f32-prelu/gen/avx512f-2x32.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002805 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128-acc2.c",
2806 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002807 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002808 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x144-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002809 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x144.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002810 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160-acc2.c",
2811 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002812 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002813 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc2.c",
2814 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc3.c",
2815 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc6.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002816 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002817 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128-acc2.c",
2818 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002819 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002820 "src/f32-raddextexp/gen/avx512f-p5-scalef-x144-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002821 "src/f32-raddextexp/gen/avx512f-p5-scalef-x144.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002822 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160-acc2.c",
2823 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002824 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002825 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc2.c",
2826 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc3.c",
2827 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc6.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002828 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002829 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128-acc2.c",
2830 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002831 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002832 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x144-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002833 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x144.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002834 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160-acc2.c",
2835 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002836 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002837 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc2.c",
2838 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc3.c",
2839 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc6.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002840 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07002841 "src/f32-relu/gen/avx512f-x16.c",
2842 "src/f32-relu/gen/avx512f-x32.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002843 "src/f32-rmax/avx512f.c",
Marat Dukhand9ca7e62020-09-23 23:45:29 -07002844 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x16.c",
2845 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x32.c",
2846 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x48.c",
2847 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x64.c",
2848 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x80.c",
2849 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x96.c",
2850 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x112.c",
2851 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x128.c",
2852 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x16.c",
2853 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x32.c",
2854 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x48.c",
2855 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x64.c",
2856 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x80.c",
2857 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x96.c",
2858 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x112.c",
2859 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002860 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x16.c",
2861 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x32.c",
2862 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x48.c",
2863 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x64.c",
2864 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x80.c",
2865 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x96.c",
2866 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x112.c",
2867 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x128.c",
2868 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x16.c",
2869 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x32.c",
2870 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x48.c",
2871 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x64.c",
2872 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x80.c",
2873 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x96.c",
2874 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x112.c",
2875 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x128.c",
Marat Dukhand9ca7e62020-09-23 23:45:29 -07002876 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x16.c",
2877 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x32.c",
2878 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x48.c",
2879 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x64.c",
2880 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x80.c",
2881 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x96.c",
2882 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x112.c",
2883 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x128.c",
2884 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x16.c",
2885 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x32.c",
2886 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x48.c",
2887 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x64.c",
2888 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x80.c",
2889 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x96.c",
2890 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x112.c",
2891 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x128.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002892 "src/f32-vbinary/gen/vadd-minmax-avx512f-x16.c",
2893 "src/f32-vbinary/gen/vadd-minmax-avx512f-x32.c",
2894 "src/f32-vbinary/gen/vaddc-minmax-avx512f-x16.c",
2895 "src/f32-vbinary/gen/vaddc-minmax-avx512f-x32.c",
2896 "src/f32-vbinary/gen/vdiv-minmax-avx512f-x16.c",
2897 "src/f32-vbinary/gen/vdiv-minmax-avx512f-x32.c",
2898 "src/f32-vbinary/gen/vdivc-minmax-avx512f-x16.c",
2899 "src/f32-vbinary/gen/vdivc-minmax-avx512f-x32.c",
Marat Dukhan9a88efe2019-12-10 15:54:24 -08002900 "src/f32-vbinary/gen/vmax-avx512f-x16.c",
2901 "src/f32-vbinary/gen/vmax-avx512f-x32.c",
2902 "src/f32-vbinary/gen/vmaxc-avx512f-x16.c",
2903 "src/f32-vbinary/gen/vmaxc-avx512f-x32.c",
2904 "src/f32-vbinary/gen/vmin-avx512f-x16.c",
2905 "src/f32-vbinary/gen/vmin-avx512f-x32.c",
2906 "src/f32-vbinary/gen/vminc-avx512f-x16.c",
2907 "src/f32-vbinary/gen/vminc-avx512f-x32.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002908 "src/f32-vbinary/gen/vmul-minmax-avx512f-x16.c",
2909 "src/f32-vbinary/gen/vmul-minmax-avx512f-x32.c",
2910 "src/f32-vbinary/gen/vmulc-minmax-avx512f-x16.c",
2911 "src/f32-vbinary/gen/vmulc-minmax-avx512f-x32.c",
2912 "src/f32-vbinary/gen/vrdivc-minmax-avx512f-x16.c",
2913 "src/f32-vbinary/gen/vrdivc-minmax-avx512f-x32.c",
2914 "src/f32-vbinary/gen/vrsubc-minmax-avx512f-x16.c",
2915 "src/f32-vbinary/gen/vrsubc-minmax-avx512f-x32.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07002916 "src/f32-vbinary/gen/vsqrdiff-avx512f-x16.c",
2917 "src/f32-vbinary/gen/vsqrdiff-avx512f-x32.c",
2918 "src/f32-vbinary/gen/vsqrdiffc-avx512f-x16.c",
2919 "src/f32-vbinary/gen/vsqrdiffc-avx512f-x32.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002920 "src/f32-vbinary/gen/vsub-minmax-avx512f-x16.c",
2921 "src/f32-vbinary/gen/vsub-minmax-avx512f-x32.c",
2922 "src/f32-vbinary/gen/vsubc-minmax-avx512f-x16.c",
2923 "src/f32-vbinary/gen/vsubc-minmax-avx512f-x32.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07002924 "src/f32-vlrelu/gen/vlrelu-avx512f-x16.c",
2925 "src/f32-vlrelu/gen/vlrelu-avx512f-x32.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002926 "src/f32-vrnd/gen/vrndd-avx512f-x16.c",
2927 "src/f32-vrnd/gen/vrndd-avx512f-x32.c",
2928 "src/f32-vrnd/gen/vrndne-avx512f-x16.c",
2929 "src/f32-vrnd/gen/vrndne-avx512f-x32.c",
2930 "src/f32-vrnd/gen/vrndu-avx512f-x16.c",
2931 "src/f32-vrnd/gen/vrndu-avx512f-x32.c",
2932 "src/f32-vrnd/gen/vrndz-avx512f-x16.c",
2933 "src/f32-vrnd/gen/vrndz-avx512f-x32.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07002934 "src/f32-vscale/avx512f-x64.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002935 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x16.c",
2936 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x32.c",
2937 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x48.c",
2938 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x64.c",
2939 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x80.c",
2940 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x96.c",
2941 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x112.c",
2942 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x128.c",
2943 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x144.c",
2944 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x160.c",
2945 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x176.c",
2946 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x192.c",
2947 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x16.c",
2948 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x32.c",
2949 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x48.c",
2950 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x64.c",
2951 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x80.c",
2952 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x96.c",
2953 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x112.c",
2954 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x128.c",
2955 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x144.c",
2956 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x160.c",
2957 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x176.c",
2958 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x192.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07002959 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x16.c",
2960 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x32.c",
2961 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x48.c",
2962 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x64.c",
2963 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x80.c",
2964 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x96.c",
2965 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x112.c",
2966 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x128.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07002967 "src/f32-vunary/gen/vabs-avx512f-x16.c",
2968 "src/f32-vunary/gen/vabs-avx512f-x32.c",
2969 "src/f32-vunary/gen/vneg-avx512f-x16.c",
2970 "src/f32-vunary/gen/vneg-avx512f-x32.c",
2971 "src/f32-vunary/gen/vsqr-avx512f-x16.c",
2972 "src/f32-vunary/gen/vsqr-avx512f-x32.c",
Marat Dukhanb7633f22020-11-20 16:34:56 -08002973 "src/math/exp-avx512f-rr2-lut16-p3-perm-scalef.c",
2974 "src/math/exp-avx512f-rr2-lut16-p3-perm.c",
2975 "src/math/exp-avx512f-rr2-lut32-p2-perm2-scalef.c",
2976 "src/math/exp-avx512f-rr2-lut32-p2-perm2.c",
2977 "src/math/exp-avx512f-rr2-p5-scalef.c",
2978 "src/math/exp-avx512f-rr2-p5.c",
Marat Dukhande390d42020-11-29 19:32:18 -08002979 "src/math/expm1minus-avx512f-rr1-lut16-p3-perm.c",
2980 "src/math/expm1minus-avx512f-rr1-p6.c",
Marat Dukhan98ba4412019-10-23 02:14:28 -07002981 "src/math/extexp-avx512f-p5.c",
Marat Dukhan6a34c5f2020-09-22 21:44:15 -07002982 "src/math/sigmoid-avx512f-rr1-lut16-p3-perm-scalef-div.c",
Marat Dukhan6a34c5f2020-09-22 21:44:15 -07002983 "src/math/sigmoid-avx512f-rr1-lut16-p3-perm-scalef-nr1fma.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002984 "src/math/sigmoid-avx512f-rr1-lut16-p3-perm-scalef-nr1fma1adj.c",
Marat Dukhan6a34c5f2020-09-22 21:44:15 -07002985 "src/math/sigmoid-avx512f-rr1-lut32-p2-perm2-scalef-div.c",
Marat Dukhan6a34c5f2020-09-22 21:44:15 -07002986 "src/math/sigmoid-avx512f-rr1-lut32-p2-perm2-scalef-nr1fma.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002987 "src/math/sigmoid-avx512f-rr1-lut32-p2-perm2-scalef-nr1fma1adj.c",
Marat Dukhan36173d22020-10-15 17:14:26 -07002988 "src/math/sigmoid-avx512f-rr1-lut64-p2-gather-scalef-div.c",
Marat Dukhan36173d22020-10-15 17:14:26 -07002989 "src/math/sigmoid-avx512f-rr1-lut64-p2-gather-scalef-nr1fma.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002990 "src/math/sigmoid-avx512f-rr1-lut64-p2-gather-scalef-nr1fma1adj.c",
2991 "src/math/sigmoid-avx512f-rr1-p5-scalef-div.c",
2992 "src/math/sigmoid-avx512f-rr1-p5-scalef-nr1fma.c",
2993 "src/math/sigmoid-avx512f-rr1-p5-scalef-nr1fma1adj.c",
2994 "src/math/sigmoid-avx512f-rr2-lut16-p3-perm-scalef-div.c",
2995 "src/math/sigmoid-avx512f-rr2-lut16-p3-perm-scalef-nr1fma.c",
2996 "src/math/sigmoid-avx512f-rr2-lut16-p3-perm-scalef-nr1fma1adj.c",
2997 "src/math/sigmoid-avx512f-rr2-lut32-p2-perm2-scalef-div.c",
2998 "src/math/sigmoid-avx512f-rr2-lut32-p2-perm2-scalef-nr1fma.c",
2999 "src/math/sigmoid-avx512f-rr2-lut32-p2-perm2-scalef-nr1fma1adj.c",
Marat Dukhan36173d22020-10-15 17:14:26 -07003000 "src/math/sigmoid-avx512f-rr2-lut64-p2-gather-scalef-div.c",
Marat Dukhan36173d22020-10-15 17:14:26 -07003001 "src/math/sigmoid-avx512f-rr2-lut64-p2-gather-scalef-nr1fma.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003002 "src/math/sigmoid-avx512f-rr2-lut64-p2-gather-scalef-nr1fma1adj.c",
3003 "src/math/sigmoid-avx512f-rr2-p5-scalef-div.c",
3004 "src/math/sigmoid-avx512f-rr2-p5-scalef-nr1fma.c",
3005 "src/math/sigmoid-avx512f-rr2-p5-scalef-nr1fma1adj.c",
Marat Dukhan84000762020-06-29 18:38:43 -07003006 "src/math/sqrt-avx512f-nr1fma.c",
Marat Dukhan84000762020-06-29 18:38:43 -07003007 "src/math/sqrt-avx512f-nr1fma1adj.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003008 "src/math/sqrt-avx512f-nr2fma.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003009]
3010
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07003011AVX512SKX_UKERNELS = [
Marat Dukhan2ffc5e62020-09-06 22:33:38 -07003012 "src/qs8-dwconv/gen/up16x9-minmax-avx512skx-mul32.c",
3013 "src/qs8-dwconv/gen/up32x9-minmax-avx512skx-mul32.c",
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07003014 "src/qs8-gemm/gen/1x16c8-minmax-avx512skx.c",
3015 "src/qs8-gemm/gen/2x16c8-minmax-avx512skx.c",
3016 "src/qs8-gemm/gen/3x16c8-minmax-avx512skx.c",
3017 "src/qs8-gemm/gen/4x16c8-minmax-avx512skx.c",
3018 "src/qs8-igemm/gen/1x16c8-minmax-avx512skx.c",
3019 "src/qs8-igemm/gen/2x16c8-minmax-avx512skx.c",
3020 "src/qs8-igemm/gen/3x16c8-minmax-avx512skx.c",
3021 "src/qs8-igemm/gen/4x16c8-minmax-avx512skx.c",
3022]
3023
Frank Barchardbcedc082020-08-17 18:00:51 -07003024WASM32_ASM_UKERNELS = [
3025 "src/f32-relu/wasm_shr_x1.S",
3026 "src/f32-relu/wasm_shr_x2.S",
3027 "src/f32-relu/wasm_shr_x4.S",
3028]
3029
Marat Dukhan08c4a432019-10-03 09:29:21 -07003030AARCH32_ASM_UKERNELS = [
Marat Dukhan32f93812020-05-17 20:31:21 -07003031 "src/f32-gemm/4x4-aarch32-vfp-ld64.S",
Marat Dukhan3b98f6b2020-05-17 10:09:22 -07003032 "src/f32-gemm/4x4-minmax-aarch32-vfp-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003033 "src/f32-gemm/4x8-minmax-aarch32-neon-cortex-a53.S",
3034 "src/f32-gemm/4x8-minmax-aarch32-neon-cortex-a55.S",
Frank Barchard490febe2020-07-16 18:42:17 -07003035 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-cortex-a7.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003036 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-cortex-a75.S",
Frank Barchard569561d2020-06-17 13:11:12 -07003037 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-ld64.S",
Frank Barchard490febe2020-07-16 18:42:17 -07003038 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-pld-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003039 "src/f32-igemm/4x8-minmax-aarch32-neon-cortex-a53.S",
3040 "src/f32-igemm/4x8-minmax-aarch32-neon-cortex-a55.S",
Frank Barchard490febe2020-07-16 18:42:17 -07003041 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-cortex-a7.S",
3042 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-cortex-a75.S",
3043 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-ld64.S",
3044 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-pld-cortex-a75.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003045]
3046
3047AARCH64_ASM_UKERNELS = [
Frank Barchardbddfbcd2020-04-15 12:32:41 -07003048 "src/f16-gemm/gen-inc/1x8inc-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchard04336c12020-10-22 16:48:55 -07003049 "src/f16-gemm/gen-inc/1x16inc-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchardbddfbcd2020-04-15 12:32:41 -07003050 "src/f16-gemm/gen-inc/4x8inc-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchard04336c12020-10-22 16:48:55 -07003051 "src/f16-gemm/gen-inc/4x16inc-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchardbddfbcd2020-04-15 12:32:41 -07003052 "src/f16-gemm/gen-inc/6x8inc-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchard04336c12020-10-22 16:48:55 -07003053 "src/f16-gemm/gen-inc/6x16inc-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchard3b8e5662020-04-20 12:12:53 -07003054 "src/f16-gemm/gen-inc/8x8inc-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchard04336c12020-10-22 16:48:55 -07003055 "src/f16-gemm/gen/1x8-minmax-aarch64-neonfp16arith-ld64.S",
3056 "src/f16-gemm/gen/1x16-minmax-aarch64-neonfp16arith-ld32.S",
3057 "src/f16-gemm/gen/4x8-minmax-aarch64-neonfp16arith-ld64.S",
3058 "src/f16-gemm/gen/4x16-minmax-aarch64-neonfp16arith-ld32.S",
3059 "src/f16-gemm/gen/6x8-minmax-aarch64-neonfp16arith-ld64.S",
3060 "src/f16-gemm/gen/6x16-minmax-aarch64-neonfp16arith-ld32.S",
3061 "src/f16-gemm/gen/8x8-minmax-aarch64-neonfp16arith-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003062 "src/f32-dwconv/up4x9-minmax-aarch64-neonfma-cortex-a55.S",
3063 "src/f32-dwconv/up4x9-minmax-aarch64-neonfma.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003064 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a53.S",
3065 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a57.S",
3066 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a75.S",
Frank Barchard04336c12020-10-22 16:48:55 -07003067 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-ld64.S",
3068 "src/f32-gemm/gen-inc/1x12inc-minmax-aarch64-neonfma-cortex-a53.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003069 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a53.S",
3070 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a55.S",
3071 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a57.S",
3072 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003073 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-ld64.S",
Frank Barchard04336c12020-10-22 16:48:55 -07003074 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-ld128.S",
3075 "src/f32-gemm/gen-inc/4x12inc-minmax-aarch64-neonfma-cortex-a53.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003076 "src/f32-gemm/gen-inc/5x8inc-minmax-aarch64-neonfma-cortex-a57.S",
3077 "src/f32-gemm/gen-inc/5x8inc-minmax-aarch64-neonfma-cortex-a75.S",
3078 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a53.S",
3079 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a55.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003080 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a57.S",
Frank Barchard04336c12020-10-22 16:48:55 -07003081 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a73.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003082 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003083 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-ld64.S",
Frank Barchard04336c12020-10-22 16:48:55 -07003084 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-ld128.S",
3085 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a53.S",
3086 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a57.S",
3087 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a75.S",
3088 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-ld64.S",
3089 "src/f32-gemm/gen/1x12-minmax-aarch64-neonfma-cortex-a53.S",
3090 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a53.S",
3091 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a55.S",
3092 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a57.S",
3093 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a75.S",
3094 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-ld64.S",
3095 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-ld128.S",
3096 "src/f32-gemm/gen/4x12-minmax-aarch64-neonfma-cortex-a53.S",
3097 "src/f32-gemm/gen/5x8-minmax-aarch64-neonfma-cortex-a57.S",
3098 "src/f32-gemm/gen/5x8-minmax-aarch64-neonfma-cortex-a75.S",
3099 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a53.S",
3100 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a55.S",
3101 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a57.S",
3102 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a73.S",
3103 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a75.S",
3104 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-ld64.S",
3105 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-ld128.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003106 "src/f32-igemm/1x8-minmax-aarch64-neonfma-cortex-a53.S",
Frank Barchard04336c12020-10-22 16:48:55 -07003107 "src/f32-igemm/1x12-minmax-aarch64-neonfma-cortex-a53.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003108 "src/f32-igemm/4x8-minmax-aarch64-neonfma-cortex-a53.S",
3109 "src/f32-igemm/4x8-minmax-aarch64-neonfma-cortex-a55.S",
Frank Barchard04336c12020-10-22 16:48:55 -07003110 "src/f32-igemm/4x12-minmax-aarch64-neonfma-cortex-a53.S",
3111 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a53.S",
3112 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a55.S",
3113 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a73.S",
3114 "src/f32-igemm/gen/1x8-minmax-aarch64-neonfma-cortex-a57.S",
3115 "src/f32-igemm/gen/1x8-minmax-aarch64-neonfma-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003116 "src/f32-igemm/gen/4x8-minmax-aarch64-neonfma-cortex-a57.S",
3117 "src/f32-igemm/gen/4x8-minmax-aarch64-neonfma-cortex-a75.S",
3118 "src/f32-igemm/gen/5x8-minmax-aarch64-neonfma-cortex-a57.S",
3119 "src/f32-igemm/gen/5x8-minmax-aarch64-neonfma-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003120 "src/f32-igemm/gen/6x8-minmax-aarch64-neonfma-cortex-a57.S",
3121 "src/f32-igemm/gen/6x8-minmax-aarch64-neonfma-cortex-a75.S",
Frank Barchardf1fd89e2020-10-06 23:27:11 -07003122 "src/qs8-gemm/1x16c4-aarch64-neondot-ld64.S",
Frank Barchardf2742c42020-10-12 21:16:46 -07003123 "src/qs8-gemm/4x16c4-aarch64-neondot-cortex-a55.S",
Frank Barcharda5237a52020-10-06 16:18:46 -07003124 "src/qs8-gemm/4x16c4-aarch64-neondot-ld64.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003125]
3126
Marat Dukhan1b354632020-03-23 12:50:22 -07003127INTERNAL_MICROKERNEL_HDRS = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003128 "src/xnnpack/argmaxpool.h",
3129 "src/xnnpack/avgpool.h",
3130 "src/xnnpack/clamp.h",
3131 "src/xnnpack/common.h",
3132 "src/xnnpack/conv.h",
Yury Kartynnike7841862020-11-04 18:22:18 -08003133 "src/xnnpack/depthtospace.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003134 "src/xnnpack/dwconv.h",
Frank Barchard04336c12020-10-22 16:48:55 -07003135 "src/xnnpack/fill.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003136 "src/xnnpack/gavgpool.h",
3137 "src/xnnpack/gemm.h",
3138 "src/xnnpack/hswish.h",
Marat Dukhan660fd192020-03-10 04:55:30 -07003139 "src/xnnpack/ibilinear.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003140 "src/xnnpack/igemm.h",
Marat Dukhancfb31342019-12-05 10:42:57 -08003141 "src/xnnpack/intrinsics-polyfill.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003142 "src/xnnpack/lut.h",
3143 "src/xnnpack/math.h",
3144 "src/xnnpack/maxpool.h",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003145 "src/xnnpack/memory.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003146 "src/xnnpack/packx.h",
3147 "src/xnnpack/pad.h",
3148 "src/xnnpack/params.h",
3149 "src/xnnpack/pavgpool.h",
3150 "src/xnnpack/ppmm.h",
3151 "src/xnnpack/prelu.h",
Marat Dukhan97579532019-10-18 16:40:39 -07003152 "src/xnnpack/raddexpminusmax.h",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07003153 "src/xnnpack/raddextexp.h",
Marat Dukhan97579532019-10-18 16:40:39 -07003154 "src/xnnpack/raddstoreexpminusmax.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003155 "src/xnnpack/rmax.h",
3156 "src/xnnpack/scalar-utils.h",
3157 "src/xnnpack/spmm.h",
3158 "src/xnnpack/unpool.h",
3159 "src/xnnpack/vadd.h",
Marat Dukhan1e782c42019-11-21 17:02:40 -08003160 "src/xnnpack/vbinary.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003161 "src/xnnpack/vmulcaddc.h",
Marat Dukhan05ac8e32019-10-21 15:39:33 -07003162 "src/xnnpack/vscale.h",
Marat Dukhan97579532019-10-18 16:40:39 -07003163 "src/xnnpack/vscaleexpminusmax.h",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07003164 "src/xnnpack/vscaleextexp.h",
Marat Dukhan1e782c42019-11-21 17:02:40 -08003165 "src/xnnpack/vunary.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003166 "src/xnnpack/zip.h",
Marat Dukhan1b354632020-03-23 12:50:22 -07003167]
3168
3169INTERNAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003170 "include/xnnpack.h",
3171 "src/xnnpack/allocator.h",
3172 "src/xnnpack/compute.h",
3173 "src/xnnpack/im2col.h",
3174 "src/xnnpack/indirection.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003175 "src/xnnpack/math-stubs.h",
Chao Mei6ddfc602020-05-13 22:29:36 -07003176 "src/xnnpack/memory-planner.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003177 "src/xnnpack/operator.h",
3178 "src/xnnpack/pack.h",
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07003179 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003180 "src/xnnpack/requantization-stubs.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003181 "src/xnnpack/requantization.h",
Marat Dukhan1d75a542020-02-03 12:23:01 -08003182 "src/xnnpack/subgraph.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003183]
3184
Marat Dukhan1b354632020-03-23 12:50:22 -07003185ACCURACY_EVAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003186 "src/xnnpack/math-stubs.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003187]
3188
Marat Dukhan1b354632020-03-23 12:50:22 -07003189MICROKERNEL_BENCHMARK_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003190 "include/xnnpack.h",
Frank Barchard35db7d02020-10-26 13:37:34 -07003191 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003192]
3193
Marat Dukhan1b354632020-03-23 12:50:22 -07003194MICROKERNEL_TEST_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Frank Barchard35db7d02020-10-26 13:37:34 -07003195 "include/xnnpack.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003196 "src/xnnpack/isa-checks.h",
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07003197 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003198 "src/xnnpack/requantization.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003199]
3200
3201OPERATOR_TEST_PARAMS_HDRS = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003202 "src/xnnpack/common.h",
Frank Barchard04336c12020-10-22 16:48:55 -07003203 "src/xnnpack/params.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003204]
3205
3206WEIGHTS_PACK_HDRS = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003207 "src/xnnpack/compute.h",
Frank Barchard04336c12020-10-22 16:48:55 -07003208 "src/xnnpack/operator.h",
3209 "src/xnnpack/pack.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003210]
3211
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07003212LOGGING_COPTS = select({
3213 # No logging in optimized mode
3214 ":optimized_build": ["-DXNN_LOG_LEVEL=0"],
3215 # Full logging in debug mode
3216 ":debug_build": ["-DXNN_LOG_LEVEL=5"],
3217 # Error-only logging in default (fastbuild) mode
3218 "//conditions:default": ["-DXNN_LOG_LEVEL=2"],
3219})
3220
Marat Dukhan3b59de22020-06-03 20:15:19 -07003221LOGGING_SRCS = select({
3222 # No logging in optimized mode
3223 ":optimized_build": [],
3224 "//conditions:default": [
3225 "src/operator-strings.c",
3226 "src/subgraph-strings.c",
3227 ],
3228})
3229
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07003230LOGGING_HDRS = [
3231 "src/xnnpack/log.h",
3232]
3233
Marat Dukhan08c4a432019-10-03 09:29:21 -07003234xnnpack_cc_library(
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003235 name = "tables",
3236 srcs = TABLE_SRCS,
3237 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003238 gcc_copts = xnnpack_gcc_std_copts(),
3239 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003240)
3241
3242xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003243 name = "scalar_ukernels",
3244 srcs = SCALAR_UKERNELS,
3245 hdrs = INTERNAL_HDRS,
3246 aarch32_copts = ["-marm"],
Marat Dukhan10a38082020-04-17 03:58:35 -07003247 gcc_copts = xnnpack_gcc_std_copts(),
3248 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003249 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003250 ":tables",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003251 "@FP16",
3252 "@FXdiv",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003253 "@pthreadpool",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003254 ],
3255)
3256
3257xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003258 name = "scalar_ukernels_test_mode",
3259 srcs = SCALAR_UKERNELS,
3260 hdrs = INTERNAL_HDRS,
3261 aarch32_copts = ["-marm"],
3262 copts = [
3263 "-UNDEBUG",
3264 "-DXNN_TEST_MODE=1",
3265 ],
3266 gcc_copts = xnnpack_gcc_std_copts(),
3267 msvc_copts = xnnpack_msvc_std_copts(),
3268 deps = [
3269 ":tables",
3270 "@FP16",
3271 "@FXdiv",
3272 "@pthreadpool",
3273 ],
3274)
3275
3276xnnpack_cc_library(
Marat Dukhan436ebe62019-12-04 15:10:12 -08003277 name = "wasm_ukernels",
3278 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003279 gcc_copts = xnnpack_gcc_std_copts(),
3280 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan436ebe62019-12-04 15:10:12 -08003281 wasm_srcs = WASM_UKERNELS,
Marat Dukhan1c6e3892020-06-25 23:56:10 -07003282 wasmsimd_srcs = WASM_UKERNELS + WASMSIMD_UKERNELS,
Marat Dukhan436ebe62019-12-04 15:10:12 -08003283 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003284 ":tables",
Marat Dukhan436ebe62019-12-04 15:10:12 -08003285 "@FP16",
3286 "@FXdiv",
3287 "@pthreadpool",
3288 ],
3289)
3290
3291xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003292 name = "wasm_ukernels_test_mode",
3293 hdrs = INTERNAL_HDRS,
3294 copts = [
3295 "-UNDEBUG",
3296 "-DXNN_TEST_MODE=1",
3297 ],
3298 gcc_copts = xnnpack_gcc_std_copts(),
3299 msvc_copts = xnnpack_msvc_std_copts(),
3300 wasm_srcs = WASM_UKERNELS,
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07003301 wasmsimd_srcs = WASM_UKERNELS + WASMSIMD_UKERNELS,
Marat Dukhan33fcf782020-05-24 14:27:15 -07003302 deps = [
3303 ":tables",
3304 "@FP16",
3305 "@FXdiv",
3306 "@pthreadpool",
3307 ],
3308)
3309
3310xnnpack_cc_library(
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08003311 name = "psimd_accmath_ukernels",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08003312 hdrs = INTERNAL_HDRS,
3313 aarch32_copts = [
3314 "-marm",
3315 "-mfpu=neon",
3316 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07003317 gcc_copts = xnnpack_gcc_std_copts(),
3318 gcc_x86_copts = ["-msse2"],
3319 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08003320 optimized_copts = [
3321 "-O3",
3322 ],
Marat Dukhan500b8892020-04-15 17:09:50 -07003323 psimd_srcs = PSIMD_ACCMATH_UKERNELS,
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08003324 deps = [
3325 ":tables",
3326 "@FP16",
3327 "@psimd",
3328 "@pthreadpool",
3329 ],
3330)
3331
3332xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003333 name = "psimd_accmath_ukernels_test_mode",
3334 hdrs = INTERNAL_HDRS,
3335 aarch32_copts = [
3336 "-marm",
3337 "-mfpu=neon",
3338 ],
3339 copts = [
3340 "-UNDEBUG",
3341 "-DXNN_TEST_MODE=1",
3342 ],
3343 gcc_copts = xnnpack_gcc_std_copts(),
3344 gcc_x86_copts = ["-msse2"],
3345 msvc_copts = xnnpack_msvc_std_copts(),
3346 optimized_copts = [
3347 "-O3",
3348 ],
3349 psimd_srcs = PSIMD_ACCMATH_UKERNELS,
3350 deps = [
3351 ":tables",
3352 "@FP16",
3353 "@psimd",
3354 "@pthreadpool",
3355 ],
3356)
3357
3358xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003359 name = "neon_ukernels",
3360 hdrs = INTERNAL_HDRS,
3361 aarch32_copts = [
3362 "-marm",
Marat Dukhan8853b822020-05-07 12:19:01 -07003363 "-march=armv7-a",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003364 "-mfpu=neon",
3365 ],
3366 aarch32_srcs = NEON_UKERNELS,
3367 aarch64_srcs = NEON_UKERNELS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003368 gcc_copts = xnnpack_gcc_std_copts(),
3369 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08003370 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003371 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003372 "@FP16",
3373 "@pthreadpool",
3374 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003375)
3376
3377xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003378 name = "neon_ukernels_test_mode",
3379 hdrs = INTERNAL_HDRS,
3380 aarch32_copts = [
3381 "-marm",
3382 "-march=armv7-a",
3383 "-mfpu=neon",
3384 ],
3385 aarch32_srcs = NEON_UKERNELS,
3386 aarch64_srcs = NEON_UKERNELS,
3387 copts = [
3388 "-UNDEBUG",
3389 "-DXNN_TEST_MODE=1",
3390 ],
3391 gcc_copts = xnnpack_gcc_std_copts(),
3392 msvc_copts = xnnpack_msvc_std_copts(),
3393 deps = [
3394 ":tables",
3395 "@FP16",
3396 "@pthreadpool",
3397 ],
3398)
3399
3400xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003401 name = "neonfma_ukernels",
3402 hdrs = INTERNAL_HDRS,
3403 aarch32_copts = [
3404 "-marm",
Marat Dukhan8853b822020-05-07 12:19:01 -07003405 "-march=armv7-a",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003406 "-mfpu=neon-vfpv4",
3407 ],
3408 aarch32_srcs = NEONFMA_UKERNELS,
3409 aarch64_srcs = NEONFMA_UKERNELS + AARCH64_NEONFMA_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07003410 apple_aarch32_copts = [
3411 "-mcpu=swift",
3412 "-mtune=generic",
3413 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07003414 gcc_copts = xnnpack_gcc_std_copts(),
3415 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08003416 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003417 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003418 "@FP16",
3419 "@pthreadpool",
3420 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003421)
3422
3423xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003424 name = "neonfma_ukernels_test_mode",
3425 hdrs = INTERNAL_HDRS,
3426 aarch32_copts = [
3427 "-marm",
3428 "-march=armv7-a",
3429 "-mfpu=neon-vfpv4",
3430 ],
3431 aarch32_srcs = NEONFMA_UKERNELS,
3432 aarch64_srcs = NEONFMA_UKERNELS + AARCH64_NEONFMA_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07003433 apple_aarch32_copts = [
3434 "-mcpu=swift",
3435 "-mtune=generic",
3436 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07003437 copts = [
3438 "-UNDEBUG",
3439 "-DXNN_TEST_MODE=1",
3440 ],
3441 gcc_copts = xnnpack_gcc_std_copts(),
3442 msvc_copts = xnnpack_msvc_std_copts(),
3443 deps = [
3444 ":tables",
3445 "@FP16",
3446 "@pthreadpool",
3447 ],
3448)
3449
3450xnnpack_cc_library(
Marat Dukhan8853b822020-05-07 12:19:01 -07003451 name = "neonv8_ukernels",
3452 hdrs = INTERNAL_HDRS,
3453 aarch32_copts = [
3454 "-marm",
3455 "-march=armv8-a",
3456 "-mfpu=neon-fp-armv8",
3457 ],
3458 aarch32_srcs = NEONV8_UKERNELS,
3459 aarch64_srcs = NEONV8_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07003460 apple_aarch32_copts = [
3461 "-mcpu=cyclone",
3462 "-mtune=generic",
3463 ],
Marat Dukhan8853b822020-05-07 12:19:01 -07003464 gcc_copts = xnnpack_gcc_std_copts(),
3465 msvc_copts = xnnpack_msvc_std_copts(),
3466 deps = [
3467 ":tables",
3468 "@FP16",
3469 "@pthreadpool",
3470 ],
3471)
3472
3473xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003474 name = "neonv8_ukernels_test_mode",
3475 hdrs = INTERNAL_HDRS,
3476 aarch32_copts = [
3477 "-marm",
3478 "-march=armv8-a",
3479 "-mfpu=neon-fp-armv8",
3480 ],
3481 aarch32_srcs = NEONV8_UKERNELS,
3482 aarch64_srcs = NEONV8_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07003483 apple_aarch32_copts = [
3484 "-mcpu=cyclone",
3485 "-mtune=generic",
3486 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07003487 copts = [
3488 "-UNDEBUG",
3489 "-DXNN_TEST_MODE=1",
3490 ],
3491 gcc_copts = xnnpack_gcc_std_copts(),
3492 msvc_copts = xnnpack_msvc_std_copts(),
3493 deps = [
3494 ":tables",
3495 "@FP16",
3496 "@pthreadpool",
3497 ],
3498)
3499
3500xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003501 name = "neonfp16arith_ukernels",
3502 hdrs = INTERNAL_HDRS,
3503 aarch64_copts = ["-march=armv8.2-a+fp16"],
3504 aarch64_srcs = AARCH64_NEONFP16ARITH_UKERNELS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003505 gcc_copts = xnnpack_gcc_std_copts(),
3506 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08003507 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003508 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003509 "@FP16",
3510 "@pthreadpool",
3511 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003512)
3513
3514xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003515 name = "neonfp16arith_ukernels_test_mode",
3516 hdrs = INTERNAL_HDRS,
3517 aarch64_copts = ["-march=armv8.2-a+fp16"],
3518 aarch64_srcs = AARCH64_NEONFP16ARITH_UKERNELS,
3519 copts = [
3520 "-UNDEBUG",
3521 "-DXNN_TEST_MODE=1",
3522 ],
3523 gcc_copts = xnnpack_gcc_std_copts(),
3524 msvc_copts = xnnpack_msvc_std_copts(),
3525 deps = [
3526 ":tables",
3527 "@FP16",
3528 "@pthreadpool",
3529 ],
3530)
3531
3532xnnpack_cc_library(
Benoit Jacoba9644732020-08-13 12:48:55 -07003533 name = "neondot_ukernels",
3534 hdrs = INTERNAL_HDRS,
Marat Dukhan799ac752020-08-13 16:08:03 -07003535 aarch32_copts = [
3536 "-marm",
3537 "-march=armv8.2-a+dotprod",
3538 "-mfpu=neon-fp-armv8",
3539 ],
Benoit Jacoba9644732020-08-13 12:48:55 -07003540 aarch32_srcs = NEONDOT_UKERNELS,
3541 aarch64_copts = ["-march=armv8.2-a+dotprod"],
3542 aarch64_srcs = NEONDOT_UKERNELS,
3543 gcc_copts = xnnpack_gcc_std_copts(),
3544 msvc_copts = xnnpack_msvc_std_copts(),
3545 deps = [
3546 ":tables",
3547 "@FP16",
3548 "@pthreadpool",
3549 ],
3550)
3551
3552xnnpack_cc_library(
3553 name = "neondot_ukernels_test_mode",
3554 hdrs = INTERNAL_HDRS,
Marat Dukhan799ac752020-08-13 16:08:03 -07003555 aarch32_copts = [
3556 "-marm",
3557 "-march=armv8.2-a+dotprod",
3558 "-mfpu=neon-fp-armv8",
3559 ],
Benoit Jacoba9644732020-08-13 12:48:55 -07003560 aarch32_srcs = NEONDOT_UKERNELS,
3561 aarch64_copts = ["-march=armv8.2-a+dotprod"],
3562 aarch64_srcs = NEONDOT_UKERNELS,
3563 copts = [
3564 "-UNDEBUG",
3565 "-DXNN_TEST_MODE=1",
3566 ],
3567 gcc_copts = xnnpack_gcc_std_copts(),
3568 msvc_copts = xnnpack_msvc_std_copts(),
3569 deps = [
3570 ":tables",
3571 "@FP16",
3572 "@pthreadpool",
3573 ],
3574)
3575
3576xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003577 name = "sse2_ukernels",
3578 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003579 gcc_copts = xnnpack_gcc_std_copts(),
3580 gcc_x86_copts = ["-msse2"],
3581 msvc_copts = xnnpack_msvc_std_copts(),
3582 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003583 x86_srcs = SSE_UKERNELS + SSE2_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08003584 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003585 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003586 "@FP16",
3587 "@pthreadpool",
3588 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003589)
3590
3591xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003592 name = "sse2_ukernels_test_mode",
3593 hdrs = INTERNAL_HDRS,
3594 copts = [
3595 "-UNDEBUG",
3596 "-DXNN_TEST_MODE=1",
3597 ],
3598 gcc_copts = xnnpack_gcc_std_copts(),
3599 gcc_x86_copts = ["-msse2"],
3600 msvc_copts = xnnpack_msvc_std_copts(),
3601 msvc_x86_32_copts = ["/arch:SSE2"],
3602 x86_srcs = SSE_UKERNELS + SSE2_UKERNELS,
3603 deps = [
3604 ":tables",
3605 "@FP16",
3606 "@pthreadpool",
3607 ],
3608)
3609
3610xnnpack_cc_library(
Marat Dukhanfe7acb62020-03-09 19:30:05 -07003611 name = "ssse3_ukernels",
3612 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003613 gcc_copts = xnnpack_gcc_std_copts(),
3614 gcc_x86_copts = ["-mssse3"],
3615 msvc_copts = xnnpack_msvc_std_copts(),
3616 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhanfe7acb62020-03-09 19:30:05 -07003617 x86_srcs = SSSE3_UKERNELS,
3618 deps = [
3619 ":tables",
3620 "@FP16",
3621 "@pthreadpool",
3622 ],
3623)
3624
3625xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003626 name = "ssse3_ukernels_test_mode",
3627 hdrs = INTERNAL_HDRS,
3628 copts = [
3629 "-UNDEBUG",
3630 "-DXNN_TEST_MODE=1",
3631 ],
3632 gcc_copts = xnnpack_gcc_std_copts(),
3633 gcc_x86_copts = ["-mssse3"],
3634 msvc_copts = xnnpack_msvc_std_copts(),
3635 msvc_x86_32_copts = ["/arch:SSE2"],
3636 x86_srcs = SSSE3_UKERNELS,
3637 deps = [
3638 ":tables",
3639 "@FP16",
3640 "@pthreadpool",
3641 ],
3642)
3643
3644xnnpack_cc_library(
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08003645 name = "sse41_ukernels",
3646 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003647 gcc_copts = xnnpack_gcc_std_copts(),
3648 gcc_x86_copts = ["-msse4.1"],
3649 msvc_copts = xnnpack_msvc_std_copts(),
3650 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08003651 x86_srcs = SSE41_UKERNELS,
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 Dukhan69c3f2c2019-11-06 12:30:01 -08003657)
3658
3659xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003660 name = "sse41_ukernels_test_mode",
3661 hdrs = INTERNAL_HDRS,
3662 copts = [
3663 "-UNDEBUG",
3664 "-DXNN_TEST_MODE=1",
3665 ],
3666 gcc_copts = xnnpack_gcc_std_copts(),
3667 gcc_x86_copts = ["-msse4.1"],
3668 msvc_copts = xnnpack_msvc_std_copts(),
3669 msvc_x86_32_copts = ["/arch:SSE2"],
3670 x86_srcs = SSE41_UKERNELS,
3671 deps = [
3672 ":tables",
3673 "@FP16",
3674 "@pthreadpool",
3675 ],
3676)
3677
3678xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003679 name = "avx_ukernels",
3680 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003681 gcc_copts = xnnpack_gcc_std_copts(),
3682 gcc_x86_copts = ["-mavx"],
3683 msvc_copts = xnnpack_msvc_std_copts(),
3684 msvc_x86_32_copts = ["/arch:AVX"],
3685 msvc_x86_64_copts = ["/arch:AVX"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003686 x86_srcs = AVX_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08003687 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003688 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003689 "@FP16",
3690 "@pthreadpool",
3691 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003692)
3693
3694xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003695 name = "avx_ukernels_test_mode",
3696 hdrs = INTERNAL_HDRS,
3697 copts = [
3698 "-UNDEBUG",
3699 "-DXNN_TEST_MODE=1",
3700 ],
3701 gcc_copts = xnnpack_gcc_std_copts(),
3702 gcc_x86_copts = ["-mavx"],
3703 msvc_copts = xnnpack_msvc_std_copts(),
3704 msvc_x86_32_copts = ["/arch:AVX"],
3705 msvc_x86_64_copts = ["/arch:AVX"],
3706 x86_srcs = AVX_UKERNELS,
3707 deps = [
3708 ":tables",
3709 "@FP16",
3710 "@pthreadpool",
3711 ],
3712)
3713
3714xnnpack_cc_library(
Marat Dukhan1566fee2020-08-02 21:55:41 -07003715 name = "xop_ukernels",
3716 hdrs = INTERNAL_HDRS,
3717 gcc_copts = xnnpack_gcc_std_copts(),
3718 gcc_x86_copts = ["-mxop"],
3719 msvc_copts = xnnpack_msvc_std_copts(),
3720 msvc_x86_32_copts = ["/arch:AVX"],
3721 msvc_x86_64_copts = ["/arch:AVX"],
3722 x86_srcs = XOP_UKERNELS,
3723 deps = [
3724 ":tables",
3725 "@FP16",
3726 "@pthreadpool",
3727 ],
3728)
3729
3730xnnpack_cc_library(
3731 name = "xop_ukernels_test_mode",
3732 hdrs = INTERNAL_HDRS,
3733 copts = [
3734 "-UNDEBUG",
3735 "-DXNN_TEST_MODE=1",
3736 ],
3737 gcc_copts = xnnpack_gcc_std_copts(),
3738 gcc_x86_copts = ["-mxop"],
3739 msvc_copts = xnnpack_msvc_std_copts(),
3740 msvc_x86_32_copts = ["/arch:AVX"],
3741 msvc_x86_64_copts = ["/arch:AVX"],
3742 x86_srcs = XOP_UKERNELS,
3743 deps = [
3744 ":tables",
3745 "@FP16",
3746 "@pthreadpool",
3747 ],
3748)
3749
3750xnnpack_cc_library(
Marat Dukhanfda12b82019-11-21 12:27:59 -08003751 name = "fma3_ukernels",
3752 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003753 gcc_copts = xnnpack_gcc_std_copts(),
3754 gcc_x86_copts = ["-mfma"],
3755 msvc_copts = xnnpack_msvc_std_copts(),
3756 msvc_x86_32_copts = ["/arch:AVX"],
3757 msvc_x86_64_copts = ["/arch:AVX"],
Marat Dukhanfda12b82019-11-21 12:27:59 -08003758 x86_srcs = FMA3_UKERNELS,
3759 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003760 ":tables",
Marat Dukhanfda12b82019-11-21 12:27:59 -08003761 "@FP16",
3762 "@pthreadpool",
3763 ],
3764)
3765
3766xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003767 name = "fma3_ukernels_test_mode",
3768 hdrs = INTERNAL_HDRS,
3769 copts = [
3770 "-UNDEBUG",
3771 "-DXNN_TEST_MODE=1",
3772 ],
3773 gcc_copts = xnnpack_gcc_std_copts(),
3774 gcc_x86_copts = ["-mfma"],
3775 msvc_copts = xnnpack_msvc_std_copts(),
3776 msvc_x86_32_copts = ["/arch:AVX"],
3777 msvc_x86_64_copts = ["/arch:AVX"],
3778 x86_srcs = FMA3_UKERNELS,
3779 deps = [
3780 ":tables",
3781 "@FP16",
3782 "@pthreadpool",
3783 ],
3784)
3785
3786xnnpack_cc_library(
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003787 name = "avx2_ukernels",
3788 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003789 gcc_copts = xnnpack_gcc_std_copts(),
3790 gcc_x86_copts = [
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003791 "-mfma",
3792 "-mavx2",
3793 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07003794 msvc_copts = xnnpack_msvc_std_copts(),
3795 msvc_x86_32_copts = ["/arch:AVX2"],
3796 msvc_x86_64_copts = ["/arch:AVX2"],
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003797 x86_srcs = AVX2_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08003798 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003799 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003800 "@FP16",
3801 "@pthreadpool",
3802 ],
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003803)
3804
3805xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003806 name = "avx2_ukernels_test_mode",
3807 hdrs = INTERNAL_HDRS,
3808 copts = [
3809 "-UNDEBUG",
3810 "-DXNN_TEST_MODE=1",
3811 ],
3812 gcc_copts = xnnpack_gcc_std_copts(),
3813 gcc_x86_copts = [
3814 "-mfma",
3815 "-mavx2",
3816 ],
3817 msvc_copts = xnnpack_msvc_std_copts(),
3818 msvc_x86_32_copts = ["/arch:AVX2"],
3819 msvc_x86_64_copts = ["/arch:AVX2"],
3820 x86_srcs = AVX2_UKERNELS,
3821 deps = [
3822 ":tables",
3823 "@FP16",
3824 "@pthreadpool",
3825 ],
3826)
3827
3828xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003829 name = "avx512f_ukernels",
3830 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003831 gcc_copts = xnnpack_gcc_std_copts(),
3832 gcc_x86_copts = ["-mavx512f"],
3833 mingw_copts = ["-fno-asynchronous-unwind-tables"],
3834 msvc_copts = xnnpack_msvc_std_copts(),
3835 msvc_x86_32_copts = ["/arch:AVX512"],
3836 msvc_x86_64_copts = ["/arch:AVX512"],
3837 msys_copts = ["-fno-asynchronous-unwind-tables"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003838 x86_srcs = AVX512F_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08003839 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003840 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003841 "@FP16",
3842 "@pthreadpool",
3843 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003844)
3845
3846xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003847 name = "avx512f_ukernels_test_mode",
3848 hdrs = INTERNAL_HDRS,
3849 copts = [
3850 "-UNDEBUG",
3851 "-DXNN_TEST_MODE=1",
3852 ],
3853 gcc_copts = xnnpack_gcc_std_copts(),
3854 gcc_x86_copts = ["-mavx512f"],
3855 mingw_copts = ["-fno-asynchronous-unwind-tables"],
3856 msvc_copts = xnnpack_msvc_std_copts(),
3857 msvc_x86_32_copts = ["/arch:AVX512"],
3858 msvc_x86_64_copts = ["/arch:AVX512"],
3859 msys_copts = ["-fno-asynchronous-unwind-tables"],
3860 x86_srcs = AVX512F_UKERNELS,
3861 deps = [
3862 ":tables",
3863 "@FP16",
3864 "@pthreadpool",
3865 ],
3866)
3867
3868xnnpack_cc_library(
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07003869 name = "avx512skx_ukernels",
3870 hdrs = INTERNAL_HDRS,
3871 gcc_copts = xnnpack_gcc_std_copts(),
3872 gcc_x86_copts = [
3873 "-mavx512f",
3874 "-mavx512cd",
3875 "-mavx512bw",
3876 "-mavx512dq",
3877 "-mavx512vl",
3878 ],
3879 mingw_copts = ["-fno-asynchronous-unwind-tables"],
3880 msvc_copts = xnnpack_msvc_std_copts(),
3881 msvc_x86_32_copts = ["/arch:AVX512"],
3882 msvc_x86_64_copts = ["/arch:AVX512"],
3883 msys_copts = ["-fno-asynchronous-unwind-tables"],
3884 x86_srcs = AVX512SKX_UKERNELS,
3885 deps = [
3886 ":tables",
3887 "@FP16",
3888 "@pthreadpool",
3889 ],
3890)
3891
3892xnnpack_cc_library(
3893 name = "avx512skx_ukernels_test_mode",
3894 hdrs = INTERNAL_HDRS,
3895 copts = [
3896 "-UNDEBUG",
3897 "-DXNN_TEST_MODE=1",
3898 ],
3899 gcc_copts = xnnpack_gcc_std_copts(),
3900 gcc_x86_copts = [
3901 "-mavx512f",
3902 "-mavx512cd",
3903 "-mavx512bw",
3904 "-mavx512dq",
3905 "-mavx512vl",
3906 ],
3907 mingw_copts = ["-fno-asynchronous-unwind-tables"],
3908 msvc_copts = xnnpack_msvc_std_copts(),
3909 msvc_x86_32_copts = ["/arch:AVX512"],
3910 msvc_x86_64_copts = ["/arch:AVX512"],
3911 msys_copts = ["-fno-asynchronous-unwind-tables"],
3912 x86_srcs = AVX512SKX_UKERNELS,
3913 deps = [
3914 ":tables",
3915 "@FP16",
3916 "@pthreadpool",
3917 ],
3918)
3919
3920xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003921 name = "asm_ukernels",
3922 hdrs = ["src/xnnpack/assembly.h"],
3923 aarch32_srcs = AARCH32_ASM_UKERNELS,
Frank Barchard31bb45b2020-10-06 00:26:33 -07003924 aarch64_copts = ["-march=armv8.2-a+fp16+dotprod"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003925 aarch64_srcs = AARCH64_ASM_UKERNELS,
Frank Barchardbcedc082020-08-17 18:00:51 -07003926 wasm_srcs = WASM32_ASM_UKERNELS,
3927 wasmsimd_srcs = WASM32_ASM_UKERNELS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003928)
3929
Marat Dukhan3b59de22020-06-03 20:15:19 -07003930xnnpack_cc_library(
3931 name = "logging_utils",
3932 srcs = LOGGING_SRCS,
3933 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
3934 copts = LOGGING_COPTS + [
3935 "-Isrc",
3936 "-Iinclude",
3937 ] + select({
3938 ":debug_build": [],
3939 "//conditions:default": xnnpack_min_size_copts(),
3940 }),
3941 gcc_copts = xnnpack_gcc_std_copts(),
3942 msvc_copts = xnnpack_msvc_std_copts(),
3943 visibility = xnnpack_visibility(),
3944 deps = [
3945 "@FP16",
3946 "@clog",
3947 "@pthreadpool",
3948 ],
3949)
3950
Marat Dukhan08c4a432019-10-03 09:29:21 -07003951xnnpack_aggregate_library(
3952 name = "ukernels",
3953 aarch32_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003954 ":neon_ukernels",
3955 ":neonfma_ukernels",
Marat Dukhan8853b822020-05-07 12:19:01 -07003956 ":neonv8_ukernels",
Benoit Jacoba9644732020-08-13 12:48:55 -07003957 ":neondot_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003958 ":asm_ukernels",
3959 ],
3960 aarch64_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003961 ":neon_ukernels",
3962 ":neonfma_ukernels",
Marat Dukhan8853b822020-05-07 12:19:01 -07003963 ":neonv8_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003964 ":neonfp16arith_ukernels",
Benoit Jacoba9644732020-08-13 12:48:55 -07003965 ":neondot_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003966 ":asm_ukernels",
3967 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07003968 generic_deps = [
3969 ":scalar_ukernels",
3970 ],
Marat Dukhan500b8892020-04-15 17:09:50 -07003971 psimd_deps = [
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08003972 ":psimd_accmath_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003973 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07003974 wasm_deps = [
3975 ":wasm_ukernels",
Frank Barchardbcedc082020-08-17 18:00:51 -07003976 ":asm_ukernels",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003977 ],
3978 wasmsimd_deps = [
3979 ":wasm_ukernels",
Frank Barchardbcedc082020-08-17 18:00:51 -07003980 ":asm_ukernels",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003981 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003982 x86_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003983 ":sse2_ukernels",
Marat Dukhanfe7acb62020-03-09 19:30:05 -07003984 ":ssse3_ukernels",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08003985 ":sse41_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003986 ":avx_ukernels",
Marat Dukhan1566fee2020-08-02 21:55:41 -07003987 ":xop_ukernels",
Marat Dukhanfda12b82019-11-21 12:27:59 -08003988 ":fma3_ukernels",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003989 ":avx2_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003990 ":avx512f_ukernels",
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07003991 ":avx512skx_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003992 ],
3993)
3994
Marat Dukhan33fcf782020-05-24 14:27:15 -07003995xnnpack_aggregate_library(
3996 name = "ukernels_test_mode",
3997 aarch32_deps = [
3998 ":neon_ukernels_test_mode",
3999 ":neonfma_ukernels_test_mode",
4000 ":neonv8_ukernels_test_mode",
Benoit Jacoba9644732020-08-13 12:48:55 -07004001 ":neondot_ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004002 ":asm_ukernels",
4003 ],
4004 aarch64_deps = [
4005 ":neon_ukernels_test_mode",
4006 ":neonfma_ukernels_test_mode",
4007 ":neonv8_ukernels_test_mode",
4008 ":neonfp16arith_ukernels_test_mode",
Benoit Jacoba9644732020-08-13 12:48:55 -07004009 ":neondot_ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004010 ":asm_ukernels",
4011 ],
4012 generic_deps = [
4013 ":scalar_ukernels_test_mode",
4014 ],
4015 psimd_deps = [
Marat Dukhan33fcf782020-05-24 14:27:15 -07004016 ":psimd_accmath_ukernels_test_mode",
4017 ],
4018 wasm_deps = [
4019 ":wasm_ukernels_test_mode",
Frank Barchardbcedc082020-08-17 18:00:51 -07004020 ":asm_ukernels",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004021 ],
4022 wasmsimd_deps = [
4023 ":wasm_ukernels_test_mode",
Frank Barchardbcedc082020-08-17 18:00:51 -07004024 ":asm_ukernels",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004025 ],
4026 x86_deps = [
4027 ":sse2_ukernels_test_mode",
4028 ":ssse3_ukernels_test_mode",
4029 ":sse41_ukernels_test_mode",
4030 ":avx_ukernels_test_mode",
Marat Dukhan1566fee2020-08-02 21:55:41 -07004031 ":xop_ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004032 ":fma3_ukernels_test_mode",
4033 ":avx2_ukernels_test_mode",
4034 ":avx512f_ukernels_test_mode",
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07004035 ":avx512skx_ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004036 ],
4037)
4038
Marat Dukhan86df3842020-06-17 14:03:31 -07004039# This target must be compatible with Android NDK r17 and compile with Clang 6.
4040# It is similar to :ukernels target, but excludes microkernels which require
4041# NEON FP16 arithmetics, NEON Dot Product, and newer intrinsics.
Marat Dukhanae046f52020-06-15 13:16:14 -07004042xnnpack_aggregate_library(
Marat Dukhan86df3842020-06-17 14:03:31 -07004043 name = "ukernels_max_compatibility",
Marat Dukhanae046f52020-06-15 13:16:14 -07004044 aarch32_deps = [
4045 ":neon_ukernels",
4046 ":neonfma_ukernels",
4047 ":neonv8_ukernels",
4048 ":asm_ukernels",
4049 ],
4050 aarch64_deps = [
4051 ":neon_ukernels",
4052 ":neonfma_ukernels",
4053 ":neonv8_ukernels",
4054 ":asm_ukernels",
4055 ],
4056 generic_deps = [
4057 ":scalar_ukernels",
4058 ],
4059 psimd_deps = [
Marat Dukhanae046f52020-06-15 13:16:14 -07004060 ":psimd_accmath_ukernels",
4061 ],
4062 wasm_deps = [
4063 ":wasm_ukernels",
4064 ],
4065 wasmsimd_deps = [
4066 ":wasm_ukernels",
4067 ],
4068 x86_deps = [
4069 ":sse2_ukernels",
4070 ":ssse3_ukernels",
4071 ":sse41_ukernels",
4072 ":avx_ukernels",
Marat Dukhan1566fee2020-08-02 21:55:41 -07004073 ":xop_ukernels",
Marat Dukhanae046f52020-06-15 13:16:14 -07004074 ":fma3_ukernels",
4075 ":avx2_ukernels",
4076 ":avx512f_ukernels",
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07004077 ":avx512skx_ukernels",
Marat Dukhanae046f52020-06-15 13:16:14 -07004078 ],
4079)
4080
Marat Dukhan08c4a432019-10-03 09:29:21 -07004081xnnpack_cc_library(
4082 name = "im2col",
4083 srcs = ["src/im2col.c"],
4084 hdrs = [
4085 "src/xnnpack/common.h",
4086 "src/xnnpack/im2col.h",
4087 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07004088 gcc_copts = xnnpack_gcc_std_copts(),
4089 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004090)
4091
4092xnnpack_cc_library(
4093 name = "indirection",
4094 srcs = ["src/indirection.c"],
4095 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004096 gcc_copts = xnnpack_gcc_std_copts(),
4097 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004098 deps = [
4099 "@FP16",
4100 "@FXdiv",
4101 "@pthreadpool",
4102 ],
4103)
4104
4105xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07004106 name = "indirection_test_mode",
4107 srcs = ["src/indirection.c"],
4108 hdrs = INTERNAL_HDRS,
4109 copts = [
4110 "-UNDEBUG",
4111 "-DXNN_TEST_MODE=1",
4112 ],
4113 gcc_copts = xnnpack_gcc_std_copts(),
4114 msvc_copts = xnnpack_msvc_std_copts(),
4115 deps = [
4116 "@FP16",
4117 "@FXdiv",
4118 "@pthreadpool",
4119 ],
4120)
4121
4122xnnpack_cc_library(
Marat Dukhanab582382020-07-06 13:32:08 -07004123 name = "packing",
4124 srcs = ["src/packing.c"],
4125 hdrs = INTERNAL_HDRS,
4126 gcc_copts = xnnpack_gcc_std_copts(),
4127 msvc_copts = xnnpack_msvc_std_copts(),
4128 deps = [
4129 "@FP16",
4130 "@FXdiv",
4131 "@pthreadpool",
4132 ],
4133)
4134
4135xnnpack_cc_library(
4136 name = "packing_test_mode",
4137 srcs = ["src/packing.c"],
4138 hdrs = INTERNAL_HDRS,
4139 copts = [
4140 "-UNDEBUG",
4141 "-DXNN_TEST_MODE=1",
4142 ],
4143 gcc_copts = xnnpack_gcc_std_copts(),
4144 msvc_copts = xnnpack_msvc_std_copts(),
4145 deps = [
4146 "@FP16",
4147 "@FXdiv",
4148 "@pthreadpool",
4149 ],
4150)
4151
4152xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004153 name = "operator_run",
4154 srcs = ["src/operator-run.c"],
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07004155 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004156 copts = LOGGING_COPTS + select({
Marat Dukhan05702cf2020-03-26 15:41:33 -07004157 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4158 "//conditions:default": [],
4159 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07004160 gcc_copts = xnnpack_gcc_std_copts(),
4161 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004162 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07004163 ":logging_utils",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004164 "@FP16",
4165 "@FXdiv",
4166 "@clog",
4167 "@pthreadpool",
4168 ],
4169)
4170
Chao Mei6ddfc602020-05-13 22:29:36 -07004171xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07004172 name = "operator_run_test_mode",
4173 srcs = ["src/operator-run.c"],
4174 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
4175 copts = LOGGING_COPTS + [
4176 "-UNDEBUG",
4177 "-DXNN_TEST_MODE=1",
4178 ] + select({
4179 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4180 "//conditions:default": [],
4181 }),
4182 gcc_copts = xnnpack_gcc_std_copts(),
4183 msvc_copts = xnnpack_msvc_std_copts(),
4184 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07004185 ":logging_utils",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004186 "@FP16",
4187 "@FXdiv",
4188 "@clog",
4189 "@pthreadpool",
4190 ],
4191)
4192
4193xnnpack_cc_library(
Chao Mei6ddfc602020-05-13 22:29:36 -07004194 name = "memory_planner",
4195 srcs = ["src/memory-planner.c"],
4196 hdrs = INTERNAL_HDRS,
4197 defines = select({
4198 ":xnn_enable_memopt_explicit_true": ["XNN_ENABLE_MEMOPT=1"],
4199 ":xnn_enable_memopt_explicit_false": ["XNN_ENABLE_MEMOPT=0"],
4200 "//conditions:default": ["XNN_ENABLE_MEMOPT=1"],
4201 }),
4202 gcc_copts = xnnpack_gcc_std_copts(),
4203 msvc_copts = xnnpack_msvc_std_copts(),
4204 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07004205 ":logging_utils",
Chao Mei6ddfc602020-05-13 22:29:36 -07004206 "@pthreadpool",
4207 ],
4208)
4209
Marat Dukhan33fcf782020-05-24 14:27:15 -07004210xnnpack_cc_library(
4211 name = "memory_planner_test_mode",
4212 srcs = ["src/memory-planner.c"],
4213 hdrs = INTERNAL_HDRS,
4214 copts = [
4215 "-UNDEBUG",
4216 "-DXNN_TEST_MODE=1",
4217 ],
4218 defines = select({
4219 ":xnn_enable_memopt_explicit_true": ["XNN_ENABLE_MEMOPT=1"],
4220 ":xnn_enable_memopt_explicit_false": ["XNN_ENABLE_MEMOPT=0"],
4221 "//conditions:default": ["XNN_ENABLE_MEMOPT=1"],
4222 }),
4223 gcc_copts = xnnpack_gcc_std_copts(),
4224 msvc_copts = xnnpack_msvc_std_copts(),
4225 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07004226 ":logging_utils",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004227 "@pthreadpool",
4228 ],
4229)
4230
Marat Dukhan08c4a432019-10-03 09:29:21 -07004231cc_library(
4232 name = "enable_assembly",
4233 defines = select({
4234 ":xnn_enable_assembly_explicit_true": ["XNN_ENABLE_ASSEMBLY=1"],
4235 ":xnn_enable_assembly_explicit_false": ["XNN_ENABLE_ASSEMBLY=0"],
Frank Barchard810171d2019-10-10 10:34:51 -07004236 "//conditions:default": ["XNN_ENABLE_ASSEMBLY=1"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004237 }),
4238)
4239
Marat Dukhan9de90e02020-06-18 16:04:12 -07004240cc_library(
4241 name = "enable_sparse",
4242 defines = select({
4243 ":xnn_enable_sparse_explicit_true": ["XNN_ENABLE_SPARSE=1"],
4244 ":xnn_enable_sparse_explicit_false": ["XNN_ENABLE_SPARSE=0"],
Marat Dukhan8b283aa2020-06-18 16:48:57 -07004245 "//conditions:default": ["XNN_ENABLE_SPARSE=0"],
Marat Dukhan9de90e02020-06-18 16:04:12 -07004246 }),
4247)
4248
Marat Dukhancf056b22019-10-07 10:26:29 -07004249xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004250 name = "operators",
4251 srcs = OPERATOR_SRCS + [
Marat Dukhan04f03be2019-11-19 12:36:47 -08004252 "src/memory.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004253 "src/operator-delete.c",
Marat Dukhancf056b22019-10-07 10:26:29 -07004254 ],
4255 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004256 copts = LOGGING_COPTS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07004257 "-Isrc",
4258 "-Iinclude",
4259 ] + select({
4260 ":debug_build": [],
4261 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07004262 }) + select({
4263 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4264 "//conditions:default": [],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004265 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07004266 gcc_copts = xnnpack_gcc_std_copts(),
4267 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004268 deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07004269 ":indirection",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004270 ":logging_utils",
Marat Dukhanab582382020-07-06 13:32:08 -07004271 ":packing",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004272 "@FP16",
4273 "@FXdiv",
4274 "@clog",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004275 "@pthreadpool",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004276 ],
4277)
4278
Marat Dukhan10a38082020-04-17 03:58:35 -07004279xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07004280 name = "operators_test_mode",
4281 srcs = OPERATOR_SRCS + [
4282 "src/memory.c",
4283 "src/operator-delete.c",
4284 ],
4285 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
4286 copts = LOGGING_COPTS + [
4287 "-Isrc",
4288 "-Iinclude",
4289 "-UNDEBUG",
4290 "-DXNN_TEST_MODE=1",
4291 ] + select({
4292 ":debug_build": [],
4293 "//conditions:default": xnnpack_min_size_copts(),
4294 }) + select({
4295 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4296 "//conditions:default": [],
4297 }),
4298 gcc_copts = xnnpack_gcc_std_copts(),
4299 msvc_copts = xnnpack_msvc_std_copts(),
4300 deps = [
4301 ":indirection_test_mode",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004302 ":logging_utils",
Marat Dukhanab582382020-07-06 13:32:08 -07004303 ":packing_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004304 "@FP16",
4305 "@FXdiv",
4306 "@clog",
4307 "@pthreadpool",
4308 ],
4309)
4310
4311xnnpack_cc_library(
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004312 name = "XNNPACK",
4313 srcs = [
4314 "src/init.c",
Marat Dukhanccfdbd12020-02-03 14:27:45 -08004315 "src/runtime.c",
4316 "src/subgraph.c",
4317 "src/tensor.c",
Marat Dukhanf03da0d2020-06-10 16:00:20 -07004318 ] + SUBGRAPH_SRCS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004319 hdrs = ["include/xnnpack.h"],
4320 copts = LOGGING_COPTS + [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004321 "-Isrc",
4322 "-Iinclude",
4323 ] + select({
4324 ":debug_build": [],
4325 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07004326 }) + select({
4327 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4328 "//conditions:default": [],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004329 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07004330 gcc_copts = xnnpack_gcc_std_copts(),
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004331 includes = ["include"],
Marat Dukhan10a38082020-04-17 03:58:35 -07004332 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004333 visibility = xnnpack_visibility(),
4334 deps = [
4335 ":enable_assembly",
Marat Dukhan9de90e02020-06-18 16:04:12 -07004336 ":enable_sparse",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004337 ":logging_utils",
Chao Mei6ddfc602020-05-13 22:29:36 -07004338 ":memory_planner",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004339 ":operator_run",
4340 ":operators",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004341 ":ukernels",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004342 "@clog",
Marat Dukhanab2946c2020-05-21 20:04:13 -07004343 "@FP16",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004344 "@pthreadpool",
Marat Dukhand343c222019-10-07 09:22:14 -07004345 ] + select({
4346 ":emscripten": [],
4347 "//conditions:default": ["@cpuinfo"],
4348 }),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004349)
4350
Marat Dukhan10a38082020-04-17 03:58:35 -07004351xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07004352 name = "XNNPACK_test_mode",
4353 srcs = [
4354 "src/init.c",
4355 "src/runtime.c",
4356 "src/subgraph.c",
4357 "src/tensor.c",
Marat Dukhanf03da0d2020-06-10 16:00:20 -07004358 ] + SUBGRAPH_SRCS,
Marat Dukhan33fcf782020-05-24 14:27:15 -07004359 hdrs = ["include/xnnpack.h"],
4360 copts = LOGGING_COPTS + [
4361 "-Isrc",
4362 "-Iinclude",
4363 "-UNDEBUG",
4364 "-DXNN_TEST_MODE=1",
4365 ] + select({
4366 ":debug_build": [],
4367 "//conditions:default": xnnpack_min_size_copts(),
4368 }) + select({
4369 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4370 "//conditions:default": [],
4371 }),
4372 gcc_copts = xnnpack_gcc_std_copts(),
4373 includes = ["include"],
4374 msvc_copts = xnnpack_msvc_std_copts(),
4375 visibility = xnnpack_visibility(),
4376 deps = [
4377 ":enable_assembly",
Marat Dukhan9de90e02020-06-18 16:04:12 -07004378 ":enable_sparse",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004379 ":logging_utils",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004380 ":memory_planner_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004381 ":operator_run_test_mode",
4382 ":operators_test_mode",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004383 ":ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004384 "@clog",
4385 "@FP16",
4386 "@pthreadpool",
4387 ] + select({
4388 ":emscripten": [],
4389 "//conditions:default": ["@cpuinfo"],
4390 }),
4391)
4392
4393xnnpack_cc_library(
Marat Dukhanae046f52020-06-15 13:16:14 -07004394 name = "xnnpack_f32",
4395 srcs = [
4396 "src/init.c",
4397 "src/runtime.c",
4398 "src/subgraph.c",
4399 "src/tensor.c",
4400 ] + SUBGRAPH_SRCS,
4401 hdrs = ["include/xnnpack.h"],
4402 copts = LOGGING_COPTS + [
4403 "-Isrc",
4404 "-Iinclude",
4405 ] + select({
4406 ":debug_build": [],
4407 "//conditions:default": xnnpack_min_size_copts(),
4408 }) + select({
4409 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4410 "//conditions:default": [],
4411 }),
4412 defines = [
Marat Dukhan16f1e1a2020-08-04 16:38:22 -07004413 "XNN_NO_QS8_OPERATORS",
Marat Dukhan08b7a972020-07-14 18:17:29 -07004414 "XNN_NO_QU8_OPERATORS",
Marat Dukhanae046f52020-06-15 13:16:14 -07004415 "XNN_NO_U8_OPERATORS",
4416 "XNN_NO_X8_OPERATORS",
4417 "XNN_NO_F16_OPERATORS",
4418 "XNN_NO_X16_OPERATORS",
4419 ],
4420 gcc_copts = xnnpack_gcc_std_copts(),
4421 includes = ["include"],
4422 msvc_copts = xnnpack_msvc_std_copts(),
4423 visibility = xnnpack_visibility(),
4424 deps = [
4425 ":enable_assembly",
Marat Dukhan9de90e02020-06-18 16:04:12 -07004426 ":enable_sparse",
Marat Dukhanae046f52020-06-15 13:16:14 -07004427 ":logging_utils",
4428 ":memory_planner",
4429 ":operator_run",
4430 ":operators",
Marat Dukhan86df3842020-06-17 14:03:31 -07004431 ":ukernels_max_compatibility",
Marat Dukhanae046f52020-06-15 13:16:14 -07004432 "@clog",
4433 "@FP16",
4434 "@pthreadpool",
4435 ] + select({
4436 ":emscripten": [],
4437 "//conditions:default": ["@cpuinfo"],
4438 }),
4439)
4440
4441xnnpack_cc_library(
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004442 name = "xnnpack_operators_nhwc_f32",
4443 srcs = [
4444 "src/init.c",
4445 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07004446 hdrs = ["include/xnnpack.h"],
4447 copts = LOGGING_COPTS + [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004448 "-Isrc",
4449 "-Iinclude",
4450 ] + select({
4451 ":debug_build": [],
4452 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07004453 }) + select({
4454 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4455 "//conditions:default": [],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004456 }),
4457 defines = [
Marat Dukhan16f1e1a2020-08-04 16:38:22 -07004458 "XNN_NO_QS8_OPERATORS",
Marat Dukhan08b7a972020-07-14 18:17:29 -07004459 "XNN_NO_QU8_OPERATORS",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004460 "XNN_NO_U8_OPERATORS",
4461 "XNN_NO_X8_OPERATORS",
Marat Dukhanefc47b82019-11-18 09:25:38 -08004462 "XNN_NO_NCHW_OPERATORS",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004463 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07004464 gcc_copts = xnnpack_gcc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004465 includes = ["include"],
Marat Dukhan10a38082020-04-17 03:58:35 -07004466 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004467 visibility = xnnpack_visibility(),
4468 deps = [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004469 ":enable_assembly",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004470 ":logging_utils",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004471 ":operator_run",
4472 ":operators",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004473 ":ukernels",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004474 "@clog",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004475 "@pthreadpool",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004476 ] + select({
4477 ":emscripten": [],
4478 "//conditions:default": ["@cpuinfo"],
4479 }),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004480)
4481
Marat Dukhancf056b22019-10-07 10:26:29 -07004482xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004483 name = "bench_utils",
4484 srcs = ["bench/utils.cc"],
4485 hdrs = ["bench/utils.h"],
Marat Dukhanbad48fe2019-11-04 10:35:22 -08004486 deps = [
4487 "@com_google_benchmark//:benchmark",
4488 "@cpuinfo",
4489 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004490)
4491
Frank Barchard7e955972019-10-11 10:34:25 -07004492######################### Benchmarks for micro-kernels #########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07004493
4494xnnpack_benchmark(
Marat Dukhan595e1702020-07-31 10:12:52 -07004495 name = "qs8_gemm_bench",
4496 srcs = [
4497 "bench/gemm.h",
4498 "bench/qs8-gemm.cc",
4499 "src/xnnpack/AlignedAllocator.h",
4500 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Frank Barchard31328cb2020-10-12 11:55:18 -07004501 copts = xnnpack_optional_ruy_copts() + xnnpack_optional_gemmlowp_copts(),
4502 deps = MICROKERNEL_BENCHMARK_DEPS + [":packing"] + xnnpack_optional_ruy_deps() + xnnpack_optional_gemmlowp_deps(),
Marat Dukhan595e1702020-07-31 10:12:52 -07004503)
4504
4505xnnpack_benchmark(
Marat Dukhan56bdd4a2020-08-03 19:47:04 -07004506 name = "qs8_requantization_bench",
4507 srcs = [
4508 "bench/qs8-requantization.cc",
Marat Dukhan56bdd4a2020-08-03 19:47:04 -07004509 "src/xnnpack/AlignedAllocator.h",
Frank Barchard04336c12020-10-22 16:48:55 -07004510 "src/xnnpack/requantization-stubs.h",
Marat Dukhan56bdd4a2020-08-03 19:47:04 -07004511 ] + MICROKERNEL_BENCHMARK_HDRS,
4512 deps = MICROKERNEL_BENCHMARK_DEPS,
4513)
4514
4515xnnpack_benchmark(
Marat Dukhan08b7a972020-07-14 18:17:29 -07004516 name = "qu8_gemm_bench",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004517 srcs = [
4518 "bench/gemm.h",
Marat Dukhan08b7a972020-07-14 18:17:29 -07004519 "bench/qu8-gemm.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004520 "src/xnnpack/AlignedAllocator.h",
4521 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004522 copts = xnnpack_optional_ruy_copts() + xnnpack_optional_gemmlowp_copts(),
Marat Dukhanab582382020-07-06 13:32:08 -07004523 deps = MICROKERNEL_BENCHMARK_DEPS + [":packing"] + xnnpack_optional_ruy_deps() + xnnpack_optional_gemmlowp_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004524)
4525
4526xnnpack_benchmark(
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07004527 name = "qu8_requantization_bench",
4528 srcs = [
4529 "bench/qu8-requantization.cc",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07004530 "src/xnnpack/AlignedAllocator.h",
Frank Barchard04336c12020-10-22 16:48:55 -07004531 "src/xnnpack/requantization-stubs.h",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07004532 ] + MICROKERNEL_BENCHMARK_HDRS,
4533 deps = MICROKERNEL_BENCHMARK_DEPS,
4534)
4535
4536xnnpack_benchmark(
Frank Barchard40d20fe2020-05-05 00:37:45 -07004537 name = "f16_igemm_bench",
4538 srcs = [
4539 "bench/f16-igemm.cc",
4540 "bench/conv.h",
4541 "bench/google/conv.h",
4542 "src/xnnpack/AlignedAllocator.h",
4543 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004544 deps = MICROKERNEL_BENCHMARK_DEPS + [
4545 ":indirection",
4546 ":packing",
4547 ],
Frank Barchard40d20fe2020-05-05 00:37:45 -07004548)
4549
4550xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004551 name = "f16_gemm_bench",
4552 srcs = [
4553 "bench/f16-gemm.cc",
4554 "bench/gemm.h",
4555 "src/xnnpack/AlignedAllocator.h",
4556 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004557 deps = MICROKERNEL_BENCHMARK_DEPS + [
4558 ":packing",
4559 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004560)
4561
4562xnnpack_benchmark(
Marat Dukhanbdb56f52020-02-05 21:42:49 -08004563 name = "f16_spmm_bench",
4564 srcs = [
4565 "bench/f16-spmm.cc",
4566 "bench/gemm.h",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08004567 "src/xnnpack/AlignedAllocator.h",
4568 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanbdb56f52020-02-05 21:42:49 -08004569 deps = MICROKERNEL_BENCHMARK_DEPS,
4570)
4571
4572xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004573 name = "f32_igemm_bench",
4574 srcs = [
4575 "bench/f32-igemm.cc",
4576 "bench/conv.h",
4577 "src/xnnpack/AlignedAllocator.h",
4578 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004579 deps = MICROKERNEL_BENCHMARK_DEPS + [
4580 ":indirection",
4581 ":packing",
4582 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004583)
4584
4585xnnpack_benchmark(
Frank Barchardc2010062020-07-19 10:04:12 -07004586 name = "f16_relu_bench",
4587 srcs = [
4588 "bench/f16-relu.cc",
4589 "src/xnnpack/AlignedAllocator.h",
4590 ] + MICROKERNEL_BENCHMARK_HDRS,
4591 deps = MICROKERNEL_BENCHMARK_DEPS,
4592)
4593
4594xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004595 name = "f32_conv_hwc_bench",
4596 srcs = [
4597 "bench/f32-conv-hwc.cc",
4598 "bench/dconv.h",
4599 "src/xnnpack/AlignedAllocator.h",
4600 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004601 deps = MICROKERNEL_BENCHMARK_DEPS + [
4602 ":packing",
4603 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004604)
4605
4606xnnpack_benchmark(
Marat Dukhan1f29b802020-05-15 23:46:39 -07004607 name = "f32_conv_hwc2chw_bench",
Erich Elsen563df5f2019-10-23 08:02:21 -07004608 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07004609 "bench/f32-conv-hwc2chw.cc",
Erich Elsen563df5f2019-10-23 08:02:21 -07004610 "bench/dconv.h",
4611 "src/xnnpack/AlignedAllocator.h",
4612 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004613 deps = MICROKERNEL_BENCHMARK_DEPS + [
4614 ":packing",
4615 ],
Erich Elsen563df5f2019-10-23 08:02:21 -07004616)
4617
4618xnnpack_benchmark(
Frank Barchard5a599a62020-06-04 20:12:44 -07004619 name = "f16_dwconv_bench",
4620 srcs = [
4621 "bench/f16-dwconv.cc",
4622 "bench/dwconv.h",
4623 "bench/google/dwconv.h",
4624 "src/xnnpack/AlignedAllocator.h",
4625 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004626 deps = MICROKERNEL_BENCHMARK_DEPS + [
4627 ":indirection",
4628 ":packing",
4629 ],
Frank Barchard5a599a62020-06-04 20:12:44 -07004630)
4631
4632xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004633 name = "f32_dwconv_bench",
4634 srcs = [
4635 "bench/f32-dwconv.cc",
4636 "bench/dwconv.h",
4637 "src/xnnpack/AlignedAllocator.h",
4638 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004639 deps = MICROKERNEL_BENCHMARK_DEPS + [
4640 ":indirection",
4641 ":packing",
4642 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004643)
4644
4645xnnpack_benchmark(
Marat Dukhanbf715f92020-10-23 20:17:00 -07004646 name = "f32_dwconv2d_chw_bench",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004647 srcs = [
Marat Dukhanbf715f92020-10-23 20:17:00 -07004648 "bench/f32-dwconv2d-chw.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004649 "bench/dwconv.h",
4650 "src/xnnpack/AlignedAllocator.h",
4651 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004652 deps = MICROKERNEL_BENCHMARK_DEPS + [
4653 ":indirection",
4654 ":packing",
4655 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004656)
4657
4658xnnpack_benchmark(
4659 name = "f32_gemm_bench",
4660 srcs = [
4661 "bench/f32-gemm.cc",
4662 "bench/gemm.h",
4663 "src/xnnpack/AlignedAllocator.h",
4664 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004665 copts = xnnpack_optional_ruy_copts(),
Marat Dukhanab582382020-07-06 13:32:08 -07004666 deps = MICROKERNEL_BENCHMARK_DEPS + [":packing"] + xnnpack_optional_ruy_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004667)
4668
4669xnnpack_benchmark(
Marat Dukhana11ca342020-06-25 23:45:07 -07004670 name = "f32_hswish_bench",
4671 srcs = [
4672 "bench/f32-hswish.cc",
4673 "src/xnnpack/AlignedAllocator.h",
4674 ] + MICROKERNEL_BENCHMARK_HDRS,
4675 deps = MICROKERNEL_BENCHMARK_DEPS,
4676)
4677
4678xnnpack_benchmark(
Marat Dukhan4c4eb002019-12-08 21:27:49 -08004679 name = "f32_raddexpminusmax_bench",
4680 srcs = [
4681 "bench/f32-raddexpminusmax.cc",
4682 "src/xnnpack/AlignedAllocator.h",
4683 ] + MICROKERNEL_BENCHMARK_HDRS,
4684 deps = MICROKERNEL_BENCHMARK_DEPS,
4685)
4686
4687xnnpack_benchmark(
4688 name = "f32_raddextexp_bench",
4689 srcs = [
4690 "bench/f32-raddextexp.cc",
4691 "src/xnnpack/AlignedAllocator.h",
4692 ] + MICROKERNEL_BENCHMARK_HDRS,
4693 deps = MICROKERNEL_BENCHMARK_DEPS,
4694)
4695
4696xnnpack_benchmark(
4697 name = "f32_raddstoreexpminusmax_bench",
4698 srcs = [
4699 "bench/f32-raddstoreexpminusmax.cc",
4700 "src/xnnpack/AlignedAllocator.h",
4701 ] + MICROKERNEL_BENCHMARK_HDRS,
4702 deps = MICROKERNEL_BENCHMARK_DEPS,
4703)
4704
4705xnnpack_benchmark(
Frank Barchard3c4a9522020-07-18 16:11:22 -07004706 name = "f32_relu_bench",
4707 srcs = [
4708 "bench/f32-relu.cc",
4709 "src/xnnpack/AlignedAllocator.h",
4710 ] + MICROKERNEL_BENCHMARK_HDRS,
4711 deps = MICROKERNEL_BENCHMARK_DEPS,
4712)
4713
4714xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004715 name = "f32_rmax_bench",
4716 srcs = [
4717 "bench/f32-rmax.cc",
4718 "src/xnnpack/AlignedAllocator.h",
4719 ] + MICROKERNEL_BENCHMARK_HDRS,
4720 deps = MICROKERNEL_BENCHMARK_DEPS,
4721)
4722
4723xnnpack_benchmark(
Marat Dukhan14bec502019-11-18 11:35:31 -08004724 name = "f32_sigmoid_bench",
4725 srcs = [
4726 "bench/f32-sigmoid.cc",
4727 "src/xnnpack/AlignedAllocator.h",
4728 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan14bec502019-11-18 11:35:31 -08004729 deps = MICROKERNEL_BENCHMARK_DEPS,
4730)
4731
4732xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004733 name = "f32_spmm_bench",
4734 srcs = [
4735 "bench/f32-spmm.cc",
4736 "bench/gemm.h",
4737 "src/xnnpack/AlignedAllocator.h",
4738 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07004739 deps = MICROKERNEL_BENCHMARK_DEPS,
4740)
4741
4742xnnpack_benchmark(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08004743 name = "f32_softmax_bench",
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07004744 srcs = [
Marat Dukhanfd8e6892020-01-27 15:25:25 -08004745 "bench/f32-softmax.cc",
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07004746 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004747 copts = xnnpack_optional_dnnl_copts(),
Marat Dukhan8d3c6932020-03-06 20:27:27 -08004748 deps = MICROKERNEL_BENCHMARK_DEPS + xnnpack_optional_dnnl_deps(),
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07004749)
4750
4751xnnpack_benchmark(
Marat Dukhan4c4eb002019-12-08 21:27:49 -08004752 name = "f32_vscaleexpminusmax_bench",
4753 srcs = [
4754 "bench/f32-vscaleexpminusmax.cc",
4755 "src/xnnpack/AlignedAllocator.h",
4756 ] + MICROKERNEL_BENCHMARK_HDRS,
4757 deps = MICROKERNEL_BENCHMARK_DEPS,
4758)
4759
4760xnnpack_benchmark(
4761 name = "f32_vscaleextexp_bench",
4762 srcs = [
4763 "bench/f32-vscaleextexp.cc",
4764 "src/xnnpack/AlignedAllocator.h",
4765 ] + MICROKERNEL_BENCHMARK_HDRS,
4766 deps = MICROKERNEL_BENCHMARK_DEPS,
4767)
4768
4769xnnpack_benchmark(
Marat Dukhanf4db2f32020-06-30 10:55:30 -07004770 name = "f32_vsqrt_bench",
4771 srcs = [
4772 "bench/f32-vsqrt.cc",
4773 "src/xnnpack/AlignedAllocator.h",
4774 ] + MICROKERNEL_BENCHMARK_HDRS,
4775 deps = MICROKERNEL_BENCHMARK_DEPS,
4776)
4777
4778xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004779 name = "f32_im2col_gemm_bench",
4780 srcs = [
4781 "bench/f32-im2col-gemm.cc",
4782 "bench/conv.h",
4783 "src/xnnpack/AlignedAllocator.h",
4784 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004785 deps = MICROKERNEL_BENCHMARK_DEPS + [
4786 ":im2col",
4787 ":packing",
4788 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004789)
4790
Marat Dukhanfe7acb62020-03-09 19:30:05 -07004791xnnpack_benchmark(
Marat Dukhanffbf96a2020-05-14 02:59:08 -07004792 name = "rounding_bench",
4793 srcs = [
4794 "bench/rounding.cc",
Marat Dukhanffbf96a2020-05-14 02:59:08 -07004795 "src/xnnpack/AlignedAllocator.h",
Frank Barchard04336c12020-10-22 16:48:55 -07004796 "src/xnnpack/math-stubs.h",
Marat Dukhanffbf96a2020-05-14 02:59:08 -07004797 ] + MICROKERNEL_BENCHMARK_HDRS,
4798 deps = MICROKERNEL_BENCHMARK_DEPS,
4799)
4800
Marat Dukhan08c4a432019-10-03 09:29:21 -07004801########################### Benchmarks for operators ###########################
4802
4803xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004804 name = "average_pooling_bench",
4805 srcs = ["bench/average-pooling.cc"],
Marat Dukhan7a16d8b2020-03-11 04:22:44 -07004806 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07004807 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004808 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004809)
4810
4811xnnpack_benchmark(
Marat Dukhan87727142020-06-24 15:24:10 -07004812 name = "bankers_rounding_bench",
4813 srcs = ["bench/bankers-rounding.cc"],
4814 copts = xnnpack_optional_tflite_copts(),
4815 tags = ["nowin32"],
4816 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
4817)
4818
4819xnnpack_benchmark(
4820 name = "ceiling_bench",
4821 srcs = ["bench/ceiling.cc"],
4822 copts = xnnpack_optional_tflite_copts(),
4823 tags = ["nowin32"],
4824 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
4825)
4826
4827xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004828 name = "channel_shuffle_bench",
4829 srcs = ["bench/channel-shuffle.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004830 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07004831)
4832
4833xnnpack_benchmark(
4834 name = "convolution_bench",
4835 srcs = ["bench/convolution.cc"],
4836 copts = xnnpack_optional_tflite_copts() + xnnpack_optional_armcl_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07004837 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004838 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps() + xnnpack_optional_armcl_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004839)
4840
4841xnnpack_benchmark(
4842 name = "deconvolution_bench",
4843 srcs = ["bench/deconvolution.cc"],
4844 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07004845 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004846 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004847)
4848
4849xnnpack_benchmark(
Marat Dukhan87727142020-06-24 15:24:10 -07004850 name = "floor_bench",
4851 srcs = ["bench/floor.cc"],
4852 copts = xnnpack_optional_tflite_copts(),
4853 tags = ["nowin32"],
4854 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
4855)
4856
4857xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004858 name = "global_average_pooling_bench",
4859 srcs = ["bench/global-average-pooling.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004860 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07004861)
4862
4863xnnpack_benchmark(
Marat Dukhanad352602020-06-25 21:50:54 -07004864 name = "hardswish_bench",
4865 srcs = ["bench/hardswish.cc"],
4866 copts = xnnpack_optional_tflite_copts(),
4867 tags = ["nowin32"],
4868 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
4869)
4870
4871xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004872 name = "max_pooling_bench",
4873 srcs = ["bench/max-pooling.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004874 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07004875)
4876
4877xnnpack_benchmark(
4878 name = "sigmoid_bench",
4879 srcs = ["bench/sigmoid.cc"],
Marat Dukhanc3b9e862019-11-17 13:18:54 -08004880 copts = xnnpack_optional_tflite_copts(),
Marat Dukhanca2ba702020-04-24 01:31:47 -07004881 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004882 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004883)
4884
4885xnnpack_benchmark(
Marat Dukhan95b22432019-10-30 16:30:14 -07004886 name = "prelu_bench",
4887 srcs = ["bench/prelu.cc"],
4888 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07004889 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004890 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan95b22432019-10-30 16:30:14 -07004891)
4892
4893xnnpack_benchmark(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08004894 name = "softmax_bench",
4895 srcs = ["bench/softmax.cc"],
Marat Dukhan9c0db962020-01-28 12:30:14 -08004896 copts = xnnpack_optional_tflite_copts(),
Marat Dukhanca2ba702020-04-24 01:31:47 -07004897 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004898 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004899)
4900
Marat Dukhan87727142020-06-24 15:24:10 -07004901xnnpack_benchmark(
Marat Dukhan6804bbd2020-06-30 19:26:11 -07004902 name = "square_root_bench",
4903 srcs = ["bench/square-root.cc"],
4904 copts = xnnpack_optional_tflite_copts(),
4905 tags = ["nowin32"],
4906 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
4907)
4908
4909xnnpack_benchmark(
Marat Dukhan87727142020-06-24 15:24:10 -07004910 name = "truncation_bench",
4911 srcs = ["bench/truncation.cc"],
4912 deps = OPERATOR_BENCHMARK_DEPS,
4913)
4914
Marat Dukhanc068bb62019-10-04 13:24:39 -07004915############################# End-to-end benchmarks ############################
4916
4917cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07004918 name = "fp32_mobilenet_v1",
4919 srcs = ["models/fp32-mobilenet-v1.cc"],
Marat Dukhanc068bb62019-10-04 13:24:39 -07004920 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08004921 copts = xnnpack_std_cxxopts(),
Marat Dukhanc068bb62019-10-04 13:24:39 -07004922 linkstatic = True,
4923 deps = [
4924 ":XNNPACK",
4925 "@pthreadpool",
4926 ],
4927)
4928
4929cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07004930 name = "fp16_mobilenet_v1",
4931 srcs = ["models/fp16-mobilenet-v1.cc"],
4932 hdrs = ["models/models.h"],
4933 copts = xnnpack_std_cxxopts(),
4934 linkstatic = True,
4935 deps = [
4936 ":XNNPACK",
4937 "@FP16",
4938 "@pthreadpool",
4939 ],
4940)
4941
4942cc_library(
Marat Dukhan0743cdf2020-08-04 18:52:07 -07004943 name = "qs8_mobilenet_v1",
4944 srcs = ["models/qs8-mobilenet-v1.cc"],
4945 hdrs = ["models/models.h"],
4946 copts = xnnpack_std_cxxopts(),
4947 linkstatic = True,
4948 deps = [
4949 ":XNNPACK",
4950 "@pthreadpool",
4951 ],
4952)
4953
4954cc_library(
Marat Dukhan70a96182020-09-03 17:13:58 -07004955 name = "qs8_mobilenet_v2",
4956 srcs = ["models/qs8-mobilenet-v2.cc"],
4957 hdrs = ["models/models.h"],
4958 copts = xnnpack_std_cxxopts(),
4959 linkstatic = True,
4960 deps = [
4961 ":XNNPACK",
4962 "@pthreadpool",
4963 ],
4964)
4965
4966cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07004967 name = "fp32_mobilenet_v2",
4968 srcs = ["models/fp32-mobilenet-v2.cc"],
Marat Dukhanc068bb62019-10-04 13:24:39 -07004969 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08004970 copts = xnnpack_std_cxxopts(),
Marat Dukhanc068bb62019-10-04 13:24:39 -07004971 linkstatic = True,
4972 deps = [
4973 ":XNNPACK",
4974 "@pthreadpool",
4975 ],
4976)
4977
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004978cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07004979 name = "fp16_mobilenet_v2",
4980 srcs = ["models/fp16-mobilenet-v2.cc"],
4981 hdrs = ["models/models.h"],
4982 copts = xnnpack_std_cxxopts(),
4983 linkstatic = True,
4984 deps = [
4985 ":XNNPACK",
4986 "@FP16",
4987 "@pthreadpool",
4988 ],
4989)
4990
4991cc_library(
4992 name = "fp32_mobilenet_v3_large",
4993 srcs = ["models/fp32-mobilenet-v3-large.cc"],
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004994 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08004995 copts = xnnpack_std_cxxopts(),
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004996 linkstatic = True,
4997 deps = [
4998 ":XNNPACK",
4999 "@pthreadpool",
5000 ],
5001)
5002
5003cc_library(
Marat Dukhan66f3ccd2020-09-14 16:09:38 -07005004 name = "fp16_mobilenet_v3_large",
5005 srcs = ["models/fp16-mobilenet-v3-large.cc"],
5006 hdrs = ["models/models.h"],
5007 copts = xnnpack_std_cxxopts(),
5008 linkstatic = True,
5009 deps = [
5010 ":XNNPACK",
5011 "@FP16",
5012 "@pthreadpool",
5013 ],
5014)
5015
5016cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07005017 name = "fp32_mobilenet_v3_small",
5018 srcs = ["models/fp32-mobilenet-v3-small.cc"],
Marat Dukhanc08cdf52019-12-09 09:17:51 -08005019 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08005020 copts = xnnpack_std_cxxopts(),
Marat Dukhanc08cdf52019-12-09 09:17:51 -08005021 linkstatic = True,
5022 deps = [
5023 ":XNNPACK",
5024 "@pthreadpool",
5025 ],
5026)
5027
Marat Dukhan66f3ccd2020-09-14 16:09:38 -07005028cc_library(
5029 name = "fp16_mobilenet_v3_small",
5030 srcs = ["models/fp16-mobilenet-v3-small.cc"],
5031 hdrs = ["models/models.h"],
5032 copts = xnnpack_std_cxxopts(),
5033 linkstatic = True,
5034 deps = [
5035 ":XNNPACK",
5036 "@FP16",
5037 "@pthreadpool",
5038 ],
5039)
5040
Marat Dukhanc068bb62019-10-04 13:24:39 -07005041xnnpack_benchmark(
Marat Dukhanef4416e2019-10-31 13:44:40 -07005042 name = "f32_dwconv_e2e_bench",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08005043 srcs = [
5044 "bench/f32-dwconv-e2e.cc",
5045 "bench/end2end.h",
5046 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanef4416e2019-10-31 13:44:40 -07005047 deps = MICROKERNEL_BENCHMARK_DEPS + [
5048 ":XNNPACK",
Marat Dukhan270a2c42020-06-26 16:45:52 -07005049 ":fp32_mobilenet_v1",
5050 ":fp32_mobilenet_v2",
5051 ":fp32_mobilenet_v3_large",
5052 ":fp32_mobilenet_v3_small",
Marat Dukhanef4416e2019-10-31 13:44:40 -07005053 ],
5054)
5055
5056xnnpack_benchmark(
Marat Dukhan5f18d262019-10-31 10:24:14 -07005057 name = "f32_gemm_e2e_bench",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08005058 srcs = [
5059 "bench/f32-gemm-e2e.cc",
5060 "bench/end2end.h",
5061 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan5f18d262019-10-31 10:24:14 -07005062 deps = MICROKERNEL_BENCHMARK_DEPS + [
5063 ":XNNPACK",
Marat Dukhan270a2c42020-06-26 16:45:52 -07005064 ":fp32_mobilenet_v1",
5065 ":fp32_mobilenet_v2",
5066 ":fp32_mobilenet_v3_large",
5067 ":fp32_mobilenet_v3_small",
Marat Dukhan5f18d262019-10-31 10:24:14 -07005068 ],
5069)
5070
5071xnnpack_benchmark(
Marat Dukhanc068bb62019-10-04 13:24:39 -07005072 name = "end2end_bench",
5073 srcs = ["bench/end2end.cc"],
5074 deps = [
5075 ":XNNPACK",
Frank Barchardc712fa42019-10-31 14:00:21 -07005076 ":bench_utils",
Marat Dukhan270a2c42020-06-26 16:45:52 -07005077 ":fp16_mobilenet_v1",
5078 ":fp16_mobilenet_v2",
Marat Dukhan66f3ccd2020-09-14 16:09:38 -07005079 ":fp16_mobilenet_v3_large",
5080 ":fp16_mobilenet_v3_small",
Marat Dukhan270a2c42020-06-26 16:45:52 -07005081 ":fp32_mobilenet_v1",
5082 ":fp32_mobilenet_v2",
5083 ":fp32_mobilenet_v3_large",
5084 ":fp32_mobilenet_v3_small",
Marat Dukhan0743cdf2020-08-04 18:52:07 -07005085 ":qs8_mobilenet_v1",
Marat Dukhan70a96182020-09-03 17:13:58 -07005086 ":qs8_mobilenet_v2",
Marat Dukhanc068bb62019-10-04 13:24:39 -07005087 "@pthreadpool",
5088 ],
5089)
5090
Marat Dukhan6adff4e2019-10-14 18:32:07 -07005091#################### Accuracy evaluation for math functions ####################
5092
5093xnnpack_benchmark(
5094 name = "f32_exp_eval",
5095 srcs = [
5096 "eval/f32-exp.cc",
5097 "src/xnnpack/AlignedAllocator.h",
5098 ] + ACCURACY_EVAL_HDRS,
5099 deps = ACCURACY_EVAL_DEPS,
5100)
5101
Marat Dukhan515c9772019-10-17 18:07:57 -07005102xnnpack_benchmark(
5103 name = "f32_expminus_eval",
5104 srcs = [
5105 "eval/f32-expminus.cc",
5106 "src/xnnpack/AlignedAllocator.h",
5107 ] + ACCURACY_EVAL_HDRS,
5108 deps = ACCURACY_EVAL_DEPS,
5109)
5110
Marat Dukhan98ba4412019-10-23 02:14:28 -07005111xnnpack_benchmark(
Marat Dukhana438aca2020-11-20 15:45:01 -08005112 name = "f32_expm1minus_eval",
5113 srcs = [
5114 "eval/f32-expm1minus.cc",
5115 "src/xnnpack/AlignedAllocator.h",
5116 ] + ACCURACY_EVAL_HDRS,
Marat Dukhande390d42020-11-29 19:32:18 -08005117 deps = ACCURACY_EVAL_DEPS + [
5118 ":bench_utils",
5119 ],
Marat Dukhana438aca2020-11-20 15:45:01 -08005120)
5121
5122xnnpack_benchmark(
Marat Dukhan98ba4412019-10-23 02:14:28 -07005123 name = "f32_extexp_eval",
5124 srcs = [
5125 "eval/f32-extexp.cc",
5126 "src/xnnpack/AlignedAllocator.h",
5127 ] + ACCURACY_EVAL_HDRS,
5128 deps = ACCURACY_EVAL_DEPS,
5129)
5130
Marat Dukhan8853b822020-05-07 12:19:01 -07005131xnnpack_unit_test(
5132 name = "f32_roundne_eval",
5133 srcs = [
5134 "eval/f32-roundne.cc",
5135 "src/xnnpack/AlignedAllocator.h",
5136 "src/xnnpack/math-stubs.h",
5137 ] + MICROKERNEL_TEST_HDRS,
Marat Dukhan22eed3d2020-05-11 20:13:37 -07005138 automatic = False,
Marat Dukhan8853b822020-05-07 12:19:01 -07005139 deps = MICROKERNEL_TEST_DEPS,
5140)
5141
Marat Dukhan2dbb9442020-05-12 20:43:43 -07005142xnnpack_unit_test(
Marat Dukhanc9852ba2020-05-13 17:21:29 -07005143 name = "f32_roundd_eval",
5144 srcs = [
5145 "eval/f32-roundd.cc",
5146 "src/xnnpack/AlignedAllocator.h",
5147 "src/xnnpack/math-stubs.h",
5148 ] + MICROKERNEL_TEST_HDRS,
5149 automatic = False,
5150 deps = MICROKERNEL_TEST_DEPS,
5151)
5152
5153xnnpack_unit_test(
5154 name = "f32_roundu_eval",
5155 srcs = [
5156 "eval/f32-roundu.cc",
5157 "src/xnnpack/AlignedAllocator.h",
5158 "src/xnnpack/math-stubs.h",
5159 ] + MICROKERNEL_TEST_HDRS,
5160 automatic = False,
5161 deps = MICROKERNEL_TEST_DEPS,
5162)
5163
5164xnnpack_unit_test(
Marat Dukhan2dbb9442020-05-12 20:43:43 -07005165 name = "f32_roundz_eval",
5166 srcs = [
5167 "eval/f32-roundz.cc",
5168 "src/xnnpack/AlignedAllocator.h",
5169 "src/xnnpack/math-stubs.h",
5170 ] + MICROKERNEL_TEST_HDRS,
Marat Dukhanc9852ba2020-05-13 17:21:29 -07005171 automatic = False,
Marat Dukhan2dbb9442020-05-12 20:43:43 -07005172 deps = MICROKERNEL_TEST_DEPS,
5173)
5174
Marat Dukhan346a9e52019-11-15 09:06:30 -08005175xnnpack_benchmark(
5176 name = "f32_sigmoid_eval",
5177 srcs = [
5178 "eval/f32-sigmoid.cc",
5179 "src/xnnpack/AlignedAllocator.h",
5180 ] + ACCURACY_EVAL_HDRS,
5181 deps = ACCURACY_EVAL_DEPS,
5182)
5183
Marat Dukhan84000762020-06-29 18:38:43 -07005184xnnpack_benchmark(
5185 name = "f32_sqrt_eval",
5186 srcs = [
5187 "eval/f32-sqrt.cc",
5188 "src/xnnpack/AlignedAllocator.h",
5189 ] + ACCURACY_EVAL_HDRS,
5190 deps = ACCURACY_EVAL_DEPS + [
5191 ":bench_utils",
5192 ],
5193)
5194
Marat Dukhan08c4a432019-10-03 09:29:21 -07005195######################### Unit tests for micro-kernels #########################
5196
5197xnnpack_unit_test(
Marat Dukhande06f492020-04-09 00:19:31 -07005198 name = "f16_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005199 srcs = [
Marat Dukhande06f492020-04-09 00:19:31 -07005200 "test/f16-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005201 "test/gemm-microkernel-tester.h",
5202 "src/xnnpack/AlignedAllocator.h",
5203 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005204 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005205)
5206
5207xnnpack_unit_test(
Marat Dukhan355ab432020-04-09 19:01:52 -07005208 name = "f16_spmm_minmax_test",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08005209 srcs = [
Marat Dukhan355ab432020-04-09 19:01:52 -07005210 "test/f16-spmm-minmax.cc",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08005211 "test/spmm-microkernel-tester.h",
5212 "src/xnnpack/AlignedAllocator.h",
5213 ] + MICROKERNEL_TEST_HDRS,
5214 deps = MICROKERNEL_TEST_DEPS,
5215)
5216
5217xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005218 name = "f32_argmaxpool_test",
5219 srcs = [
5220 "test/f32-argmaxpool.cc",
5221 "test/argmaxpool-microkernel-tester.h",
5222 "src/xnnpack/AlignedAllocator.h",
5223 ] + MICROKERNEL_TEST_HDRS,
5224 deps = MICROKERNEL_TEST_DEPS,
5225)
5226
5227xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07005228 name = "f32_avgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005229 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07005230 "test/f32-avgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005231 "test/avgpool-microkernel-tester.h",
5232 "src/xnnpack/AlignedAllocator.h",
5233 ] + MICROKERNEL_TEST_HDRS,
5234 deps = MICROKERNEL_TEST_DEPS,
5235)
5236
5237xnnpack_unit_test(
Marat Dukhan660fd192020-03-10 04:55:30 -07005238 name = "f32_ibilinear_test",
Marat Dukhan35dacfb2019-11-07 19:18:16 -08005239 srcs = [
Marat Dukhan660fd192020-03-10 04:55:30 -07005240 "test/f32-ibilinear.cc",
5241 "test/ibilinear-microkernel-tester.h",
Marat Dukhan35dacfb2019-11-07 19:18:16 -08005242 "src/xnnpack/AlignedAllocator.h",
5243 ] + MICROKERNEL_TEST_HDRS,
5244 deps = MICROKERNEL_TEST_DEPS,
5245)
5246
5247xnnpack_unit_test(
XNNPACK Team21432672020-10-19 19:58:48 -07005248 name = "f32_ibilinear_chw_test",
5249 srcs = [
5250 "test/f32-ibilinear-chw.cc",
XNNPACK Team6be46b22020-10-22 23:34:54 -07005251 "test/ibilinear-microkernel-tester.h",
XNNPACK Team21432672020-10-19 19:58:48 -07005252 "src/xnnpack/AlignedAllocator.h",
5253 ] + MICROKERNEL_TEST_HDRS,
5254 deps = MICROKERNEL_TEST_DEPS,
5255)
5256
5257xnnpack_unit_test(
Frank Barchardb1966592020-05-12 13:47:06 -07005258 name = "f16_clamp_test",
5259 srcs = [
5260 "test/f16-clamp.cc",
5261 "test/clamp-microkernel-tester.h",
5262 ] + MICROKERNEL_TEST_HDRS,
5263 deps = MICROKERNEL_TEST_DEPS,
5264)
5265
5266xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005267 name = "f32_clamp_test",
5268 srcs = [
5269 "test/f32-clamp.cc",
5270 "test/clamp-microkernel-tester.h",
5271 ] + MICROKERNEL_TEST_HDRS,
5272 deps = MICROKERNEL_TEST_DEPS,
5273)
5274
5275xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07005276 name = "f32_igemm_test",
5277 srcs = [
5278 "test/f32-igemm.cc",
5279 "test/gemm-microkernel-tester.h",
5280 "src/xnnpack/AlignedAllocator.h",
5281 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005282 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan163a7e62020-04-09 04:19:26 -07005283)
5284
5285xnnpack_unit_test(
Marat Dukhan467f6362020-05-22 23:21:55 -07005286 name = "f32_igemm_relu_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005287 srcs = [
Marat Dukhan467f6362020-05-22 23:21:55 -07005288 "test/f32-igemm-relu.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005289 "test/gemm-microkernel-tester.h",
5290 "src/xnnpack/AlignedAllocator.h",
5291 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005292 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005293)
5294
5295xnnpack_unit_test(
Marat Dukhane207b7b2020-05-28 16:27:42 -07005296 name = "f32_igemm_minmax_test",
5297 srcs = [
5298 "test/f32-igemm-minmax.cc",
5299 "test/gemm-microkernel-tester.h",
5300 "src/xnnpack/AlignedAllocator.h",
5301 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005302 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhane207b7b2020-05-28 16:27:42 -07005303)
5304
5305xnnpack_unit_test(
Frank Barchardb0e4fae2020-05-04 15:27:51 -07005306 name = "f16_igemm_minmax_test",
5307 srcs = [
5308 "test/f16-igemm-minmax.cc",
5309 "test/gemm-microkernel-tester.h",
5310 "src/xnnpack/AlignedAllocator.h",
5311 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005312 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Frank Barchardb0e4fae2020-05-04 15:27:51 -07005313)
5314
5315xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005316 name = "f32_conv_hwc_test",
5317 srcs = [
5318 "test/f32-conv-hwc.cc",
5319 "test/conv-hwc-microkernel-tester.h",
5320 "src/xnnpack/AlignedAllocator.h",
5321 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005322 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005323)
5324
5325xnnpack_unit_test(
Marat Dukhan1f29b802020-05-15 23:46:39 -07005326 name = "f32_conv_hwc2chw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005327 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07005328 "test/f32-conv-hwc2chw.cc",
5329 "test/conv-hwc2chw-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005330 "src/xnnpack/AlignedAllocator.h",
5331 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005332 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005333)
5334
5335xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07005336 name = "f32_dwconv_test",
5337 srcs = [
5338 "test/f32-dwconv.cc",
5339 "test/dwconv-microkernel-tester.h",
5340 "src/xnnpack/AlignedAllocator.h",
5341 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005342 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan163a7e62020-04-09 04:19:26 -07005343)
5344
5345xnnpack_unit_test(
Frank Barchard5a599a62020-06-04 20:12:44 -07005346 name = "f16_dwconv_minmax_test",
5347 srcs = [
5348 "test/f16-dwconv-minmax.cc",
5349 "test/dwconv-microkernel-tester.h",
5350 "src/xnnpack/AlignedAllocator.h",
5351 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005352 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Frank Barchard5a599a62020-06-04 20:12:44 -07005353)
5354
5355xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07005356 name = "f32_dwconv_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005357 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07005358 "test/f32-dwconv-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005359 "test/dwconv-microkernel-tester.h",
5360 "src/xnnpack/AlignedAllocator.h",
5361 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005362 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005363)
5364
5365xnnpack_unit_test(
Marat Dukhanbf715f92020-10-23 20:17:00 -07005366 name = "f32_dwconv2d_chw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005367 srcs = [
Marat Dukhanbf715f92020-10-23 20:17:00 -07005368 "test/f32-dwconv2d-chw.cc",
5369 "test/dwconv2d-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005370 "src/xnnpack/AlignedAllocator.h",
5371 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005372 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005373)
5374
5375xnnpack_unit_test(
Frank Barchard0bb49a72020-06-04 11:35:11 -07005376 name = "f16_gavgpool_minmax_test",
5377 srcs = [
5378 "test/f16-gavgpool-minmax.cc",
5379 "test/gavgpool-microkernel-tester.h",
5380 "src/xnnpack/AlignedAllocator.h",
5381 ] + MICROKERNEL_TEST_HDRS,
5382 deps = MICROKERNEL_TEST_DEPS,
5383)
5384
5385xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07005386 name = "f32_gavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005387 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07005388 "test/f32-gavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005389 "test/gavgpool-microkernel-tester.h",
5390 "src/xnnpack/AlignedAllocator.h",
5391 ] + MICROKERNEL_TEST_HDRS,
5392 deps = MICROKERNEL_TEST_DEPS,
5393)
5394
5395xnnpack_unit_test(
Marat Dukhan1f29b802020-05-15 23:46:39 -07005396 name = "f32_gavgpool_cw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005397 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07005398 "test/f32-gavgpool-cw.cc",
5399 "test/gavgpool-cw-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005400 "src/xnnpack/AlignedAllocator.h",
5401 ] + MICROKERNEL_TEST_HDRS,
5402 deps = MICROKERNEL_TEST_DEPS,
5403)
5404
5405xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07005406 name = "f32_gemm_test",
5407 srcs = [
5408 "test/f32-gemm.cc",
5409 "test/gemm-microkernel-tester.h",
5410 "src/xnnpack/AlignedAllocator.h",
5411 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005412 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan163a7e62020-04-09 04:19:26 -07005413)
5414
5415xnnpack_unit_test(
Marat Dukhan467f6362020-05-22 23:21:55 -07005416 name = "f32_gemm_relu_test",
5417 srcs = [
5418 "test/f32-gemm-relu.cc",
5419 "test/gemm-microkernel-tester.h",
5420 "src/xnnpack/AlignedAllocator.h",
5421 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005422 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan467f6362020-05-22 23:21:55 -07005423)
5424
5425xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07005426 name = "f32_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005427 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07005428 "test/f32-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005429 "test/gemm-microkernel-tester.h",
5430 "src/xnnpack/AlignedAllocator.h",
5431 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005432 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005433)
5434
5435xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07005436 name = "f32_gemminc_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005437 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07005438 "test/f32-gemminc-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005439 "test/gemm-microkernel-tester.h",
5440 "src/xnnpack/AlignedAllocator.h",
5441 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005442 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005443)
5444
5445xnnpack_unit_test(
Frank Barchardb1966592020-05-12 13:47:06 -07005446 name = "f16_hswish_test",
5447 srcs = [
5448 "test/f16-hswish.cc",
5449 "test/hswish-microkernel-tester.h",
5450 ] + MICROKERNEL_TEST_HDRS,
5451 deps = MICROKERNEL_TEST_DEPS,
5452)
5453
5454xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005455 name = "f32_hswish_test",
5456 srcs = [
5457 "test/f32-hswish.cc",
5458 "test/hswish-microkernel-tester.h",
5459 ] + MICROKERNEL_TEST_HDRS,
5460 deps = MICROKERNEL_TEST_DEPS,
5461)
5462
5463xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07005464 name = "f32_maxpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005465 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07005466 "test/f32-maxpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005467 "test/maxpool-microkernel-tester.h",
5468 ] + MICROKERNEL_TEST_HDRS,
5469 deps = MICROKERNEL_TEST_DEPS,
5470)
5471
5472xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07005473 name = "f32_pavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005474 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07005475 "test/f32-pavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005476 "test/avgpool-microkernel-tester.h",
5477 "src/xnnpack/AlignedAllocator.h",
5478 ] + MICROKERNEL_TEST_HDRS,
5479 deps = MICROKERNEL_TEST_DEPS,
5480)
5481
5482xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07005483 name = "f32_ppmm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005484 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07005485 "test/f32-ppmm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005486 "test/gemm-microkernel-tester.h",
5487 "src/xnnpack/AlignedAllocator.h",
5488 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005489 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005490)
5491
5492xnnpack_unit_test(
Frank Barchardb1966592020-05-12 13:47:06 -07005493 name = "f16_prelu_test",
5494 srcs = [
5495 "test/f16-prelu.cc",
5496 "test/prelu-microkernel-tester.h",
5497 "src/xnnpack/AlignedAllocator.h",
5498 ] + MICROKERNEL_TEST_HDRS,
5499 deps = MICROKERNEL_TEST_DEPS,
5500)
5501
5502xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005503 name = "f32_prelu_test",
5504 srcs = [
5505 "test/f32-prelu.cc",
5506 "test/prelu-microkernel-tester.h",
5507 "src/xnnpack/AlignedAllocator.h",
5508 ] + MICROKERNEL_TEST_HDRS,
5509 deps = MICROKERNEL_TEST_DEPS,
5510)
5511
5512xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07005513 name = "f32_raddexpminusmax_test",
5514 srcs = [
5515 "test/f32-raddexpminusmax.cc",
5516 "test/raddexpminusmax-microkernel-tester.h",
5517 ] + MICROKERNEL_TEST_HDRS,
5518 deps = MICROKERNEL_TEST_DEPS,
5519)
5520
5521xnnpack_unit_test(
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07005522 name = "f32_raddextexp_test",
5523 srcs = [
5524 "test/f32-raddextexp.cc",
5525 "test/raddextexp-microkernel-tester.h",
5526 ] + MICROKERNEL_TEST_HDRS,
5527 deps = MICROKERNEL_TEST_DEPS,
5528)
5529
5530xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07005531 name = "f32_raddstoreexpminusmax_test",
5532 srcs = [
5533 "test/f32-raddstoreexpminusmax.cc",
5534 "test/raddstoreexpminusmax-microkernel-tester.h",
5535 ] + MICROKERNEL_TEST_HDRS,
5536 deps = MICROKERNEL_TEST_DEPS,
5537)
5538
5539xnnpack_unit_test(
Frank Barchardfb158e22020-07-15 16:10:10 -07005540 name = "f32_relu_test",
5541 srcs = [
5542 "test/f32-relu.cc",
5543 "test/vunary-microkernel-tester.h",
5544 ] + MICROKERNEL_TEST_HDRS,
5545 deps = MICROKERNEL_TEST_DEPS,
5546)
5547
5548xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005549 name = "f32_rmax_test",
5550 srcs = [
5551 "test/f32-rmax.cc",
5552 "test/rmax-microkernel-tester.h",
5553 ] + MICROKERNEL_TEST_HDRS,
5554 deps = MICROKERNEL_TEST_DEPS,
5555)
5556
5557xnnpack_unit_test(
Marat Dukhan346a9e52019-11-15 09:06:30 -08005558 name = "f32_sigmoid_test",
5559 srcs = [
5560 "test/f32-sigmoid.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005561 "test/vunary-microkernel-tester.h",
Marat Dukhan346a9e52019-11-15 09:06:30 -08005562 ] + MICROKERNEL_TEST_HDRS,
5563 deps = MICROKERNEL_TEST_DEPS,
5564)
5565
5566xnnpack_unit_test(
Marat Dukhan355ab432020-04-09 19:01:52 -07005567 name = "f32_spmm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005568 srcs = [
Marat Dukhan355ab432020-04-09 19:01:52 -07005569 "test/f32-spmm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005570 "test/spmm-microkernel-tester.h",
5571 "src/xnnpack/AlignedAllocator.h",
5572 ] + MICROKERNEL_TEST_HDRS,
5573 deps = MICROKERNEL_TEST_DEPS,
5574)
5575
5576xnnpack_unit_test(
Frank Barchardd793f6c2020-05-08 13:37:43 -07005577 name = "f16_vadd_minmax_test",
5578 srcs = [
5579 "test/f16-vadd-minmax.cc",
5580 "test/vbinary-microkernel-tester.h",
5581 ] + MICROKERNEL_TEST_HDRS,
5582 deps = MICROKERNEL_TEST_DEPS,
5583)
5584
5585xnnpack_unit_test(
5586 name = "f16_vaddc_minmax_test",
5587 srcs = [
5588 "test/f16-vaddc-minmax.cc",
5589 "test/vbinaryc-microkernel-tester.h",
5590 ] + MICROKERNEL_TEST_HDRS,
5591 deps = MICROKERNEL_TEST_DEPS,
5592)
5593
5594xnnpack_unit_test(
5595 name = "f16_vdiv_minmax_test",
5596 srcs = [
5597 "test/f16-vdiv-minmax.cc",
5598 "test/vbinary-microkernel-tester.h",
5599 ] + MICROKERNEL_TEST_HDRS,
5600 deps = MICROKERNEL_TEST_DEPS,
5601)
5602
5603xnnpack_unit_test(
5604 name = "f16_vdivc_minmax_test",
5605 srcs = [
5606 "test/f16-vdivc-minmax.cc",
5607 "test/vbinaryc-microkernel-tester.h",
5608 ] + MICROKERNEL_TEST_HDRS,
5609 deps = MICROKERNEL_TEST_DEPS,
5610)
5611
5612xnnpack_unit_test(
5613 name = "f16_vrdivc_minmax_test",
5614 srcs = [
5615 "test/f16-vrdivc-minmax.cc",
5616 "test/vbinaryc-microkernel-tester.h",
5617 ] + MICROKERNEL_TEST_HDRS,
5618 deps = MICROKERNEL_TEST_DEPS,
5619)
5620
5621xnnpack_unit_test(
5622 name = "f16_vmax_test",
5623 srcs = [
5624 "test/f16-vmax.cc",
5625 "test/vbinary-microkernel-tester.h",
5626 ] + MICROKERNEL_TEST_HDRS,
5627 deps = MICROKERNEL_TEST_DEPS,
5628)
5629
5630xnnpack_unit_test(
5631 name = "f16_vmaxc_test",
5632 srcs = [
5633 "test/f16-vmaxc.cc",
5634 "test/vbinaryc-microkernel-tester.h",
5635 ] + MICROKERNEL_TEST_HDRS,
5636 deps = MICROKERNEL_TEST_DEPS,
5637)
5638
5639xnnpack_unit_test(
5640 name = "f16_vmin_test",
5641 srcs = [
5642 "test/f16-vmin.cc",
5643 "test/vbinary-microkernel-tester.h",
5644 ] + MICROKERNEL_TEST_HDRS,
5645 deps = MICROKERNEL_TEST_DEPS,
5646)
5647
5648xnnpack_unit_test(
5649 name = "f16_vminc_test",
5650 srcs = [
5651 "test/f16-vminc.cc",
5652 "test/vbinaryc-microkernel-tester.h",
5653 ] + MICROKERNEL_TEST_HDRS,
5654 deps = MICROKERNEL_TEST_DEPS,
5655)
5656
5657xnnpack_unit_test(
5658 name = "f16_vmul_minmax_test",
5659 srcs = [
5660 "test/f16-vmul-minmax.cc",
5661 "test/vbinary-microkernel-tester.h",
5662 ] + MICROKERNEL_TEST_HDRS,
5663 deps = MICROKERNEL_TEST_DEPS,
5664)
5665
5666xnnpack_unit_test(
5667 name = "f16_vmulc_minmax_test",
5668 srcs = [
5669 "test/f16-vmulc-minmax.cc",
5670 "test/vbinaryc-microkernel-tester.h",
5671 ] + MICROKERNEL_TEST_HDRS,
5672 deps = MICROKERNEL_TEST_DEPS,
5673)
5674
5675xnnpack_unit_test(
5676 name = "f16_vsub_minmax_test",
5677 srcs = [
5678 "test/f16-vsub-minmax.cc",
5679 "test/vbinary-microkernel-tester.h",
5680 ] + MICROKERNEL_TEST_HDRS,
5681 deps = MICROKERNEL_TEST_DEPS,
5682)
5683
5684xnnpack_unit_test(
5685 name = "f16_vsubc_minmax_test",
5686 srcs = [
5687 "test/f16-vsubc-minmax.cc",
5688 "test/vbinaryc-microkernel-tester.h",
5689 ] + MICROKERNEL_TEST_HDRS,
5690 deps = MICROKERNEL_TEST_DEPS,
5691)
5692
5693xnnpack_unit_test(
5694 name = "f16_vrsubc_minmax_test",
5695 srcs = [
5696 "test/f16-vrsubc-minmax.cc",
5697 "test/vbinaryc-microkernel-tester.h",
5698 ] + MICROKERNEL_TEST_HDRS,
5699 deps = MICROKERNEL_TEST_DEPS,
5700)
5701
5702xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07005703 name = "f32_vabs_test",
5704 srcs = [
5705 "test/f32-vabs.cc",
5706 "test/vunary-microkernel-tester.h",
5707 ] + MICROKERNEL_TEST_HDRS,
5708 deps = MICROKERNEL_TEST_DEPS,
5709)
5710
5711xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005712 name = "f32_vadd_test",
5713 srcs = [
5714 "test/f32-vadd.cc",
5715 "test/vbinary-microkernel-tester.h",
5716 ] + MICROKERNEL_TEST_HDRS,
5717 deps = MICROKERNEL_TEST_DEPS,
5718)
5719
5720xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005721 name = "f32_vadd_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005722 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005723 "test/f32-vadd-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005724 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005725 ] + MICROKERNEL_TEST_HDRS,
5726 deps = MICROKERNEL_TEST_DEPS,
5727)
5728
5729xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005730 name = "f32_vadd_relu_test",
5731 srcs = [
5732 "test/f32-vadd-relu.cc",
5733 "test/vbinary-microkernel-tester.h",
5734 ] + MICROKERNEL_TEST_HDRS,
5735 deps = MICROKERNEL_TEST_DEPS,
5736)
5737
5738xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005739 name = "f32_vaddc_test",
5740 srcs = [
5741 "test/f32-vaddc.cc",
5742 "test/vbinaryc-microkernel-tester.h",
5743 ] + MICROKERNEL_TEST_HDRS,
5744 deps = MICROKERNEL_TEST_DEPS,
5745)
5746
5747xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005748 name = "f32_vaddc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005749 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005750 "test/f32-vaddc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005751 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005752 ] + MICROKERNEL_TEST_HDRS,
5753 deps = MICROKERNEL_TEST_DEPS,
5754)
5755
5756xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005757 name = "f32_vaddc_relu_test",
5758 srcs = [
5759 "test/f32-vaddc-relu.cc",
5760 "test/vbinaryc-microkernel-tester.h",
5761 ] + MICROKERNEL_TEST_HDRS,
5762 deps = MICROKERNEL_TEST_DEPS,
5763)
5764
5765xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005766 name = "f32_vdiv_test",
5767 srcs = [
5768 "test/f32-vdiv.cc",
5769 "test/vbinary-microkernel-tester.h",
5770 ] + MICROKERNEL_TEST_HDRS,
5771 deps = MICROKERNEL_TEST_DEPS,
5772)
5773
5774xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005775 name = "f32_vdiv_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08005776 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005777 "test/f32-vdiv-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08005778 "test/vbinary-microkernel-tester.h",
5779 ] + MICROKERNEL_TEST_HDRS,
5780 deps = MICROKERNEL_TEST_DEPS,
5781)
5782
5783xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005784 name = "f32_vdiv_relu_test",
5785 srcs = [
5786 "test/f32-vdiv-relu.cc",
5787 "test/vbinary-microkernel-tester.h",
5788 ] + MICROKERNEL_TEST_HDRS,
5789 deps = MICROKERNEL_TEST_DEPS,
5790)
5791
5792xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005793 name = "f32_vdivc_test",
5794 srcs = [
5795 "test/f32-vdivc.cc",
5796 "test/vbinaryc-microkernel-tester.h",
5797 ] + MICROKERNEL_TEST_HDRS,
5798 deps = MICROKERNEL_TEST_DEPS,
5799)
5800
5801xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005802 name = "f32_vdivc_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08005803 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005804 "test/f32-vdivc-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08005805 "test/vbinaryc-microkernel-tester.h",
5806 ] + MICROKERNEL_TEST_HDRS,
5807 deps = MICROKERNEL_TEST_DEPS,
5808)
5809
5810xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005811 name = "f32_vdivc_relu_test",
5812 srcs = [
5813 "test/f32-vdivc-relu.cc",
5814 "test/vbinaryc-microkernel-tester.h",
5815 ] + MICROKERNEL_TEST_HDRS,
5816 deps = MICROKERNEL_TEST_DEPS,
5817)
5818
5819xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005820 name = "f32_vrdivc_test",
5821 srcs = [
5822 "test/f32-vrdivc.cc",
5823 "test/vbinaryc-microkernel-tester.h",
5824 ] + MICROKERNEL_TEST_HDRS,
5825 deps = MICROKERNEL_TEST_DEPS,
5826)
5827
5828xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005829 name = "f32_vrdivc_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08005830 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005831 "test/f32-vrdivc-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08005832 "test/vbinaryc-microkernel-tester.h",
5833 ] + MICROKERNEL_TEST_HDRS,
5834 deps = MICROKERNEL_TEST_DEPS,
5835)
5836
5837xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005838 name = "f32_vrdivc_relu_test",
5839 srcs = [
5840 "test/f32-vrdivc-relu.cc",
5841 "test/vbinaryc-microkernel-tester.h",
5842 ] + MICROKERNEL_TEST_HDRS,
5843 deps = MICROKERNEL_TEST_DEPS,
5844)
5845
5846xnnpack_unit_test(
Marat Dukhan403b7d42019-12-05 12:49:11 -08005847 name = "f32_vmax_test",
5848 srcs = [
5849 "test/f32-vmax.cc",
5850 "test/vbinary-microkernel-tester.h",
5851 ] + MICROKERNEL_TEST_HDRS,
5852 deps = MICROKERNEL_TEST_DEPS,
5853)
5854
5855xnnpack_unit_test(
5856 name = "f32_vmaxc_test",
5857 srcs = [
5858 "test/f32-vmaxc.cc",
5859 "test/vbinaryc-microkernel-tester.h",
5860 ] + MICROKERNEL_TEST_HDRS,
5861 deps = MICROKERNEL_TEST_DEPS,
5862)
5863
5864xnnpack_unit_test(
5865 name = "f32_vmin_test",
5866 srcs = [
5867 "test/f32-vmin.cc",
5868 "test/vbinary-microkernel-tester.h",
5869 ] + MICROKERNEL_TEST_HDRS,
5870 deps = MICROKERNEL_TEST_DEPS,
5871)
5872
5873xnnpack_unit_test(
5874 name = "f32_vminc_test",
5875 srcs = [
5876 "test/f32-vminc.cc",
5877 "test/vbinaryc-microkernel-tester.h",
5878 ] + MICROKERNEL_TEST_HDRS,
5879 deps = MICROKERNEL_TEST_DEPS,
5880)
5881
5882xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005883 name = "f32_vmul_test",
5884 srcs = [
5885 "test/f32-vmul.cc",
5886 "test/vbinary-microkernel-tester.h",
5887 ] + MICROKERNEL_TEST_HDRS,
5888 deps = MICROKERNEL_TEST_DEPS,
5889)
5890
5891xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005892 name = "f32_vmul_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005893 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005894 "test/f32-vmul-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005895 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005896 ] + MICROKERNEL_TEST_HDRS,
5897 deps = MICROKERNEL_TEST_DEPS,
5898)
5899
5900xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005901 name = "f32_vmul_relu_test",
5902 srcs = [
5903 "test/f32-vmul-relu.cc",
5904 "test/vbinary-microkernel-tester.h",
5905 ] + MICROKERNEL_TEST_HDRS,
5906 deps = MICROKERNEL_TEST_DEPS,
5907)
5908
5909xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005910 name = "f32_vmulc_test",
5911 srcs = [
5912 "test/f32-vmulc.cc",
5913 "test/vbinaryc-microkernel-tester.h",
5914 ] + MICROKERNEL_TEST_HDRS,
5915 deps = MICROKERNEL_TEST_DEPS,
5916)
5917
5918xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005919 name = "f32_vmulc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005920 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005921 "test/f32-vmulc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005922 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005923 ] + MICROKERNEL_TEST_HDRS,
5924 deps = MICROKERNEL_TEST_DEPS,
5925)
5926
5927xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005928 name = "f32_vmulc_relu_test",
5929 srcs = [
5930 "test/f32-vmulc-relu.cc",
5931 "test/vbinaryc-microkernel-tester.h",
5932 ] + MICROKERNEL_TEST_HDRS,
5933 deps = MICROKERNEL_TEST_DEPS,
5934)
5935
5936xnnpack_unit_test(
Frank Barchard2a1049c2020-06-03 02:31:27 -07005937 name = "f16_vmulcaddc_minmax_test",
5938 srcs = [
5939 "test/f16-vmulcaddc-minmax.cc",
5940 "test/vmulcaddc-microkernel-tester.h",
5941 "src/xnnpack/AlignedAllocator.h",
5942 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005943 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Frank Barchard2a1049c2020-06-03 02:31:27 -07005944)
5945
5946xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07005947 name = "f32_vmulcaddc_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005948 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07005949 "test/f32-vmulcaddc-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005950 "test/vmulcaddc-microkernel-tester.h",
5951 "src/xnnpack/AlignedAllocator.h",
5952 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005953 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005954)
5955
5956xnnpack_unit_test(
Marat Dukhan8cc7efe2020-06-10 16:24:27 -07005957 name = "f32_vlrelu_test",
5958 srcs = [
5959 "test/f32-vlrelu.cc",
5960 "test/vunary-microkernel-tester.h",
5961 ] + MICROKERNEL_TEST_HDRS,
5962 deps = MICROKERNEL_TEST_DEPS,
5963)
5964
5965xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07005966 name = "f32_vneg_test",
5967 srcs = [
5968 "test/f32-vneg.cc",
5969 "test/vunary-microkernel-tester.h",
5970 ] + MICROKERNEL_TEST_HDRS,
5971 deps = MICROKERNEL_TEST_DEPS,
5972)
5973
5974xnnpack_unit_test(
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07005975 name = "f32_vrndne_test",
5976 srcs = [
5977 "test/f32-vrndne.cc",
5978 "test/vunary-microkernel-tester.h",
5979 ] + MICROKERNEL_TEST_HDRS,
5980 deps = MICROKERNEL_TEST_DEPS,
5981)
5982
5983xnnpack_unit_test(
5984 name = "f32_vrndz_test",
5985 srcs = [
5986 "test/f32-vrndz.cc",
5987 "test/vunary-microkernel-tester.h",
5988 ] + MICROKERNEL_TEST_HDRS,
5989 deps = MICROKERNEL_TEST_DEPS,
5990)
5991
5992xnnpack_unit_test(
5993 name = "f32_vrndu_test",
5994 srcs = [
5995 "test/f32-vrndu.cc",
5996 "test/vunary-microkernel-tester.h",
5997 ] + MICROKERNEL_TEST_HDRS,
5998 deps = MICROKERNEL_TEST_DEPS,
5999)
6000
6001xnnpack_unit_test(
6002 name = "f32_vrndd_test",
6003 srcs = [
6004 "test/f32-vrndd.cc",
6005 "test/vunary-microkernel-tester.h",
6006 ] + MICROKERNEL_TEST_HDRS,
6007 deps = MICROKERNEL_TEST_DEPS,
6008)
6009
6010xnnpack_unit_test(
Marat Dukhan05ac8e32019-10-21 15:39:33 -07006011 name = "f32_vscale_test",
6012 srcs = [
6013 "test/f32-vscale.cc",
6014 "test/vscale-microkernel-tester.h",
6015 ] + MICROKERNEL_TEST_HDRS,
6016 deps = MICROKERNEL_TEST_DEPS,
6017)
6018
6019xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07006020 name = "f32_vscaleexpminusmax_test",
6021 srcs = [
6022 "test/f32-vscaleexpminusmax.cc",
6023 "test/vscaleexpminusmax-microkernel-tester.h",
6024 ] + MICROKERNEL_TEST_HDRS,
6025 deps = MICROKERNEL_TEST_DEPS,
6026)
6027
6028xnnpack_unit_test(
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07006029 name = "f32_vscaleextexp_test",
6030 srcs = [
6031 "test/f32-vscaleextexp.cc",
6032 "test/vscaleextexp-microkernel-tester.h",
6033 ] + MICROKERNEL_TEST_HDRS,
6034 deps = MICROKERNEL_TEST_DEPS,
6035)
6036
6037xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07006038 name = "f32_vsqr_test",
6039 srcs = [
6040 "test/f32-vsqr.cc",
6041 "test/vunary-microkernel-tester.h",
6042 ] + MICROKERNEL_TEST_HDRS,
6043 deps = MICROKERNEL_TEST_DEPS,
6044)
6045
6046xnnpack_unit_test(
Marat Dukhan13bafb02020-06-05 00:43:11 -07006047 name = "f32_vsqrdiff_test",
6048 srcs = [
6049 "test/f32-vsqrdiff.cc",
6050 "test/vbinary-microkernel-tester.h",
6051 ] + MICROKERNEL_TEST_HDRS,
6052 deps = MICROKERNEL_TEST_DEPS,
6053)
6054
6055xnnpack_unit_test(
6056 name = "f32_vsqrdiffc_test",
6057 srcs = [
6058 "test/f32-vsqrdiffc.cc",
6059 "test/vbinaryc-microkernel-tester.h",
6060 ] + MICROKERNEL_TEST_HDRS,
6061 deps = MICROKERNEL_TEST_DEPS,
6062)
6063
6064xnnpack_unit_test(
Marat Dukhanf4db2f32020-06-30 10:55:30 -07006065 name = "f32_vsqrt_test",
6066 srcs = [
6067 "test/f32-vsqrt.cc",
6068 "test/vunary-microkernel-tester.h",
6069 ] + MICROKERNEL_TEST_HDRS,
6070 deps = MICROKERNEL_TEST_DEPS,
6071)
6072
6073xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07006074 name = "f32_vsub_test",
6075 srcs = [
6076 "test/f32-vsub.cc",
6077 "test/vbinary-microkernel-tester.h",
6078 ] + MICROKERNEL_TEST_HDRS,
6079 deps = MICROKERNEL_TEST_DEPS,
6080)
6081
6082xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006083 name = "f32_vsub_minmax_test",
Marat Dukhan97579532019-10-18 16:40:39 -07006084 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006085 "test/f32-vsub-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08006086 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08006087 ] + MICROKERNEL_TEST_HDRS,
6088 deps = MICROKERNEL_TEST_DEPS,
6089)
6090
6091xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07006092 name = "f32_vsub_relu_test",
6093 srcs = [
6094 "test/f32-vsub-relu.cc",
6095 "test/vbinary-microkernel-tester.h",
6096 ] + MICROKERNEL_TEST_HDRS,
6097 deps = MICROKERNEL_TEST_DEPS,
6098)
6099
6100xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07006101 name = "f32_vsubc_test",
6102 srcs = [
6103 "test/f32-vsubc.cc",
6104 "test/vbinaryc-microkernel-tester.h",
6105 ] + MICROKERNEL_TEST_HDRS,
6106 deps = MICROKERNEL_TEST_DEPS,
6107)
6108
6109xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006110 name = "f32_vsubc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08006111 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006112 "test/f32-vsubc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08006113 "test/vbinaryc-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08006114 ] + MICROKERNEL_TEST_HDRS,
6115 deps = MICROKERNEL_TEST_DEPS,
6116)
6117
6118xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07006119 name = "f32_vsubc_relu_test",
6120 srcs = [
6121 "test/f32-vsubc-relu.cc",
6122 "test/vbinaryc-microkernel-tester.h",
6123 ] + MICROKERNEL_TEST_HDRS,
6124 deps = MICROKERNEL_TEST_DEPS,
6125)
6126
6127xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07006128 name = "f32_vrsubc_test",
6129 srcs = [
6130 "test/f32-vrsubc.cc",
6131 "test/vbinaryc-microkernel-tester.h",
6132 ] + MICROKERNEL_TEST_HDRS,
6133 deps = MICROKERNEL_TEST_DEPS,
6134)
6135
6136xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006137 name = "f32_vrsubc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08006138 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006139 "test/f32-vrsubc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08006140 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan97579532019-10-18 16:40:39 -07006141 ] + MICROKERNEL_TEST_HDRS,
6142 deps = MICROKERNEL_TEST_DEPS,
6143)
6144
6145xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07006146 name = "f32_vrsubc_relu_test",
6147 srcs = [
6148 "test/f32-vrsubc-relu.cc",
6149 "test/vbinaryc-microkernel-tester.h",
6150 ] + MICROKERNEL_TEST_HDRS,
6151 deps = MICROKERNEL_TEST_DEPS,
6152)
6153
6154xnnpack_unit_test(
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07006155 name = "qs8_dwconv_minmax_test",
6156 srcs = [
6157 "test/qs8-dwconv-minmax.cc",
6158 "test/dwconv-microkernel-tester.h",
6159 "src/xnnpack/AlignedAllocator.h",
6160 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
6161 deps = MICROKERNEL_TEST_DEPS + [":packing"],
6162)
6163
6164xnnpack_unit_test(
Marat Dukhan4ed53f42020-08-06 01:12:55 -07006165 name = "qs8_gavgpool_minmax_test",
6166 srcs = [
6167 "test/qs8-gavgpool-minmax.cc",
6168 "test/gavgpool-microkernel-tester.h",
6169 "src/xnnpack/AlignedAllocator.h",
6170 ] + MICROKERNEL_TEST_HDRS,
6171 deps = MICROKERNEL_TEST_DEPS,
6172)
6173
6174xnnpack_unit_test(
Marat Dukhan595e1702020-07-31 10:12:52 -07006175 name = "qs8_gemm_minmax_test",
6176 srcs = [
6177 "test/qs8-gemm-minmax.cc",
6178 "test/gemm-microkernel-tester.h",
6179 "src/xnnpack/AlignedAllocator.h",
6180 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
6181 deps = MICROKERNEL_TEST_DEPS + [":packing"],
6182)
6183
6184xnnpack_unit_test(
Marat Dukhanf9480682020-07-31 14:50:24 -07006185 name = "qs8_igemm_minmax_test",
6186 srcs = [
6187 "test/qs8-igemm-minmax.cc",
6188 "test/gemm-microkernel-tester.h",
6189 "src/xnnpack/AlignedAllocator.h",
6190 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
6191 deps = MICROKERNEL_TEST_DEPS + [":packing"],
6192)
6193
6194xnnpack_unit_test(
6195 name = "qs8_requantization_test",
6196 srcs = [
6197 "src/xnnpack/requantization-stubs.h",
6198 "test/qs8-requantization.cc",
6199 "test/requantization-tester.h",
6200 ] + MICROKERNEL_TEST_HDRS,
6201 deps = MICROKERNEL_TEST_DEPS,
6202)
6203
6204xnnpack_unit_test(
Marat Dukhand9f3ad42020-08-10 12:30:58 -07006205 name = "qs8_vadd_minmax_test",
6206 srcs = [
6207 "test/qs8-vadd-minmax.cc",
6208 "test/vadd-microkernel-tester.h",
6209 ] + MICROKERNEL_TEST_HDRS,
6210 deps = MICROKERNEL_TEST_DEPS,
6211)
6212
6213xnnpack_unit_test(
Marat Dukhan0270d9f2020-08-11 00:56:46 -07006214 name = "qs8_vaddc_minmax_test",
6215 srcs = [
6216 "test/qs8-vaddc-minmax.cc",
6217 "test/vaddc-microkernel-tester.h",
6218 ] + MICROKERNEL_TEST_HDRS,
6219 deps = MICROKERNEL_TEST_DEPS,
6220)
6221
6222xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07006223 name = "qu8_avgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006224 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07006225 "test/qu8-avgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006226 "test/avgpool-microkernel-tester.h",
6227 "src/xnnpack/AlignedAllocator.h",
6228 ] + MICROKERNEL_TEST_HDRS,
6229 deps = MICROKERNEL_TEST_DEPS,
6230)
6231
6232xnnpack_unit_test(
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07006233 name = "qu8_dwconv_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006234 srcs = [
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07006235 "test/qu8-dwconv-minmax.cc",
6236 "test/dwconv-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006237 "src/xnnpack/AlignedAllocator.h",
6238 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07006239 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07006240)
6241
6242xnnpack_unit_test(
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07006243 name = "qu8_igemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006244 srcs = [
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07006245 "test/qu8-igemm-minmax.cc",
6246 "test/gemm-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006247 "src/xnnpack/AlignedAllocator.h",
6248 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07006249 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07006250)
6251
6252xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07006253 name = "qu8_gavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006254 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07006255 "test/qu8-gavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006256 "test/gavgpool-microkernel-tester.h",
6257 "src/xnnpack/AlignedAllocator.h",
6258 ] + MICROKERNEL_TEST_HDRS,
6259 deps = MICROKERNEL_TEST_DEPS,
6260)
6261
6262xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07006263 name = "qu8_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006264 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07006265 "test/qu8-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006266 "test/gemm-microkernel-tester.h",
6267 "src/xnnpack/AlignedAllocator.h",
6268 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07006269 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07006270)
6271
6272xnnpack_unit_test(
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07006273 name = "qu8_requantization_test",
6274 srcs = [
6275 "src/xnnpack/requantization-stubs.h",
6276 "test/qu8-requantization.cc",
6277 "test/requantization-tester.h",
6278 ] + MICROKERNEL_TEST_HDRS,
6279 deps = MICROKERNEL_TEST_DEPS,
6280)
6281
6282xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07006283 name = "qu8_vadd_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006284 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07006285 "test/qu8-vadd-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006286 "test/vadd-microkernel-tester.h",
6287 ] + MICROKERNEL_TEST_HDRS,
6288 deps = MICROKERNEL_TEST_DEPS,
6289)
6290
6291xnnpack_unit_test(
6292 name = "u8_clamp_test",
6293 srcs = [
6294 "test/u8-clamp.cc",
6295 "test/clamp-microkernel-tester.h",
6296 ] + MICROKERNEL_TEST_HDRS,
6297 deps = MICROKERNEL_TEST_DEPS,
6298)
6299
6300xnnpack_unit_test(
6301 name = "u8_lut32norm_test",
6302 srcs = [
6303 "test/u8-lut32norm.cc",
6304 "test/lut-norm-microkernel-tester.h",
6305 ] + MICROKERNEL_TEST_HDRS,
6306 deps = MICROKERNEL_TEST_DEPS,
6307)
6308
6309xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07006310 name = "u8_maxpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006311 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07006312 "test/u8-maxpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006313 "test/maxpool-microkernel-tester.h",
6314 ] + MICROKERNEL_TEST_HDRS,
6315 deps = MICROKERNEL_TEST_DEPS,
6316)
6317
6318xnnpack_unit_test(
6319 name = "u8_rmax_test",
6320 srcs = [
6321 "test/u8-rmax.cc",
6322 "test/rmax-microkernel-tester.h",
6323 ] + MICROKERNEL_TEST_HDRS,
6324 deps = MICROKERNEL_TEST_DEPS,
6325)
6326
6327xnnpack_unit_test(
Marat Dukhanad71b9a2020-11-20 00:01:51 -08006328 name = "x32_depthtospace2d_chw2hwc_test",
Yury Kartynnike7841862020-11-04 18:22:18 -08006329 srcs = [
Marat Dukhanad71b9a2020-11-20 00:01:51 -08006330 "test/x32-depthtospace2d-chw2hwc.cc",
6331 "test/depthtospace-microkernel-tester.h",
Yury Kartynnike7841862020-11-04 18:22:18 -08006332 ] + MICROKERNEL_TEST_HDRS,
6333 deps = MICROKERNEL_TEST_DEPS,
6334)
6335
6336xnnpack_unit_test(
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -07006337 name = "x32_fill_test",
6338 srcs = [
6339 "test/x32-fill.cc",
6340 "test/fill-microkernel-tester.h",
6341 ] + MICROKERNEL_TEST_HDRS,
6342 deps = MICROKERNEL_TEST_DEPS,
6343)
6344
6345xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07006346 name = "x32_packx_test",
6347 srcs = [
6348 "test/x32-packx.cc",
6349 "test/pack-microkernel-tester.h",
6350 "src/xnnpack/AlignedAllocator.h",
6351 ] + MICROKERNEL_TEST_HDRS,
6352 deps = MICROKERNEL_TEST_DEPS,
6353)
6354
6355xnnpack_unit_test(
6356 name = "x32_pad_test",
6357 srcs = [
6358 "test/x32-pad.cc",
6359 "test/pad-microkernel-tester.h",
6360 ] + MICROKERNEL_TEST_HDRS,
6361 deps = MICROKERNEL_TEST_DEPS,
6362)
6363
6364xnnpack_unit_test(
6365 name = "x32_unpool_test",
6366 srcs = [
6367 "test/x32-unpool.cc",
6368 "test/unpool-microkernel-tester.h",
6369 ] + MICROKERNEL_TEST_HDRS,
6370 deps = MICROKERNEL_TEST_DEPS,
6371)
6372
6373xnnpack_unit_test(
6374 name = "x32_zip_test",
6375 srcs = [
6376 "test/x32-zip.cc",
6377 "test/zip-microkernel-tester.h",
6378 ] + MICROKERNEL_TEST_HDRS,
6379 deps = MICROKERNEL_TEST_DEPS,
6380)
6381
6382xnnpack_unit_test(
6383 name = "x8_lut_test",
6384 srcs = [
6385 "test/x8-lut.cc",
6386 "test/lut-microkernel-tester.h",
6387 ] + MICROKERNEL_TEST_HDRS,
6388 deps = MICROKERNEL_TEST_DEPS,
6389)
6390
6391xnnpack_unit_test(
6392 name = "x8_zip_test",
6393 srcs = [
6394 "test/x8-zip.cc",
6395 "test/zip-microkernel-tester.h",
6396 ] + MICROKERNEL_TEST_HDRS,
6397 deps = MICROKERNEL_TEST_DEPS,
6398)
6399
Marat Dukhan20c3b922020-03-10 03:45:06 -07006400########################## Size tests for the library #########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07006401
6402xnnpack_binary(
Marat Dukhan20c3b922020-03-10 03:45:06 -07006403 name = "operator_size_test",
6404 srcs = ["test/operator-size.c"],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07006405 deps = [":xnnpack_operators_nhwc_f32"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07006406)
6407
Marat Dukhan20c3b922020-03-10 03:45:06 -07006408xnnpack_binary(
6409 name = "subgraph_size_test",
6410 srcs = ["test/subgraph-size.c"],
6411 deps = [":XNNPACK"],
6412)
6413
6414########################### Unit tests for operators ##########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07006415
6416xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07006417 name = "abs_nc_test",
6418 srcs = [
6419 "test/abs-nc.cc",
6420 "test/abs-operator-tester.h",
6421 ],
6422 deps = OPERATOR_TEST_DEPS,
6423)
6424
6425xnnpack_unit_test(
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08006426 name = "add_nd_test",
6427 srcs = [
6428 "test/add-nd.cc",
6429 "test/binary-elementwise-operator-tester.h",
6430 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006431 deps = OPERATOR_TEST_DEPS,
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08006432)
6433
6434xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006435 name = "argmax_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006436 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006437 "test/argmax-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006438 "test/argmax-pooling-operator-tester.h",
6439 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07006440 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006441)
6442
6443xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006444 name = "average_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006445 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006446 "test/average-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006447 "test/average-pooling-operator-tester.h",
6448 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07006449 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006450)
6451
6452xnnpack_unit_test(
Marat Dukhan64e52512020-06-09 13:41:16 -07006453 name = "bankers_rounding_nc_test",
6454 srcs = [
6455 "test/bankers-rounding-nc.cc",
6456 "test/bankers-rounding-operator-tester.h",
6457 ],
6458 deps = OPERATOR_TEST_DEPS,
6459)
6460
6461xnnpack_unit_test(
6462 name = "ceiling_nc_test",
6463 srcs = [
6464 "test/ceiling-nc.cc",
6465 "test/ceiling-operator-tester.h",
6466 ],
6467 deps = OPERATOR_TEST_DEPS,
6468)
6469
6470xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006471 name = "channel_shuffle_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006472 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006473 "test/channel-shuffle-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006474 "test/channel-shuffle-operator-tester.h",
6475 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006476 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006477)
6478
6479xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006480 name = "clamp_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006481 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006482 "test/clamp-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006483 "test/clamp-operator-tester.h",
6484 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006485 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006486)
6487
6488xnnpack_unit_test(
Marat Dukhan065b11e2020-05-22 09:49:41 -07006489 name = "constant_pad_nd_test",
6490 srcs = [
6491 "test/constant-pad-nd.cc",
6492 "test/constant-pad-operator-tester.h",
6493 ],
6494 deps = OPERATOR_TEST_DEPS,
6495)
6496
6497xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006498 name = "convolution_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006499 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006500 "test/convolution-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006501 "test/convolution-operator-tester.h",
6502 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006503 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006504)
6505
6506xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006507 name = "convolution_nchw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006508 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006509 "test/convolution-nchw.cc",
6510 "test/convolution-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006511 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006512 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006513)
6514
6515xnnpack_unit_test(
Marat Dukhan4e21b272020-06-04 18:45:01 -07006516 name = "copy_nc_test",
6517 srcs = [
6518 "test/copy-nc.cc",
6519 "test/copy-operator-tester.h",
6520 ],
6521 deps = OPERATOR_TEST_DEPS,
6522)
6523
6524xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006525 name = "deconvolution_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006526 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006527 "test/deconvolution-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006528 "test/deconvolution-operator-tester.h",
6529 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07006530 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006531)
6532
6533xnnpack_unit_test(
Marat Dukhan188d1042020-11-24 23:39:40 -08006534 name = "depth_to_space_nchw2nhwc_test",
Artsiom Ablavatski0f1dc182020-11-05 19:21:50 -08006535 srcs = [
Marat Dukhan188d1042020-11-24 23:39:40 -08006536 "test/depth-to-space-nchw2nhwc.cc",
Artsiom Ablavatski0f1dc182020-11-05 19:21:50 -08006537 "test/depth-to-space-operator-tester.h",
6538 ] + OPERATOR_TEST_PARAMS_HDRS,
6539 deps = OPERATOR_TEST_DEPS,
6540)
6541
6542xnnpack_unit_test(
Marat Dukhan0e521172020-11-25 13:10:04 -08006543 name = "depth_to_space_nhwc_test",
6544 srcs = [
6545 "test/depth-to-space-nhwc.cc",
6546 "test/depth-to-space-operator-tester.h",
6547 ] + OPERATOR_TEST_PARAMS_HDRS,
6548 deps = OPERATOR_TEST_DEPS,
6549)
6550
6551xnnpack_unit_test(
Marat Dukhan69180502019-12-06 15:00:31 -08006552 name = "divide_nd_test",
6553 srcs = [
6554 "test/binary-elementwise-operator-tester.h",
6555 "test/divide-nd.cc",
6556 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006557 deps = OPERATOR_TEST_DEPS,
Marat Dukhan69180502019-12-06 15:00:31 -08006558)
6559
6560xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006561 name = "fully_connected_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006562 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006563 "test/fully-connected-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006564 "test/fully-connected-operator-tester.h",
6565 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006566 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006567)
6568
6569xnnpack_unit_test(
Marat Dukhan64e52512020-06-09 13:41:16 -07006570 name = "floor_nc_test",
6571 srcs = [
6572 "test/floor-nc.cc",
6573 "test/floor-operator-tester.h",
6574 ],
6575 deps = OPERATOR_TEST_DEPS,
6576)
6577
6578xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006579 name = "global_average_pooling_nwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006580 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006581 "test/global-average-pooling-nwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006582 "test/global-average-pooling-operator-tester.h",
Marat Dukhanef61d022020-06-19 13:54:49 -07006583 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07006584 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006585)
6586
6587xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006588 name = "global_average_pooling_ncw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006589 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006590 "test/global-average-pooling-ncw.cc",
6591 "test/global-average-pooling-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006592 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006593 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006594)
6595
6596xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006597 name = "hardswish_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006598 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006599 "test/hardswish-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006600 "test/hardswish-operator-tester.h",
6601 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006602 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006603)
6604
6605xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006606 name = "leaky_relu_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006607 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006608 "test/leaky-relu-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006609 "test/leaky-relu-operator-tester.h",
6610 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006611 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006612)
6613
6614xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006615 name = "max_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006616 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006617 "test/max-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006618 "test/max-pooling-operator-tester.h",
6619 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07006620 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006621)
6622
6623xnnpack_unit_test(
Marat Dukhan79e7f842019-12-05 14:35:50 -08006624 name = "maximum_nd_test",
6625 srcs = [
6626 "test/binary-elementwise-operator-tester.h",
6627 "test/maximum-nd.cc",
6628 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006629 deps = OPERATOR_TEST_DEPS,
Marat Dukhan79e7f842019-12-05 14:35:50 -08006630)
6631
6632xnnpack_unit_test(
6633 name = "minimum_nd_test",
6634 srcs = [
6635 "test/binary-elementwise-operator-tester.h",
6636 "test/minimum-nd.cc",
6637 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006638 deps = OPERATOR_TEST_DEPS,
Marat Dukhan79e7f842019-12-05 14:35:50 -08006639)
6640
6641xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006642 name = "multiply_nd_test",
Marat Dukhanca2733c2019-11-15 23:21:17 -08006643 srcs = [
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08006644 "test/binary-elementwise-operator-tester.h",
Marat Dukhanefc47b82019-11-18 09:25:38 -08006645 "test/multiply-nd.cc",
Marat Dukhanca2733c2019-11-15 23:21:17 -08006646 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006647 deps = OPERATOR_TEST_DEPS,
Marat Dukhanca2733c2019-11-15 23:21:17 -08006648)
6649
6650xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07006651 name = "negate_nc_test",
6652 srcs = [
6653 "test/negate-nc.cc",
6654 "test/negate-operator-tester.h",
6655 ],
6656 deps = OPERATOR_TEST_DEPS,
6657)
6658
6659xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006660 name = "prelu_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006661 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006662 "test/prelu-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006663 "test/prelu-operator-tester.h",
6664 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07006665 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006666)
6667
6668xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006669 name = "resize_bilinear_nhwc_test",
Marat Dukhan69722492019-11-11 19:55:50 -08006670 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006671 "test/resize-bilinear-nhwc.cc",
Marat Dukhan69722492019-11-11 19:55:50 -08006672 "test/resize-bilinear-operator-tester.h",
6673 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07006674 deps = OPERATOR_TEST_DEPS,
Marat Dukhan69722492019-11-11 19:55:50 -08006675)
6676
6677xnnpack_unit_test(
Artsiom Ablavatski97918102020-10-27 15:52:59 -07006678 name = "resize_bilinear_nchw_test",
6679 srcs = [
6680 "test/resize-bilinear-nchw.cc",
6681 "test/resize-bilinear-operator-tester.h",
6682 ] + OPERATOR_TEST_PARAMS_HDRS,
6683 deps = OPERATOR_TEST_DEPS,
6684)
6685
6686xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006687 name = "sigmoid_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006688 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006689 "test/sigmoid-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006690 "test/sigmoid-operator-tester.h",
6691 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006692 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006693)
6694
6695xnnpack_unit_test(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08006696 name = "softmax_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006697 srcs = [
Marat Dukhanfd8e6892020-01-27 15:25:25 -08006698 "test/softmax-nc.cc",
6699 "test/softmax-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006700 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006701 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006702)
6703
6704xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07006705 name = "square_nc_test",
6706 srcs = [
6707 "test/square-nc.cc",
6708 "test/square-operator-tester.h",
6709 ],
6710 deps = OPERATOR_TEST_DEPS,
6711)
6712
6713xnnpack_unit_test(
Marat Dukhan6804bbd2020-06-30 19:26:11 -07006714 name = "square_root_nc_test",
6715 srcs = [
6716 "test/square-root-nc.cc",
6717 "test/square-root-operator-tester.h",
6718 ],
6719 deps = OPERATOR_TEST_DEPS,
6720)
6721
6722xnnpack_unit_test(
Marat Dukhanf7399262020-06-05 10:58:44 -07006723 name = "squared_difference_nd_test",
6724 srcs = [
6725 "test/binary-elementwise-operator-tester.h",
6726 "test/squared-difference-nd.cc",
6727 ],
6728 deps = OPERATOR_TEST_DEPS,
6729)
6730
6731xnnpack_unit_test(
Marat Dukhan05f3f6d2019-12-03 15:13:53 -08006732 name = "subtract_nd_test",
6733 srcs = [
6734 "test/binary-elementwise-operator-tester.h",
6735 "test/subtract-nd.cc",
6736 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006737 deps = OPERATOR_TEST_DEPS,
Marat Dukhan05f3f6d2019-12-03 15:13:53 -08006738)
6739
6740xnnpack_unit_test(
Marat Dukhan64e52512020-06-09 13:41:16 -07006741 name = "truncation_nc_test",
6742 srcs = [
6743 "test/truncation-nc.cc",
6744 "test/truncation-operator-tester.h",
6745 ],
6746 deps = OPERATOR_TEST_DEPS,
6747)
6748
6749xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006750 name = "unpooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006751 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006752 "test/unpooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006753 "test/unpooling-operator-tester.h",
6754 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006755 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006756)
6757
Chao Mei6ddfc602020-05-13 22:29:36 -07006758############################### Misc unit tests ###############################
6759
6760xnnpack_unit_test(
6761 name = "memory_planner_test",
6762 srcs = [
6763 "test/memory-planner-test.cc",
6764 ],
6765 deps = [
6766 ":XNNPACK",
6767 ":memory_planner",
6768 ],
6769)
6770
XNNPACK Teamab8c4c82020-10-09 08:05:51 -07006771xnnpack_unit_test(
6772 name = "subgraph_nchw_test",
6773 srcs = [
6774 "src/xnnpack/subgraph.h",
6775 "test/subgraph-nchw.cc",
6776 "test/subgraph-tester.h",
6777 ],
6778 deps = [
6779 ":XNNPACK",
6780 ],
6781)
6782
Marat Dukhan08c4a432019-10-03 09:29:21 -07006783############################# Build configurations #############################
6784
Marat Dukhanb8642352019-10-30 15:43:02 -07006785# Enables usage of assembly kernels.
Marat Dukhan08c4a432019-10-03 09:29:21 -07006786config_setting(
Marat Dukhanb8642352019-10-30 15:43:02 -07006787 name = "xnn_enable_assembly_explicit_true",
6788 define_values = {"xnn_enable_assembly": "true"},
6789)
6790
6791# Disables usage of assembly kernels.
6792config_setting(
6793 name = "xnn_enable_assembly_explicit_false",
6794 define_values = {"xnn_enable_assembly": "false"},
6795)
6796
Marat Dukhan9de90e02020-06-18 16:04:12 -07006797# Enables usage of sparse inference.
6798config_setting(
6799 name = "xnn_enable_sparse_explicit_true",
6800 define_values = {"xnn_enable_sparse": "true"},
6801)
6802
6803# Disables usage of sparse inference.
6804config_setting(
6805 name = "xnn_enable_sparse_explicit_false",
6806 define_values = {"xnn_enable_sparse": "false"},
6807)
6808
Marat Dukhan05702cf2020-03-26 15:41:33 -07006809# Disables usage of HMP-aware optimizations.
6810config_setting(
6811 name = "xnn_enable_hmp_explicit_false",
6812 define_values = {"xnn_enable_hmp": "false"},
6813)
6814
Chao Mei6ddfc602020-05-13 22:29:36 -07006815# Enable usage of optimized memory allocation
6816config_setting(
6817 name = "xnn_enable_memopt_explicit_true",
6818 define_values = {"xnnpack_enable_memopt": "true"},
6819)
6820
6821# Disable usage of optimized memory allocation
6822config_setting(
6823 name = "xnn_enable_memopt_explicit_false",
6824 define_values = {"xnnpack_enable_memopt": "false"},
6825)
6826
Marat Dukhanb8642352019-10-30 15:43:02 -07006827# Builds with -c dbg
6828config_setting(
6829 name = "debug_build",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006830 values = {
Marat Dukhanb8642352019-10-30 15:43:02 -07006831 "compilation_mode": "dbg",
6832 },
6833)
6834
6835# Builds with -c opt
6836config_setting(
6837 name = "optimized_build",
6838 values = {
6839 "compilation_mode": "opt",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006840 },
6841)
6842
6843config_setting(
Marat Dukhanb8642352019-10-30 15:43:02 -07006844 name = "linux_k8",
6845 values = {"cpu": "k8"},
6846)
6847
6848config_setting(
Marat Dukhan582094e2020-04-30 17:21:25 -07006849 name = "linux_arm",
6850 values = {"cpu": "arm"},
Marat Dukhan4e45e662019-10-03 15:40:24 -07006851)
6852
6853config_setting(
Marat Dukhanf0bd4de2020-06-15 15:53:19 -07006854 name = "linux_armeabi",
6855 values = {"cpu": "armeabi"},
6856)
6857
6858config_setting(
Terry Heo68eef3f2020-04-13 22:53:52 -07006859 name = "linux_armhf",
6860 values = {"cpu": "armhf"},
6861)
6862
6863config_setting(
Marat Dukhana720e932020-06-10 13:01:11 -07006864 name = "linux_armv7a",
6865 values = {"cpu": "armv7a"},
6866)
6867
6868config_setting(
Marat Dukhan582094e2020-04-30 17:21:25 -07006869 name = "linux_aarch64",
6870 values = {"cpu": "aarch64"},
6871)
6872
6873config_setting(
Marat Dukhan08c4a432019-10-03 09:29:21 -07006874 name = "android",
6875 values = {"crosstool_top": "//external:android/crosstool"},
6876)
6877
6878config_setting(
6879 name = "android_armv7",
6880 values = {
6881 "crosstool_top": "//external:android/crosstool",
6882 "cpu": "armeabi-v7a",
6883 },
6884)
6885
6886config_setting(
6887 name = "android_arm64",
6888 values = {
6889 "crosstool_top": "//external:android/crosstool",
6890 "cpu": "arm64-v8a",
6891 },
6892)
6893
6894config_setting(
6895 name = "android_x86",
6896 values = {
6897 "crosstool_top": "//external:android/crosstool",
6898 "cpu": "x86",
6899 },
6900)
6901
6902config_setting(
6903 name = "android_x86_64",
6904 values = {
6905 "crosstool_top": "//external:android/crosstool",
6906 "cpu": "x86_64",
6907 },
6908)
6909
6910config_setting(
Marat Dukhan10a38082020-04-17 03:58:35 -07006911 name = "windows_x86_64",
6912 values = {"cpu": "x64_windows"},
Marat Dukhan9fe932e2020-04-11 17:14:15 -07006913)
6914
6915config_setting(
Marat Dukhan10a38082020-04-17 03:58:35 -07006916 name = "windows_x86_64_clang",
6917 values = {
6918 "compiler": "clang-cl",
6919 "cpu": "x64_windows",
6920 },
6921)
6922
6923config_setting(
6924 name = "windows_x86_64_mingw",
6925 values = {
6926 "compiler": "mingw-gcc",
6927 "cpu": "x64_windows",
6928 },
6929)
6930
6931config_setting(
6932 name = "windows_x86_64_msys",
6933 values = {
6934 "compiler": "msys-gcc",
6935 "cpu": "x64_windows",
6936 },
Marat Dukhan9fe932e2020-04-11 17:14:15 -07006937)
6938
6939config_setting(
Marat Dukhan885ca242019-10-07 09:17:32 -07006940 name = "macos_x86_64",
6941 values = {
6942 "apple_platform_type": "macos",
6943 "cpu": "darwin",
6944 },
6945)
6946
6947config_setting(
Marat Dukhan08c4a432019-10-03 09:29:21 -07006948 name = "emscripten",
Marat Dukhan1a729ec2019-10-07 09:31:44 -07006949 values = {"crosstool_top": "//toolchain:emscripten"},
Marat Dukhan08c4a432019-10-03 09:29:21 -07006950)
6951
6952config_setting(
6953 name = "emscripten_wasm",
6954 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07006955 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006956 "cpu": "wasm",
6957 },
6958)
6959
6960config_setting(
6961 name = "emscripten_wasmsimd",
6962 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07006963 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006964 "cpu": "wasm",
Marat Dukhan81c62602020-05-29 13:22:49 -07006965 "copt": "-msimd128",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006966 },
6967)
6968
6969config_setting(
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006970 name = "ios_armv7",
6971 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006972 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006973 "cpu": "ios_armv7",
6974 },
6975)
6976
6977config_setting(
6978 name = "ios_arm64",
6979 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006980 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006981 "cpu": "ios_arm64",
6982 },
6983)
6984
6985config_setting(
6986 name = "ios_arm64e",
6987 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006988 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006989 "cpu": "ios_arm64e",
6990 },
6991)
6992
6993config_setting(
6994 name = "ios_x86",
6995 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006996 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006997 "cpu": "ios_i386",
6998 },
6999)
7000
7001config_setting(
7002 name = "ios_x86_64",
7003 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08007004 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08007005 "cpu": "ios_x86_64",
7006 },
7007)
7008
7009config_setting(
7010 name = "watchos_armv7k",
7011 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08007012 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08007013 "cpu": "watchos_armv7k",
7014 },
7015)
7016
7017config_setting(
7018 name = "watchos_arm64_32",
7019 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08007020 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08007021 "cpu": "watchos_arm64_32",
7022 },
7023)
7024
7025config_setting(
7026 name = "watchos_x86",
7027 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08007028 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08007029 "cpu": "watchos_i386",
7030 },
7031)
7032
7033config_setting(
7034 name = "watchos_x86_64",
7035 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08007036 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08007037 "cpu": "watchos_x86_64",
7038 },
7039)
7040
7041config_setting(
7042 name = "tvos_arm64",
7043 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08007044 "apple_platform_type": "tvos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08007045 "cpu": "tvos_arm64",
7046 },
7047)
7048
7049config_setting(
7050 name = "tvos_x86_64",
7051 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08007052 "apple_platform_type": "tvos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08007053 "cpu": "tvos_x86_64",
7054 },
7055)