blob: a1829bcd972155e3c84330ddd86122f3f65be6bf [file] [log] [blame]
Frank Barchard1a953052020-11-16 18:44:58 -08001# Copyright 2020 Google LLC
Marat Dukhan08c4a432019-10-03 09:29:21 -07002#
3# This source code is licensed under the BSD-style license found in the
4# LICENSE file in the root directory of this source tree.
5#
6# Description:
7# XNNPACK - optimized floating-point neural network operators library
8
Marat Dukhan10a38082020-04-17 03:58:35 -07009load(":build_defs.bzl", "xnnpack_aggregate_library", "xnnpack_benchmark", "xnnpack_binary", "xnnpack_cc_library", "xnnpack_gcc_std_copts", "xnnpack_min_size_copts", "xnnpack_msvc_std_copts", "xnnpack_optional_armcl_copts", "xnnpack_optional_armcl_deps", "xnnpack_optional_dnnl_copts", "xnnpack_optional_dnnl_deps", "xnnpack_optional_gemmlowp_copts", "xnnpack_optional_gemmlowp_deps", "xnnpack_optional_ruy_copts", "xnnpack_optional_ruy_deps", "xnnpack_optional_tflite_copts", "xnnpack_optional_tflite_deps", "xnnpack_std_cxxopts", "xnnpack_unit_test", "xnnpack_visibility")
Marat Dukhan69c3f2c2019-11-06 12:30:01 -080010
Marat Dukhan08c4a432019-10-03 09:29:21 -070011licenses(["notice"])
12
13exports_files(["LICENSE"])
14
Marat Dukhan1b354632020-03-23 12:50:22 -070015OPERATOR_BENCHMARK_DEPS = [
16 ":XNNPACK",
17 ":bench_utils",
18 "@cpuinfo",
Frank Barchard0c849732020-06-12 13:31:32 -070019 "@FP16",
Marat Dukhan1b354632020-03-23 12:50:22 -070020 "@pthreadpool",
21]
22
Marat Dukhan08c4a432019-10-03 09:29:21 -070023MICROKERNEL_BENCHMARK_DEPS = [
24 ":ukernels",
25 ":bench_utils",
Frank Barchard7e955972019-10-11 10:34:25 -070026 ":enable_assembly",
Marat Dukhan08c4a432019-10-03 09:29:21 -070027 "@cpuinfo",
28 "@FP16",
29 "@pthreadpool",
30]
31
Marat Dukhan6adff4e2019-10-14 18:32:07 -070032ACCURACY_EVAL_DEPS = [
33 ":XNNPACK",
34 ":ukernels",
35 "@FP16",
36 "@pthreadpool",
37]
38
Marat Dukhan08c4a432019-10-03 09:29:21 -070039MICROKERNEL_TEST_DEPS = [
Marat Dukhan33fcf782020-05-24 14:27:15 -070040 ":ukernels_test_mode",
Frank Barchard7e955972019-10-11 10:34:25 -070041 ":enable_assembly",
Marat Dukhan08c4a432019-10-03 09:29:21 -070042 "@cpuinfo",
43 "@FP16",
44 "@pthreadpool",
45]
46
Marat Dukhan1b354632020-03-23 12:50:22 -070047OPERATOR_TEST_DEPS = [
Marat Dukhan33fcf782020-05-24 14:27:15 -070048 ":XNNPACK_test_mode",
Marat Dukhan1b354632020-03-23 12:50:22 -070049 "@pthreadpool",
50 "@FP16",
51]
52
Marat Dukhan08c4a432019-10-03 09:29:21 -070053OPERATOR_SRCS = [
Marat Dukhane8265432020-04-28 18:42:59 -070054 "src/operators/argmax-pooling-nhwc.c",
55 "src/operators/average-pooling-nhwc.c",
56 "src/operators/binary-elementwise-nd.c",
Marat Dukhane8265432020-04-28 18:42:59 -070057 "src/operators/channel-shuffle-nc.c",
Marat Dukhan065b11e2020-05-22 09:49:41 -070058 "src/operators/constant-pad-nd.c",
Marat Dukhane8265432020-04-28 18:42:59 -070059 "src/operators/convolution-nchw.c",
60 "src/operators/convolution-nhwc.c",
61 "src/operators/deconvolution-nhwc.c",
Marat Dukhan13b68f22020-11-12 11:55:19 -080062 "src/operators/depth-to-space-nchw2nhwc.c",
Marat Dukhan0e521172020-11-25 13:10:04 -080063 "src/operators/depth-to-space-nhwc.c",
Marat Dukhane8265432020-04-28 18:42:59 -070064 "src/operators/fully-connected-nc.c",
65 "src/operators/global-average-pooling-ncw.c",
66 "src/operators/global-average-pooling-nwc.c",
Marat Dukhane8265432020-04-28 18:42:59 -070067 "src/operators/leaky-relu-nc.c",
68 "src/operators/max-pooling-nhwc.c",
69 "src/operators/prelu-nc.c",
Artsiom Ablavatski97918102020-10-27 15:52:59 -070070 "src/operators/resize-bilinear-nchw.c",
Marat Dukhane8265432020-04-28 18:42:59 -070071 "src/operators/resize-bilinear-nhwc.c",
72 "src/operators/sigmoid-nc.c",
73 "src/operators/softmax-nc.c",
Marat Dukhanc3065f52020-06-04 13:33:32 -070074 "src/operators/unary-elementwise-nc.c",
Marat Dukhane8265432020-04-28 18:42:59 -070075 "src/operators/unpooling-nhwc.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -070076]
77
Marat Dukhan95e8b7a2020-06-03 12:46:26 -070078SUBGRAPH_SRCS = [
Marat Dukhan5fab4092020-06-10 01:28:28 -070079 "src/subgraph/abs.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -070080 "src/subgraph/add2.c",
81 "src/subgraph/argmax-pooling-2d.c",
82 "src/subgraph/average-pooling-2d.c",
Marat Dukhan5fab4092020-06-10 01:28:28 -070083 "src/subgraph/bankers-rounding.c",
84 "src/subgraph/ceiling.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -070085 "src/subgraph/clamp.c",
86 "src/subgraph/convolution-2d.c",
87 "src/subgraph/deconvolution-2d.c",
Artsiom Ablavatskibbe85062020-11-05 14:07:37 -080088 "src/subgraph/depth-to-space.c",
Frank Barchard9cef5ea2020-11-18 14:52:08 -080089 "src/subgraph/depthwise-convolution-2d.c",
Marat Dukhan9d3a4592020-06-05 16:52:42 -070090 "src/subgraph/divide.c",
Marat Dukhana1600202020-12-01 22:17:16 -080091 "src/subgraph/elu.c",
Marat Dukhan5fab4092020-06-10 01:28:28 -070092 "src/subgraph/floor.c",
Frank Barchard04336c12020-10-22 16:48:55 -070093 "src/subgraph/fully-connected.c",
Marat Dukhana059b7d2020-06-11 11:41:27 -070094 "src/subgraph/global-average-pooling-2d.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -070095 "src/subgraph/hardswish.c",
Marat Dukhan5bbebac2020-06-10 19:42:15 -070096 "src/subgraph/leaky-relu.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -070097 "src/subgraph/max-pooling-2d.c",
Marat Dukhan9d3a4592020-06-05 16:52:42 -070098 "src/subgraph/maximum2.c",
99 "src/subgraph/minimum2.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -0700100 "src/subgraph/multiply2.c",
Marat Dukhan5fab4092020-06-10 01:28:28 -0700101 "src/subgraph/negate.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -0700102 "src/subgraph/prelu.c",
103 "src/subgraph/sigmoid.c",
104 "src/subgraph/softmax.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700105 "src/subgraph/square-root.c",
106 "src/subgraph/square.c",
107 "src/subgraph/squared-difference.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -0700108 "src/subgraph/static-constant-pad.c",
Marat Dukhand27202d2020-07-09 23:43:40 -0700109 "src/subgraph/static-reshape.c",
Marat Dukhanaff24e22020-07-23 01:43:58 -0700110 "src/subgraph/static-resize-bilinear-2d.c",
Marat Dukhan9d3a4592020-06-05 16:52:42 -0700111 "src/subgraph/subtract.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -0700112 "src/subgraph/unpooling-2d.c",
113]
114
Marat Dukhan3a77ea72019-12-23 12:10:24 -0800115TABLE_SRCS = [
116 "src/tables/exp2-k-over-64.c",
117 "src/tables/exp2-k-over-2048.c",
Marat Dukhande390d42020-11-29 19:32:18 -0800118 "src/tables/exp2minus-k-over-4.c",
119 "src/tables/exp2minus-k-over-8.c",
Marat Dukhanc60742b2020-11-23 12:33:27 -0800120 "src/tables/exp2minus-k-over-16.c",
Marat Dukhan1f256fc2020-09-24 21:27:14 -0700121 "src/tables/exp2minus-k-over-64.c",
122 "src/tables/exp2minus-k-over-2048.c",
Marat Dukhan3a77ea72019-12-23 12:10:24 -0800123]
124
Marat Dukhan08c4a432019-10-03 09:29:21 -0700125SCALAR_UKERNELS = [
Marat Dukhan329da642019-11-19 21:44:39 -0800126 "src/f32-argmaxpool/4x-scalar-c1.c",
Marat Dukhan1e782c42019-11-21 17:02:40 -0800127 "src/f32-argmaxpool/9p8x-scalar-c1.c",
Marat Dukhan329da642019-11-19 21:44:39 -0800128 "src/f32-argmaxpool/9x-scalar-c1.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700129 "src/f32-avgpool/9p8x-minmax-scalar-c1.c",
130 "src/f32-avgpool/9x-minmax-scalar-c1.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -0700131 "src/f32-clamp/gen/scalar-x1.c",
132 "src/f32-clamp/gen/scalar-x2.c",
133 "src/f32-clamp/gen/scalar-x4.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700134 "src/f32-conv-hwc/3x3s2p0p1c3x4-scalar-1x1.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700135 "src/f32-conv-hwc/3x3s2p1c3x4-scalar-1x1.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -0700136 "src/f32-conv-hwc2chw/3x3s2p1c3x4-scalar-1x1.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700137 "src/f32-dwconv/gen/up1x4-minmax-scalar-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700138 "src/f32-dwconv/gen/up1x4-minmax-scalar.c",
139 "src/f32-dwconv/gen/up1x4-scalar-acc2.c",
140 "src/f32-dwconv/gen/up1x4-scalar.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700141 "src/f32-dwconv/gen/up1x9-minmax-scalar-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700142 "src/f32-dwconv/gen/up1x9-minmax-scalar.c",
143 "src/f32-dwconv/gen/up1x9-scalar-acc2.c",
144 "src/f32-dwconv/gen/up1x9-scalar.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700145 "src/f32-dwconv/gen/up1x25-minmax-scalar-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700146 "src/f32-dwconv/gen/up1x25-minmax-scalar.c",
147 "src/f32-dwconv/gen/up1x25-scalar-acc2.c",
148 "src/f32-dwconv/gen/up1x25-scalar.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700149 "src/f32-dwconv/gen/up2x4-minmax-scalar-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700150 "src/f32-dwconv/gen/up2x4-minmax-scalar.c",
151 "src/f32-dwconv/gen/up2x4-scalar-acc2.c",
152 "src/f32-dwconv/gen/up2x4-scalar.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700153 "src/f32-dwconv/gen/up2x9-minmax-scalar-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700154 "src/f32-dwconv/gen/up2x9-minmax-scalar.c",
155 "src/f32-dwconv/gen/up2x9-scalar-acc2.c",
156 "src/f32-dwconv/gen/up2x9-scalar.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700157 "src/f32-dwconv/gen/up2x25-minmax-scalar-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700158 "src/f32-dwconv/gen/up2x25-minmax-scalar.c",
159 "src/f32-dwconv/gen/up2x25-scalar-acc2.c",
160 "src/f32-dwconv/gen/up2x25-scalar.c",
Frank Barchard35db7d02020-10-26 13:37:34 -0700161 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-scalar-1x1-acc2.c",
162 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-scalar-1x1-acc3.c",
163 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-scalar-1x1-acc4.c",
Marat Dukhan91249d22020-10-24 12:02:51 -0700164 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-scalar-1x1.c",
Frank Barchard35db7d02020-10-26 13:37:34 -0700165 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-scalar-2x1-acc2.c",
Marat Dukhan91249d22020-10-24 12:02:51 -0700166 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-scalar-2x1.c",
167 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-scalar-3x1.c",
168 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-scalar-4x1.c",
169 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-scalar-5x1.c",
170 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-scalar-6x1.c",
Marat Dukhancf5b3c32020-10-25 19:21:10 -0700171 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-scalar-1x1-acc2.c",
172 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-scalar-1x1-acc3.c",
173 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-scalar-1x1-acc4.c",
Frank Barchard35db7d02020-10-26 13:37:34 -0700174 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-scalar-1x1.c",
Marat Dukhancf5b3c32020-10-25 19:21:10 -0700175 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-scalar-2x1-acc2.c",
Frank Barchard35db7d02020-10-26 13:37:34 -0700176 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-scalar-2x1.c",
177 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-scalar-3x1.c",
178 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-scalar-4x1.c",
Marat Dukhanc4efb002020-10-25 23:14:47 -0700179 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-scalar-1x1-acc2.c",
180 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-scalar-1x1-acc3.c",
181 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-scalar-1x1-acc4.c",
182 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-scalar-1x1-acc5.c",
Frank Barchard35db7d02020-10-26 13:37:34 -0700183 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-scalar-1x1.c",
Marat Dukhanc4efb002020-10-25 23:14:47 -0700184 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-scalar-2x1-acc2.c",
185 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-scalar-2x1-acc3.c",
Frank Barchard35db7d02020-10-26 13:37:34 -0700186 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-scalar-2x1.c",
Marat Dukhanc4efb002020-10-25 23:14:47 -0700187 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-scalar-3x1-acc2.c",
Frank Barchard35db7d02020-10-26 13:37:34 -0700188 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-scalar-3x1.c",
Marat Dukhan29c0c332020-10-28 22:11:00 -0700189 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-scalar-1x1-acc2.c",
190 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-scalar-1x1-acc3.c",
191 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-scalar-1x1-acc4.c",
192 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-scalar-1x1-acc5.c",
193 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-scalar-1x1.c",
194 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-scalar-2x1-acc2.c",
195 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-scalar-2x1-acc3.c",
196 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-scalar-2x1.c",
197 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-scalar-3x1-acc2.c",
198 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-scalar-3x1.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -0700199 "src/f32-gavgpool-cw/scalar-x1.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700200 "src/f32-gavgpool/7p7x-minmax-scalar-c1.c",
201 "src/f32-gavgpool/7x-minmax-scalar-c1.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700202 "src/f32-gemm/gen-inc/1x4inc-minmax-scalar.c",
203 "src/f32-gemm/gen-inc/2x4inc-minmax-scalar.c",
204 "src/f32-gemm/gen-inc/4x4inc-minmax-scalar.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700205 "src/f32-gemm/gen/1x4-minmax-scalar.c",
206 "src/f32-gemm/gen/1x4-relu-scalar.c",
207 "src/f32-gemm/gen/1x4-scalar.c",
208 "src/f32-gemm/gen/2x4-minmax-scalar.c",
209 "src/f32-gemm/gen/2x4-relu-scalar.c",
210 "src/f32-gemm/gen/2x4-scalar.c",
211 "src/f32-gemm/gen/4x2-minmax-scalar.c",
212 "src/f32-gemm/gen/4x2-relu-scalar.c",
213 "src/f32-gemm/gen/4x2-scalar.c",
214 "src/f32-gemm/gen/4x4-minmax-scalar.c",
215 "src/f32-gemm/gen/4x4-relu-scalar.c",
216 "src/f32-gemm/gen/4x4-scalar.c",
Marat Dukhanefaac272020-07-16 16:58:07 -0700217 "src/f32-hswish/gen/hswish-scalar-x1.c",
218 "src/f32-hswish/gen/hswish-scalar-x2.c",
219 "src/f32-hswish/gen/hswish-scalar-x4.c",
XNNPACK Team21432672020-10-19 19:58:48 -0700220 "src/f32-ibilinear-chw/gen/scalar-p1.c",
221 "src/f32-ibilinear-chw/gen/scalar-p2.c",
222 "src/f32-ibilinear-chw/gen/scalar-p4.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700223 "src/f32-ibilinear/gen/scalar-c1.c",
224 "src/f32-ibilinear/gen/scalar-c2.c",
225 "src/f32-ibilinear/gen/scalar-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700226 "src/f32-igemm/gen/1x4-minmax-scalar.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700227 "src/f32-igemm/gen/1x4-relu-scalar.c",
228 "src/f32-igemm/gen/1x4-scalar.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700229 "src/f32-igemm/gen/2x4-minmax-scalar.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700230 "src/f32-igemm/gen/2x4-relu-scalar.c",
231 "src/f32-igemm/gen/2x4-scalar.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700232 "src/f32-igemm/gen/4x2-minmax-scalar.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700233 "src/f32-igemm/gen/4x2-relu-scalar.c",
234 "src/f32-igemm/gen/4x2-scalar.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700235 "src/f32-igemm/gen/4x4-minmax-scalar.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700236 "src/f32-igemm/gen/4x4-relu-scalar.c",
237 "src/f32-igemm/gen/4x4-scalar.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700238 "src/f32-maxpool/9p8x-minmax-scalar-c1.c",
239 "src/f32-pavgpool/9p8x-minmax-scalar-c1.c",
240 "src/f32-pavgpool/9x-minmax-scalar-c1.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700241 "src/f32-ppmm/gen/2x4-minmax-scalar.c",
242 "src/f32-ppmm/gen/3x3-minmax-scalar.c",
243 "src/f32-ppmm/gen/4x2-minmax-scalar.c",
244 "src/f32-ppmm/gen/4x4-minmax-scalar.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800245 "src/f32-prelu/gen/scalar-2x1.c",
246 "src/f32-prelu/gen/scalar-2x4.c",
Marat Dukhanf46f6752020-01-21 11:03:49 -0800247 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x1.c",
Marat Dukhanf46f6752020-01-21 11:03:49 -0800248 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x2-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700249 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x2.c",
Marat Dukhanf46f6752020-01-21 11:03:49 -0800250 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x4-acc2.c",
251 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x4-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700252 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x4.c",
Marat Dukhanf46f6752020-01-21 11:03:49 -0800253 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x1.c",
Marat Dukhanf46f6752020-01-21 11:03:49 -0800254 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x2-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700255 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x2.c",
Marat Dukhanf46f6752020-01-21 11:03:49 -0800256 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x4-acc2.c",
257 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x4-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700258 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x4.c",
Frank Barchardfb158e22020-07-15 16:10:10 -0700259 "src/f32-relu/gen/scalar-x1.c",
260 "src/f32-relu/gen/scalar-x2.c",
261 "src/f32-relu/gen/scalar-x4.c",
Frank Barchard37297a62020-08-27 12:15:10 -0700262 "src/f32-relu/gen/scalar-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700263 "src/f32-rmax/scalar.c",
Marat Dukhan3a77ea72019-12-23 12:10:24 -0800264 "src/f32-sigmoid/gen/scalar-lut64-p2-div-x1.c",
265 "src/f32-sigmoid/gen/scalar-lut64-p2-div-x2.c",
266 "src/f32-sigmoid/gen/scalar-lut64-p2-div-x4.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700267 "src/f32-sigmoid/gen/scalar-lut2048-p1-div-x1.c",
268 "src/f32-sigmoid/gen/scalar-lut2048-p1-div-x2.c",
269 "src/f32-sigmoid/gen/scalar-lut2048-p1-div-x4.c",
Marat Dukhan3a77ea72019-12-23 12:10:24 -0800270 "src/f32-sigmoid/gen/scalar-p5-div-x1.c",
271 "src/f32-sigmoid/gen/scalar-p5-div-x2.c",
272 "src/f32-sigmoid/gen/scalar-p5-div-x4.c",
Marat Dukhan355ab432020-04-09 19:01:52 -0700273 "src/f32-spmm/gen/1x1-minmax-scalar-pipelined.c",
274 "src/f32-spmm/gen/1x1-minmax-scalar.c",
275 "src/f32-spmm/gen/2x1-minmax-scalar-pipelined.c",
276 "src/f32-spmm/gen/2x1-minmax-scalar.c",
277 "src/f32-spmm/gen/4x1-minmax-scalar-pipelined.c",
278 "src/f32-spmm/gen/4x1-minmax-scalar.c",
279 "src/f32-spmm/gen/8x1-minmax-scalar-pipelined.c",
280 "src/f32-spmm/gen/8x1-minmax-scalar.c",
281 "src/f32-spmm/gen/8x2-minmax-scalar.c",
282 "src/f32-spmm/gen/8x4-minmax-scalar.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700283 "src/f32-vbinary/gen/vadd-minmax-scalar-x1.c",
284 "src/f32-vbinary/gen/vadd-minmax-scalar-x2.c",
285 "src/f32-vbinary/gen/vadd-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700286 "src/f32-vbinary/gen/vadd-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700287 "src/f32-vbinary/gen/vadd-relu-scalar-x1.c",
288 "src/f32-vbinary/gen/vadd-relu-scalar-x2.c",
289 "src/f32-vbinary/gen/vadd-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700290 "src/f32-vbinary/gen/vadd-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700291 "src/f32-vbinary/gen/vadd-scalar-x1.c",
292 "src/f32-vbinary/gen/vadd-scalar-x2.c",
293 "src/f32-vbinary/gen/vadd-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700294 "src/f32-vbinary/gen/vadd-scalar-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700295 "src/f32-vbinary/gen/vaddc-minmax-scalar-x1.c",
296 "src/f32-vbinary/gen/vaddc-minmax-scalar-x2.c",
297 "src/f32-vbinary/gen/vaddc-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700298 "src/f32-vbinary/gen/vaddc-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700299 "src/f32-vbinary/gen/vaddc-relu-scalar-x1.c",
300 "src/f32-vbinary/gen/vaddc-relu-scalar-x2.c",
301 "src/f32-vbinary/gen/vaddc-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700302 "src/f32-vbinary/gen/vaddc-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700303 "src/f32-vbinary/gen/vaddc-scalar-x1.c",
304 "src/f32-vbinary/gen/vaddc-scalar-x2.c",
305 "src/f32-vbinary/gen/vaddc-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700306 "src/f32-vbinary/gen/vaddc-scalar-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700307 "src/f32-vbinary/gen/vdiv-minmax-scalar-x1.c",
308 "src/f32-vbinary/gen/vdiv-minmax-scalar-x2.c",
309 "src/f32-vbinary/gen/vdiv-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700310 "src/f32-vbinary/gen/vdiv-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700311 "src/f32-vbinary/gen/vdiv-relu-scalar-x1.c",
312 "src/f32-vbinary/gen/vdiv-relu-scalar-x2.c",
313 "src/f32-vbinary/gen/vdiv-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700314 "src/f32-vbinary/gen/vdiv-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700315 "src/f32-vbinary/gen/vdiv-scalar-x1.c",
316 "src/f32-vbinary/gen/vdiv-scalar-x2.c",
317 "src/f32-vbinary/gen/vdiv-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700318 "src/f32-vbinary/gen/vdiv-scalar-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700319 "src/f32-vbinary/gen/vdivc-minmax-scalar-x1.c",
320 "src/f32-vbinary/gen/vdivc-minmax-scalar-x2.c",
321 "src/f32-vbinary/gen/vdivc-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700322 "src/f32-vbinary/gen/vdivc-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700323 "src/f32-vbinary/gen/vdivc-relu-scalar-x1.c",
324 "src/f32-vbinary/gen/vdivc-relu-scalar-x2.c",
325 "src/f32-vbinary/gen/vdivc-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700326 "src/f32-vbinary/gen/vdivc-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700327 "src/f32-vbinary/gen/vdivc-scalar-x1.c",
328 "src/f32-vbinary/gen/vdivc-scalar-x2.c",
329 "src/f32-vbinary/gen/vdivc-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700330 "src/f32-vbinary/gen/vdivc-scalar-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800331 "src/f32-vbinary/gen/vmax-scalar-x1.c",
332 "src/f32-vbinary/gen/vmax-scalar-x2.c",
333 "src/f32-vbinary/gen/vmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700334 "src/f32-vbinary/gen/vmax-scalar-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800335 "src/f32-vbinary/gen/vmaxc-scalar-x1.c",
336 "src/f32-vbinary/gen/vmaxc-scalar-x2.c",
337 "src/f32-vbinary/gen/vmaxc-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700338 "src/f32-vbinary/gen/vmaxc-scalar-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800339 "src/f32-vbinary/gen/vmin-scalar-x1.c",
340 "src/f32-vbinary/gen/vmin-scalar-x2.c",
341 "src/f32-vbinary/gen/vmin-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700342 "src/f32-vbinary/gen/vmin-scalar-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800343 "src/f32-vbinary/gen/vminc-scalar-x1.c",
344 "src/f32-vbinary/gen/vminc-scalar-x2.c",
345 "src/f32-vbinary/gen/vminc-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700346 "src/f32-vbinary/gen/vminc-scalar-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700347 "src/f32-vbinary/gen/vmul-minmax-scalar-x1.c",
348 "src/f32-vbinary/gen/vmul-minmax-scalar-x2.c",
349 "src/f32-vbinary/gen/vmul-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700350 "src/f32-vbinary/gen/vmul-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700351 "src/f32-vbinary/gen/vmul-relu-scalar-x1.c",
352 "src/f32-vbinary/gen/vmul-relu-scalar-x2.c",
353 "src/f32-vbinary/gen/vmul-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700354 "src/f32-vbinary/gen/vmul-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700355 "src/f32-vbinary/gen/vmul-scalar-x1.c",
356 "src/f32-vbinary/gen/vmul-scalar-x2.c",
357 "src/f32-vbinary/gen/vmul-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700358 "src/f32-vbinary/gen/vmul-scalar-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700359 "src/f32-vbinary/gen/vmulc-minmax-scalar-x1.c",
360 "src/f32-vbinary/gen/vmulc-minmax-scalar-x2.c",
361 "src/f32-vbinary/gen/vmulc-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700362 "src/f32-vbinary/gen/vmulc-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700363 "src/f32-vbinary/gen/vmulc-relu-scalar-x1.c",
364 "src/f32-vbinary/gen/vmulc-relu-scalar-x2.c",
365 "src/f32-vbinary/gen/vmulc-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700366 "src/f32-vbinary/gen/vmulc-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700367 "src/f32-vbinary/gen/vmulc-scalar-x1.c",
368 "src/f32-vbinary/gen/vmulc-scalar-x2.c",
369 "src/f32-vbinary/gen/vmulc-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700370 "src/f32-vbinary/gen/vmulc-scalar-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700371 "src/f32-vbinary/gen/vrdivc-minmax-scalar-x1.c",
372 "src/f32-vbinary/gen/vrdivc-minmax-scalar-x2.c",
373 "src/f32-vbinary/gen/vrdivc-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700374 "src/f32-vbinary/gen/vrdivc-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700375 "src/f32-vbinary/gen/vrdivc-relu-scalar-x1.c",
376 "src/f32-vbinary/gen/vrdivc-relu-scalar-x2.c",
377 "src/f32-vbinary/gen/vrdivc-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700378 "src/f32-vbinary/gen/vrdivc-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700379 "src/f32-vbinary/gen/vrdivc-scalar-x1.c",
380 "src/f32-vbinary/gen/vrdivc-scalar-x2.c",
381 "src/f32-vbinary/gen/vrdivc-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700382 "src/f32-vbinary/gen/vrdivc-scalar-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700383 "src/f32-vbinary/gen/vrsubc-minmax-scalar-x1.c",
384 "src/f32-vbinary/gen/vrsubc-minmax-scalar-x2.c",
385 "src/f32-vbinary/gen/vrsubc-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700386 "src/f32-vbinary/gen/vrsubc-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700387 "src/f32-vbinary/gen/vrsubc-relu-scalar-x1.c",
388 "src/f32-vbinary/gen/vrsubc-relu-scalar-x2.c",
389 "src/f32-vbinary/gen/vrsubc-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700390 "src/f32-vbinary/gen/vrsubc-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700391 "src/f32-vbinary/gen/vrsubc-scalar-x1.c",
392 "src/f32-vbinary/gen/vrsubc-scalar-x2.c",
393 "src/f32-vbinary/gen/vrsubc-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700394 "src/f32-vbinary/gen/vrsubc-scalar-x8.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -0700395 "src/f32-vbinary/gen/vsqrdiff-scalar-x1.c",
396 "src/f32-vbinary/gen/vsqrdiff-scalar-x2.c",
397 "src/f32-vbinary/gen/vsqrdiff-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700398 "src/f32-vbinary/gen/vsqrdiff-scalar-x8.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -0700399 "src/f32-vbinary/gen/vsqrdiffc-scalar-x1.c",
400 "src/f32-vbinary/gen/vsqrdiffc-scalar-x2.c",
401 "src/f32-vbinary/gen/vsqrdiffc-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700402 "src/f32-vbinary/gen/vsqrdiffc-scalar-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700403 "src/f32-vbinary/gen/vsub-minmax-scalar-x1.c",
404 "src/f32-vbinary/gen/vsub-minmax-scalar-x2.c",
405 "src/f32-vbinary/gen/vsub-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700406 "src/f32-vbinary/gen/vsub-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700407 "src/f32-vbinary/gen/vsub-relu-scalar-x1.c",
408 "src/f32-vbinary/gen/vsub-relu-scalar-x2.c",
409 "src/f32-vbinary/gen/vsub-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700410 "src/f32-vbinary/gen/vsub-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700411 "src/f32-vbinary/gen/vsub-scalar-x1.c",
412 "src/f32-vbinary/gen/vsub-scalar-x2.c",
413 "src/f32-vbinary/gen/vsub-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700414 "src/f32-vbinary/gen/vsub-scalar-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700415 "src/f32-vbinary/gen/vsubc-minmax-scalar-x1.c",
416 "src/f32-vbinary/gen/vsubc-minmax-scalar-x2.c",
417 "src/f32-vbinary/gen/vsubc-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700418 "src/f32-vbinary/gen/vsubc-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700419 "src/f32-vbinary/gen/vsubc-relu-scalar-x1.c",
420 "src/f32-vbinary/gen/vsubc-relu-scalar-x2.c",
421 "src/f32-vbinary/gen/vsubc-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700422 "src/f32-vbinary/gen/vsubc-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700423 "src/f32-vbinary/gen/vsubc-scalar-x1.c",
424 "src/f32-vbinary/gen/vsubc-scalar-x2.c",
425 "src/f32-vbinary/gen/vsubc-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700426 "src/f32-vbinary/gen/vsubc-scalar-x8.c",
Marat Dukhaned6baaf2020-12-01 15:07:08 -0800427 "src/f32-velu/gen/velu-scalar-rr2-lut16-p3-x1.c",
428 "src/f32-velu/gen/velu-scalar-rr2-lut16-p3-x2.c",
429 "src/f32-velu/gen/velu-scalar-rr2-lut16-p3-x3.c",
430 "src/f32-velu/gen/velu-scalar-rr2-lut16-p3-x4.c",
431 "src/f32-velu/gen/velu-scalar-rr2-lut16-p3-x5.c",
432 "src/f32-velu/gen/velu-scalar-rr2-lut16-p3-x6.c",
433 "src/f32-velu/gen/velu-scalar-rr2-p6-x1.c",
434 "src/f32-velu/gen/velu-scalar-rr2-p6-x2.c",
435 "src/f32-velu/gen/velu-scalar-rr2-p6-x3.c",
436 "src/f32-velu/gen/velu-scalar-rr2-p6-x4.c",
437 "src/f32-velu/gen/velu-scalar-rr2-p6-x5.c",
438 "src/f32-velu/gen/velu-scalar-rr2-p6-x6.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700439 "src/f32-vlrelu/gen/vlrelu-scalar-x1.c",
440 "src/f32-vlrelu/gen/vlrelu-scalar-x2.c",
441 "src/f32-vlrelu/gen/vlrelu-scalar-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700442 "src/f32-vmulcaddc/gen/c1-minmax-scalar-2x.c",
443 "src/f32-vmulcaddc/gen/c2-minmax-scalar-2x.c",
444 "src/f32-vmulcaddc/gen/c4-minmax-scalar-2x.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -0700445 "src/f32-vrnd/gen/vrndd-scalar-libm-x1.c",
446 "src/f32-vrnd/gen/vrndd-scalar-libm-x2.c",
447 "src/f32-vrnd/gen/vrndd-scalar-libm-x4.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700448 "src/f32-vrnd/gen/vrndne-scalar-libm-x1.c",
449 "src/f32-vrnd/gen/vrndne-scalar-libm-x2.c",
450 "src/f32-vrnd/gen/vrndne-scalar-libm-x4.c",
451 "src/f32-vrnd/gen/vrndu-scalar-libm-x1.c",
452 "src/f32-vrnd/gen/vrndu-scalar-libm-x2.c",
453 "src/f32-vrnd/gen/vrndu-scalar-libm-x4.c",
454 "src/f32-vrnd/gen/vrndz-scalar-libm-x1.c",
455 "src/f32-vrnd/gen/vrndz-scalar-libm-x2.c",
456 "src/f32-vrnd/gen/vrndz-scalar-libm-x4.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -0700457 "src/f32-vsqrt/gen/scalar-sqrt-x1.c",
458 "src/f32-vsqrt/gen/scalar-sqrt-x2.c",
459 "src/f32-vsqrt/gen/scalar-sqrt-x4.c",
Marat Dukhan5020b962020-06-08 13:30:10 -0700460 "src/f32-vunary/gen/vabs-scalar-x1.c",
461 "src/f32-vunary/gen/vabs-scalar-x2.c",
462 "src/f32-vunary/gen/vabs-scalar-x4.c",
463 "src/f32-vunary/gen/vneg-scalar-x1.c",
464 "src/f32-vunary/gen/vneg-scalar-x2.c",
465 "src/f32-vunary/gen/vneg-scalar-x4.c",
466 "src/f32-vunary/gen/vsqr-scalar-x1.c",
467 "src/f32-vunary/gen/vsqr-scalar-x2.c",
468 "src/f32-vunary/gen/vsqr-scalar-x4.c",
Marat Dukhande390d42020-11-29 19:32:18 -0800469 "src/math/expm1minus-scalar-rr2-lut4-p4.c",
470 "src/math/expm1minus-scalar-rr2-lut8-p3.c",
471 "src/math/expm1minus-scalar-rr2-lut8-p4.c",
Marat Dukhanc60742b2020-11-23 12:33:27 -0800472 "src/math/expm1minus-scalar-rr2-lut16-p3.c",
473 "src/math/expm1minus-scalar-rr2-lut16-p4.c",
474 "src/math/expm1minus-scalar-rr2-p5.c",
475 "src/math/expm1minus-scalar-rr2-p6.c",
Frank Barchard22136062020-11-24 18:44:46 -0800476 "src/math/expminus-scalar-rr2-lut64-p2.c",
477 "src/math/expminus-scalar-rr2-lut2048-p1.c",
478 "src/math/expminus-scalar-rr2-p5.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -0700479 "src/math/roundd-scalar-addsub.c",
480 "src/math/roundd-scalar-cvt.c",
Marat Dukhanffbf96a2020-05-14 02:59:08 -0700481 "src/math/roundd-scalar-floor.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700482 "src/math/roundne-scalar-addsub.c",
483 "src/math/roundne-scalar-nearbyint.c",
484 "src/math/roundne-scalar-rint.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -0700485 "src/math/roundu-scalar-addsub.c",
Marat Dukhanffbf96a2020-05-14 02:59:08 -0700486 "src/math/roundu-scalar-ceil.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -0700487 "src/math/roundu-scalar-cvt.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -0700488 "src/math/roundz-scalar-addsub.c",
489 "src/math/roundz-scalar-cvt.c",
Marat Dukhanffbf96a2020-05-14 02:59:08 -0700490 "src/math/roundz-scalar-trunc.c",
Marat Dukhanf8475d62020-09-17 15:01:43 -0700491 "src/math/sigmoid-scalar-rr2-lut64-p2-div.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700492 "src/math/sigmoid-scalar-rr2-lut2048-p1-div.c",
Marat Dukhanf8475d62020-09-17 15:01:43 -0700493 "src/math/sigmoid-scalar-rr2-p5-div.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -0700494 "src/qs8-requantization/fp32-scalar-lrintf.c",
495 "src/qs8-requantization/fp32-scalar-magic.c",
496 "src/qs8-requantization/precise-scalar-signed64.c",
497 "src/qs8-requantization/precise-scalar-unsigned32.c",
498 "src/qs8-requantization/precise-scalar-unsigned64.c",
499 "src/qs8-requantization/q31-scalar.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -0700500 "src/qu8-avgpool/9p8x-minmax-scalar-c1.c",
501 "src/qu8-avgpool/9x-minmax-scalar-c1.c",
502 "src/qu8-dwconv/up1x9-minmax-scalar.c",
503 "src/qu8-gavgpool/7p7x-minmax-scalar-c1.c",
504 "src/qu8-gavgpool/7x-minmax-scalar-c1.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700505 "src/qu8-gemm/2x2-minmax-scalar.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -0700506 "src/qu8-igemm/2x2-minmax-scalar.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -0700507 "src/qu8-requantization/fp32-scalar-lrintf.c",
508 "src/qu8-requantization/fp32-scalar-magic.c",
509 "src/qu8-requantization/precise-scalar-signed64.c",
510 "src/qu8-requantization/precise-scalar-unsigned32.c",
511 "src/qu8-requantization/precise-scalar-unsigned64.c",
512 "src/qu8-requantization/q31-scalar.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -0700513 "src/qu8-vadd/minmax-scalar.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -0700514 "src/u8-clamp/scalar-x4.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700515 "src/u8-lut32norm/scalar.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700516 "src/u8-maxpool/9p8x-minmax-scalar-c1.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700517 "src/u8-rmax/scalar.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700518 "src/x8-lut/scalar.c",
519 "src/x8-zip/x2-scalar.c",
520 "src/x8-zip/x3-scalar.c",
521 "src/x8-zip/x4-scalar.c",
522 "src/x8-zip/xm-scalar.c",
Marat Dukhanad71b9a2020-11-20 00:01:51 -0800523 "src/x32-depthtospace2d-chw2hwc/scalar.c",
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -0700524 "src/x32-fill/scalar-float.c",
525 "src/x32-fill/scalar-int.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700526 "src/x32-packx/x2-scalar.c",
527 "src/x32-packx/x3-scalar.c",
528 "src/x32-packx/x4-scalar.c",
Marat Dukhan63523d42020-05-22 17:07:33 -0700529 "src/x32-pad/scalar-float.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700530 "src/x32-pad/scalar-int.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700531 "src/x32-unpool/scalar.c",
532 "src/x32-zip/x2-scalar.c",
533 "src/x32-zip/x3-scalar.c",
534 "src/x32-zip/x4-scalar.c",
535 "src/x32-zip/xm-scalar.c",
Marat Dukhan048931b2020-11-24 20:53:54 -0800536 "src/xx-copy/memcpy.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700537]
538
Marat Dukhan436ebe62019-12-04 15:10:12 -0800539WASM_UKERNELS = [
Marat Dukhan99936602020-04-11 16:47:01 -0700540 "src/f32-avgpool/9p8x-minmax-wasm-c1.c",
541 "src/f32-avgpool/9x-minmax-wasm-c1.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -0700542 "src/f32-clamp/gen/wasm-x1.c",
543 "src/f32-clamp/gen/wasm-x2.c",
544 "src/f32-clamp/gen/wasm-x4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700545 "src/f32-dwconv/gen/up1x4-minmax-wasm-acc2.c",
546 "src/f32-dwconv/gen/up1x4-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700547 "src/f32-dwconv/gen/up1x4-wasm-acc2.c",
548 "src/f32-dwconv/gen/up1x4-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700549 "src/f32-dwconv/gen/up1x9-minmax-wasm-acc2.c",
550 "src/f32-dwconv/gen/up1x9-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700551 "src/f32-dwconv/gen/up1x9-wasm-acc2.c",
552 "src/f32-dwconv/gen/up1x9-wasm.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700553 "src/f32-dwconv/gen/up1x25-minmax-wasm-acc2.c",
554 "src/f32-dwconv/gen/up1x25-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700555 "src/f32-dwconv/gen/up1x25-wasm-acc2.c",
556 "src/f32-dwconv/gen/up1x25-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700557 "src/f32-dwconv/gen/up2x4-minmax-wasm-acc2.c",
558 "src/f32-dwconv/gen/up2x4-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700559 "src/f32-dwconv/gen/up2x4-wasm-acc2.c",
560 "src/f32-dwconv/gen/up2x4-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700561 "src/f32-dwconv/gen/up2x9-minmax-wasm-acc2.c",
562 "src/f32-dwconv/gen/up2x9-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700563 "src/f32-dwconv/gen/up2x9-wasm-acc2.c",
564 "src/f32-dwconv/gen/up2x9-wasm.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700565 "src/f32-dwconv/gen/up2x25-minmax-wasm-acc2.c",
566 "src/f32-dwconv/gen/up2x25-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700567 "src/f32-dwconv/gen/up2x25-wasm-acc2.c",
568 "src/f32-dwconv/gen/up2x25-wasm.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700569 "src/f32-gavgpool/7p7x-minmax-wasm-c1.c",
570 "src/f32-gavgpool/7x-minmax-wasm-c1.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700571 "src/f32-gemm/gen-inc/1x4inc-minmax-wasm.c",
572 "src/f32-gemm/gen-inc/2x4inc-minmax-wasm.c",
573 "src/f32-gemm/gen-inc/4x4inc-minmax-wasm.c",
574 "src/f32-gemm/gen/1x4-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700575 "src/f32-gemm/gen/1x4-relu-wasm.c",
576 "src/f32-gemm/gen/1x4-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700577 "src/f32-gemm/gen/2x4-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700578 "src/f32-gemm/gen/2x4-relu-wasm.c",
579 "src/f32-gemm/gen/2x4-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700580 "src/f32-gemm/gen/4x2-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700581 "src/f32-gemm/gen/4x2-relu-wasm.c",
582 "src/f32-gemm/gen/4x2-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700583 "src/f32-gemm/gen/4x4-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700584 "src/f32-gemm/gen/4x4-relu-wasm.c",
585 "src/f32-gemm/gen/4x4-wasm.c",
Marat Dukhanefaac272020-07-16 16:58:07 -0700586 "src/f32-hswish/gen/hswish-wasm-x1.c",
587 "src/f32-hswish/gen/hswish-wasm-x2.c",
588 "src/f32-hswish/gen/hswish-wasm-x4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700589 "src/f32-igemm/gen/1x4-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700590 "src/f32-igemm/gen/1x4-relu-wasm.c",
591 "src/f32-igemm/gen/1x4-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700592 "src/f32-igemm/gen/2x4-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700593 "src/f32-igemm/gen/2x4-relu-wasm.c",
594 "src/f32-igemm/gen/2x4-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700595 "src/f32-igemm/gen/4x2-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700596 "src/f32-igemm/gen/4x2-relu-wasm.c",
597 "src/f32-igemm/gen/4x2-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700598 "src/f32-igemm/gen/4x4-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700599 "src/f32-igemm/gen/4x4-relu-wasm.c",
600 "src/f32-igemm/gen/4x4-wasm.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700601 "src/f32-maxpool/9p8x-minmax-wasm-c1.c",
602 "src/f32-pavgpool/9p8x-minmax-wasm-c1.c",
603 "src/f32-pavgpool/9x-minmax-wasm-c1.c",
Marat Dukhan7c1f8082020-06-25 13:26:20 -0700604 "src/f32-prelu/gen/wasm-2x1.c",
605 "src/f32-prelu/gen/wasm-2x4.c",
Frank Barchardfb158e22020-07-15 16:10:10 -0700606 "src/f32-relu/gen/wasm-x1.c",
607 "src/f32-relu/gen/wasm-x2.c",
608 "src/f32-relu/gen/wasm-x4.c",
Frank Barchard37297a62020-08-27 12:15:10 -0700609 "src/f32-relu/gen/wasm-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700610 "src/f32-vbinary/gen/vadd-minmax-wasm-x1.c",
611 "src/f32-vbinary/gen/vadd-minmax-wasm-x2.c",
612 "src/f32-vbinary/gen/vadd-minmax-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700613 "src/f32-vbinary/gen/vadd-minmax-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700614 "src/f32-vbinary/gen/vadd-relu-wasm-x1.c",
615 "src/f32-vbinary/gen/vadd-relu-wasm-x2.c",
616 "src/f32-vbinary/gen/vadd-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700617 "src/f32-vbinary/gen/vadd-relu-wasm-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700618 "src/f32-vbinary/gen/vaddc-minmax-wasm-x1.c",
619 "src/f32-vbinary/gen/vaddc-minmax-wasm-x2.c",
620 "src/f32-vbinary/gen/vaddc-minmax-wasm-x4.c",
621 "src/f32-vbinary/gen/vaddc-minmax-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700622 "src/f32-vbinary/gen/vaddc-relu-wasm-x1.c",
623 "src/f32-vbinary/gen/vaddc-relu-wasm-x2.c",
624 "src/f32-vbinary/gen/vaddc-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700625 "src/f32-vbinary/gen/vaddc-relu-wasm-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700626 "src/f32-vbinary/gen/vdiv-minmax-wasm-x1.c",
627 "src/f32-vbinary/gen/vdiv-minmax-wasm-x2.c",
628 "src/f32-vbinary/gen/vdiv-minmax-wasm-x4.c",
629 "src/f32-vbinary/gen/vdiv-minmax-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700630 "src/f32-vbinary/gen/vdiv-relu-wasm-x1.c",
631 "src/f32-vbinary/gen/vdiv-relu-wasm-x2.c",
632 "src/f32-vbinary/gen/vdiv-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700633 "src/f32-vbinary/gen/vdiv-relu-wasm-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700634 "src/f32-vbinary/gen/vdivc-minmax-wasm-x1.c",
635 "src/f32-vbinary/gen/vdivc-minmax-wasm-x2.c",
636 "src/f32-vbinary/gen/vdivc-minmax-wasm-x4.c",
637 "src/f32-vbinary/gen/vdivc-minmax-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700638 "src/f32-vbinary/gen/vdivc-relu-wasm-x1.c",
639 "src/f32-vbinary/gen/vdivc-relu-wasm-x2.c",
640 "src/f32-vbinary/gen/vdivc-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700641 "src/f32-vbinary/gen/vdivc-relu-wasm-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800642 "src/f32-vbinary/gen/vmax-wasm-x1.c",
643 "src/f32-vbinary/gen/vmax-wasm-x2.c",
644 "src/f32-vbinary/gen/vmax-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700645 "src/f32-vbinary/gen/vmax-wasm-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800646 "src/f32-vbinary/gen/vmaxc-wasm-x1.c",
647 "src/f32-vbinary/gen/vmaxc-wasm-x2.c",
648 "src/f32-vbinary/gen/vmaxc-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700649 "src/f32-vbinary/gen/vmaxc-wasm-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800650 "src/f32-vbinary/gen/vmin-wasm-x1.c",
651 "src/f32-vbinary/gen/vmin-wasm-x2.c",
652 "src/f32-vbinary/gen/vmin-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700653 "src/f32-vbinary/gen/vmin-wasm-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800654 "src/f32-vbinary/gen/vminc-wasm-x1.c",
655 "src/f32-vbinary/gen/vminc-wasm-x2.c",
656 "src/f32-vbinary/gen/vminc-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700657 "src/f32-vbinary/gen/vminc-wasm-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700658 "src/f32-vbinary/gen/vmul-minmax-wasm-x1.c",
659 "src/f32-vbinary/gen/vmul-minmax-wasm-x2.c",
660 "src/f32-vbinary/gen/vmul-minmax-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700661 "src/f32-vbinary/gen/vmul-minmax-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700662 "src/f32-vbinary/gen/vmul-relu-wasm-x1.c",
663 "src/f32-vbinary/gen/vmul-relu-wasm-x2.c",
664 "src/f32-vbinary/gen/vmul-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700665 "src/f32-vbinary/gen/vmul-relu-wasm-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700666 "src/f32-vbinary/gen/vmulc-minmax-wasm-x1.c",
667 "src/f32-vbinary/gen/vmulc-minmax-wasm-x2.c",
668 "src/f32-vbinary/gen/vmulc-minmax-wasm-x4.c",
669 "src/f32-vbinary/gen/vmulc-minmax-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700670 "src/f32-vbinary/gen/vmulc-relu-wasm-x1.c",
671 "src/f32-vbinary/gen/vmulc-relu-wasm-x2.c",
672 "src/f32-vbinary/gen/vmulc-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700673 "src/f32-vbinary/gen/vmulc-relu-wasm-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700674 "src/f32-vbinary/gen/vrdivc-minmax-wasm-x1.c",
675 "src/f32-vbinary/gen/vrdivc-minmax-wasm-x2.c",
676 "src/f32-vbinary/gen/vrdivc-minmax-wasm-x4.c",
677 "src/f32-vbinary/gen/vrdivc-minmax-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700678 "src/f32-vbinary/gen/vrdivc-relu-wasm-x1.c",
679 "src/f32-vbinary/gen/vrdivc-relu-wasm-x2.c",
680 "src/f32-vbinary/gen/vrdivc-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700681 "src/f32-vbinary/gen/vrdivc-relu-wasm-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700682 "src/f32-vbinary/gen/vrsubc-minmax-wasm-x1.c",
683 "src/f32-vbinary/gen/vrsubc-minmax-wasm-x2.c",
684 "src/f32-vbinary/gen/vrsubc-minmax-wasm-x4.c",
685 "src/f32-vbinary/gen/vrsubc-minmax-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700686 "src/f32-vbinary/gen/vrsubc-relu-wasm-x1.c",
687 "src/f32-vbinary/gen/vrsubc-relu-wasm-x2.c",
688 "src/f32-vbinary/gen/vrsubc-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700689 "src/f32-vbinary/gen/vrsubc-relu-wasm-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700690 "src/f32-vbinary/gen/vsub-minmax-wasm-x1.c",
691 "src/f32-vbinary/gen/vsub-minmax-wasm-x2.c",
692 "src/f32-vbinary/gen/vsub-minmax-wasm-x4.c",
693 "src/f32-vbinary/gen/vsub-minmax-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700694 "src/f32-vbinary/gen/vsub-relu-wasm-x1.c",
695 "src/f32-vbinary/gen/vsub-relu-wasm-x2.c",
696 "src/f32-vbinary/gen/vsub-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700697 "src/f32-vbinary/gen/vsub-relu-wasm-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700698 "src/f32-vbinary/gen/vsubc-minmax-wasm-x1.c",
699 "src/f32-vbinary/gen/vsubc-minmax-wasm-x2.c",
700 "src/f32-vbinary/gen/vsubc-minmax-wasm-x4.c",
701 "src/f32-vbinary/gen/vsubc-minmax-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700702 "src/f32-vbinary/gen/vsubc-relu-wasm-x1.c",
703 "src/f32-vbinary/gen/vsubc-relu-wasm-x2.c",
704 "src/f32-vbinary/gen/vsubc-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700705 "src/f32-vbinary/gen/vsubc-relu-wasm-x8.c",
Marat Dukhaned6baaf2020-12-01 15:07:08 -0800706 "src/f32-velu/gen/velu-wasm-rr2-lut16-p3-x1.c",
707 "src/f32-velu/gen/velu-wasm-rr2-lut16-p3-x2.c",
708 "src/f32-velu/gen/velu-wasm-rr2-lut16-p3-x3.c",
709 "src/f32-velu/gen/velu-wasm-rr2-lut16-p3-x4.c",
710 "src/f32-velu/gen/velu-wasm-rr2-lut16-p3-x5.c",
711 "src/f32-velu/gen/velu-wasm-rr2-lut16-p3-x6.c",
712 "src/f32-velu/gen/velu-wasm-rr2-p6-x1.c",
713 "src/f32-velu/gen/velu-wasm-rr2-p6-x2.c",
714 "src/f32-velu/gen/velu-wasm-rr2-p6-x3.c",
715 "src/f32-velu/gen/velu-wasm-rr2-p6-x4.c",
716 "src/f32-velu/gen/velu-wasm-rr2-p6-x5.c",
717 "src/f32-velu/gen/velu-wasm-rr2-p6-x6.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -0700718 "src/f32-vlrelu/gen/vlrelu-wasm-x1.c",
719 "src/f32-vlrelu/gen/vlrelu-wasm-x2.c",
720 "src/f32-vlrelu/gen/vlrelu-wasm-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700721 "src/f32-vmulcaddc/gen/c1-minmax-wasm-2x.c",
722 "src/f32-vmulcaddc/gen/c2-minmax-wasm-2x.c",
723 "src/f32-vmulcaddc/gen/c4-minmax-wasm-2x.c",
Marat Dukhan436ebe62019-12-04 15:10:12 -0800724]
725
Marat Dukhan290055c2020-06-09 12:24:29 -0700726WASMSIMD_UKERNELS = [
Marat Dukhan40f05522020-07-16 22:33:12 -0700727 "src/f32-argmaxpool/4x-wasmsimd-c4.c",
728 "src/f32-argmaxpool/9p8x-wasmsimd-c4.c",
729 "src/f32-argmaxpool/9x-wasmsimd-c4.c",
Marat Dukhan3b7432d2020-07-16 17:46:32 -0700730 "src/f32-avgpool/9p8x-minmax-wasmsimd-arm-c4.c",
731 "src/f32-avgpool/9p8x-minmax-wasmsimd-x86-c4.c",
732 "src/f32-avgpool/9x-minmax-wasmsimd-arm-c4.c",
733 "src/f32-avgpool/9x-minmax-wasmsimd-x86-c4.c",
Marat Dukhan3fa52c82020-07-08 12:54:33 -0700734 "src/f32-clamp/gen/wasmsimd-arm-x4.c",
735 "src/f32-clamp/gen/wasmsimd-arm-x8.c",
736 "src/f32-clamp/gen/wasmsimd-x86-x4.c",
737 "src/f32-clamp/gen/wasmsimd-x86-x8.c",
Frank Barchard22136062020-11-24 18:44:46 -0800738 "src/f32-conv-hwc2chw/3x3s2p1c3x4-wasmsimd-2x2.c",
Frank Barchard0725b8d2020-12-07 11:07:35 -0800739 "src/f32-dwconv/gen/up4x4-minmax-wasmsimd-arm-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700740 "src/f32-dwconv/gen/up4x4-minmax-wasmsimd-arm.c",
Frank Barchard0725b8d2020-12-07 11:07:35 -0800741 "src/f32-dwconv/gen/up4x4-minmax-wasmsimd-x86-acc2.c",
Marat Dukhanac014d72020-06-16 08:36:47 -0700742 "src/f32-dwconv/gen/up4x4-minmax-wasmsimd-x86.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700743 "src/f32-dwconv/gen/up4x4-wasmsimd.c",
Frank Barchard0725b8d2020-12-07 11:07:35 -0800744 "src/f32-dwconv/gen/up4x9-minmax-wasmsimd-arm-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700745 "src/f32-dwconv/gen/up4x9-minmax-wasmsimd-arm.c",
Frank Barchard0725b8d2020-12-07 11:07:35 -0800746 "src/f32-dwconv/gen/up4x9-minmax-wasmsimd-x86-acc2.c",
Marat Dukhanac014d72020-06-16 08:36:47 -0700747 "src/f32-dwconv/gen/up4x9-minmax-wasmsimd-x86.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700748 "src/f32-dwconv/gen/up4x9-wasmsimd.c",
Frank Barchard0725b8d2020-12-07 11:07:35 -0800749 "src/f32-dwconv/gen/up4x25-minmax-wasmsimd-arm-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700750 "src/f32-dwconv/gen/up4x25-minmax-wasmsimd-arm.c",
Frank Barchard0725b8d2020-12-07 11:07:35 -0800751 "src/f32-dwconv/gen/up4x25-minmax-wasmsimd-x86-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700752 "src/f32-dwconv/gen/up4x25-minmax-wasmsimd-x86.c",
753 "src/f32-dwconv/gen/up4x25-wasmsimd.c",
Frank Barchard0725b8d2020-12-07 11:07:35 -0800754 "src/f32-dwconv/gen/up8x4-minmax-wasmsimd-arm-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700755 "src/f32-dwconv/gen/up8x4-minmax-wasmsimd-arm.c",
Frank Barchard0725b8d2020-12-07 11:07:35 -0800756 "src/f32-dwconv/gen/up8x4-minmax-wasmsimd-x86-acc2.c",
Marat Dukhanac014d72020-06-16 08:36:47 -0700757 "src/f32-dwconv/gen/up8x4-minmax-wasmsimd-x86.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700758 "src/f32-dwconv/gen/up8x4-wasmsimd.c",
Frank Barchard0725b8d2020-12-07 11:07:35 -0800759 "src/f32-dwconv/gen/up8x9-minmax-wasmsimd-arm-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700760 "src/f32-dwconv/gen/up8x9-minmax-wasmsimd-arm.c",
Frank Barchard0725b8d2020-12-07 11:07:35 -0800761 "src/f32-dwconv/gen/up8x9-minmax-wasmsimd-x86-acc2.c",
Marat Dukhanac014d72020-06-16 08:36:47 -0700762 "src/f32-dwconv/gen/up8x9-minmax-wasmsimd-x86.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700763 "src/f32-dwconv/gen/up8x9-wasmsimd.c",
Frank Barchard0725b8d2020-12-07 11:07:35 -0800764 "src/f32-dwconv/gen/up8x25-minmax-wasmsimd-arm-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700765 "src/f32-dwconv/gen/up8x25-minmax-wasmsimd-arm.c",
Frank Barchard0725b8d2020-12-07 11:07:35 -0800766 "src/f32-dwconv/gen/up8x25-minmax-wasmsimd-x86-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700767 "src/f32-dwconv/gen/up8x25-minmax-wasmsimd-x86.c",
768 "src/f32-dwconv/gen/up8x25-wasmsimd.c",
Frank Barchard412e2f42020-12-11 11:40:50 -0800769 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-loadsplat-1x4-acc2.c",
770 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-loadsplat-1x4-acc3.c",
771 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-loadsplat-1x4-acc4.c",
772 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-loadsplat-1x4.c",
773 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-loadsplat-2x4-acc2.c",
774 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-loadsplat-2x4.c",
775 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-loadsplat-3x4.c",
776 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-loadsplat-4x4.c",
777 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-loadsplat-5x4.c",
778 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-loadsplat-6x4.c",
Frank Barchard02bb4292020-12-15 18:25:32 -0800779 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-splat-1x4-acc2.c",
780 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-splat-1x4-acc3.c",
781 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-splat-1x4-acc4.c",
782 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-splat-1x4.c",
783 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-splat-2x4-acc2.c",
784 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-splat-2x4.c",
785 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-splat-3x4.c",
786 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-splat-4x4.c",
787 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-splat-5x4.c",
788 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-splat-6x4.c",
Frank Barchard412e2f42020-12-11 11:40:50 -0800789 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-loadsplat-1x4-acc2.c",
790 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-loadsplat-1x4-acc3.c",
791 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-loadsplat-1x4-acc4.c",
792 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-loadsplat-1x4.c",
793 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-loadsplat-2x4-acc2.c",
794 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-loadsplat-2x4.c",
795 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-loadsplat-3x4.c",
796 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-loadsplat-4x4.c",
797 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-loadsplat-5x4.c",
798 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-loadsplat-6x4.c",
Frank Barchard02bb4292020-12-15 18:25:32 -0800799 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-splat-1x4-acc2.c",
800 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-splat-1x4-acc3.c",
801 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-splat-1x4-acc4.c",
802 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-splat-1x4.c",
803 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-splat-2x4-acc2.c",
804 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-splat-2x4.c",
805 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-splat-3x4.c",
806 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-splat-4x4.c",
807 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-splat-5x4.c",
808 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-splat-6x4.c",
Frank Barchardc5704bf2020-12-21 23:09:00 -0800809 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-arm-loadsplat-1x4-acc2.c",
810 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-arm-loadsplat-1x4-acc3.c",
811 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-arm-loadsplat-1x4-acc4.c",
812 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-arm-loadsplat-1x4.c",
813 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-arm-loadsplat-2x4-acc2.c",
814 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-arm-loadsplat-2x4.c",
815 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-arm-loadsplat-3x4.c",
816 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-arm-loadsplat-4x4.c",
Frank Barchard412e2f42020-12-11 11:40:50 -0800817 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-arm-splat-1x4-acc2.c",
818 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-arm-splat-1x4-acc3.c",
819 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-arm-splat-1x4-acc4.c",
820 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-arm-splat-1x4.c",
821 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-arm-splat-2x4-acc2.c",
822 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-arm-splat-2x4.c",
823 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-arm-splat-3x4.c",
824 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-arm-splat-4x4.c",
Frank Barchardcadd4222021-01-20 16:27:25 -0800825 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-x86-loadsplat-1x4-acc2.c",
826 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-x86-loadsplat-1x4-acc3.c",
827 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-x86-loadsplat-1x4-acc4.c",
828 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-x86-loadsplat-1x4.c",
829 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-x86-loadsplat-2x4-acc2.c",
830 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-x86-loadsplat-2x4.c",
831 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-x86-loadsplat-3x4.c",
832 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-x86-loadsplat-4x4.c",
Frank Barchard412e2f42020-12-11 11:40:50 -0800833 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-x86-splat-1x4-acc2.c",
834 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-x86-splat-1x4-acc3.c",
835 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-x86-splat-1x4-acc4.c",
836 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-x86-splat-1x4.c",
837 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-x86-splat-2x4-acc2.c",
838 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-x86-splat-2x4.c",
839 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-x86-splat-3x4.c",
840 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-x86-splat-4x4.c",
Frank Barchardb20dcd62020-12-15 16:46:14 -0800841 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-loadsplat-1x4-acc2.c",
842 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-loadsplat-1x4-acc3.c",
843 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-loadsplat-1x4-acc4.c",
844 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-loadsplat-1x4-acc5.c",
845 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-loadsplat-1x4.c",
846 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-loadsplat-2x4-acc2.c",
847 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-loadsplat-2x4-acc3.c",
848 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-loadsplat-2x4.c",
849 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-loadsplat-3x4-acc2.c",
850 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-loadsplat-3x4.c",
851 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-loadsplat-4x4-acc2.c",
852 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-loadsplat-4x4.c",
853 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-loadsplat-5x4.c",
Frank Barchard412e2f42020-12-11 11:40:50 -0800854 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-splat-1x4-acc2.c",
855 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-splat-1x4-acc3.c",
856 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-splat-1x4-acc4.c",
857 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-splat-1x4-acc5.c",
858 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-splat-1x4.c",
859 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-splat-2x4-acc2.c",
860 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-splat-2x4-acc3.c",
861 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-splat-2x4.c",
862 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-splat-3x4-acc2.c",
863 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-splat-3x4.c",
864 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-splat-4x4-acc2.c",
865 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-splat-4x4.c",
866 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-splat-5x4.c",
Frank Barchardb20dcd62020-12-15 16:46:14 -0800867 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-loadsplat-1x4-acc2.c",
868 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-loadsplat-1x4-acc3.c",
869 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-loadsplat-1x4-acc4.c",
870 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-loadsplat-1x4-acc5.c",
871 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-loadsplat-1x4.c",
872 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-loadsplat-2x4-acc2.c",
873 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-loadsplat-2x4-acc3.c",
874 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-loadsplat-2x4.c",
875 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-loadsplat-3x4-acc2.c",
876 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-loadsplat-3x4.c",
877 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-loadsplat-4x4-acc2.c",
878 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-loadsplat-4x4.c",
879 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-loadsplat-5x4.c",
Frank Barchard412e2f42020-12-11 11:40:50 -0800880 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-splat-1x4-acc2.c",
881 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-splat-1x4-acc3.c",
882 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-splat-1x4-acc4.c",
883 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-splat-1x4-acc5.c",
884 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-splat-1x4.c",
885 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-splat-2x4-acc2.c",
886 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-splat-2x4-acc3.c",
887 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-splat-2x4.c",
888 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-splat-3x4-acc2.c",
889 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-splat-3x4.c",
890 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-splat-4x4-acc2.c",
891 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-splat-4x4.c",
892 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-splat-5x4.c",
Frank Barchardc6889b32020-12-21 11:27:22 -0800893 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-arm-loadsplat-1x4-acc2.c",
894 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-arm-loadsplat-1x4-acc3.c",
895 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-arm-loadsplat-1x4-acc4.c",
896 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-arm-loadsplat-1x4-acc5.c",
897 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-arm-loadsplat-1x4.c",
898 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-arm-loadsplat-2x4-acc2.c",
899 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-arm-loadsplat-2x4-acc3.c",
900 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-arm-loadsplat-2x4.c",
901 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-arm-loadsplat-3x4-acc2.c",
902 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-arm-loadsplat-3x4.c",
Frank Barchard412e2f42020-12-11 11:40:50 -0800903 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-arm-splat-1x4-acc2.c",
904 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-arm-splat-1x4-acc3.c",
905 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-arm-splat-1x4-acc4.c",
906 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-arm-splat-1x4-acc5.c",
907 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-arm-splat-1x4.c",
908 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-arm-splat-2x4-acc2.c",
909 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-arm-splat-2x4-acc3.c",
910 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-arm-splat-2x4.c",
911 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-arm-splat-3x4-acc2.c",
912 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-arm-splat-3x4.c",
Frank Barchardc6889b32020-12-21 11:27:22 -0800913 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-x86-loadsplat-1x4-acc2.c",
914 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-x86-loadsplat-1x4-acc3.c",
915 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-x86-loadsplat-1x4-acc4.c",
916 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-x86-loadsplat-1x4-acc5.c",
917 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-x86-loadsplat-1x4.c",
918 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-x86-loadsplat-2x4-acc2.c",
919 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-x86-loadsplat-2x4-acc3.c",
920 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-x86-loadsplat-2x4.c",
921 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-x86-loadsplat-3x4-acc2.c",
922 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-x86-loadsplat-3x4.c",
Frank Barchard412e2f42020-12-11 11:40:50 -0800923 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-x86-splat-1x4-acc2.c",
924 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-x86-splat-1x4-acc3.c",
925 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-x86-splat-1x4-acc4.c",
926 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-x86-splat-1x4-acc5.c",
927 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-x86-splat-1x4.c",
928 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-x86-splat-2x4-acc2.c",
929 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-x86-splat-2x4-acc3.c",
930 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-x86-splat-2x4.c",
931 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-x86-splat-3x4-acc2.c",
932 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-x86-splat-3x4.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700933 "src/f32-gavgpool-cw/wasmsimd-arm-x4.c",
934 "src/f32-gavgpool-cw/wasmsimd-x86-x4.c",
Marat Dukhanc6016802020-07-16 18:51:28 -0700935 "src/f32-gavgpool/7p7x-minmax-wasmsimd-arm-c4.c",
936 "src/f32-gavgpool/7p7x-minmax-wasmsimd-x86-c4.c",
937 "src/f32-gavgpool/7x-minmax-wasmsimd-arm-c4.c",
938 "src/f32-gavgpool/7x-minmax-wasmsimd-x86-c4.c",
Frank Barchard0725b8d2020-12-07 11:07:35 -0800939 "src/f32-gemm/gen-inc/1x8inc-minmax-wasmsimd-arm-loadsplat.c",
940 "src/f32-gemm/gen-inc/1x8inc-minmax-wasmsimd-arm-splat.c",
941 "src/f32-gemm/gen-inc/1x8inc-minmax-wasmsimd-x86-loadsplat.c",
942 "src/f32-gemm/gen-inc/1x8inc-minmax-wasmsimd-x86-splat.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700943 "src/f32-gemm/gen-inc/1x8s4inc-minmax-wasmsimd-arm.c",
944 "src/f32-gemm/gen-inc/1x8s4inc-minmax-wasmsimd-x86.c",
Frank Barchard0725b8d2020-12-07 11:07:35 -0800945 "src/f32-gemm/gen-inc/3x8inc-minmax-wasmsimd-arm-loadsplat.c",
946 "src/f32-gemm/gen-inc/3x8inc-minmax-wasmsimd-arm-splat.c",
947 "src/f32-gemm/gen-inc/3x8inc-minmax-wasmsimd-x86-loadsplat.c",
948 "src/f32-gemm/gen-inc/3x8inc-minmax-wasmsimd-x86-splat.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700949 "src/f32-gemm/gen-inc/3x8s4inc-minmax-wasmsimd-arm.c",
950 "src/f32-gemm/gen-inc/3x8s4inc-minmax-wasmsimd-x86.c",
Frank Barchard0725b8d2020-12-07 11:07:35 -0800951 "src/f32-gemm/gen-inc/4x8inc-minmax-wasmsimd-arm-loadsplat.c",
952 "src/f32-gemm/gen-inc/4x8inc-minmax-wasmsimd-arm-splat.c",
953 "src/f32-gemm/gen-inc/4x8inc-minmax-wasmsimd-x86-loadsplat.c",
954 "src/f32-gemm/gen-inc/4x8inc-minmax-wasmsimd-x86-splat.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700955 "src/f32-gemm/gen-inc/4x8s4inc-minmax-wasmsimd-arm.c",
956 "src/f32-gemm/gen-inc/4x8s4inc-minmax-wasmsimd-x86.c",
Frank Barchard0725b8d2020-12-07 11:07:35 -0800957 "src/f32-gemm/gen-inc/5x8inc-minmax-wasmsimd-arm-loadsplat.c",
958 "src/f32-gemm/gen-inc/5x8inc-minmax-wasmsimd-arm-splat.c",
959 "src/f32-gemm/gen-inc/5x8inc-minmax-wasmsimd-x86-loadsplat.c",
960 "src/f32-gemm/gen-inc/5x8inc-minmax-wasmsimd-x86-splat.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700961 "src/f32-gemm/gen-inc/5x8s4inc-minmax-wasmsimd-arm.c",
962 "src/f32-gemm/gen-inc/5x8s4inc-minmax-wasmsimd-x86.c",
Frank Barchard0725b8d2020-12-07 11:07:35 -0800963 "src/f32-gemm/gen-inc/6x8inc-minmax-wasmsimd-arm-loadsplat.c",
964 "src/f32-gemm/gen-inc/6x8inc-minmax-wasmsimd-arm-splat.c",
965 "src/f32-gemm/gen-inc/6x8inc-minmax-wasmsimd-x86-loadsplat.c",
966 "src/f32-gemm/gen-inc/6x8inc-minmax-wasmsimd-x86-splat.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700967 "src/f32-gemm/gen-inc/6x8s4inc-minmax-wasmsimd-arm.c",
968 "src/f32-gemm/gen-inc/6x8s4inc-minmax-wasmsimd-x86.c",
Frank Barchard0725b8d2020-12-07 11:07:35 -0800969 "src/f32-gemm/gen/1x8-minmax-wasmsimd-arm-loadsplat.c",
970 "src/f32-gemm/gen/1x8-minmax-wasmsimd-arm-splat.c",
971 "src/f32-gemm/gen/1x8-minmax-wasmsimd-x86-loadsplat.c",
972 "src/f32-gemm/gen/1x8-minmax-wasmsimd-x86-splat.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700973 "src/f32-gemm/gen/1x8-relu-wasmsimd-splat.c",
974 "src/f32-gemm/gen/1x8-wasmsimd-splat.c",
975 "src/f32-gemm/gen/1x8s4-minmax-wasmsimd-arm.c",
976 "src/f32-gemm/gen/1x8s4-minmax-wasmsimd-x86.c",
Frank Barchard0725b8d2020-12-07 11:07:35 -0800977 "src/f32-gemm/gen/3x8-minmax-wasmsimd-arm-loadsplat.c",
978 "src/f32-gemm/gen/3x8-minmax-wasmsimd-arm-splat.c",
979 "src/f32-gemm/gen/3x8-minmax-wasmsimd-x86-loadsplat.c",
980 "src/f32-gemm/gen/3x8-minmax-wasmsimd-x86-splat.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700981 "src/f32-gemm/gen/3x8s4-minmax-wasmsimd-arm.c",
982 "src/f32-gemm/gen/3x8s4-minmax-wasmsimd-x86.c",
983 "src/f32-gemm/gen/4x2c4-minmax-wasmsimd-arm.c",
984 "src/f32-gemm/gen/4x2c4-minmax-wasmsimd-x86.c",
985 "src/f32-gemm/gen/4x2c4-relu-wasmsimd.c",
986 "src/f32-gemm/gen/4x2c4-wasmsimd.c",
Frank Barchard0725b8d2020-12-07 11:07:35 -0800987 "src/f32-gemm/gen/4x8-minmax-wasmsimd-arm-loadsplat.c",
988 "src/f32-gemm/gen/4x8-minmax-wasmsimd-arm-splat.c",
989 "src/f32-gemm/gen/4x8-minmax-wasmsimd-x86-loadsplat.c",
990 "src/f32-gemm/gen/4x8-minmax-wasmsimd-x86-splat.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700991 "src/f32-gemm/gen/4x8-relu-wasmsimd-splat.c",
992 "src/f32-gemm/gen/4x8-wasmsimd-splat.c",
993 "src/f32-gemm/gen/4x8s4-minmax-wasmsimd-arm.c",
994 "src/f32-gemm/gen/4x8s4-minmax-wasmsimd-x86.c",
Frank Barchard0725b8d2020-12-07 11:07:35 -0800995 "src/f32-gemm/gen/5x8-minmax-wasmsimd-arm-loadsplat.c",
996 "src/f32-gemm/gen/5x8-minmax-wasmsimd-arm-splat.c",
997 "src/f32-gemm/gen/5x8-minmax-wasmsimd-x86-loadsplat.c",
998 "src/f32-gemm/gen/5x8-minmax-wasmsimd-x86-splat.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700999 "src/f32-gemm/gen/5x8-relu-wasmsimd-splat.c",
1000 "src/f32-gemm/gen/5x8-wasmsimd-splat.c",
1001 "src/f32-gemm/gen/5x8s4-minmax-wasmsimd-arm.c",
1002 "src/f32-gemm/gen/5x8s4-minmax-wasmsimd-x86.c",
Frank Barchard0725b8d2020-12-07 11:07:35 -08001003 "src/f32-gemm/gen/6x8-minmax-wasmsimd-arm-loadsplat.c",
1004 "src/f32-gemm/gen/6x8-minmax-wasmsimd-arm-splat.c",
1005 "src/f32-gemm/gen/6x8-minmax-wasmsimd-x86-loadsplat.c",
1006 "src/f32-gemm/gen/6x8-minmax-wasmsimd-x86-splat.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001007 "src/f32-gemm/gen/6x8s4-minmax-wasmsimd-arm.c",
1008 "src/f32-gemm/gen/6x8s4-minmax-wasmsimd-x86.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07001009 "src/f32-hswish/gen/hswish-wasmsimd-x4.c",
1010 "src/f32-hswish/gen/hswish-wasmsimd-x8.c",
1011 "src/f32-hswish/gen/hswish-wasmsimd-x16.c",
XNNPACK Team965272b2020-10-23 21:10:15 -07001012 "src/f32-ibilinear-chw/gen/wasmsimd-p4.c",
1013 "src/f32-ibilinear-chw/gen/wasmsimd-p8.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001014 "src/f32-ibilinear/gen/wasmsimd-c4.c",
1015 "src/f32-ibilinear/gen/wasmsimd-c8.c",
Frank Barchard0725b8d2020-12-07 11:07:35 -08001016 "src/f32-igemm/gen/1x8-minmax-wasmsimd-arm-loadsplat.c",
1017 "src/f32-igemm/gen/1x8-minmax-wasmsimd-arm-splat.c",
1018 "src/f32-igemm/gen/1x8-minmax-wasmsimd-x86-loadsplat.c",
1019 "src/f32-igemm/gen/1x8-minmax-wasmsimd-x86-splat.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001020 "src/f32-igemm/gen/1x8-relu-wasmsimd-splat.c",
1021 "src/f32-igemm/gen/1x8-wasmsimd-splat.c",
1022 "src/f32-igemm/gen/1x8s4-minmax-wasmsimd-arm.c",
1023 "src/f32-igemm/gen/1x8s4-minmax-wasmsimd-x86.c",
Frank Barchard0725b8d2020-12-07 11:07:35 -08001024 "src/f32-igemm/gen/3x8-minmax-wasmsimd-arm-loadsplat.c",
1025 "src/f32-igemm/gen/3x8-minmax-wasmsimd-arm-splat.c",
1026 "src/f32-igemm/gen/3x8-minmax-wasmsimd-x86-loadsplat.c",
1027 "src/f32-igemm/gen/3x8-minmax-wasmsimd-x86-splat.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001028 "src/f32-igemm/gen/3x8s4-minmax-wasmsimd-arm.c",
1029 "src/f32-igemm/gen/3x8s4-minmax-wasmsimd-x86.c",
1030 "src/f32-igemm/gen/4x2c4-minmax-wasmsimd-arm.c",
1031 "src/f32-igemm/gen/4x2c4-minmax-wasmsimd-x86.c",
1032 "src/f32-igemm/gen/4x2c4-relu-wasmsimd.c",
1033 "src/f32-igemm/gen/4x2c4-wasmsimd.c",
Frank Barchard0725b8d2020-12-07 11:07:35 -08001034 "src/f32-igemm/gen/4x8-minmax-wasmsimd-arm-loadsplat.c",
1035 "src/f32-igemm/gen/4x8-minmax-wasmsimd-arm-splat.c",
1036 "src/f32-igemm/gen/4x8-minmax-wasmsimd-x86-loadsplat.c",
1037 "src/f32-igemm/gen/4x8-minmax-wasmsimd-x86-splat.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001038 "src/f32-igemm/gen/4x8-relu-wasmsimd-splat.c",
1039 "src/f32-igemm/gen/4x8-wasmsimd-splat.c",
1040 "src/f32-igemm/gen/4x8s4-minmax-wasmsimd-arm.c",
1041 "src/f32-igemm/gen/4x8s4-minmax-wasmsimd-x86.c",
Frank Barchard0725b8d2020-12-07 11:07:35 -08001042 "src/f32-igemm/gen/5x8-minmax-wasmsimd-arm-loadsplat.c",
1043 "src/f32-igemm/gen/5x8-minmax-wasmsimd-arm-splat.c",
1044 "src/f32-igemm/gen/5x8-minmax-wasmsimd-x86-loadsplat.c",
1045 "src/f32-igemm/gen/5x8-minmax-wasmsimd-x86-splat.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001046 "src/f32-igemm/gen/5x8-relu-wasmsimd-splat.c",
1047 "src/f32-igemm/gen/5x8-wasmsimd-splat.c",
1048 "src/f32-igemm/gen/5x8s4-minmax-wasmsimd-arm.c",
1049 "src/f32-igemm/gen/5x8s4-minmax-wasmsimd-x86.c",
Frank Barchard0725b8d2020-12-07 11:07:35 -08001050 "src/f32-igemm/gen/6x8-minmax-wasmsimd-arm-loadsplat.c",
1051 "src/f32-igemm/gen/6x8-minmax-wasmsimd-arm-splat.c",
1052 "src/f32-igemm/gen/6x8-minmax-wasmsimd-x86-loadsplat.c",
1053 "src/f32-igemm/gen/6x8-minmax-wasmsimd-x86-splat.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001054 "src/f32-igemm/gen/6x8s4-minmax-wasmsimd-arm.c",
1055 "src/f32-igemm/gen/6x8s4-minmax-wasmsimd-x86.c",
Marat Dukhanf6e24802020-07-08 22:20:40 -07001056 "src/f32-maxpool/9p8x-minmax-wasmsimd-arm-c4.c",
1057 "src/f32-maxpool/9p8x-minmax-wasmsimd-x86-c4.c",
Marat Dukhan1483c532020-07-16 18:08:19 -07001058 "src/f32-pavgpool/9p8x-minmax-wasmsimd-arm-c4.c",
1059 "src/f32-pavgpool/9p8x-minmax-wasmsimd-x86-c4.c",
1060 "src/f32-pavgpool/9x-minmax-wasmsimd-arm-c4.c",
1061 "src/f32-pavgpool/9x-minmax-wasmsimd-x86-c4.c",
Frank Barchard0725b8d2020-12-07 11:07:35 -08001062 "src/f32-ppmm/gen/4x8-minmax-wasmsimd-arm-splat.c",
1063 "src/f32-ppmm/gen/4x8-minmax-wasmsimd-x86-splat.c",
Frank Barcharda5316982020-07-23 13:19:28 -07001064 "src/f32-prelu/gen/wasmsimd-bitselect-1x4.c",
1065 "src/f32-prelu/gen/wasmsimd-bitselect-1x8.c",
1066 "src/f32-prelu/gen/wasmsimd-bitselect-1x16.c",
Marat Dukhan195f8eb2020-06-25 12:50:57 -07001067 "src/f32-prelu/gen/wasmsimd-bitselect-2x4.c",
1068 "src/f32-prelu/gen/wasmsimd-bitselect-2x8.c",
Frank Barcharda5316982020-07-23 13:19:28 -07001069 "src/f32-prelu/gen/wasmsimd-bitselect-2x16.c",
1070 "src/f32-prelu/gen/wasmsimd-bitselect-4x4.c",
1071 "src/f32-prelu/gen/wasmsimd-bitselect-4x8.c",
1072 "src/f32-prelu/gen/wasmsimd-bitselect-4x16.c",
1073 "src/f32-prelu/gen/wasmsimd-minmax-1x4.c",
1074 "src/f32-prelu/gen/wasmsimd-minmax-1x8.c",
1075 "src/f32-prelu/gen/wasmsimd-minmax-1x16.c",
Marat Dukhan195f8eb2020-06-25 12:50:57 -07001076 "src/f32-prelu/gen/wasmsimd-minmax-2x4.c",
1077 "src/f32-prelu/gen/wasmsimd-minmax-2x8.c",
Frank Barcharda5316982020-07-23 13:19:28 -07001078 "src/f32-prelu/gen/wasmsimd-minmax-2x16.c",
1079 "src/f32-prelu/gen/wasmsimd-minmax-4x4.c",
1080 "src/f32-prelu/gen/wasmsimd-minmax-4x8.c",
1081 "src/f32-prelu/gen/wasmsimd-minmax-4x16.c",
Marat Dukhan52238f02020-07-16 15:30:28 -07001082 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x4.c",
Marat Dukhan52238f02020-07-16 15:30:28 -07001083 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x8-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001084 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x8.c",
Marat Dukhan52238f02020-07-16 15:30:28 -07001085 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x12-acc2.c",
1086 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x12-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001087 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x12.c",
Marat Dukhan52238f02020-07-16 15:30:28 -07001088 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x16-acc2.c",
1089 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x16-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001090 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x16.c",
Marat Dukhan52238f02020-07-16 15:30:28 -07001091 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x20-acc2.c",
1092 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x20-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001093 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x20.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07001094 "src/f32-relu/gen/wasmsimd-x4.c",
1095 "src/f32-relu/gen/wasmsimd-x8.c",
Frank Barchard37297a62020-08-27 12:15:10 -07001096 "src/f32-relu/gen/wasmsimd-x16.c",
Marat Dukhan8c417962020-07-08 12:27:50 -07001097 "src/f32-rmax/wasmsimd-arm.c",
1098 "src/f32-rmax/wasmsimd-x86.c",
Marat Dukhand187a5b2020-07-20 01:07:17 -07001099 "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x4.c",
1100 "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x8.c",
1101 "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x12.c",
1102 "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x16.c",
1103 "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x20.c",
1104 "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x24.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001105 "src/f32-sigmoid/gen/wasmsimd-p5-div-x4.c",
1106 "src/f32-sigmoid/gen/wasmsimd-p5-div-x8.c",
1107 "src/f32-sigmoid/gen/wasmsimd-p5-div-x12.c",
1108 "src/f32-sigmoid/gen/wasmsimd-p5-div-x16.c",
1109 "src/f32-sigmoid/gen/wasmsimd-p5-div-x20.c",
1110 "src/f32-sigmoid/gen/wasmsimd-p5-div-x24.c",
Frank Barchard8ef44cd2020-11-03 12:30:23 -08001111 "src/f32-spmm/gen/4x1-minmax-wasmsimd-arm-pipelined-x2.c",
1112 "src/f32-spmm/gen/4x1-minmax-wasmsimd-arm-pipelined.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07001113 "src/f32-spmm/gen/4x1-minmax-wasmsimd-arm-x2.c",
1114 "src/f32-spmm/gen/4x1-minmax-wasmsimd-arm-x4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001115 "src/f32-spmm/gen/4x1-minmax-wasmsimd-arm.c",
Frank Barchard8ef44cd2020-11-03 12:30:23 -08001116 "src/f32-spmm/gen/4x1-minmax-wasmsimd-x86-pipelined-x2.c",
1117 "src/f32-spmm/gen/4x1-minmax-wasmsimd-x86-pipelined.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07001118 "src/f32-spmm/gen/4x1-minmax-wasmsimd-x86-x2.c",
1119 "src/f32-spmm/gen/4x1-minmax-wasmsimd-x86-x4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001120 "src/f32-spmm/gen/4x1-minmax-wasmsimd-x86.c",
Frank Barchard8ef44cd2020-11-03 12:30:23 -08001121 "src/f32-spmm/gen/8x1-minmax-wasmsimd-arm-pipelined-x2.c",
1122 "src/f32-spmm/gen/8x1-minmax-wasmsimd-arm-pipelined.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07001123 "src/f32-spmm/gen/8x1-minmax-wasmsimd-arm-x2.c",
1124 "src/f32-spmm/gen/8x1-minmax-wasmsimd-arm-x4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001125 "src/f32-spmm/gen/8x1-minmax-wasmsimd-arm.c",
Frank Barchard8ef44cd2020-11-03 12:30:23 -08001126 "src/f32-spmm/gen/8x1-minmax-wasmsimd-x86-pipelined-x2.c",
1127 "src/f32-spmm/gen/8x1-minmax-wasmsimd-x86-pipelined.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07001128 "src/f32-spmm/gen/8x1-minmax-wasmsimd-x86-x2.c",
1129 "src/f32-spmm/gen/8x1-minmax-wasmsimd-x86-x4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001130 "src/f32-spmm/gen/8x1-minmax-wasmsimd-x86.c",
Frank Barchard8ef44cd2020-11-03 12:30:23 -08001131 "src/f32-spmm/gen/16x1-minmax-wasmsimd-arm-pipelined-x2.c",
1132 "src/f32-spmm/gen/16x1-minmax-wasmsimd-arm-pipelined.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07001133 "src/f32-spmm/gen/16x1-minmax-wasmsimd-arm-x2.c",
1134 "src/f32-spmm/gen/16x1-minmax-wasmsimd-arm-x4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001135 "src/f32-spmm/gen/16x1-minmax-wasmsimd-arm.c",
Frank Barchard8ef44cd2020-11-03 12:30:23 -08001136 "src/f32-spmm/gen/16x1-minmax-wasmsimd-x86-pipelined-x2.c",
1137 "src/f32-spmm/gen/16x1-minmax-wasmsimd-x86-pipelined.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07001138 "src/f32-spmm/gen/16x1-minmax-wasmsimd-x86-x2.c",
1139 "src/f32-spmm/gen/16x1-minmax-wasmsimd-x86-x4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001140 "src/f32-spmm/gen/16x1-minmax-wasmsimd-x86.c",
Frank Barchard8ef44cd2020-11-03 12:30:23 -08001141 "src/f32-spmm/gen/32x1-minmax-wasmsimd-arm-pipelined-x2.c",
1142 "src/f32-spmm/gen/32x1-minmax-wasmsimd-arm-pipelined.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07001143 "src/f32-spmm/gen/32x1-minmax-wasmsimd-arm-x2.c",
1144 "src/f32-spmm/gen/32x1-minmax-wasmsimd-arm-x4.c",
Frank Barchard846c0c62020-10-26 15:01:39 -07001145 "src/f32-spmm/gen/32x1-minmax-wasmsimd-arm.c",
Frank Barchard8ef44cd2020-11-03 12:30:23 -08001146 "src/f32-spmm/gen/32x1-minmax-wasmsimd-x86-pipelined-x2.c",
1147 "src/f32-spmm/gen/32x1-minmax-wasmsimd-x86-pipelined.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07001148 "src/f32-spmm/gen/32x1-minmax-wasmsimd-x86-x2.c",
1149 "src/f32-spmm/gen/32x1-minmax-wasmsimd-x86-x4.c",
Frank Barchard846c0c62020-10-26 15:01:39 -07001150 "src/f32-spmm/gen/32x1-minmax-wasmsimd-x86.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -07001151 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-arm-x4.c",
1152 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001153 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-arm-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -07001154 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-x86-x4.c",
1155 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001156 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -07001157 "src/f32-vbinary/gen/vadd-relu-wasmsimd-x4.c",
1158 "src/f32-vbinary/gen/vadd-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001159 "src/f32-vbinary/gen/vadd-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07001160 "src/f32-vbinary/gen/vadd-wasmsimd-x4.c",
1161 "src/f32-vbinary/gen/vadd-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001162 "src/f32-vbinary/gen/vadd-wasmsimd-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001163 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-arm-x4.c",
1164 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001165 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-arm-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001166 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-x86-x4.c",
1167 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001168 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -07001169 "src/f32-vbinary/gen/vaddc-relu-wasmsimd-x4.c",
1170 "src/f32-vbinary/gen/vaddc-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001171 "src/f32-vbinary/gen/vaddc-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07001172 "src/f32-vbinary/gen/vaddc-wasmsimd-x4.c",
1173 "src/f32-vbinary/gen/vaddc-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001174 "src/f32-vbinary/gen/vaddc-wasmsimd-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -07001175 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-arm-x4.c",
1176 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001177 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-arm-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -07001178 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-x86-x4.c",
1179 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001180 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -07001181 "src/f32-vbinary/gen/vdiv-relu-wasmsimd-x4.c",
1182 "src/f32-vbinary/gen/vdiv-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001183 "src/f32-vbinary/gen/vdiv-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07001184 "src/f32-vbinary/gen/vdiv-wasmsimd-x4.c",
1185 "src/f32-vbinary/gen/vdiv-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001186 "src/f32-vbinary/gen/vdiv-wasmsimd-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001187 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-arm-x4.c",
1188 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001189 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-arm-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001190 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-x86-x4.c",
1191 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001192 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -07001193 "src/f32-vbinary/gen/vdivc-relu-wasmsimd-x4.c",
1194 "src/f32-vbinary/gen/vdivc-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001195 "src/f32-vbinary/gen/vdivc-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07001196 "src/f32-vbinary/gen/vdivc-wasmsimd-x4.c",
1197 "src/f32-vbinary/gen/vdivc-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001198 "src/f32-vbinary/gen/vdivc-wasmsimd-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -07001199 "src/f32-vbinary/gen/vmax-wasmsimd-arm-x4.c",
1200 "src/f32-vbinary/gen/vmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001201 "src/f32-vbinary/gen/vmax-wasmsimd-arm-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -07001202 "src/f32-vbinary/gen/vmax-wasmsimd-x86-x4.c",
1203 "src/f32-vbinary/gen/vmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001204 "src/f32-vbinary/gen/vmax-wasmsimd-x86-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001205 "src/f32-vbinary/gen/vmaxc-wasmsimd-arm-x4.c",
1206 "src/f32-vbinary/gen/vmaxc-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001207 "src/f32-vbinary/gen/vmaxc-wasmsimd-arm-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001208 "src/f32-vbinary/gen/vmaxc-wasmsimd-x86-x4.c",
1209 "src/f32-vbinary/gen/vmaxc-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001210 "src/f32-vbinary/gen/vmaxc-wasmsimd-x86-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -07001211 "src/f32-vbinary/gen/vmin-wasmsimd-arm-x4.c",
1212 "src/f32-vbinary/gen/vmin-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001213 "src/f32-vbinary/gen/vmin-wasmsimd-arm-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -07001214 "src/f32-vbinary/gen/vmin-wasmsimd-x86-x4.c",
1215 "src/f32-vbinary/gen/vmin-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001216 "src/f32-vbinary/gen/vmin-wasmsimd-x86-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001217 "src/f32-vbinary/gen/vminc-wasmsimd-arm-x4.c",
1218 "src/f32-vbinary/gen/vminc-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001219 "src/f32-vbinary/gen/vminc-wasmsimd-arm-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001220 "src/f32-vbinary/gen/vminc-wasmsimd-x86-x4.c",
1221 "src/f32-vbinary/gen/vminc-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001222 "src/f32-vbinary/gen/vminc-wasmsimd-x86-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -07001223 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-arm-x4.c",
1224 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001225 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-arm-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -07001226 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-x86-x4.c",
1227 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001228 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -07001229 "src/f32-vbinary/gen/vmul-relu-wasmsimd-x4.c",
1230 "src/f32-vbinary/gen/vmul-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001231 "src/f32-vbinary/gen/vmul-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07001232 "src/f32-vbinary/gen/vmul-wasmsimd-x4.c",
1233 "src/f32-vbinary/gen/vmul-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001234 "src/f32-vbinary/gen/vmul-wasmsimd-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001235 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-arm-x4.c",
1236 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001237 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-arm-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001238 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-x86-x4.c",
1239 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001240 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -07001241 "src/f32-vbinary/gen/vmulc-relu-wasmsimd-x4.c",
1242 "src/f32-vbinary/gen/vmulc-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001243 "src/f32-vbinary/gen/vmulc-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07001244 "src/f32-vbinary/gen/vmulc-wasmsimd-x4.c",
1245 "src/f32-vbinary/gen/vmulc-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001246 "src/f32-vbinary/gen/vmulc-wasmsimd-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001247 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-arm-x4.c",
1248 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001249 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-arm-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001250 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-x86-x4.c",
1251 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001252 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -07001253 "src/f32-vbinary/gen/vrdivc-relu-wasmsimd-x4.c",
1254 "src/f32-vbinary/gen/vrdivc-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001255 "src/f32-vbinary/gen/vrdivc-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07001256 "src/f32-vbinary/gen/vrdivc-wasmsimd-x4.c",
1257 "src/f32-vbinary/gen/vrdivc-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001258 "src/f32-vbinary/gen/vrdivc-wasmsimd-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001259 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-arm-x4.c",
1260 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001261 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-arm-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001262 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-x86-x4.c",
1263 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001264 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -07001265 "src/f32-vbinary/gen/vrsubc-relu-wasmsimd-x4.c",
1266 "src/f32-vbinary/gen/vrsubc-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001267 "src/f32-vbinary/gen/vrsubc-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07001268 "src/f32-vbinary/gen/vrsubc-wasmsimd-x4.c",
1269 "src/f32-vbinary/gen/vrsubc-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001270 "src/f32-vbinary/gen/vrsubc-wasmsimd-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -07001271 "src/f32-vbinary/gen/vsqrdiff-wasmsimd-x4.c",
1272 "src/f32-vbinary/gen/vsqrdiff-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001273 "src/f32-vbinary/gen/vsqrdiff-wasmsimd-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001274 "src/f32-vbinary/gen/vsqrdiffc-wasmsimd-x4.c",
1275 "src/f32-vbinary/gen/vsqrdiffc-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001276 "src/f32-vbinary/gen/vsqrdiffc-wasmsimd-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -07001277 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-arm-x4.c",
1278 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001279 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-arm-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -07001280 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-x86-x4.c",
1281 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001282 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -07001283 "src/f32-vbinary/gen/vsub-relu-wasmsimd-x4.c",
1284 "src/f32-vbinary/gen/vsub-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001285 "src/f32-vbinary/gen/vsub-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07001286 "src/f32-vbinary/gen/vsub-wasmsimd-x4.c",
1287 "src/f32-vbinary/gen/vsub-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001288 "src/f32-vbinary/gen/vsub-wasmsimd-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001289 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-arm-x4.c",
1290 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001291 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-arm-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001292 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-x86-x4.c",
1293 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001294 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -07001295 "src/f32-vbinary/gen/vsubc-relu-wasmsimd-x4.c",
1296 "src/f32-vbinary/gen/vsubc-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001297 "src/f32-vbinary/gen/vsubc-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07001298 "src/f32-vbinary/gen/vsubc-wasmsimd-x4.c",
1299 "src/f32-vbinary/gen/vsubc-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001300 "src/f32-vbinary/gen/vsubc-wasmsimd-x16.c",
Marat Dukhaned6baaf2020-12-01 15:07:08 -08001301 "src/f32-velu/gen/velu-wasmsimd-arm-rr2-lut16-p3-x4.c",
1302 "src/f32-velu/gen/velu-wasmsimd-arm-rr2-lut16-p3-x8.c",
1303 "src/f32-velu/gen/velu-wasmsimd-arm-rr2-lut16-p3-x12.c",
1304 "src/f32-velu/gen/velu-wasmsimd-arm-rr2-lut16-p3-x16.c",
1305 "src/f32-velu/gen/velu-wasmsimd-arm-rr2-lut16-p3-x20.c",
1306 "src/f32-velu/gen/velu-wasmsimd-arm-rr2-lut16-p3-x24.c",
Marat Dukhaned6baaf2020-12-01 15:07:08 -08001307 "src/f32-velu/gen/velu-wasmsimd-arm-rr2-p6-x4.c",
1308 "src/f32-velu/gen/velu-wasmsimd-arm-rr2-p6-x8.c",
1309 "src/f32-velu/gen/velu-wasmsimd-arm-rr2-p6-x12.c",
1310 "src/f32-velu/gen/velu-wasmsimd-arm-rr2-p6-x16.c",
1311 "src/f32-velu/gen/velu-wasmsimd-arm-rr2-p6-x20.c",
1312 "src/f32-velu/gen/velu-wasmsimd-arm-rr2-p6-x24.c",
Frank Barcharde7223ee2020-12-04 19:04:01 -08001313 "src/f32-velu/gen/velu-wasmsimd-x86-rr2-lut16-p3-x4.c",
1314 "src/f32-velu/gen/velu-wasmsimd-x86-rr2-lut16-p3-x8.c",
1315 "src/f32-velu/gen/velu-wasmsimd-x86-rr2-lut16-p3-x12.c",
1316 "src/f32-velu/gen/velu-wasmsimd-x86-rr2-lut16-p3-x16.c",
1317 "src/f32-velu/gen/velu-wasmsimd-x86-rr2-lut16-p3-x20.c",
1318 "src/f32-velu/gen/velu-wasmsimd-x86-rr2-lut16-p3-x24.c",
Marat Dukhaned6baaf2020-12-01 15:07:08 -08001319 "src/f32-velu/gen/velu-wasmsimd-x86-rr2-p6-x4.c",
1320 "src/f32-velu/gen/velu-wasmsimd-x86-rr2-p6-x8.c",
1321 "src/f32-velu/gen/velu-wasmsimd-x86-rr2-p6-x12.c",
1322 "src/f32-velu/gen/velu-wasmsimd-x86-rr2-p6-x16.c",
1323 "src/f32-velu/gen/velu-wasmsimd-x86-rr2-p6-x20.c",
1324 "src/f32-velu/gen/velu-wasmsimd-x86-rr2-p6-x24.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07001325 "src/f32-vlrelu/gen/vlrelu-wasmsimd-bitselect-x4.c",
1326 "src/f32-vlrelu/gen/vlrelu-wasmsimd-bitselect-x8.c",
1327 "src/f32-vlrelu/gen/vlrelu-wasmsimd-minmax-x4.c",
1328 "src/f32-vlrelu/gen/vlrelu-wasmsimd-minmax-x8.c",
Marat Dukhand816f622020-07-15 10:14:39 -07001329 "src/f32-vmulcaddc/gen/c4-minmax-wasmsimd-arm-2x.c",
Marat Dukhand816f622020-07-15 10:14:39 -07001330 "src/f32-vmulcaddc/gen/c4-minmax-wasmsimd-x86-2x.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001331 "src/f32-vmulcaddc/gen/c8-minmax-wasmsimd-arm-2x.c",
Marat Dukhand816f622020-07-15 10:14:39 -07001332 "src/f32-vmulcaddc/gen/c8-minmax-wasmsimd-x86-2x.c",
Marat Dukhanb82b2cd2020-07-16 02:23:42 -07001333 "src/f32-vrnd/gen/vrndd-wasmsimd-addsub-x4.c",
1334 "src/f32-vrnd/gen/vrndd-wasmsimd-addsub-x8.c",
1335 "src/f32-vrnd/gen/vrndd-wasmsimd-cvt-x4.c",
1336 "src/f32-vrnd/gen/vrndd-wasmsimd-cvt-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001337 "src/f32-vrnd/gen/vrndne-wasmsimd-addsub-x4.c",
1338 "src/f32-vrnd/gen/vrndne-wasmsimd-addsub-x8.c",
1339 "src/f32-vrnd/gen/vrndu-wasmsimd-addsub-x4.c",
1340 "src/f32-vrnd/gen/vrndu-wasmsimd-addsub-x8.c",
1341 "src/f32-vrnd/gen/vrndu-wasmsimd-cvt-x4.c",
1342 "src/f32-vrnd/gen/vrndu-wasmsimd-cvt-x8.c",
1343 "src/f32-vrnd/gen/vrndz-wasmsimd-addsub-x4.c",
1344 "src/f32-vrnd/gen/vrndz-wasmsimd-addsub-x8.c",
1345 "src/f32-vrnd/gen/vrndz-wasmsimd-cvt-x4.c",
1346 "src/f32-vrnd/gen/vrndz-wasmsimd-cvt-x8.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07001347 "src/f32-vsqrt/gen/wasmsimd-sqrt-x4.c",
1348 "src/f32-vsqrt/gen/wasmsimd-sqrt-x8.c",
Marat Dukhan37c83512020-06-29 13:25:53 -07001349 "src/f32-vunary/gen/vabs-wasmsimd-x4.c",
1350 "src/f32-vunary/gen/vabs-wasmsimd-x8.c",
1351 "src/f32-vunary/gen/vneg-wasmsimd-x4.c",
1352 "src/f32-vunary/gen/vneg-wasmsimd-x8.c",
1353 "src/f32-vunary/gen/vsqr-wasmsimd-x4.c",
1354 "src/f32-vunary/gen/vsqr-wasmsimd-x8.c",
Marat Dukhande390d42020-11-29 19:32:18 -08001355 "src/math/expm1minus-wasmsimd-rr2-lut16-p3-andnot.c",
1356 "src/math/expm1minus-wasmsimd-rr2-lut16-p3-max.c",
1357 "src/math/expm1minus-wasmsimd-rr2-p6-andnot.c",
1358 "src/math/expm1minus-wasmsimd-rr2-p6-max.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001359 "src/math/roundd-wasmsimd-addsub.c",
1360 "src/math/roundd-wasmsimd-cvt.c",
1361 "src/math/roundne-wasmsimd-addsub.c",
1362 "src/math/roundu-wasmsimd-addsub.c",
1363 "src/math/roundu-wasmsimd-cvt.c",
1364 "src/math/roundz-wasmsimd-addsub.c",
1365 "src/math/roundz-wasmsimd-cvt.c",
1366 "src/math/sigmoid-wasmsimd-rr2-lut64-p2-div.c",
1367 "src/math/sigmoid-wasmsimd-rr2-p5-div.c",
Marat Dukhancc8f34c2020-08-05 16:36:38 -07001368 "src/qs8-dwconv/gen/up8x9-minmax-wasmsimd-mul16.c",
1369 "src/qs8-dwconv/gen/up16x9-minmax-wasmsimd-mul16.c",
1370 "src/qs8-dwconv/gen/up24x9-minmax-wasmsimd-mul16.c",
Marat Dukhanb5e3d172020-08-06 13:29:53 -07001371 "src/qs8-gavgpool/gen/7p7x-minmax-wasmsimd-c8-acc2.c",
1372 "src/qs8-gavgpool/gen/7p7x-minmax-wasmsimd-c16-acc2.c",
1373 "src/qs8-gavgpool/gen/7p7x-minmax-wasmsimd-c24-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001374 "src/qs8-gavgpool/gen/7x-minmax-wasmsimd-c8-acc2.c",
1375 "src/qs8-gavgpool/gen/7x-minmax-wasmsimd-c16-acc2.c",
1376 "src/qs8-gavgpool/gen/7x-minmax-wasmsimd-c24-acc2.c",
Marat Dukhan27203da2020-08-05 15:19:03 -07001377 "src/qs8-gemm/gen/1x4c8-minmax-wasmsimd-ld64.c",
Marat Dukhan27203da2020-08-05 15:19:03 -07001378 "src/qs8-gemm/gen/1x4c8-minmax-wasmsimd-ld128.c",
Marat Dukhan27203da2020-08-05 15:19:03 -07001379 "src/qs8-gemm/gen/1x4c8-xw-minmax-wasmsimd.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001380 "src/qs8-gemm/gen/2x4c8-minmax-wasmsimd-ld64.c",
1381 "src/qs8-gemm/gen/2x4c8-minmax-wasmsimd-ld128.c",
Marat Dukhan27203da2020-08-05 15:19:03 -07001382 "src/qs8-gemm/gen/2x4c8-xw-minmax-wasmsimd.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001383 "src/qs8-gemm/gen/3x4c8-minmax-wasmsimd-ld64.c",
1384 "src/qs8-gemm/gen/3x4c8-minmax-wasmsimd-ld128.c",
Marat Dukhan27203da2020-08-05 15:19:03 -07001385 "src/qs8-gemm/gen/3x4c8-xw-minmax-wasmsimd.c",
1386 "src/qs8-igemm/gen/1x4c8-minmax-wasmsimd-ld64.c",
Marat Dukhan27203da2020-08-05 15:19:03 -07001387 "src/qs8-igemm/gen/1x4c8-minmax-wasmsimd-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001388 "src/qs8-igemm/gen/2x4c8-minmax-wasmsimd-ld64.c",
Marat Dukhan27203da2020-08-05 15:19:03 -07001389 "src/qs8-igemm/gen/2x4c8-minmax-wasmsimd-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001390 "src/qs8-igemm/gen/3x4c8-minmax-wasmsimd-ld64.c",
Marat Dukhan27203da2020-08-05 15:19:03 -07001391 "src/qs8-igemm/gen/3x4c8-minmax-wasmsimd-ld128.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -07001392 "src/qs8-requantization/fp32-wasmsimd.c",
Marat Dukhan138560c2020-08-03 18:57:34 -07001393 "src/qs8-requantization/q31-wasmsimd.c",
Marat Dukhan5df27f82020-09-02 23:59:21 -07001394 "src/qs8-vadd/gen/minmax-wasmsimd-x8.c",
1395 "src/qs8-vadd/gen/minmax-wasmsimd-x16.c",
1396 "src/qs8-vadd/gen/minmax-wasmsimd-x24.c",
1397 "src/qs8-vadd/gen/minmax-wasmsimd-x32.c",
1398 "src/qs8-vaddc/gen/minmax-wasmsimd-x8.c",
1399 "src/qs8-vaddc/gen/minmax-wasmsimd-x16.c",
1400 "src/qs8-vaddc/gen/minmax-wasmsimd-x24.c",
1401 "src/qs8-vaddc/gen/minmax-wasmsimd-x32.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07001402 "src/qu8-requantization/fp32-wasmsimd.c",
Marat Dukhan22de5e72020-08-03 21:53:18 -07001403 "src/qu8-requantization/q31-wasmsimd.c",
Marat Dukhan8ee37012020-07-16 13:17:13 -07001404 "src/x32-fill/wasmsimd.c",
Marat Dukhan66d99e92020-07-16 12:56:21 -07001405 "src/x32-packx/x4-wasmsimd.c",
Marat Dukhan9306ae02020-07-16 15:51:13 -07001406 "src/x32-pad/wasmsimd.c",
Marat Dukhan9d4bfa22020-07-16 19:07:04 -07001407 "src/x32-unpool/wasmsimd.c",
Marat Dukhane3b78762020-07-16 20:02:58 -07001408 "src/x32-zip/x2-wasmsimd.c",
1409 "src/x32-zip/x3-wasmsimd.c",
1410 "src/x32-zip/x4-wasmsimd.c",
1411 "src/x32-zip/xm-wasmsimd.c",
Marat Dukhan290055c2020-06-09 12:24:29 -07001412]
1413
Marat Dukhan08c4a432019-10-03 09:29:21 -07001414# ISA-specific micro-kernels
1415NEON_UKERNELS = [
Marat Dukhanef25c6d2020-07-24 00:59:40 -07001416 "src/f32-argmaxpool/4x-neon-c4.c",
1417 "src/f32-argmaxpool/9p8x-neon-c4.c",
1418 "src/f32-argmaxpool/9x-neon-c4.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001419 "src/f32-avgpool/9p8x-minmax-neon-c4.c",
1420 "src/f32-avgpool/9x-minmax-neon-c4.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001421 "src/f32-clamp/gen/neon-x4.c",
1422 "src/f32-clamp/gen/neon-x8.c",
Marat Dukhan56b10cd2020-05-18 09:35:49 -07001423 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neon-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001424 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neon-2x2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001425 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neon-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001426 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neon-2x2.c",
Marat Dukhan56b10cd2020-05-18 09:35:49 -07001427 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neon-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001428 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neon-2x2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001429 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neon-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001430 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neon-2x2.c",
Marat Dukhanc7634882020-12-07 15:11:12 -08001431 "src/f32-conv-hwc2chw/3x3s2p1c3x4-neon-2x2.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -07001432 "src/f32-dwconv/gen/up4x4-minmax-neon-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001433 "src/f32-dwconv/gen/up4x4-minmax-neon.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001434 "src/f32-dwconv/gen/up4x9-minmax-neon-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001435 "src/f32-dwconv/gen/up4x9-minmax-neon.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -07001436 "src/f32-dwconv/gen/up4x25-minmax-neon-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001437 "src/f32-dwconv/gen/up4x25-minmax-neon.c",
1438 "src/f32-dwconv/gen/up8x4-minmax-neon-acc2.c",
1439 "src/f32-dwconv/gen/up8x4-minmax-neon.c",
1440 "src/f32-dwconv/gen/up8x9-minmax-neon-acc2.c",
1441 "src/f32-dwconv/gen/up8x9-minmax-neon.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -07001442 "src/f32-dwconv/gen/up8x25-minmax-neon-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001443 "src/f32-dwconv/gen/up8x25-minmax-neon.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001444 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-1x4-acc2.c",
1445 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-1x4-acc3.c",
1446 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-1x4-acc4.c",
Marat Dukhanc581e482020-10-24 01:28:11 -07001447 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-1x4.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001448 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-2x4-acc2.c",
Marat Dukhanc581e482020-10-24 01:28:11 -07001449 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-2x4.c",
1450 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-3x4.c",
1451 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-4x4.c",
1452 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-5x4.c",
1453 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-6x4.c",
Marat Dukhan82f0c322020-10-25 19:17:35 -07001454 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neon-1x4-acc2.c",
1455 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neon-1x4-acc3.c",
1456 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neon-1x4-acc4.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001457 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neon-1x4.c",
Marat Dukhan82f0c322020-10-25 19:17:35 -07001458 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neon-2x4-acc2.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001459 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neon-2x4.c",
1460 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neon-3x4.c",
1461 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neon-4x4.c",
Marat Dukhan149f0ea2020-10-26 12:50:33 -07001462 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-1x4-acc2.c",
1463 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-1x4-acc3.c",
1464 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-1x4-acc4.c",
1465 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-1x4-acc5.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001466 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-1x4.c",
Marat Dukhan149f0ea2020-10-26 12:50:33 -07001467 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-2x4-acc2.c",
1468 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-2x4-acc3.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001469 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-2x4.c",
Marat Dukhan149f0ea2020-10-26 12:50:33 -07001470 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-3x4-acc2.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001471 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-3x4.c",
Marat Dukhan149f0ea2020-10-26 12:50:33 -07001472 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-4x4-acc2.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001473 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-4x4.c",
1474 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-5x4.c",
Marat Dukhan30d4b252020-10-29 16:33:22 -07001475 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neon-1x4-acc2.c",
1476 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neon-1x4-acc3.c",
1477 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neon-1x4-acc4.c",
1478 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neon-1x4-acc5.c",
1479 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neon-1x4.c",
1480 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neon-2x4-acc2.c",
1481 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neon-2x4-acc3.c",
1482 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neon-2x4.c",
Marat Dukhan30d4b252020-10-29 16:33:22 -07001483 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neon-3x4-acc2.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07001484 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neon-3x4.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -07001485 "src/f32-gavgpool-cw/neon-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001486 "src/f32-gavgpool/7p7x-minmax-neon-c4.c",
1487 "src/f32-gavgpool/7x-minmax-neon-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001488 "src/f32-gemm/gen-inc/1x8inc-minmax-neon-dup-ld64.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001489 "src/f32-gemm/gen-inc/1x8inc-minmax-neon-lane-ld64.c",
1490 "src/f32-gemm/gen-inc/1x8s4inc-minmax-neon.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001491 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-dup-ld64.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001492 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-dup-ld128.c",
1493 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-lane-ld64.c",
1494 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-lane-ld128.c",
1495 "src/f32-gemm/gen-inc/4x8s4inc-minmax-neon.c",
1496 "src/f32-gemm/gen-inc/5x8inc-minmax-neon-lane-ld64.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001497 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-dup-ld64.c",
1498 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-dup-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001499 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-lane-ld64.c",
1500 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-lane-ld128.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001501 "src/f32-gemm/gen-inc/6x8s4inc-minmax-neon.c",
1502 "src/f32-gemm/gen-inc/8x8s4inc-minmax-neon.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001503 "src/f32-gemm/gen/1x8-minmax-neon-dup-ld64.c",
1504 "src/f32-gemm/gen/1x8-minmax-neon-lane-ld64.c",
1505 "src/f32-gemm/gen/1x8s4-minmax-neon.c",
1506 "src/f32-gemm/gen/4x2-minmax-neon-lane-ld64.c",
1507 "src/f32-gemm/gen/4x8-minmax-neon-dup-ld64.c",
1508 "src/f32-gemm/gen/4x8-minmax-neon-dup-ld128.c",
1509 "src/f32-gemm/gen/4x8-minmax-neon-lane-ld64.c",
1510 "src/f32-gemm/gen/4x8-minmax-neon-lane-ld128.c",
1511 "src/f32-gemm/gen/4x8s4-minmax-neon.c",
1512 "src/f32-gemm/gen/5x8-minmax-neon-lane-ld64.c",
1513 "src/f32-gemm/gen/6x8-minmax-neon-dup-ld64.c",
1514 "src/f32-gemm/gen/6x8-minmax-neon-dup-ld128.c",
1515 "src/f32-gemm/gen/6x8-minmax-neon-lane-ld64.c",
1516 "src/f32-gemm/gen/6x8-minmax-neon-lane-ld128.c",
1517 "src/f32-gemm/gen/6x8s4-minmax-neon.c",
1518 "src/f32-gemm/gen/8x8s4-minmax-neon.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07001519 "src/f32-hswish/gen/hswish-neon-x4.c",
1520 "src/f32-hswish/gen/hswish-neon-x8.c",
1521 "src/f32-hswish/gen/hswish-neon-x16.c",
Artsiom Ablavatski2202c812021-01-22 14:16:43 -08001522 "src/f32-ibilinear-chw/gen/neon-p4.c",
1523 "src/f32-ibilinear-chw/gen/neon-p8.c",
Frank Barchard8247e212021-02-03 18:12:33 -08001524 "src/f32-ibilinear/gen/neon-c4.c",
1525 "src/f32-ibilinear/gen/neon-c8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001526 "src/f32-igemm/gen/1x8-minmax-neon-dup-ld64.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001527 "src/f32-igemm/gen/1x8-minmax-neon-lane-ld64.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001528 "src/f32-igemm/gen/1x8s4-minmax-neon.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001529 "src/f32-igemm/gen/4x2-minmax-neon-lane-ld64.c",
1530 "src/f32-igemm/gen/4x4-minmax-neon-lane-ld64.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001531 "src/f32-igemm/gen/4x8-minmax-neon-dup-ld64.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001532 "src/f32-igemm/gen/4x8-minmax-neon-dup-ld128.c",
1533 "src/f32-igemm/gen/4x8-minmax-neon-lane-ld64.c",
1534 "src/f32-igemm/gen/4x8-minmax-neon-lane-ld128.c",
1535 "src/f32-igemm/gen/4x8s4-minmax-neon.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001536 "src/f32-igemm/gen/6x8-minmax-neon-dup-ld64.c",
1537 "src/f32-igemm/gen/6x8-minmax-neon-dup-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001538 "src/f32-igemm/gen/6x8-minmax-neon-lane-ld64.c",
1539 "src/f32-igemm/gen/6x8-minmax-neon-lane-ld128.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001540 "src/f32-igemm/gen/6x8s4-minmax-neon.c",
1541 "src/f32-igemm/gen/8x8s4-minmax-neon.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001542 "src/f32-maxpool/9p8x-minmax-neon-c4.c",
1543 "src/f32-pavgpool/9p8x-minmax-neon-c4.c",
1544 "src/f32-pavgpool/9x-minmax-neon-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001545 "src/f32-ppmm/gen/4x8-minmax-neon.c",
1546 "src/f32-ppmm/gen/8x8-minmax-neon.c",
Frank Barcharda5316982020-07-23 13:19:28 -07001547 "src/f32-prelu/gen/neon-1x4.c",
1548 "src/f32-prelu/gen/neon-1x8.c",
1549 "src/f32-prelu/gen/neon-1x16.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001550 "src/f32-prelu/gen/neon-2x4.c",
1551 "src/f32-prelu/gen/neon-2x8.c",
Frank Barcharda5316982020-07-23 13:19:28 -07001552 "src/f32-prelu/gen/neon-2x16.c",
1553 "src/f32-prelu/gen/neon-4x4.c",
1554 "src/f32-prelu/gen/neon-4x8.c",
1555 "src/f32-prelu/gen/neon-4x16.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001556 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x4.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001557 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x8-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001558 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x8.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001559 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12-acc2.c",
1560 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001561 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001562 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16-acc2.c",
1563 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001564 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001565 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20-acc2.c",
1566 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001567 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20.c",
1568 "src/f32-raddstoreexpminusmax/gen/neon-p5-x4.c",
1569 "src/f32-raddstoreexpminusmax/gen/neon-p5-x8-acc2.c",
1570 "src/f32-raddstoreexpminusmax/gen/neon-p5-x8.c",
1571 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12-acc2.c",
1572 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12-acc3.c",
1573 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12.c",
1574 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16-acc2.c",
1575 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16-acc4.c",
1576 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16.c",
1577 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20-acc2.c",
1578 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20-acc5.c",
1579 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07001580 "src/f32-relu/gen/neon-x4.c",
1581 "src/f32-relu/gen/neon-x8.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001582 "src/f32-rmax/neon.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -08001583 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x4.c",
1584 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x8.c",
1585 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x12.c",
1586 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x16.c",
1587 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x20.c",
1588 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x24.c",
1589 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x4.c",
1590 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x8.c",
1591 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x12.c",
1592 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x16.c",
1593 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x20.c",
1594 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x24.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001595 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x4.c",
1596 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x8.c",
1597 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x12.c",
1598 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x16.c",
1599 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x20.c",
1600 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x24.c",
Marat Dukhan5b86c432020-12-06 19:15:03 -08001601 "src/f32-spmm/gen/4x1-minmax-neon-pipelined.c",
1602 "src/f32-spmm/gen/4x1-minmax-neon-x2.c",
1603 "src/f32-spmm/gen/4x1-minmax-neon.c",
1604 "src/f32-spmm/gen/8x1-minmax-neon-pipelined.c",
1605 "src/f32-spmm/gen/8x1-minmax-neon-x2.c",
1606 "src/f32-spmm/gen/8x1-minmax-neon.c",
1607 "src/f32-spmm/gen/12x1-minmax-neon.c",
1608 "src/f32-spmm/gen/16x1-minmax-neon-pipelined.c",
1609 "src/f32-spmm/gen/16x1-minmax-neon-x2.c",
1610 "src/f32-spmm/gen/16x1-minmax-neon.c",
1611 "src/f32-spmm/gen/32x1-minmax-neon-pipelined.c",
1612 "src/f32-spmm/gen/32x1-minmax-neon-x2.c",
1613 "src/f32-spmm/gen/32x1-minmax-neon.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001614 "src/f32-vbinary/gen/vadd-minmax-neon-x4.c",
1615 "src/f32-vbinary/gen/vadd-minmax-neon-x8.c",
1616 "src/f32-vbinary/gen/vaddc-minmax-neon-x4.c",
1617 "src/f32-vbinary/gen/vaddc-minmax-neon-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -08001618 "src/f32-vbinary/gen/vmax-neon-x4.c",
1619 "src/f32-vbinary/gen/vmax-neon-x8.c",
1620 "src/f32-vbinary/gen/vmaxc-neon-x4.c",
1621 "src/f32-vbinary/gen/vmaxc-neon-x8.c",
1622 "src/f32-vbinary/gen/vmin-neon-x4.c",
1623 "src/f32-vbinary/gen/vmin-neon-x8.c",
1624 "src/f32-vbinary/gen/vminc-neon-x4.c",
1625 "src/f32-vbinary/gen/vminc-neon-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001626 "src/f32-vbinary/gen/vmul-minmax-neon-x4.c",
1627 "src/f32-vbinary/gen/vmul-minmax-neon-x8.c",
1628 "src/f32-vbinary/gen/vmulc-minmax-neon-x4.c",
1629 "src/f32-vbinary/gen/vmulc-minmax-neon-x8.c",
1630 "src/f32-vbinary/gen/vrsubc-minmax-neon-x4.c",
1631 "src/f32-vbinary/gen/vrsubc-minmax-neon-x8.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07001632 "src/f32-vbinary/gen/vsqrdiff-neon-x4.c",
1633 "src/f32-vbinary/gen/vsqrdiff-neon-x8.c",
1634 "src/f32-vbinary/gen/vsqrdiffc-neon-x4.c",
1635 "src/f32-vbinary/gen/vsqrdiffc-neon-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001636 "src/f32-vbinary/gen/vsub-minmax-neon-x4.c",
1637 "src/f32-vbinary/gen/vsub-minmax-neon-x8.c",
1638 "src/f32-vbinary/gen/vsubc-minmax-neon-x4.c",
1639 "src/f32-vbinary/gen/vsubc-minmax-neon-x8.c",
Marat Dukhaned6baaf2020-12-01 15:07:08 -08001640 "src/f32-velu/gen/velu-neon-rr2-lut16-p3-x4.c",
1641 "src/f32-velu/gen/velu-neon-rr2-lut16-p3-x8.c",
1642 "src/f32-velu/gen/velu-neon-rr2-lut16-p3-x12.c",
1643 "src/f32-velu/gen/velu-neon-rr2-lut16-p3-x16.c",
1644 "src/f32-velu/gen/velu-neon-rr2-lut16-p3-x20.c",
1645 "src/f32-velu/gen/velu-neon-rr2-lut16-p3-x24.c",
1646 "src/f32-velu/gen/velu-neon-rr2-p6-x4.c",
1647 "src/f32-velu/gen/velu-neon-rr2-p6-x8.c",
1648 "src/f32-velu/gen/velu-neon-rr2-p6-x12.c",
1649 "src/f32-velu/gen/velu-neon-rr2-p6-x16.c",
1650 "src/f32-velu/gen/velu-neon-rr2-p6-x20.c",
1651 "src/f32-velu/gen/velu-neon-rr2-p6-x24.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07001652 "src/f32-vlrelu/gen/vlrelu-neon-x4.c",
1653 "src/f32-vlrelu/gen/vlrelu-neon-x8.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001654 "src/f32-vmulcaddc/gen/c4-minmax-neon-2x.c",
1655 "src/f32-vmulcaddc/gen/c8-minmax-neon-2x.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07001656 "src/f32-vrnd/gen/vrndd-neon-x4.c",
1657 "src/f32-vrnd/gen/vrndd-neon-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001658 "src/f32-vrnd/gen/vrndne-neon-x4.c",
1659 "src/f32-vrnd/gen/vrndne-neon-x8.c",
1660 "src/f32-vrnd/gen/vrndu-neon-x4.c",
1661 "src/f32-vrnd/gen/vrndu-neon-x8.c",
1662 "src/f32-vrnd/gen/vrndz-neon-x4.c",
1663 "src/f32-vrnd/gen/vrndz-neon-x8.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07001664 "src/f32-vunary/gen/vabs-neon-x4.c",
1665 "src/f32-vunary/gen/vabs-neon-x8.c",
1666 "src/f32-vunary/gen/vneg-neon-x4.c",
1667 "src/f32-vunary/gen/vneg-neon-x8.c",
1668 "src/f32-vunary/gen/vsqr-neon-x4.c",
1669 "src/f32-vunary/gen/vsqr-neon-x8.c",
Marat Dukhande390d42020-11-29 19:32:18 -08001670 "src/math/expm1minus-neon-rr2-lut16-p3.c",
1671 "src/math/expm1minus-neon-rr2-p6.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001672 "src/math/roundd-neon-addsub.c",
1673 "src/math/roundd-neon-cvt.c",
1674 "src/math/roundne-neon-addsub.c",
1675 "src/math/roundu-neon-addsub.c",
1676 "src/math/roundu-neon-cvt.c",
1677 "src/math/roundz-neon-addsub.c",
1678 "src/math/roundz-neon-cvt.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001679 "src/math/sigmoid-neon-rr2-lut64-p2-nr2recps.c",
1680 "src/math/sigmoid-neon-rr2-lut2048-p1-nr2recps.c",
1681 "src/math/sigmoid-neon-rr2-p5-nr2recps.c",
1682 "src/math/sqrt-neon-nr1rsqrts.c",
1683 "src/math/sqrt-neon-nr2rsqrts.c",
1684 "src/math/sqrt-neon-nr3rsqrts.c",
Marat Dukhan023bcf92020-08-10 12:40:50 -07001685 "src/qs8-dwconv/gen/up8x9-minmax-neon-mul16.c",
1686 "src/qs8-dwconv/gen/up16x9-minmax-neon-mul16.c",
1687 "src/qs8-dwconv/gen/up24x9-minmax-neon-mul16.c",
1688 "src/qs8-dwconv/gen/up32x9-minmax-neon-mul16.c",
Marat Dukhan281262d2020-08-10 13:23:21 -07001689 "src/qs8-gavgpool/gen/7p7x-minmax-neon-c8-acc2.c",
1690 "src/qs8-gavgpool/gen/7p7x-minmax-neon-c16-acc2.c",
1691 "src/qs8-gavgpool/gen/7p7x-minmax-neon-c24-acc2.c",
1692 "src/qs8-gavgpool/gen/7p7x-minmax-neon-c32-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001693 "src/qs8-gavgpool/gen/7x-minmax-neon-c8-acc2.c",
1694 "src/qs8-gavgpool/gen/7x-minmax-neon-c16-acc2.c",
1695 "src/qs8-gavgpool/gen/7x-minmax-neon-c24-acc2.c",
1696 "src/qs8-gavgpool/gen/7x-minmax-neon-c32-acc2.c",
Frank Barchardec0bf142021-01-15 10:14:05 -08001697 "src/qs8-gemm/gen/1x8-minmax-neon-mlal-lane.c",
Frank Barchardec0bf142021-01-15 10:14:05 -08001698 "src/qs8-gemm/gen/1x8-minmax-neon-mull-addw-dup.c",
Frank Barchard8247e212021-02-03 18:12:33 -08001699 "src/qs8-gemm/gen/1x8c2-minmax-neon-mlal-padal-dup.c",
Frank Barchard2302ffd2021-01-22 14:08:02 -08001700 "src/qs8-gemm/gen/1x8c2-minmax-neon-mull-padal-dup.c",
Frank Barchardda78da12021-03-02 14:28:00 -08001701 "src/qs8-gemm/gen/1x8c8-minmax-neon-mlal-padal.c",
Frank Barchard8247e212021-02-03 18:12:33 -08001702 "src/qs8-gemm/gen/1x8c8-minmax-neon-mull-padal.c",
Frank Barchard671d1b02021-03-10 15:42:08 -08001703 "src/qs8-gemm/gen/1x8c16-minmax-neon-mlal-padal.c",
Frank Barchardcadd4222021-01-20 16:27:25 -08001704 "src/qs8-gemm/gen/1x16-minmax-neon-mlal-lane.c",
1705 "src/qs8-gemm/gen/1x16-minmax-neon-mull-addw-dup.c",
Frank Barchard8247e212021-02-03 18:12:33 -08001706 "src/qs8-gemm/gen/1x16c2-minmax-neon-mlal-padal-dup.c",
Frank Barchard2302ffd2021-01-22 14:08:02 -08001707 "src/qs8-gemm/gen/1x16c2-minmax-neon-mull-padal-dup.c",
Frank Barchardda78da12021-03-02 14:28:00 -08001708 "src/qs8-gemm/gen/1x16c8-minmax-neon-mlal-padal.c",
Frank Barchard8247e212021-02-03 18:12:33 -08001709 "src/qs8-gemm/gen/1x16c8-minmax-neon-mull-padal.c",
Frank Barchard671d1b02021-03-10 15:42:08 -08001710 "src/qs8-gemm/gen/1x16c16-minmax-neon-mlal-padal.c",
Frank Barchardcadd4222021-01-20 16:27:25 -08001711 "src/qs8-gemm/gen/2x8-minmax-neon-mlal-lane.c",
Frank Barchardec0bf142021-01-15 10:14:05 -08001712 "src/qs8-gemm/gen/2x8-minmax-neon-mull-addw-dup.c",
Frank Barchard8247e212021-02-03 18:12:33 -08001713 "src/qs8-gemm/gen/2x8c2-minmax-neon-mlal-padal-dup.c",
Frank Barchard2302ffd2021-01-22 14:08:02 -08001714 "src/qs8-gemm/gen/2x8c2-minmax-neon-mull-padal-dup.c",
Frank Barchardda78da12021-03-02 14:28:00 -08001715 "src/qs8-gemm/gen/2x8c8-minmax-neon-mlal-padal.c",
Frank Barchard8247e212021-02-03 18:12:33 -08001716 "src/qs8-gemm/gen/2x8c8-minmax-neon-mull-padal.c",
Frank Barchard671d1b02021-03-10 15:42:08 -08001717 "src/qs8-gemm/gen/2x8c16-minmax-neon-mlal-padal.c",
Frank Barchardcadd4222021-01-20 16:27:25 -08001718 "src/qs8-gemm/gen/2x16-minmax-neon-mlal-lane.c",
1719 "src/qs8-gemm/gen/2x16-minmax-neon-mull-addw-dup.c",
Frank Barchard8247e212021-02-03 18:12:33 -08001720 "src/qs8-gemm/gen/2x16c2-minmax-neon-mlal-padal-dup.c",
Frank Barchard2302ffd2021-01-22 14:08:02 -08001721 "src/qs8-gemm/gen/2x16c2-minmax-neon-mull-padal-dup.c",
Frank Barchardda78da12021-03-02 14:28:00 -08001722 "src/qs8-gemm/gen/2x16c8-minmax-neon-mlal-padal.c",
Frank Barchard8247e212021-02-03 18:12:33 -08001723 "src/qs8-gemm/gen/2x16c8-minmax-neon-mull-padal.c",
Frank Barchard671d1b02021-03-10 15:42:08 -08001724 "src/qs8-gemm/gen/2x16c16-minmax-neon-mlal-padal.c",
Frank Barchard2302ffd2021-01-22 14:08:02 -08001725 "src/qs8-gemm/gen/3x8-minmax-neon-mlal-lane.c",
1726 "src/qs8-gemm/gen/3x8-minmax-neon-mull-addw-dup.c",
Frank Barchard8247e212021-02-03 18:12:33 -08001727 "src/qs8-gemm/gen/3x8c2-minmax-neon-mlal-padal-dup.c",
Frank Barchard2302ffd2021-01-22 14:08:02 -08001728 "src/qs8-gemm/gen/3x8c2-minmax-neon-mull-padal-dup.c",
Frank Barchardda78da12021-03-02 14:28:00 -08001729 "src/qs8-gemm/gen/3x8c8-minmax-neon-mlal-padal.c",
Frank Barchard8247e212021-02-03 18:12:33 -08001730 "src/qs8-gemm/gen/3x8c8-minmax-neon-mull-padal.c",
Frank Barchard671d1b02021-03-10 15:42:08 -08001731 "src/qs8-gemm/gen/3x8c16-minmax-neon-mlal-padal.c",
Frank Barchard8247e212021-02-03 18:12:33 -08001732 "src/qs8-gemm/gen/3x16-minmax-neon-mlal-lane.c",
1733 "src/qs8-gemm/gen/3x16-minmax-neon-mull-addw-dup.c",
1734 "src/qs8-gemm/gen/3x16c2-minmax-neon-mlal-padal-dup.c",
1735 "src/qs8-gemm/gen/3x16c2-minmax-neon-mull-padal-dup.c",
Frank Barchardda78da12021-03-02 14:28:00 -08001736 "src/qs8-gemm/gen/3x16c8-minmax-neon-mlal-padal.c",
Frank Barchard8247e212021-02-03 18:12:33 -08001737 "src/qs8-gemm/gen/3x16c8-minmax-neon-mull-padal.c",
Frank Barchard671d1b02021-03-10 15:42:08 -08001738 "src/qs8-gemm/gen/3x16c16-minmax-neon-mlal-padal.c",
Frank Barchardcadd4222021-01-20 16:27:25 -08001739 "src/qs8-gemm/gen/4x8-minmax-neon-mlal-lane.c",
Frank Barchardec0bf142021-01-15 10:14:05 -08001740 "src/qs8-gemm/gen/4x8-minmax-neon-mull-addw-dup.c",
Frank Barchard8247e212021-02-03 18:12:33 -08001741 "src/qs8-gemm/gen/4x8c2-minmax-neon-mlal-padal-dup.c",
Frank Barchard2302ffd2021-01-22 14:08:02 -08001742 "src/qs8-gemm/gen/4x8c2-minmax-neon-mull-padal-dup.c",
Frank Barchardda78da12021-03-02 14:28:00 -08001743 "src/qs8-gemm/gen/4x8c8-minmax-neon-mlal-padal.c",
Frank Barchard8247e212021-02-03 18:12:33 -08001744 "src/qs8-gemm/gen/4x8c8-minmax-neon-mull-padal.c",
Frank Barchard671d1b02021-03-10 15:42:08 -08001745 "src/qs8-gemm/gen/4x8c16-minmax-neon-mlal-padal.c",
Frank Barchardcadd4222021-01-20 16:27:25 -08001746 "src/qs8-gemm/gen/4x16-minmax-neon-mlal-lane.c",
1747 "src/qs8-gemm/gen/4x16-minmax-neon-mull-addw-dup.c",
Frank Barchard8247e212021-02-03 18:12:33 -08001748 "src/qs8-gemm/gen/4x16c2-minmax-neon-mlal-padal-dup.c",
Frank Barchard2302ffd2021-01-22 14:08:02 -08001749 "src/qs8-gemm/gen/4x16c2-minmax-neon-mull-padal-dup.c",
Frank Barchardda78da12021-03-02 14:28:00 -08001750 "src/qs8-gemm/gen/4x16c8-minmax-neon-mlal-padal.c",
Frank Barchard58990122021-01-29 16:09:56 -08001751 "src/qs8-gemm/gen/4x16c8-minmax-neon-mull-padal.c",
Frank Barchard71c4d1a2021-02-19 10:07:23 -08001752 "src/qs8-gemm/gen/4x16c16-minmax-neon-mlal-padal.c",
Frank Barchardec0bf142021-01-15 10:14:05 -08001753 "src/qs8-igemm/gen/1x8-minmax-neon-mlal-lane.c",
Frank Barchard671d1b02021-03-10 15:42:08 -08001754 "src/qs8-igemm/gen/1x8-minmax-neon-mull-addw-dup.c",
1755 "src/qs8-igemm/gen/1x8c2-minmax-neon-mlal-padal-dup.c",
1756 "src/qs8-igemm/gen/1x8c2-minmax-neon-mull-padal-dup.c",
Frank Barchardda78da12021-03-02 14:28:00 -08001757 "src/qs8-igemm/gen/1x8c8-minmax-neon-mlal-padal.c",
1758 "src/qs8-igemm/gen/1x8c8-minmax-neon-mull-padal.c",
Frank Barchard671d1b02021-03-10 15:42:08 -08001759 "src/qs8-igemm/gen/1x8c16-minmax-neon-mlal-padal.c",
1760 "src/qs8-igemm/gen/1x16-minmax-neon-mlal-lane.c",
1761 "src/qs8-igemm/gen/1x16-minmax-neon-mull-addw-dup.c",
1762 "src/qs8-igemm/gen/1x16c2-minmax-neon-mlal-padal-dup.c",
1763 "src/qs8-igemm/gen/1x16c2-minmax-neon-mull-padal-dup.c",
1764 "src/qs8-igemm/gen/1x16c8-minmax-neon-mlal-padal.c",
1765 "src/qs8-igemm/gen/1x16c8-minmax-neon-mull-padal.c",
1766 "src/qs8-igemm/gen/1x16c16-minmax-neon-mlal-padal.c",
1767 "src/qs8-igemm/gen/2x8-minmax-neon-mlal-lane.c",
1768 "src/qs8-igemm/gen/2x8-minmax-neon-mull-addw-dup.c",
1769 "src/qs8-igemm/gen/2x8c2-minmax-neon-mlal-padal-dup.c",
1770 "src/qs8-igemm/gen/2x8c2-minmax-neon-mull-padal-dup.c",
Frank Barchardda78da12021-03-02 14:28:00 -08001771 "src/qs8-igemm/gen/2x8c8-minmax-neon-mlal-padal.c",
1772 "src/qs8-igemm/gen/2x8c8-minmax-neon-mull-padal.c",
Frank Barchard671d1b02021-03-10 15:42:08 -08001773 "src/qs8-igemm/gen/2x8c16-minmax-neon-mlal-padal.c",
1774 "src/qs8-igemm/gen/2x16-minmax-neon-mlal-lane.c",
1775 "src/qs8-igemm/gen/2x16-minmax-neon-mull-addw-dup.c",
1776 "src/qs8-igemm/gen/2x16c2-minmax-neon-mlal-padal-dup.c",
1777 "src/qs8-igemm/gen/2x16c2-minmax-neon-mull-padal-dup.c",
1778 "src/qs8-igemm/gen/2x16c8-minmax-neon-mlal-padal.c",
1779 "src/qs8-igemm/gen/2x16c8-minmax-neon-mull-padal.c",
1780 "src/qs8-igemm/gen/2x16c16-minmax-neon-mlal-padal.c",
1781 "src/qs8-igemm/gen/3x8-minmax-neon-mlal-lane.c",
1782 "src/qs8-igemm/gen/3x8-minmax-neon-mull-addw-dup.c",
1783 "src/qs8-igemm/gen/3x8c2-minmax-neon-mlal-padal-dup.c",
1784 "src/qs8-igemm/gen/3x8c2-minmax-neon-mull-padal-dup.c",
Frank Barchardda78da12021-03-02 14:28:00 -08001785 "src/qs8-igemm/gen/3x8c8-minmax-neon-mlal-padal.c",
1786 "src/qs8-igemm/gen/3x8c8-minmax-neon-mull-padal.c",
Frank Barchard671d1b02021-03-10 15:42:08 -08001787 "src/qs8-igemm/gen/3x8c16-minmax-neon-mlal-padal.c",
1788 "src/qs8-igemm/gen/3x16-minmax-neon-mlal-lane.c",
1789 "src/qs8-igemm/gen/3x16-minmax-neon-mull-addw-dup.c",
1790 "src/qs8-igemm/gen/3x16c2-minmax-neon-mlal-padal-dup.c",
1791 "src/qs8-igemm/gen/3x16c2-minmax-neon-mull-padal-dup.c",
1792 "src/qs8-igemm/gen/3x16c8-minmax-neon-mlal-padal.c",
1793 "src/qs8-igemm/gen/3x16c8-minmax-neon-mull-padal.c",
1794 "src/qs8-igemm/gen/3x16c16-minmax-neon-mlal-padal.c",
1795 "src/qs8-igemm/gen/4x8-minmax-neon-mlal-lane.c",
1796 "src/qs8-igemm/gen/4x8-minmax-neon-mull-addw-dup.c",
1797 "src/qs8-igemm/gen/4x8c2-minmax-neon-mlal-padal-dup.c",
1798 "src/qs8-igemm/gen/4x8c2-minmax-neon-mull-padal-dup.c",
Frank Barchardda78da12021-03-02 14:28:00 -08001799 "src/qs8-igemm/gen/4x8c8-minmax-neon-mlal-padal.c",
1800 "src/qs8-igemm/gen/4x8c8-minmax-neon-mull-padal.c",
Frank Barchardda78da12021-03-02 14:28:00 -08001801 "src/qs8-igemm/gen/4x8c16-minmax-neon-mlal-padal.c",
Frank Barchard671d1b02021-03-10 15:42:08 -08001802 "src/qs8-igemm/gen/4x16-minmax-neon-mlal-lane.c",
Frank Barchardda78da12021-03-02 14:28:00 -08001803 "src/qs8-igemm/gen/4x16-minmax-neon-mull-addw-dup.c",
Frank Barchard671d1b02021-03-10 15:42:08 -08001804 "src/qs8-igemm/gen/4x16c2-minmax-neon-mlal-padal-dup.c",
1805 "src/qs8-igemm/gen/4x16c2-minmax-neon-mull-padal-dup.c",
1806 "src/qs8-igemm/gen/4x16c8-minmax-neon-mlal-padal.c",
1807 "src/qs8-igemm/gen/4x16c8-minmax-neon-mull-padal.c",
1808 "src/qs8-igemm/gen/4x16c16-minmax-neon-mlal-padal.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -07001809 "src/qs8-requantization/fp32-neon.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001810 "src/qs8-requantization/precise-neon.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -07001811 "src/qs8-requantization/q31-neon.c",
Marat Dukhanba7b2792020-09-02 14:26:45 -07001812 "src/qs8-vadd/gen/minmax-neon-ld64-x8.c",
1813 "src/qs8-vadd/gen/minmax-neon-ld64-x16.c",
1814 "src/qs8-vadd/gen/minmax-neon-ld64-x24.c",
1815 "src/qs8-vadd/gen/minmax-neon-ld64-x32.c",
1816 "src/qs8-vaddc/gen/minmax-neon-ld64-x8.c",
1817 "src/qs8-vaddc/gen/minmax-neon-ld64-x16.c",
1818 "src/qs8-vaddc/gen/minmax-neon-ld64-x24.c",
1819 "src/qs8-vaddc/gen/minmax-neon-ld64-x32.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -07001820 "src/qu8-avgpool/9p8x-minmax-neon-c8.c",
1821 "src/qu8-avgpool/9x-minmax-neon-c8.c",
1822 "src/qu8-dwconv/up8x9-minmax-neon.c",
1823 "src/qu8-gavgpool/7p7x-minmax-neon-c8.c",
1824 "src/qu8-gavgpool/7x-minmax-neon-c8.c",
1825 "src/qu8-gemm/4x8-minmax-neon.c",
1826 "src/qu8-gemm/8x8-minmax-neon.c",
1827 "src/qu8-igemm/4x8-minmax-neon.c",
1828 "src/qu8-igemm/8x8-minmax-neon.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07001829 "src/qu8-requantization/fp32-neon.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001830 "src/qu8-requantization/precise-neon.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07001831 "src/qu8-requantization/q31-neon.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -07001832 "src/qu8-vadd/minmax-neon.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001833 "src/u8-clamp/neon-x64.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001834 "src/u8-maxpool/9p8x-minmax-neon-c16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001835 "src/u8-rmax/neon.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001836 "src/x8-zip/x2-neon.c",
1837 "src/x8-zip/x3-neon.c",
1838 "src/x8-zip/x4-neon.c",
1839 "src/x8-zip/xm-neon.c",
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -07001840 "src/x32-fill/neon.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001841 "src/x32-packx/x4-neon-st4.c",
Marat Dukhan63523d42020-05-22 17:07:33 -07001842 "src/x32-pad/neon.c",
Marat Dukhan57dccd82020-04-14 00:53:10 -07001843 "src/x32-unpool/neon.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001844 "src/x32-zip/x2-neon.c",
1845 "src/x32-zip/x3-neon.c",
1846 "src/x32-zip/x4-neon.c",
1847 "src/x32-zip/xm-neon.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001848]
1849
1850NEONFMA_UKERNELS = [
Frank Barchard04336c12020-10-22 16:48:55 -07001851 "src/f32-dwconv/gen/up4x4-minmax-neonfma-acc2.c",
1852 "src/f32-dwconv/gen/up4x4-minmax-neonfma.c",
1853 "src/f32-dwconv/gen/up4x9-minmax-neonfma-acc2.c",
1854 "src/f32-dwconv/gen/up4x9-minmax-neonfma.c",
1855 "src/f32-dwconv/gen/up4x25-minmax-neonfma-acc2.c",
1856 "src/f32-dwconv/gen/up4x25-minmax-neonfma.c",
1857 "src/f32-dwconv/gen/up8x4-minmax-neonfma-acc2.c",
1858 "src/f32-dwconv/gen/up8x4-minmax-neonfma.c",
1859 "src/f32-dwconv/gen/up8x9-minmax-neonfma-acc2.c",
1860 "src/f32-dwconv/gen/up8x9-minmax-neonfma.c",
1861 "src/f32-dwconv/gen/up8x25-minmax-neonfma-acc2.c",
1862 "src/f32-dwconv/gen/up8x25-minmax-neonfma.c",
1863 "src/f32-gemm/gen-inc/1x8inc-minmax-neonfma-dup-ld64.c",
1864 "src/f32-gemm/gen-inc/1x8s4inc-minmax-neonfma.c",
1865 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-dup-ld64.c",
1866 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-dup-ld128.c",
1867 "src/f32-gemm/gen-inc/4x8s4inc-minmax-neonfma.c",
1868 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-dup-ld64.c",
1869 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-dup-ld128.c",
1870 "src/f32-gemm/gen-inc/6x8s4inc-minmax-neonfma.c",
1871 "src/f32-gemm/gen-inc/8x8s4inc-minmax-neonfma.c",
1872 "src/f32-gemm/gen/1x8-minmax-neonfma-dup-ld64.c",
1873 "src/f32-gemm/gen/1x8s4-minmax-neonfma.c",
1874 "src/f32-gemm/gen/4x8-minmax-neonfma-dup-ld64.c",
1875 "src/f32-gemm/gen/4x8-minmax-neonfma-dup-ld128.c",
1876 "src/f32-gemm/gen/4x8s4-minmax-neonfma.c",
1877 "src/f32-gemm/gen/6x8-minmax-neonfma-dup-ld64.c",
1878 "src/f32-gemm/gen/6x8-minmax-neonfma-dup-ld128.c",
1879 "src/f32-gemm/gen/6x8s4-minmax-neonfma.c",
1880 "src/f32-gemm/gen/8x8s4-minmax-neonfma.c",
Artsiom Ablavatski2202c812021-01-22 14:16:43 -08001881 "src/f32-ibilinear-chw/gen/neonfma-p4.c",
1882 "src/f32-ibilinear-chw/gen/neonfma-p8.c",
Frank Barchard8247e212021-02-03 18:12:33 -08001883 "src/f32-ibilinear/gen/neonfma-c4.c",
1884 "src/f32-ibilinear/gen/neonfma-c8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001885 "src/f32-igemm/gen/1x8-minmax-neonfma-dup-ld64.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001886 "src/f32-igemm/gen/1x8s4-minmax-neonfma.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001887 "src/f32-igemm/gen/4x8-minmax-neonfma-dup-ld64.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001888 "src/f32-igemm/gen/4x8-minmax-neonfma-dup-ld128.c",
1889 "src/f32-igemm/gen/4x8s4-minmax-neonfma.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001890 "src/f32-igemm/gen/6x8-minmax-neonfma-dup-ld64.c",
1891 "src/f32-igemm/gen/6x8-minmax-neonfma-dup-ld128.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001892 "src/f32-igemm/gen/6x8s4-minmax-neonfma.c",
1893 "src/f32-igemm/gen/8x8s4-minmax-neonfma.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001894 "src/f32-ppmm/gen/4x8-minmax-neonfma.c",
1895 "src/f32-ppmm/gen/8x8-minmax-neonfma.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001896 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x4.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001897 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x8-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001898 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x8.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001899 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12-acc2.c",
1900 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001901 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001902 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16-acc2.c",
1903 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001904 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001905 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20-acc2.c",
1906 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001907 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20.c",
1908 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x4.c",
1909 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x8-acc2.c",
1910 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x8.c",
1911 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12-acc2.c",
1912 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12-acc3.c",
1913 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12.c",
1914 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16-acc2.c",
1915 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16-acc4.c",
1916 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16.c",
1917 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20-acc2.c",
1918 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20-acc5.c",
1919 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -08001920 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x4.c",
1921 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x8.c",
1922 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x12.c",
1923 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x16.c",
1924 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x20.c",
1925 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x24.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001926 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x4.c",
1927 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x8.c",
1928 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x12.c",
1929 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x16.c",
1930 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x20.c",
1931 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x24.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -08001932 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x4.c",
1933 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x8.c",
1934 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x12.c",
1935 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x16.c",
1936 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x20.c",
1937 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x24.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -08001938 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x4.c",
1939 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x8.c",
1940 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x12.c",
1941 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x16.c",
1942 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x20.c",
1943 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x24.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001944 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x4.c",
1945 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x8.c",
1946 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x12.c",
1947 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x16.c",
1948 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x20.c",
1949 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x24.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -08001950 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x4.c",
1951 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x8.c",
1952 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x12.c",
1953 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x16.c",
1954 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x20.c",
1955 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x24.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001956 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x4.c",
1957 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x8.c",
1958 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x12.c",
1959 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x16.c",
1960 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x20.c",
1961 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x24.c",
1962 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x4.c",
1963 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x8.c",
1964 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x12.c",
1965 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x16.c",
1966 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x20.c",
1967 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x24.c",
1968 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x4.c",
1969 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x8.c",
1970 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x12.c",
1971 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x16.c",
1972 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x20.c",
1973 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x24.c",
Marat Dukhan2fa7a0c2020-12-06 19:09:02 -08001974 "src/f32-spmm/gen/4x1-minmax-neonfma-pipelined.c",
1975 "src/f32-spmm/gen/4x1-minmax-neonfma-x2.c",
1976 "src/f32-spmm/gen/4x1-minmax-neonfma.c",
1977 "src/f32-spmm/gen/8x1-minmax-neonfma-pipelined.c",
1978 "src/f32-spmm/gen/8x1-minmax-neonfma-x2.c",
1979 "src/f32-spmm/gen/8x1-minmax-neonfma.c",
1980 "src/f32-spmm/gen/12x1-minmax-neonfma.c",
1981 "src/f32-spmm/gen/16x1-minmax-neonfma-pipelined.c",
1982 "src/f32-spmm/gen/16x1-minmax-neonfma-x2.c",
1983 "src/f32-spmm/gen/16x1-minmax-neonfma.c",
1984 "src/f32-spmm/gen/32x1-minmax-neonfma-pipelined.c",
1985 "src/f32-spmm/gen/32x1-minmax-neonfma-x2.c",
1986 "src/f32-spmm/gen/32x1-minmax-neonfma.c",
Marat Dukhaned6baaf2020-12-01 15:07:08 -08001987 "src/f32-velu/gen/velu-neonfma-rr1-lut16-p3-x4.c",
1988 "src/f32-velu/gen/velu-neonfma-rr1-lut16-p3-x8.c",
1989 "src/f32-velu/gen/velu-neonfma-rr1-lut16-p3-x12.c",
1990 "src/f32-velu/gen/velu-neonfma-rr1-lut16-p3-x16.c",
1991 "src/f32-velu/gen/velu-neonfma-rr1-lut16-p3-x20.c",
1992 "src/f32-velu/gen/velu-neonfma-rr1-lut16-p3-x24.c",
1993 "src/f32-velu/gen/velu-neonfma-rr1-p6-x4.c",
1994 "src/f32-velu/gen/velu-neonfma-rr1-p6-x8.c",
1995 "src/f32-velu/gen/velu-neonfma-rr1-p6-x12.c",
1996 "src/f32-velu/gen/velu-neonfma-rr1-p6-x16.c",
1997 "src/f32-velu/gen/velu-neonfma-rr1-p6-x20.c",
1998 "src/f32-velu/gen/velu-neonfma-rr1-p6-x24.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001999 "src/f32-vmulcaddc/gen/c4-minmax-neonfma-2x.c",
2000 "src/f32-vmulcaddc/gen/c8-minmax-neonfma-2x.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07002001 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x4.c",
2002 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x8.c",
2003 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x12.c",
2004 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x16.c",
2005 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x20.c",
2006 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x24.c",
2007 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x28.c",
2008 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x32.c",
2009 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x36.c",
2010 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x40.c",
2011 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x4.c",
2012 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x8.c",
2013 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x12.c",
2014 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x16.c",
2015 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x20.c",
2016 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x24.c",
2017 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x28.c",
2018 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x32.c",
2019 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x36.c",
2020 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x40.c",
Marat Dukhanb7633f22020-11-20 16:34:56 -08002021 "src/math/exp-neonfma-rr2-lut64-p2.c",
2022 "src/math/exp-neonfma-rr2-p5.c",
Frank Barcharde7223ee2020-12-04 19:04:01 -08002023 "src/math/expm1minus-neonfma-rr1-lut16-p3.c",
2024 "src/math/expm1minus-neonfma-rr1-p6.c",
Marat Dukhan9dd119a2020-11-20 18:20:04 -08002025 "src/math/expminus-neonfma-rr2-lut64-p2.c",
2026 "src/math/expminus-neonfma-rr2-lut2048-p1.c",
2027 "src/math/expminus-neonfma-rr2-p5.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08002028 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr1recps1fma.c",
2029 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr2fma.c",
2030 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr2recps.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002031 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr1recps1fma.c",
2032 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr2fma.c",
2033 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr2recps.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08002034 "src/math/sigmoid-neonfma-rr1-p5-nr1recps1fma.c",
2035 "src/math/sigmoid-neonfma-rr1-p5-nr2fma.c",
2036 "src/math/sigmoid-neonfma-rr1-p5-nr2recps.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08002037 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr1recps1fma.c",
2038 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr2fma.c",
2039 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr2recps.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002040 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr1recps1fma.c",
2041 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr2fma.c",
2042 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr2recps.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08002043 "src/math/sigmoid-neonfma-rr2-p5-nr1recps1fma.c",
2044 "src/math/sigmoid-neonfma-rr2-p5-nr2fma.c",
2045 "src/math/sigmoid-neonfma-rr2-p5-nr2recps.c",
Marat Dukhan84000762020-06-29 18:38:43 -07002046 "src/math/sqrt-neonfma-nr1fma.c",
Marat Dukhan84000762020-06-29 18:38:43 -07002047 "src/math/sqrt-neonfma-nr1rsqrts1fma1adj.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002048 "src/math/sqrt-neonfma-nr2fma.c",
2049 "src/math/sqrt-neonfma-nr2fma1adj.c",
2050 "src/math/sqrt-neonfma-nr3fma.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002051]
2052
2053AARCH64_NEONFMA_UKERNELS = [
Marat Dukhan56b10cd2020-05-18 09:35:49 -07002054 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neonfma-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07002055 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neonfma-2x2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002056 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neonfma-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07002057 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neonfma-2x2.c",
Marat Dukhan56b10cd2020-05-18 09:35:49 -07002058 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neonfma-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07002059 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neonfma-2x2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002060 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neonfma-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07002061 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neonfma-2x2.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -07002062 "src/f32-conv-hwc2chw/3x3s2p1c3x4-neonfma-2x2.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07002063 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-1x4-acc2.c",
2064 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-1x4-acc3.c",
2065 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-1x4-acc4.c",
Marat Dukhan1268a242020-10-24 00:36:32 -07002066 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-1x4.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07002067 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-2x4-acc2.c",
Marat Dukhan1268a242020-10-24 00:36:32 -07002068 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-2x4.c",
2069 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-3x4.c",
2070 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-4x4.c",
2071 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-5x4.c",
2072 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-6x4.c",
Marat Dukhan82f0c322020-10-25 19:17:35 -07002073 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neonfma-1x4-acc2.c",
2074 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neonfma-1x4-acc3.c",
2075 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neonfma-1x4-acc4.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07002076 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neonfma-1x4.c",
Marat Dukhan82f0c322020-10-25 19:17:35 -07002077 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neonfma-2x4-acc2.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07002078 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neonfma-2x4.c",
2079 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neonfma-3x4.c",
2080 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neonfma-4x4.c",
Marat Dukhan149f0ea2020-10-26 12:50:33 -07002081 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-1x4-acc2.c",
2082 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-1x4-acc3.c",
2083 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-1x4-acc4.c",
2084 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-1x4-acc5.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07002085 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-1x4.c",
Marat Dukhan149f0ea2020-10-26 12:50:33 -07002086 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-2x4-acc2.c",
2087 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-2x4-acc3.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07002088 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-2x4.c",
Marat Dukhan149f0ea2020-10-26 12:50:33 -07002089 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-3x4-acc2.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07002090 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-3x4.c",
Marat Dukhan149f0ea2020-10-26 12:50:33 -07002091 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-4x4-acc2.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07002092 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-4x4.c",
2093 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-5x4.c",
Marat Dukhan30d4b252020-10-29 16:33:22 -07002094 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neonfma-1x4-acc2.c",
2095 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neonfma-1x4-acc3.c",
2096 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neonfma-1x4-acc4.c",
2097 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neonfma-1x4-acc5.c",
2098 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neonfma-1x4.c",
2099 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neonfma-2x4-acc2.c",
2100 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neonfma-2x4-acc3.c",
2101 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neonfma-2x4.c",
Marat Dukhan30d4b252020-10-29 16:33:22 -07002102 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neonfma-3x4-acc2.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07002103 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neonfma-3x4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002104 "src/f32-gemm/gen-inc/1x8inc-minmax-neonfma-lane-ld64.c",
2105 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-lane-ld64.c",
2106 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-lane-ld128.c",
2107 "src/f32-gemm/gen-inc/5x8inc-minmax-neonfma-lane-ld64.c",
2108 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-lane-ld64.c",
2109 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-lane-ld128.c",
2110 "src/f32-gemm/gen/1x8-minmax-neonfma-lane-ld64.c",
2111 "src/f32-gemm/gen/4x2-minmax-neonfma-lane-ld64.c",
2112 "src/f32-gemm/gen/4x8-minmax-neonfma-lane-ld64.c",
2113 "src/f32-gemm/gen/4x8-minmax-neonfma-lane-ld128.c",
2114 "src/f32-gemm/gen/5x8-minmax-neonfma-lane-ld64.c",
2115 "src/f32-gemm/gen/6x8-minmax-neonfma-lane-ld64.c",
2116 "src/f32-gemm/gen/6x8-minmax-neonfma-lane-ld128.c",
2117 "src/f32-igemm/gen/1x8-minmax-neonfma-lane-ld64.c",
2118 "src/f32-igemm/gen/4x2-minmax-neonfma-lane-ld64.c",
2119 "src/f32-igemm/gen/4x4-minmax-neonfma-lane-ld64.c",
2120 "src/f32-igemm/gen/4x8-minmax-neonfma-lane-ld64.c",
2121 "src/f32-igemm/gen/4x8-minmax-neonfma-lane-ld128.c",
2122 "src/f32-igemm/gen/6x8-minmax-neonfma-lane-ld64.c",
2123 "src/f32-igemm/gen/6x8-minmax-neonfma-lane-ld128.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -08002124 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x4.c",
2125 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x8.c",
2126 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x12.c",
2127 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x16.c",
2128 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x20.c",
2129 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x24.c",
2130 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x4.c",
2131 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x8.c",
2132 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x12.c",
2133 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x16.c",
2134 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x20.c",
2135 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x24.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002136 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x4.c",
2137 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x8.c",
2138 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x12.c",
2139 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x16.c",
2140 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x20.c",
2141 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x24.c",
Marat Dukhan355ab432020-04-09 19:01:52 -07002142 "src/f32-spmm/gen/4x2-minmax-neonfma.c",
2143 "src/f32-spmm/gen/4x4-minmax-neonfma.c",
Marat Dukhan355ab432020-04-09 19:01:52 -07002144 "src/f32-spmm/gen/8x2-minmax-neonfma.c",
2145 "src/f32-spmm/gen/8x4-minmax-neonfma.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002146 "src/f32-spmm/gen/12x2-minmax-neonfma.c",
2147 "src/f32-spmm/gen/12x4-minmax-neonfma.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002148 "src/f32-spmm/gen/16x2-minmax-neonfma.c",
2149 "src/f32-spmm/gen/16x4-minmax-neonfma.c",
Frank Barchard846c0c62020-10-26 15:01:39 -07002150 "src/f32-spmm/gen/32x2-minmax-neonfma.c",
2151 "src/f32-spmm/gen/32x4-minmax-neonfma.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002152 "src/f32-vbinary/gen/vdiv-minmax-neon-x4.c",
2153 "src/f32-vbinary/gen/vdiv-minmax-neon-x8.c",
2154 "src/f32-vbinary/gen/vdivc-minmax-neon-x4.c",
2155 "src/f32-vbinary/gen/vdivc-minmax-neon-x8.c",
2156 "src/f32-vbinary/gen/vrdivc-minmax-neon-x4.c",
2157 "src/f32-vbinary/gen/vrdivc-minmax-neon-x8.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07002158 "src/f32-vsqrt/gen/neon-sqrt-x4.c",
2159 "src/f32-vsqrt/gen/neon-sqrt-x8.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08002160 "src/math/sigmoid-neonfma-rr1-lut64-p2-div.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002161 "src/math/sigmoid-neonfma-rr1-lut2048-p1-div.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08002162 "src/math/sigmoid-neonfma-rr1-p5-div.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08002163 "src/math/sigmoid-neonfma-rr2-lut64-p2-div.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002164 "src/math/sigmoid-neonfma-rr2-lut2048-p1-div.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08002165 "src/math/sigmoid-neonfma-rr2-p5-div.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002166]
2167
Marat Dukhan8853b822020-05-07 12:19:01 -07002168NEONV8_UKERNELS = [
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07002169 "src/f32-vrnd/gen/vrndd-neonv8-x4.c",
2170 "src/f32-vrnd/gen/vrndd-neonv8-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002171 "src/f32-vrnd/gen/vrndne-neonv8-x4.c",
2172 "src/f32-vrnd/gen/vrndne-neonv8-x8.c",
2173 "src/f32-vrnd/gen/vrndu-neonv8-x4.c",
2174 "src/f32-vrnd/gen/vrndu-neonv8-x8.c",
2175 "src/f32-vrnd/gen/vrndz-neonv8-x4.c",
2176 "src/f32-vrnd/gen/vrndz-neonv8-x8.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07002177 "src/math/roundd-neonv8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002178 "src/math/roundne-neonv8.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07002179 "src/math/roundu-neonv8.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07002180 "src/math/roundz-neonv8.c",
Marat Dukhan8853b822020-05-07 12:19:01 -07002181]
2182
Marat Dukhan08c4a432019-10-03 09:29:21 -07002183AARCH64_NEONFP16ARITH_UKERNELS = [
Frank Barchardb1966592020-05-12 13:47:06 -07002184 "src/f16-clamp/gen/neonfp16arith-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002185 "src/f16-clamp/gen/neonfp16arith-x16.c",
Frank Barchard5a599a62020-06-04 20:12:44 -07002186 "src/f16-dwconv/gen/up8x4-minmax-neonfp16arith-acc2.c",
2187 "src/f16-dwconv/gen/up8x4-minmax-neonfp16arith.c",
2188 "src/f16-dwconv/gen/up8x9-minmax-neonfp16arith-acc2.c",
2189 "src/f16-dwconv/gen/up8x9-minmax-neonfp16arith.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002190 "src/f16-dwconv/gen/up8x25-minmax-neonfp16arith-acc2.c",
2191 "src/f16-dwconv/gen/up8x25-minmax-neonfp16arith.c",
2192 "src/f16-dwconv/gen/up16x4-minmax-neonfp16arith-acc2.c",
2193 "src/f16-dwconv/gen/up16x4-minmax-neonfp16arith.c",
2194 "src/f16-dwconv/gen/up16x9-minmax-neonfp16arith-acc2.c",
2195 "src/f16-dwconv/gen/up16x9-minmax-neonfp16arith.c",
2196 "src/f16-dwconv/gen/up16x25-minmax-neonfp16arith-acc2.c",
2197 "src/f16-dwconv/gen/up16x25-minmax-neonfp16arith.c",
Frank Barchard0bb49a72020-06-04 11:35:11 -07002198 "src/f16-gavgpool/7p7x-minmax-neonfp16arith-c8.c",
2199 "src/f16-gavgpool/7x-minmax-neonfp16arith-c8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002200 "src/f16-gemm/gen-inc/1x8inc-minmax-neonfp16arith-ld64.c",
2201 "src/f16-gemm/gen-inc/1x16inc-minmax-neonfp16arith-ld64.c",
2202 "src/f16-gemm/gen-inc/4x8inc-minmax-neonfp16arith-ld64.c",
2203 "src/f16-gemm/gen-inc/4x16inc-minmax-neonfp16arith-ld64.c",
2204 "src/f16-gemm/gen-inc/6x8inc-minmax-neonfp16arith-ld64.c",
2205 "src/f16-gemm/gen-inc/6x16inc-minmax-neonfp16arith-ld64.c",
2206 "src/f16-gemm/gen-inc/8x8inc-minmax-neonfp16arith-ld64.c",
2207 "src/f16-gemm/gen-inc/8x16inc-minmax-neonfp16arith-ld64.c",
2208 "src/f16-gemm/gen/1x8-minmax-neonfp16arith-ld64.c",
2209 "src/f16-gemm/gen/1x16-minmax-neonfp16arith-ld64.c",
2210 "src/f16-gemm/gen/4x8-minmax-neonfp16arith-ld64.c",
2211 "src/f16-gemm/gen/4x16-minmax-neonfp16arith-ld64.c",
2212 "src/f16-gemm/gen/6x8-minmax-neonfp16arith-ld64.c",
2213 "src/f16-gemm/gen/6x16-minmax-neonfp16arith-ld64.c",
2214 "src/f16-gemm/gen/8x8-minmax-neonfp16arith-ld64.c",
2215 "src/f16-gemm/gen/8x16-minmax-neonfp16arith-ld64.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07002216 "src/f16-hswish/gen/hswish-neonfp16arith-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002217 "src/f16-hswish/gen/hswish-neonfp16arith-x16.c",
2218 "src/f16-igemm/gen/1x8-minmax-neonfp16arith-ld64.c",
2219 "src/f16-igemm/gen/1x16-minmax-neonfp16arith-ld64.c",
2220 "src/f16-igemm/gen/4x8-minmax-neonfp16arith-ld64.c",
2221 "src/f16-igemm/gen/4x16-minmax-neonfp16arith-ld64.c",
2222 "src/f16-igemm/gen/6x8-minmax-neonfp16arith-ld64.c",
2223 "src/f16-igemm/gen/6x16-minmax-neonfp16arith-ld64.c",
2224 "src/f16-igemm/gen/8x8-minmax-neonfp16arith-ld64.c",
2225 "src/f16-igemm/gen/8x16-minmax-neonfp16arith-ld64.c",
Frank Barchardb1966592020-05-12 13:47:06 -07002226 "src/f16-prelu/gen/neonfp16arith-2x8.c",
Frank Barcharda5316982020-07-23 13:19:28 -07002227 "src/f16-prelu/gen/neonfp16arith-2x16.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07002228 "src/f16-relu/gen/neonfp16arith-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002229 "src/f16-relu/gen/neonfp16arith-x16.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07002230 "src/f16-spmm/gen/8x1-minmax-neonfp16arith-x2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002231 "src/f16-spmm/gen/8x1-minmax-neonfp16arith.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07002232 "src/f16-spmm/gen/16x1-minmax-neonfp16arith-x2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002233 "src/f16-spmm/gen/16x1-minmax-neonfp16arith.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07002234 "src/f16-spmm/gen/24x1-minmax-neonfp16arith-x2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002235 "src/f16-spmm/gen/24x1-minmax-neonfp16arith.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07002236 "src/f16-spmm/gen/32x1-minmax-neonfp16arith-x2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002237 "src/f16-spmm/gen/32x1-minmax-neonfp16arith.c",
2238 "src/f16-vbinary/gen/vadd-minmax-neonfp16arith-x8.c",
2239 "src/f16-vbinary/gen/vadd-minmax-neonfp16arith-x16.c",
2240 "src/f16-vbinary/gen/vaddc-minmax-neonfp16arith-x8.c",
2241 "src/f16-vbinary/gen/vaddc-minmax-neonfp16arith-x16.c",
2242 "src/f16-vbinary/gen/vdiv-minmax-neonfp16arith-x8.c",
2243 "src/f16-vbinary/gen/vdiv-minmax-neonfp16arith-x16.c",
2244 "src/f16-vbinary/gen/vdivc-minmax-neonfp16arith-x8.c",
2245 "src/f16-vbinary/gen/vdivc-minmax-neonfp16arith-x16.c",
2246 "src/f16-vbinary/gen/vmax-neonfp16arith-x8.c",
2247 "src/f16-vbinary/gen/vmax-neonfp16arith-x16.c",
2248 "src/f16-vbinary/gen/vmaxc-neonfp16arith-x8.c",
2249 "src/f16-vbinary/gen/vmaxc-neonfp16arith-x16.c",
2250 "src/f16-vbinary/gen/vmin-neonfp16arith-x8.c",
2251 "src/f16-vbinary/gen/vmin-neonfp16arith-x16.c",
2252 "src/f16-vbinary/gen/vminc-neonfp16arith-x8.c",
2253 "src/f16-vbinary/gen/vminc-neonfp16arith-x16.c",
2254 "src/f16-vbinary/gen/vmul-minmax-neonfp16arith-x8.c",
2255 "src/f16-vbinary/gen/vmul-minmax-neonfp16arith-x16.c",
2256 "src/f16-vbinary/gen/vmulc-minmax-neonfp16arith-x8.c",
2257 "src/f16-vbinary/gen/vmulc-minmax-neonfp16arith-x16.c",
2258 "src/f16-vbinary/gen/vrdivc-minmax-neonfp16arith-x8.c",
2259 "src/f16-vbinary/gen/vrdivc-minmax-neonfp16arith-x16.c",
2260 "src/f16-vbinary/gen/vrsubc-minmax-neonfp16arith-x8.c",
2261 "src/f16-vbinary/gen/vrsubc-minmax-neonfp16arith-x16.c",
2262 "src/f16-vbinary/gen/vsub-minmax-neonfp16arith-x8.c",
2263 "src/f16-vbinary/gen/vsub-minmax-neonfp16arith-x16.c",
2264 "src/f16-vbinary/gen/vsubc-minmax-neonfp16arith-x8.c",
2265 "src/f16-vbinary/gen/vsubc-minmax-neonfp16arith-x16.c",
2266 "src/f16-vmulcaddc/gen/c8-minmax-neonfp16arith-2x.c",
2267 "src/f16-vmulcaddc/gen/c16-minmax-neonfp16arith-2x.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002268]
2269
Benoit Jacoba9644732020-08-13 12:48:55 -07002270NEONDOT_UKERNELS = [
Marat Dukhand4c83032020-09-09 23:20:12 -07002271 "src/qs8-gemm/gen/1x8c4-minmax-neondot.c",
Marat Dukhand4c83032020-09-09 23:20:12 -07002272 "src/qs8-gemm/gen/1x16c4-minmax-neondot.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002273 "src/qs8-gemm/gen/4x8c4-minmax-neondot.c",
Benoit Jacoba9644732020-08-13 12:48:55 -07002274 "src/qs8-gemm/gen/4x16c4-minmax-neondot.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002275 "src/qs8-gemm/gen/6x8c4-minmax-neondot.c",
Frank Barchard2fa17452020-09-22 09:51:46 -07002276 "src/qs8-gemm/gen/6x16c4-minmax-neondot.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002277 "src/qs8-gemm/gen/8x8c4-minmax-neondot.c",
Frank Barcharda48848f2020-09-24 00:59:24 -07002278 "src/qs8-gemm/gen/8x16c4-minmax-neondot.c",
Marat Dukhand4c83032020-09-09 23:20:12 -07002279 "src/qs8-igemm/gen/1x8c4-minmax-neondot.c",
Marat Dukhand4c83032020-09-09 23:20:12 -07002280 "src/qs8-igemm/gen/1x16c4-minmax-neondot.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002281 "src/qs8-igemm/gen/4x8c4-minmax-neondot.c",
Benoit Jacoba05487f2020-08-14 13:37:21 -07002282 "src/qs8-igemm/gen/4x16c4-minmax-neondot.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002283 "src/qs8-igemm/gen/6x8c4-minmax-neondot.c",
Frank Barchard2fa17452020-09-22 09:51:46 -07002284 "src/qs8-igemm/gen/6x16c4-minmax-neondot.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002285 "src/qs8-igemm/gen/8x8c4-minmax-neondot.c",
Frank Barcharda48848f2020-09-24 00:59:24 -07002286 "src/qs8-igemm/gen/8x16c4-minmax-neondot.c",
Benoit Jacoba9644732020-08-13 12:48:55 -07002287]
2288
Marat Dukhan08c4a432019-10-03 09:29:21 -07002289SSE_UKERNELS = [
Marat Dukhan99936602020-04-11 16:47:01 -07002290 "src/f32-avgpool/9p8x-minmax-sse-c4.c",
2291 "src/f32-avgpool/9x-minmax-sse-c4.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -07002292 "src/f32-clamp/gen/sse-x4.c",
2293 "src/f32-clamp/gen/sse-x8.c",
Erich Elsenb1233402020-06-08 15:53:15 -07002294 "src/f32-conv-hwc2chw/3x3s2p1c3x4-sse-1x1.c",
2295 "src/f32-conv-hwc2chw/3x3s2p1c3x4-sse-2x2.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002296 "src/f32-dwconv/gen/up4x4-minmax-sse-acc2.c",
2297 "src/f32-dwconv/gen/up4x4-minmax-sse.c",
2298 "src/f32-dwconv/gen/up4x9-minmax-sse-acc2.c",
2299 "src/f32-dwconv/gen/up4x9-minmax-sse.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002300 "src/f32-dwconv/gen/up4x25-minmax-sse-acc2.c",
2301 "src/f32-dwconv/gen/up4x25-minmax-sse.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002302 "src/f32-dwconv/gen/up8x4-minmax-sse-acc2.c",
2303 "src/f32-dwconv/gen/up8x4-minmax-sse.c",
2304 "src/f32-dwconv/gen/up8x9-minmax-sse-acc2.c",
2305 "src/f32-dwconv/gen/up8x9-minmax-sse.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002306 "src/f32-dwconv/gen/up8x25-minmax-sse-acc2.c",
2307 "src/f32-dwconv/gen/up8x25-minmax-sse.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07002308 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-1x4-acc2.c",
2309 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-1x4-acc3.c",
2310 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-1x4-acc4.c",
Marat Dukhan470078a2020-10-23 22:36:52 -07002311 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-1x4.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07002312 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-2x4-acc2.c",
Marat Dukhan470078a2020-10-23 22:36:52 -07002313 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-2x4.c",
2314 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-3x4.c",
2315 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-4x4.c",
2316 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-5x4.c",
2317 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-6x4.c",
Marat Dukhan0ff97182020-10-25 19:14:03 -07002318 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-sse-1x4-acc2.c",
2319 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-sse-1x4-acc3.c",
2320 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-sse-1x4-acc4.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07002321 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-sse-1x4.c",
Marat Dukhan0ff97182020-10-25 19:14:03 -07002322 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-sse-2x4-acc2.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07002323 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-sse-2x4.c",
2324 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-sse-3x4.c",
2325 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-sse-4x4.c",
Marat Dukhand0503892020-10-30 08:22:04 -07002326 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-1x4-acc2.c",
2327 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-1x4-acc3.c",
2328 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-1x4-acc4.c",
2329 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-1x4-acc5.c",
2330 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-1x4.c",
2331 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-2x4-acc2.c",
2332 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-2x4-acc3.c",
2333 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-2x4.c",
2334 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-3x4-acc2.c",
2335 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-3x4.c",
2336 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-4x4-acc2.c",
2337 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-4x4.c",
2338 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-5x4.c",
Marat Dukhanccca2142020-10-30 17:32:45 -07002339 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-sse-1x4-acc2.c",
2340 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-sse-1x4-acc3.c",
2341 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-sse-1x4-acc4.c",
2342 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-sse-1x4-acc5.c",
2343 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-sse-1x4.c",
2344 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-sse-2x4-acc2.c",
2345 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-sse-2x4-acc3.c",
2346 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-sse-2x4.c",
Marat Dukhanccca2142020-10-30 17:32:45 -07002347 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-sse-3x4-acc2.c",
Frank Barchard8ef44cd2020-11-03 12:30:23 -08002348 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-sse-3x4.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -07002349 "src/f32-gavgpool-cw/sse-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -07002350 "src/f32-gavgpool/7p7x-minmax-sse-c4.c",
2351 "src/f32-gavgpool/7x-minmax-sse-c4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002352 "src/f32-gemm/gen-inc/1x8inc-minmax-sse-dup.c",
2353 "src/f32-gemm/gen-inc/1x8inc-minmax-sse-load1.c",
2354 "src/f32-gemm/gen-inc/1x8s4inc-minmax-sse.c",
Marat Dukhan802fcae2020-12-11 14:37:25 -08002355 "src/f32-gemm/gen-inc/3x8inc-minmax-sse-dup.c",
2356 "src/f32-gemm/gen-inc/3x8inc-minmax-sse-load1.c",
2357 "src/f32-gemm/gen-inc/3x8s4inc-minmax-sse.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002358 "src/f32-gemm/gen-inc/4x8inc-minmax-sse-dup.c",
2359 "src/f32-gemm/gen-inc/4x8inc-minmax-sse-load1.c",
2360 "src/f32-gemm/gen-inc/4x8s4inc-minmax-sse.c",
Marat Dukhan802fcae2020-12-11 14:37:25 -08002361 "src/f32-gemm/gen-inc/5x8inc-minmax-sse-dup.c",
2362 "src/f32-gemm/gen-inc/5x8inc-minmax-sse-load1.c",
2363 "src/f32-gemm/gen-inc/5x8s4inc-minmax-sse.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002364 "src/f32-gemm/gen/1x8-minmax-sse-dup.c",
2365 "src/f32-gemm/gen/1x8-minmax-sse-load1.c",
2366 "src/f32-gemm/gen/1x8s4-minmax-sse.c",
Marat Dukhan802fcae2020-12-11 14:37:25 -08002367 "src/f32-gemm/gen/3x8-minmax-sse-dup.c",
2368 "src/f32-gemm/gen/3x8-minmax-sse-load1.c",
2369 "src/f32-gemm/gen/3x8s4-minmax-sse.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002370 "src/f32-gemm/gen/4x2c4-minmax-sse.c",
2371 "src/f32-gemm/gen/4x8-minmax-sse-dup.c",
2372 "src/f32-gemm/gen/4x8-minmax-sse-load1.c",
2373 "src/f32-gemm/gen/4x8s4-minmax-sse.c",
Marat Dukhan802fcae2020-12-11 14:37:25 -08002374 "src/f32-gemm/gen/5x8-minmax-sse-dup.c",
2375 "src/f32-gemm/gen/5x8-minmax-sse-load1.c",
2376 "src/f32-gemm/gen/5x8s4-minmax-sse.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07002377 "src/f32-hswish/gen/hswish-sse-x4.c",
2378 "src/f32-hswish/gen/hswish-sse-x8.c",
Marat Dukhan660fd192020-03-10 04:55:30 -07002379 "src/f32-ibilinear/gen/sse-c4.c",
2380 "src/f32-ibilinear/gen/sse-c8.c",
Artsiom Ablavatskib3ffd582021-03-31 13:00:08 -07002381 "src/f32-ibilinear-chw/gen/sse-p4.c",
2382 "src/f32-ibilinear-chw/gen/sse-p8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002383 "src/f32-igemm/gen/1x8-minmax-sse-dup.c",
2384 "src/f32-igemm/gen/1x8-minmax-sse-load1.c",
2385 "src/f32-igemm/gen/1x8s4-minmax-sse.c",
Marat Dukhan802fcae2020-12-11 14:37:25 -08002386 "src/f32-igemm/gen/3x8-minmax-sse-dup.c",
2387 "src/f32-igemm/gen/3x8-minmax-sse-load1.c",
2388 "src/f32-igemm/gen/3x8s4-minmax-sse.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002389 "src/f32-igemm/gen/4x2c4-minmax-sse.c",
2390 "src/f32-igemm/gen/4x8-minmax-sse-dup.c",
2391 "src/f32-igemm/gen/4x8-minmax-sse-load1.c",
2392 "src/f32-igemm/gen/4x8s4-minmax-sse.c",
Marat Dukhan802fcae2020-12-11 14:37:25 -08002393 "src/f32-igemm/gen/5x8-minmax-sse-dup.c",
2394 "src/f32-igemm/gen/5x8-minmax-sse-load1.c",
2395 "src/f32-igemm/gen/5x8s4-minmax-sse.c",
Marat Dukhan99936602020-04-11 16:47:01 -07002396 "src/f32-maxpool/9p8x-minmax-sse-c4.c",
2397 "src/f32-pavgpool/9p8x-minmax-sse-c4.c",
2398 "src/f32-pavgpool/9x-minmax-sse-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002399 "src/f32-ppmm/gen/4x8-minmax-sse.c",
Marat Dukhan39b5e942020-06-24 15:03:48 -07002400 "src/f32-prelu/gen/sse-2x4.c",
2401 "src/f32-prelu/gen/sse-2x8.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07002402 "src/f32-relu/gen/sse-x4.c",
2403 "src/f32-relu/gen/sse-x8.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002404 "src/f32-rmax/sse.c",
Marat Dukhan355ab432020-04-09 19:01:52 -07002405 "src/f32-spmm/gen/4x1-minmax-sse.c",
2406 "src/f32-spmm/gen/8x1-minmax-sse.c",
Erich Elsen6e80fdc2020-06-09 15:35:37 -07002407 "src/f32-spmm/gen/16x1-minmax-sse.c",
Frank Barchard846c0c62020-10-26 15:01:39 -07002408 "src/f32-spmm/gen/32x1-minmax-sse.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002409 "src/f32-vbinary/gen/vadd-minmax-sse-x4.c",
2410 "src/f32-vbinary/gen/vadd-minmax-sse-x8.c",
2411 "src/f32-vbinary/gen/vaddc-minmax-sse-x4.c",
2412 "src/f32-vbinary/gen/vaddc-minmax-sse-x8.c",
2413 "src/f32-vbinary/gen/vdiv-minmax-sse-x4.c",
2414 "src/f32-vbinary/gen/vdiv-minmax-sse-x8.c",
2415 "src/f32-vbinary/gen/vdivc-minmax-sse-x4.c",
2416 "src/f32-vbinary/gen/vdivc-minmax-sse-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -08002417 "src/f32-vbinary/gen/vmax-sse-x4.c",
2418 "src/f32-vbinary/gen/vmax-sse-x8.c",
2419 "src/f32-vbinary/gen/vmaxc-sse-x4.c",
2420 "src/f32-vbinary/gen/vmaxc-sse-x8.c",
2421 "src/f32-vbinary/gen/vmin-sse-x4.c",
2422 "src/f32-vbinary/gen/vmin-sse-x8.c",
2423 "src/f32-vbinary/gen/vminc-sse-x4.c",
2424 "src/f32-vbinary/gen/vminc-sse-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002425 "src/f32-vbinary/gen/vmul-minmax-sse-x4.c",
2426 "src/f32-vbinary/gen/vmul-minmax-sse-x8.c",
2427 "src/f32-vbinary/gen/vmulc-minmax-sse-x4.c",
2428 "src/f32-vbinary/gen/vmulc-minmax-sse-x8.c",
2429 "src/f32-vbinary/gen/vrdivc-minmax-sse-x4.c",
2430 "src/f32-vbinary/gen/vrdivc-minmax-sse-x8.c",
2431 "src/f32-vbinary/gen/vrsubc-minmax-sse-x4.c",
2432 "src/f32-vbinary/gen/vrsubc-minmax-sse-x8.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07002433 "src/f32-vbinary/gen/vsqrdiff-sse-x4.c",
2434 "src/f32-vbinary/gen/vsqrdiff-sse-x8.c",
2435 "src/f32-vbinary/gen/vsqrdiffc-sse-x4.c",
2436 "src/f32-vbinary/gen/vsqrdiffc-sse-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002437 "src/f32-vbinary/gen/vsub-minmax-sse-x4.c",
2438 "src/f32-vbinary/gen/vsub-minmax-sse-x8.c",
2439 "src/f32-vbinary/gen/vsubc-minmax-sse-x4.c",
2440 "src/f32-vbinary/gen/vsubc-minmax-sse-x8.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07002441 "src/f32-vlrelu/gen/vlrelu-sse-x4.c",
2442 "src/f32-vlrelu/gen/vlrelu-sse-x8.c",
Marat Dukhan99936602020-04-11 16:47:01 -07002443 "src/f32-vmulcaddc/gen/c4-minmax-sse-2x.c",
2444 "src/f32-vmulcaddc/gen/c8-minmax-sse-2x.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07002445 "src/f32-vsqrt/gen/sse-sqrt-x4.c",
2446 "src/f32-vsqrt/gen/sse-sqrt-x8.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07002447 "src/f32-vunary/gen/vabs-sse-x4.c",
2448 "src/f32-vunary/gen/vabs-sse-x8.c",
2449 "src/f32-vunary/gen/vneg-sse-x4.c",
2450 "src/f32-vunary/gen/vneg-sse-x8.c",
2451 "src/f32-vunary/gen/vsqr-sse-x4.c",
2452 "src/f32-vunary/gen/vsqr-sse-x8.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07002453 "src/math/roundd-sse-addsub.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002454 "src/math/roundne-sse-addsub.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07002455 "src/math/roundu-sse-addsub.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07002456 "src/math/roundz-sse-addsub.c",
Marat Dukhan84000762020-06-29 18:38:43 -07002457 "src/math/sqrt-sse-hh1mac.c",
2458 "src/math/sqrt-sse-nr1mac.c",
2459 "src/math/sqrt-sse-nr2mac.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002460 "src/x32-fill/sse.c",
2461 "src/x32-packx/x4-sse.c",
2462 "src/x32-pad/sse.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002463]
2464
2465SSE2_UKERNELS = [
Marat Dukhan329da642019-11-19 21:44:39 -08002466 "src/f32-argmaxpool/4x-sse2-c4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002467 "src/f32-argmaxpool/9p8x-sse2-c4.c",
Marat Dukhan329da642019-11-19 21:44:39 -08002468 "src/f32-argmaxpool/9x-sse2-c4.c",
Marat Dukhan802fcae2020-12-11 14:37:25 -08002469 "src/f32-gemm/gen-inc/1x8inc-minmax-sse2-dup.c",
2470 "src/f32-gemm/gen-inc/3x8inc-minmax-sse2-dup.c",
2471 "src/f32-gemm/gen-inc/4x8inc-minmax-sse2-dup.c",
2472 "src/f32-gemm/gen-inc/5x8inc-minmax-sse2-dup.c",
2473 "src/f32-gemm/gen/1x8-minmax-sse2-dup.c",
2474 "src/f32-gemm/gen/3x8-minmax-sse2-dup.c",
2475 "src/f32-gemm/gen/4x8-minmax-sse2-dup.c",
2476 "src/f32-gemm/gen/5x8-minmax-sse2-dup.c",
2477 "src/f32-igemm/gen/1x8-minmax-sse2-dup.c",
2478 "src/f32-igemm/gen/3x8-minmax-sse2-dup.c",
2479 "src/f32-igemm/gen/4x8-minmax-sse2-dup.c",
2480 "src/f32-igemm/gen/5x8-minmax-sse2-dup.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08002481 "src/f32-prelu/gen/sse2-2x4.c",
2482 "src/f32-prelu/gen/sse2-2x8.c",
Marat Dukhanb39689d2020-01-24 13:32:20 -08002483 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x4.c",
Marat Dukhanb39689d2020-01-24 13:32:20 -08002484 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x8-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002485 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x8.c",
Marat Dukhanb39689d2020-01-24 13:32:20 -08002486 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12-acc2.c",
2487 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002488 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12.c",
Marat Dukhanb39689d2020-01-24 13:32:20 -08002489 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16-acc2.c",
2490 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002491 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16.c",
Marat Dukhanb39689d2020-01-24 13:32:20 -08002492 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20-acc2.c",
2493 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002494 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20.c",
Marat Dukhand243c1a2020-09-17 18:48:10 -07002495 "src/f32-sigmoid/gen/sse2-lut64-p2-div-x4.c",
2496 "src/f32-sigmoid/gen/sse2-lut64-p2-div-x8.c",
2497 "src/f32-sigmoid/gen/sse2-lut64-p2-div-x12.c",
2498 "src/f32-sigmoid/gen/sse2-lut64-p2-div-x16.c",
2499 "src/f32-sigmoid/gen/sse2-lut64-p2-div-x20.c",
2500 "src/f32-sigmoid/gen/sse2-lut64-p2-div-x24.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002501 "src/f32-sigmoid/gen/sse2-p5-div-x4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08002502 "src/f32-sigmoid/gen/sse2-p5-div-x8.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002503 "src/f32-sigmoid/gen/sse2-p5-div-x12.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08002504 "src/f32-sigmoid/gen/sse2-p5-div-x16.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002505 "src/f32-sigmoid/gen/sse2-p5-div-x20.c",
2506 "src/f32-sigmoid/gen/sse2-p5-div-x24.c",
Marat Dukhaned6baaf2020-12-01 15:07:08 -08002507 "src/f32-velu/gen/velu-sse2-rr2-lut16-p3-x4.c",
2508 "src/f32-velu/gen/velu-sse2-rr2-lut16-p3-x8.c",
2509 "src/f32-velu/gen/velu-sse2-rr2-lut16-p3-x12.c",
2510 "src/f32-velu/gen/velu-sse2-rr2-lut16-p3-x16.c",
2511 "src/f32-velu/gen/velu-sse2-rr2-lut16-p3-x20.c",
2512 "src/f32-velu/gen/velu-sse2-rr2-lut16-p3-x24.c",
2513 "src/f32-velu/gen/velu-sse2-rr2-p6-x4.c",
2514 "src/f32-velu/gen/velu-sse2-rr2-p6-x8.c",
2515 "src/f32-velu/gen/velu-sse2-rr2-p6-x12.c",
2516 "src/f32-velu/gen/velu-sse2-rr2-p6-x16.c",
2517 "src/f32-velu/gen/velu-sse2-rr2-p6-x20.c",
2518 "src/f32-velu/gen/velu-sse2-rr2-p6-x24.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07002519 "src/f32-vlrelu/gen/vlrelu-sse2-x4.c",
2520 "src/f32-vlrelu/gen/vlrelu-sse2-x8.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07002521 "src/f32-vrnd/gen/vrndd-sse2-x4.c",
2522 "src/f32-vrnd/gen/vrndd-sse2-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002523 "src/f32-vrnd/gen/vrndne-sse2-x4.c",
2524 "src/f32-vrnd/gen/vrndne-sse2-x8.c",
2525 "src/f32-vrnd/gen/vrndu-sse2-x4.c",
2526 "src/f32-vrnd/gen/vrndu-sse2-x8.c",
2527 "src/f32-vrnd/gen/vrndz-sse2-x4.c",
2528 "src/f32-vrnd/gen/vrndz-sse2-x8.c",
Marat Dukhanb7633f22020-11-20 16:34:56 -08002529 "src/math/exp-sse2-rr2-lut64-p2.c",
2530 "src/math/exp-sse2-rr2-p5.c",
Marat Dukhande390d42020-11-29 19:32:18 -08002531 "src/math/expm1minus-sse2-rr2-lut16-p3.c",
Marat Dukhana438aca2020-11-20 15:45:01 -08002532 "src/math/expm1minus-sse2-rr2-p6.c",
Frank Barchard3b800452020-11-22 12:12:35 -08002533 "src/math/expminus-sse2-rr2-p5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002534 "src/math/roundd-sse2-cvt.c",
2535 "src/math/roundne-sse2-cvt.c",
2536 "src/math/roundu-sse2-cvt.c",
2537 "src/math/roundz-sse2-cvt.c",
2538 "src/math/sigmoid-sse2-rr2-lut64-p2-div.c",
2539 "src/math/sigmoid-sse2-rr2-lut64-p2-nr1.c",
2540 "src/math/sigmoid-sse2-rr2-lut64-p2-nr2.c",
2541 "src/math/sigmoid-sse2-rr2-p5-div.c",
2542 "src/math/sigmoid-sse2-rr2-p5-nr1.c",
2543 "src/math/sigmoid-sse2-rr2-p5-nr2.c",
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07002544 "src/qs8-dwconv/gen/up8x9-minmax-sse2-mul16.c",
2545 "src/qs8-dwconv/gen/up16x9-minmax-sse2-mul16.c",
2546 "src/qs8-dwconv/gen/up24x9-minmax-sse2-mul16.c",
Marat Dukhan159688f2020-08-06 10:34:29 -07002547 "src/qs8-gavgpool/gen/7p7x-minmax-sse2-c8-acc2.c",
2548 "src/qs8-gavgpool/gen/7p7x-minmax-sse2-c16-acc2.c",
2549 "src/qs8-gavgpool/gen/7p7x-minmax-sse2-c24-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002550 "src/qs8-gavgpool/gen/7x-minmax-sse2-c8-acc2.c",
2551 "src/qs8-gavgpool/gen/7x-minmax-sse2-c16-acc2.c",
2552 "src/qs8-gavgpool/gen/7x-minmax-sse2-c24-acc2.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002553 "src/qs8-gemm/gen/1x4c2-minmax-sse2-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002554 "src/qs8-gemm/gen/1x4c2-minmax-sse2-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002555 "src/qs8-gemm/gen/1x4c2-xw-minmax-sse2.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002556 "src/qs8-gemm/gen/1x4c8-minmax-sse2-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002557 "src/qs8-gemm/gen/1x4c8-minmax-sse2-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002558 "src/qs8-gemm/gen/1x4c8-xw-minmax-sse2.c",
Marat Dukhan2e427872021-04-09 23:40:07 -07002559 "src/qs8-gemm/gen/2x4c2-minmax-sse2-ld64.c",
2560 "src/qs8-gemm/gen/2x4c2-minmax-sse2-ld128.c",
2561 "src/qs8-gemm/gen/2x4c2-xw-minmax-sse2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002562 "src/qs8-gemm/gen/2x4c8-minmax-sse2-ld64.c",
2563 "src/qs8-gemm/gen/2x4c8-minmax-sse2-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002564 "src/qs8-gemm/gen/2x4c8-xw-minmax-sse2.c",
Marat Dukhan2e427872021-04-09 23:40:07 -07002565 "src/qs8-gemm/gen/3x4c2-minmax-sse2-ld64.c",
2566 "src/qs8-gemm/gen/3x4c2-minmax-sse2-ld128.c",
2567 "src/qs8-gemm/gen/3x4c2-xw-minmax-sse2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002568 "src/qs8-gemm/gen/3x4c8-minmax-sse2-ld64.c",
2569 "src/qs8-gemm/gen/3x4c8-minmax-sse2-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002570 "src/qs8-gemm/gen/3x4c8-xw-minmax-sse2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002571 "src/qs8-gemm/gen/4x4c2-minmax-sse2-ld64.c",
2572 "src/qs8-gemm/gen/4x4c2-minmax-sse2-ld128.c",
2573 "src/qs8-gemm/gen/4x4c2-xw-minmax-sse2.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002574 "src/qs8-igemm/gen/1x4c2-minmax-sse2-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002575 "src/qs8-igemm/gen/1x4c2-minmax-sse2-ld128.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002576 "src/qs8-igemm/gen/1x4c8-minmax-sse2-ld64.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002577 "src/qs8-igemm/gen/1x4c8-minmax-sse2-ld128.c",
Marat Dukhan2e427872021-04-09 23:40:07 -07002578 "src/qs8-igemm/gen/2x4c2-minmax-sse2-ld64.c",
2579 "src/qs8-igemm/gen/2x4c2-minmax-sse2-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002580 "src/qs8-igemm/gen/2x4c8-minmax-sse2-ld64.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002581 "src/qs8-igemm/gen/2x4c8-minmax-sse2-ld128.c",
Marat Dukhan2e427872021-04-09 23:40:07 -07002582 "src/qs8-igemm/gen/3x4c2-minmax-sse2-ld64.c",
2583 "src/qs8-igemm/gen/3x4c2-minmax-sse2-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002584 "src/qs8-igemm/gen/3x4c8-minmax-sse2-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002585 "src/qs8-igemm/gen/3x4c8-minmax-sse2-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002586 "src/qs8-igemm/gen/4x4c2-minmax-sse2-ld64.c",
2587 "src/qs8-igemm/gen/4x4c2-minmax-sse2-ld128.c",
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07002588 "src/qs8-requantization/fp32-sse2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002589 "src/qs8-requantization/precise-sse2.c",
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07002590 "src/qs8-requantization/q31-sse2.c",
Marat Dukhand9f3ad42020-08-10 12:30:58 -07002591 "src/qs8-vadd/gen/minmax-sse2-mul16-ld64-x8.c",
2592 "src/qs8-vadd/gen/minmax-sse2-mul16-ld64-x16.c",
2593 "src/qs8-vadd/gen/minmax-sse2-mul16-ld64-x24.c",
2594 "src/qs8-vadd/gen/minmax-sse2-mul16-ld64-x32.c",
Marat Dukhan0270d9f2020-08-11 00:56:46 -07002595 "src/qs8-vaddc/gen/minmax-sse2-mul16-ld64-x8.c",
2596 "src/qs8-vaddc/gen/minmax-sse2-mul16-ld64-x16.c",
2597 "src/qs8-vaddc/gen/minmax-sse2-mul16-ld64-x24.c",
2598 "src/qs8-vaddc/gen/minmax-sse2-mul16-ld64-x32.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -07002599 "src/qu8-avgpool/9p8x-minmax-sse2-c8.c",
2600 "src/qu8-avgpool/9x-minmax-sse2-c8.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -07002601 "src/qu8-dwconv/up8x9-minmax-sse2.c",
2602 "src/qu8-gavgpool/7p7x-minmax-sse2-c8.c",
2603 "src/qu8-gavgpool/7x-minmax-sse2-c8.c",
2604 "src/qu8-gemm/2x4c8-minmax-sse2.c",
2605 "src/qu8-gemm/4x4c2-minmax-sse2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002606 "src/qu8-igemm/4x4c2-minmax-sse2.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07002607 "src/qu8-requantization/fp32-sse2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002608 "src/qu8-requantization/precise-sse2.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07002609 "src/qu8-requantization/q31-sse2.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -07002610 "src/qu8-vadd/minmax-sse2.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -07002611 "src/u8-clamp/sse2-x64.c",
Marat Dukhan99936602020-04-11 16:47:01 -07002612 "src/u8-maxpool/9p8x-minmax-sse2-c16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002613 "src/u8-rmax/sse2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002614 "src/x8-zip/x2-sse2.c",
2615 "src/x8-zip/x3-sse2.c",
2616 "src/x8-zip/x4-sse2.c",
2617 "src/x8-zip/xm-sse2.c",
Marat Dukhan57dccd82020-04-14 00:53:10 -07002618 "src/x32-unpool/sse2.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002619 "src/x32-zip/x2-sse2.c",
2620 "src/x32-zip/x3-sse2.c",
2621 "src/x32-zip/x4-sse2.c",
2622 "src/x32-zip/xm-sse2.c",
Marat Dukhanfe7acb62020-03-09 19:30:05 -07002623]
2624
2625SSSE3_UKERNELS = [
Frank Barchard35db7d02020-10-26 13:37:34 -07002626 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-1x4-acc2.c",
2627 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-1x4-acc3.c",
2628 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-1x4-acc4.c",
Marat Dukhan98f2eeb2020-10-23 23:13:41 -07002629 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-1x4.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07002630 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-2x4-acc2.c",
Marat Dukhan98f2eeb2020-10-23 23:13:41 -07002631 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-2x4.c",
2632 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-3x4.c",
2633 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-4x4.c",
2634 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-5x4.c",
2635 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-6x4.c",
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07002636 "src/qs8-dwconv/gen/up8x9-minmax-ssse3-mul16.c",
2637 "src/qs8-dwconv/gen/up16x9-minmax-ssse3-mul16.c",
2638 "src/qs8-dwconv/gen/up24x9-minmax-ssse3-mul16.c",
Marat Dukhan159688f2020-08-06 10:34:29 -07002639 "src/qs8-gavgpool/gen/7p7x-minmax-ssse3-c8-acc2.c",
2640 "src/qs8-gavgpool/gen/7p7x-minmax-ssse3-c16-acc2.c",
2641 "src/qs8-gavgpool/gen/7p7x-minmax-ssse3-c24-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002642 "src/qs8-gavgpool/gen/7x-minmax-ssse3-c8-acc2.c",
2643 "src/qs8-gavgpool/gen/7x-minmax-ssse3-c16-acc2.c",
2644 "src/qs8-gavgpool/gen/7x-minmax-ssse3-c24-acc2.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002645 "src/qs8-gemm/gen/1x4c2-minmax-ssse3-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002646 "src/qs8-gemm/gen/1x4c2-minmax-ssse3-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002647 "src/qs8-gemm/gen/1x4c2-xw-minmax-ssse3.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002648 "src/qs8-gemm/gen/1x4c8-minmax-ssse3-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002649 "src/qs8-gemm/gen/1x4c8-minmax-ssse3-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002650 "src/qs8-gemm/gen/1x4c8-xw-minmax-ssse3.c",
Marat Dukhan2e427872021-04-09 23:40:07 -07002651 "src/qs8-gemm/gen/2x4c2-minmax-ssse3-ld64.c",
2652 "src/qs8-gemm/gen/2x4c2-minmax-ssse3-ld128.c",
2653 "src/qs8-gemm/gen/2x4c2-xw-minmax-ssse3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002654 "src/qs8-gemm/gen/2x4c8-minmax-ssse3-ld64.c",
2655 "src/qs8-gemm/gen/2x4c8-minmax-ssse3-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002656 "src/qs8-gemm/gen/2x4c8-xw-minmax-ssse3.c",
Marat Dukhan2e427872021-04-09 23:40:07 -07002657 "src/qs8-gemm/gen/3x4c2-minmax-ssse3-ld64.c",
2658 "src/qs8-gemm/gen/3x4c2-minmax-ssse3-ld128.c",
2659 "src/qs8-gemm/gen/3x4c2-xw-minmax-ssse3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002660 "src/qs8-gemm/gen/3x4c8-minmax-ssse3-ld64.c",
2661 "src/qs8-gemm/gen/3x4c8-minmax-ssse3-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002662 "src/qs8-gemm/gen/3x4c8-xw-minmax-ssse3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002663 "src/qs8-gemm/gen/4x4c2-minmax-ssse3-ld64.c",
2664 "src/qs8-gemm/gen/4x4c2-minmax-ssse3-ld128.c",
2665 "src/qs8-gemm/gen/4x4c2-xw-minmax-ssse3.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002666 "src/qs8-igemm/gen/1x4c2-minmax-ssse3-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002667 "src/qs8-igemm/gen/1x4c2-minmax-ssse3-ld128.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002668 "src/qs8-igemm/gen/1x4c8-minmax-ssse3-ld64.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002669 "src/qs8-igemm/gen/1x4c8-minmax-ssse3-ld128.c",
Marat Dukhan2e427872021-04-09 23:40:07 -07002670 "src/qs8-igemm/gen/2x4c2-minmax-ssse3-ld64.c",
2671 "src/qs8-igemm/gen/2x4c2-minmax-ssse3-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002672 "src/qs8-igemm/gen/2x4c8-minmax-ssse3-ld64.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002673 "src/qs8-igemm/gen/2x4c8-minmax-ssse3-ld128.c",
Marat Dukhan2e427872021-04-09 23:40:07 -07002674 "src/qs8-igemm/gen/3x4c2-minmax-ssse3-ld64.c",
2675 "src/qs8-igemm/gen/3x4c2-minmax-ssse3-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002676 "src/qs8-igemm/gen/3x4c8-minmax-ssse3-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002677 "src/qs8-igemm/gen/3x4c8-minmax-ssse3-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002678 "src/qs8-igemm/gen/4x4c2-minmax-ssse3-ld64.c",
2679 "src/qs8-igemm/gen/4x4c2-minmax-ssse3-ld128.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -07002680 "src/qs8-requantization/precise-ssse3.c",
2681 "src/qs8-requantization/q31-ssse3.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07002682 "src/qu8-requantization/precise-ssse3.c",
2683 "src/qu8-requantization/q31-ssse3.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002684]
2685
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08002686SSE41_UKERNELS = [
Marat Dukhan40a672f2019-11-25 03:08:22 -08002687 "src/f32-prelu/gen/sse41-2x4.c",
2688 "src/f32-prelu/gen/sse41-2x8.c",
Marat Dukhand243c1a2020-09-17 18:48:10 -07002689 "src/f32-sigmoid/gen/sse41-lut64-p2-div-x4.c",
2690 "src/f32-sigmoid/gen/sse41-lut64-p2-div-x8.c",
2691 "src/f32-sigmoid/gen/sse41-lut64-p2-div-x12.c",
2692 "src/f32-sigmoid/gen/sse41-lut64-p2-div-x16.c",
2693 "src/f32-sigmoid/gen/sse41-lut64-p2-div-x20.c",
2694 "src/f32-sigmoid/gen/sse41-lut64-p2-div-x24.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002695 "src/f32-sigmoid/gen/sse41-p5-div-x4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08002696 "src/f32-sigmoid/gen/sse41-p5-div-x8.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002697 "src/f32-sigmoid/gen/sse41-p5-div-x12.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08002698 "src/f32-sigmoid/gen/sse41-p5-div-x16.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002699 "src/f32-sigmoid/gen/sse41-p5-div-x20.c",
2700 "src/f32-sigmoid/gen/sse41-p5-div-x24.c",
Marat Dukhaned6baaf2020-12-01 15:07:08 -08002701 "src/f32-velu/gen/velu-sse41-rr2-lut16-p3-x4.c",
2702 "src/f32-velu/gen/velu-sse41-rr2-lut16-p3-x8.c",
2703 "src/f32-velu/gen/velu-sse41-rr2-lut16-p3-x12.c",
2704 "src/f32-velu/gen/velu-sse41-rr2-lut16-p3-x16.c",
2705 "src/f32-velu/gen/velu-sse41-rr2-lut16-p3-x20.c",
2706 "src/f32-velu/gen/velu-sse41-rr2-lut16-p3-x24.c",
2707 "src/f32-velu/gen/velu-sse41-rr2-p6-x4.c",
2708 "src/f32-velu/gen/velu-sse41-rr2-p6-x8.c",
2709 "src/f32-velu/gen/velu-sse41-rr2-p6-x12.c",
2710 "src/f32-velu/gen/velu-sse41-rr2-p6-x16.c",
2711 "src/f32-velu/gen/velu-sse41-rr2-p6-x20.c",
2712 "src/f32-velu/gen/velu-sse41-rr2-p6-x24.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07002713 "src/f32-vlrelu/gen/vlrelu-sse41-x4.c",
2714 "src/f32-vlrelu/gen/vlrelu-sse41-x8.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07002715 "src/f32-vrnd/gen/vrndd-sse41-x4.c",
2716 "src/f32-vrnd/gen/vrndd-sse41-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002717 "src/f32-vrnd/gen/vrndne-sse41-x4.c",
2718 "src/f32-vrnd/gen/vrndne-sse41-x8.c",
2719 "src/f32-vrnd/gen/vrndu-sse41-x4.c",
2720 "src/f32-vrnd/gen/vrndu-sse41-x8.c",
2721 "src/f32-vrnd/gen/vrndz-sse41-x4.c",
2722 "src/f32-vrnd/gen/vrndz-sse41-x8.c",
2723 "src/math/roundd-sse41.c",
2724 "src/math/roundne-sse41.c",
2725 "src/math/roundu-sse41.c",
2726 "src/math/roundz-sse41.c",
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07002727 "src/qs8-dwconv/gen/up8x9-minmax-sse41-mul16.c",
2728 "src/qs8-dwconv/gen/up16x9-minmax-sse41-mul16.c",
2729 "src/qs8-dwconv/gen/up24x9-minmax-sse41-mul16.c",
Marat Dukhan07feec82021-04-02 22:41:15 -07002730 "src/qs8-dwconv/gen/up8x9-minmax-sse41-mul32.c",
2731 "src/qs8-dwconv/gen/up16x9-minmax-sse41-mul32.c",
2732 "src/qs8-dwconv/gen/up24x9-minmax-sse41-mul32.c",
Marat Dukhan159688f2020-08-06 10:34:29 -07002733 "src/qs8-gavgpool/gen/7p7x-minmax-sse41-c8-acc2.c",
2734 "src/qs8-gavgpool/gen/7p7x-minmax-sse41-c16-acc2.c",
2735 "src/qs8-gavgpool/gen/7p7x-minmax-sse41-c24-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002736 "src/qs8-gavgpool/gen/7x-minmax-sse41-c8-acc2.c",
2737 "src/qs8-gavgpool/gen/7x-minmax-sse41-c16-acc2.c",
2738 "src/qs8-gavgpool/gen/7x-minmax-sse41-c24-acc2.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002739 "src/qs8-gemm/gen/1x4c2-minmax-sse41-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002740 "src/qs8-gemm/gen/1x4c2-minmax-sse41-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002741 "src/qs8-gemm/gen/1x4c2-xw-minmax-sse41.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002742 "src/qs8-gemm/gen/1x4c8-minmax-sse41-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002743 "src/qs8-gemm/gen/1x4c8-minmax-sse41-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002744 "src/qs8-gemm/gen/1x4c8-xw-minmax-sse41.c",
Marat Dukhan2e427872021-04-09 23:40:07 -07002745 "src/qs8-gemm/gen/2x4c2-minmax-sse41-ld64.c",
2746 "src/qs8-gemm/gen/2x4c2-minmax-sse41-ld128.c",
2747 "src/qs8-gemm/gen/2x4c2-xw-minmax-sse41.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002748 "src/qs8-gemm/gen/2x4c8-minmax-sse41-ld64.c",
2749 "src/qs8-gemm/gen/2x4c8-minmax-sse41-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002750 "src/qs8-gemm/gen/2x4c8-xw-minmax-sse41.c",
Marat Dukhan2e427872021-04-09 23:40:07 -07002751 "src/qs8-gemm/gen/3x4c2-minmax-sse41-ld64.c",
2752 "src/qs8-gemm/gen/3x4c2-minmax-sse41-ld128.c",
2753 "src/qs8-gemm/gen/3x4c2-xw-minmax-sse41.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002754 "src/qs8-gemm/gen/3x4c8-minmax-sse41-ld64.c",
2755 "src/qs8-gemm/gen/3x4c8-minmax-sse41-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002756 "src/qs8-gemm/gen/3x4c8-xw-minmax-sse41.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002757 "src/qs8-gemm/gen/4x4c2-minmax-sse41-ld64.c",
2758 "src/qs8-gemm/gen/4x4c2-minmax-sse41-ld128.c",
2759 "src/qs8-gemm/gen/4x4c2-xw-minmax-sse41.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002760 "src/qs8-igemm/gen/1x4c2-minmax-sse41-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002761 "src/qs8-igemm/gen/1x4c2-minmax-sse41-ld128.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002762 "src/qs8-igemm/gen/1x4c8-minmax-sse41-ld64.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002763 "src/qs8-igemm/gen/1x4c8-minmax-sse41-ld128.c",
Marat Dukhan2e427872021-04-09 23:40:07 -07002764 "src/qs8-igemm/gen/2x4c2-minmax-sse41-ld64.c",
2765 "src/qs8-igemm/gen/2x4c2-minmax-sse41-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002766 "src/qs8-igemm/gen/2x4c8-minmax-sse41-ld64.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002767 "src/qs8-igemm/gen/2x4c8-minmax-sse41-ld128.c",
Marat Dukhan2e427872021-04-09 23:40:07 -07002768 "src/qs8-igemm/gen/3x4c2-minmax-sse41-ld64.c",
2769 "src/qs8-igemm/gen/3x4c2-minmax-sse41-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002770 "src/qs8-igemm/gen/3x4c8-minmax-sse41-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002771 "src/qs8-igemm/gen/3x4c8-minmax-sse41-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002772 "src/qs8-igemm/gen/4x4c2-minmax-sse41-ld64.c",
2773 "src/qs8-igemm/gen/4x4c2-minmax-sse41-ld128.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -07002774 "src/qs8-requantization/fp32-sse4.c",
2775 "src/qs8-requantization/precise-sse4.c",
2776 "src/qs8-requantization/q31-sse4.c",
Marat Dukhand9f3ad42020-08-10 12:30:58 -07002777 "src/qs8-vadd/gen/minmax-sse41-mul16-ld64-x8.c",
2778 "src/qs8-vadd/gen/minmax-sse41-mul16-ld64-x16.c",
2779 "src/qs8-vadd/gen/minmax-sse41-mul16-ld64-x24.c",
2780 "src/qs8-vadd/gen/minmax-sse41-mul16-ld64-x32.c",
Marat Dukhanbb9225e2020-09-06 22:40:56 -07002781 "src/qs8-vadd/gen/minmax-sse41-mul32-ld32-x8.c",
2782 "src/qs8-vadd/gen/minmax-sse41-mul32-ld32-x16.c",
2783 "src/qs8-vadd/gen/minmax-sse41-mul32-ld32-x24.c",
2784 "src/qs8-vadd/gen/minmax-sse41-mul32-ld32-x32.c",
Marat Dukhan0270d9f2020-08-11 00:56:46 -07002785 "src/qs8-vaddc/gen/minmax-sse41-mul16-ld64-x8.c",
2786 "src/qs8-vaddc/gen/minmax-sse41-mul16-ld64-x16.c",
2787 "src/qs8-vaddc/gen/minmax-sse41-mul16-ld64-x24.c",
2788 "src/qs8-vaddc/gen/minmax-sse41-mul16-ld64-x32.c",
Marat Dukhanbb9225e2020-09-06 22:40:56 -07002789 "src/qs8-vaddc/gen/minmax-sse41-mul32-ld32-x8.c",
2790 "src/qs8-vaddc/gen/minmax-sse41-mul32-ld32-x16.c",
2791 "src/qs8-vaddc/gen/minmax-sse41-mul32-ld32-x24.c",
2792 "src/qs8-vaddc/gen/minmax-sse41-mul32-ld32-x32.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07002793 "src/qu8-requantization/precise-sse4.c",
2794 "src/qu8-requantization/q31-sse4.c",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08002795]
2796
Marat Dukhan08c4a432019-10-03 09:29:21 -07002797AVX_UKERNELS = [
Marat Dukhan5c5fa962020-03-10 18:38:33 -07002798 "src/f32-clamp/gen/avx-x8.c",
2799 "src/f32-clamp/gen/avx-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002800 "src/f32-dwconv/gen/up8x4-minmax-avx-acc2.c",
2801 "src/f32-dwconv/gen/up8x4-minmax-avx.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002802 "src/f32-dwconv/gen/up8x9-minmax-avx-acc2.c",
2803 "src/f32-dwconv/gen/up8x9-minmax-avx.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002804 "src/f32-dwconv/gen/up8x25-minmax-avx-acc2.c",
2805 "src/f32-dwconv/gen/up8x25-minmax-avx.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002806 "src/f32-dwconv/gen/up16x4-minmax-avx-acc2.c",
2807 "src/f32-dwconv/gen/up16x4-minmax-avx.c",
2808 "src/f32-dwconv/gen/up16x9-minmax-avx-acc2.c",
2809 "src/f32-dwconv/gen/up16x9-minmax-avx.c",
2810 "src/f32-dwconv/gen/up16x25-minmax-avx-acc2.c",
2811 "src/f32-dwconv/gen/up16x25-minmax-avx.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002812 "src/f32-gemm/gen-inc/1x8inc-minmax-avx-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002813 "src/f32-gemm/gen-inc/1x16inc-minmax-avx-broadcast.c",
2814 "src/f32-gemm/gen-inc/3x16inc-minmax-avx-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002815 "src/f32-gemm/gen-inc/4x8inc-minmax-avx-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002816 "src/f32-gemm/gen-inc/4x16inc-minmax-avx-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002817 "src/f32-gemm/gen-inc/5x8inc-minmax-avx-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002818 "src/f32-gemm/gen-inc/5x16inc-minmax-avx-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002819 "src/f32-gemm/gen-inc/6x8inc-minmax-avx-broadcast.c",
2820 "src/f32-gemm/gen-inc/7x8inc-minmax-avx-broadcast.c",
2821 "src/f32-gemm/gen/1x8-minmax-avx-broadcast.c",
2822 "src/f32-gemm/gen/1x16-minmax-avx-broadcast.c",
2823 "src/f32-gemm/gen/3x16-minmax-avx-broadcast.c",
2824 "src/f32-gemm/gen/4x8-minmax-avx-broadcast.c",
2825 "src/f32-gemm/gen/4x16-minmax-avx-broadcast.c",
2826 "src/f32-gemm/gen/5x8-minmax-avx-broadcast.c",
2827 "src/f32-gemm/gen/5x16-minmax-avx-broadcast.c",
2828 "src/f32-gemm/gen/6x8-minmax-avx-broadcast.c",
2829 "src/f32-gemm/gen/7x8-minmax-avx-broadcast.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07002830 "src/f32-hswish/gen/hswish-avx-x8.c",
2831 "src/f32-hswish/gen/hswish-avx-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002832 "src/f32-igemm/gen/1x8-minmax-avx-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002833 "src/f32-igemm/gen/1x16-minmax-avx-broadcast.c",
2834 "src/f32-igemm/gen/3x16-minmax-avx-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002835 "src/f32-igemm/gen/4x8-minmax-avx-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002836 "src/f32-igemm/gen/4x16-minmax-avx-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002837 "src/f32-igemm/gen/5x8-minmax-avx-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002838 "src/f32-igemm/gen/5x16-minmax-avx-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002839 "src/f32-igemm/gen/6x8-minmax-avx-broadcast.c",
2840 "src/f32-igemm/gen/7x8-minmax-avx-broadcast.c",
Marat Dukhan90eca0a2020-03-11 00:52:23 -07002841 "src/f32-prelu/gen/avx-2x8.c",
2842 "src/f32-prelu/gen/avx-2x16.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07002843 "src/f32-relu/gen/avx-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002844 "src/f32-relu/gen/avx-x16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002845 "src/f32-rmax/avx.c",
T.J. Alumbaughdc2b29c2020-10-14 13:56:08 -07002846 "src/f32-sigmoid/gen/avx-rr2-p5-div-x8.c",
2847 "src/f32-sigmoid/gen/avx-rr2-p5-div-x16.c",
2848 "src/f32-sigmoid/gen/avx-rr2-p5-div-x24.c",
2849 "src/f32-sigmoid/gen/avx-rr2-p5-div-x32.c",
2850 "src/f32-sigmoid/gen/avx-rr2-p5-div-x40.c",
2851 "src/f32-sigmoid/gen/avx-rr2-p5-div-x48.c",
2852 "src/f32-sigmoid/gen/avx-rr2-p5-div-x56.c",
2853 "src/f32-sigmoid/gen/avx-rr2-p5-div-x64.c",
2854 "src/f32-sigmoid/gen/avx-rr2-p5-div-x72.c",
2855 "src/f32-sigmoid/gen/avx-rr2-p5-div-x80.c",
2856 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x8.c",
2857 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x16.c",
2858 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x24.c",
2859 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x32.c",
2860 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x40.c",
2861 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x48.c",
2862 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x56.c",
2863 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x64.c",
2864 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x72.c",
2865 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x80.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002866 "src/f32-vbinary/gen/vadd-minmax-avx-x8.c",
2867 "src/f32-vbinary/gen/vadd-minmax-avx-x16.c",
2868 "src/f32-vbinary/gen/vaddc-minmax-avx-x8.c",
2869 "src/f32-vbinary/gen/vaddc-minmax-avx-x16.c",
2870 "src/f32-vbinary/gen/vdiv-minmax-avx-x8.c",
2871 "src/f32-vbinary/gen/vdiv-minmax-avx-x16.c",
2872 "src/f32-vbinary/gen/vdivc-minmax-avx-x8.c",
2873 "src/f32-vbinary/gen/vdivc-minmax-avx-x16.c",
Marat Dukhan9a88efe2019-12-10 15:54:24 -08002874 "src/f32-vbinary/gen/vmax-avx-x8.c",
2875 "src/f32-vbinary/gen/vmax-avx-x16.c",
2876 "src/f32-vbinary/gen/vmaxc-avx-x8.c",
2877 "src/f32-vbinary/gen/vmaxc-avx-x16.c",
2878 "src/f32-vbinary/gen/vmin-avx-x8.c",
2879 "src/f32-vbinary/gen/vmin-avx-x16.c",
2880 "src/f32-vbinary/gen/vminc-avx-x8.c",
2881 "src/f32-vbinary/gen/vminc-avx-x16.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002882 "src/f32-vbinary/gen/vmul-minmax-avx-x8.c",
2883 "src/f32-vbinary/gen/vmul-minmax-avx-x16.c",
2884 "src/f32-vbinary/gen/vmulc-minmax-avx-x8.c",
2885 "src/f32-vbinary/gen/vmulc-minmax-avx-x16.c",
2886 "src/f32-vbinary/gen/vrdivc-minmax-avx-x8.c",
2887 "src/f32-vbinary/gen/vrdivc-minmax-avx-x16.c",
2888 "src/f32-vbinary/gen/vrsubc-minmax-avx-x8.c",
2889 "src/f32-vbinary/gen/vrsubc-minmax-avx-x16.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07002890 "src/f32-vbinary/gen/vsqrdiff-avx-x8.c",
2891 "src/f32-vbinary/gen/vsqrdiff-avx-x16.c",
2892 "src/f32-vbinary/gen/vsqrdiffc-avx-x8.c",
2893 "src/f32-vbinary/gen/vsqrdiffc-avx-x16.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002894 "src/f32-vbinary/gen/vsub-minmax-avx-x8.c",
2895 "src/f32-vbinary/gen/vsub-minmax-avx-x16.c",
2896 "src/f32-vbinary/gen/vsubc-minmax-avx-x8.c",
2897 "src/f32-vbinary/gen/vsubc-minmax-avx-x16.c",
Marat Dukhaned6baaf2020-12-01 15:07:08 -08002898 "src/f32-velu/gen/velu-avx-rr2-lut4-p4-perm-x8.c",
2899 "src/f32-velu/gen/velu-avx-rr2-lut4-p4-perm-x16.c",
2900 "src/f32-velu/gen/velu-avx-rr2-lut4-p4-perm-x24.c",
2901 "src/f32-velu/gen/velu-avx-rr2-lut4-p4-perm-x32.c",
2902 "src/f32-velu/gen/velu-avx-rr2-lut4-p4-perm-x40.c",
2903 "src/f32-velu/gen/velu-avx-rr2-lut4-p4-perm-x48.c",
2904 "src/f32-velu/gen/velu-avx-rr2-lut16-p3-x8.c",
2905 "src/f32-velu/gen/velu-avx-rr2-lut16-p3-x16.c",
2906 "src/f32-velu/gen/velu-avx-rr2-lut16-p3-x24.c",
2907 "src/f32-velu/gen/velu-avx-rr2-lut16-p3-x32.c",
2908 "src/f32-velu/gen/velu-avx-rr2-lut16-p3-x40.c",
2909 "src/f32-velu/gen/velu-avx-rr2-lut16-p3-x48.c",
2910 "src/f32-velu/gen/velu-avx-rr2-p6-x8.c",
2911 "src/f32-velu/gen/velu-avx-rr2-p6-x16.c",
2912 "src/f32-velu/gen/velu-avx-rr2-p6-x24.c",
2913 "src/f32-velu/gen/velu-avx-rr2-p6-x32.c",
2914 "src/f32-velu/gen/velu-avx-rr2-p6-x40.c",
2915 "src/f32-velu/gen/velu-avx-rr2-p6-x48.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07002916 "src/f32-vlrelu/gen/vlrelu-avx-x8.c",
2917 "src/f32-vlrelu/gen/vlrelu-avx-x16.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07002918 "src/f32-vrnd/gen/vrndd-avx-x8.c",
2919 "src/f32-vrnd/gen/vrndd-avx-x16.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002920 "src/f32-vrnd/gen/vrndne-avx-x8.c",
2921 "src/f32-vrnd/gen/vrndne-avx-x16.c",
2922 "src/f32-vrnd/gen/vrndu-avx-x8.c",
2923 "src/f32-vrnd/gen/vrndu-avx-x16.c",
2924 "src/f32-vrnd/gen/vrndz-avx-x8.c",
2925 "src/f32-vrnd/gen/vrndz-avx-x16.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07002926 "src/f32-vscale/avx-x32.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07002927 "src/f32-vsqrt/gen/avx-sqrt-x8.c",
2928 "src/f32-vsqrt/gen/avx-sqrt-x16.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07002929 "src/f32-vunary/gen/vabs-avx-x8.c",
2930 "src/f32-vunary/gen/vabs-avx-x16.c",
2931 "src/f32-vunary/gen/vneg-avx-x8.c",
2932 "src/f32-vunary/gen/vneg-avx-x16.c",
2933 "src/f32-vunary/gen/vsqr-avx-x8.c",
2934 "src/f32-vunary/gen/vsqr-avx-x16.c",
Marat Dukhanfa0ab852021-04-02 17:30:49 -07002935 "src/qs8-dwconv/gen/up8x9-minmax-avx-mul16.c",
2936 "src/qs8-dwconv/gen/up16x9-minmax-avx-mul16.c",
2937 "src/qs8-dwconv/gen/up24x9-minmax-avx-mul16.c",
Marat Dukhan07feec82021-04-02 22:41:15 -07002938 "src/qs8-dwconv/gen/up8x9-minmax-avx-mul32.c",
2939 "src/qs8-dwconv/gen/up16x9-minmax-avx-mul32.c",
2940 "src/qs8-dwconv/gen/up24x9-minmax-avx-mul32.c",
Marat Dukhana3c16332021-04-02 15:03:27 -07002941 "src/qs8-gemm/gen/1x4c2-minmax-avx-ld64.c",
2942 "src/qs8-gemm/gen/1x4c2-minmax-avx-ld128.c",
2943 "src/qs8-gemm/gen/1x4c2-xw-minmax-avx.c",
2944 "src/qs8-gemm/gen/1x4c8-minmax-avx-ld64.c",
2945 "src/qs8-gemm/gen/1x4c8-minmax-avx-ld128.c",
2946 "src/qs8-gemm/gen/1x4c8-xw-minmax-avx.c",
Marat Dukhan2e427872021-04-09 23:40:07 -07002947 "src/qs8-gemm/gen/2x4c2-minmax-avx-ld64.c",
2948 "src/qs8-gemm/gen/2x4c2-minmax-avx-ld128.c",
2949 "src/qs8-gemm/gen/2x4c2-xw-minmax-avx.c",
Marat Dukhana3c16332021-04-02 15:03:27 -07002950 "src/qs8-gemm/gen/2x4c8-minmax-avx-ld64.c",
2951 "src/qs8-gemm/gen/2x4c8-minmax-avx-ld128.c",
2952 "src/qs8-gemm/gen/2x4c8-xw-minmax-avx.c",
Marat Dukhan2e427872021-04-09 23:40:07 -07002953 "src/qs8-gemm/gen/3x4c2-minmax-avx-ld64.c",
2954 "src/qs8-gemm/gen/3x4c2-minmax-avx-ld128.c",
2955 "src/qs8-gemm/gen/3x4c2-xw-minmax-avx.c",
Marat Dukhana3c16332021-04-02 15:03:27 -07002956 "src/qs8-gemm/gen/3x4c8-minmax-avx-ld64.c",
2957 "src/qs8-gemm/gen/3x4c8-minmax-avx-ld128.c",
2958 "src/qs8-gemm/gen/3x4c8-xw-minmax-avx.c",
2959 "src/qs8-gemm/gen/4x4c2-minmax-avx-ld64.c",
2960 "src/qs8-gemm/gen/4x4c2-minmax-avx-ld128.c",
2961 "src/qs8-gemm/gen/4x4c2-xw-minmax-avx.c",
2962 "src/qs8-igemm/gen/1x4c2-minmax-avx-ld64.c",
2963 "src/qs8-igemm/gen/1x4c2-minmax-avx-ld128.c",
2964 "src/qs8-igemm/gen/1x4c8-minmax-avx-ld64.c",
2965 "src/qs8-igemm/gen/1x4c8-minmax-avx-ld128.c",
Marat Dukhan2e427872021-04-09 23:40:07 -07002966 "src/qs8-igemm/gen/2x4c2-minmax-avx-ld64.c",
2967 "src/qs8-igemm/gen/2x4c2-minmax-avx-ld128.c",
Marat Dukhana3c16332021-04-02 15:03:27 -07002968 "src/qs8-igemm/gen/2x4c8-minmax-avx-ld64.c",
2969 "src/qs8-igemm/gen/2x4c8-minmax-avx-ld128.c",
Marat Dukhan2e427872021-04-09 23:40:07 -07002970 "src/qs8-igemm/gen/3x4c2-minmax-avx-ld64.c",
2971 "src/qs8-igemm/gen/3x4c2-minmax-avx-ld128.c",
Marat Dukhana3c16332021-04-02 15:03:27 -07002972 "src/qs8-igemm/gen/3x4c8-minmax-avx-ld64.c",
2973 "src/qs8-igemm/gen/3x4c8-minmax-avx-ld128.c",
2974 "src/qs8-igemm/gen/4x4c2-minmax-avx-ld64.c",
2975 "src/qs8-igemm/gen/4x4c2-minmax-avx-ld128.c",
Marat Dukhane9c4b962021-04-02 16:56:55 -07002976 "src/qs8-vadd/gen/minmax-avx-mul16-ld64-x8.c",
2977 "src/qs8-vadd/gen/minmax-avx-mul16-ld64-x16.c",
2978 "src/qs8-vadd/gen/minmax-avx-mul16-ld64-x24.c",
2979 "src/qs8-vadd/gen/minmax-avx-mul16-ld64-x32.c",
2980 "src/qs8-vadd/gen/minmax-avx-mul32-ld32-x8.c",
2981 "src/qs8-vadd/gen/minmax-avx-mul32-ld32-x16.c",
2982 "src/qs8-vadd/gen/minmax-avx-mul32-ld32-x24.c",
2983 "src/qs8-vadd/gen/minmax-avx-mul32-ld32-x32.c",
2984 "src/qs8-vaddc/gen/minmax-avx-mul16-ld64-x8.c",
2985 "src/qs8-vaddc/gen/minmax-avx-mul16-ld64-x16.c",
2986 "src/qs8-vaddc/gen/minmax-avx-mul16-ld64-x24.c",
2987 "src/qs8-vaddc/gen/minmax-avx-mul16-ld64-x32.c",
2988 "src/qs8-vaddc/gen/minmax-avx-mul32-ld32-x8.c",
2989 "src/qs8-vaddc/gen/minmax-avx-mul32-ld32-x16.c",
2990 "src/qs8-vaddc/gen/minmax-avx-mul32-ld32-x24.c",
2991 "src/qs8-vaddc/gen/minmax-avx-mul32-ld32-x32.c",
Marat Dukhanb7633f22020-11-20 16:34:56 -08002992 "src/math/exp-avx-rr2-p5.c",
Marat Dukhande390d42020-11-29 19:32:18 -08002993 "src/math/expm1minus-avx-rr2-lut4-p4-perm.c",
2994 "src/math/expm1minus-avx-rr2-lut16-p3.c",
2995 "src/math/expm1minus-avx-rr2-p6.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002996 "src/math/sigmoid-avx-rr2-lut64-p2-div.c",
Marat Dukhan2e4727b2020-09-24 00:48:39 -07002997 "src/math/sigmoid-avx-rr2-p5-div.c",
Marat Dukhan39ab16f2020-10-16 14:30:35 -07002998 "src/math/sigmoid-avx-rr2-p5-nr1.c",
2999 "src/math/sigmoid-avx-rr2-p5-nr2.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003000]
3001
Marat Dukhan1566fee2020-08-02 21:55:41 -07003002XOP_UKERNELS = [
3003 "src/qs8-gemm/gen/1x4c2-minmax-xop-ld64.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07003004 "src/qs8-gemm/gen/1x4c2-minmax-xop-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07003005 "src/qs8-gemm/gen/1x4c2-xw-minmax-xop.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07003006 "src/qs8-gemm/gen/1x4c8-minmax-xop-ld64.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07003007 "src/qs8-gemm/gen/1x4c8-minmax-xop-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07003008 "src/qs8-gemm/gen/1x4c8-xw-minmax-xop.c",
Marat Dukhan2e427872021-04-09 23:40:07 -07003009 "src/qs8-gemm/gen/2x4c2-minmax-xop-ld64.c",
3010 "src/qs8-gemm/gen/2x4c2-minmax-xop-ld128.c",
3011 "src/qs8-gemm/gen/2x4c2-xw-minmax-xop.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003012 "src/qs8-gemm/gen/2x4c8-minmax-xop-ld64.c",
3013 "src/qs8-gemm/gen/2x4c8-minmax-xop-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07003014 "src/qs8-gemm/gen/2x4c8-xw-minmax-xop.c",
Marat Dukhan2e427872021-04-09 23:40:07 -07003015 "src/qs8-gemm/gen/3x4c2-minmax-xop-ld64.c",
3016 "src/qs8-gemm/gen/3x4c2-minmax-xop-ld128.c",
3017 "src/qs8-gemm/gen/3x4c2-xw-minmax-xop.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003018 "src/qs8-gemm/gen/3x4c8-minmax-xop-ld64.c",
3019 "src/qs8-gemm/gen/3x4c8-minmax-xop-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07003020 "src/qs8-gemm/gen/3x4c8-xw-minmax-xop.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003021 "src/qs8-gemm/gen/4x4c2-minmax-xop-ld64.c",
3022 "src/qs8-gemm/gen/4x4c2-minmax-xop-ld128.c",
3023 "src/qs8-gemm/gen/4x4c2-xw-minmax-xop.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07003024 "src/qs8-igemm/gen/1x4c2-minmax-xop-ld64.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07003025 "src/qs8-igemm/gen/1x4c2-minmax-xop-ld128.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07003026 "src/qs8-igemm/gen/1x4c8-minmax-xop-ld64.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07003027 "src/qs8-igemm/gen/1x4c8-minmax-xop-ld128.c",
Marat Dukhan2e427872021-04-09 23:40:07 -07003028 "src/qs8-igemm/gen/2x4c2-minmax-xop-ld64.c",
3029 "src/qs8-igemm/gen/2x4c2-minmax-xop-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003030 "src/qs8-igemm/gen/2x4c8-minmax-xop-ld64.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07003031 "src/qs8-igemm/gen/2x4c8-minmax-xop-ld128.c",
Marat Dukhan2e427872021-04-09 23:40:07 -07003032 "src/qs8-igemm/gen/3x4c2-minmax-xop-ld64.c",
3033 "src/qs8-igemm/gen/3x4c2-minmax-xop-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003034 "src/qs8-igemm/gen/3x4c8-minmax-xop-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07003035 "src/qs8-igemm/gen/3x4c8-minmax-xop-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003036 "src/qs8-igemm/gen/4x4c2-minmax-xop-ld64.c",
3037 "src/qs8-igemm/gen/4x4c2-minmax-xop-ld128.c",
Marat Dukhanbb9225e2020-09-06 22:40:56 -07003038 "src/qs8-vadd/gen/minmax-xop-mul32-ld32-x8.c",
3039 "src/qs8-vadd/gen/minmax-xop-mul32-ld32-x16.c",
3040 "src/qs8-vadd/gen/minmax-xop-mul32-ld32-x24.c",
3041 "src/qs8-vadd/gen/minmax-xop-mul32-ld32-x32.c",
3042 "src/qs8-vaddc/gen/minmax-xop-mul32-ld32-x8.c",
3043 "src/qs8-vaddc/gen/minmax-xop-mul32-ld32-x16.c",
3044 "src/qs8-vaddc/gen/minmax-xop-mul32-ld32-x24.c",
3045 "src/qs8-vaddc/gen/minmax-xop-mul32-ld32-x32.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07003046]
3047
Marat Dukhanfda12b82019-11-21 12:27:59 -08003048FMA3_UKERNELS = [
Marat Dukhan1c587112020-04-08 20:04:28 -07003049 "src/f32-dwconv/gen/up8x4-minmax-fma3-acc2.c",
3050 "src/f32-dwconv/gen/up8x4-minmax-fma3.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07003051 "src/f32-dwconv/gen/up8x9-minmax-fma3-acc2.c",
3052 "src/f32-dwconv/gen/up8x9-minmax-fma3.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07003053 "src/f32-dwconv/gen/up8x25-minmax-fma3-acc2.c",
3054 "src/f32-dwconv/gen/up8x25-minmax-fma3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003055 "src/f32-dwconv/gen/up16x4-minmax-fma3-acc2.c",
3056 "src/f32-dwconv/gen/up16x4-minmax-fma3.c",
3057 "src/f32-dwconv/gen/up16x9-minmax-fma3-acc2.c",
3058 "src/f32-dwconv/gen/up16x9-minmax-fma3.c",
3059 "src/f32-dwconv/gen/up16x25-minmax-fma3-acc2.c",
3060 "src/f32-dwconv/gen/up16x25-minmax-fma3.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07003061 "src/f32-gemm/gen-inc/1x8inc-minmax-fma3-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003062 "src/f32-gemm/gen-inc/1x16inc-minmax-fma3-broadcast.c",
3063 "src/f32-gemm/gen-inc/1x16s4inc-minmax-fma3-broadcast.c",
3064 "src/f32-gemm/gen-inc/3x16inc-minmax-fma3-broadcast.c",
3065 "src/f32-gemm/gen-inc/3x16s4inc-minmax-fma3-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07003066 "src/f32-gemm/gen-inc/4x8inc-minmax-fma3-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003067 "src/f32-gemm/gen-inc/4x16inc-minmax-fma3-broadcast.c",
3068 "src/f32-gemm/gen-inc/4x16s4inc-minmax-fma3-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07003069 "src/f32-gemm/gen-inc/5x8inc-minmax-fma3-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003070 "src/f32-gemm/gen-inc/5x16inc-minmax-fma3-broadcast.c",
3071 "src/f32-gemm/gen-inc/5x16s4inc-minmax-fma3-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07003072 "src/f32-gemm/gen-inc/6x8inc-minmax-fma3-broadcast.c",
3073 "src/f32-gemm/gen-inc/7x8inc-minmax-fma3-broadcast.c",
3074 "src/f32-gemm/gen-inc/8x8inc-minmax-fma3-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003075 "src/f32-gemm/gen/1x8-minmax-fma3-broadcast.c",
3076 "src/f32-gemm/gen/1x16-minmax-fma3-broadcast.c",
3077 "src/f32-gemm/gen/1x16s4-minmax-fma3-broadcast.c",
3078 "src/f32-gemm/gen/3x16-minmax-fma3-broadcast.c",
3079 "src/f32-gemm/gen/3x16s4-minmax-fma3-broadcast.c",
3080 "src/f32-gemm/gen/4x8-minmax-fma3-broadcast.c",
3081 "src/f32-gemm/gen/4x16-minmax-fma3-broadcast.c",
3082 "src/f32-gemm/gen/4x16s4-minmax-fma3-broadcast.c",
3083 "src/f32-gemm/gen/5x8-minmax-fma3-broadcast.c",
3084 "src/f32-gemm/gen/5x16-minmax-fma3-broadcast.c",
3085 "src/f32-gemm/gen/5x16s4-minmax-fma3-broadcast.c",
3086 "src/f32-gemm/gen/6x8-minmax-fma3-broadcast.c",
3087 "src/f32-gemm/gen/7x8-minmax-fma3-broadcast.c",
3088 "src/f32-gemm/gen/8x8-minmax-fma3-broadcast.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07003089 "src/f32-hswish/gen/hswish-fma3-x8.c",
3090 "src/f32-hswish/gen/hswish-fma3-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07003091 "src/f32-igemm/gen/1x8-minmax-fma3-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003092 "src/f32-igemm/gen/1x16-minmax-fma3-broadcast.c",
3093 "src/f32-igemm/gen/1x16s4-minmax-fma3-broadcast.c",
3094 "src/f32-igemm/gen/3x16-minmax-fma3-broadcast.c",
3095 "src/f32-igemm/gen/3x16s4-minmax-fma3-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07003096 "src/f32-igemm/gen/4x8-minmax-fma3-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003097 "src/f32-igemm/gen/4x16-minmax-fma3-broadcast.c",
3098 "src/f32-igemm/gen/4x16s4-minmax-fma3-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07003099 "src/f32-igemm/gen/5x8-minmax-fma3-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003100 "src/f32-igemm/gen/5x16-minmax-fma3-broadcast.c",
3101 "src/f32-igemm/gen/5x16s4-minmax-fma3-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07003102 "src/f32-igemm/gen/6x8-minmax-fma3-broadcast.c",
3103 "src/f32-igemm/gen/7x8-minmax-fma3-broadcast.c",
3104 "src/f32-igemm/gen/8x8-minmax-fma3-broadcast.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07003105 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x8.c",
3106 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x16.c",
3107 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x24.c",
3108 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x32.c",
3109 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x40.c",
3110 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x48.c",
3111 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x56.c",
3112 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x64.c",
Marat Dukhan84000762020-06-29 18:38:43 -07003113 "src/math/sqrt-fma3-nr1fma.c",
Marat Dukhan84000762020-06-29 18:38:43 -07003114 "src/math/sqrt-fma3-nr1fma1adj.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003115 "src/math/sqrt-fma3-nr2fma.c",
Marat Dukhanfda12b82019-11-21 12:27:59 -08003116]
3117
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003118AVX2_UKERNELS = [
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003119 "src/f32-raddexpminusmax/gen/avx2-p5-x64-acc2.c",
3120 "src/f32-raddexpminusmax/gen/avx2-p5-x64-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003121 "src/f32-raddexpminusmax/gen/avx2-p5-x64.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003122 "src/f32-raddexpminusmax/gen/avx2-p5-x72-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003123 "src/f32-raddexpminusmax/gen/avx2-p5-x72.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003124 "src/f32-raddexpminusmax/gen/avx2-p5-x80-acc2.c",
3125 "src/f32-raddexpminusmax/gen/avx2-p5-x80-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003126 "src/f32-raddexpminusmax/gen/avx2-p5-x80.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003127 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc2.c",
3128 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc3.c",
3129 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc6.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003130 "src/f32-raddexpminusmax/gen/avx2-p5-x96.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003131 "src/f32-raddextexp/gen/avx2-p5-x64-acc2.c",
3132 "src/f32-raddextexp/gen/avx2-p5-x64-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003133 "src/f32-raddextexp/gen/avx2-p5-x64.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003134 "src/f32-raddextexp/gen/avx2-p5-x72-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003135 "src/f32-raddextexp/gen/avx2-p5-x72.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003136 "src/f32-raddextexp/gen/avx2-p5-x80-acc2.c",
3137 "src/f32-raddextexp/gen/avx2-p5-x80-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003138 "src/f32-raddextexp/gen/avx2-p5-x80.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003139 "src/f32-raddextexp/gen/avx2-p5-x96-acc2.c",
3140 "src/f32-raddextexp/gen/avx2-p5-x96-acc3.c",
3141 "src/f32-raddextexp/gen/avx2-p5-x96-acc6.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003142 "src/f32-raddextexp/gen/avx2-p5-x96.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003143 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64-acc2.c",
3144 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003145 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003146 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x72-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003147 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x72.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003148 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80-acc2.c",
3149 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003150 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003151 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc2.c",
3152 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc3.c",
3153 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc6.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003154 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96.c",
Marat Dukhanfa0a4322020-01-06 16:14:29 -08003155 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x8.c",
3156 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x16.c",
3157 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x24.c",
3158 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x32.c",
3159 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x40.c",
3160 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x48.c",
3161 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x56.c",
3162 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x64.c",
3163 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x72.c",
3164 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x80.c",
3165 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x8.c",
3166 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x16.c",
3167 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x24.c",
3168 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x32.c",
3169 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x40.c",
3170 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x48.c",
3171 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x56.c",
3172 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x64.c",
3173 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x72.c",
3174 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x80.c",
3175 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x8.c",
3176 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x16.c",
3177 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x24.c",
3178 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x32.c",
3179 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x40.c",
3180 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x48.c",
3181 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x56.c",
3182 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x64.c",
3183 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x72.c",
3184 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x80.c",
Marat Dukhaned6baaf2020-12-01 15:07:08 -08003185 "src/f32-velu/gen/velu-avx2-rr1-lut4-p4-perm-x8.c",
3186 "src/f32-velu/gen/velu-avx2-rr1-lut4-p4-perm-x16.c",
3187 "src/f32-velu/gen/velu-avx2-rr1-lut4-p4-perm-x24.c",
3188 "src/f32-velu/gen/velu-avx2-rr1-lut4-p4-perm-x32.c",
3189 "src/f32-velu/gen/velu-avx2-rr1-lut4-p4-perm-x40.c",
3190 "src/f32-velu/gen/velu-avx2-rr1-lut4-p4-perm-x48.c",
3191 "src/f32-velu/gen/velu-avx2-rr1-lut4-p4-perm-x56.c",
3192 "src/f32-velu/gen/velu-avx2-rr1-lut4-p4-perm-x64.c",
3193 "src/f32-velu/gen/velu-avx2-rr1-lut4-p4-perm-x72.c",
3194 "src/f32-velu/gen/velu-avx2-rr1-lut4-p4-perm-x80.c",
3195 "src/f32-velu/gen/velu-avx2-rr1-lut8-p4-perm-x8.c",
3196 "src/f32-velu/gen/velu-avx2-rr1-lut8-p4-perm-x16.c",
3197 "src/f32-velu/gen/velu-avx2-rr1-lut8-p4-perm-x24.c",
3198 "src/f32-velu/gen/velu-avx2-rr1-lut8-p4-perm-x32.c",
3199 "src/f32-velu/gen/velu-avx2-rr1-lut8-p4-perm-x40.c",
3200 "src/f32-velu/gen/velu-avx2-rr1-lut8-p4-perm-x48.c",
3201 "src/f32-velu/gen/velu-avx2-rr1-lut8-p4-perm-x56.c",
3202 "src/f32-velu/gen/velu-avx2-rr1-lut8-p4-perm-x64.c",
3203 "src/f32-velu/gen/velu-avx2-rr1-lut8-p4-perm-x72.c",
3204 "src/f32-velu/gen/velu-avx2-rr1-lut8-p4-perm-x80.c",
3205 "src/f32-velu/gen/velu-avx2-rr1-lut16-p3-gather-x8.c",
3206 "src/f32-velu/gen/velu-avx2-rr1-lut16-p3-gather-x16.c",
3207 "src/f32-velu/gen/velu-avx2-rr1-lut16-p3-gather-x24.c",
3208 "src/f32-velu/gen/velu-avx2-rr1-lut16-p3-gather-x32.c",
3209 "src/f32-velu/gen/velu-avx2-rr1-lut16-p3-gather-x40.c",
3210 "src/f32-velu/gen/velu-avx2-rr1-lut16-p3-gather-x48.c",
3211 "src/f32-velu/gen/velu-avx2-rr1-lut16-p3-gather-x56.c",
3212 "src/f32-velu/gen/velu-avx2-rr1-lut16-p3-gather-x64.c",
3213 "src/f32-velu/gen/velu-avx2-rr1-lut16-p3-gather-x72.c",
3214 "src/f32-velu/gen/velu-avx2-rr1-lut16-p3-gather-x80.c",
3215 "src/f32-velu/gen/velu-avx2-rr1-p6-x8.c",
3216 "src/f32-velu/gen/velu-avx2-rr1-p6-x16.c",
3217 "src/f32-velu/gen/velu-avx2-rr1-p6-x24.c",
3218 "src/f32-velu/gen/velu-avx2-rr1-p6-x32.c",
3219 "src/f32-velu/gen/velu-avx2-rr1-p6-x40.c",
3220 "src/f32-velu/gen/velu-avx2-rr1-p6-x48.c",
3221 "src/f32-velu/gen/velu-avx2-rr1-p6-x56.c",
3222 "src/f32-velu/gen/velu-avx2-rr1-p6-x64.c",
3223 "src/f32-velu/gen/velu-avx2-rr1-p6-x72.c",
3224 "src/f32-velu/gen/velu-avx2-rr1-p6-x80.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003225 "src/f32-vscaleexpminusmax/gen/avx2-p5-x8.c",
3226 "src/f32-vscaleexpminusmax/gen/avx2-p5-x16.c",
3227 "src/f32-vscaleexpminusmax/gen/avx2-p5-x24.c",
3228 "src/f32-vscaleexpminusmax/gen/avx2-p5-x32.c",
3229 "src/f32-vscaleexpminusmax/gen/avx2-p5-x40.c",
3230 "src/f32-vscaleexpminusmax/gen/avx2-p5-x48.c",
3231 "src/f32-vscaleexpminusmax/gen/avx2-p5-x56.c",
3232 "src/f32-vscaleexpminusmax/gen/avx2-p5-x64.c",
3233 "src/f32-vscaleexpminusmax/gen/avx2-p5-x72.c",
3234 "src/f32-vscaleexpminusmax/gen/avx2-p5-x80.c",
3235 "src/f32-vscaleexpminusmax/gen/avx2-p5-x88.c",
3236 "src/f32-vscaleexpminusmax/gen/avx2-p5-x96.c",
3237 "src/f32-vscaleextexp/gen/avx2-p5-x8.c",
3238 "src/f32-vscaleextexp/gen/avx2-p5-x16.c",
3239 "src/f32-vscaleextexp/gen/avx2-p5-x24.c",
3240 "src/f32-vscaleextexp/gen/avx2-p5-x32.c",
3241 "src/f32-vscaleextexp/gen/avx2-p5-x40.c",
3242 "src/f32-vscaleextexp/gen/avx2-p5-x48.c",
3243 "src/f32-vscaleextexp/gen/avx2-p5-x56.c",
3244 "src/f32-vscaleextexp/gen/avx2-p5-x64.c",
3245 "src/f32-vscaleextexp/gen/avx2-p5-x72.c",
3246 "src/f32-vscaleextexp/gen/avx2-p5-x80.c",
3247 "src/f32-vscaleextexp/gen/avx2-p5-x88.c",
3248 "src/f32-vscaleextexp/gen/avx2-p5-x96.c",
Marat Dukhanb7633f22020-11-20 16:34:56 -08003249 "src/math/exp-avx2-rr2-lut8-p3-perm.c",
3250 "src/math/exp-avx2-rr2-lut8-p4-perm.c",
3251 "src/math/exp-avx2-rr2-p5.c",
Marat Dukhande390d42020-11-29 19:32:18 -08003252 "src/math/expm1minus-avx2-rr1-lut4-p4-perm.c",
3253 "src/math/expm1minus-avx2-rr1-lut8-p4-perm.c",
3254 "src/math/expm1minus-avx2-rr1-lut16-p3-gather.c",
3255 "src/math/expm1minus-avx2-rr1-p6.c",
Frank Barcharde7223ee2020-12-04 19:04:01 -08003256 "src/math/expminus-avx2-rr2-p5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003257 "src/math/extexp-avx2-p5.c",
3258 "src/math/sigmoid-avx2-rr1-lut64-p2-gather-div.c",
3259 "src/math/sigmoid-avx2-rr1-lut64-p2-gather-nr1fma.c",
3260 "src/math/sigmoid-avx2-rr1-lut64-p2-gather-nr2fma.c",
3261 "src/math/sigmoid-avx2-rr1-lut64-p2-gather-nr2fma1adj.c",
3262 "src/math/sigmoid-avx2-rr1-p5-div.c",
3263 "src/math/sigmoid-avx2-rr1-p5-nr1fma.c",
3264 "src/math/sigmoid-avx2-rr1-p5-nr2fma.c",
3265 "src/math/sigmoid-avx2-rr2-lut64-p2-gather-div.c",
3266 "src/math/sigmoid-avx2-rr2-lut64-p2-gather-nr1fma.c",
3267 "src/math/sigmoid-avx2-rr2-lut64-p2-gather-nr2fma.c",
3268 "src/math/sigmoid-avx2-rr2-lut64-p2-gather-nr2fma1adj.c",
3269 "src/math/sigmoid-avx2-rr2-p5-div.c",
3270 "src/math/sigmoid-avx2-rr2-p5-nr1fma.c",
3271 "src/math/sigmoid-avx2-rr2-p5-nr2fma.c",
Marat Dukhand65a1522020-08-04 19:28:18 -07003272 "src/qs8-dwconv/gen/up8x9-minmax-avx2-mul32.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003273 "src/qs8-dwconv/gen/up16x9-minmax-avx2-mul16.c",
Marat Dukhand65a1522020-08-04 19:28:18 -07003274 "src/qs8-dwconv/gen/up16x9-minmax-avx2-mul32.c",
3275 "src/qs8-dwconv/gen/up24x9-minmax-avx2-mul32.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003276 "src/qs8-dwconv/gen/up32x9-minmax-avx2-mul16.c",
Marat Dukhand65a1522020-08-04 19:28:18 -07003277 "src/qs8-dwconv/gen/up32x9-minmax-avx2-mul32.c",
Marat Dukhan12809522020-08-02 22:23:51 -07003278 "src/qs8-gemm/gen/1x8c8-minmax-avx2.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07003279 "src/qs8-gemm/gen/1x8c8-xw-minmax-avx2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003280 "src/qs8-gemm/gen/2x8c8-minmax-avx2.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07003281 "src/qs8-gemm/gen/2x8c8-xw-minmax-avx2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003282 "src/qs8-gemm/gen/3x8c8-minmax-avx2.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07003283 "src/qs8-gemm/gen/3x8c8-xw-minmax-avx2.c",
Marat Dukhan12809522020-08-02 22:23:51 -07003284 "src/qs8-igemm/gen/1x8c8-minmax-avx2.c",
3285 "src/qs8-igemm/gen/2x8c8-minmax-avx2.c",
3286 "src/qs8-igemm/gen/3x8c8-minmax-avx2.c",
Marat Dukhane6dc0b62020-09-08 23:57:14 -07003287 "src/qs8-vadd/gen/minmax-avx2-mul32-ld64-x8.c",
3288 "src/qs8-vadd/gen/minmax-avx2-mul32-ld64-x16.c",
3289 "src/qs8-vadd/gen/minmax-avx2-mul32-ld64-x24.c",
3290 "src/qs8-vadd/gen/minmax-avx2-mul32-ld64-x32.c",
3291 "src/qs8-vaddc/gen/minmax-avx2-mul32-ld64-x8.c",
3292 "src/qs8-vaddc/gen/minmax-avx2-mul32-ld64-x16.c",
3293 "src/qs8-vaddc/gen/minmax-avx2-mul32-ld64-x24.c",
3294 "src/qs8-vaddc/gen/minmax-avx2-mul32-ld64-x32.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003295]
3296
Marat Dukhan08c4a432019-10-03 09:29:21 -07003297AVX512F_UKERNELS = [
Marat Dukhan5c5fa962020-03-10 18:38:33 -07003298 "src/f32-clamp/gen/avx512f-x16.c",
3299 "src/f32-clamp/gen/avx512f-x32.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07003300 "src/f32-dwconv/gen/up16x4-minmax-avx512f-acc2.c",
3301 "src/f32-dwconv/gen/up16x4-minmax-avx512f.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07003302 "src/f32-dwconv/gen/up16x9-minmax-avx512f-acc2.c",
3303 "src/f32-dwconv/gen/up16x9-minmax-avx512f.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07003304 "src/f32-dwconv/gen/up16x25-minmax-avx512f-acc2.c",
3305 "src/f32-dwconv/gen/up16x25-minmax-avx512f.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003306 "src/f32-dwconv/gen/up32x4-minmax-avx512f-acc2.c",
3307 "src/f32-dwconv/gen/up32x4-minmax-avx512f.c",
3308 "src/f32-dwconv/gen/up32x9-minmax-avx512f-acc2.c",
3309 "src/f32-dwconv/gen/up32x9-minmax-avx512f.c",
3310 "src/f32-dwconv/gen/up32x25-minmax-avx512f-acc2.c",
3311 "src/f32-dwconv/gen/up32x25-minmax-avx512f.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07003312 "src/f32-gemm/gen-inc/1x16inc-minmax-avx512f-broadcast.c",
3313 "src/f32-gemm/gen-inc/4x16inc-minmax-avx512f-broadcast.c",
3314 "src/f32-gemm/gen-inc/5x16inc-minmax-avx512f-broadcast.c",
3315 "src/f32-gemm/gen-inc/6x16inc-minmax-avx512f-broadcast.c",
3316 "src/f32-gemm/gen-inc/7x16inc-minmax-avx512f-broadcast.c",
3317 "src/f32-gemm/gen-inc/8x16inc-minmax-avx512f-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003318 "src/f32-gemm/gen/1x16-minmax-avx512f-broadcast.c",
3319 "src/f32-gemm/gen/4x16-minmax-avx512f-broadcast.c",
3320 "src/f32-gemm/gen/5x16-minmax-avx512f-broadcast.c",
3321 "src/f32-gemm/gen/6x16-minmax-avx512f-broadcast.c",
3322 "src/f32-gemm/gen/7x16-minmax-avx512f-broadcast.c",
3323 "src/f32-gemm/gen/8x16-minmax-avx512f-broadcast.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07003324 "src/f32-hswish/gen/hswish-avx512f-x16.c",
3325 "src/f32-hswish/gen/hswish-avx512f-x32.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07003326 "src/f32-igemm/gen/1x16-minmax-avx512f-broadcast.c",
3327 "src/f32-igemm/gen/4x16-minmax-avx512f-broadcast.c",
3328 "src/f32-igemm/gen/5x16-minmax-avx512f-broadcast.c",
3329 "src/f32-igemm/gen/6x16-minmax-avx512f-broadcast.c",
3330 "src/f32-igemm/gen/7x16-minmax-avx512f-broadcast.c",
3331 "src/f32-igemm/gen/8x16-minmax-avx512f-broadcast.c",
Marat Dukhan90eca0a2020-03-11 00:52:23 -07003332 "src/f32-prelu/gen/avx512f-2x16.c",
3333 "src/f32-prelu/gen/avx512f-2x32.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003334 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128-acc2.c",
3335 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003336 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003337 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x144-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003338 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x144.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003339 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160-acc2.c",
3340 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003341 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003342 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc2.c",
3343 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc3.c",
3344 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc6.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003345 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003346 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128-acc2.c",
3347 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003348 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003349 "src/f32-raddextexp/gen/avx512f-p5-scalef-x144-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003350 "src/f32-raddextexp/gen/avx512f-p5-scalef-x144.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003351 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160-acc2.c",
3352 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003353 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003354 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc2.c",
3355 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc3.c",
3356 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc6.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003357 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003358 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128-acc2.c",
3359 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003360 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003361 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x144-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003362 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x144.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003363 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160-acc2.c",
3364 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003365 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003366 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc2.c",
3367 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc3.c",
3368 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc6.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003369 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07003370 "src/f32-relu/gen/avx512f-x16.c",
3371 "src/f32-relu/gen/avx512f-x32.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003372 "src/f32-rmax/avx512f.c",
Marat Dukhand9ca7e62020-09-23 23:45:29 -07003373 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x16.c",
3374 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x32.c",
3375 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x48.c",
3376 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x64.c",
3377 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x80.c",
3378 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x96.c",
3379 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x112.c",
3380 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x128.c",
3381 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x16.c",
3382 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x32.c",
3383 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x48.c",
3384 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x64.c",
3385 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x80.c",
3386 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x96.c",
3387 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x112.c",
3388 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003389 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x16.c",
3390 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x32.c",
3391 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x48.c",
3392 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x64.c",
3393 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x80.c",
3394 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x96.c",
3395 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x112.c",
3396 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x128.c",
3397 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x16.c",
3398 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x32.c",
3399 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x48.c",
3400 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x64.c",
3401 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x80.c",
3402 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x96.c",
3403 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x112.c",
3404 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x128.c",
Marat Dukhand9ca7e62020-09-23 23:45:29 -07003405 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x16.c",
3406 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x32.c",
3407 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x48.c",
3408 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x64.c",
3409 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x80.c",
3410 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x96.c",
3411 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x112.c",
3412 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x128.c",
3413 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x16.c",
3414 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x32.c",
3415 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x48.c",
3416 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x64.c",
3417 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x80.c",
3418 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x96.c",
3419 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x112.c",
3420 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x128.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07003421 "src/f32-vbinary/gen/vadd-minmax-avx512f-x16.c",
3422 "src/f32-vbinary/gen/vadd-minmax-avx512f-x32.c",
3423 "src/f32-vbinary/gen/vaddc-minmax-avx512f-x16.c",
3424 "src/f32-vbinary/gen/vaddc-minmax-avx512f-x32.c",
3425 "src/f32-vbinary/gen/vdiv-minmax-avx512f-x16.c",
3426 "src/f32-vbinary/gen/vdiv-minmax-avx512f-x32.c",
3427 "src/f32-vbinary/gen/vdivc-minmax-avx512f-x16.c",
3428 "src/f32-vbinary/gen/vdivc-minmax-avx512f-x32.c",
Marat Dukhan9a88efe2019-12-10 15:54:24 -08003429 "src/f32-vbinary/gen/vmax-avx512f-x16.c",
3430 "src/f32-vbinary/gen/vmax-avx512f-x32.c",
3431 "src/f32-vbinary/gen/vmaxc-avx512f-x16.c",
3432 "src/f32-vbinary/gen/vmaxc-avx512f-x32.c",
3433 "src/f32-vbinary/gen/vmin-avx512f-x16.c",
3434 "src/f32-vbinary/gen/vmin-avx512f-x32.c",
3435 "src/f32-vbinary/gen/vminc-avx512f-x16.c",
3436 "src/f32-vbinary/gen/vminc-avx512f-x32.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07003437 "src/f32-vbinary/gen/vmul-minmax-avx512f-x16.c",
3438 "src/f32-vbinary/gen/vmul-minmax-avx512f-x32.c",
3439 "src/f32-vbinary/gen/vmulc-minmax-avx512f-x16.c",
3440 "src/f32-vbinary/gen/vmulc-minmax-avx512f-x32.c",
3441 "src/f32-vbinary/gen/vrdivc-minmax-avx512f-x16.c",
3442 "src/f32-vbinary/gen/vrdivc-minmax-avx512f-x32.c",
3443 "src/f32-vbinary/gen/vrsubc-minmax-avx512f-x16.c",
3444 "src/f32-vbinary/gen/vrsubc-minmax-avx512f-x32.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07003445 "src/f32-vbinary/gen/vsqrdiff-avx512f-x16.c",
3446 "src/f32-vbinary/gen/vsqrdiff-avx512f-x32.c",
3447 "src/f32-vbinary/gen/vsqrdiffc-avx512f-x16.c",
3448 "src/f32-vbinary/gen/vsqrdiffc-avx512f-x32.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07003449 "src/f32-vbinary/gen/vsub-minmax-avx512f-x16.c",
3450 "src/f32-vbinary/gen/vsub-minmax-avx512f-x32.c",
3451 "src/f32-vbinary/gen/vsubc-minmax-avx512f-x16.c",
3452 "src/f32-vbinary/gen/vsubc-minmax-avx512f-x32.c",
Marat Dukhaned6baaf2020-12-01 15:07:08 -08003453 "src/f32-velu/gen/velu-avx512f-rr1-lut16-p3-perm-x16.c",
3454 "src/f32-velu/gen/velu-avx512f-rr1-lut16-p3-perm-x32.c",
3455 "src/f32-velu/gen/velu-avx512f-rr1-lut16-p3-perm-x48.c",
3456 "src/f32-velu/gen/velu-avx512f-rr1-lut16-p3-perm-x64.c",
3457 "src/f32-velu/gen/velu-avx512f-rr1-lut16-p3-perm-x80.c",
3458 "src/f32-velu/gen/velu-avx512f-rr1-lut16-p3-perm-x96.c",
3459 "src/f32-velu/gen/velu-avx512f-rr1-lut16-p3-perm-x112.c",
3460 "src/f32-velu/gen/velu-avx512f-rr1-lut16-p3-perm-x128.c",
3461 "src/f32-velu/gen/velu-avx512f-rr1-p6-x16.c",
3462 "src/f32-velu/gen/velu-avx512f-rr1-p6-x32.c",
3463 "src/f32-velu/gen/velu-avx512f-rr1-p6-x48.c",
3464 "src/f32-velu/gen/velu-avx512f-rr1-p6-x64.c",
3465 "src/f32-velu/gen/velu-avx512f-rr1-p6-x80.c",
3466 "src/f32-velu/gen/velu-avx512f-rr1-p6-x96.c",
3467 "src/f32-velu/gen/velu-avx512f-rr1-p6-x112.c",
3468 "src/f32-velu/gen/velu-avx512f-rr1-p6-x128.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07003469 "src/f32-vlrelu/gen/vlrelu-avx512f-x16.c",
3470 "src/f32-vlrelu/gen/vlrelu-avx512f-x32.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003471 "src/f32-vrnd/gen/vrndd-avx512f-x16.c",
3472 "src/f32-vrnd/gen/vrndd-avx512f-x32.c",
3473 "src/f32-vrnd/gen/vrndne-avx512f-x16.c",
3474 "src/f32-vrnd/gen/vrndne-avx512f-x32.c",
3475 "src/f32-vrnd/gen/vrndu-avx512f-x16.c",
3476 "src/f32-vrnd/gen/vrndu-avx512f-x32.c",
3477 "src/f32-vrnd/gen/vrndz-avx512f-x16.c",
3478 "src/f32-vrnd/gen/vrndz-avx512f-x32.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07003479 "src/f32-vscale/avx512f-x64.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003480 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x16.c",
3481 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x32.c",
3482 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x48.c",
3483 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x64.c",
3484 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x80.c",
3485 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x96.c",
3486 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x112.c",
3487 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x128.c",
3488 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x144.c",
3489 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x160.c",
3490 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x176.c",
3491 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x192.c",
3492 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x16.c",
3493 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x32.c",
3494 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x48.c",
3495 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x64.c",
3496 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x80.c",
3497 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x96.c",
3498 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x112.c",
3499 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x128.c",
3500 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x144.c",
3501 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x160.c",
3502 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x176.c",
3503 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x192.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07003504 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x16.c",
3505 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x32.c",
3506 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x48.c",
3507 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x64.c",
3508 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x80.c",
3509 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x96.c",
3510 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x112.c",
3511 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x128.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07003512 "src/f32-vunary/gen/vabs-avx512f-x16.c",
3513 "src/f32-vunary/gen/vabs-avx512f-x32.c",
3514 "src/f32-vunary/gen/vneg-avx512f-x16.c",
3515 "src/f32-vunary/gen/vneg-avx512f-x32.c",
3516 "src/f32-vunary/gen/vsqr-avx512f-x16.c",
3517 "src/f32-vunary/gen/vsqr-avx512f-x32.c",
Marat Dukhanb7633f22020-11-20 16:34:56 -08003518 "src/math/exp-avx512f-rr2-lut16-p3-perm-scalef.c",
3519 "src/math/exp-avx512f-rr2-lut16-p3-perm.c",
3520 "src/math/exp-avx512f-rr2-lut32-p2-perm2-scalef.c",
3521 "src/math/exp-avx512f-rr2-lut32-p2-perm2.c",
3522 "src/math/exp-avx512f-rr2-p5-scalef.c",
3523 "src/math/exp-avx512f-rr2-p5.c",
Marat Dukhande390d42020-11-29 19:32:18 -08003524 "src/math/expm1minus-avx512f-rr1-lut16-p3-perm.c",
3525 "src/math/expm1minus-avx512f-rr1-p6.c",
Marat Dukhan98ba4412019-10-23 02:14:28 -07003526 "src/math/extexp-avx512f-p5.c",
Marat Dukhan6a34c5f2020-09-22 21:44:15 -07003527 "src/math/sigmoid-avx512f-rr1-lut16-p3-perm-scalef-div.c",
Marat Dukhan6a34c5f2020-09-22 21:44:15 -07003528 "src/math/sigmoid-avx512f-rr1-lut16-p3-perm-scalef-nr1fma.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003529 "src/math/sigmoid-avx512f-rr1-lut16-p3-perm-scalef-nr1fma1adj.c",
Marat Dukhan6a34c5f2020-09-22 21:44:15 -07003530 "src/math/sigmoid-avx512f-rr1-lut32-p2-perm2-scalef-div.c",
Marat Dukhan6a34c5f2020-09-22 21:44:15 -07003531 "src/math/sigmoid-avx512f-rr1-lut32-p2-perm2-scalef-nr1fma.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003532 "src/math/sigmoid-avx512f-rr1-lut32-p2-perm2-scalef-nr1fma1adj.c",
Marat Dukhan36173d22020-10-15 17:14:26 -07003533 "src/math/sigmoid-avx512f-rr1-lut64-p2-gather-scalef-div.c",
Marat Dukhan36173d22020-10-15 17:14:26 -07003534 "src/math/sigmoid-avx512f-rr1-lut64-p2-gather-scalef-nr1fma.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003535 "src/math/sigmoid-avx512f-rr1-lut64-p2-gather-scalef-nr1fma1adj.c",
3536 "src/math/sigmoid-avx512f-rr1-p5-scalef-div.c",
3537 "src/math/sigmoid-avx512f-rr1-p5-scalef-nr1fma.c",
3538 "src/math/sigmoid-avx512f-rr1-p5-scalef-nr1fma1adj.c",
3539 "src/math/sigmoid-avx512f-rr2-lut16-p3-perm-scalef-div.c",
3540 "src/math/sigmoid-avx512f-rr2-lut16-p3-perm-scalef-nr1fma.c",
3541 "src/math/sigmoid-avx512f-rr2-lut16-p3-perm-scalef-nr1fma1adj.c",
3542 "src/math/sigmoid-avx512f-rr2-lut32-p2-perm2-scalef-div.c",
3543 "src/math/sigmoid-avx512f-rr2-lut32-p2-perm2-scalef-nr1fma.c",
3544 "src/math/sigmoid-avx512f-rr2-lut32-p2-perm2-scalef-nr1fma1adj.c",
Marat Dukhan36173d22020-10-15 17:14:26 -07003545 "src/math/sigmoid-avx512f-rr2-lut64-p2-gather-scalef-div.c",
Marat Dukhan36173d22020-10-15 17:14:26 -07003546 "src/math/sigmoid-avx512f-rr2-lut64-p2-gather-scalef-nr1fma.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003547 "src/math/sigmoid-avx512f-rr2-lut64-p2-gather-scalef-nr1fma1adj.c",
3548 "src/math/sigmoid-avx512f-rr2-p5-scalef-div.c",
3549 "src/math/sigmoid-avx512f-rr2-p5-scalef-nr1fma.c",
3550 "src/math/sigmoid-avx512f-rr2-p5-scalef-nr1fma1adj.c",
Marat Dukhan84000762020-06-29 18:38:43 -07003551 "src/math/sqrt-avx512f-nr1fma.c",
Marat Dukhan84000762020-06-29 18:38:43 -07003552 "src/math/sqrt-avx512f-nr1fma1adj.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003553 "src/math/sqrt-avx512f-nr2fma.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003554]
3555
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07003556AVX512SKX_UKERNELS = [
Marat Dukhan2ffc5e62020-09-06 22:33:38 -07003557 "src/qs8-dwconv/gen/up16x9-minmax-avx512skx-mul32.c",
3558 "src/qs8-dwconv/gen/up32x9-minmax-avx512skx-mul32.c",
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07003559 "src/qs8-gemm/gen/1x16c8-minmax-avx512skx.c",
3560 "src/qs8-gemm/gen/2x16c8-minmax-avx512skx.c",
3561 "src/qs8-gemm/gen/3x16c8-minmax-avx512skx.c",
3562 "src/qs8-gemm/gen/4x16c8-minmax-avx512skx.c",
3563 "src/qs8-igemm/gen/1x16c8-minmax-avx512skx.c",
3564 "src/qs8-igemm/gen/2x16c8-minmax-avx512skx.c",
3565 "src/qs8-igemm/gen/3x16c8-minmax-avx512skx.c",
3566 "src/qs8-igemm/gen/4x16c8-minmax-avx512skx.c",
3567]
3568
Frank Barchardbcedc082020-08-17 18:00:51 -07003569WASM32_ASM_UKERNELS = [
3570 "src/f32-relu/wasm_shr_x1.S",
3571 "src/f32-relu/wasm_shr_x2.S",
3572 "src/f32-relu/wasm_shr_x4.S",
3573]
3574
Marat Dukhan08c4a432019-10-03 09:29:21 -07003575AARCH32_ASM_UKERNELS = [
Marat Dukhan32f93812020-05-17 20:31:21 -07003576 "src/f32-gemm/4x4-aarch32-vfp-ld64.S",
Marat Dukhan3b98f6b2020-05-17 10:09:22 -07003577 "src/f32-gemm/4x4-minmax-aarch32-vfp-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003578 "src/f32-gemm/4x8-minmax-aarch32-neon-cortex-a53.S",
3579 "src/f32-gemm/4x8-minmax-aarch32-neon-cortex-a55.S",
Frank Barchard490febe2020-07-16 18:42:17 -07003580 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-cortex-a7.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003581 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-cortex-a75.S",
Frank Barchard569561d2020-06-17 13:11:12 -07003582 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-ld64.S",
Frank Barchard490febe2020-07-16 18:42:17 -07003583 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-pld-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003584 "src/f32-igemm/4x8-minmax-aarch32-neon-cortex-a53.S",
3585 "src/f32-igemm/4x8-minmax-aarch32-neon-cortex-a55.S",
Frank Barchard490febe2020-07-16 18:42:17 -07003586 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-cortex-a7.S",
3587 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-cortex-a75.S",
3588 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-ld64.S",
3589 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-pld-cortex-a75.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003590]
3591
3592AARCH64_ASM_UKERNELS = [
Frank Barchardbddfbcd2020-04-15 12:32:41 -07003593 "src/f16-gemm/gen-inc/1x8inc-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchard04336c12020-10-22 16:48:55 -07003594 "src/f16-gemm/gen-inc/1x16inc-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchardbddfbcd2020-04-15 12:32:41 -07003595 "src/f16-gemm/gen-inc/4x8inc-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchard04336c12020-10-22 16:48:55 -07003596 "src/f16-gemm/gen-inc/4x16inc-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchardbddfbcd2020-04-15 12:32:41 -07003597 "src/f16-gemm/gen-inc/6x8inc-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchard04336c12020-10-22 16:48:55 -07003598 "src/f16-gemm/gen-inc/6x16inc-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchard3b8e5662020-04-20 12:12:53 -07003599 "src/f16-gemm/gen-inc/8x8inc-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchard04336c12020-10-22 16:48:55 -07003600 "src/f16-gemm/gen/1x8-minmax-aarch64-neonfp16arith-ld64.S",
3601 "src/f16-gemm/gen/1x16-minmax-aarch64-neonfp16arith-ld32.S",
3602 "src/f16-gemm/gen/4x8-minmax-aarch64-neonfp16arith-ld64.S",
3603 "src/f16-gemm/gen/4x16-minmax-aarch64-neonfp16arith-ld32.S",
3604 "src/f16-gemm/gen/6x8-minmax-aarch64-neonfp16arith-ld64.S",
3605 "src/f16-gemm/gen/6x16-minmax-aarch64-neonfp16arith-ld32.S",
3606 "src/f16-gemm/gen/8x8-minmax-aarch64-neonfp16arith-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003607 "src/f32-dwconv/up4x9-minmax-aarch64-neonfma-cortex-a55.S",
3608 "src/f32-dwconv/up4x9-minmax-aarch64-neonfma.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003609 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a53.S",
3610 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a57.S",
3611 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a75.S",
Frank Barchard04336c12020-10-22 16:48:55 -07003612 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-ld64.S",
3613 "src/f32-gemm/gen-inc/1x12inc-minmax-aarch64-neonfma-cortex-a53.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003614 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a53.S",
3615 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a55.S",
3616 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a57.S",
3617 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003618 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-ld64.S",
Frank Barchard04336c12020-10-22 16:48:55 -07003619 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-ld128.S",
3620 "src/f32-gemm/gen-inc/4x12inc-minmax-aarch64-neonfma-cortex-a53.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003621 "src/f32-gemm/gen-inc/5x8inc-minmax-aarch64-neonfma-cortex-a57.S",
3622 "src/f32-gemm/gen-inc/5x8inc-minmax-aarch64-neonfma-cortex-a75.S",
3623 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a53.S",
3624 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a55.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003625 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a57.S",
Frank Barchard04336c12020-10-22 16:48:55 -07003626 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a73.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003627 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003628 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-ld64.S",
Frank Barchard04336c12020-10-22 16:48:55 -07003629 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-ld128.S",
3630 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a53.S",
3631 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a57.S",
3632 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a75.S",
3633 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-ld64.S",
3634 "src/f32-gemm/gen/1x12-minmax-aarch64-neonfma-cortex-a53.S",
3635 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a53.S",
3636 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a55.S",
3637 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a57.S",
3638 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a75.S",
3639 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-ld64.S",
3640 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-ld128.S",
3641 "src/f32-gemm/gen/4x12-minmax-aarch64-neonfma-cortex-a53.S",
3642 "src/f32-gemm/gen/5x8-minmax-aarch64-neonfma-cortex-a57.S",
3643 "src/f32-gemm/gen/5x8-minmax-aarch64-neonfma-cortex-a75.S",
3644 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a53.S",
3645 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a55.S",
3646 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a57.S",
3647 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a73.S",
3648 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a75.S",
3649 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-ld64.S",
3650 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-ld128.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003651 "src/f32-igemm/1x8-minmax-aarch64-neonfma-cortex-a53.S",
Frank Barchard04336c12020-10-22 16:48:55 -07003652 "src/f32-igemm/1x12-minmax-aarch64-neonfma-cortex-a53.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003653 "src/f32-igemm/4x8-minmax-aarch64-neonfma-cortex-a53.S",
3654 "src/f32-igemm/4x8-minmax-aarch64-neonfma-cortex-a55.S",
Frank Barchard04336c12020-10-22 16:48:55 -07003655 "src/f32-igemm/4x12-minmax-aarch64-neonfma-cortex-a53.S",
3656 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a53.S",
3657 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a55.S",
3658 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a73.S",
3659 "src/f32-igemm/gen/1x8-minmax-aarch64-neonfma-cortex-a57.S",
3660 "src/f32-igemm/gen/1x8-minmax-aarch64-neonfma-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003661 "src/f32-igemm/gen/4x8-minmax-aarch64-neonfma-cortex-a57.S",
3662 "src/f32-igemm/gen/4x8-minmax-aarch64-neonfma-cortex-a75.S",
3663 "src/f32-igemm/gen/5x8-minmax-aarch64-neonfma-cortex-a57.S",
3664 "src/f32-igemm/gen/5x8-minmax-aarch64-neonfma-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003665 "src/f32-igemm/gen/6x8-minmax-aarch64-neonfma-cortex-a57.S",
3666 "src/f32-igemm/gen/6x8-minmax-aarch64-neonfma-cortex-a75.S",
Frank Barchard4a4be4e2021-03-01 15:00:28 -08003667 "src/qs8-gemm/1x16c4-aarch64-neondot-ld32.S",
Frank Barchardf1fd89e2020-10-06 23:27:11 -07003668 "src/qs8-gemm/1x16c4-aarch64-neondot-ld64.S",
Frank Barchardb8762632021-03-25 04:00:12 -07003669 "src/qs8-gemm/1x8c8-aarch64-neon-mlal-padal.S",
Frank Barchard1dc9fef2021-03-10 23:50:19 -08003670 "src/qs8-gemm/2x8c8-aarch64-neon-mull-padal.S",
Frank Barchardcbb8e702021-03-12 01:22:55 -08003671 "src/qs8-gemm/2x8c8-aarch64-neon-mlal-padal.S",
Frank Barchard5655cb72021-03-09 20:57:14 -08003672 "src/qs8-gemm/2x8c16-aarch64-neon-mlal-padal.S",
Frank Barchard671d1b02021-03-10 15:42:08 -08003673 "src/qs8-gemm/4x16c4-aarch64-neondot-cortex-a55.S",
Frank Barchardda78da12021-03-02 14:28:00 -08003674 "src/qs8-gemm/4x16c4-aarch64-neondot-ld32.S",
Frank Barchard671d1b02021-03-10 15:42:08 -08003675 "src/qs8-gemm/4x16c4-aarch64-neondot-ld64.S",
Frank Barchard6e35de52021-03-25 04:33:21 -07003676 "src/qs8-igemm/1x8c8-aarch64-neon-mlal-padal.S",
Frank Barchardcbb8e702021-03-12 01:22:55 -08003677 "src/qs8-igemm/2x8c8-aarch64-neon-mlal-padal.S",
Frank Barchard7ca54df2021-03-12 00:09:02 -08003678 "src/qs8-igemm/2x8c16-aarch64-neon-mlal-padal.S",
Frank Barchard89e12f82021-03-08 11:42:44 -08003679 "src/qs8-igemm/4x16c4-aarch64-neondot-cortex-a55.S",
Frank Barchard671d1b02021-03-10 15:42:08 -08003680 "src/qs8-igemm/4x16c4-aarch64-neondot-ld64.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003681]
3682
Marat Dukhan1b354632020-03-23 12:50:22 -07003683INTERNAL_MICROKERNEL_HDRS = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003684 "src/xnnpack/argmaxpool.h",
3685 "src/xnnpack/avgpool.h",
3686 "src/xnnpack/clamp.h",
3687 "src/xnnpack/common.h",
3688 "src/xnnpack/conv.h",
Yury Kartynnike7841862020-11-04 18:22:18 -08003689 "src/xnnpack/depthtospace.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003690 "src/xnnpack/dwconv.h",
Frank Barchard04336c12020-10-22 16:48:55 -07003691 "src/xnnpack/fill.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003692 "src/xnnpack/gavgpool.h",
3693 "src/xnnpack/gemm.h",
3694 "src/xnnpack/hswish.h",
Marat Dukhan660fd192020-03-10 04:55:30 -07003695 "src/xnnpack/ibilinear.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003696 "src/xnnpack/igemm.h",
Marat Dukhancfb31342019-12-05 10:42:57 -08003697 "src/xnnpack/intrinsics-polyfill.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003698 "src/xnnpack/lut.h",
3699 "src/xnnpack/math.h",
3700 "src/xnnpack/maxpool.h",
3701 "src/xnnpack/packx.h",
3702 "src/xnnpack/pad.h",
3703 "src/xnnpack/params.h",
3704 "src/xnnpack/pavgpool.h",
3705 "src/xnnpack/ppmm.h",
3706 "src/xnnpack/prelu.h",
Marat Dukhan97579532019-10-18 16:40:39 -07003707 "src/xnnpack/raddexpminusmax.h",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07003708 "src/xnnpack/raddextexp.h",
Marat Dukhan97579532019-10-18 16:40:39 -07003709 "src/xnnpack/raddstoreexpminusmax.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003710 "src/xnnpack/rmax.h",
3711 "src/xnnpack/scalar-utils.h",
3712 "src/xnnpack/spmm.h",
3713 "src/xnnpack/unpool.h",
3714 "src/xnnpack/vadd.h",
Marat Dukhan1e782c42019-11-21 17:02:40 -08003715 "src/xnnpack/vbinary.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003716 "src/xnnpack/vmulcaddc.h",
Marat Dukhan05ac8e32019-10-21 15:39:33 -07003717 "src/xnnpack/vscale.h",
Marat Dukhan97579532019-10-18 16:40:39 -07003718 "src/xnnpack/vscaleexpminusmax.h",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07003719 "src/xnnpack/vscaleextexp.h",
Marat Dukhan1e782c42019-11-21 17:02:40 -08003720 "src/xnnpack/vunary.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003721 "src/xnnpack/zip.h",
Marat Dukhan1b354632020-03-23 12:50:22 -07003722]
3723
3724INTERNAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003725 "include/xnnpack.h",
3726 "src/xnnpack/allocator.h",
3727 "src/xnnpack/compute.h",
3728 "src/xnnpack/im2col.h",
3729 "src/xnnpack/indirection.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003730 "src/xnnpack/math-stubs.h",
Chao Mei6ddfc602020-05-13 22:29:36 -07003731 "src/xnnpack/memory-planner.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003732 "src/xnnpack/operator.h",
3733 "src/xnnpack/pack.h",
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07003734 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003735 "src/xnnpack/requantization-stubs.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003736 "src/xnnpack/requantization.h",
Marat Dukhan1d75a542020-02-03 12:23:01 -08003737 "src/xnnpack/subgraph.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003738]
3739
Marat Dukhan1b354632020-03-23 12:50:22 -07003740ACCURACY_EVAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003741 "src/xnnpack/math-stubs.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003742]
3743
Marat Dukhan1b354632020-03-23 12:50:22 -07003744MICROKERNEL_BENCHMARK_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003745 "include/xnnpack.h",
Frank Barchard35db7d02020-10-26 13:37:34 -07003746 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003747]
3748
Marat Dukhan1b354632020-03-23 12:50:22 -07003749MICROKERNEL_TEST_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Frank Barchard35db7d02020-10-26 13:37:34 -07003750 "include/xnnpack.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003751 "src/xnnpack/isa-checks.h",
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07003752 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003753 "src/xnnpack/requantization.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003754]
3755
3756OPERATOR_TEST_PARAMS_HDRS = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003757 "src/xnnpack/common.h",
Frank Barchard04336c12020-10-22 16:48:55 -07003758 "src/xnnpack/params.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003759]
3760
3761WEIGHTS_PACK_HDRS = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003762 "src/xnnpack/compute.h",
Frank Barchard04336c12020-10-22 16:48:55 -07003763 "src/xnnpack/operator.h",
3764 "src/xnnpack/pack.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003765]
3766
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07003767LOGGING_COPTS = select({
3768 # No logging in optimized mode
3769 ":optimized_build": ["-DXNN_LOG_LEVEL=0"],
3770 # Full logging in debug mode
3771 ":debug_build": ["-DXNN_LOG_LEVEL=5"],
3772 # Error-only logging in default (fastbuild) mode
3773 "//conditions:default": ["-DXNN_LOG_LEVEL=2"],
3774})
3775
Marat Dukhan3b59de22020-06-03 20:15:19 -07003776LOGGING_SRCS = select({
3777 # No logging in optimized mode
3778 ":optimized_build": [],
3779 "//conditions:default": [
Marat Dukhanccd3a1d2021-03-29 16:03:12 -07003780 "src/datatype-strings.c",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003781 "src/operator-strings.c",
3782 "src/subgraph-strings.c",
3783 ],
3784})
3785
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07003786LOGGING_HDRS = [
3787 "src/xnnpack/log.h",
3788]
3789
Marat Dukhan08c4a432019-10-03 09:29:21 -07003790xnnpack_cc_library(
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003791 name = "tables",
3792 srcs = TABLE_SRCS,
3793 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003794 gcc_copts = xnnpack_gcc_std_copts(),
3795 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003796)
3797
3798xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003799 name = "scalar_ukernels",
3800 srcs = SCALAR_UKERNELS,
3801 hdrs = INTERNAL_HDRS,
3802 aarch32_copts = ["-marm"],
Marat Dukhan10a38082020-04-17 03:58:35 -07003803 gcc_copts = xnnpack_gcc_std_copts(),
3804 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003805 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003806 ":tables",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003807 "@FP16",
3808 "@FXdiv",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003809 "@pthreadpool",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003810 ],
3811)
3812
3813xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003814 name = "scalar_ukernels_test_mode",
3815 srcs = SCALAR_UKERNELS,
3816 hdrs = INTERNAL_HDRS,
3817 aarch32_copts = ["-marm"],
3818 copts = [
3819 "-UNDEBUG",
3820 "-DXNN_TEST_MODE=1",
3821 ],
3822 gcc_copts = xnnpack_gcc_std_copts(),
3823 msvc_copts = xnnpack_msvc_std_copts(),
3824 deps = [
3825 ":tables",
3826 "@FP16",
3827 "@FXdiv",
3828 "@pthreadpool",
3829 ],
3830)
3831
3832xnnpack_cc_library(
Marat Dukhan436ebe62019-12-04 15:10:12 -08003833 name = "wasm_ukernels",
3834 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003835 gcc_copts = xnnpack_gcc_std_copts(),
3836 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan436ebe62019-12-04 15:10:12 -08003837 wasm_srcs = WASM_UKERNELS,
Marat Dukhan1c6e3892020-06-25 23:56:10 -07003838 wasmsimd_srcs = WASM_UKERNELS + WASMSIMD_UKERNELS,
Marat Dukhan436ebe62019-12-04 15:10:12 -08003839 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003840 ":tables",
Marat Dukhan436ebe62019-12-04 15:10:12 -08003841 "@FP16",
3842 "@FXdiv",
3843 "@pthreadpool",
3844 ],
3845)
3846
3847xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003848 name = "wasm_ukernels_test_mode",
3849 hdrs = INTERNAL_HDRS,
3850 copts = [
3851 "-UNDEBUG",
3852 "-DXNN_TEST_MODE=1",
3853 ],
3854 gcc_copts = xnnpack_gcc_std_copts(),
3855 msvc_copts = xnnpack_msvc_std_copts(),
3856 wasm_srcs = WASM_UKERNELS,
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07003857 wasmsimd_srcs = WASM_UKERNELS + WASMSIMD_UKERNELS,
Marat Dukhan33fcf782020-05-24 14:27:15 -07003858 deps = [
3859 ":tables",
3860 "@FP16",
3861 "@FXdiv",
3862 "@pthreadpool",
3863 ],
3864)
3865
3866xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003867 name = "neon_ukernels",
3868 hdrs = INTERNAL_HDRS,
3869 aarch32_copts = [
3870 "-marm",
Marat Dukhan8853b822020-05-07 12:19:01 -07003871 "-march=armv7-a",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003872 "-mfpu=neon",
3873 ],
3874 aarch32_srcs = NEON_UKERNELS,
3875 aarch64_srcs = NEON_UKERNELS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003876 gcc_copts = xnnpack_gcc_std_copts(),
3877 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08003878 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003879 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003880 "@FP16",
3881 "@pthreadpool",
3882 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003883)
3884
3885xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003886 name = "neon_ukernels_test_mode",
3887 hdrs = INTERNAL_HDRS,
3888 aarch32_copts = [
3889 "-marm",
3890 "-march=armv7-a",
3891 "-mfpu=neon",
3892 ],
3893 aarch32_srcs = NEON_UKERNELS,
3894 aarch64_srcs = NEON_UKERNELS,
3895 copts = [
3896 "-UNDEBUG",
3897 "-DXNN_TEST_MODE=1",
3898 ],
3899 gcc_copts = xnnpack_gcc_std_copts(),
3900 msvc_copts = xnnpack_msvc_std_copts(),
3901 deps = [
3902 ":tables",
3903 "@FP16",
3904 "@pthreadpool",
3905 ],
3906)
3907
3908xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003909 name = "neonfma_ukernels",
3910 hdrs = INTERNAL_HDRS,
3911 aarch32_copts = [
3912 "-marm",
Marat Dukhan8853b822020-05-07 12:19:01 -07003913 "-march=armv7-a",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003914 "-mfpu=neon-vfpv4",
3915 ],
3916 aarch32_srcs = NEONFMA_UKERNELS,
3917 aarch64_srcs = NEONFMA_UKERNELS + AARCH64_NEONFMA_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07003918 apple_aarch32_copts = [
3919 "-mcpu=swift",
3920 "-mtune=generic",
3921 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07003922 gcc_copts = xnnpack_gcc_std_copts(),
3923 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08003924 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003925 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003926 "@FP16",
3927 "@pthreadpool",
3928 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003929)
3930
3931xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003932 name = "neonfma_ukernels_test_mode",
3933 hdrs = INTERNAL_HDRS,
3934 aarch32_copts = [
3935 "-marm",
3936 "-march=armv7-a",
3937 "-mfpu=neon-vfpv4",
3938 ],
3939 aarch32_srcs = NEONFMA_UKERNELS,
3940 aarch64_srcs = NEONFMA_UKERNELS + AARCH64_NEONFMA_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07003941 apple_aarch32_copts = [
3942 "-mcpu=swift",
3943 "-mtune=generic",
3944 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07003945 copts = [
3946 "-UNDEBUG",
3947 "-DXNN_TEST_MODE=1",
3948 ],
3949 gcc_copts = xnnpack_gcc_std_copts(),
3950 msvc_copts = xnnpack_msvc_std_copts(),
3951 deps = [
3952 ":tables",
3953 "@FP16",
3954 "@pthreadpool",
3955 ],
3956)
3957
3958xnnpack_cc_library(
Marat Dukhan8853b822020-05-07 12:19:01 -07003959 name = "neonv8_ukernels",
3960 hdrs = INTERNAL_HDRS,
3961 aarch32_copts = [
3962 "-marm",
3963 "-march=armv8-a",
3964 "-mfpu=neon-fp-armv8",
3965 ],
3966 aarch32_srcs = NEONV8_UKERNELS,
3967 aarch64_srcs = NEONV8_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07003968 apple_aarch32_copts = [
3969 "-mcpu=cyclone",
3970 "-mtune=generic",
3971 ],
Marat Dukhan8853b822020-05-07 12:19:01 -07003972 gcc_copts = xnnpack_gcc_std_copts(),
3973 msvc_copts = xnnpack_msvc_std_copts(),
3974 deps = [
3975 ":tables",
3976 "@FP16",
3977 "@pthreadpool",
3978 ],
3979)
3980
3981xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003982 name = "neonv8_ukernels_test_mode",
3983 hdrs = INTERNAL_HDRS,
3984 aarch32_copts = [
3985 "-marm",
3986 "-march=armv8-a",
3987 "-mfpu=neon-fp-armv8",
3988 ],
3989 aarch32_srcs = NEONV8_UKERNELS,
3990 aarch64_srcs = NEONV8_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07003991 apple_aarch32_copts = [
3992 "-mcpu=cyclone",
3993 "-mtune=generic",
3994 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07003995 copts = [
3996 "-UNDEBUG",
3997 "-DXNN_TEST_MODE=1",
3998 ],
3999 gcc_copts = xnnpack_gcc_std_copts(),
4000 msvc_copts = xnnpack_msvc_std_copts(),
4001 deps = [
4002 ":tables",
4003 "@FP16",
4004 "@pthreadpool",
4005 ],
4006)
4007
4008xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004009 name = "neonfp16arith_ukernels",
4010 hdrs = INTERNAL_HDRS,
4011 aarch64_copts = ["-march=armv8.2-a+fp16"],
4012 aarch64_srcs = AARCH64_NEONFP16ARITH_UKERNELS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004013 gcc_copts = xnnpack_gcc_std_copts(),
4014 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08004015 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08004016 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08004017 "@FP16",
4018 "@pthreadpool",
4019 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004020)
4021
4022xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07004023 name = "neonfp16arith_ukernels_test_mode",
4024 hdrs = INTERNAL_HDRS,
4025 aarch64_copts = ["-march=armv8.2-a+fp16"],
4026 aarch64_srcs = AARCH64_NEONFP16ARITH_UKERNELS,
4027 copts = [
4028 "-UNDEBUG",
4029 "-DXNN_TEST_MODE=1",
4030 ],
4031 gcc_copts = xnnpack_gcc_std_copts(),
4032 msvc_copts = xnnpack_msvc_std_copts(),
4033 deps = [
4034 ":tables",
4035 "@FP16",
4036 "@pthreadpool",
4037 ],
4038)
4039
4040xnnpack_cc_library(
Benoit Jacoba9644732020-08-13 12:48:55 -07004041 name = "neondot_ukernels",
4042 hdrs = INTERNAL_HDRS,
Marat Dukhan799ac752020-08-13 16:08:03 -07004043 aarch32_copts = [
4044 "-marm",
4045 "-march=armv8.2-a+dotprod",
4046 "-mfpu=neon-fp-armv8",
4047 ],
Benoit Jacoba9644732020-08-13 12:48:55 -07004048 aarch32_srcs = NEONDOT_UKERNELS,
4049 aarch64_copts = ["-march=armv8.2-a+dotprod"],
4050 aarch64_srcs = NEONDOT_UKERNELS,
4051 gcc_copts = xnnpack_gcc_std_copts(),
4052 msvc_copts = xnnpack_msvc_std_copts(),
4053 deps = [
4054 ":tables",
4055 "@FP16",
4056 "@pthreadpool",
4057 ],
4058)
4059
4060xnnpack_cc_library(
4061 name = "neondot_ukernels_test_mode",
4062 hdrs = INTERNAL_HDRS,
Marat Dukhan799ac752020-08-13 16:08:03 -07004063 aarch32_copts = [
4064 "-marm",
4065 "-march=armv8.2-a+dotprod",
4066 "-mfpu=neon-fp-armv8",
4067 ],
Benoit Jacoba9644732020-08-13 12:48:55 -07004068 aarch32_srcs = NEONDOT_UKERNELS,
4069 aarch64_copts = ["-march=armv8.2-a+dotprod"],
4070 aarch64_srcs = NEONDOT_UKERNELS,
4071 copts = [
4072 "-UNDEBUG",
4073 "-DXNN_TEST_MODE=1",
4074 ],
4075 gcc_copts = xnnpack_gcc_std_copts(),
4076 msvc_copts = xnnpack_msvc_std_copts(),
4077 deps = [
4078 ":tables",
4079 "@FP16",
4080 "@pthreadpool",
4081 ],
4082)
4083
4084xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004085 name = "sse2_ukernels",
4086 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004087 gcc_copts = xnnpack_gcc_std_copts(),
4088 gcc_x86_copts = ["-msse2"],
4089 msvc_copts = xnnpack_msvc_std_copts(),
4090 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004091 x86_srcs = SSE_UKERNELS + SSE2_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08004092 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08004093 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08004094 "@FP16",
4095 "@pthreadpool",
4096 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004097)
4098
4099xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07004100 name = "sse2_ukernels_test_mode",
4101 hdrs = INTERNAL_HDRS,
4102 copts = [
4103 "-UNDEBUG",
4104 "-DXNN_TEST_MODE=1",
4105 ],
4106 gcc_copts = xnnpack_gcc_std_copts(),
4107 gcc_x86_copts = ["-msse2"],
4108 msvc_copts = xnnpack_msvc_std_copts(),
4109 msvc_x86_32_copts = ["/arch:SSE2"],
4110 x86_srcs = SSE_UKERNELS + SSE2_UKERNELS,
4111 deps = [
4112 ":tables",
4113 "@FP16",
4114 "@pthreadpool",
4115 ],
4116)
4117
4118xnnpack_cc_library(
Marat Dukhanfe7acb62020-03-09 19:30:05 -07004119 name = "ssse3_ukernels",
4120 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004121 gcc_copts = xnnpack_gcc_std_copts(),
4122 gcc_x86_copts = ["-mssse3"],
4123 msvc_copts = xnnpack_msvc_std_copts(),
4124 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhanfe7acb62020-03-09 19:30:05 -07004125 x86_srcs = SSSE3_UKERNELS,
4126 deps = [
4127 ":tables",
4128 "@FP16",
4129 "@pthreadpool",
4130 ],
4131)
4132
4133xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07004134 name = "ssse3_ukernels_test_mode",
4135 hdrs = INTERNAL_HDRS,
4136 copts = [
4137 "-UNDEBUG",
4138 "-DXNN_TEST_MODE=1",
4139 ],
4140 gcc_copts = xnnpack_gcc_std_copts(),
4141 gcc_x86_copts = ["-mssse3"],
4142 msvc_copts = xnnpack_msvc_std_copts(),
4143 msvc_x86_32_copts = ["/arch:SSE2"],
4144 x86_srcs = SSSE3_UKERNELS,
4145 deps = [
4146 ":tables",
4147 "@FP16",
4148 "@pthreadpool",
4149 ],
4150)
4151
4152xnnpack_cc_library(
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08004153 name = "sse41_ukernels",
4154 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004155 gcc_copts = xnnpack_gcc_std_copts(),
4156 gcc_x86_copts = ["-msse4.1"],
4157 msvc_copts = xnnpack_msvc_std_copts(),
4158 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08004159 x86_srcs = SSE41_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08004160 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08004161 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08004162 "@FP16",
4163 "@pthreadpool",
4164 ],
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08004165)
4166
4167xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07004168 name = "sse41_ukernels_test_mode",
4169 hdrs = INTERNAL_HDRS,
4170 copts = [
4171 "-UNDEBUG",
4172 "-DXNN_TEST_MODE=1",
4173 ],
4174 gcc_copts = xnnpack_gcc_std_copts(),
4175 gcc_x86_copts = ["-msse4.1"],
4176 msvc_copts = xnnpack_msvc_std_copts(),
4177 msvc_x86_32_copts = ["/arch:SSE2"],
4178 x86_srcs = SSE41_UKERNELS,
4179 deps = [
4180 ":tables",
4181 "@FP16",
4182 "@pthreadpool",
4183 ],
4184)
4185
4186xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004187 name = "avx_ukernels",
4188 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004189 gcc_copts = xnnpack_gcc_std_copts(),
4190 gcc_x86_copts = ["-mavx"],
4191 msvc_copts = xnnpack_msvc_std_copts(),
4192 msvc_x86_32_copts = ["/arch:AVX"],
4193 msvc_x86_64_copts = ["/arch:AVX"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004194 x86_srcs = AVX_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08004195 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08004196 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08004197 "@FP16",
4198 "@pthreadpool",
4199 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004200)
4201
4202xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07004203 name = "avx_ukernels_test_mode",
4204 hdrs = INTERNAL_HDRS,
4205 copts = [
4206 "-UNDEBUG",
4207 "-DXNN_TEST_MODE=1",
4208 ],
4209 gcc_copts = xnnpack_gcc_std_copts(),
4210 gcc_x86_copts = ["-mavx"],
4211 msvc_copts = xnnpack_msvc_std_copts(),
4212 msvc_x86_32_copts = ["/arch:AVX"],
4213 msvc_x86_64_copts = ["/arch:AVX"],
4214 x86_srcs = AVX_UKERNELS,
4215 deps = [
4216 ":tables",
4217 "@FP16",
4218 "@pthreadpool",
4219 ],
4220)
4221
4222xnnpack_cc_library(
Marat Dukhan1566fee2020-08-02 21:55:41 -07004223 name = "xop_ukernels",
4224 hdrs = INTERNAL_HDRS,
4225 gcc_copts = xnnpack_gcc_std_copts(),
4226 gcc_x86_copts = ["-mxop"],
4227 msvc_copts = xnnpack_msvc_std_copts(),
4228 msvc_x86_32_copts = ["/arch:AVX"],
4229 msvc_x86_64_copts = ["/arch:AVX"],
4230 x86_srcs = XOP_UKERNELS,
4231 deps = [
4232 ":tables",
4233 "@FP16",
4234 "@pthreadpool",
4235 ],
4236)
4237
4238xnnpack_cc_library(
4239 name = "xop_ukernels_test_mode",
4240 hdrs = INTERNAL_HDRS,
4241 copts = [
4242 "-UNDEBUG",
4243 "-DXNN_TEST_MODE=1",
4244 ],
4245 gcc_copts = xnnpack_gcc_std_copts(),
4246 gcc_x86_copts = ["-mxop"],
4247 msvc_copts = xnnpack_msvc_std_copts(),
4248 msvc_x86_32_copts = ["/arch:AVX"],
4249 msvc_x86_64_copts = ["/arch:AVX"],
4250 x86_srcs = XOP_UKERNELS,
4251 deps = [
4252 ":tables",
4253 "@FP16",
4254 "@pthreadpool",
4255 ],
4256)
4257
4258xnnpack_cc_library(
Marat Dukhanfda12b82019-11-21 12:27:59 -08004259 name = "fma3_ukernels",
4260 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004261 gcc_copts = xnnpack_gcc_std_copts(),
4262 gcc_x86_copts = ["-mfma"],
4263 msvc_copts = xnnpack_msvc_std_copts(),
4264 msvc_x86_32_copts = ["/arch:AVX"],
4265 msvc_x86_64_copts = ["/arch:AVX"],
Marat Dukhanfda12b82019-11-21 12:27:59 -08004266 x86_srcs = FMA3_UKERNELS,
4267 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08004268 ":tables",
Marat Dukhanfda12b82019-11-21 12:27:59 -08004269 "@FP16",
4270 "@pthreadpool",
4271 ],
4272)
4273
4274xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07004275 name = "fma3_ukernels_test_mode",
4276 hdrs = INTERNAL_HDRS,
4277 copts = [
4278 "-UNDEBUG",
4279 "-DXNN_TEST_MODE=1",
4280 ],
4281 gcc_copts = xnnpack_gcc_std_copts(),
4282 gcc_x86_copts = ["-mfma"],
4283 msvc_copts = xnnpack_msvc_std_copts(),
4284 msvc_x86_32_copts = ["/arch:AVX"],
4285 msvc_x86_64_copts = ["/arch:AVX"],
4286 x86_srcs = FMA3_UKERNELS,
4287 deps = [
4288 ":tables",
4289 "@FP16",
4290 "@pthreadpool",
4291 ],
4292)
4293
4294xnnpack_cc_library(
Marat Dukhan6adff4e2019-10-14 18:32:07 -07004295 name = "avx2_ukernels",
4296 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004297 gcc_copts = xnnpack_gcc_std_copts(),
4298 gcc_x86_copts = [
Marat Dukhan6adff4e2019-10-14 18:32:07 -07004299 "-mfma",
4300 "-mavx2",
4301 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07004302 msvc_copts = xnnpack_msvc_std_copts(),
4303 msvc_x86_32_copts = ["/arch:AVX2"],
4304 msvc_x86_64_copts = ["/arch:AVX2"],
Marat Dukhan6adff4e2019-10-14 18:32:07 -07004305 x86_srcs = AVX2_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08004306 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08004307 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08004308 "@FP16",
4309 "@pthreadpool",
4310 ],
Marat Dukhan6adff4e2019-10-14 18:32:07 -07004311)
4312
4313xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07004314 name = "avx2_ukernels_test_mode",
4315 hdrs = INTERNAL_HDRS,
4316 copts = [
4317 "-UNDEBUG",
4318 "-DXNN_TEST_MODE=1",
4319 ],
4320 gcc_copts = xnnpack_gcc_std_copts(),
4321 gcc_x86_copts = [
4322 "-mfma",
4323 "-mavx2",
4324 ],
4325 msvc_copts = xnnpack_msvc_std_copts(),
4326 msvc_x86_32_copts = ["/arch:AVX2"],
4327 msvc_x86_64_copts = ["/arch:AVX2"],
4328 x86_srcs = AVX2_UKERNELS,
4329 deps = [
4330 ":tables",
4331 "@FP16",
4332 "@pthreadpool",
4333 ],
4334)
4335
4336xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004337 name = "avx512f_ukernels",
4338 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004339 gcc_copts = xnnpack_gcc_std_copts(),
4340 gcc_x86_copts = ["-mavx512f"],
4341 mingw_copts = ["-fno-asynchronous-unwind-tables"],
4342 msvc_copts = xnnpack_msvc_std_copts(),
4343 msvc_x86_32_copts = ["/arch:AVX512"],
4344 msvc_x86_64_copts = ["/arch:AVX512"],
4345 msys_copts = ["-fno-asynchronous-unwind-tables"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004346 x86_srcs = AVX512F_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08004347 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08004348 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08004349 "@FP16",
4350 "@pthreadpool",
4351 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004352)
4353
4354xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07004355 name = "avx512f_ukernels_test_mode",
4356 hdrs = INTERNAL_HDRS,
4357 copts = [
4358 "-UNDEBUG",
4359 "-DXNN_TEST_MODE=1",
4360 ],
4361 gcc_copts = xnnpack_gcc_std_copts(),
4362 gcc_x86_copts = ["-mavx512f"],
4363 mingw_copts = ["-fno-asynchronous-unwind-tables"],
4364 msvc_copts = xnnpack_msvc_std_copts(),
4365 msvc_x86_32_copts = ["/arch:AVX512"],
4366 msvc_x86_64_copts = ["/arch:AVX512"],
4367 msys_copts = ["-fno-asynchronous-unwind-tables"],
4368 x86_srcs = AVX512F_UKERNELS,
4369 deps = [
4370 ":tables",
4371 "@FP16",
4372 "@pthreadpool",
4373 ],
4374)
4375
4376xnnpack_cc_library(
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07004377 name = "avx512skx_ukernels",
4378 hdrs = INTERNAL_HDRS,
4379 gcc_copts = xnnpack_gcc_std_copts(),
4380 gcc_x86_copts = [
4381 "-mavx512f",
4382 "-mavx512cd",
4383 "-mavx512bw",
4384 "-mavx512dq",
4385 "-mavx512vl",
4386 ],
4387 mingw_copts = ["-fno-asynchronous-unwind-tables"],
4388 msvc_copts = xnnpack_msvc_std_copts(),
4389 msvc_x86_32_copts = ["/arch:AVX512"],
4390 msvc_x86_64_copts = ["/arch:AVX512"],
4391 msys_copts = ["-fno-asynchronous-unwind-tables"],
4392 x86_srcs = AVX512SKX_UKERNELS,
4393 deps = [
4394 ":tables",
4395 "@FP16",
4396 "@pthreadpool",
4397 ],
4398)
4399
4400xnnpack_cc_library(
4401 name = "avx512skx_ukernels_test_mode",
4402 hdrs = INTERNAL_HDRS,
4403 copts = [
4404 "-UNDEBUG",
4405 "-DXNN_TEST_MODE=1",
4406 ],
4407 gcc_copts = xnnpack_gcc_std_copts(),
4408 gcc_x86_copts = [
4409 "-mavx512f",
4410 "-mavx512cd",
4411 "-mavx512bw",
4412 "-mavx512dq",
4413 "-mavx512vl",
4414 ],
4415 mingw_copts = ["-fno-asynchronous-unwind-tables"],
4416 msvc_copts = xnnpack_msvc_std_copts(),
4417 msvc_x86_32_copts = ["/arch:AVX512"],
4418 msvc_x86_64_copts = ["/arch:AVX512"],
4419 msys_copts = ["-fno-asynchronous-unwind-tables"],
4420 x86_srcs = AVX512SKX_UKERNELS,
4421 deps = [
4422 ":tables",
4423 "@FP16",
4424 "@pthreadpool",
4425 ],
4426)
4427
4428xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004429 name = "asm_ukernels",
4430 hdrs = ["src/xnnpack/assembly.h"],
4431 aarch32_srcs = AARCH32_ASM_UKERNELS,
Frank Barchard31bb45b2020-10-06 00:26:33 -07004432 aarch64_copts = ["-march=armv8.2-a+fp16+dotprod"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004433 aarch64_srcs = AARCH64_ASM_UKERNELS,
Frank Barchardbcedc082020-08-17 18:00:51 -07004434 wasm_srcs = WASM32_ASM_UKERNELS,
4435 wasmsimd_srcs = WASM32_ASM_UKERNELS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07004436)
4437
Marat Dukhan3b59de22020-06-03 20:15:19 -07004438xnnpack_cc_library(
4439 name = "logging_utils",
4440 srcs = LOGGING_SRCS,
4441 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
4442 copts = LOGGING_COPTS + [
4443 "-Isrc",
4444 "-Iinclude",
4445 ] + select({
4446 ":debug_build": [],
4447 "//conditions:default": xnnpack_min_size_copts(),
4448 }),
4449 gcc_copts = xnnpack_gcc_std_copts(),
4450 msvc_copts = xnnpack_msvc_std_copts(),
4451 visibility = xnnpack_visibility(),
4452 deps = [
4453 "@FP16",
4454 "@clog",
4455 "@pthreadpool",
4456 ],
4457)
4458
Marat Dukhan08c4a432019-10-03 09:29:21 -07004459xnnpack_aggregate_library(
4460 name = "ukernels",
4461 aarch32_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07004462 ":neon_ukernels",
4463 ":neonfma_ukernels",
Marat Dukhan8853b822020-05-07 12:19:01 -07004464 ":neonv8_ukernels",
Benoit Jacoba9644732020-08-13 12:48:55 -07004465 ":neondot_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004466 ":asm_ukernels",
4467 ],
4468 aarch64_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07004469 ":neon_ukernels",
4470 ":neonfma_ukernels",
Marat Dukhan8853b822020-05-07 12:19:01 -07004471 ":neonv8_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004472 ":neonfp16arith_ukernels",
Benoit Jacoba9644732020-08-13 12:48:55 -07004473 ":neondot_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004474 ":asm_ukernels",
4475 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07004476 generic_deps = [
4477 ":scalar_ukernels",
4478 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07004479 wasm_deps = [
4480 ":wasm_ukernels",
Frank Barchardbcedc082020-08-17 18:00:51 -07004481 ":asm_ukernels",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004482 ],
4483 wasmsimd_deps = [
4484 ":wasm_ukernels",
Frank Barchardbcedc082020-08-17 18:00:51 -07004485 ":asm_ukernels",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004486 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004487 x86_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07004488 ":sse2_ukernels",
Marat Dukhanfe7acb62020-03-09 19:30:05 -07004489 ":ssse3_ukernels",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08004490 ":sse41_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004491 ":avx_ukernels",
Marat Dukhan1566fee2020-08-02 21:55:41 -07004492 ":xop_ukernels",
Marat Dukhanfda12b82019-11-21 12:27:59 -08004493 ":fma3_ukernels",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07004494 ":avx2_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004495 ":avx512f_ukernels",
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07004496 ":avx512skx_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004497 ],
4498)
4499
Marat Dukhan33fcf782020-05-24 14:27:15 -07004500xnnpack_aggregate_library(
4501 name = "ukernels_test_mode",
4502 aarch32_deps = [
4503 ":neon_ukernels_test_mode",
4504 ":neonfma_ukernels_test_mode",
4505 ":neonv8_ukernels_test_mode",
Benoit Jacoba9644732020-08-13 12:48:55 -07004506 ":neondot_ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004507 ":asm_ukernels",
4508 ],
4509 aarch64_deps = [
4510 ":neon_ukernels_test_mode",
4511 ":neonfma_ukernels_test_mode",
4512 ":neonv8_ukernels_test_mode",
4513 ":neonfp16arith_ukernels_test_mode",
Benoit Jacoba9644732020-08-13 12:48:55 -07004514 ":neondot_ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004515 ":asm_ukernels",
4516 ],
4517 generic_deps = [
4518 ":scalar_ukernels_test_mode",
4519 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07004520 wasm_deps = [
4521 ":wasm_ukernels_test_mode",
Frank Barchardbcedc082020-08-17 18:00:51 -07004522 ":asm_ukernels",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004523 ],
4524 wasmsimd_deps = [
4525 ":wasm_ukernels_test_mode",
Frank Barchardbcedc082020-08-17 18:00:51 -07004526 ":asm_ukernels",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004527 ],
4528 x86_deps = [
4529 ":sse2_ukernels_test_mode",
4530 ":ssse3_ukernels_test_mode",
4531 ":sse41_ukernels_test_mode",
4532 ":avx_ukernels_test_mode",
Marat Dukhan1566fee2020-08-02 21:55:41 -07004533 ":xop_ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004534 ":fma3_ukernels_test_mode",
4535 ":avx2_ukernels_test_mode",
4536 ":avx512f_ukernels_test_mode",
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07004537 ":avx512skx_ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004538 ],
4539)
4540
Marat Dukhan08c4a432019-10-03 09:29:21 -07004541xnnpack_cc_library(
4542 name = "im2col",
4543 srcs = ["src/im2col.c"],
4544 hdrs = [
4545 "src/xnnpack/common.h",
4546 "src/xnnpack/im2col.h",
4547 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07004548 gcc_copts = xnnpack_gcc_std_copts(),
4549 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004550)
4551
4552xnnpack_cc_library(
4553 name = "indirection",
4554 srcs = ["src/indirection.c"],
4555 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004556 gcc_copts = xnnpack_gcc_std_copts(),
4557 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004558 deps = [
4559 "@FP16",
4560 "@FXdiv",
4561 "@pthreadpool",
4562 ],
4563)
4564
4565xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07004566 name = "indirection_test_mode",
4567 srcs = ["src/indirection.c"],
4568 hdrs = INTERNAL_HDRS,
4569 copts = [
4570 "-UNDEBUG",
4571 "-DXNN_TEST_MODE=1",
4572 ],
4573 gcc_copts = xnnpack_gcc_std_copts(),
4574 msvc_copts = xnnpack_msvc_std_copts(),
4575 deps = [
4576 "@FP16",
4577 "@FXdiv",
4578 "@pthreadpool",
4579 ],
4580)
4581
4582xnnpack_cc_library(
Marat Dukhanab582382020-07-06 13:32:08 -07004583 name = "packing",
4584 srcs = ["src/packing.c"],
4585 hdrs = INTERNAL_HDRS,
4586 gcc_copts = xnnpack_gcc_std_copts(),
4587 msvc_copts = xnnpack_msvc_std_copts(),
4588 deps = [
4589 "@FP16",
4590 "@FXdiv",
4591 "@pthreadpool",
4592 ],
4593)
4594
4595xnnpack_cc_library(
4596 name = "packing_test_mode",
4597 srcs = ["src/packing.c"],
4598 hdrs = INTERNAL_HDRS,
4599 copts = [
4600 "-UNDEBUG",
4601 "-DXNN_TEST_MODE=1",
4602 ],
4603 gcc_copts = xnnpack_gcc_std_copts(),
4604 msvc_copts = xnnpack_msvc_std_copts(),
4605 deps = [
4606 "@FP16",
4607 "@FXdiv",
4608 "@pthreadpool",
4609 ],
4610)
4611
4612xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004613 name = "operator_run",
4614 srcs = ["src/operator-run.c"],
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07004615 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004616 copts = LOGGING_COPTS + select({
Marat Dukhan05702cf2020-03-26 15:41:33 -07004617 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4618 "//conditions:default": [],
4619 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07004620 gcc_copts = xnnpack_gcc_std_copts(),
4621 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004622 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07004623 ":logging_utils",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004624 "@FP16",
4625 "@FXdiv",
4626 "@clog",
4627 "@pthreadpool",
4628 ],
4629)
4630
Chao Mei6ddfc602020-05-13 22:29:36 -07004631xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07004632 name = "operator_run_test_mode",
4633 srcs = ["src/operator-run.c"],
4634 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
4635 copts = LOGGING_COPTS + [
4636 "-UNDEBUG",
4637 "-DXNN_TEST_MODE=1",
4638 ] + select({
4639 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4640 "//conditions:default": [],
4641 }),
4642 gcc_copts = xnnpack_gcc_std_copts(),
4643 msvc_copts = xnnpack_msvc_std_copts(),
4644 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07004645 ":logging_utils",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004646 "@FP16",
4647 "@FXdiv",
4648 "@clog",
4649 "@pthreadpool",
4650 ],
4651)
4652
4653xnnpack_cc_library(
Chao Mei6ddfc602020-05-13 22:29:36 -07004654 name = "memory_planner",
4655 srcs = ["src/memory-planner.c"],
4656 hdrs = INTERNAL_HDRS,
4657 defines = select({
4658 ":xnn_enable_memopt_explicit_true": ["XNN_ENABLE_MEMOPT=1"],
4659 ":xnn_enable_memopt_explicit_false": ["XNN_ENABLE_MEMOPT=0"],
4660 "//conditions:default": ["XNN_ENABLE_MEMOPT=1"],
4661 }),
4662 gcc_copts = xnnpack_gcc_std_copts(),
4663 msvc_copts = xnnpack_msvc_std_copts(),
4664 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07004665 ":logging_utils",
Chao Mei6ddfc602020-05-13 22:29:36 -07004666 "@pthreadpool",
4667 ],
4668)
4669
Marat Dukhan33fcf782020-05-24 14:27:15 -07004670xnnpack_cc_library(
4671 name = "memory_planner_test_mode",
4672 srcs = ["src/memory-planner.c"],
4673 hdrs = INTERNAL_HDRS,
4674 copts = [
4675 "-UNDEBUG",
4676 "-DXNN_TEST_MODE=1",
4677 ],
4678 defines = select({
4679 ":xnn_enable_memopt_explicit_true": ["XNN_ENABLE_MEMOPT=1"],
4680 ":xnn_enable_memopt_explicit_false": ["XNN_ENABLE_MEMOPT=0"],
4681 "//conditions:default": ["XNN_ENABLE_MEMOPT=1"],
4682 }),
4683 gcc_copts = xnnpack_gcc_std_copts(),
4684 msvc_copts = xnnpack_msvc_std_copts(),
4685 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07004686 ":logging_utils",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004687 "@pthreadpool",
4688 ],
4689)
4690
Marat Dukhan08c4a432019-10-03 09:29:21 -07004691cc_library(
4692 name = "enable_assembly",
4693 defines = select({
4694 ":xnn_enable_assembly_explicit_true": ["XNN_ENABLE_ASSEMBLY=1"],
4695 ":xnn_enable_assembly_explicit_false": ["XNN_ENABLE_ASSEMBLY=0"],
Frank Barchard810171d2019-10-10 10:34:51 -07004696 "//conditions:default": ["XNN_ENABLE_ASSEMBLY=1"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004697 }),
4698)
4699
Marat Dukhan9de90e02020-06-18 16:04:12 -07004700cc_library(
4701 name = "enable_sparse",
4702 defines = select({
4703 ":xnn_enable_sparse_explicit_true": ["XNN_ENABLE_SPARSE=1"],
4704 ":xnn_enable_sparse_explicit_false": ["XNN_ENABLE_SPARSE=0"],
Marat Dukhanb36582b2020-12-08 11:16:28 -08004705 "//conditions:default": ["XNN_ENABLE_SPARSE=1"],
Marat Dukhan9de90e02020-06-18 16:04:12 -07004706 }),
4707)
4708
Marat Dukhancf056b22019-10-07 10:26:29 -07004709xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004710 name = "operators",
4711 srcs = OPERATOR_SRCS + [
Marat Dukhan496389f2021-04-07 15:47:12 -07004712 "src/allocator.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004713 "src/operator-delete.c",
Marat Dukhancf056b22019-10-07 10:26:29 -07004714 ],
4715 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004716 copts = LOGGING_COPTS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07004717 "-Isrc",
4718 "-Iinclude",
4719 ] + select({
4720 ":debug_build": [],
4721 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07004722 }) + select({
4723 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4724 "//conditions:default": [],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004725 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07004726 gcc_copts = xnnpack_gcc_std_copts(),
4727 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004728 deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07004729 ":indirection",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004730 ":logging_utils",
Marat Dukhanab582382020-07-06 13:32:08 -07004731 ":packing",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004732 "@FP16",
4733 "@FXdiv",
4734 "@clog",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004735 "@pthreadpool",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004736 ],
4737)
4738
Marat Dukhan10a38082020-04-17 03:58:35 -07004739xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07004740 name = "operators_test_mode",
4741 srcs = OPERATOR_SRCS + [
Marat Dukhan496389f2021-04-07 15:47:12 -07004742 "src/allocator.c",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004743 "src/operator-delete.c",
4744 ],
4745 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
4746 copts = LOGGING_COPTS + [
4747 "-Isrc",
4748 "-Iinclude",
4749 "-UNDEBUG",
4750 "-DXNN_TEST_MODE=1",
4751 ] + select({
4752 ":debug_build": [],
4753 "//conditions:default": xnnpack_min_size_copts(),
4754 }) + select({
4755 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4756 "//conditions:default": [],
4757 }),
4758 gcc_copts = xnnpack_gcc_std_copts(),
4759 msvc_copts = xnnpack_msvc_std_copts(),
4760 deps = [
4761 ":indirection_test_mode",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004762 ":logging_utils",
Marat Dukhanab582382020-07-06 13:32:08 -07004763 ":packing_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004764 "@FP16",
4765 "@FXdiv",
4766 "@clog",
4767 "@pthreadpool",
4768 ],
4769)
4770
4771xnnpack_cc_library(
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004772 name = "XNNPACK",
4773 srcs = [
4774 "src/init.c",
Marat Dukhanccfdbd12020-02-03 14:27:45 -08004775 "src/runtime.c",
4776 "src/subgraph.c",
4777 "src/tensor.c",
Marat Dukhanf03da0d2020-06-10 16:00:20 -07004778 ] + SUBGRAPH_SRCS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004779 hdrs = ["include/xnnpack.h"],
4780 copts = LOGGING_COPTS + [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004781 "-Isrc",
4782 "-Iinclude",
4783 ] + select({
4784 ":debug_build": [],
4785 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07004786 }) + select({
4787 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4788 "//conditions:default": [],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004789 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07004790 gcc_copts = xnnpack_gcc_std_copts(),
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004791 includes = ["include"],
Marat Dukhan10a38082020-04-17 03:58:35 -07004792 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004793 visibility = xnnpack_visibility(),
4794 deps = [
4795 ":enable_assembly",
Marat Dukhan9de90e02020-06-18 16:04:12 -07004796 ":enable_sparse",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004797 ":logging_utils",
Chao Mei6ddfc602020-05-13 22:29:36 -07004798 ":memory_planner",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004799 ":operator_run",
4800 ":operators",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004801 ":ukernels",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004802 "@clog",
Marat Dukhanab2946c2020-05-21 20:04:13 -07004803 "@FP16",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004804 "@pthreadpool",
Marat Dukhand343c222019-10-07 09:22:14 -07004805 ] + select({
4806 ":emscripten": [],
4807 "//conditions:default": ["@cpuinfo"],
4808 }),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004809)
4810
Marat Dukhan10a38082020-04-17 03:58:35 -07004811xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07004812 name = "XNNPACK_test_mode",
4813 srcs = [
4814 "src/init.c",
4815 "src/runtime.c",
4816 "src/subgraph.c",
4817 "src/tensor.c",
Marat Dukhanf03da0d2020-06-10 16:00:20 -07004818 ] + SUBGRAPH_SRCS,
Marat Dukhan33fcf782020-05-24 14:27:15 -07004819 hdrs = ["include/xnnpack.h"],
4820 copts = LOGGING_COPTS + [
4821 "-Isrc",
4822 "-Iinclude",
4823 "-UNDEBUG",
4824 "-DXNN_TEST_MODE=1",
4825 ] + select({
4826 ":debug_build": [],
4827 "//conditions:default": xnnpack_min_size_copts(),
4828 }) + select({
4829 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4830 "//conditions:default": [],
4831 }),
4832 gcc_copts = xnnpack_gcc_std_copts(),
4833 includes = ["include"],
4834 msvc_copts = xnnpack_msvc_std_copts(),
4835 visibility = xnnpack_visibility(),
4836 deps = [
4837 ":enable_assembly",
Marat Dukhan9de90e02020-06-18 16:04:12 -07004838 ":enable_sparse",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004839 ":logging_utils",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004840 ":memory_planner_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004841 ":operator_run_test_mode",
4842 ":operators_test_mode",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004843 ":ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004844 "@clog",
4845 "@FP16",
4846 "@pthreadpool",
4847 ] + select({
4848 ":emscripten": [],
4849 "//conditions:default": ["@cpuinfo"],
4850 }),
4851)
4852
Marat Dukhanf0cb70a2021-03-30 15:45:15 -07004853# Specialized XNNPACK version for TensorFlow Lite. Excludes operators currently
4854# not used by the TensorFlow Lite XNNPACK delegate to minimize code size.
Marat Dukhanae046f52020-06-15 13:16:14 -07004855xnnpack_cc_library(
Marat Dukhanf0cb70a2021-03-30 15:45:15 -07004856 name = "xnnpack_for_tflite",
4857 srcs = [
4858 "src/init.c",
4859 "src/runtime.c",
4860 "src/subgraph.c",
4861 "src/tensor.c",
4862 ] + SUBGRAPH_SRCS,
4863 hdrs = ["include/xnnpack.h"],
4864 copts = LOGGING_COPTS + [
4865 "-Isrc",
4866 "-Iinclude",
4867 ] + select({
4868 ":debug_build": [],
4869 "//conditions:default": xnnpack_min_size_copts(),
4870 }) + select({
4871 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4872 "//conditions:default": [],
4873 }),
4874 defines = [
Marat Dukhanf0cb70a2021-03-30 15:45:15 -07004875 "XNN_NO_QU8_OPERATORS",
4876 "XNN_NO_U8_OPERATORS",
4877 "XNN_NO_X8_OPERATORS",
4878 "XNN_NO_F16_OPERATORS",
4879 "XNN_NO_X16_OPERATORS",
Marat Dukhanb939cdb2021-03-30 18:51:51 -07004880 ] + select({
4881 ":xnn_enable_qs8_explicit_true": [],
4882 ":xnn_enable_qs8_explicit_false": ["XNN_NO_QS8_OPERATORS"],
4883 "//conditions:default": ["XNN_NO_QS8_OPERATORS"],
4884 }),
Marat Dukhanf0cb70a2021-03-30 15:45:15 -07004885 gcc_copts = xnnpack_gcc_std_copts(),
4886 includes = ["include"],
4887 msvc_copts = xnnpack_msvc_std_copts(),
4888 visibility = xnnpack_visibility(),
4889 deps = [
4890 ":enable_assembly",
4891 ":enable_sparse",
4892 ":logging_utils",
4893 ":memory_planner",
4894 ":operator_run",
4895 ":operators",
Marat Dukhand09ca262021-03-30 16:17:12 -07004896 ":ukernels",
Marat Dukhanf0cb70a2021-03-30 15:45:15 -07004897 "@clog",
4898 "@FP16",
4899 "@pthreadpool",
4900 ] + select({
4901 ":emscripten": [],
4902 "//conditions:default": ["@cpuinfo"],
4903 }),
4904)
4905
4906# Specialized XNNPACK version for TensorFlow.js. Excludes operators currently
4907# not used by the TensorFlow.js WebAssembly backend to minimize code size.
4908xnnpack_cc_library(
4909 name = "xnnpack_for_tfjs",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004910 srcs = [
4911 "src/init.c",
4912 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07004913 hdrs = ["include/xnnpack.h"],
4914 copts = LOGGING_COPTS + [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004915 "-Isrc",
4916 "-Iinclude",
4917 ] + select({
4918 ":debug_build": [],
4919 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07004920 }) + select({
4921 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4922 "//conditions:default": [],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004923 }),
4924 defines = [
Marat Dukhan16f1e1a2020-08-04 16:38:22 -07004925 "XNN_NO_QS8_OPERATORS",
Marat Dukhan08b7a972020-07-14 18:17:29 -07004926 "XNN_NO_QU8_OPERATORS",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004927 "XNN_NO_U8_OPERATORS",
4928 "XNN_NO_X8_OPERATORS",
Marat Dukhanefc47b82019-11-18 09:25:38 -08004929 "XNN_NO_NCHW_OPERATORS",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004930 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07004931 gcc_copts = xnnpack_gcc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004932 includes = ["include"],
Marat Dukhan10a38082020-04-17 03:58:35 -07004933 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004934 visibility = xnnpack_visibility(),
4935 deps = [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004936 ":enable_assembly",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004937 ":logging_utils",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004938 ":operator_run",
4939 ":operators",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004940 ":ukernels",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004941 "@clog",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004942 "@pthreadpool",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004943 ] + select({
4944 ":emscripten": [],
4945 "//conditions:default": ["@cpuinfo"],
4946 }),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004947)
4948
Marat Dukhancf056b22019-10-07 10:26:29 -07004949xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004950 name = "bench_utils",
4951 srcs = ["bench/utils.cc"],
4952 hdrs = ["bench/utils.h"],
Marat Dukhanbad48fe2019-11-04 10:35:22 -08004953 deps = [
4954 "@com_google_benchmark//:benchmark",
4955 "@cpuinfo",
4956 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004957)
4958
Frank Barchard7e955972019-10-11 10:34:25 -07004959######################### Benchmarks for micro-kernels #########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07004960
4961xnnpack_benchmark(
Marat Dukhan595e1702020-07-31 10:12:52 -07004962 name = "qs8_gemm_bench",
4963 srcs = [
4964 "bench/gemm.h",
4965 "bench/qs8-gemm.cc",
4966 "src/xnnpack/AlignedAllocator.h",
4967 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Frank Barchard31328cb2020-10-12 11:55:18 -07004968 copts = xnnpack_optional_ruy_copts() + xnnpack_optional_gemmlowp_copts(),
4969 deps = MICROKERNEL_BENCHMARK_DEPS + [":packing"] + xnnpack_optional_ruy_deps() + xnnpack_optional_gemmlowp_deps(),
Marat Dukhan595e1702020-07-31 10:12:52 -07004970)
4971
4972xnnpack_benchmark(
Marat Dukhan56bdd4a2020-08-03 19:47:04 -07004973 name = "qs8_requantization_bench",
4974 srcs = [
4975 "bench/qs8-requantization.cc",
Marat Dukhan56bdd4a2020-08-03 19:47:04 -07004976 "src/xnnpack/AlignedAllocator.h",
Frank Barchard04336c12020-10-22 16:48:55 -07004977 "src/xnnpack/requantization-stubs.h",
Marat Dukhan56bdd4a2020-08-03 19:47:04 -07004978 ] + MICROKERNEL_BENCHMARK_HDRS,
4979 deps = MICROKERNEL_BENCHMARK_DEPS,
4980)
4981
4982xnnpack_benchmark(
Marat Dukhan08b7a972020-07-14 18:17:29 -07004983 name = "qu8_gemm_bench",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004984 srcs = [
4985 "bench/gemm.h",
Marat Dukhan08b7a972020-07-14 18:17:29 -07004986 "bench/qu8-gemm.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004987 "src/xnnpack/AlignedAllocator.h",
4988 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004989 copts = xnnpack_optional_ruy_copts() + xnnpack_optional_gemmlowp_copts(),
Marat Dukhanab582382020-07-06 13:32:08 -07004990 deps = MICROKERNEL_BENCHMARK_DEPS + [":packing"] + xnnpack_optional_ruy_deps() + xnnpack_optional_gemmlowp_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004991)
4992
4993xnnpack_benchmark(
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07004994 name = "qu8_requantization_bench",
4995 srcs = [
4996 "bench/qu8-requantization.cc",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07004997 "src/xnnpack/AlignedAllocator.h",
Frank Barchard04336c12020-10-22 16:48:55 -07004998 "src/xnnpack/requantization-stubs.h",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07004999 ] + MICROKERNEL_BENCHMARK_HDRS,
5000 deps = MICROKERNEL_BENCHMARK_DEPS,
5001)
5002
5003xnnpack_benchmark(
Frank Barchard40d20fe2020-05-05 00:37:45 -07005004 name = "f16_igemm_bench",
5005 srcs = [
5006 "bench/f16-igemm.cc",
5007 "bench/conv.h",
5008 "bench/google/conv.h",
5009 "src/xnnpack/AlignedAllocator.h",
5010 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005011 deps = MICROKERNEL_BENCHMARK_DEPS + [
5012 ":indirection",
5013 ":packing",
5014 ],
Frank Barchard40d20fe2020-05-05 00:37:45 -07005015)
5016
5017xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005018 name = "f16_gemm_bench",
5019 srcs = [
5020 "bench/f16-gemm.cc",
5021 "bench/gemm.h",
5022 "src/xnnpack/AlignedAllocator.h",
5023 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005024 deps = MICROKERNEL_BENCHMARK_DEPS + [
5025 ":packing",
5026 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005027)
5028
5029xnnpack_benchmark(
Marat Dukhanbdb56f52020-02-05 21:42:49 -08005030 name = "f16_spmm_bench",
5031 srcs = [
5032 "bench/f16-spmm.cc",
Marat Dukhan1631e3e2020-12-06 19:29:31 -08005033 "bench/spmm.h",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08005034 "src/xnnpack/AlignedAllocator.h",
5035 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanbdb56f52020-02-05 21:42:49 -08005036 deps = MICROKERNEL_BENCHMARK_DEPS,
5037)
5038
5039xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005040 name = "f32_igemm_bench",
5041 srcs = [
5042 "bench/f32-igemm.cc",
5043 "bench/conv.h",
5044 "src/xnnpack/AlignedAllocator.h",
5045 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005046 deps = MICROKERNEL_BENCHMARK_DEPS + [
5047 ":indirection",
5048 ":packing",
5049 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005050)
5051
5052xnnpack_benchmark(
Frank Barchardc2010062020-07-19 10:04:12 -07005053 name = "f16_relu_bench",
5054 srcs = [
5055 "bench/f16-relu.cc",
5056 "src/xnnpack/AlignedAllocator.h",
5057 ] + MICROKERNEL_BENCHMARK_HDRS,
5058 deps = MICROKERNEL_BENCHMARK_DEPS,
5059)
5060
5061xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005062 name = "f32_conv_hwc_bench",
5063 srcs = [
5064 "bench/f32-conv-hwc.cc",
5065 "bench/dconv.h",
5066 "src/xnnpack/AlignedAllocator.h",
5067 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005068 deps = MICROKERNEL_BENCHMARK_DEPS + [
5069 ":packing",
5070 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005071)
5072
5073xnnpack_benchmark(
Marat Dukhan1f29b802020-05-15 23:46:39 -07005074 name = "f32_conv_hwc2chw_bench",
Erich Elsen563df5f2019-10-23 08:02:21 -07005075 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07005076 "bench/f32-conv-hwc2chw.cc",
Erich Elsen563df5f2019-10-23 08:02:21 -07005077 "bench/dconv.h",
5078 "src/xnnpack/AlignedAllocator.h",
5079 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005080 deps = MICROKERNEL_BENCHMARK_DEPS + [
5081 ":packing",
5082 ],
Erich Elsen563df5f2019-10-23 08:02:21 -07005083)
5084
5085xnnpack_benchmark(
Frank Barchard5a599a62020-06-04 20:12:44 -07005086 name = "f16_dwconv_bench",
5087 srcs = [
5088 "bench/f16-dwconv.cc",
5089 "bench/dwconv.h",
5090 "bench/google/dwconv.h",
5091 "src/xnnpack/AlignedAllocator.h",
5092 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005093 deps = MICROKERNEL_BENCHMARK_DEPS + [
5094 ":indirection",
5095 ":packing",
5096 ],
Frank Barchard5a599a62020-06-04 20:12:44 -07005097)
5098
5099xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005100 name = "f32_dwconv_bench",
5101 srcs = [
5102 "bench/f32-dwconv.cc",
5103 "bench/dwconv.h",
5104 "src/xnnpack/AlignedAllocator.h",
5105 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005106 deps = MICROKERNEL_BENCHMARK_DEPS + [
5107 ":indirection",
5108 ":packing",
5109 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005110)
5111
5112xnnpack_benchmark(
Marat Dukhanbf715f92020-10-23 20:17:00 -07005113 name = "f32_dwconv2d_chw_bench",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005114 srcs = [
Marat Dukhanbf715f92020-10-23 20:17:00 -07005115 "bench/f32-dwconv2d-chw.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005116 "bench/dwconv.h",
5117 "src/xnnpack/AlignedAllocator.h",
5118 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005119 deps = MICROKERNEL_BENCHMARK_DEPS + [
5120 ":indirection",
5121 ":packing",
5122 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005123)
5124
5125xnnpack_benchmark(
5126 name = "f32_gemm_bench",
5127 srcs = [
5128 "bench/f32-gemm.cc",
5129 "bench/gemm.h",
5130 "src/xnnpack/AlignedAllocator.h",
5131 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07005132 copts = xnnpack_optional_ruy_copts(),
Marat Dukhanab582382020-07-06 13:32:08 -07005133 deps = MICROKERNEL_BENCHMARK_DEPS + [":packing"] + xnnpack_optional_ruy_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07005134)
5135
5136xnnpack_benchmark(
Marat Dukhana11ca342020-06-25 23:45:07 -07005137 name = "f32_hswish_bench",
5138 srcs = [
5139 "bench/f32-hswish.cc",
5140 "src/xnnpack/AlignedAllocator.h",
5141 ] + MICROKERNEL_BENCHMARK_HDRS,
5142 deps = MICROKERNEL_BENCHMARK_DEPS,
5143)
5144
5145xnnpack_benchmark(
Marat Dukhan4c4eb002019-12-08 21:27:49 -08005146 name = "f32_raddexpminusmax_bench",
5147 srcs = [
5148 "bench/f32-raddexpminusmax.cc",
5149 "src/xnnpack/AlignedAllocator.h",
5150 ] + MICROKERNEL_BENCHMARK_HDRS,
5151 deps = MICROKERNEL_BENCHMARK_DEPS,
5152)
5153
5154xnnpack_benchmark(
5155 name = "f32_raddextexp_bench",
5156 srcs = [
5157 "bench/f32-raddextexp.cc",
5158 "src/xnnpack/AlignedAllocator.h",
5159 ] + MICROKERNEL_BENCHMARK_HDRS,
5160 deps = MICROKERNEL_BENCHMARK_DEPS,
5161)
5162
5163xnnpack_benchmark(
5164 name = "f32_raddstoreexpminusmax_bench",
5165 srcs = [
5166 "bench/f32-raddstoreexpminusmax.cc",
5167 "src/xnnpack/AlignedAllocator.h",
5168 ] + MICROKERNEL_BENCHMARK_HDRS,
5169 deps = MICROKERNEL_BENCHMARK_DEPS,
5170)
5171
5172xnnpack_benchmark(
Frank Barchard3c4a9522020-07-18 16:11:22 -07005173 name = "f32_relu_bench",
5174 srcs = [
5175 "bench/f32-relu.cc",
5176 "src/xnnpack/AlignedAllocator.h",
5177 ] + MICROKERNEL_BENCHMARK_HDRS,
5178 deps = MICROKERNEL_BENCHMARK_DEPS,
5179)
5180
5181xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005182 name = "f32_rmax_bench",
5183 srcs = [
5184 "bench/f32-rmax.cc",
5185 "src/xnnpack/AlignedAllocator.h",
5186 ] + MICROKERNEL_BENCHMARK_HDRS,
5187 deps = MICROKERNEL_BENCHMARK_DEPS,
5188)
5189
5190xnnpack_benchmark(
Marat Dukhan14bec502019-11-18 11:35:31 -08005191 name = "f32_sigmoid_bench",
5192 srcs = [
5193 "bench/f32-sigmoid.cc",
5194 "src/xnnpack/AlignedAllocator.h",
5195 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan14bec502019-11-18 11:35:31 -08005196 deps = MICROKERNEL_BENCHMARK_DEPS,
5197)
5198
5199xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005200 name = "f32_spmm_bench",
5201 srcs = [
5202 "bench/f32-spmm.cc",
Marat Dukhan1631e3e2020-12-06 19:29:31 -08005203 "bench/spmm.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005204 "src/xnnpack/AlignedAllocator.h",
5205 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005206 deps = MICROKERNEL_BENCHMARK_DEPS,
5207)
5208
5209xnnpack_benchmark(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08005210 name = "f32_softmax_bench",
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07005211 srcs = [
Marat Dukhanfd8e6892020-01-27 15:25:25 -08005212 "bench/f32-softmax.cc",
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07005213 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07005214 copts = xnnpack_optional_dnnl_copts(),
Marat Dukhan8d3c6932020-03-06 20:27:27 -08005215 deps = MICROKERNEL_BENCHMARK_DEPS + xnnpack_optional_dnnl_deps(),
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07005216)
5217
5218xnnpack_benchmark(
Marat Dukhaned6baaf2020-12-01 15:07:08 -08005219 name = "f32_velu_bench",
5220 srcs = [
5221 "bench/f32-velu.cc",
5222 "src/xnnpack/AlignedAllocator.h",
5223 ] + MICROKERNEL_BENCHMARK_HDRS,
5224 deps = MICROKERNEL_BENCHMARK_DEPS,
5225)
5226
5227xnnpack_benchmark(
Marat Dukhan4c4eb002019-12-08 21:27:49 -08005228 name = "f32_vscaleexpminusmax_bench",
5229 srcs = [
5230 "bench/f32-vscaleexpminusmax.cc",
5231 "src/xnnpack/AlignedAllocator.h",
5232 ] + MICROKERNEL_BENCHMARK_HDRS,
5233 deps = MICROKERNEL_BENCHMARK_DEPS,
5234)
5235
5236xnnpack_benchmark(
5237 name = "f32_vscaleextexp_bench",
5238 srcs = [
5239 "bench/f32-vscaleextexp.cc",
5240 "src/xnnpack/AlignedAllocator.h",
5241 ] + MICROKERNEL_BENCHMARK_HDRS,
5242 deps = MICROKERNEL_BENCHMARK_DEPS,
5243)
5244
5245xnnpack_benchmark(
Marat Dukhanf4db2f32020-06-30 10:55:30 -07005246 name = "f32_vsqrt_bench",
5247 srcs = [
5248 "bench/f32-vsqrt.cc",
5249 "src/xnnpack/AlignedAllocator.h",
5250 ] + MICROKERNEL_BENCHMARK_HDRS,
5251 deps = MICROKERNEL_BENCHMARK_DEPS,
5252)
5253
5254xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005255 name = "f32_im2col_gemm_bench",
5256 srcs = [
5257 "bench/f32-im2col-gemm.cc",
5258 "bench/conv.h",
5259 "src/xnnpack/AlignedAllocator.h",
5260 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005261 deps = MICROKERNEL_BENCHMARK_DEPS + [
5262 ":im2col",
5263 ":packing",
5264 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005265)
5266
Marat Dukhanfe7acb62020-03-09 19:30:05 -07005267xnnpack_benchmark(
Marat Dukhanffbf96a2020-05-14 02:59:08 -07005268 name = "rounding_bench",
5269 srcs = [
5270 "bench/rounding.cc",
Marat Dukhanffbf96a2020-05-14 02:59:08 -07005271 "src/xnnpack/AlignedAllocator.h",
Frank Barchard04336c12020-10-22 16:48:55 -07005272 "src/xnnpack/math-stubs.h",
Marat Dukhanffbf96a2020-05-14 02:59:08 -07005273 ] + MICROKERNEL_BENCHMARK_HDRS,
5274 deps = MICROKERNEL_BENCHMARK_DEPS,
5275)
5276
Marat Dukhan08c4a432019-10-03 09:29:21 -07005277########################### Benchmarks for operators ###########################
5278
5279xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005280 name = "average_pooling_bench",
5281 srcs = ["bench/average-pooling.cc"],
Marat Dukhan7a16d8b2020-03-11 04:22:44 -07005282 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07005283 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07005284 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07005285)
5286
5287xnnpack_benchmark(
Marat Dukhan87727142020-06-24 15:24:10 -07005288 name = "bankers_rounding_bench",
5289 srcs = ["bench/bankers-rounding.cc"],
5290 copts = xnnpack_optional_tflite_copts(),
5291 tags = ["nowin32"],
5292 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
5293)
5294
5295xnnpack_benchmark(
5296 name = "ceiling_bench",
5297 srcs = ["bench/ceiling.cc"],
5298 copts = xnnpack_optional_tflite_copts(),
5299 tags = ["nowin32"],
5300 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
5301)
5302
5303xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005304 name = "channel_shuffle_bench",
5305 srcs = ["bench/channel-shuffle.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07005306 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005307)
5308
5309xnnpack_benchmark(
5310 name = "convolution_bench",
5311 srcs = ["bench/convolution.cc"],
5312 copts = xnnpack_optional_tflite_copts() + xnnpack_optional_armcl_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07005313 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07005314 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps() + xnnpack_optional_armcl_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07005315)
5316
5317xnnpack_benchmark(
5318 name = "deconvolution_bench",
5319 srcs = ["bench/deconvolution.cc"],
5320 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07005321 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07005322 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07005323)
5324
5325xnnpack_benchmark(
Marat Dukhanb6bd4bc2020-12-01 17:01:40 -08005326 name = "elu_bench",
5327 srcs = ["bench/elu.cc"],
5328 copts = xnnpack_optional_tflite_copts(),
5329 tags = ["nowin32"],
5330 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
5331)
5332
5333xnnpack_benchmark(
Marat Dukhan87727142020-06-24 15:24:10 -07005334 name = "floor_bench",
5335 srcs = ["bench/floor.cc"],
5336 copts = xnnpack_optional_tflite_copts(),
5337 tags = ["nowin32"],
5338 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
5339)
5340
5341xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005342 name = "global_average_pooling_bench",
5343 srcs = ["bench/global-average-pooling.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07005344 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005345)
5346
5347xnnpack_benchmark(
Marat Dukhanad352602020-06-25 21:50:54 -07005348 name = "hardswish_bench",
5349 srcs = ["bench/hardswish.cc"],
5350 copts = xnnpack_optional_tflite_copts(),
5351 tags = ["nowin32"],
5352 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
5353)
5354
5355xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005356 name = "max_pooling_bench",
5357 srcs = ["bench/max-pooling.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07005358 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005359)
5360
5361xnnpack_benchmark(
5362 name = "sigmoid_bench",
5363 srcs = ["bench/sigmoid.cc"],
Marat Dukhanc3b9e862019-11-17 13:18:54 -08005364 copts = xnnpack_optional_tflite_copts(),
Marat Dukhanca2ba702020-04-24 01:31:47 -07005365 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07005366 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07005367)
5368
5369xnnpack_benchmark(
Marat Dukhan95b22432019-10-30 16:30:14 -07005370 name = "prelu_bench",
5371 srcs = ["bench/prelu.cc"],
5372 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07005373 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07005374 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan95b22432019-10-30 16:30:14 -07005375)
5376
5377xnnpack_benchmark(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08005378 name = "softmax_bench",
5379 srcs = ["bench/softmax.cc"],
Marat Dukhan9c0db962020-01-28 12:30:14 -08005380 copts = xnnpack_optional_tflite_copts(),
Marat Dukhanca2ba702020-04-24 01:31:47 -07005381 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07005382 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07005383)
5384
Marat Dukhan87727142020-06-24 15:24:10 -07005385xnnpack_benchmark(
Marat Dukhan6804bbd2020-06-30 19:26:11 -07005386 name = "square_root_bench",
5387 srcs = ["bench/square-root.cc"],
5388 copts = xnnpack_optional_tflite_copts(),
5389 tags = ["nowin32"],
5390 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
5391)
5392
5393xnnpack_benchmark(
Marat Dukhan87727142020-06-24 15:24:10 -07005394 name = "truncation_bench",
5395 srcs = ["bench/truncation.cc"],
5396 deps = OPERATOR_BENCHMARK_DEPS,
5397)
5398
Marat Dukhanc068bb62019-10-04 13:24:39 -07005399############################# End-to-end benchmarks ############################
5400
5401cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07005402 name = "fp32_mobilenet_v1",
5403 srcs = ["models/fp32-mobilenet-v1.cc"],
Marat Dukhanc068bb62019-10-04 13:24:39 -07005404 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08005405 copts = xnnpack_std_cxxopts(),
Marat Dukhanc068bb62019-10-04 13:24:39 -07005406 linkstatic = True,
5407 deps = [
5408 ":XNNPACK",
5409 "@pthreadpool",
5410 ],
5411)
5412
5413cc_library(
Marat Dukhan4cea2322021-03-09 09:35:36 -08005414 name = "fp32_sparse_mobilenet_v1",
5415 srcs = ["models/fp32-sparse-mobilenet-v1.cc"],
5416 hdrs = ["models/models.h"],
5417 copts = xnnpack_std_cxxopts(),
5418 linkstatic = True,
5419 deps = [
5420 ":XNNPACK",
5421 "@pthreadpool",
5422 ],
5423)
5424
5425cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07005426 name = "fp16_mobilenet_v1",
5427 srcs = ["models/fp16-mobilenet-v1.cc"],
5428 hdrs = ["models/models.h"],
5429 copts = xnnpack_std_cxxopts(),
5430 linkstatic = True,
5431 deps = [
5432 ":XNNPACK",
5433 "@FP16",
5434 "@pthreadpool",
5435 ],
5436)
5437
5438cc_library(
Marat Dukhan0743cdf2020-08-04 18:52:07 -07005439 name = "qs8_mobilenet_v1",
5440 srcs = ["models/qs8-mobilenet-v1.cc"],
5441 hdrs = ["models/models.h"],
5442 copts = xnnpack_std_cxxopts(),
5443 linkstatic = True,
5444 deps = [
5445 ":XNNPACK",
5446 "@pthreadpool",
5447 ],
5448)
5449
5450cc_library(
Marat Dukhan70a96182020-09-03 17:13:58 -07005451 name = "qs8_mobilenet_v2",
5452 srcs = ["models/qs8-mobilenet-v2.cc"],
5453 hdrs = ["models/models.h"],
5454 copts = xnnpack_std_cxxopts(),
5455 linkstatic = True,
5456 deps = [
5457 ":XNNPACK",
5458 "@pthreadpool",
5459 ],
5460)
5461
5462cc_library(
Marat Dukhan12a23bb2021-03-08 08:13:21 -08005463 name = "qu8_mobilenet_v1",
5464 srcs = ["models/qu8-mobilenet-v1.cc"],
5465 hdrs = ["models/models.h"],
5466 copts = xnnpack_std_cxxopts(),
5467 linkstatic = True,
5468 deps = [
5469 ":XNNPACK",
5470 "@pthreadpool",
5471 ],
5472)
5473
5474cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07005475 name = "fp32_mobilenet_v2",
5476 srcs = ["models/fp32-mobilenet-v2.cc"],
Marat Dukhanc068bb62019-10-04 13:24:39 -07005477 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08005478 copts = xnnpack_std_cxxopts(),
Marat Dukhanc068bb62019-10-04 13:24:39 -07005479 linkstatic = True,
5480 deps = [
5481 ":XNNPACK",
5482 "@pthreadpool",
5483 ],
5484)
5485
Marat Dukhanc08cdf52019-12-09 09:17:51 -08005486cc_library(
Marat Dukhan4cea2322021-03-09 09:35:36 -08005487 name = "fp32_sparse_mobilenet_v2",
5488 srcs = ["models/fp32-sparse-mobilenet-v2.cc"],
5489 hdrs = ["models/models.h"],
5490 copts = xnnpack_std_cxxopts(),
5491 linkstatic = True,
5492 deps = [
5493 ":XNNPACK",
5494 "@pthreadpool",
5495 ],
5496)
5497
5498cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07005499 name = "fp16_mobilenet_v2",
5500 srcs = ["models/fp16-mobilenet-v2.cc"],
5501 hdrs = ["models/models.h"],
5502 copts = xnnpack_std_cxxopts(),
5503 linkstatic = True,
5504 deps = [
5505 ":XNNPACK",
5506 "@FP16",
5507 "@pthreadpool",
5508 ],
5509)
5510
5511cc_library(
5512 name = "fp32_mobilenet_v3_large",
5513 srcs = ["models/fp32-mobilenet-v3-large.cc"],
Marat Dukhanc08cdf52019-12-09 09:17:51 -08005514 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08005515 copts = xnnpack_std_cxxopts(),
Marat Dukhanc08cdf52019-12-09 09:17:51 -08005516 linkstatic = True,
5517 deps = [
5518 ":XNNPACK",
5519 "@pthreadpool",
5520 ],
5521)
5522
5523cc_library(
Marat Dukhan4cea2322021-03-09 09:35:36 -08005524 name = "fp32_sparse_mobilenet_v3_large",
5525 srcs = ["models/fp32-sparse-mobilenet-v3-large.cc"],
5526 hdrs = ["models/models.h"],
5527 copts = xnnpack_std_cxxopts(),
5528 linkstatic = True,
5529 deps = [
5530 ":XNNPACK",
5531 "@pthreadpool",
5532 ],
5533)
5534
5535cc_library(
Marat Dukhan66f3ccd2020-09-14 16:09:38 -07005536 name = "fp16_mobilenet_v3_large",
5537 srcs = ["models/fp16-mobilenet-v3-large.cc"],
5538 hdrs = ["models/models.h"],
5539 copts = xnnpack_std_cxxopts(),
5540 linkstatic = True,
5541 deps = [
5542 ":XNNPACK",
5543 "@FP16",
5544 "@pthreadpool",
5545 ],
5546)
5547
5548cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07005549 name = "fp32_mobilenet_v3_small",
5550 srcs = ["models/fp32-mobilenet-v3-small.cc"],
Marat Dukhanc08cdf52019-12-09 09:17:51 -08005551 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08005552 copts = xnnpack_std_cxxopts(),
Marat Dukhanc08cdf52019-12-09 09:17:51 -08005553 linkstatic = True,
5554 deps = [
5555 ":XNNPACK",
5556 "@pthreadpool",
5557 ],
5558)
5559
Marat Dukhan66f3ccd2020-09-14 16:09:38 -07005560cc_library(
Marat Dukhan4cea2322021-03-09 09:35:36 -08005561 name = "fp32_sparse_mobilenet_v3_small",
5562 srcs = ["models/fp32-sparse-mobilenet-v3-small.cc"],
5563 hdrs = ["models/models.h"],
5564 copts = xnnpack_std_cxxopts(),
5565 linkstatic = True,
5566 deps = [
5567 ":XNNPACK",
5568 "@pthreadpool",
5569 ],
5570)
5571
5572cc_library(
Marat Dukhan66f3ccd2020-09-14 16:09:38 -07005573 name = "fp16_mobilenet_v3_small",
5574 srcs = ["models/fp16-mobilenet-v3-small.cc"],
5575 hdrs = ["models/models.h"],
5576 copts = xnnpack_std_cxxopts(),
5577 linkstatic = True,
5578 deps = [
5579 ":XNNPACK",
5580 "@FP16",
5581 "@pthreadpool",
5582 ],
5583)
5584
Marat Dukhanc068bb62019-10-04 13:24:39 -07005585xnnpack_benchmark(
Marat Dukhanef4416e2019-10-31 13:44:40 -07005586 name = "f32_dwconv_e2e_bench",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08005587 srcs = [
5588 "bench/f32-dwconv-e2e.cc",
5589 "bench/end2end.h",
5590 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanef4416e2019-10-31 13:44:40 -07005591 deps = MICROKERNEL_BENCHMARK_DEPS + [
5592 ":XNNPACK",
Marat Dukhan270a2c42020-06-26 16:45:52 -07005593 ":fp32_mobilenet_v1",
5594 ":fp32_mobilenet_v2",
5595 ":fp32_mobilenet_v3_large",
5596 ":fp32_mobilenet_v3_small",
Marat Dukhanef4416e2019-10-31 13:44:40 -07005597 ],
5598)
5599
5600xnnpack_benchmark(
Marat Dukhan5f18d262019-10-31 10:24:14 -07005601 name = "f32_gemm_e2e_bench",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08005602 srcs = [
5603 "bench/f32-gemm-e2e.cc",
5604 "bench/end2end.h",
5605 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan5f18d262019-10-31 10:24:14 -07005606 deps = MICROKERNEL_BENCHMARK_DEPS + [
5607 ":XNNPACK",
Marat Dukhan270a2c42020-06-26 16:45:52 -07005608 ":fp32_mobilenet_v1",
5609 ":fp32_mobilenet_v2",
5610 ":fp32_mobilenet_v3_large",
5611 ":fp32_mobilenet_v3_small",
Marat Dukhan5f18d262019-10-31 10:24:14 -07005612 ],
5613)
5614
5615xnnpack_benchmark(
Frank Barcharddc909cb2021-02-08 13:59:31 -08005616 name = "qs8_gemm_e2e_bench",
5617 srcs = [
5618 "bench/qs8-gemm-e2e.cc",
5619 "bench/end2end.h",
5620 ] + MICROKERNEL_BENCHMARK_HDRS,
5621 deps = MICROKERNEL_BENCHMARK_DEPS + [
5622 ":XNNPACK",
5623 ":qs8_mobilenet_v1",
5624 ":qs8_mobilenet_v2",
5625 ],
5626)
5627
5628xnnpack_benchmark(
Marat Dukhanc068bb62019-10-04 13:24:39 -07005629 name = "end2end_bench",
5630 srcs = ["bench/end2end.cc"],
5631 deps = [
5632 ":XNNPACK",
Frank Barchardc712fa42019-10-31 14:00:21 -07005633 ":bench_utils",
Marat Dukhan270a2c42020-06-26 16:45:52 -07005634 ":fp16_mobilenet_v1",
5635 ":fp16_mobilenet_v2",
Marat Dukhan66f3ccd2020-09-14 16:09:38 -07005636 ":fp16_mobilenet_v3_large",
5637 ":fp16_mobilenet_v3_small",
Marat Dukhan270a2c42020-06-26 16:45:52 -07005638 ":fp32_mobilenet_v1",
5639 ":fp32_mobilenet_v2",
5640 ":fp32_mobilenet_v3_large",
5641 ":fp32_mobilenet_v3_small",
Marat Dukhan4cea2322021-03-09 09:35:36 -08005642 ":fp32_sparse_mobilenet_v1",
5643 ":fp32_sparse_mobilenet_v2",
5644 ":fp32_sparse_mobilenet_v3_large",
5645 ":fp32_sparse_mobilenet_v3_small",
Marat Dukhan0743cdf2020-08-04 18:52:07 -07005646 ":qs8_mobilenet_v1",
Marat Dukhan70a96182020-09-03 17:13:58 -07005647 ":qs8_mobilenet_v2",
Marat Dukhan12a23bb2021-03-08 08:13:21 -08005648 ":qu8_mobilenet_v1",
Marat Dukhanc068bb62019-10-04 13:24:39 -07005649 "@pthreadpool",
5650 ],
5651)
5652
Marat Dukhan6adff4e2019-10-14 18:32:07 -07005653#################### Accuracy evaluation for math functions ####################
5654
5655xnnpack_benchmark(
Marat Dukhancf67ec62020-12-14 09:56:37 -08005656 name = "f32_exp_ulp_eval",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07005657 srcs = [
Marat Dukhancf67ec62020-12-14 09:56:37 -08005658 "eval/f32-exp-ulp.cc",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07005659 "src/xnnpack/AlignedAllocator.h",
5660 ] + ACCURACY_EVAL_HDRS,
Marat Dukhan3a305212020-12-06 19:24:27 -08005661 deps = ACCURACY_EVAL_DEPS + [
5662 ":bench_utils",
5663 "@cpuinfo",
5664 ],
Marat Dukhan6adff4e2019-10-14 18:32:07 -07005665)
5666
Marat Dukhan515c9772019-10-17 18:07:57 -07005667xnnpack_benchmark(
Marat Dukhancf67ec62020-12-14 09:56:37 -08005668 name = "f32_expminus_ulp_eval",
Marat Dukhan515c9772019-10-17 18:07:57 -07005669 srcs = [
Marat Dukhancf67ec62020-12-14 09:56:37 -08005670 "eval/f32-expminus-ulp.cc",
Marat Dukhan515c9772019-10-17 18:07:57 -07005671 "src/xnnpack/AlignedAllocator.h",
5672 ] + ACCURACY_EVAL_HDRS,
Marat Dukhan3a305212020-12-06 19:24:27 -08005673 deps = ACCURACY_EVAL_DEPS + [
5674 ":bench_utils",
5675 "@cpuinfo",
5676 ],
Marat Dukhan515c9772019-10-17 18:07:57 -07005677)
5678
Marat Dukhan98ba4412019-10-23 02:14:28 -07005679xnnpack_benchmark(
Marat Dukhancf67ec62020-12-14 09:56:37 -08005680 name = "f32_expm1minus_ulp_eval",
Marat Dukhana438aca2020-11-20 15:45:01 -08005681 srcs = [
Marat Dukhancf67ec62020-12-14 09:56:37 -08005682 "eval/f32-expm1minus-ulp.cc",
Marat Dukhana438aca2020-11-20 15:45:01 -08005683 "src/xnnpack/AlignedAllocator.h",
5684 ] + ACCURACY_EVAL_HDRS,
Marat Dukhande390d42020-11-29 19:32:18 -08005685 deps = ACCURACY_EVAL_DEPS + [
5686 ":bench_utils",
Marat Dukhan3a305212020-12-06 19:24:27 -08005687 "@cpuinfo",
Marat Dukhande390d42020-11-29 19:32:18 -08005688 ],
Marat Dukhana438aca2020-11-20 15:45:01 -08005689)
5690
5691xnnpack_benchmark(
Marat Dukhancf67ec62020-12-14 09:56:37 -08005692 name = "f32_extexp_ulp_eval",
Marat Dukhan98ba4412019-10-23 02:14:28 -07005693 srcs = [
Marat Dukhancf67ec62020-12-14 09:56:37 -08005694 "eval/f32-extexp-ulp.cc",
Marat Dukhan98ba4412019-10-23 02:14:28 -07005695 "src/xnnpack/AlignedAllocator.h",
5696 ] + ACCURACY_EVAL_HDRS,
Marat Dukhan3a305212020-12-06 19:24:27 -08005697 deps = ACCURACY_EVAL_DEPS + [
5698 ":bench_utils",
5699 "@cpuinfo",
5700 ],
Marat Dukhan98ba4412019-10-23 02:14:28 -07005701)
5702
Marat Dukhanf44f0222020-12-14 11:53:27 -08005703xnnpack_benchmark(
5704 name = "f32_sigmoid_ulp_eval",
5705 srcs = [
5706 "eval/f32-sigmoid-ulp.cc",
5707 "src/xnnpack/AlignedAllocator.h",
5708 ] + ACCURACY_EVAL_HDRS,
5709 deps = ACCURACY_EVAL_DEPS + [
5710 ":bench_utils",
5711 "@cpuinfo",
5712 ],
5713)
5714
5715xnnpack_benchmark(
5716 name = "f32_sqrt_ulp_eval",
5717 srcs = [
5718 "eval/f32-sqrt-ulp.cc",
5719 "src/xnnpack/AlignedAllocator.h",
5720 ] + ACCURACY_EVAL_HDRS,
5721 deps = ACCURACY_EVAL_DEPS + [
5722 ":bench_utils",
5723 "@cpuinfo",
5724 ],
5725)
5726
5727################### Accuracy verification for math functions ##################
5728
5729xnnpack_unit_test(
Marat Dukhanf7291fc2020-12-15 11:02:50 -08005730 name = "f32_exp_eval",
5731 srcs = [
5732 "eval/f32-exp.cc",
5733 "src/xnnpack/AlignedAllocator.h",
5734 "src/xnnpack/math-stubs.h",
5735 ] + MICROKERNEL_TEST_HDRS,
5736 automatic = False,
5737 deps = MICROKERNEL_TEST_DEPS,
5738)
5739
5740xnnpack_unit_test(
Marat Dukhanf44f0222020-12-14 11:53:27 -08005741 name = "f32_expm1minus_eval",
5742 srcs = [
5743 "eval/f32-expm1minus.cc",
5744 "src/xnnpack/AlignedAllocator.h",
5745 "src/xnnpack/math-stubs.h",
5746 ] + MICROKERNEL_TEST_HDRS,
5747 automatic = False,
5748 deps = MICROKERNEL_TEST_DEPS,
5749)
5750
Marat Dukhan8853b822020-05-07 12:19:01 -07005751xnnpack_unit_test(
Marat Dukhand28a5a22020-12-14 15:27:22 -08005752 name = "f32_expminus_eval",
5753 srcs = [
5754 "eval/f32-expminus.cc",
5755 "src/xnnpack/AlignedAllocator.h",
5756 "src/xnnpack/math-stubs.h",
5757 ] + MICROKERNEL_TEST_HDRS,
5758 automatic = False,
5759 deps = MICROKERNEL_TEST_DEPS,
5760)
5761
5762xnnpack_unit_test(
Marat Dukhan8853b822020-05-07 12:19:01 -07005763 name = "f32_roundne_eval",
5764 srcs = [
5765 "eval/f32-roundne.cc",
5766 "src/xnnpack/AlignedAllocator.h",
5767 "src/xnnpack/math-stubs.h",
5768 ] + MICROKERNEL_TEST_HDRS,
Marat Dukhan22eed3d2020-05-11 20:13:37 -07005769 automatic = False,
Marat Dukhan8853b822020-05-07 12:19:01 -07005770 deps = MICROKERNEL_TEST_DEPS,
5771)
5772
Marat Dukhan2dbb9442020-05-12 20:43:43 -07005773xnnpack_unit_test(
Marat Dukhanc9852ba2020-05-13 17:21:29 -07005774 name = "f32_roundd_eval",
5775 srcs = [
5776 "eval/f32-roundd.cc",
5777 "src/xnnpack/AlignedAllocator.h",
5778 "src/xnnpack/math-stubs.h",
5779 ] + MICROKERNEL_TEST_HDRS,
5780 automatic = False,
5781 deps = MICROKERNEL_TEST_DEPS,
5782)
5783
5784xnnpack_unit_test(
5785 name = "f32_roundu_eval",
5786 srcs = [
5787 "eval/f32-roundu.cc",
5788 "src/xnnpack/AlignedAllocator.h",
5789 "src/xnnpack/math-stubs.h",
5790 ] + MICROKERNEL_TEST_HDRS,
5791 automatic = False,
5792 deps = MICROKERNEL_TEST_DEPS,
5793)
5794
5795xnnpack_unit_test(
Marat Dukhan2dbb9442020-05-12 20:43:43 -07005796 name = "f32_roundz_eval",
5797 srcs = [
5798 "eval/f32-roundz.cc",
5799 "src/xnnpack/AlignedAllocator.h",
5800 "src/xnnpack/math-stubs.h",
5801 ] + MICROKERNEL_TEST_HDRS,
Marat Dukhanc9852ba2020-05-13 17:21:29 -07005802 automatic = False,
Marat Dukhan2dbb9442020-05-12 20:43:43 -07005803 deps = MICROKERNEL_TEST_DEPS,
5804)
5805
Marat Dukhan08c4a432019-10-03 09:29:21 -07005806######################### Unit tests for micro-kernels #########################
5807
5808xnnpack_unit_test(
Marat Dukhande06f492020-04-09 00:19:31 -07005809 name = "f16_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005810 srcs = [
Marat Dukhande06f492020-04-09 00:19:31 -07005811 "test/f16-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005812 "test/gemm-microkernel-tester.h",
5813 "src/xnnpack/AlignedAllocator.h",
5814 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005815 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005816)
5817
5818xnnpack_unit_test(
Marat Dukhan355ab432020-04-09 19:01:52 -07005819 name = "f16_spmm_minmax_test",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08005820 srcs = [
Marat Dukhan355ab432020-04-09 19:01:52 -07005821 "test/f16-spmm-minmax.cc",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08005822 "test/spmm-microkernel-tester.h",
5823 "src/xnnpack/AlignedAllocator.h",
5824 ] + MICROKERNEL_TEST_HDRS,
5825 deps = MICROKERNEL_TEST_DEPS,
5826)
5827
5828xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005829 name = "f32_argmaxpool_test",
5830 srcs = [
5831 "test/f32-argmaxpool.cc",
5832 "test/argmaxpool-microkernel-tester.h",
5833 "src/xnnpack/AlignedAllocator.h",
5834 ] + MICROKERNEL_TEST_HDRS,
5835 deps = MICROKERNEL_TEST_DEPS,
5836)
5837
5838xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07005839 name = "f32_avgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005840 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07005841 "test/f32-avgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005842 "test/avgpool-microkernel-tester.h",
5843 "src/xnnpack/AlignedAllocator.h",
5844 ] + MICROKERNEL_TEST_HDRS,
5845 deps = MICROKERNEL_TEST_DEPS,
5846)
5847
5848xnnpack_unit_test(
Marat Dukhan660fd192020-03-10 04:55:30 -07005849 name = "f32_ibilinear_test",
Marat Dukhan35dacfb2019-11-07 19:18:16 -08005850 srcs = [
Marat Dukhan660fd192020-03-10 04:55:30 -07005851 "test/f32-ibilinear.cc",
5852 "test/ibilinear-microkernel-tester.h",
Marat Dukhan35dacfb2019-11-07 19:18:16 -08005853 "src/xnnpack/AlignedAllocator.h",
5854 ] + MICROKERNEL_TEST_HDRS,
5855 deps = MICROKERNEL_TEST_DEPS,
5856)
5857
5858xnnpack_unit_test(
XNNPACK Team21432672020-10-19 19:58:48 -07005859 name = "f32_ibilinear_chw_test",
5860 srcs = [
5861 "test/f32-ibilinear-chw.cc",
XNNPACK Team6be46b22020-10-22 23:34:54 -07005862 "test/ibilinear-microkernel-tester.h",
XNNPACK Team21432672020-10-19 19:58:48 -07005863 "src/xnnpack/AlignedAllocator.h",
5864 ] + MICROKERNEL_TEST_HDRS,
5865 deps = MICROKERNEL_TEST_DEPS,
5866)
5867
5868xnnpack_unit_test(
Frank Barchardb1966592020-05-12 13:47:06 -07005869 name = "f16_clamp_test",
5870 srcs = [
5871 "test/f16-clamp.cc",
5872 "test/clamp-microkernel-tester.h",
5873 ] + MICROKERNEL_TEST_HDRS,
5874 deps = MICROKERNEL_TEST_DEPS,
5875)
5876
5877xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005878 name = "f32_clamp_test",
5879 srcs = [
5880 "test/f32-clamp.cc",
5881 "test/clamp-microkernel-tester.h",
5882 ] + MICROKERNEL_TEST_HDRS,
5883 deps = MICROKERNEL_TEST_DEPS,
5884)
5885
5886xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07005887 name = "f32_igemm_test",
5888 srcs = [
5889 "test/f32-igemm.cc",
5890 "test/gemm-microkernel-tester.h",
5891 "src/xnnpack/AlignedAllocator.h",
5892 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005893 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan163a7e62020-04-09 04:19:26 -07005894)
5895
5896xnnpack_unit_test(
Marat Dukhan467f6362020-05-22 23:21:55 -07005897 name = "f32_igemm_relu_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005898 srcs = [
Marat Dukhan467f6362020-05-22 23:21:55 -07005899 "test/f32-igemm-relu.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005900 "test/gemm-microkernel-tester.h",
5901 "src/xnnpack/AlignedAllocator.h",
5902 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005903 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005904)
5905
5906xnnpack_unit_test(
Marat Dukhane207b7b2020-05-28 16:27:42 -07005907 name = "f32_igemm_minmax_test",
5908 srcs = [
5909 "test/f32-igemm-minmax.cc",
5910 "test/gemm-microkernel-tester.h",
5911 "src/xnnpack/AlignedAllocator.h",
5912 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005913 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhane207b7b2020-05-28 16:27:42 -07005914)
5915
5916xnnpack_unit_test(
Frank Barchardb0e4fae2020-05-04 15:27:51 -07005917 name = "f16_igemm_minmax_test",
5918 srcs = [
5919 "test/f16-igemm-minmax.cc",
5920 "test/gemm-microkernel-tester.h",
5921 "src/xnnpack/AlignedAllocator.h",
5922 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005923 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Frank Barchardb0e4fae2020-05-04 15:27:51 -07005924)
5925
5926xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005927 name = "f32_conv_hwc_test",
5928 srcs = [
5929 "test/f32-conv-hwc.cc",
5930 "test/conv-hwc-microkernel-tester.h",
5931 "src/xnnpack/AlignedAllocator.h",
5932 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005933 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005934)
5935
5936xnnpack_unit_test(
Marat Dukhan1f29b802020-05-15 23:46:39 -07005937 name = "f32_conv_hwc2chw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005938 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07005939 "test/f32-conv-hwc2chw.cc",
5940 "test/conv-hwc2chw-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005941 "src/xnnpack/AlignedAllocator.h",
5942 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005943 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005944)
5945
5946xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07005947 name = "f32_dwconv_test",
5948 srcs = [
5949 "test/f32-dwconv.cc",
5950 "test/dwconv-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 Dukhan163a7e62020-04-09 04:19:26 -07005954)
5955
5956xnnpack_unit_test(
Frank Barchard5a599a62020-06-04 20:12:44 -07005957 name = "f16_dwconv_minmax_test",
5958 srcs = [
5959 "test/f16-dwconv-minmax.cc",
5960 "test/dwconv-microkernel-tester.h",
5961 "src/xnnpack/AlignedAllocator.h",
5962 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005963 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Frank Barchard5a599a62020-06-04 20:12:44 -07005964)
5965
5966xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07005967 name = "f32_dwconv_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005968 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07005969 "test/f32-dwconv-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005970 "test/dwconv-microkernel-tester.h",
5971 "src/xnnpack/AlignedAllocator.h",
5972 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005973 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005974)
5975
5976xnnpack_unit_test(
Marat Dukhanbf715f92020-10-23 20:17:00 -07005977 name = "f32_dwconv2d_chw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005978 srcs = [
Marat Dukhanbf715f92020-10-23 20:17:00 -07005979 "test/f32-dwconv2d-chw.cc",
5980 "test/dwconv2d-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005981 "src/xnnpack/AlignedAllocator.h",
5982 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005983 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005984)
5985
5986xnnpack_unit_test(
Frank Barchard0bb49a72020-06-04 11:35:11 -07005987 name = "f16_gavgpool_minmax_test",
5988 srcs = [
5989 "test/f16-gavgpool-minmax.cc",
5990 "test/gavgpool-microkernel-tester.h",
5991 "src/xnnpack/AlignedAllocator.h",
5992 ] + MICROKERNEL_TEST_HDRS,
5993 deps = MICROKERNEL_TEST_DEPS,
5994)
5995
5996xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07005997 name = "f32_gavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005998 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07005999 "test/f32-gavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006000 "test/gavgpool-microkernel-tester.h",
6001 "src/xnnpack/AlignedAllocator.h",
6002 ] + MICROKERNEL_TEST_HDRS,
6003 deps = MICROKERNEL_TEST_DEPS,
6004)
6005
6006xnnpack_unit_test(
Marat Dukhan1f29b802020-05-15 23:46:39 -07006007 name = "f32_gavgpool_cw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006008 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07006009 "test/f32-gavgpool-cw.cc",
6010 "test/gavgpool-cw-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006011 "src/xnnpack/AlignedAllocator.h",
6012 ] + MICROKERNEL_TEST_HDRS,
6013 deps = MICROKERNEL_TEST_DEPS,
6014)
6015
6016xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07006017 name = "f32_gemm_test",
6018 srcs = [
6019 "test/f32-gemm.cc",
6020 "test/gemm-microkernel-tester.h",
6021 "src/xnnpack/AlignedAllocator.h",
6022 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07006023 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan163a7e62020-04-09 04:19:26 -07006024)
6025
6026xnnpack_unit_test(
Marat Dukhan467f6362020-05-22 23:21:55 -07006027 name = "f32_gemm_relu_test",
6028 srcs = [
6029 "test/f32-gemm-relu.cc",
6030 "test/gemm-microkernel-tester.h",
6031 "src/xnnpack/AlignedAllocator.h",
6032 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07006033 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan467f6362020-05-22 23:21:55 -07006034)
6035
6036xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07006037 name = "f32_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006038 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07006039 "test/f32-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006040 "test/gemm-microkernel-tester.h",
6041 "src/xnnpack/AlignedAllocator.h",
6042 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07006043 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07006044)
6045
6046xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07006047 name = "f32_gemminc_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006048 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07006049 "test/f32-gemminc-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006050 "test/gemm-microkernel-tester.h",
6051 "src/xnnpack/AlignedAllocator.h",
6052 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07006053 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07006054)
6055
6056xnnpack_unit_test(
Frank Barchardb1966592020-05-12 13:47:06 -07006057 name = "f16_hswish_test",
6058 srcs = [
6059 "test/f16-hswish.cc",
6060 "test/hswish-microkernel-tester.h",
6061 ] + MICROKERNEL_TEST_HDRS,
6062 deps = MICROKERNEL_TEST_DEPS,
6063)
6064
6065xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07006066 name = "f32_hswish_test",
6067 srcs = [
6068 "test/f32-hswish.cc",
6069 "test/hswish-microkernel-tester.h",
6070 ] + MICROKERNEL_TEST_HDRS,
6071 deps = MICROKERNEL_TEST_DEPS,
6072)
6073
6074xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07006075 name = "f32_maxpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006076 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07006077 "test/f32-maxpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006078 "test/maxpool-microkernel-tester.h",
6079 ] + MICROKERNEL_TEST_HDRS,
6080 deps = MICROKERNEL_TEST_DEPS,
6081)
6082
6083xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07006084 name = "f32_pavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006085 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07006086 "test/f32-pavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006087 "test/avgpool-microkernel-tester.h",
6088 "src/xnnpack/AlignedAllocator.h",
6089 ] + MICROKERNEL_TEST_HDRS,
6090 deps = MICROKERNEL_TEST_DEPS,
6091)
6092
6093xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07006094 name = "f32_ppmm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006095 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07006096 "test/f32-ppmm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006097 "test/gemm-microkernel-tester.h",
6098 "src/xnnpack/AlignedAllocator.h",
6099 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07006100 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07006101)
6102
6103xnnpack_unit_test(
Frank Barchardb1966592020-05-12 13:47:06 -07006104 name = "f16_prelu_test",
6105 srcs = [
6106 "test/f16-prelu.cc",
6107 "test/prelu-microkernel-tester.h",
6108 "src/xnnpack/AlignedAllocator.h",
6109 ] + MICROKERNEL_TEST_HDRS,
6110 deps = MICROKERNEL_TEST_DEPS,
6111)
6112
6113xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07006114 name = "f32_prelu_test",
6115 srcs = [
6116 "test/f32-prelu.cc",
6117 "test/prelu-microkernel-tester.h",
6118 "src/xnnpack/AlignedAllocator.h",
6119 ] + MICROKERNEL_TEST_HDRS,
6120 deps = MICROKERNEL_TEST_DEPS,
6121)
6122
6123xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07006124 name = "f32_raddexpminusmax_test",
6125 srcs = [
6126 "test/f32-raddexpminusmax.cc",
6127 "test/raddexpminusmax-microkernel-tester.h",
6128 ] + MICROKERNEL_TEST_HDRS,
6129 deps = MICROKERNEL_TEST_DEPS,
6130)
6131
6132xnnpack_unit_test(
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07006133 name = "f32_raddextexp_test",
6134 srcs = [
6135 "test/f32-raddextexp.cc",
6136 "test/raddextexp-microkernel-tester.h",
6137 ] + MICROKERNEL_TEST_HDRS,
6138 deps = MICROKERNEL_TEST_DEPS,
6139)
6140
6141xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07006142 name = "f32_raddstoreexpminusmax_test",
6143 srcs = [
6144 "test/f32-raddstoreexpminusmax.cc",
6145 "test/raddstoreexpminusmax-microkernel-tester.h",
6146 ] + MICROKERNEL_TEST_HDRS,
6147 deps = MICROKERNEL_TEST_DEPS,
6148)
6149
6150xnnpack_unit_test(
Frank Barchardfb158e22020-07-15 16:10:10 -07006151 name = "f32_relu_test",
6152 srcs = [
6153 "test/f32-relu.cc",
6154 "test/vunary-microkernel-tester.h",
6155 ] + MICROKERNEL_TEST_HDRS,
6156 deps = MICROKERNEL_TEST_DEPS,
6157)
6158
6159xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07006160 name = "f32_rmax_test",
6161 srcs = [
6162 "test/f32-rmax.cc",
6163 "test/rmax-microkernel-tester.h",
6164 ] + MICROKERNEL_TEST_HDRS,
6165 deps = MICROKERNEL_TEST_DEPS,
6166)
6167
6168xnnpack_unit_test(
Marat Dukhan346a9e52019-11-15 09:06:30 -08006169 name = "f32_sigmoid_test",
6170 srcs = [
6171 "test/f32-sigmoid.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08006172 "test/vunary-microkernel-tester.h",
Marat Dukhan346a9e52019-11-15 09:06:30 -08006173 ] + MICROKERNEL_TEST_HDRS,
6174 deps = MICROKERNEL_TEST_DEPS,
6175)
6176
6177xnnpack_unit_test(
Marat Dukhan355ab432020-04-09 19:01:52 -07006178 name = "f32_spmm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006179 srcs = [
Marat Dukhan355ab432020-04-09 19:01:52 -07006180 "test/f32-spmm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006181 "test/spmm-microkernel-tester.h",
6182 "src/xnnpack/AlignedAllocator.h",
6183 ] + MICROKERNEL_TEST_HDRS,
6184 deps = MICROKERNEL_TEST_DEPS,
6185)
6186
6187xnnpack_unit_test(
Frank Barchardd793f6c2020-05-08 13:37:43 -07006188 name = "f16_vadd_minmax_test",
6189 srcs = [
6190 "test/f16-vadd-minmax.cc",
6191 "test/vbinary-microkernel-tester.h",
6192 ] + MICROKERNEL_TEST_HDRS,
6193 deps = MICROKERNEL_TEST_DEPS,
6194)
6195
6196xnnpack_unit_test(
6197 name = "f16_vaddc_minmax_test",
6198 srcs = [
6199 "test/f16-vaddc-minmax.cc",
6200 "test/vbinaryc-microkernel-tester.h",
6201 ] + MICROKERNEL_TEST_HDRS,
6202 deps = MICROKERNEL_TEST_DEPS,
6203)
6204
6205xnnpack_unit_test(
6206 name = "f16_vdiv_minmax_test",
6207 srcs = [
6208 "test/f16-vdiv-minmax.cc",
6209 "test/vbinary-microkernel-tester.h",
6210 ] + MICROKERNEL_TEST_HDRS,
6211 deps = MICROKERNEL_TEST_DEPS,
6212)
6213
6214xnnpack_unit_test(
6215 name = "f16_vdivc_minmax_test",
6216 srcs = [
6217 "test/f16-vdivc-minmax.cc",
6218 "test/vbinaryc-microkernel-tester.h",
6219 ] + MICROKERNEL_TEST_HDRS,
6220 deps = MICROKERNEL_TEST_DEPS,
6221)
6222
6223xnnpack_unit_test(
6224 name = "f16_vrdivc_minmax_test",
6225 srcs = [
6226 "test/f16-vrdivc-minmax.cc",
6227 "test/vbinaryc-microkernel-tester.h",
6228 ] + MICROKERNEL_TEST_HDRS,
6229 deps = MICROKERNEL_TEST_DEPS,
6230)
6231
6232xnnpack_unit_test(
6233 name = "f16_vmax_test",
6234 srcs = [
6235 "test/f16-vmax.cc",
6236 "test/vbinary-microkernel-tester.h",
6237 ] + MICROKERNEL_TEST_HDRS,
6238 deps = MICROKERNEL_TEST_DEPS,
6239)
6240
6241xnnpack_unit_test(
6242 name = "f16_vmaxc_test",
6243 srcs = [
6244 "test/f16-vmaxc.cc",
6245 "test/vbinaryc-microkernel-tester.h",
6246 ] + MICROKERNEL_TEST_HDRS,
6247 deps = MICROKERNEL_TEST_DEPS,
6248)
6249
6250xnnpack_unit_test(
6251 name = "f16_vmin_test",
6252 srcs = [
6253 "test/f16-vmin.cc",
6254 "test/vbinary-microkernel-tester.h",
6255 ] + MICROKERNEL_TEST_HDRS,
6256 deps = MICROKERNEL_TEST_DEPS,
6257)
6258
6259xnnpack_unit_test(
6260 name = "f16_vminc_test",
6261 srcs = [
6262 "test/f16-vminc.cc",
6263 "test/vbinaryc-microkernel-tester.h",
6264 ] + MICROKERNEL_TEST_HDRS,
6265 deps = MICROKERNEL_TEST_DEPS,
6266)
6267
6268xnnpack_unit_test(
6269 name = "f16_vmul_minmax_test",
6270 srcs = [
6271 "test/f16-vmul-minmax.cc",
6272 "test/vbinary-microkernel-tester.h",
6273 ] + MICROKERNEL_TEST_HDRS,
6274 deps = MICROKERNEL_TEST_DEPS,
6275)
6276
6277xnnpack_unit_test(
6278 name = "f16_vmulc_minmax_test",
6279 srcs = [
6280 "test/f16-vmulc-minmax.cc",
6281 "test/vbinaryc-microkernel-tester.h",
6282 ] + MICROKERNEL_TEST_HDRS,
6283 deps = MICROKERNEL_TEST_DEPS,
6284)
6285
6286xnnpack_unit_test(
6287 name = "f16_vsub_minmax_test",
6288 srcs = [
6289 "test/f16-vsub-minmax.cc",
6290 "test/vbinary-microkernel-tester.h",
6291 ] + MICROKERNEL_TEST_HDRS,
6292 deps = MICROKERNEL_TEST_DEPS,
6293)
6294
6295xnnpack_unit_test(
6296 name = "f16_vsubc_minmax_test",
6297 srcs = [
6298 "test/f16-vsubc-minmax.cc",
6299 "test/vbinaryc-microkernel-tester.h",
6300 ] + MICROKERNEL_TEST_HDRS,
6301 deps = MICROKERNEL_TEST_DEPS,
6302)
6303
6304xnnpack_unit_test(
6305 name = "f16_vrsubc_minmax_test",
6306 srcs = [
6307 "test/f16-vrsubc-minmax.cc",
6308 "test/vbinaryc-microkernel-tester.h",
6309 ] + MICROKERNEL_TEST_HDRS,
6310 deps = MICROKERNEL_TEST_DEPS,
6311)
6312
6313xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07006314 name = "f32_vabs_test",
6315 srcs = [
6316 "test/f32-vabs.cc",
6317 "test/vunary-microkernel-tester.h",
6318 ] + MICROKERNEL_TEST_HDRS,
6319 deps = MICROKERNEL_TEST_DEPS,
6320)
6321
6322xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07006323 name = "f32_vadd_test",
6324 srcs = [
6325 "test/f32-vadd.cc",
6326 "test/vbinary-microkernel-tester.h",
6327 ] + MICROKERNEL_TEST_HDRS,
6328 deps = MICROKERNEL_TEST_DEPS,
6329)
6330
6331xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006332 name = "f32_vadd_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006333 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006334 "test/f32-vadd-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08006335 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08006336 ] + MICROKERNEL_TEST_HDRS,
6337 deps = MICROKERNEL_TEST_DEPS,
6338)
6339
6340xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07006341 name = "f32_vadd_relu_test",
6342 srcs = [
6343 "test/f32-vadd-relu.cc",
6344 "test/vbinary-microkernel-tester.h",
6345 ] + MICROKERNEL_TEST_HDRS,
6346 deps = MICROKERNEL_TEST_DEPS,
6347)
6348
6349xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07006350 name = "f32_vaddc_test",
6351 srcs = [
6352 "test/f32-vaddc.cc",
6353 "test/vbinaryc-microkernel-tester.h",
6354 ] + MICROKERNEL_TEST_HDRS,
6355 deps = MICROKERNEL_TEST_DEPS,
6356)
6357
6358xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006359 name = "f32_vaddc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08006360 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006361 "test/f32-vaddc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08006362 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006363 ] + MICROKERNEL_TEST_HDRS,
6364 deps = MICROKERNEL_TEST_DEPS,
6365)
6366
6367xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07006368 name = "f32_vaddc_relu_test",
6369 srcs = [
6370 "test/f32-vaddc-relu.cc",
6371 "test/vbinaryc-microkernel-tester.h",
6372 ] + MICROKERNEL_TEST_HDRS,
6373 deps = MICROKERNEL_TEST_DEPS,
6374)
6375
6376xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07006377 name = "f32_vdiv_test",
6378 srcs = [
6379 "test/f32-vdiv.cc",
6380 "test/vbinary-microkernel-tester.h",
6381 ] + MICROKERNEL_TEST_HDRS,
6382 deps = MICROKERNEL_TEST_DEPS,
6383)
6384
6385xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006386 name = "f32_vdiv_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08006387 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006388 "test/f32-vdiv-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08006389 "test/vbinary-microkernel-tester.h",
6390 ] + MICROKERNEL_TEST_HDRS,
6391 deps = MICROKERNEL_TEST_DEPS,
6392)
6393
6394xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07006395 name = "f32_vdiv_relu_test",
6396 srcs = [
6397 "test/f32-vdiv-relu.cc",
6398 "test/vbinary-microkernel-tester.h",
6399 ] + MICROKERNEL_TEST_HDRS,
6400 deps = MICROKERNEL_TEST_DEPS,
6401)
6402
6403xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07006404 name = "f32_vdivc_test",
6405 srcs = [
6406 "test/f32-vdivc.cc",
6407 "test/vbinaryc-microkernel-tester.h",
6408 ] + MICROKERNEL_TEST_HDRS,
6409 deps = MICROKERNEL_TEST_DEPS,
6410)
6411
6412xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006413 name = "f32_vdivc_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08006414 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006415 "test/f32-vdivc-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08006416 "test/vbinaryc-microkernel-tester.h",
6417 ] + MICROKERNEL_TEST_HDRS,
6418 deps = MICROKERNEL_TEST_DEPS,
6419)
6420
6421xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07006422 name = "f32_vdivc_relu_test",
6423 srcs = [
6424 "test/f32-vdivc-relu.cc",
6425 "test/vbinaryc-microkernel-tester.h",
6426 ] + MICROKERNEL_TEST_HDRS,
6427 deps = MICROKERNEL_TEST_DEPS,
6428)
6429
6430xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07006431 name = "f32_vrdivc_test",
6432 srcs = [
6433 "test/f32-vrdivc.cc",
6434 "test/vbinaryc-microkernel-tester.h",
6435 ] + MICROKERNEL_TEST_HDRS,
6436 deps = MICROKERNEL_TEST_DEPS,
6437)
6438
6439xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006440 name = "f32_vrdivc_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08006441 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006442 "test/f32-vrdivc-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08006443 "test/vbinaryc-microkernel-tester.h",
6444 ] + MICROKERNEL_TEST_HDRS,
6445 deps = MICROKERNEL_TEST_DEPS,
6446)
6447
6448xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07006449 name = "f32_vrdivc_relu_test",
6450 srcs = [
6451 "test/f32-vrdivc-relu.cc",
6452 "test/vbinaryc-microkernel-tester.h",
6453 ] + MICROKERNEL_TEST_HDRS,
6454 deps = MICROKERNEL_TEST_DEPS,
6455)
6456
6457xnnpack_unit_test(
Marat Dukhaned6baaf2020-12-01 15:07:08 -08006458 name = "f32_velu_test",
6459 srcs = [
6460 "test/f32-velu.cc",
6461 "test/vunary-microkernel-tester.h",
6462 ] + MICROKERNEL_TEST_HDRS,
6463 deps = MICROKERNEL_TEST_DEPS,
6464)
6465
6466xnnpack_unit_test(
Marat Dukhan403b7d42019-12-05 12:49:11 -08006467 name = "f32_vmax_test",
6468 srcs = [
6469 "test/f32-vmax.cc",
6470 "test/vbinary-microkernel-tester.h",
6471 ] + MICROKERNEL_TEST_HDRS,
6472 deps = MICROKERNEL_TEST_DEPS,
6473)
6474
6475xnnpack_unit_test(
6476 name = "f32_vmaxc_test",
6477 srcs = [
6478 "test/f32-vmaxc.cc",
6479 "test/vbinaryc-microkernel-tester.h",
6480 ] + MICROKERNEL_TEST_HDRS,
6481 deps = MICROKERNEL_TEST_DEPS,
6482)
6483
6484xnnpack_unit_test(
6485 name = "f32_vmin_test",
6486 srcs = [
6487 "test/f32-vmin.cc",
6488 "test/vbinary-microkernel-tester.h",
6489 ] + MICROKERNEL_TEST_HDRS,
6490 deps = MICROKERNEL_TEST_DEPS,
6491)
6492
6493xnnpack_unit_test(
6494 name = "f32_vminc_test",
6495 srcs = [
6496 "test/f32-vminc.cc",
6497 "test/vbinaryc-microkernel-tester.h",
6498 ] + MICROKERNEL_TEST_HDRS,
6499 deps = MICROKERNEL_TEST_DEPS,
6500)
6501
6502xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07006503 name = "f32_vmul_test",
6504 srcs = [
6505 "test/f32-vmul.cc",
6506 "test/vbinary-microkernel-tester.h",
6507 ] + MICROKERNEL_TEST_HDRS,
6508 deps = MICROKERNEL_TEST_DEPS,
6509)
6510
6511xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006512 name = "f32_vmul_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006513 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006514 "test/f32-vmul-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08006515 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08006516 ] + MICROKERNEL_TEST_HDRS,
6517 deps = MICROKERNEL_TEST_DEPS,
6518)
6519
6520xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07006521 name = "f32_vmul_relu_test",
6522 srcs = [
6523 "test/f32-vmul-relu.cc",
6524 "test/vbinary-microkernel-tester.h",
6525 ] + MICROKERNEL_TEST_HDRS,
6526 deps = MICROKERNEL_TEST_DEPS,
6527)
6528
6529xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07006530 name = "f32_vmulc_test",
6531 srcs = [
6532 "test/f32-vmulc.cc",
6533 "test/vbinaryc-microkernel-tester.h",
6534 ] + MICROKERNEL_TEST_HDRS,
6535 deps = MICROKERNEL_TEST_DEPS,
6536)
6537
6538xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006539 name = "f32_vmulc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08006540 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006541 "test/f32-vmulc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08006542 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006543 ] + MICROKERNEL_TEST_HDRS,
6544 deps = MICROKERNEL_TEST_DEPS,
6545)
6546
6547xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07006548 name = "f32_vmulc_relu_test",
6549 srcs = [
6550 "test/f32-vmulc-relu.cc",
6551 "test/vbinaryc-microkernel-tester.h",
6552 ] + MICROKERNEL_TEST_HDRS,
6553 deps = MICROKERNEL_TEST_DEPS,
6554)
6555
6556xnnpack_unit_test(
Frank Barchard2a1049c2020-06-03 02:31:27 -07006557 name = "f16_vmulcaddc_minmax_test",
6558 srcs = [
6559 "test/f16-vmulcaddc-minmax.cc",
6560 "test/vmulcaddc-microkernel-tester.h",
6561 "src/xnnpack/AlignedAllocator.h",
6562 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07006563 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Frank Barchard2a1049c2020-06-03 02:31:27 -07006564)
6565
6566xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07006567 name = "f32_vmulcaddc_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006568 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07006569 "test/f32-vmulcaddc-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006570 "test/vmulcaddc-microkernel-tester.h",
6571 "src/xnnpack/AlignedAllocator.h",
6572 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07006573 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07006574)
6575
6576xnnpack_unit_test(
Marat Dukhan8cc7efe2020-06-10 16:24:27 -07006577 name = "f32_vlrelu_test",
6578 srcs = [
6579 "test/f32-vlrelu.cc",
6580 "test/vunary-microkernel-tester.h",
6581 ] + MICROKERNEL_TEST_HDRS,
6582 deps = MICROKERNEL_TEST_DEPS,
6583)
6584
6585xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07006586 name = "f32_vneg_test",
6587 srcs = [
6588 "test/f32-vneg.cc",
6589 "test/vunary-microkernel-tester.h",
6590 ] + MICROKERNEL_TEST_HDRS,
6591 deps = MICROKERNEL_TEST_DEPS,
6592)
6593
6594xnnpack_unit_test(
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07006595 name = "f32_vrndne_test",
6596 srcs = [
6597 "test/f32-vrndne.cc",
6598 "test/vunary-microkernel-tester.h",
6599 ] + MICROKERNEL_TEST_HDRS,
6600 deps = MICROKERNEL_TEST_DEPS,
6601)
6602
6603xnnpack_unit_test(
6604 name = "f32_vrndz_test",
6605 srcs = [
6606 "test/f32-vrndz.cc",
6607 "test/vunary-microkernel-tester.h",
6608 ] + MICROKERNEL_TEST_HDRS,
6609 deps = MICROKERNEL_TEST_DEPS,
6610)
6611
6612xnnpack_unit_test(
6613 name = "f32_vrndu_test",
6614 srcs = [
6615 "test/f32-vrndu.cc",
6616 "test/vunary-microkernel-tester.h",
6617 ] + MICROKERNEL_TEST_HDRS,
6618 deps = MICROKERNEL_TEST_DEPS,
6619)
6620
6621xnnpack_unit_test(
6622 name = "f32_vrndd_test",
6623 srcs = [
6624 "test/f32-vrndd.cc",
6625 "test/vunary-microkernel-tester.h",
6626 ] + MICROKERNEL_TEST_HDRS,
6627 deps = MICROKERNEL_TEST_DEPS,
6628)
6629
6630xnnpack_unit_test(
Marat Dukhan05ac8e32019-10-21 15:39:33 -07006631 name = "f32_vscale_test",
6632 srcs = [
6633 "test/f32-vscale.cc",
6634 "test/vscale-microkernel-tester.h",
6635 ] + MICROKERNEL_TEST_HDRS,
6636 deps = MICROKERNEL_TEST_DEPS,
6637)
6638
6639xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07006640 name = "f32_vscaleexpminusmax_test",
6641 srcs = [
6642 "test/f32-vscaleexpminusmax.cc",
6643 "test/vscaleexpminusmax-microkernel-tester.h",
6644 ] + MICROKERNEL_TEST_HDRS,
6645 deps = MICROKERNEL_TEST_DEPS,
6646)
6647
6648xnnpack_unit_test(
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07006649 name = "f32_vscaleextexp_test",
6650 srcs = [
6651 "test/f32-vscaleextexp.cc",
6652 "test/vscaleextexp-microkernel-tester.h",
6653 ] + MICROKERNEL_TEST_HDRS,
6654 deps = MICROKERNEL_TEST_DEPS,
6655)
6656
6657xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07006658 name = "f32_vsqr_test",
6659 srcs = [
6660 "test/f32-vsqr.cc",
6661 "test/vunary-microkernel-tester.h",
6662 ] + MICROKERNEL_TEST_HDRS,
6663 deps = MICROKERNEL_TEST_DEPS,
6664)
6665
6666xnnpack_unit_test(
Marat Dukhan13bafb02020-06-05 00:43:11 -07006667 name = "f32_vsqrdiff_test",
6668 srcs = [
6669 "test/f32-vsqrdiff.cc",
6670 "test/vbinary-microkernel-tester.h",
6671 ] + MICROKERNEL_TEST_HDRS,
6672 deps = MICROKERNEL_TEST_DEPS,
6673)
6674
6675xnnpack_unit_test(
6676 name = "f32_vsqrdiffc_test",
6677 srcs = [
6678 "test/f32-vsqrdiffc.cc",
6679 "test/vbinaryc-microkernel-tester.h",
6680 ] + MICROKERNEL_TEST_HDRS,
6681 deps = MICROKERNEL_TEST_DEPS,
6682)
6683
6684xnnpack_unit_test(
Marat Dukhanf4db2f32020-06-30 10:55:30 -07006685 name = "f32_vsqrt_test",
6686 srcs = [
6687 "test/f32-vsqrt.cc",
6688 "test/vunary-microkernel-tester.h",
6689 ] + MICROKERNEL_TEST_HDRS,
6690 deps = MICROKERNEL_TEST_DEPS,
6691)
6692
6693xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07006694 name = "f32_vsub_test",
6695 srcs = [
6696 "test/f32-vsub.cc",
6697 "test/vbinary-microkernel-tester.h",
6698 ] + MICROKERNEL_TEST_HDRS,
6699 deps = MICROKERNEL_TEST_DEPS,
6700)
6701
6702xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006703 name = "f32_vsub_minmax_test",
Marat Dukhan97579532019-10-18 16:40:39 -07006704 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006705 "test/f32-vsub-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08006706 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08006707 ] + MICROKERNEL_TEST_HDRS,
6708 deps = MICROKERNEL_TEST_DEPS,
6709)
6710
6711xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07006712 name = "f32_vsub_relu_test",
6713 srcs = [
6714 "test/f32-vsub-relu.cc",
6715 "test/vbinary-microkernel-tester.h",
6716 ] + MICROKERNEL_TEST_HDRS,
6717 deps = MICROKERNEL_TEST_DEPS,
6718)
6719
6720xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07006721 name = "f32_vsubc_test",
6722 srcs = [
6723 "test/f32-vsubc.cc",
6724 "test/vbinaryc-microkernel-tester.h",
6725 ] + MICROKERNEL_TEST_HDRS,
6726 deps = MICROKERNEL_TEST_DEPS,
6727)
6728
6729xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006730 name = "f32_vsubc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08006731 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006732 "test/f32-vsubc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08006733 "test/vbinaryc-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08006734 ] + MICROKERNEL_TEST_HDRS,
6735 deps = MICROKERNEL_TEST_DEPS,
6736)
6737
6738xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07006739 name = "f32_vsubc_relu_test",
6740 srcs = [
6741 "test/f32-vsubc-relu.cc",
6742 "test/vbinaryc-microkernel-tester.h",
6743 ] + MICROKERNEL_TEST_HDRS,
6744 deps = MICROKERNEL_TEST_DEPS,
6745)
6746
6747xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07006748 name = "f32_vrsubc_test",
6749 srcs = [
6750 "test/f32-vrsubc.cc",
6751 "test/vbinaryc-microkernel-tester.h",
6752 ] + MICROKERNEL_TEST_HDRS,
6753 deps = MICROKERNEL_TEST_DEPS,
6754)
6755
6756xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006757 name = "f32_vrsubc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08006758 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006759 "test/f32-vrsubc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08006760 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan97579532019-10-18 16:40:39 -07006761 ] + MICROKERNEL_TEST_HDRS,
6762 deps = MICROKERNEL_TEST_DEPS,
6763)
6764
6765xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07006766 name = "f32_vrsubc_relu_test",
6767 srcs = [
6768 "test/f32-vrsubc-relu.cc",
6769 "test/vbinaryc-microkernel-tester.h",
6770 ] + MICROKERNEL_TEST_HDRS,
6771 deps = MICROKERNEL_TEST_DEPS,
6772)
6773
6774xnnpack_unit_test(
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07006775 name = "qs8_dwconv_minmax_test",
6776 srcs = [
6777 "test/qs8-dwconv-minmax.cc",
6778 "test/dwconv-microkernel-tester.h",
6779 "src/xnnpack/AlignedAllocator.h",
6780 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
6781 deps = MICROKERNEL_TEST_DEPS + [":packing"],
6782)
6783
6784xnnpack_unit_test(
Marat Dukhan4ed53f42020-08-06 01:12:55 -07006785 name = "qs8_gavgpool_minmax_test",
6786 srcs = [
6787 "test/qs8-gavgpool-minmax.cc",
6788 "test/gavgpool-microkernel-tester.h",
6789 "src/xnnpack/AlignedAllocator.h",
6790 ] + MICROKERNEL_TEST_HDRS,
6791 deps = MICROKERNEL_TEST_DEPS,
6792)
6793
6794xnnpack_unit_test(
Marat Dukhan595e1702020-07-31 10:12:52 -07006795 name = "qs8_gemm_minmax_test",
6796 srcs = [
6797 "test/qs8-gemm-minmax.cc",
6798 "test/gemm-microkernel-tester.h",
6799 "src/xnnpack/AlignedAllocator.h",
6800 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
6801 deps = MICROKERNEL_TEST_DEPS + [":packing"],
6802)
6803
6804xnnpack_unit_test(
Marat Dukhanf9480682020-07-31 14:50:24 -07006805 name = "qs8_igemm_minmax_test",
6806 srcs = [
6807 "test/qs8-igemm-minmax.cc",
6808 "test/gemm-microkernel-tester.h",
6809 "src/xnnpack/AlignedAllocator.h",
6810 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
6811 deps = MICROKERNEL_TEST_DEPS + [":packing"],
6812)
6813
6814xnnpack_unit_test(
6815 name = "qs8_requantization_test",
6816 srcs = [
6817 "src/xnnpack/requantization-stubs.h",
6818 "test/qs8-requantization.cc",
6819 "test/requantization-tester.h",
6820 ] + MICROKERNEL_TEST_HDRS,
6821 deps = MICROKERNEL_TEST_DEPS,
6822)
6823
6824xnnpack_unit_test(
Marat Dukhand9f3ad42020-08-10 12:30:58 -07006825 name = "qs8_vadd_minmax_test",
6826 srcs = [
6827 "test/qs8-vadd-minmax.cc",
6828 "test/vadd-microkernel-tester.h",
6829 ] + MICROKERNEL_TEST_HDRS,
6830 deps = MICROKERNEL_TEST_DEPS,
6831)
6832
6833xnnpack_unit_test(
Marat Dukhan0270d9f2020-08-11 00:56:46 -07006834 name = "qs8_vaddc_minmax_test",
6835 srcs = [
6836 "test/qs8-vaddc-minmax.cc",
6837 "test/vaddc-microkernel-tester.h",
6838 ] + MICROKERNEL_TEST_HDRS,
6839 deps = MICROKERNEL_TEST_DEPS,
6840)
6841
6842xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07006843 name = "qu8_avgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006844 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07006845 "test/qu8-avgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006846 "test/avgpool-microkernel-tester.h",
6847 "src/xnnpack/AlignedAllocator.h",
6848 ] + MICROKERNEL_TEST_HDRS,
6849 deps = MICROKERNEL_TEST_DEPS,
6850)
6851
6852xnnpack_unit_test(
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07006853 name = "qu8_dwconv_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006854 srcs = [
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07006855 "test/qu8-dwconv-minmax.cc",
6856 "test/dwconv-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006857 "src/xnnpack/AlignedAllocator.h",
6858 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07006859 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07006860)
6861
6862xnnpack_unit_test(
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07006863 name = "qu8_igemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006864 srcs = [
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07006865 "test/qu8-igemm-minmax.cc",
6866 "test/gemm-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006867 "src/xnnpack/AlignedAllocator.h",
6868 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07006869 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07006870)
6871
6872xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07006873 name = "qu8_gavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006874 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07006875 "test/qu8-gavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006876 "test/gavgpool-microkernel-tester.h",
6877 "src/xnnpack/AlignedAllocator.h",
6878 ] + MICROKERNEL_TEST_HDRS,
6879 deps = MICROKERNEL_TEST_DEPS,
6880)
6881
6882xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07006883 name = "qu8_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006884 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07006885 "test/qu8-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006886 "test/gemm-microkernel-tester.h",
6887 "src/xnnpack/AlignedAllocator.h",
6888 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07006889 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07006890)
6891
6892xnnpack_unit_test(
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07006893 name = "qu8_requantization_test",
6894 srcs = [
6895 "src/xnnpack/requantization-stubs.h",
6896 "test/qu8-requantization.cc",
6897 "test/requantization-tester.h",
6898 ] + MICROKERNEL_TEST_HDRS,
6899 deps = MICROKERNEL_TEST_DEPS,
6900)
6901
6902xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07006903 name = "qu8_vadd_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006904 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07006905 "test/qu8-vadd-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006906 "test/vadd-microkernel-tester.h",
6907 ] + MICROKERNEL_TEST_HDRS,
6908 deps = MICROKERNEL_TEST_DEPS,
6909)
6910
6911xnnpack_unit_test(
6912 name = "u8_clamp_test",
6913 srcs = [
6914 "test/u8-clamp.cc",
6915 "test/clamp-microkernel-tester.h",
6916 ] + MICROKERNEL_TEST_HDRS,
6917 deps = MICROKERNEL_TEST_DEPS,
6918)
6919
6920xnnpack_unit_test(
6921 name = "u8_lut32norm_test",
6922 srcs = [
6923 "test/u8-lut32norm.cc",
6924 "test/lut-norm-microkernel-tester.h",
6925 ] + MICROKERNEL_TEST_HDRS,
6926 deps = MICROKERNEL_TEST_DEPS,
6927)
6928
6929xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07006930 name = "u8_maxpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006931 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07006932 "test/u8-maxpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006933 "test/maxpool-microkernel-tester.h",
6934 ] + MICROKERNEL_TEST_HDRS,
6935 deps = MICROKERNEL_TEST_DEPS,
6936)
6937
6938xnnpack_unit_test(
6939 name = "u8_rmax_test",
6940 srcs = [
6941 "test/u8-rmax.cc",
6942 "test/rmax-microkernel-tester.h",
6943 ] + MICROKERNEL_TEST_HDRS,
6944 deps = MICROKERNEL_TEST_DEPS,
6945)
6946
6947xnnpack_unit_test(
Marat Dukhanad71b9a2020-11-20 00:01:51 -08006948 name = "x32_depthtospace2d_chw2hwc_test",
Yury Kartynnike7841862020-11-04 18:22:18 -08006949 srcs = [
Marat Dukhanad71b9a2020-11-20 00:01:51 -08006950 "test/x32-depthtospace2d-chw2hwc.cc",
6951 "test/depthtospace-microkernel-tester.h",
Yury Kartynnike7841862020-11-04 18:22:18 -08006952 ] + MICROKERNEL_TEST_HDRS,
6953 deps = MICROKERNEL_TEST_DEPS,
6954)
6955
6956xnnpack_unit_test(
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -07006957 name = "x32_fill_test",
6958 srcs = [
6959 "test/x32-fill.cc",
6960 "test/fill-microkernel-tester.h",
6961 ] + MICROKERNEL_TEST_HDRS,
6962 deps = MICROKERNEL_TEST_DEPS,
6963)
6964
6965xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07006966 name = "x32_packx_test",
6967 srcs = [
6968 "test/x32-packx.cc",
6969 "test/pack-microkernel-tester.h",
6970 "src/xnnpack/AlignedAllocator.h",
6971 ] + MICROKERNEL_TEST_HDRS,
6972 deps = MICROKERNEL_TEST_DEPS,
6973)
6974
6975xnnpack_unit_test(
6976 name = "x32_pad_test",
6977 srcs = [
6978 "test/x32-pad.cc",
6979 "test/pad-microkernel-tester.h",
6980 ] + MICROKERNEL_TEST_HDRS,
6981 deps = MICROKERNEL_TEST_DEPS,
6982)
6983
6984xnnpack_unit_test(
6985 name = "x32_unpool_test",
6986 srcs = [
6987 "test/x32-unpool.cc",
6988 "test/unpool-microkernel-tester.h",
6989 ] + MICROKERNEL_TEST_HDRS,
6990 deps = MICROKERNEL_TEST_DEPS,
6991)
6992
6993xnnpack_unit_test(
6994 name = "x32_zip_test",
6995 srcs = [
6996 "test/x32-zip.cc",
6997 "test/zip-microkernel-tester.h",
6998 ] + MICROKERNEL_TEST_HDRS,
6999 deps = MICROKERNEL_TEST_DEPS,
7000)
7001
7002xnnpack_unit_test(
7003 name = "x8_lut_test",
7004 srcs = [
7005 "test/x8-lut.cc",
7006 "test/lut-microkernel-tester.h",
7007 ] + MICROKERNEL_TEST_HDRS,
7008 deps = MICROKERNEL_TEST_DEPS,
7009)
7010
7011xnnpack_unit_test(
7012 name = "x8_zip_test",
7013 srcs = [
7014 "test/x8-zip.cc",
7015 "test/zip-microkernel-tester.h",
7016 ] + MICROKERNEL_TEST_HDRS,
7017 deps = MICROKERNEL_TEST_DEPS,
7018)
7019
Marat Dukhan20c3b922020-03-10 03:45:06 -07007020########################## Size tests for the library #########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07007021
7022xnnpack_binary(
Marat Dukhan20c3b922020-03-10 03:45:06 -07007023 name = "operator_size_test",
7024 srcs = ["test/operator-size.c"],
Marat Dukhanf0cb70a2021-03-30 15:45:15 -07007025 deps = [":xnnpack_for_tfjs"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07007026)
7027
Marat Dukhan20c3b922020-03-10 03:45:06 -07007028xnnpack_binary(
7029 name = "subgraph_size_test",
7030 srcs = ["test/subgraph-size.c"],
7031 deps = [":XNNPACK"],
7032)
7033
7034########################### Unit tests for operators ##########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07007035
7036xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07007037 name = "abs_nc_test",
7038 srcs = [
7039 "test/abs-nc.cc",
7040 "test/abs-operator-tester.h",
7041 ],
7042 deps = OPERATOR_TEST_DEPS,
7043)
7044
7045xnnpack_unit_test(
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08007046 name = "add_nd_test",
Artsiom Ablavatski2202c812021-01-22 14:16:43 -08007047 timeout = "moderate",
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08007048 srcs = [
7049 "test/add-nd.cc",
7050 "test/binary-elementwise-operator-tester.h",
7051 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07007052 deps = OPERATOR_TEST_DEPS,
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08007053)
7054
7055xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08007056 name = "argmax_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007057 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08007058 "test/argmax-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007059 "test/argmax-pooling-operator-tester.h",
7060 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07007061 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07007062)
7063
7064xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08007065 name = "average_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007066 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08007067 "test/average-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007068 "test/average-pooling-operator-tester.h",
7069 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07007070 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07007071)
7072
7073xnnpack_unit_test(
Marat Dukhan64e52512020-06-09 13:41:16 -07007074 name = "bankers_rounding_nc_test",
7075 srcs = [
7076 "test/bankers-rounding-nc.cc",
7077 "test/bankers-rounding-operator-tester.h",
7078 ],
7079 deps = OPERATOR_TEST_DEPS,
7080)
7081
7082xnnpack_unit_test(
7083 name = "ceiling_nc_test",
7084 srcs = [
7085 "test/ceiling-nc.cc",
7086 "test/ceiling-operator-tester.h",
7087 ],
7088 deps = OPERATOR_TEST_DEPS,
7089)
7090
7091xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08007092 name = "channel_shuffle_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007093 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08007094 "test/channel-shuffle-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007095 "test/channel-shuffle-operator-tester.h",
7096 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07007097 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07007098)
7099
7100xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08007101 name = "clamp_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007102 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08007103 "test/clamp-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007104 "test/clamp-operator-tester.h",
7105 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07007106 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07007107)
7108
7109xnnpack_unit_test(
Marat Dukhan065b11e2020-05-22 09:49:41 -07007110 name = "constant_pad_nd_test",
7111 srcs = [
7112 "test/constant-pad-nd.cc",
7113 "test/constant-pad-operator-tester.h",
7114 ],
7115 deps = OPERATOR_TEST_DEPS,
7116)
7117
7118xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08007119 name = "convolution_nhwc_test",
Artsiom Ablavatski2202c812021-01-22 14:16:43 -08007120 timeout = "moderate",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007121 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08007122 "test/convolution-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007123 "test/convolution-operator-tester.h",
7124 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07007125 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07007126)
7127
7128xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08007129 name = "convolution_nchw_test",
Artsiom Ablavatski2202c812021-01-22 14:16:43 -08007130 timeout = "moderate",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007131 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08007132 "test/convolution-nchw.cc",
7133 "test/convolution-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007134 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07007135 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07007136)
7137
7138xnnpack_unit_test(
Marat Dukhan4e21b272020-06-04 18:45:01 -07007139 name = "copy_nc_test",
7140 srcs = [
7141 "test/copy-nc.cc",
7142 "test/copy-operator-tester.h",
7143 ],
7144 deps = OPERATOR_TEST_DEPS,
7145)
7146
7147xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08007148 name = "deconvolution_nhwc_test",
Artsiom Ablavatskic1aa2972020-12-08 11:23:34 -08007149 timeout = "moderate",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007150 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08007151 "test/deconvolution-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007152 "test/deconvolution-operator-tester.h",
7153 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07007154 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07007155)
7156
7157xnnpack_unit_test(
Marat Dukhan188d1042020-11-24 23:39:40 -08007158 name = "depth_to_space_nchw2nhwc_test",
Artsiom Ablavatski0f1dc182020-11-05 19:21:50 -08007159 srcs = [
Marat Dukhan188d1042020-11-24 23:39:40 -08007160 "test/depth-to-space-nchw2nhwc.cc",
Artsiom Ablavatski0f1dc182020-11-05 19:21:50 -08007161 "test/depth-to-space-operator-tester.h",
7162 ] + OPERATOR_TEST_PARAMS_HDRS,
7163 deps = OPERATOR_TEST_DEPS,
7164)
7165
7166xnnpack_unit_test(
Marat Dukhan0e521172020-11-25 13:10:04 -08007167 name = "depth_to_space_nhwc_test",
7168 srcs = [
7169 "test/depth-to-space-nhwc.cc",
7170 "test/depth-to-space-operator-tester.h",
7171 ] + OPERATOR_TEST_PARAMS_HDRS,
7172 deps = OPERATOR_TEST_DEPS,
7173)
7174
7175xnnpack_unit_test(
Marat Dukhan69180502019-12-06 15:00:31 -08007176 name = "divide_nd_test",
7177 srcs = [
7178 "test/binary-elementwise-operator-tester.h",
7179 "test/divide-nd.cc",
7180 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07007181 deps = OPERATOR_TEST_DEPS,
Marat Dukhan69180502019-12-06 15:00:31 -08007182)
7183
7184xnnpack_unit_test(
Marat Dukhanb6bd4bc2020-12-01 17:01:40 -08007185 name = "elu_nc_test",
7186 srcs = [
7187 "test/elu-nc.cc",
7188 "test/elu-operator-tester.h",
7189 ],
7190 deps = OPERATOR_TEST_DEPS,
7191)
7192
7193xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08007194 name = "fully_connected_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007195 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08007196 "test/fully-connected-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007197 "test/fully-connected-operator-tester.h",
7198 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07007199 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07007200)
7201
7202xnnpack_unit_test(
Marat Dukhan64e52512020-06-09 13:41:16 -07007203 name = "floor_nc_test",
7204 srcs = [
7205 "test/floor-nc.cc",
7206 "test/floor-operator-tester.h",
7207 ],
7208 deps = OPERATOR_TEST_DEPS,
7209)
7210
7211xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08007212 name = "global_average_pooling_nwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007213 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08007214 "test/global-average-pooling-nwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007215 "test/global-average-pooling-operator-tester.h",
Marat Dukhanef61d022020-06-19 13:54:49 -07007216 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07007217 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07007218)
7219
7220xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08007221 name = "global_average_pooling_ncw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007222 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08007223 "test/global-average-pooling-ncw.cc",
7224 "test/global-average-pooling-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007225 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07007226 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07007227)
7228
7229xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08007230 name = "hardswish_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007231 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08007232 "test/hardswish-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007233 "test/hardswish-operator-tester.h",
7234 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07007235 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07007236)
7237
7238xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08007239 name = "leaky_relu_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007240 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08007241 "test/leaky-relu-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007242 "test/leaky-relu-operator-tester.h",
7243 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07007244 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07007245)
7246
7247xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08007248 name = "max_pooling_nhwc_test",
Artsiom Ablavatski2202c812021-01-22 14:16:43 -08007249 timeout = "moderate",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007250 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08007251 "test/max-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007252 "test/max-pooling-operator-tester.h",
7253 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07007254 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07007255)
7256
7257xnnpack_unit_test(
Marat Dukhan79e7f842019-12-05 14:35:50 -08007258 name = "maximum_nd_test",
7259 srcs = [
7260 "test/binary-elementwise-operator-tester.h",
7261 "test/maximum-nd.cc",
7262 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07007263 deps = OPERATOR_TEST_DEPS,
Marat Dukhan79e7f842019-12-05 14:35:50 -08007264)
7265
7266xnnpack_unit_test(
7267 name = "minimum_nd_test",
7268 srcs = [
7269 "test/binary-elementwise-operator-tester.h",
7270 "test/minimum-nd.cc",
7271 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07007272 deps = OPERATOR_TEST_DEPS,
Marat Dukhan79e7f842019-12-05 14:35:50 -08007273)
7274
7275xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08007276 name = "multiply_nd_test",
Marat Dukhanca2733c2019-11-15 23:21:17 -08007277 srcs = [
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08007278 "test/binary-elementwise-operator-tester.h",
Marat Dukhanefc47b82019-11-18 09:25:38 -08007279 "test/multiply-nd.cc",
Marat Dukhanca2733c2019-11-15 23:21:17 -08007280 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07007281 deps = OPERATOR_TEST_DEPS,
Marat Dukhanca2733c2019-11-15 23:21:17 -08007282)
7283
7284xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07007285 name = "negate_nc_test",
7286 srcs = [
7287 "test/negate-nc.cc",
7288 "test/negate-operator-tester.h",
7289 ],
7290 deps = OPERATOR_TEST_DEPS,
7291)
7292
7293xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08007294 name = "prelu_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007295 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08007296 "test/prelu-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007297 "test/prelu-operator-tester.h",
7298 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07007299 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07007300)
7301
7302xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08007303 name = "resize_bilinear_nhwc_test",
Marat Dukhan69722492019-11-11 19:55:50 -08007304 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08007305 "test/resize-bilinear-nhwc.cc",
Marat Dukhan69722492019-11-11 19:55:50 -08007306 "test/resize-bilinear-operator-tester.h",
7307 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07007308 deps = OPERATOR_TEST_DEPS,
Marat Dukhan69722492019-11-11 19:55:50 -08007309)
7310
7311xnnpack_unit_test(
Artsiom Ablavatski97918102020-10-27 15:52:59 -07007312 name = "resize_bilinear_nchw_test",
7313 srcs = [
7314 "test/resize-bilinear-nchw.cc",
7315 "test/resize-bilinear-operator-tester.h",
7316 ] + OPERATOR_TEST_PARAMS_HDRS,
7317 deps = OPERATOR_TEST_DEPS,
7318)
7319
7320xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08007321 name = "sigmoid_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007322 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08007323 "test/sigmoid-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007324 "test/sigmoid-operator-tester.h",
7325 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07007326 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07007327)
7328
7329xnnpack_unit_test(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08007330 name = "softmax_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007331 srcs = [
Marat Dukhanfd8e6892020-01-27 15:25:25 -08007332 "test/softmax-nc.cc",
7333 "test/softmax-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007334 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07007335 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07007336)
7337
7338xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07007339 name = "square_nc_test",
7340 srcs = [
7341 "test/square-nc.cc",
7342 "test/square-operator-tester.h",
7343 ],
7344 deps = OPERATOR_TEST_DEPS,
7345)
7346
7347xnnpack_unit_test(
Marat Dukhan6804bbd2020-06-30 19:26:11 -07007348 name = "square_root_nc_test",
7349 srcs = [
7350 "test/square-root-nc.cc",
7351 "test/square-root-operator-tester.h",
7352 ],
7353 deps = OPERATOR_TEST_DEPS,
7354)
7355
7356xnnpack_unit_test(
Marat Dukhanf7399262020-06-05 10:58:44 -07007357 name = "squared_difference_nd_test",
7358 srcs = [
7359 "test/binary-elementwise-operator-tester.h",
7360 "test/squared-difference-nd.cc",
7361 ],
7362 deps = OPERATOR_TEST_DEPS,
7363)
7364
7365xnnpack_unit_test(
Marat Dukhan05f3f6d2019-12-03 15:13:53 -08007366 name = "subtract_nd_test",
7367 srcs = [
7368 "test/binary-elementwise-operator-tester.h",
7369 "test/subtract-nd.cc",
7370 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07007371 deps = OPERATOR_TEST_DEPS,
Marat Dukhan05f3f6d2019-12-03 15:13:53 -08007372)
7373
7374xnnpack_unit_test(
Marat Dukhan64e52512020-06-09 13:41:16 -07007375 name = "truncation_nc_test",
7376 srcs = [
7377 "test/truncation-nc.cc",
7378 "test/truncation-operator-tester.h",
7379 ],
7380 deps = OPERATOR_TEST_DEPS,
7381)
7382
7383xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08007384 name = "unpooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007385 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08007386 "test/unpooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007387 "test/unpooling-operator-tester.h",
7388 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07007389 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07007390)
7391
Chao Mei6ddfc602020-05-13 22:29:36 -07007392############################### Misc unit tests ###############################
7393
7394xnnpack_unit_test(
7395 name = "memory_planner_test",
7396 srcs = [
7397 "test/memory-planner-test.cc",
7398 ],
7399 deps = [
7400 ":XNNPACK",
7401 ":memory_planner",
7402 ],
7403)
7404
XNNPACK Teamab8c4c82020-10-09 08:05:51 -07007405xnnpack_unit_test(
7406 name = "subgraph_nchw_test",
7407 srcs = [
7408 "src/xnnpack/subgraph.h",
7409 "test/subgraph-nchw.cc",
7410 "test/subgraph-tester.h",
7411 ],
7412 deps = [
7413 ":XNNPACK",
7414 ],
7415)
7416
Marat Dukhan08c4a432019-10-03 09:29:21 -07007417############################# Build configurations #############################
7418
Marat Dukhanb8642352019-10-30 15:43:02 -07007419# Enables usage of assembly kernels.
Marat Dukhan08c4a432019-10-03 09:29:21 -07007420config_setting(
Marat Dukhanb8642352019-10-30 15:43:02 -07007421 name = "xnn_enable_assembly_explicit_true",
7422 define_values = {"xnn_enable_assembly": "true"},
7423)
7424
7425# Disables usage of assembly kernels.
7426config_setting(
7427 name = "xnn_enable_assembly_explicit_false",
7428 define_values = {"xnn_enable_assembly": "false"},
7429)
7430
Marat Dukhan9de90e02020-06-18 16:04:12 -07007431# Enables usage of sparse inference.
7432config_setting(
7433 name = "xnn_enable_sparse_explicit_true",
7434 define_values = {"xnn_enable_sparse": "true"},
7435)
7436
7437# Disables usage of sparse inference.
7438config_setting(
7439 name = "xnn_enable_sparse_explicit_false",
7440 define_values = {"xnn_enable_sparse": "false"},
7441)
7442
Marat Dukhan05702cf2020-03-26 15:41:33 -07007443# Disables usage of HMP-aware optimizations.
7444config_setting(
7445 name = "xnn_enable_hmp_explicit_false",
7446 define_values = {"xnn_enable_hmp": "false"},
7447)
7448
Chao Mei6ddfc602020-05-13 22:29:36 -07007449# Enable usage of optimized memory allocation
7450config_setting(
7451 name = "xnn_enable_memopt_explicit_true",
Marat Dukhan03f46212021-03-30 21:29:49 -07007452 define_values = {"xnn_enable_memopt": "true"},
Chao Mei6ddfc602020-05-13 22:29:36 -07007453)
7454
7455# Disable usage of optimized memory allocation
7456config_setting(
7457 name = "xnn_enable_memopt_explicit_false",
Marat Dukhan03f46212021-03-30 21:29:49 -07007458 define_values = {"xnn_enable_memopt": "false"},
Chao Mei6ddfc602020-05-13 22:29:36 -07007459)
7460
Marat Dukhanb939cdb2021-03-30 18:51:51 -07007461# Enable QS8 inference in TFLite-specific version
7462config_setting(
7463 name = "xnn_enable_qs8_explicit_true",
7464 define_values = {"xnn_enable_qs8": "true"},
7465)
7466
7467# Disable QS8 inference in TFLite-specific version
7468config_setting(
7469 name = "xnn_enable_qs8_explicit_false",
7470 define_values = {"xnn_enable_qs8": "false"},
7471)
7472
Marat Dukhanb8642352019-10-30 15:43:02 -07007473# Builds with -c dbg
7474config_setting(
7475 name = "debug_build",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007476 values = {
Marat Dukhanb8642352019-10-30 15:43:02 -07007477 "compilation_mode": "dbg",
7478 },
7479)
7480
7481# Builds with -c opt
7482config_setting(
7483 name = "optimized_build",
7484 values = {
7485 "compilation_mode": "opt",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007486 },
7487)
7488
7489config_setting(
Marat Dukhanb8642352019-10-30 15:43:02 -07007490 name = "linux_k8",
7491 values = {"cpu": "k8"},
7492)
7493
7494config_setting(
Marat Dukhan582094e2020-04-30 17:21:25 -07007495 name = "linux_arm",
7496 values = {"cpu": "arm"},
Marat Dukhan4e45e662019-10-03 15:40:24 -07007497)
7498
7499config_setting(
Marat Dukhanf0bd4de2020-06-15 15:53:19 -07007500 name = "linux_armeabi",
7501 values = {"cpu": "armeabi"},
7502)
7503
7504config_setting(
Terry Heo68eef3f2020-04-13 22:53:52 -07007505 name = "linux_armhf",
7506 values = {"cpu": "armhf"},
7507)
7508
7509config_setting(
Marat Dukhana720e932020-06-10 13:01:11 -07007510 name = "linux_armv7a",
7511 values = {"cpu": "armv7a"},
7512)
7513
7514config_setting(
Marat Dukhan582094e2020-04-30 17:21:25 -07007515 name = "linux_aarch64",
7516 values = {"cpu": "aarch64"},
7517)
7518
7519config_setting(
Marat Dukhan08c4a432019-10-03 09:29:21 -07007520 name = "android",
7521 values = {"crosstool_top": "//external:android/crosstool"},
7522)
7523
7524config_setting(
7525 name = "android_armv7",
7526 values = {
7527 "crosstool_top": "//external:android/crosstool",
7528 "cpu": "armeabi-v7a",
7529 },
7530)
7531
7532config_setting(
7533 name = "android_arm64",
7534 values = {
7535 "crosstool_top": "//external:android/crosstool",
7536 "cpu": "arm64-v8a",
7537 },
7538)
7539
7540config_setting(
7541 name = "android_x86",
7542 values = {
7543 "crosstool_top": "//external:android/crosstool",
7544 "cpu": "x86",
7545 },
7546)
7547
7548config_setting(
7549 name = "android_x86_64",
7550 values = {
7551 "crosstool_top": "//external:android/crosstool",
7552 "cpu": "x86_64",
7553 },
7554)
7555
7556config_setting(
Marat Dukhan10a38082020-04-17 03:58:35 -07007557 name = "windows_x86_64",
7558 values = {"cpu": "x64_windows"},
Marat Dukhan9fe932e2020-04-11 17:14:15 -07007559)
7560
7561config_setting(
Marat Dukhan10a38082020-04-17 03:58:35 -07007562 name = "windows_x86_64_clang",
7563 values = {
7564 "compiler": "clang-cl",
7565 "cpu": "x64_windows",
7566 },
7567)
7568
7569config_setting(
7570 name = "windows_x86_64_mingw",
7571 values = {
7572 "compiler": "mingw-gcc",
7573 "cpu": "x64_windows",
7574 },
7575)
7576
7577config_setting(
7578 name = "windows_x86_64_msys",
7579 values = {
7580 "compiler": "msys-gcc",
7581 "cpu": "x64_windows",
7582 },
Marat Dukhan9fe932e2020-04-11 17:14:15 -07007583)
7584
7585config_setting(
Marat Dukhan885ca242019-10-07 09:17:32 -07007586 name = "macos_x86_64",
7587 values = {
7588 "apple_platform_type": "macos",
7589 "cpu": "darwin",
7590 },
7591)
7592
7593config_setting(
Simon Maurerae33ab82021-03-03 23:38:22 +01007594 name = "macos_arm64",
7595 values = {
7596 "apple_platform_type": "macos",
7597 "cpu": "darwin_arm64",
7598 },
7599)
7600
7601config_setting(
Marat Dukhan08c4a432019-10-03 09:29:21 -07007602 name = "emscripten",
XNNPACK Team3bfbdaf2021-03-29 15:26:23 -07007603 values = {"crosstool_top": "@emsdk//emscripten_toolchain:everything"},
Marat Dukhan08c4a432019-10-03 09:29:21 -07007604)
7605
7606config_setting(
7607 name = "emscripten_wasm",
7608 values = {
XNNPACK Team3bfbdaf2021-03-29 15:26:23 -07007609 "crosstool_top": "@emsdk//emscripten_toolchain:everything",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007610 "cpu": "wasm",
7611 },
7612)
7613
7614config_setting(
7615 name = "emscripten_wasmsimd",
7616 values = {
XNNPACK Team3bfbdaf2021-03-29 15:26:23 -07007617 "crosstool_top": "@emsdk//emscripten_toolchain:everything",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007618 "cpu": "wasm",
Marat Dukhan81c62602020-05-29 13:22:49 -07007619 "copt": "-msimd128",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007620 },
7621)
7622
7623config_setting(
Marat Dukhan1498d1d2020-02-11 20:00:05 -08007624 name = "ios_armv7",
7625 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08007626 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08007627 "cpu": "ios_armv7",
7628 },
7629)
7630
7631config_setting(
7632 name = "ios_arm64",
7633 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08007634 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08007635 "cpu": "ios_arm64",
7636 },
7637)
7638
7639config_setting(
7640 name = "ios_arm64e",
7641 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08007642 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08007643 "cpu": "ios_arm64e",
7644 },
7645)
7646
7647config_setting(
7648 name = "ios_x86",
7649 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08007650 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08007651 "cpu": "ios_i386",
7652 },
7653)
7654
7655config_setting(
7656 name = "ios_x86_64",
7657 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08007658 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08007659 "cpu": "ios_x86_64",
7660 },
7661)
7662
7663config_setting(
7664 name = "watchos_armv7k",
7665 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08007666 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08007667 "cpu": "watchos_armv7k",
7668 },
7669)
7670
7671config_setting(
7672 name = "watchos_arm64_32",
7673 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08007674 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08007675 "cpu": "watchos_arm64_32",
7676 },
7677)
7678
7679config_setting(
7680 name = "watchos_x86",
7681 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08007682 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08007683 "cpu": "watchos_i386",
7684 },
7685)
7686
7687config_setting(
7688 name = "watchos_x86_64",
7689 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08007690 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08007691 "cpu": "watchos_x86_64",
7692 },
7693)
7694
7695config_setting(
7696 name = "tvos_arm64",
7697 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08007698 "apple_platform_type": "tvos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08007699 "cpu": "tvos_arm64",
7700 },
7701)
7702
7703config_setting(
7704 name = "tvos_x86_64",
7705 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08007706 "apple_platform_type": "tvos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08007707 "cpu": "tvos_x86_64",
7708 },
7709)