blob: 45fa8c769a951c1d02d521783ce0a11c1dcd6b92 [file] [log] [blame]
Frank Barchard1a953052020-11-16 18:44:58 -08001# Copyright 2020 Google LLC
Marat Dukhan08c4a432019-10-03 09:29:21 -07002#
3# This source code is licensed under the BSD-style license found in the
4# LICENSE file in the root directory of this source tree.
5#
6# Description:
7# XNNPACK - optimized floating-point neural network operators library
8
Marat Dukhan10a38082020-04-17 03:58:35 -07009load(":build_defs.bzl", "xnnpack_aggregate_library", "xnnpack_benchmark", "xnnpack_binary", "xnnpack_cc_library", "xnnpack_gcc_std_copts", "xnnpack_min_size_copts", "xnnpack_msvc_std_copts", "xnnpack_optional_armcl_copts", "xnnpack_optional_armcl_deps", "xnnpack_optional_dnnl_copts", "xnnpack_optional_dnnl_deps", "xnnpack_optional_gemmlowp_copts", "xnnpack_optional_gemmlowp_deps", "xnnpack_optional_ruy_copts", "xnnpack_optional_ruy_deps", "xnnpack_optional_tflite_copts", "xnnpack_optional_tflite_deps", "xnnpack_std_cxxopts", "xnnpack_unit_test", "xnnpack_visibility")
Marat Dukhan69c3f2c2019-11-06 12:30:01 -080010
Marat Dukhan08c4a432019-10-03 09:29:21 -070011licenses(["notice"])
12
13exports_files(["LICENSE"])
14
Marat Dukhan1b354632020-03-23 12:50:22 -070015OPERATOR_BENCHMARK_DEPS = [
16 ":XNNPACK",
17 ":bench_utils",
18 "@cpuinfo",
Frank Barchard0c849732020-06-12 13:31:32 -070019 "@FP16",
Marat Dukhan1b354632020-03-23 12:50:22 -070020 "@pthreadpool",
21]
22
Marat Dukhan08c4a432019-10-03 09:29:21 -070023MICROKERNEL_BENCHMARK_DEPS = [
24 ":ukernels",
25 ":bench_utils",
Frank Barchard7e955972019-10-11 10:34:25 -070026 ":enable_assembly",
Marat Dukhan08c4a432019-10-03 09:29:21 -070027 "@cpuinfo",
28 "@FP16",
29 "@pthreadpool",
30]
31
Marat Dukhan6adff4e2019-10-14 18:32:07 -070032ACCURACY_EVAL_DEPS = [
33 ":XNNPACK",
34 ":ukernels",
35 "@FP16",
36 "@pthreadpool",
37]
38
Marat Dukhan08c4a432019-10-03 09:29:21 -070039MICROKERNEL_TEST_DEPS = [
Marat Dukhan33fcf782020-05-24 14:27:15 -070040 ":ukernels_test_mode",
Frank Barchard7e955972019-10-11 10:34:25 -070041 ":enable_assembly",
Marat Dukhan08c4a432019-10-03 09:29:21 -070042 "@cpuinfo",
43 "@FP16",
44 "@pthreadpool",
45]
46
Marat Dukhan1b354632020-03-23 12:50:22 -070047OPERATOR_TEST_DEPS = [
Marat Dukhan33fcf782020-05-24 14:27:15 -070048 ":XNNPACK_test_mode",
Marat Dukhan1b354632020-03-23 12:50:22 -070049 "@pthreadpool",
50 "@FP16",
51]
52
Marat Dukhan08c4a432019-10-03 09:29:21 -070053OPERATOR_SRCS = [
Marat Dukhane8265432020-04-28 18:42:59 -070054 "src/operators/argmax-pooling-nhwc.c",
55 "src/operators/average-pooling-nhwc.c",
56 "src/operators/binary-elementwise-nd.c",
Marat Dukhane8265432020-04-28 18:42:59 -070057 "src/operators/channel-shuffle-nc.c",
Marat Dukhan065b11e2020-05-22 09:49:41 -070058 "src/operators/constant-pad-nd.c",
Marat Dukhane8265432020-04-28 18:42:59 -070059 "src/operators/convolution-nchw.c",
60 "src/operators/convolution-nhwc.c",
61 "src/operators/deconvolution-nhwc.c",
Marat Dukhan13b68f22020-11-12 11:55:19 -080062 "src/operators/depth-to-space-nchw2nhwc.c",
Marat Dukhan0e521172020-11-25 13:10:04 -080063 "src/operators/depth-to-space-nhwc.c",
Marat Dukhane8265432020-04-28 18:42:59 -070064 "src/operators/fully-connected-nc.c",
65 "src/operators/global-average-pooling-ncw.c",
66 "src/operators/global-average-pooling-nwc.c",
Marat Dukhane8265432020-04-28 18:42:59 -070067 "src/operators/leaky-relu-nc.c",
68 "src/operators/max-pooling-nhwc.c",
69 "src/operators/prelu-nc.c",
Artsiom Ablavatski97918102020-10-27 15:52:59 -070070 "src/operators/resize-bilinear-nchw.c",
Marat Dukhane8265432020-04-28 18:42:59 -070071 "src/operators/resize-bilinear-nhwc.c",
72 "src/operators/sigmoid-nc.c",
73 "src/operators/softmax-nc.c",
Marat Dukhanc3065f52020-06-04 13:33:32 -070074 "src/operators/unary-elementwise-nc.c",
Marat Dukhane8265432020-04-28 18:42:59 -070075 "src/operators/unpooling-nhwc.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -070076]
77
Marat Dukhan95e8b7a2020-06-03 12:46:26 -070078SUBGRAPH_SRCS = [
Marat Dukhan5fab4092020-06-10 01:28:28 -070079 "src/subgraph/abs.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -070080 "src/subgraph/add2.c",
81 "src/subgraph/argmax-pooling-2d.c",
82 "src/subgraph/average-pooling-2d.c",
Marat Dukhan5fab4092020-06-10 01:28:28 -070083 "src/subgraph/bankers-rounding.c",
84 "src/subgraph/ceiling.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -070085 "src/subgraph/clamp.c",
86 "src/subgraph/convolution-2d.c",
87 "src/subgraph/deconvolution-2d.c",
Artsiom Ablavatskibbe85062020-11-05 14:07:37 -080088 "src/subgraph/depth-to-space.c",
Frank Barchard9cef5ea2020-11-18 14:52:08 -080089 "src/subgraph/depthwise-convolution-2d.c",
Marat Dukhan9d3a4592020-06-05 16:52:42 -070090 "src/subgraph/divide.c",
Marat Dukhana1600202020-12-01 22:17:16 -080091 "src/subgraph/elu.c",
Marat Dukhan5fab4092020-06-10 01:28:28 -070092 "src/subgraph/floor.c",
Frank Barchard04336c12020-10-22 16:48:55 -070093 "src/subgraph/fully-connected.c",
Marat Dukhana059b7d2020-06-11 11:41:27 -070094 "src/subgraph/global-average-pooling-2d.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -070095 "src/subgraph/hardswish.c",
Marat Dukhan5bbebac2020-06-10 19:42:15 -070096 "src/subgraph/leaky-relu.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -070097 "src/subgraph/max-pooling-2d.c",
Marat Dukhan9d3a4592020-06-05 16:52:42 -070098 "src/subgraph/maximum2.c",
99 "src/subgraph/minimum2.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -0700100 "src/subgraph/multiply2.c",
Marat Dukhan5fab4092020-06-10 01:28:28 -0700101 "src/subgraph/negate.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -0700102 "src/subgraph/prelu.c",
103 "src/subgraph/sigmoid.c",
104 "src/subgraph/softmax.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700105 "src/subgraph/square-root.c",
106 "src/subgraph/square.c",
107 "src/subgraph/squared-difference.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -0700108 "src/subgraph/static-constant-pad.c",
Marat Dukhand27202d2020-07-09 23:43:40 -0700109 "src/subgraph/static-reshape.c",
Marat Dukhanaff24e22020-07-23 01:43:58 -0700110 "src/subgraph/static-resize-bilinear-2d.c",
Marat Dukhan9d3a4592020-06-05 16:52:42 -0700111 "src/subgraph/subtract.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -0700112 "src/subgraph/unpooling-2d.c",
113]
114
Marat Dukhan3a77ea72019-12-23 12:10:24 -0800115TABLE_SRCS = [
116 "src/tables/exp2-k-over-64.c",
117 "src/tables/exp2-k-over-2048.c",
Marat Dukhande390d42020-11-29 19:32:18 -0800118 "src/tables/exp2minus-k-over-4.c",
119 "src/tables/exp2minus-k-over-8.c",
Marat Dukhanc60742b2020-11-23 12:33:27 -0800120 "src/tables/exp2minus-k-over-16.c",
Marat Dukhan1f256fc2020-09-24 21:27:14 -0700121 "src/tables/exp2minus-k-over-64.c",
122 "src/tables/exp2minus-k-over-2048.c",
Marat Dukhan3a77ea72019-12-23 12:10:24 -0800123]
124
Marat Dukhan08c4a432019-10-03 09:29:21 -0700125SCALAR_UKERNELS = [
Marat Dukhan329da642019-11-19 21:44:39 -0800126 "src/f32-argmaxpool/4x-scalar-c1.c",
Marat Dukhan1e782c42019-11-21 17:02:40 -0800127 "src/f32-argmaxpool/9p8x-scalar-c1.c",
Marat Dukhan329da642019-11-19 21:44:39 -0800128 "src/f32-argmaxpool/9x-scalar-c1.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700129 "src/f32-avgpool/9p8x-minmax-scalar-c1.c",
130 "src/f32-avgpool/9x-minmax-scalar-c1.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -0700131 "src/f32-clamp/gen/scalar-x1.c",
132 "src/f32-clamp/gen/scalar-x2.c",
133 "src/f32-clamp/gen/scalar-x4.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700134 "src/f32-conv-hwc/3x3s2p0p1c3x4-scalar-1x1.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700135 "src/f32-conv-hwc/3x3s2p1c3x4-scalar-1x1.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -0700136 "src/f32-conv-hwc2chw/3x3s2p1c3x4-scalar-1x1.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700137 "src/f32-dwconv/gen/up1x4-minmax-scalar-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700138 "src/f32-dwconv/gen/up1x4-minmax-scalar.c",
139 "src/f32-dwconv/gen/up1x4-scalar-acc2.c",
140 "src/f32-dwconv/gen/up1x4-scalar.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700141 "src/f32-dwconv/gen/up1x9-minmax-scalar-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700142 "src/f32-dwconv/gen/up1x9-minmax-scalar.c",
143 "src/f32-dwconv/gen/up1x9-scalar-acc2.c",
144 "src/f32-dwconv/gen/up1x9-scalar.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700145 "src/f32-dwconv/gen/up1x25-minmax-scalar-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700146 "src/f32-dwconv/gen/up1x25-minmax-scalar.c",
147 "src/f32-dwconv/gen/up1x25-scalar-acc2.c",
148 "src/f32-dwconv/gen/up1x25-scalar.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700149 "src/f32-dwconv/gen/up2x4-minmax-scalar-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700150 "src/f32-dwconv/gen/up2x4-minmax-scalar.c",
151 "src/f32-dwconv/gen/up2x4-scalar-acc2.c",
152 "src/f32-dwconv/gen/up2x4-scalar.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700153 "src/f32-dwconv/gen/up2x9-minmax-scalar-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700154 "src/f32-dwconv/gen/up2x9-minmax-scalar.c",
155 "src/f32-dwconv/gen/up2x9-scalar-acc2.c",
156 "src/f32-dwconv/gen/up2x9-scalar.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700157 "src/f32-dwconv/gen/up2x25-minmax-scalar-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700158 "src/f32-dwconv/gen/up2x25-minmax-scalar.c",
159 "src/f32-dwconv/gen/up2x25-scalar-acc2.c",
160 "src/f32-dwconv/gen/up2x25-scalar.c",
Frank Barchard35db7d02020-10-26 13:37:34 -0700161 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-scalar-1x1-acc2.c",
162 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-scalar-1x1-acc3.c",
163 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-scalar-1x1-acc4.c",
Marat Dukhan91249d22020-10-24 12:02:51 -0700164 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-scalar-1x1.c",
Frank Barchard35db7d02020-10-26 13:37:34 -0700165 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-scalar-2x1-acc2.c",
Marat Dukhan91249d22020-10-24 12:02:51 -0700166 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-scalar-2x1.c",
167 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-scalar-3x1.c",
168 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-scalar-4x1.c",
169 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-scalar-5x1.c",
170 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-scalar-6x1.c",
Marat Dukhancf5b3c32020-10-25 19:21:10 -0700171 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-scalar-1x1-acc2.c",
172 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-scalar-1x1-acc3.c",
173 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-scalar-1x1-acc4.c",
Frank Barchard35db7d02020-10-26 13:37:34 -0700174 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-scalar-1x1.c",
Marat Dukhancf5b3c32020-10-25 19:21:10 -0700175 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-scalar-2x1-acc2.c",
Frank Barchard35db7d02020-10-26 13:37:34 -0700176 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-scalar-2x1.c",
177 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-scalar-3x1.c",
178 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-scalar-4x1.c",
Marat Dukhanc4efb002020-10-25 23:14:47 -0700179 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-scalar-1x1-acc2.c",
180 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-scalar-1x1-acc3.c",
181 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-scalar-1x1-acc4.c",
182 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-scalar-1x1-acc5.c",
Frank Barchard35db7d02020-10-26 13:37:34 -0700183 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-scalar-1x1.c",
Marat Dukhanc4efb002020-10-25 23:14:47 -0700184 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-scalar-2x1-acc2.c",
185 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-scalar-2x1-acc3.c",
Frank Barchard35db7d02020-10-26 13:37:34 -0700186 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-scalar-2x1.c",
Marat Dukhanc4efb002020-10-25 23:14:47 -0700187 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-scalar-3x1-acc2.c",
Frank Barchard35db7d02020-10-26 13:37:34 -0700188 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-scalar-3x1.c",
Marat Dukhan29c0c332020-10-28 22:11:00 -0700189 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-scalar-1x1-acc2.c",
190 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-scalar-1x1-acc3.c",
191 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-scalar-1x1-acc4.c",
192 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-scalar-1x1-acc5.c",
193 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-scalar-1x1.c",
194 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-scalar-2x1-acc2.c",
195 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-scalar-2x1-acc3.c",
196 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-scalar-2x1.c",
197 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-scalar-3x1-acc2.c",
198 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-scalar-3x1.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -0700199 "src/f32-gavgpool-cw/scalar-x1.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700200 "src/f32-gavgpool/7p7x-minmax-scalar-c1.c",
201 "src/f32-gavgpool/7x-minmax-scalar-c1.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700202 "src/f32-gemm/gen-inc/1x4inc-minmax-scalar.c",
203 "src/f32-gemm/gen-inc/2x4inc-minmax-scalar.c",
204 "src/f32-gemm/gen-inc/4x4inc-minmax-scalar.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700205 "src/f32-gemm/gen/1x4-minmax-scalar.c",
206 "src/f32-gemm/gen/1x4-relu-scalar.c",
207 "src/f32-gemm/gen/1x4-scalar.c",
208 "src/f32-gemm/gen/2x4-minmax-scalar.c",
209 "src/f32-gemm/gen/2x4-relu-scalar.c",
210 "src/f32-gemm/gen/2x4-scalar.c",
211 "src/f32-gemm/gen/4x2-minmax-scalar.c",
212 "src/f32-gemm/gen/4x2-relu-scalar.c",
213 "src/f32-gemm/gen/4x2-scalar.c",
214 "src/f32-gemm/gen/4x4-minmax-scalar.c",
215 "src/f32-gemm/gen/4x4-relu-scalar.c",
216 "src/f32-gemm/gen/4x4-scalar.c",
Marat Dukhanefaac272020-07-16 16:58:07 -0700217 "src/f32-hswish/gen/hswish-scalar-x1.c",
218 "src/f32-hswish/gen/hswish-scalar-x2.c",
219 "src/f32-hswish/gen/hswish-scalar-x4.c",
XNNPACK Team21432672020-10-19 19:58:48 -0700220 "src/f32-ibilinear-chw/gen/scalar-p1.c",
221 "src/f32-ibilinear-chw/gen/scalar-p2.c",
222 "src/f32-ibilinear-chw/gen/scalar-p4.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700223 "src/f32-ibilinear/gen/scalar-c1.c",
224 "src/f32-ibilinear/gen/scalar-c2.c",
225 "src/f32-ibilinear/gen/scalar-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700226 "src/f32-igemm/gen/1x4-minmax-scalar.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700227 "src/f32-igemm/gen/1x4-relu-scalar.c",
228 "src/f32-igemm/gen/1x4-scalar.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700229 "src/f32-igemm/gen/2x4-minmax-scalar.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700230 "src/f32-igemm/gen/2x4-relu-scalar.c",
231 "src/f32-igemm/gen/2x4-scalar.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700232 "src/f32-igemm/gen/4x2-minmax-scalar.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700233 "src/f32-igemm/gen/4x2-relu-scalar.c",
234 "src/f32-igemm/gen/4x2-scalar.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700235 "src/f32-igemm/gen/4x4-minmax-scalar.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700236 "src/f32-igemm/gen/4x4-relu-scalar.c",
237 "src/f32-igemm/gen/4x4-scalar.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700238 "src/f32-maxpool/9p8x-minmax-scalar-c1.c",
239 "src/f32-pavgpool/9p8x-minmax-scalar-c1.c",
240 "src/f32-pavgpool/9x-minmax-scalar-c1.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700241 "src/f32-ppmm/gen/2x4-minmax-scalar.c",
242 "src/f32-ppmm/gen/3x3-minmax-scalar.c",
243 "src/f32-ppmm/gen/4x2-minmax-scalar.c",
244 "src/f32-ppmm/gen/4x4-minmax-scalar.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800245 "src/f32-prelu/gen/scalar-2x1.c",
246 "src/f32-prelu/gen/scalar-2x4.c",
Marat Dukhanf46f6752020-01-21 11:03:49 -0800247 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x1.c",
Marat Dukhanf46f6752020-01-21 11:03:49 -0800248 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x2-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700249 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x2.c",
Marat Dukhanf46f6752020-01-21 11:03:49 -0800250 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x4-acc2.c",
251 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x4-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700252 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x4.c",
Marat Dukhanf46f6752020-01-21 11:03:49 -0800253 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x1.c",
Marat Dukhanf46f6752020-01-21 11:03:49 -0800254 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x2-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700255 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x2.c",
Marat Dukhanf46f6752020-01-21 11:03:49 -0800256 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x4-acc2.c",
257 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x4-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700258 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x4.c",
Frank Barchardfb158e22020-07-15 16:10:10 -0700259 "src/f32-relu/gen/scalar-x1.c",
260 "src/f32-relu/gen/scalar-x2.c",
261 "src/f32-relu/gen/scalar-x4.c",
Frank Barchard37297a62020-08-27 12:15:10 -0700262 "src/f32-relu/gen/scalar-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700263 "src/f32-rmax/scalar.c",
Marat Dukhan3a77ea72019-12-23 12:10:24 -0800264 "src/f32-sigmoid/gen/scalar-lut64-p2-div-x1.c",
265 "src/f32-sigmoid/gen/scalar-lut64-p2-div-x2.c",
266 "src/f32-sigmoid/gen/scalar-lut64-p2-div-x4.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700267 "src/f32-sigmoid/gen/scalar-lut2048-p1-div-x1.c",
268 "src/f32-sigmoid/gen/scalar-lut2048-p1-div-x2.c",
269 "src/f32-sigmoid/gen/scalar-lut2048-p1-div-x4.c",
Marat Dukhan3a77ea72019-12-23 12:10:24 -0800270 "src/f32-sigmoid/gen/scalar-p5-div-x1.c",
271 "src/f32-sigmoid/gen/scalar-p5-div-x2.c",
272 "src/f32-sigmoid/gen/scalar-p5-div-x4.c",
Marat Dukhan355ab432020-04-09 19:01:52 -0700273 "src/f32-spmm/gen/1x1-minmax-scalar-pipelined.c",
274 "src/f32-spmm/gen/1x1-minmax-scalar.c",
275 "src/f32-spmm/gen/2x1-minmax-scalar-pipelined.c",
276 "src/f32-spmm/gen/2x1-minmax-scalar.c",
277 "src/f32-spmm/gen/4x1-minmax-scalar-pipelined.c",
278 "src/f32-spmm/gen/4x1-minmax-scalar.c",
279 "src/f32-spmm/gen/8x1-minmax-scalar-pipelined.c",
280 "src/f32-spmm/gen/8x1-minmax-scalar.c",
281 "src/f32-spmm/gen/8x2-minmax-scalar.c",
282 "src/f32-spmm/gen/8x4-minmax-scalar.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700283 "src/f32-vbinary/gen/vadd-minmax-scalar-x1.c",
284 "src/f32-vbinary/gen/vadd-minmax-scalar-x2.c",
285 "src/f32-vbinary/gen/vadd-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700286 "src/f32-vbinary/gen/vadd-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700287 "src/f32-vbinary/gen/vadd-relu-scalar-x1.c",
288 "src/f32-vbinary/gen/vadd-relu-scalar-x2.c",
289 "src/f32-vbinary/gen/vadd-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700290 "src/f32-vbinary/gen/vadd-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700291 "src/f32-vbinary/gen/vadd-scalar-x1.c",
292 "src/f32-vbinary/gen/vadd-scalar-x2.c",
293 "src/f32-vbinary/gen/vadd-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700294 "src/f32-vbinary/gen/vadd-scalar-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700295 "src/f32-vbinary/gen/vaddc-minmax-scalar-x1.c",
296 "src/f32-vbinary/gen/vaddc-minmax-scalar-x2.c",
297 "src/f32-vbinary/gen/vaddc-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700298 "src/f32-vbinary/gen/vaddc-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700299 "src/f32-vbinary/gen/vaddc-relu-scalar-x1.c",
300 "src/f32-vbinary/gen/vaddc-relu-scalar-x2.c",
301 "src/f32-vbinary/gen/vaddc-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700302 "src/f32-vbinary/gen/vaddc-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700303 "src/f32-vbinary/gen/vaddc-scalar-x1.c",
304 "src/f32-vbinary/gen/vaddc-scalar-x2.c",
305 "src/f32-vbinary/gen/vaddc-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700306 "src/f32-vbinary/gen/vaddc-scalar-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700307 "src/f32-vbinary/gen/vdiv-minmax-scalar-x1.c",
308 "src/f32-vbinary/gen/vdiv-minmax-scalar-x2.c",
309 "src/f32-vbinary/gen/vdiv-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700310 "src/f32-vbinary/gen/vdiv-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700311 "src/f32-vbinary/gen/vdiv-relu-scalar-x1.c",
312 "src/f32-vbinary/gen/vdiv-relu-scalar-x2.c",
313 "src/f32-vbinary/gen/vdiv-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700314 "src/f32-vbinary/gen/vdiv-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700315 "src/f32-vbinary/gen/vdiv-scalar-x1.c",
316 "src/f32-vbinary/gen/vdiv-scalar-x2.c",
317 "src/f32-vbinary/gen/vdiv-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700318 "src/f32-vbinary/gen/vdiv-scalar-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700319 "src/f32-vbinary/gen/vdivc-minmax-scalar-x1.c",
320 "src/f32-vbinary/gen/vdivc-minmax-scalar-x2.c",
321 "src/f32-vbinary/gen/vdivc-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700322 "src/f32-vbinary/gen/vdivc-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700323 "src/f32-vbinary/gen/vdivc-relu-scalar-x1.c",
324 "src/f32-vbinary/gen/vdivc-relu-scalar-x2.c",
325 "src/f32-vbinary/gen/vdivc-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700326 "src/f32-vbinary/gen/vdivc-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700327 "src/f32-vbinary/gen/vdivc-scalar-x1.c",
328 "src/f32-vbinary/gen/vdivc-scalar-x2.c",
329 "src/f32-vbinary/gen/vdivc-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700330 "src/f32-vbinary/gen/vdivc-scalar-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800331 "src/f32-vbinary/gen/vmax-scalar-x1.c",
332 "src/f32-vbinary/gen/vmax-scalar-x2.c",
333 "src/f32-vbinary/gen/vmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700334 "src/f32-vbinary/gen/vmax-scalar-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800335 "src/f32-vbinary/gen/vmaxc-scalar-x1.c",
336 "src/f32-vbinary/gen/vmaxc-scalar-x2.c",
337 "src/f32-vbinary/gen/vmaxc-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700338 "src/f32-vbinary/gen/vmaxc-scalar-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800339 "src/f32-vbinary/gen/vmin-scalar-x1.c",
340 "src/f32-vbinary/gen/vmin-scalar-x2.c",
341 "src/f32-vbinary/gen/vmin-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700342 "src/f32-vbinary/gen/vmin-scalar-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800343 "src/f32-vbinary/gen/vminc-scalar-x1.c",
344 "src/f32-vbinary/gen/vminc-scalar-x2.c",
345 "src/f32-vbinary/gen/vminc-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700346 "src/f32-vbinary/gen/vminc-scalar-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700347 "src/f32-vbinary/gen/vmul-minmax-scalar-x1.c",
348 "src/f32-vbinary/gen/vmul-minmax-scalar-x2.c",
349 "src/f32-vbinary/gen/vmul-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700350 "src/f32-vbinary/gen/vmul-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700351 "src/f32-vbinary/gen/vmul-relu-scalar-x1.c",
352 "src/f32-vbinary/gen/vmul-relu-scalar-x2.c",
353 "src/f32-vbinary/gen/vmul-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700354 "src/f32-vbinary/gen/vmul-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700355 "src/f32-vbinary/gen/vmul-scalar-x1.c",
356 "src/f32-vbinary/gen/vmul-scalar-x2.c",
357 "src/f32-vbinary/gen/vmul-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700358 "src/f32-vbinary/gen/vmul-scalar-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700359 "src/f32-vbinary/gen/vmulc-minmax-scalar-x1.c",
360 "src/f32-vbinary/gen/vmulc-minmax-scalar-x2.c",
361 "src/f32-vbinary/gen/vmulc-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700362 "src/f32-vbinary/gen/vmulc-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700363 "src/f32-vbinary/gen/vmulc-relu-scalar-x1.c",
364 "src/f32-vbinary/gen/vmulc-relu-scalar-x2.c",
365 "src/f32-vbinary/gen/vmulc-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700366 "src/f32-vbinary/gen/vmulc-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700367 "src/f32-vbinary/gen/vmulc-scalar-x1.c",
368 "src/f32-vbinary/gen/vmulc-scalar-x2.c",
369 "src/f32-vbinary/gen/vmulc-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700370 "src/f32-vbinary/gen/vmulc-scalar-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700371 "src/f32-vbinary/gen/vrdivc-minmax-scalar-x1.c",
372 "src/f32-vbinary/gen/vrdivc-minmax-scalar-x2.c",
373 "src/f32-vbinary/gen/vrdivc-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700374 "src/f32-vbinary/gen/vrdivc-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700375 "src/f32-vbinary/gen/vrdivc-relu-scalar-x1.c",
376 "src/f32-vbinary/gen/vrdivc-relu-scalar-x2.c",
377 "src/f32-vbinary/gen/vrdivc-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700378 "src/f32-vbinary/gen/vrdivc-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700379 "src/f32-vbinary/gen/vrdivc-scalar-x1.c",
380 "src/f32-vbinary/gen/vrdivc-scalar-x2.c",
381 "src/f32-vbinary/gen/vrdivc-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700382 "src/f32-vbinary/gen/vrdivc-scalar-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700383 "src/f32-vbinary/gen/vrsubc-minmax-scalar-x1.c",
384 "src/f32-vbinary/gen/vrsubc-minmax-scalar-x2.c",
385 "src/f32-vbinary/gen/vrsubc-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700386 "src/f32-vbinary/gen/vrsubc-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700387 "src/f32-vbinary/gen/vrsubc-relu-scalar-x1.c",
388 "src/f32-vbinary/gen/vrsubc-relu-scalar-x2.c",
389 "src/f32-vbinary/gen/vrsubc-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700390 "src/f32-vbinary/gen/vrsubc-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700391 "src/f32-vbinary/gen/vrsubc-scalar-x1.c",
392 "src/f32-vbinary/gen/vrsubc-scalar-x2.c",
393 "src/f32-vbinary/gen/vrsubc-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700394 "src/f32-vbinary/gen/vrsubc-scalar-x8.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -0700395 "src/f32-vbinary/gen/vsqrdiff-scalar-x1.c",
396 "src/f32-vbinary/gen/vsqrdiff-scalar-x2.c",
397 "src/f32-vbinary/gen/vsqrdiff-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700398 "src/f32-vbinary/gen/vsqrdiff-scalar-x8.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -0700399 "src/f32-vbinary/gen/vsqrdiffc-scalar-x1.c",
400 "src/f32-vbinary/gen/vsqrdiffc-scalar-x2.c",
401 "src/f32-vbinary/gen/vsqrdiffc-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700402 "src/f32-vbinary/gen/vsqrdiffc-scalar-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700403 "src/f32-vbinary/gen/vsub-minmax-scalar-x1.c",
404 "src/f32-vbinary/gen/vsub-minmax-scalar-x2.c",
405 "src/f32-vbinary/gen/vsub-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700406 "src/f32-vbinary/gen/vsub-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700407 "src/f32-vbinary/gen/vsub-relu-scalar-x1.c",
408 "src/f32-vbinary/gen/vsub-relu-scalar-x2.c",
409 "src/f32-vbinary/gen/vsub-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700410 "src/f32-vbinary/gen/vsub-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700411 "src/f32-vbinary/gen/vsub-scalar-x1.c",
412 "src/f32-vbinary/gen/vsub-scalar-x2.c",
413 "src/f32-vbinary/gen/vsub-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700414 "src/f32-vbinary/gen/vsub-scalar-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700415 "src/f32-vbinary/gen/vsubc-minmax-scalar-x1.c",
416 "src/f32-vbinary/gen/vsubc-minmax-scalar-x2.c",
417 "src/f32-vbinary/gen/vsubc-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700418 "src/f32-vbinary/gen/vsubc-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700419 "src/f32-vbinary/gen/vsubc-relu-scalar-x1.c",
420 "src/f32-vbinary/gen/vsubc-relu-scalar-x2.c",
421 "src/f32-vbinary/gen/vsubc-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700422 "src/f32-vbinary/gen/vsubc-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700423 "src/f32-vbinary/gen/vsubc-scalar-x1.c",
424 "src/f32-vbinary/gen/vsubc-scalar-x2.c",
425 "src/f32-vbinary/gen/vsubc-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700426 "src/f32-vbinary/gen/vsubc-scalar-x8.c",
Marat Dukhaned6baaf2020-12-01 15:07:08 -0800427 "src/f32-velu/gen/velu-scalar-rr2-lut16-p3-x1.c",
428 "src/f32-velu/gen/velu-scalar-rr2-lut16-p3-x2.c",
429 "src/f32-velu/gen/velu-scalar-rr2-lut16-p3-x3.c",
430 "src/f32-velu/gen/velu-scalar-rr2-lut16-p3-x4.c",
431 "src/f32-velu/gen/velu-scalar-rr2-lut16-p3-x5.c",
432 "src/f32-velu/gen/velu-scalar-rr2-lut16-p3-x6.c",
433 "src/f32-velu/gen/velu-scalar-rr2-p6-x1.c",
434 "src/f32-velu/gen/velu-scalar-rr2-p6-x2.c",
435 "src/f32-velu/gen/velu-scalar-rr2-p6-x3.c",
436 "src/f32-velu/gen/velu-scalar-rr2-p6-x4.c",
437 "src/f32-velu/gen/velu-scalar-rr2-p6-x5.c",
438 "src/f32-velu/gen/velu-scalar-rr2-p6-x6.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700439 "src/f32-vlrelu/gen/vlrelu-scalar-x1.c",
440 "src/f32-vlrelu/gen/vlrelu-scalar-x2.c",
441 "src/f32-vlrelu/gen/vlrelu-scalar-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700442 "src/f32-vmulcaddc/gen/c1-minmax-scalar-2x.c",
443 "src/f32-vmulcaddc/gen/c2-minmax-scalar-2x.c",
444 "src/f32-vmulcaddc/gen/c4-minmax-scalar-2x.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -0700445 "src/f32-vrnd/gen/vrndd-scalar-libm-x1.c",
446 "src/f32-vrnd/gen/vrndd-scalar-libm-x2.c",
447 "src/f32-vrnd/gen/vrndd-scalar-libm-x4.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700448 "src/f32-vrnd/gen/vrndne-scalar-libm-x1.c",
449 "src/f32-vrnd/gen/vrndne-scalar-libm-x2.c",
450 "src/f32-vrnd/gen/vrndne-scalar-libm-x4.c",
451 "src/f32-vrnd/gen/vrndu-scalar-libm-x1.c",
452 "src/f32-vrnd/gen/vrndu-scalar-libm-x2.c",
453 "src/f32-vrnd/gen/vrndu-scalar-libm-x4.c",
454 "src/f32-vrnd/gen/vrndz-scalar-libm-x1.c",
455 "src/f32-vrnd/gen/vrndz-scalar-libm-x2.c",
456 "src/f32-vrnd/gen/vrndz-scalar-libm-x4.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -0700457 "src/f32-vsqrt/gen/scalar-sqrt-x1.c",
458 "src/f32-vsqrt/gen/scalar-sqrt-x2.c",
459 "src/f32-vsqrt/gen/scalar-sqrt-x4.c",
Marat Dukhan5020b962020-06-08 13:30:10 -0700460 "src/f32-vunary/gen/vabs-scalar-x1.c",
461 "src/f32-vunary/gen/vabs-scalar-x2.c",
462 "src/f32-vunary/gen/vabs-scalar-x4.c",
463 "src/f32-vunary/gen/vneg-scalar-x1.c",
464 "src/f32-vunary/gen/vneg-scalar-x2.c",
465 "src/f32-vunary/gen/vneg-scalar-x4.c",
466 "src/f32-vunary/gen/vsqr-scalar-x1.c",
467 "src/f32-vunary/gen/vsqr-scalar-x2.c",
468 "src/f32-vunary/gen/vsqr-scalar-x4.c",
Marat Dukhande390d42020-11-29 19:32:18 -0800469 "src/math/expm1minus-scalar-rr2-lut4-p4.c",
470 "src/math/expm1minus-scalar-rr2-lut8-p3.c",
471 "src/math/expm1minus-scalar-rr2-lut8-p4.c",
Marat Dukhanc60742b2020-11-23 12:33:27 -0800472 "src/math/expm1minus-scalar-rr2-lut16-p3.c",
473 "src/math/expm1minus-scalar-rr2-lut16-p4.c",
474 "src/math/expm1minus-scalar-rr2-p5.c",
475 "src/math/expm1minus-scalar-rr2-p6.c",
Frank Barchard22136062020-11-24 18:44:46 -0800476 "src/math/expminus-scalar-rr2-lut64-p2.c",
477 "src/math/expminus-scalar-rr2-lut2048-p1.c",
478 "src/math/expminus-scalar-rr2-p5.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -0700479 "src/math/roundd-scalar-addsub.c",
480 "src/math/roundd-scalar-cvt.c",
Marat Dukhanffbf96a2020-05-14 02:59:08 -0700481 "src/math/roundd-scalar-floor.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700482 "src/math/roundne-scalar-addsub.c",
483 "src/math/roundne-scalar-nearbyint.c",
484 "src/math/roundne-scalar-rint.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -0700485 "src/math/roundu-scalar-addsub.c",
Marat Dukhanffbf96a2020-05-14 02:59:08 -0700486 "src/math/roundu-scalar-ceil.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -0700487 "src/math/roundu-scalar-cvt.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -0700488 "src/math/roundz-scalar-addsub.c",
489 "src/math/roundz-scalar-cvt.c",
Marat Dukhanffbf96a2020-05-14 02:59:08 -0700490 "src/math/roundz-scalar-trunc.c",
Marat Dukhanf8475d62020-09-17 15:01:43 -0700491 "src/math/sigmoid-scalar-rr2-lut64-p2-div.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700492 "src/math/sigmoid-scalar-rr2-lut2048-p1-div.c",
Marat Dukhanf8475d62020-09-17 15:01:43 -0700493 "src/math/sigmoid-scalar-rr2-p5-div.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700494 "src/qs8-gemm/gen/8x8c4-minmax-scalar.c",
495 "src/qs8-gemm/gen/12x4c4-minmax-scalar.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -0700496 "src/qs8-requantization/fp32-scalar-lrintf.c",
497 "src/qs8-requantization/fp32-scalar-magic.c",
498 "src/qs8-requantization/precise-scalar-signed64.c",
499 "src/qs8-requantization/precise-scalar-unsigned32.c",
500 "src/qs8-requantization/precise-scalar-unsigned64.c",
501 "src/qs8-requantization/q31-scalar.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -0700502 "src/qu8-avgpool/9p8x-minmax-scalar-c1.c",
503 "src/qu8-avgpool/9x-minmax-scalar-c1.c",
504 "src/qu8-dwconv/up1x9-minmax-scalar.c",
505 "src/qu8-gavgpool/7p7x-minmax-scalar-c1.c",
506 "src/qu8-gavgpool/7x-minmax-scalar-c1.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700507 "src/qu8-gemm/2x2-minmax-scalar.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700508 "src/qu8-gemm/gen/8x8c4-minmax-scalar.c",
509 "src/qu8-gemm/gen/12x4c4-minmax-scalar.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -0700510 "src/qu8-igemm/2x2-minmax-scalar.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -0700511 "src/qu8-requantization/fp32-scalar-lrintf.c",
512 "src/qu8-requantization/fp32-scalar-magic.c",
513 "src/qu8-requantization/precise-scalar-signed64.c",
514 "src/qu8-requantization/precise-scalar-unsigned32.c",
515 "src/qu8-requantization/precise-scalar-unsigned64.c",
516 "src/qu8-requantization/q31-scalar.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -0700517 "src/qu8-vadd/minmax-scalar.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -0700518 "src/u8-clamp/scalar-x4.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700519 "src/u8-lut32norm/scalar.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700520 "src/u8-maxpool/9p8x-minmax-scalar-c1.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700521 "src/u8-rmax/scalar.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700522 "src/x8-lut/scalar.c",
523 "src/x8-zip/x2-scalar.c",
524 "src/x8-zip/x3-scalar.c",
525 "src/x8-zip/x4-scalar.c",
526 "src/x8-zip/xm-scalar.c",
Marat Dukhanad71b9a2020-11-20 00:01:51 -0800527 "src/x32-depthtospace2d-chw2hwc/scalar.c",
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -0700528 "src/x32-fill/scalar-float.c",
529 "src/x32-fill/scalar-int.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700530 "src/x32-packx/x2-scalar.c",
531 "src/x32-packx/x3-scalar.c",
532 "src/x32-packx/x4-scalar.c",
Marat Dukhan63523d42020-05-22 17:07:33 -0700533 "src/x32-pad/scalar-float.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700534 "src/x32-pad/scalar-int.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700535 "src/x32-unpool/scalar.c",
536 "src/x32-zip/x2-scalar.c",
537 "src/x32-zip/x3-scalar.c",
538 "src/x32-zip/x4-scalar.c",
539 "src/x32-zip/xm-scalar.c",
Marat Dukhan048931b2020-11-24 20:53:54 -0800540 "src/xx-copy/memcpy.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700541]
542
Marat Dukhan436ebe62019-12-04 15:10:12 -0800543WASM_UKERNELS = [
Marat Dukhan99936602020-04-11 16:47:01 -0700544 "src/f32-avgpool/9p8x-minmax-wasm-c1.c",
545 "src/f32-avgpool/9x-minmax-wasm-c1.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -0700546 "src/f32-clamp/gen/wasm-x1.c",
547 "src/f32-clamp/gen/wasm-x2.c",
548 "src/f32-clamp/gen/wasm-x4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700549 "src/f32-dwconv/gen/up1x4-minmax-wasm-acc2.c",
550 "src/f32-dwconv/gen/up1x4-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700551 "src/f32-dwconv/gen/up1x4-wasm-acc2.c",
552 "src/f32-dwconv/gen/up1x4-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700553 "src/f32-dwconv/gen/up1x9-minmax-wasm-acc2.c",
554 "src/f32-dwconv/gen/up1x9-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700555 "src/f32-dwconv/gen/up1x9-wasm-acc2.c",
556 "src/f32-dwconv/gen/up1x9-wasm.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700557 "src/f32-dwconv/gen/up1x25-minmax-wasm-acc2.c",
558 "src/f32-dwconv/gen/up1x25-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700559 "src/f32-dwconv/gen/up1x25-wasm-acc2.c",
560 "src/f32-dwconv/gen/up1x25-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700561 "src/f32-dwconv/gen/up2x4-minmax-wasm-acc2.c",
562 "src/f32-dwconv/gen/up2x4-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700563 "src/f32-dwconv/gen/up2x4-wasm-acc2.c",
564 "src/f32-dwconv/gen/up2x4-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700565 "src/f32-dwconv/gen/up2x9-minmax-wasm-acc2.c",
566 "src/f32-dwconv/gen/up2x9-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700567 "src/f32-dwconv/gen/up2x9-wasm-acc2.c",
568 "src/f32-dwconv/gen/up2x9-wasm.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700569 "src/f32-dwconv/gen/up2x25-minmax-wasm-acc2.c",
570 "src/f32-dwconv/gen/up2x25-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700571 "src/f32-dwconv/gen/up2x25-wasm-acc2.c",
572 "src/f32-dwconv/gen/up2x25-wasm.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700573 "src/f32-gavgpool/7p7x-minmax-wasm-c1.c",
574 "src/f32-gavgpool/7x-minmax-wasm-c1.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700575 "src/f32-gemm/gen-inc/1x4inc-minmax-wasm.c",
576 "src/f32-gemm/gen-inc/2x4inc-minmax-wasm.c",
577 "src/f32-gemm/gen-inc/4x4inc-minmax-wasm.c",
578 "src/f32-gemm/gen/1x4-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700579 "src/f32-gemm/gen/1x4-relu-wasm.c",
580 "src/f32-gemm/gen/1x4-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700581 "src/f32-gemm/gen/2x4-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700582 "src/f32-gemm/gen/2x4-relu-wasm.c",
583 "src/f32-gemm/gen/2x4-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700584 "src/f32-gemm/gen/4x2-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700585 "src/f32-gemm/gen/4x2-relu-wasm.c",
586 "src/f32-gemm/gen/4x2-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700587 "src/f32-gemm/gen/4x4-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700588 "src/f32-gemm/gen/4x4-relu-wasm.c",
589 "src/f32-gemm/gen/4x4-wasm.c",
Marat Dukhanefaac272020-07-16 16:58:07 -0700590 "src/f32-hswish/gen/hswish-wasm-x1.c",
591 "src/f32-hswish/gen/hswish-wasm-x2.c",
592 "src/f32-hswish/gen/hswish-wasm-x4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700593 "src/f32-igemm/gen/1x4-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700594 "src/f32-igemm/gen/1x4-relu-wasm.c",
595 "src/f32-igemm/gen/1x4-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700596 "src/f32-igemm/gen/2x4-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700597 "src/f32-igemm/gen/2x4-relu-wasm.c",
598 "src/f32-igemm/gen/2x4-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700599 "src/f32-igemm/gen/4x2-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700600 "src/f32-igemm/gen/4x2-relu-wasm.c",
601 "src/f32-igemm/gen/4x2-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700602 "src/f32-igemm/gen/4x4-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700603 "src/f32-igemm/gen/4x4-relu-wasm.c",
604 "src/f32-igemm/gen/4x4-wasm.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700605 "src/f32-maxpool/9p8x-minmax-wasm-c1.c",
606 "src/f32-pavgpool/9p8x-minmax-wasm-c1.c",
607 "src/f32-pavgpool/9x-minmax-wasm-c1.c",
Marat Dukhan7c1f8082020-06-25 13:26:20 -0700608 "src/f32-prelu/gen/wasm-2x1.c",
609 "src/f32-prelu/gen/wasm-2x4.c",
Frank Barchardfb158e22020-07-15 16:10:10 -0700610 "src/f32-relu/gen/wasm-x1.c",
611 "src/f32-relu/gen/wasm-x2.c",
612 "src/f32-relu/gen/wasm-x4.c",
Frank Barchard37297a62020-08-27 12:15:10 -0700613 "src/f32-relu/gen/wasm-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700614 "src/f32-vbinary/gen/vadd-minmax-wasm-x1.c",
615 "src/f32-vbinary/gen/vadd-minmax-wasm-x2.c",
616 "src/f32-vbinary/gen/vadd-minmax-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700617 "src/f32-vbinary/gen/vadd-minmax-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700618 "src/f32-vbinary/gen/vadd-relu-wasm-x1.c",
619 "src/f32-vbinary/gen/vadd-relu-wasm-x2.c",
620 "src/f32-vbinary/gen/vadd-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700621 "src/f32-vbinary/gen/vadd-relu-wasm-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700622 "src/f32-vbinary/gen/vaddc-minmax-wasm-x1.c",
623 "src/f32-vbinary/gen/vaddc-minmax-wasm-x2.c",
624 "src/f32-vbinary/gen/vaddc-minmax-wasm-x4.c",
625 "src/f32-vbinary/gen/vaddc-minmax-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700626 "src/f32-vbinary/gen/vaddc-relu-wasm-x1.c",
627 "src/f32-vbinary/gen/vaddc-relu-wasm-x2.c",
628 "src/f32-vbinary/gen/vaddc-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700629 "src/f32-vbinary/gen/vaddc-relu-wasm-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700630 "src/f32-vbinary/gen/vdiv-minmax-wasm-x1.c",
631 "src/f32-vbinary/gen/vdiv-minmax-wasm-x2.c",
632 "src/f32-vbinary/gen/vdiv-minmax-wasm-x4.c",
633 "src/f32-vbinary/gen/vdiv-minmax-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700634 "src/f32-vbinary/gen/vdiv-relu-wasm-x1.c",
635 "src/f32-vbinary/gen/vdiv-relu-wasm-x2.c",
636 "src/f32-vbinary/gen/vdiv-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700637 "src/f32-vbinary/gen/vdiv-relu-wasm-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700638 "src/f32-vbinary/gen/vdivc-minmax-wasm-x1.c",
639 "src/f32-vbinary/gen/vdivc-minmax-wasm-x2.c",
640 "src/f32-vbinary/gen/vdivc-minmax-wasm-x4.c",
641 "src/f32-vbinary/gen/vdivc-minmax-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700642 "src/f32-vbinary/gen/vdivc-relu-wasm-x1.c",
643 "src/f32-vbinary/gen/vdivc-relu-wasm-x2.c",
644 "src/f32-vbinary/gen/vdivc-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700645 "src/f32-vbinary/gen/vdivc-relu-wasm-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800646 "src/f32-vbinary/gen/vmax-wasm-x1.c",
647 "src/f32-vbinary/gen/vmax-wasm-x2.c",
648 "src/f32-vbinary/gen/vmax-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700649 "src/f32-vbinary/gen/vmax-wasm-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800650 "src/f32-vbinary/gen/vmaxc-wasm-x1.c",
651 "src/f32-vbinary/gen/vmaxc-wasm-x2.c",
652 "src/f32-vbinary/gen/vmaxc-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700653 "src/f32-vbinary/gen/vmaxc-wasm-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800654 "src/f32-vbinary/gen/vmin-wasm-x1.c",
655 "src/f32-vbinary/gen/vmin-wasm-x2.c",
656 "src/f32-vbinary/gen/vmin-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700657 "src/f32-vbinary/gen/vmin-wasm-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800658 "src/f32-vbinary/gen/vminc-wasm-x1.c",
659 "src/f32-vbinary/gen/vminc-wasm-x2.c",
660 "src/f32-vbinary/gen/vminc-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700661 "src/f32-vbinary/gen/vminc-wasm-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700662 "src/f32-vbinary/gen/vmul-minmax-wasm-x1.c",
663 "src/f32-vbinary/gen/vmul-minmax-wasm-x2.c",
664 "src/f32-vbinary/gen/vmul-minmax-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700665 "src/f32-vbinary/gen/vmul-minmax-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700666 "src/f32-vbinary/gen/vmul-relu-wasm-x1.c",
667 "src/f32-vbinary/gen/vmul-relu-wasm-x2.c",
668 "src/f32-vbinary/gen/vmul-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700669 "src/f32-vbinary/gen/vmul-relu-wasm-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700670 "src/f32-vbinary/gen/vmulc-minmax-wasm-x1.c",
671 "src/f32-vbinary/gen/vmulc-minmax-wasm-x2.c",
672 "src/f32-vbinary/gen/vmulc-minmax-wasm-x4.c",
673 "src/f32-vbinary/gen/vmulc-minmax-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700674 "src/f32-vbinary/gen/vmulc-relu-wasm-x1.c",
675 "src/f32-vbinary/gen/vmulc-relu-wasm-x2.c",
676 "src/f32-vbinary/gen/vmulc-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700677 "src/f32-vbinary/gen/vmulc-relu-wasm-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700678 "src/f32-vbinary/gen/vrdivc-minmax-wasm-x1.c",
679 "src/f32-vbinary/gen/vrdivc-minmax-wasm-x2.c",
680 "src/f32-vbinary/gen/vrdivc-minmax-wasm-x4.c",
681 "src/f32-vbinary/gen/vrdivc-minmax-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700682 "src/f32-vbinary/gen/vrdivc-relu-wasm-x1.c",
683 "src/f32-vbinary/gen/vrdivc-relu-wasm-x2.c",
684 "src/f32-vbinary/gen/vrdivc-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700685 "src/f32-vbinary/gen/vrdivc-relu-wasm-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700686 "src/f32-vbinary/gen/vrsubc-minmax-wasm-x1.c",
687 "src/f32-vbinary/gen/vrsubc-minmax-wasm-x2.c",
688 "src/f32-vbinary/gen/vrsubc-minmax-wasm-x4.c",
689 "src/f32-vbinary/gen/vrsubc-minmax-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700690 "src/f32-vbinary/gen/vrsubc-relu-wasm-x1.c",
691 "src/f32-vbinary/gen/vrsubc-relu-wasm-x2.c",
692 "src/f32-vbinary/gen/vrsubc-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700693 "src/f32-vbinary/gen/vrsubc-relu-wasm-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700694 "src/f32-vbinary/gen/vsub-minmax-wasm-x1.c",
695 "src/f32-vbinary/gen/vsub-minmax-wasm-x2.c",
696 "src/f32-vbinary/gen/vsub-minmax-wasm-x4.c",
697 "src/f32-vbinary/gen/vsub-minmax-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700698 "src/f32-vbinary/gen/vsub-relu-wasm-x1.c",
699 "src/f32-vbinary/gen/vsub-relu-wasm-x2.c",
700 "src/f32-vbinary/gen/vsub-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700701 "src/f32-vbinary/gen/vsub-relu-wasm-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700702 "src/f32-vbinary/gen/vsubc-minmax-wasm-x1.c",
703 "src/f32-vbinary/gen/vsubc-minmax-wasm-x2.c",
704 "src/f32-vbinary/gen/vsubc-minmax-wasm-x4.c",
705 "src/f32-vbinary/gen/vsubc-minmax-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700706 "src/f32-vbinary/gen/vsubc-relu-wasm-x1.c",
707 "src/f32-vbinary/gen/vsubc-relu-wasm-x2.c",
708 "src/f32-vbinary/gen/vsubc-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700709 "src/f32-vbinary/gen/vsubc-relu-wasm-x8.c",
Marat Dukhaned6baaf2020-12-01 15:07:08 -0800710 "src/f32-velu/gen/velu-wasm-rr2-lut16-p3-x1.c",
711 "src/f32-velu/gen/velu-wasm-rr2-lut16-p3-x2.c",
712 "src/f32-velu/gen/velu-wasm-rr2-lut16-p3-x3.c",
713 "src/f32-velu/gen/velu-wasm-rr2-lut16-p3-x4.c",
714 "src/f32-velu/gen/velu-wasm-rr2-lut16-p3-x5.c",
715 "src/f32-velu/gen/velu-wasm-rr2-lut16-p3-x6.c",
716 "src/f32-velu/gen/velu-wasm-rr2-p6-x1.c",
717 "src/f32-velu/gen/velu-wasm-rr2-p6-x2.c",
718 "src/f32-velu/gen/velu-wasm-rr2-p6-x3.c",
719 "src/f32-velu/gen/velu-wasm-rr2-p6-x4.c",
720 "src/f32-velu/gen/velu-wasm-rr2-p6-x5.c",
721 "src/f32-velu/gen/velu-wasm-rr2-p6-x6.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -0700722 "src/f32-vlrelu/gen/vlrelu-wasm-x1.c",
723 "src/f32-vlrelu/gen/vlrelu-wasm-x2.c",
724 "src/f32-vlrelu/gen/vlrelu-wasm-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700725 "src/f32-vmulcaddc/gen/c1-minmax-wasm-2x.c",
726 "src/f32-vmulcaddc/gen/c2-minmax-wasm-2x.c",
727 "src/f32-vmulcaddc/gen/c4-minmax-wasm-2x.c",
Marat Dukhan436ebe62019-12-04 15:10:12 -0800728]
729
Marat Dukhan290055c2020-06-09 12:24:29 -0700730WASMSIMD_UKERNELS = [
Marat Dukhan40f05522020-07-16 22:33:12 -0700731 "src/f32-argmaxpool/4x-wasmsimd-c4.c",
732 "src/f32-argmaxpool/9p8x-wasmsimd-c4.c",
733 "src/f32-argmaxpool/9x-wasmsimd-c4.c",
Marat Dukhan3b7432d2020-07-16 17:46:32 -0700734 "src/f32-avgpool/9p8x-minmax-wasmsimd-arm-c4.c",
735 "src/f32-avgpool/9p8x-minmax-wasmsimd-x86-c4.c",
736 "src/f32-avgpool/9x-minmax-wasmsimd-arm-c4.c",
737 "src/f32-avgpool/9x-minmax-wasmsimd-x86-c4.c",
Marat Dukhan3fa52c82020-07-08 12:54:33 -0700738 "src/f32-clamp/gen/wasmsimd-arm-x4.c",
739 "src/f32-clamp/gen/wasmsimd-arm-x8.c",
740 "src/f32-clamp/gen/wasmsimd-x86-x4.c",
741 "src/f32-clamp/gen/wasmsimd-x86-x8.c",
Frank Barchard22136062020-11-24 18:44:46 -0800742 "src/f32-conv-hwc2chw/3x3s2p1c3x4-wasmsimd-2x2.c",
Marat Dukhanac014d72020-06-16 08:36:47 -0700743 "src/f32-dwconv/gen/up4x4-minmax-wasmsimd-acc2-arm.c",
Marat Dukhanac014d72020-06-16 08:36:47 -0700744 "src/f32-dwconv/gen/up4x4-minmax-wasmsimd-acc2-x86.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700745 "src/f32-dwconv/gen/up4x4-minmax-wasmsimd-arm.c",
Marat Dukhanac014d72020-06-16 08:36:47 -0700746 "src/f32-dwconv/gen/up4x4-minmax-wasmsimd-x86.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700747 "src/f32-dwconv/gen/up4x4-wasmsimd.c",
748 "src/f32-dwconv/gen/up4x9-minmax-wasmsimd-acc2-arm.c",
Marat Dukhanac014d72020-06-16 08:36:47 -0700749 "src/f32-dwconv/gen/up4x9-minmax-wasmsimd-acc2-x86.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700750 "src/f32-dwconv/gen/up4x9-minmax-wasmsimd-arm.c",
Marat Dukhanac014d72020-06-16 08:36:47 -0700751 "src/f32-dwconv/gen/up4x9-minmax-wasmsimd-x86.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700752 "src/f32-dwconv/gen/up4x9-wasmsimd.c",
753 "src/f32-dwconv/gen/up4x25-minmax-wasmsimd-acc2-arm.c",
754 "src/f32-dwconv/gen/up4x25-minmax-wasmsimd-acc2-x86.c",
755 "src/f32-dwconv/gen/up4x25-minmax-wasmsimd-arm.c",
756 "src/f32-dwconv/gen/up4x25-minmax-wasmsimd-x86.c",
757 "src/f32-dwconv/gen/up4x25-wasmsimd.c",
758 "src/f32-dwconv/gen/up8x4-minmax-wasmsimd-acc2-arm.c",
Marat Dukhanac014d72020-06-16 08:36:47 -0700759 "src/f32-dwconv/gen/up8x4-minmax-wasmsimd-acc2-x86.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700760 "src/f32-dwconv/gen/up8x4-minmax-wasmsimd-arm.c",
Marat Dukhanac014d72020-06-16 08:36:47 -0700761 "src/f32-dwconv/gen/up8x4-minmax-wasmsimd-x86.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700762 "src/f32-dwconv/gen/up8x4-wasmsimd.c",
763 "src/f32-dwconv/gen/up8x9-minmax-wasmsimd-acc2-arm.c",
Marat Dukhanac014d72020-06-16 08:36:47 -0700764 "src/f32-dwconv/gen/up8x9-minmax-wasmsimd-acc2-x86.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700765 "src/f32-dwconv/gen/up8x9-minmax-wasmsimd-arm.c",
Marat Dukhanac014d72020-06-16 08:36:47 -0700766 "src/f32-dwconv/gen/up8x9-minmax-wasmsimd-x86.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700767 "src/f32-dwconv/gen/up8x9-wasmsimd.c",
768 "src/f32-dwconv/gen/up8x25-minmax-wasmsimd-acc2-arm.c",
769 "src/f32-dwconv/gen/up8x25-minmax-wasmsimd-acc2-x86.c",
770 "src/f32-dwconv/gen/up8x25-minmax-wasmsimd-arm.c",
771 "src/f32-dwconv/gen/up8x25-minmax-wasmsimd-x86.c",
772 "src/f32-dwconv/gen/up8x25-wasmsimd.c",
Frank Barchard3b800452020-11-22 12:12:35 -0800773 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-1x4-acc2.c",
774 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-1x4-acc3.c",
775 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-1x4-acc4.c",
776 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-1x4.c",
777 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-2x4-acc2.c",
778 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-2x4.c",
779 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-3x4.c",
780 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-4x4.c",
781 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-5x4.c",
782 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-6x4.c",
783 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-1x4-acc2.c",
784 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-1x4-acc3.c",
785 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-1x4-acc4.c",
786 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-1x4.c",
787 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-2x4-acc2.c",
788 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-2x4.c",
789 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-3x4.c",
790 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-4x4.c",
791 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-5x4.c",
792 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-6x4.c",
Frank Barchard9cef5ea2020-11-18 14:52:08 -0800793 "src/f32-dwconv2d-chw/gen/3x3s2p1-wasmsimd-1x4-acc3-arm.c",
Frank Barcharddb5c32d2020-11-16 23:58:42 -0800794 "src/f32-dwconv2d-chw/gen/3x3s2p1-wasmsimd-1x4-acc3-x86.c",
Frank Barchard20a07412020-11-30 23:30:00 -0800795 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-1x4-acc2.c",
796 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-1x4-acc3.c",
797 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-1x4-acc4.c",
798 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-1x4-acc5.c",
799 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-1x4.c",
800 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-2x4-acc2.c",
801 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-2x4-acc3.c",
802 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-2x4.c",
803 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-3x4-acc2.c",
804 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-3x4.c",
805 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-4x4-acc2.c",
806 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-4x4.c",
807 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-5x4.c",
808 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-1x4-acc2.c",
809 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-1x4-acc3.c",
810 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-1x4-acc4.c",
811 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-1x4-acc5.c",
812 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-1x4.c",
813 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-2x4-acc2.c",
814 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-2x4-acc3.c",
815 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-2x4.c",
816 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-3x4-acc2.c",
817 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-3x4.c",
818 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-4x4-acc2.c",
819 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-4x4.c",
820 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-5x4.c",
Frank Barchard9cef5ea2020-11-18 14:52:08 -0800821 "src/f32-dwconv2d-chw/gen/5x5s2p2-wasmsimd-1x4-acc2-arm.c",
Frank Barcharddb5c32d2020-11-16 23:58:42 -0800822 "src/f32-dwconv2d-chw/gen/5x5s2p2-wasmsimd-1x4-acc2-x86.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700823 "src/f32-gavgpool-cw/wasmsimd-arm-x4.c",
824 "src/f32-gavgpool-cw/wasmsimd-x86-x4.c",
Marat Dukhanc6016802020-07-16 18:51:28 -0700825 "src/f32-gavgpool/7p7x-minmax-wasmsimd-arm-c4.c",
826 "src/f32-gavgpool/7p7x-minmax-wasmsimd-x86-c4.c",
827 "src/f32-gavgpool/7x-minmax-wasmsimd-arm-c4.c",
828 "src/f32-gavgpool/7x-minmax-wasmsimd-x86-c4.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700829 "src/f32-gemm/gen-inc/1x8inc-minmax-wasmsimd-loadsplat-arm.c",
830 "src/f32-gemm/gen-inc/1x8inc-minmax-wasmsimd-loadsplat-x86.c",
831 "src/f32-gemm/gen-inc/1x8inc-minmax-wasmsimd-splat-arm.c",
832 "src/f32-gemm/gen-inc/1x8inc-minmax-wasmsimd-splat-x86.c",
833 "src/f32-gemm/gen-inc/1x8s4inc-minmax-wasmsimd-arm.c",
834 "src/f32-gemm/gen-inc/1x8s4inc-minmax-wasmsimd-x86.c",
835 "src/f32-gemm/gen-inc/3x8inc-minmax-wasmsimd-loadsplat-arm.c",
836 "src/f32-gemm/gen-inc/3x8inc-minmax-wasmsimd-loadsplat-x86.c",
837 "src/f32-gemm/gen-inc/3x8inc-minmax-wasmsimd-splat-arm.c",
838 "src/f32-gemm/gen-inc/3x8inc-minmax-wasmsimd-splat-x86.c",
839 "src/f32-gemm/gen-inc/3x8s4inc-minmax-wasmsimd-arm.c",
840 "src/f32-gemm/gen-inc/3x8s4inc-minmax-wasmsimd-x86.c",
841 "src/f32-gemm/gen-inc/4x8inc-minmax-wasmsimd-loadsplat-arm.c",
842 "src/f32-gemm/gen-inc/4x8inc-minmax-wasmsimd-loadsplat-x86.c",
843 "src/f32-gemm/gen-inc/4x8inc-minmax-wasmsimd-splat-arm.c",
844 "src/f32-gemm/gen-inc/4x8inc-minmax-wasmsimd-splat-x86.c",
845 "src/f32-gemm/gen-inc/4x8s4inc-minmax-wasmsimd-arm.c",
846 "src/f32-gemm/gen-inc/4x8s4inc-minmax-wasmsimd-x86.c",
847 "src/f32-gemm/gen-inc/5x8inc-minmax-wasmsimd-loadsplat-arm.c",
848 "src/f32-gemm/gen-inc/5x8inc-minmax-wasmsimd-loadsplat-x86.c",
849 "src/f32-gemm/gen-inc/5x8inc-minmax-wasmsimd-splat-arm.c",
850 "src/f32-gemm/gen-inc/5x8inc-minmax-wasmsimd-splat-x86.c",
851 "src/f32-gemm/gen-inc/5x8s4inc-minmax-wasmsimd-arm.c",
852 "src/f32-gemm/gen-inc/5x8s4inc-minmax-wasmsimd-x86.c",
853 "src/f32-gemm/gen-inc/6x8inc-minmax-wasmsimd-loadsplat-arm.c",
854 "src/f32-gemm/gen-inc/6x8inc-minmax-wasmsimd-loadsplat-x86.c",
855 "src/f32-gemm/gen-inc/6x8inc-minmax-wasmsimd-splat-arm.c",
856 "src/f32-gemm/gen-inc/6x8inc-minmax-wasmsimd-splat-x86.c",
857 "src/f32-gemm/gen-inc/6x8s4inc-minmax-wasmsimd-arm.c",
858 "src/f32-gemm/gen-inc/6x8s4inc-minmax-wasmsimd-x86.c",
859 "src/f32-gemm/gen/1x8-minmax-wasmsimd-loadsplat-arm.c",
860 "src/f32-gemm/gen/1x8-minmax-wasmsimd-loadsplat-x86.c",
861 "src/f32-gemm/gen/1x8-minmax-wasmsimd-splat-arm.c",
862 "src/f32-gemm/gen/1x8-minmax-wasmsimd-splat-x86.c",
863 "src/f32-gemm/gen/1x8-relu-wasmsimd-splat.c",
864 "src/f32-gemm/gen/1x8-wasmsimd-splat.c",
865 "src/f32-gemm/gen/1x8s4-minmax-wasmsimd-arm.c",
866 "src/f32-gemm/gen/1x8s4-minmax-wasmsimd-x86.c",
867 "src/f32-gemm/gen/3x8-minmax-wasmsimd-loadsplat-arm.c",
868 "src/f32-gemm/gen/3x8-minmax-wasmsimd-loadsplat-x86.c",
869 "src/f32-gemm/gen/3x8-minmax-wasmsimd-splat-arm.c",
870 "src/f32-gemm/gen/3x8-minmax-wasmsimd-splat-x86.c",
871 "src/f32-gemm/gen/3x8s4-minmax-wasmsimd-arm.c",
872 "src/f32-gemm/gen/3x8s4-minmax-wasmsimd-x86.c",
873 "src/f32-gemm/gen/4x2c4-minmax-wasmsimd-arm.c",
874 "src/f32-gemm/gen/4x2c4-minmax-wasmsimd-x86.c",
875 "src/f32-gemm/gen/4x2c4-relu-wasmsimd.c",
876 "src/f32-gemm/gen/4x2c4-wasmsimd.c",
877 "src/f32-gemm/gen/4x8-minmax-wasmsimd-loadsplat-arm.c",
878 "src/f32-gemm/gen/4x8-minmax-wasmsimd-loadsplat-x86.c",
879 "src/f32-gemm/gen/4x8-minmax-wasmsimd-splat-arm.c",
880 "src/f32-gemm/gen/4x8-minmax-wasmsimd-splat-x86.c",
881 "src/f32-gemm/gen/4x8-relu-wasmsimd-splat.c",
882 "src/f32-gemm/gen/4x8-wasmsimd-splat.c",
883 "src/f32-gemm/gen/4x8s4-minmax-wasmsimd-arm.c",
884 "src/f32-gemm/gen/4x8s4-minmax-wasmsimd-x86.c",
885 "src/f32-gemm/gen/5x8-minmax-wasmsimd-loadsplat-arm.c",
886 "src/f32-gemm/gen/5x8-minmax-wasmsimd-loadsplat-x86.c",
887 "src/f32-gemm/gen/5x8-minmax-wasmsimd-splat-arm.c",
888 "src/f32-gemm/gen/5x8-minmax-wasmsimd-splat-x86.c",
889 "src/f32-gemm/gen/5x8-relu-wasmsimd-splat.c",
890 "src/f32-gemm/gen/5x8-wasmsimd-splat.c",
891 "src/f32-gemm/gen/5x8s4-minmax-wasmsimd-arm.c",
892 "src/f32-gemm/gen/5x8s4-minmax-wasmsimd-x86.c",
893 "src/f32-gemm/gen/6x8-minmax-wasmsimd-loadsplat-arm.c",
894 "src/f32-gemm/gen/6x8-minmax-wasmsimd-loadsplat-x86.c",
895 "src/f32-gemm/gen/6x8-minmax-wasmsimd-splat-arm.c",
896 "src/f32-gemm/gen/6x8-minmax-wasmsimd-splat-x86.c",
897 "src/f32-gemm/gen/6x8s4-minmax-wasmsimd-arm.c",
898 "src/f32-gemm/gen/6x8s4-minmax-wasmsimd-x86.c",
Marat Dukhanefaac272020-07-16 16:58:07 -0700899 "src/f32-hswish/gen/hswish-wasmsimd-x4.c",
900 "src/f32-hswish/gen/hswish-wasmsimd-x8.c",
901 "src/f32-hswish/gen/hswish-wasmsimd-x16.c",
XNNPACK Team965272b2020-10-23 21:10:15 -0700902 "src/f32-ibilinear-chw/gen/wasmsimd-p4.c",
903 "src/f32-ibilinear-chw/gen/wasmsimd-p8.c",
Frank Barchard35db7d02020-10-26 13:37:34 -0700904 "src/f32-ibilinear/gen/wasmsimd-c4.c",
905 "src/f32-ibilinear/gen/wasmsimd-c8.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700906 "src/f32-igemm/gen/1x8-minmax-wasmsimd-loadsplat-arm.c",
907 "src/f32-igemm/gen/1x8-minmax-wasmsimd-loadsplat-x86.c",
908 "src/f32-igemm/gen/1x8-minmax-wasmsimd-splat-arm.c",
909 "src/f32-igemm/gen/1x8-minmax-wasmsimd-splat-x86.c",
910 "src/f32-igemm/gen/1x8-relu-wasmsimd-splat.c",
911 "src/f32-igemm/gen/1x8-wasmsimd-splat.c",
912 "src/f32-igemm/gen/1x8s4-minmax-wasmsimd-arm.c",
913 "src/f32-igemm/gen/1x8s4-minmax-wasmsimd-x86.c",
914 "src/f32-igemm/gen/3x8-minmax-wasmsimd-loadsplat-arm.c",
915 "src/f32-igemm/gen/3x8-minmax-wasmsimd-loadsplat-x86.c",
916 "src/f32-igemm/gen/3x8-minmax-wasmsimd-splat-arm.c",
917 "src/f32-igemm/gen/3x8-minmax-wasmsimd-splat-x86.c",
918 "src/f32-igemm/gen/3x8s4-minmax-wasmsimd-arm.c",
919 "src/f32-igemm/gen/3x8s4-minmax-wasmsimd-x86.c",
920 "src/f32-igemm/gen/4x2c4-minmax-wasmsimd-arm.c",
921 "src/f32-igemm/gen/4x2c4-minmax-wasmsimd-x86.c",
922 "src/f32-igemm/gen/4x2c4-relu-wasmsimd.c",
923 "src/f32-igemm/gen/4x2c4-wasmsimd.c",
924 "src/f32-igemm/gen/4x8-minmax-wasmsimd-loadsplat-arm.c",
925 "src/f32-igemm/gen/4x8-minmax-wasmsimd-loadsplat-x86.c",
926 "src/f32-igemm/gen/4x8-minmax-wasmsimd-splat-arm.c",
927 "src/f32-igemm/gen/4x8-minmax-wasmsimd-splat-x86.c",
928 "src/f32-igemm/gen/4x8-relu-wasmsimd-splat.c",
929 "src/f32-igemm/gen/4x8-wasmsimd-splat.c",
930 "src/f32-igemm/gen/4x8s4-minmax-wasmsimd-arm.c",
931 "src/f32-igemm/gen/4x8s4-minmax-wasmsimd-x86.c",
932 "src/f32-igemm/gen/5x8-minmax-wasmsimd-loadsplat-arm.c",
933 "src/f32-igemm/gen/5x8-minmax-wasmsimd-loadsplat-x86.c",
934 "src/f32-igemm/gen/5x8-minmax-wasmsimd-splat-arm.c",
935 "src/f32-igemm/gen/5x8-minmax-wasmsimd-splat-x86.c",
936 "src/f32-igemm/gen/5x8-relu-wasmsimd-splat.c",
937 "src/f32-igemm/gen/5x8-wasmsimd-splat.c",
938 "src/f32-igemm/gen/5x8s4-minmax-wasmsimd-arm.c",
939 "src/f32-igemm/gen/5x8s4-minmax-wasmsimd-x86.c",
940 "src/f32-igemm/gen/6x8-minmax-wasmsimd-loadsplat-arm.c",
941 "src/f32-igemm/gen/6x8-minmax-wasmsimd-loadsplat-x86.c",
942 "src/f32-igemm/gen/6x8-minmax-wasmsimd-splat-arm.c",
943 "src/f32-igemm/gen/6x8-minmax-wasmsimd-splat-x86.c",
944 "src/f32-igemm/gen/6x8s4-minmax-wasmsimd-arm.c",
945 "src/f32-igemm/gen/6x8s4-minmax-wasmsimd-x86.c",
Marat Dukhanf6e24802020-07-08 22:20:40 -0700946 "src/f32-maxpool/9p8x-minmax-wasmsimd-arm-c4.c",
947 "src/f32-maxpool/9p8x-minmax-wasmsimd-x86-c4.c",
Marat Dukhan1483c532020-07-16 18:08:19 -0700948 "src/f32-pavgpool/9p8x-minmax-wasmsimd-arm-c4.c",
949 "src/f32-pavgpool/9p8x-minmax-wasmsimd-x86-c4.c",
950 "src/f32-pavgpool/9x-minmax-wasmsimd-arm-c4.c",
951 "src/f32-pavgpool/9x-minmax-wasmsimd-x86-c4.c",
Marat Dukhan0d0d8822020-07-23 23:37:56 -0700952 "src/f32-ppmm/gen/4x8-minmax-wasmsimd-splat-arm.c",
953 "src/f32-ppmm/gen/4x8-minmax-wasmsimd-splat-x86.c",
Frank Barcharda5316982020-07-23 13:19:28 -0700954 "src/f32-prelu/gen/wasmsimd-bitselect-1x4.c",
955 "src/f32-prelu/gen/wasmsimd-bitselect-1x8.c",
956 "src/f32-prelu/gen/wasmsimd-bitselect-1x16.c",
Marat Dukhan195f8eb2020-06-25 12:50:57 -0700957 "src/f32-prelu/gen/wasmsimd-bitselect-2x4.c",
958 "src/f32-prelu/gen/wasmsimd-bitselect-2x8.c",
Frank Barcharda5316982020-07-23 13:19:28 -0700959 "src/f32-prelu/gen/wasmsimd-bitselect-2x16.c",
960 "src/f32-prelu/gen/wasmsimd-bitselect-4x4.c",
961 "src/f32-prelu/gen/wasmsimd-bitselect-4x8.c",
962 "src/f32-prelu/gen/wasmsimd-bitselect-4x16.c",
963 "src/f32-prelu/gen/wasmsimd-minmax-1x4.c",
964 "src/f32-prelu/gen/wasmsimd-minmax-1x8.c",
965 "src/f32-prelu/gen/wasmsimd-minmax-1x16.c",
Marat Dukhan195f8eb2020-06-25 12:50:57 -0700966 "src/f32-prelu/gen/wasmsimd-minmax-2x4.c",
967 "src/f32-prelu/gen/wasmsimd-minmax-2x8.c",
Frank Barcharda5316982020-07-23 13:19:28 -0700968 "src/f32-prelu/gen/wasmsimd-minmax-2x16.c",
969 "src/f32-prelu/gen/wasmsimd-minmax-4x4.c",
970 "src/f32-prelu/gen/wasmsimd-minmax-4x8.c",
971 "src/f32-prelu/gen/wasmsimd-minmax-4x16.c",
Marat Dukhan52238f02020-07-16 15:30:28 -0700972 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x4.c",
Marat Dukhan52238f02020-07-16 15:30:28 -0700973 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x8-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700974 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x8.c",
Marat Dukhan52238f02020-07-16 15:30:28 -0700975 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x12-acc2.c",
976 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x12-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700977 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x12.c",
Marat Dukhan52238f02020-07-16 15:30:28 -0700978 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x16-acc2.c",
979 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x16-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700980 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x16.c",
Marat Dukhan52238f02020-07-16 15:30:28 -0700981 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x20-acc2.c",
982 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x20-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700983 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x20.c",
Frank Barchardfb158e22020-07-15 16:10:10 -0700984 "src/f32-relu/gen/wasmsimd-x4.c",
985 "src/f32-relu/gen/wasmsimd-x8.c",
Frank Barchard37297a62020-08-27 12:15:10 -0700986 "src/f32-relu/gen/wasmsimd-x16.c",
Marat Dukhan8c417962020-07-08 12:27:50 -0700987 "src/f32-rmax/wasmsimd-arm.c",
988 "src/f32-rmax/wasmsimd-x86.c",
Marat Dukhand187a5b2020-07-20 01:07:17 -0700989 "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x4.c",
990 "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x8.c",
991 "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x12.c",
992 "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x16.c",
993 "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x20.c",
994 "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x24.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700995 "src/f32-sigmoid/gen/wasmsimd-p5-div-x4.c",
996 "src/f32-sigmoid/gen/wasmsimd-p5-div-x8.c",
997 "src/f32-sigmoid/gen/wasmsimd-p5-div-x12.c",
998 "src/f32-sigmoid/gen/wasmsimd-p5-div-x16.c",
999 "src/f32-sigmoid/gen/wasmsimd-p5-div-x20.c",
1000 "src/f32-sigmoid/gen/wasmsimd-p5-div-x24.c",
Frank Barchard8ef44cd2020-11-03 12:30:23 -08001001 "src/f32-spmm/gen/4x1-minmax-wasmsimd-arm-pipelined-x2.c",
1002 "src/f32-spmm/gen/4x1-minmax-wasmsimd-arm-pipelined.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07001003 "src/f32-spmm/gen/4x1-minmax-wasmsimd-arm-x2.c",
1004 "src/f32-spmm/gen/4x1-minmax-wasmsimd-arm-x4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001005 "src/f32-spmm/gen/4x1-minmax-wasmsimd-arm.c",
Frank Barchard8ef44cd2020-11-03 12:30:23 -08001006 "src/f32-spmm/gen/4x1-minmax-wasmsimd-x86-pipelined-x2.c",
1007 "src/f32-spmm/gen/4x1-minmax-wasmsimd-x86-pipelined.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07001008 "src/f32-spmm/gen/4x1-minmax-wasmsimd-x86-x2.c",
1009 "src/f32-spmm/gen/4x1-minmax-wasmsimd-x86-x4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001010 "src/f32-spmm/gen/4x1-minmax-wasmsimd-x86.c",
Frank Barchard8ef44cd2020-11-03 12:30:23 -08001011 "src/f32-spmm/gen/8x1-minmax-wasmsimd-arm-pipelined-x2.c",
1012 "src/f32-spmm/gen/8x1-minmax-wasmsimd-arm-pipelined.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07001013 "src/f32-spmm/gen/8x1-minmax-wasmsimd-arm-x2.c",
1014 "src/f32-spmm/gen/8x1-minmax-wasmsimd-arm-x4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001015 "src/f32-spmm/gen/8x1-minmax-wasmsimd-arm.c",
Frank Barchard8ef44cd2020-11-03 12:30:23 -08001016 "src/f32-spmm/gen/8x1-minmax-wasmsimd-x86-pipelined-x2.c",
1017 "src/f32-spmm/gen/8x1-minmax-wasmsimd-x86-pipelined.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07001018 "src/f32-spmm/gen/8x1-minmax-wasmsimd-x86-x2.c",
1019 "src/f32-spmm/gen/8x1-minmax-wasmsimd-x86-x4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001020 "src/f32-spmm/gen/8x1-minmax-wasmsimd-x86.c",
Frank Barchard8ef44cd2020-11-03 12:30:23 -08001021 "src/f32-spmm/gen/16x1-minmax-wasmsimd-arm-pipelined-x2.c",
1022 "src/f32-spmm/gen/16x1-minmax-wasmsimd-arm-pipelined.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07001023 "src/f32-spmm/gen/16x1-minmax-wasmsimd-arm-x2.c",
1024 "src/f32-spmm/gen/16x1-minmax-wasmsimd-arm-x4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001025 "src/f32-spmm/gen/16x1-minmax-wasmsimd-arm.c",
Frank Barchard8ef44cd2020-11-03 12:30:23 -08001026 "src/f32-spmm/gen/16x1-minmax-wasmsimd-x86-pipelined-x2.c",
1027 "src/f32-spmm/gen/16x1-minmax-wasmsimd-x86-pipelined.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07001028 "src/f32-spmm/gen/16x1-minmax-wasmsimd-x86-x2.c",
1029 "src/f32-spmm/gen/16x1-minmax-wasmsimd-x86-x4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001030 "src/f32-spmm/gen/16x1-minmax-wasmsimd-x86.c",
Frank Barchard8ef44cd2020-11-03 12:30:23 -08001031 "src/f32-spmm/gen/32x1-minmax-wasmsimd-arm-pipelined-x2.c",
1032 "src/f32-spmm/gen/32x1-minmax-wasmsimd-arm-pipelined.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07001033 "src/f32-spmm/gen/32x1-minmax-wasmsimd-arm-x2.c",
1034 "src/f32-spmm/gen/32x1-minmax-wasmsimd-arm-x4.c",
Frank Barchard846c0c62020-10-26 15:01:39 -07001035 "src/f32-spmm/gen/32x1-minmax-wasmsimd-arm.c",
Frank Barchard8ef44cd2020-11-03 12:30:23 -08001036 "src/f32-spmm/gen/32x1-minmax-wasmsimd-x86-pipelined-x2.c",
1037 "src/f32-spmm/gen/32x1-minmax-wasmsimd-x86-pipelined.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07001038 "src/f32-spmm/gen/32x1-minmax-wasmsimd-x86-x2.c",
1039 "src/f32-spmm/gen/32x1-minmax-wasmsimd-x86-x4.c",
Frank Barchard846c0c62020-10-26 15:01:39 -07001040 "src/f32-spmm/gen/32x1-minmax-wasmsimd-x86.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -07001041 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-arm-x4.c",
1042 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001043 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-arm-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -07001044 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-x86-x4.c",
1045 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001046 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -07001047 "src/f32-vbinary/gen/vadd-relu-wasmsimd-x4.c",
1048 "src/f32-vbinary/gen/vadd-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001049 "src/f32-vbinary/gen/vadd-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07001050 "src/f32-vbinary/gen/vadd-wasmsimd-x4.c",
1051 "src/f32-vbinary/gen/vadd-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001052 "src/f32-vbinary/gen/vadd-wasmsimd-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001053 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-arm-x4.c",
1054 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001055 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-arm-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001056 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-x86-x4.c",
1057 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001058 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -07001059 "src/f32-vbinary/gen/vaddc-relu-wasmsimd-x4.c",
1060 "src/f32-vbinary/gen/vaddc-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001061 "src/f32-vbinary/gen/vaddc-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07001062 "src/f32-vbinary/gen/vaddc-wasmsimd-x4.c",
1063 "src/f32-vbinary/gen/vaddc-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001064 "src/f32-vbinary/gen/vaddc-wasmsimd-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -07001065 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-arm-x4.c",
1066 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001067 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-arm-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -07001068 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-x86-x4.c",
1069 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001070 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -07001071 "src/f32-vbinary/gen/vdiv-relu-wasmsimd-x4.c",
1072 "src/f32-vbinary/gen/vdiv-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001073 "src/f32-vbinary/gen/vdiv-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07001074 "src/f32-vbinary/gen/vdiv-wasmsimd-x4.c",
1075 "src/f32-vbinary/gen/vdiv-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001076 "src/f32-vbinary/gen/vdiv-wasmsimd-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001077 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-arm-x4.c",
1078 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001079 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-arm-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001080 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-x86-x4.c",
1081 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001082 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -07001083 "src/f32-vbinary/gen/vdivc-relu-wasmsimd-x4.c",
1084 "src/f32-vbinary/gen/vdivc-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001085 "src/f32-vbinary/gen/vdivc-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07001086 "src/f32-vbinary/gen/vdivc-wasmsimd-x4.c",
1087 "src/f32-vbinary/gen/vdivc-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001088 "src/f32-vbinary/gen/vdivc-wasmsimd-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -07001089 "src/f32-vbinary/gen/vmax-wasmsimd-arm-x4.c",
1090 "src/f32-vbinary/gen/vmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001091 "src/f32-vbinary/gen/vmax-wasmsimd-arm-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -07001092 "src/f32-vbinary/gen/vmax-wasmsimd-x86-x4.c",
1093 "src/f32-vbinary/gen/vmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001094 "src/f32-vbinary/gen/vmax-wasmsimd-x86-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001095 "src/f32-vbinary/gen/vmaxc-wasmsimd-arm-x4.c",
1096 "src/f32-vbinary/gen/vmaxc-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001097 "src/f32-vbinary/gen/vmaxc-wasmsimd-arm-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001098 "src/f32-vbinary/gen/vmaxc-wasmsimd-x86-x4.c",
1099 "src/f32-vbinary/gen/vmaxc-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001100 "src/f32-vbinary/gen/vmaxc-wasmsimd-x86-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -07001101 "src/f32-vbinary/gen/vmin-wasmsimd-arm-x4.c",
1102 "src/f32-vbinary/gen/vmin-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001103 "src/f32-vbinary/gen/vmin-wasmsimd-arm-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -07001104 "src/f32-vbinary/gen/vmin-wasmsimd-x86-x4.c",
1105 "src/f32-vbinary/gen/vmin-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001106 "src/f32-vbinary/gen/vmin-wasmsimd-x86-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001107 "src/f32-vbinary/gen/vminc-wasmsimd-arm-x4.c",
1108 "src/f32-vbinary/gen/vminc-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001109 "src/f32-vbinary/gen/vminc-wasmsimd-arm-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001110 "src/f32-vbinary/gen/vminc-wasmsimd-x86-x4.c",
1111 "src/f32-vbinary/gen/vminc-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001112 "src/f32-vbinary/gen/vminc-wasmsimd-x86-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -07001113 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-arm-x4.c",
1114 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001115 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-arm-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -07001116 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-x86-x4.c",
1117 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001118 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -07001119 "src/f32-vbinary/gen/vmul-relu-wasmsimd-x4.c",
1120 "src/f32-vbinary/gen/vmul-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001121 "src/f32-vbinary/gen/vmul-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07001122 "src/f32-vbinary/gen/vmul-wasmsimd-x4.c",
1123 "src/f32-vbinary/gen/vmul-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001124 "src/f32-vbinary/gen/vmul-wasmsimd-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001125 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-arm-x4.c",
1126 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001127 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-arm-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001128 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-x86-x4.c",
1129 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001130 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -07001131 "src/f32-vbinary/gen/vmulc-relu-wasmsimd-x4.c",
1132 "src/f32-vbinary/gen/vmulc-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001133 "src/f32-vbinary/gen/vmulc-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07001134 "src/f32-vbinary/gen/vmulc-wasmsimd-x4.c",
1135 "src/f32-vbinary/gen/vmulc-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001136 "src/f32-vbinary/gen/vmulc-wasmsimd-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001137 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-arm-x4.c",
1138 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001139 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-arm-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001140 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-x86-x4.c",
1141 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001142 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -07001143 "src/f32-vbinary/gen/vrdivc-relu-wasmsimd-x4.c",
1144 "src/f32-vbinary/gen/vrdivc-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001145 "src/f32-vbinary/gen/vrdivc-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07001146 "src/f32-vbinary/gen/vrdivc-wasmsimd-x4.c",
1147 "src/f32-vbinary/gen/vrdivc-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001148 "src/f32-vbinary/gen/vrdivc-wasmsimd-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001149 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-arm-x4.c",
1150 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001151 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-arm-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001152 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-x86-x4.c",
1153 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001154 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -07001155 "src/f32-vbinary/gen/vrsubc-relu-wasmsimd-x4.c",
1156 "src/f32-vbinary/gen/vrsubc-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001157 "src/f32-vbinary/gen/vrsubc-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07001158 "src/f32-vbinary/gen/vrsubc-wasmsimd-x4.c",
1159 "src/f32-vbinary/gen/vrsubc-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001160 "src/f32-vbinary/gen/vrsubc-wasmsimd-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -07001161 "src/f32-vbinary/gen/vsqrdiff-wasmsimd-x4.c",
1162 "src/f32-vbinary/gen/vsqrdiff-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001163 "src/f32-vbinary/gen/vsqrdiff-wasmsimd-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001164 "src/f32-vbinary/gen/vsqrdiffc-wasmsimd-x4.c",
1165 "src/f32-vbinary/gen/vsqrdiffc-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001166 "src/f32-vbinary/gen/vsqrdiffc-wasmsimd-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -07001167 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-arm-x4.c",
1168 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001169 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-arm-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -07001170 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-x86-x4.c",
1171 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001172 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -07001173 "src/f32-vbinary/gen/vsub-relu-wasmsimd-x4.c",
1174 "src/f32-vbinary/gen/vsub-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001175 "src/f32-vbinary/gen/vsub-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07001176 "src/f32-vbinary/gen/vsub-wasmsimd-x4.c",
1177 "src/f32-vbinary/gen/vsub-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001178 "src/f32-vbinary/gen/vsub-wasmsimd-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001179 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-arm-x4.c",
1180 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001181 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-arm-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001182 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-x86-x4.c",
1183 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001184 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -07001185 "src/f32-vbinary/gen/vsubc-relu-wasmsimd-x4.c",
1186 "src/f32-vbinary/gen/vsubc-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001187 "src/f32-vbinary/gen/vsubc-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07001188 "src/f32-vbinary/gen/vsubc-wasmsimd-x4.c",
1189 "src/f32-vbinary/gen/vsubc-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001190 "src/f32-vbinary/gen/vsubc-wasmsimd-x16.c",
Marat Dukhaned6baaf2020-12-01 15:07:08 -08001191 "src/f32-velu/gen/velu-wasmsimd-arm-rr2-lut16-p3-x4.c",
1192 "src/f32-velu/gen/velu-wasmsimd-arm-rr2-lut16-p3-x8.c",
1193 "src/f32-velu/gen/velu-wasmsimd-arm-rr2-lut16-p3-x12.c",
1194 "src/f32-velu/gen/velu-wasmsimd-arm-rr2-lut16-p3-x16.c",
1195 "src/f32-velu/gen/velu-wasmsimd-arm-rr2-lut16-p3-x20.c",
1196 "src/f32-velu/gen/velu-wasmsimd-arm-rr2-lut16-p3-x24.c",
1197 "src/f32-velu/gen/velu-wasmsimd-x86-rr2-lut16-p3-x4.c",
1198 "src/f32-velu/gen/velu-wasmsimd-x86-rr2-lut16-p3-x8.c",
1199 "src/f32-velu/gen/velu-wasmsimd-x86-rr2-lut16-p3-x12.c",
1200 "src/f32-velu/gen/velu-wasmsimd-x86-rr2-lut16-p3-x16.c",
1201 "src/f32-velu/gen/velu-wasmsimd-x86-rr2-lut16-p3-x20.c",
1202 "src/f32-velu/gen/velu-wasmsimd-x86-rr2-lut16-p3-x24.c",
1203 "src/f32-velu/gen/velu-wasmsimd-arm-rr2-p6-x4.c",
1204 "src/f32-velu/gen/velu-wasmsimd-arm-rr2-p6-x8.c",
1205 "src/f32-velu/gen/velu-wasmsimd-arm-rr2-p6-x12.c",
1206 "src/f32-velu/gen/velu-wasmsimd-arm-rr2-p6-x16.c",
1207 "src/f32-velu/gen/velu-wasmsimd-arm-rr2-p6-x20.c",
1208 "src/f32-velu/gen/velu-wasmsimd-arm-rr2-p6-x24.c",
1209 "src/f32-velu/gen/velu-wasmsimd-x86-rr2-p6-x4.c",
1210 "src/f32-velu/gen/velu-wasmsimd-x86-rr2-p6-x8.c",
1211 "src/f32-velu/gen/velu-wasmsimd-x86-rr2-p6-x12.c",
1212 "src/f32-velu/gen/velu-wasmsimd-x86-rr2-p6-x16.c",
1213 "src/f32-velu/gen/velu-wasmsimd-x86-rr2-p6-x20.c",
1214 "src/f32-velu/gen/velu-wasmsimd-x86-rr2-p6-x24.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07001215 "src/f32-vlrelu/gen/vlrelu-wasmsimd-bitselect-x4.c",
1216 "src/f32-vlrelu/gen/vlrelu-wasmsimd-bitselect-x8.c",
1217 "src/f32-vlrelu/gen/vlrelu-wasmsimd-minmax-x4.c",
1218 "src/f32-vlrelu/gen/vlrelu-wasmsimd-minmax-x8.c",
Marat Dukhand816f622020-07-15 10:14:39 -07001219 "src/f32-vmulcaddc/gen/c4-minmax-wasmsimd-arm-2x.c",
Marat Dukhand816f622020-07-15 10:14:39 -07001220 "src/f32-vmulcaddc/gen/c4-minmax-wasmsimd-x86-2x.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001221 "src/f32-vmulcaddc/gen/c8-minmax-wasmsimd-arm-2x.c",
Marat Dukhand816f622020-07-15 10:14:39 -07001222 "src/f32-vmulcaddc/gen/c8-minmax-wasmsimd-x86-2x.c",
Marat Dukhanb82b2cd2020-07-16 02:23:42 -07001223 "src/f32-vrnd/gen/vrndd-wasmsimd-addsub-x4.c",
1224 "src/f32-vrnd/gen/vrndd-wasmsimd-addsub-x8.c",
1225 "src/f32-vrnd/gen/vrndd-wasmsimd-cvt-x4.c",
1226 "src/f32-vrnd/gen/vrndd-wasmsimd-cvt-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001227 "src/f32-vrnd/gen/vrndne-wasmsimd-addsub-x4.c",
1228 "src/f32-vrnd/gen/vrndne-wasmsimd-addsub-x8.c",
1229 "src/f32-vrnd/gen/vrndu-wasmsimd-addsub-x4.c",
1230 "src/f32-vrnd/gen/vrndu-wasmsimd-addsub-x8.c",
1231 "src/f32-vrnd/gen/vrndu-wasmsimd-cvt-x4.c",
1232 "src/f32-vrnd/gen/vrndu-wasmsimd-cvt-x8.c",
1233 "src/f32-vrnd/gen/vrndz-wasmsimd-addsub-x4.c",
1234 "src/f32-vrnd/gen/vrndz-wasmsimd-addsub-x8.c",
1235 "src/f32-vrnd/gen/vrndz-wasmsimd-cvt-x4.c",
1236 "src/f32-vrnd/gen/vrndz-wasmsimd-cvt-x8.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07001237 "src/f32-vsqrt/gen/wasmsimd-sqrt-x4.c",
1238 "src/f32-vsqrt/gen/wasmsimd-sqrt-x8.c",
Marat Dukhan37c83512020-06-29 13:25:53 -07001239 "src/f32-vunary/gen/vabs-wasmsimd-x4.c",
1240 "src/f32-vunary/gen/vabs-wasmsimd-x8.c",
1241 "src/f32-vunary/gen/vneg-wasmsimd-x4.c",
1242 "src/f32-vunary/gen/vneg-wasmsimd-x8.c",
1243 "src/f32-vunary/gen/vsqr-wasmsimd-x4.c",
1244 "src/f32-vunary/gen/vsqr-wasmsimd-x8.c",
Marat Dukhande390d42020-11-29 19:32:18 -08001245 "src/math/expm1minus-wasmsimd-rr2-lut16-p3-andnot.c",
1246 "src/math/expm1minus-wasmsimd-rr2-lut16-p3-max.c",
1247 "src/math/expm1minus-wasmsimd-rr2-p6-andnot.c",
1248 "src/math/expm1minus-wasmsimd-rr2-p6-max.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001249 "src/math/roundd-wasmsimd-addsub.c",
1250 "src/math/roundd-wasmsimd-cvt.c",
1251 "src/math/roundne-wasmsimd-addsub.c",
1252 "src/math/roundu-wasmsimd-addsub.c",
1253 "src/math/roundu-wasmsimd-cvt.c",
1254 "src/math/roundz-wasmsimd-addsub.c",
1255 "src/math/roundz-wasmsimd-cvt.c",
1256 "src/math/sigmoid-wasmsimd-rr2-lut64-p2-div.c",
1257 "src/math/sigmoid-wasmsimd-rr2-p5-div.c",
Marat Dukhancc8f34c2020-08-05 16:36:38 -07001258 "src/qs8-dwconv/gen/up8x9-minmax-wasmsimd-mul16.c",
1259 "src/qs8-dwconv/gen/up16x9-minmax-wasmsimd-mul16.c",
1260 "src/qs8-dwconv/gen/up24x9-minmax-wasmsimd-mul16.c",
Marat Dukhanb5e3d172020-08-06 13:29:53 -07001261 "src/qs8-gavgpool/gen/7p7x-minmax-wasmsimd-c8-acc2.c",
1262 "src/qs8-gavgpool/gen/7p7x-minmax-wasmsimd-c16-acc2.c",
1263 "src/qs8-gavgpool/gen/7p7x-minmax-wasmsimd-c24-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001264 "src/qs8-gavgpool/gen/7x-minmax-wasmsimd-c8-acc2.c",
1265 "src/qs8-gavgpool/gen/7x-minmax-wasmsimd-c16-acc2.c",
1266 "src/qs8-gavgpool/gen/7x-minmax-wasmsimd-c24-acc2.c",
Marat Dukhan27203da2020-08-05 15:19:03 -07001267 "src/qs8-gemm/gen/1x4c8-minmax-wasmsimd-ld64.c",
Marat Dukhan27203da2020-08-05 15:19:03 -07001268 "src/qs8-gemm/gen/1x4c8-minmax-wasmsimd-ld128.c",
Marat Dukhan27203da2020-08-05 15:19:03 -07001269 "src/qs8-gemm/gen/1x4c8-xw-minmax-wasmsimd.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001270 "src/qs8-gemm/gen/2x4c8-minmax-wasmsimd-ld64.c",
1271 "src/qs8-gemm/gen/2x4c8-minmax-wasmsimd-ld128.c",
Marat Dukhan27203da2020-08-05 15:19:03 -07001272 "src/qs8-gemm/gen/2x4c8-xw-minmax-wasmsimd.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001273 "src/qs8-gemm/gen/3x4c8-minmax-wasmsimd-ld64.c",
1274 "src/qs8-gemm/gen/3x4c8-minmax-wasmsimd-ld128.c",
Marat Dukhan27203da2020-08-05 15:19:03 -07001275 "src/qs8-gemm/gen/3x4c8-xw-minmax-wasmsimd.c",
1276 "src/qs8-igemm/gen/1x4c8-minmax-wasmsimd-ld64.c",
Marat Dukhan27203da2020-08-05 15:19:03 -07001277 "src/qs8-igemm/gen/1x4c8-minmax-wasmsimd-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001278 "src/qs8-igemm/gen/2x4c8-minmax-wasmsimd-ld64.c",
Marat Dukhan27203da2020-08-05 15:19:03 -07001279 "src/qs8-igemm/gen/2x4c8-minmax-wasmsimd-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001280 "src/qs8-igemm/gen/3x4c8-minmax-wasmsimd-ld64.c",
Marat Dukhan27203da2020-08-05 15:19:03 -07001281 "src/qs8-igemm/gen/3x4c8-minmax-wasmsimd-ld128.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -07001282 "src/qs8-requantization/fp32-wasmsimd.c",
Marat Dukhan138560c2020-08-03 18:57:34 -07001283 "src/qs8-requantization/q31-wasmsimd.c",
Marat Dukhan5df27f82020-09-02 23:59:21 -07001284 "src/qs8-vadd/gen/minmax-wasmsimd-x8.c",
1285 "src/qs8-vadd/gen/minmax-wasmsimd-x16.c",
1286 "src/qs8-vadd/gen/minmax-wasmsimd-x24.c",
1287 "src/qs8-vadd/gen/minmax-wasmsimd-x32.c",
1288 "src/qs8-vaddc/gen/minmax-wasmsimd-x8.c",
1289 "src/qs8-vaddc/gen/minmax-wasmsimd-x16.c",
1290 "src/qs8-vaddc/gen/minmax-wasmsimd-x24.c",
1291 "src/qs8-vaddc/gen/minmax-wasmsimd-x32.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07001292 "src/qu8-requantization/fp32-wasmsimd.c",
Marat Dukhan22de5e72020-08-03 21:53:18 -07001293 "src/qu8-requantization/q31-wasmsimd.c",
Marat Dukhan8ee37012020-07-16 13:17:13 -07001294 "src/x32-fill/wasmsimd.c",
Marat Dukhan66d99e92020-07-16 12:56:21 -07001295 "src/x32-packx/x4-wasmsimd.c",
Marat Dukhan9306ae02020-07-16 15:51:13 -07001296 "src/x32-pad/wasmsimd.c",
Marat Dukhan9d4bfa22020-07-16 19:07:04 -07001297 "src/x32-unpool/wasmsimd.c",
Marat Dukhane3b78762020-07-16 20:02:58 -07001298 "src/x32-zip/x2-wasmsimd.c",
1299 "src/x32-zip/x3-wasmsimd.c",
1300 "src/x32-zip/x4-wasmsimd.c",
1301 "src/x32-zip/xm-wasmsimd.c",
Marat Dukhan290055c2020-06-09 12:24:29 -07001302]
1303
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001304PSIMD_ACCMATH_UKERNELS = [
Marat Dukhan2e23d2b2020-07-29 16:01:37 -07001305 "src/qs8-requantization/fp32-psimd.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001306 "src/qs8-requantization/precise-psimd.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07001307 "src/qu8-requantization/fp32-psimd.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001308 "src/qu8-requantization/precise-psimd.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001309]
1310
Marat Dukhan08c4a432019-10-03 09:29:21 -07001311# ISA-specific micro-kernels
1312NEON_UKERNELS = [
Marat Dukhanef25c6d2020-07-24 00:59:40 -07001313 "src/f32-argmaxpool/4x-neon-c4.c",
1314 "src/f32-argmaxpool/9p8x-neon-c4.c",
1315 "src/f32-argmaxpool/9x-neon-c4.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001316 "src/f32-avgpool/9p8x-minmax-neon-c4.c",
1317 "src/f32-avgpool/9x-minmax-neon-c4.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001318 "src/f32-clamp/gen/neon-x4.c",
1319 "src/f32-clamp/gen/neon-x8.c",
Marat Dukhan56b10cd2020-05-18 09:35:49 -07001320 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neon-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001321 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neon-2x2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001322 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neon-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001323 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neon-2x2.c",
Marat Dukhan56b10cd2020-05-18 09:35:49 -07001324 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neon-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001325 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neon-2x2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001326 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neon-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001327 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neon-2x2.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -07001328 "src/f32-dwconv/gen/up4x4-minmax-neon-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001329 "src/f32-dwconv/gen/up4x4-minmax-neon.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001330 "src/f32-dwconv/gen/up4x9-minmax-neon-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001331 "src/f32-dwconv/gen/up4x9-minmax-neon.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -07001332 "src/f32-dwconv/gen/up4x25-minmax-neon-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001333 "src/f32-dwconv/gen/up4x25-minmax-neon.c",
1334 "src/f32-dwconv/gen/up8x4-minmax-neon-acc2.c",
1335 "src/f32-dwconv/gen/up8x4-minmax-neon.c",
1336 "src/f32-dwconv/gen/up8x9-minmax-neon-acc2.c",
1337 "src/f32-dwconv/gen/up8x9-minmax-neon.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -07001338 "src/f32-dwconv/gen/up8x25-minmax-neon-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001339 "src/f32-dwconv/gen/up8x25-minmax-neon.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001340 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-1x4-acc2.c",
1341 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-1x4-acc3.c",
1342 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-1x4-acc4.c",
Marat Dukhanc581e482020-10-24 01:28:11 -07001343 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-1x4.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001344 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-2x4-acc2.c",
Marat Dukhanc581e482020-10-24 01:28:11 -07001345 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-2x4.c",
1346 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-3x4.c",
1347 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-4x4.c",
1348 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-5x4.c",
1349 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-6x4.c",
Marat Dukhan82f0c322020-10-25 19:17:35 -07001350 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neon-1x4-acc2.c",
1351 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neon-1x4-acc3.c",
1352 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neon-1x4-acc4.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001353 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neon-1x4.c",
Marat Dukhan82f0c322020-10-25 19:17:35 -07001354 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neon-2x4-acc2.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001355 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neon-2x4.c",
1356 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neon-3x4.c",
1357 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neon-4x4.c",
Marat Dukhan149f0ea2020-10-26 12:50:33 -07001358 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-1x4-acc2.c",
1359 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-1x4-acc3.c",
1360 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-1x4-acc4.c",
1361 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-1x4-acc5.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001362 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-1x4.c",
Marat Dukhan149f0ea2020-10-26 12:50:33 -07001363 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-2x4-acc2.c",
1364 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-2x4-acc3.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001365 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-2x4.c",
Marat Dukhan149f0ea2020-10-26 12:50:33 -07001366 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-3x4-acc2.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001367 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-3x4.c",
Marat Dukhan149f0ea2020-10-26 12:50:33 -07001368 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-4x4-acc2.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001369 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-4x4.c",
1370 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-5x4.c",
Marat Dukhan30d4b252020-10-29 16:33:22 -07001371 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neon-1x4-acc2.c",
1372 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neon-1x4-acc3.c",
1373 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neon-1x4-acc4.c",
1374 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neon-1x4-acc5.c",
1375 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neon-1x4.c",
1376 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neon-2x4-acc2.c",
1377 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neon-2x4-acc3.c",
1378 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neon-2x4.c",
Marat Dukhan30d4b252020-10-29 16:33:22 -07001379 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neon-3x4-acc2.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07001380 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neon-3x4.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -07001381 "src/f32-gavgpool-cw/neon-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001382 "src/f32-gavgpool/7p7x-minmax-neon-c4.c",
1383 "src/f32-gavgpool/7x-minmax-neon-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001384 "src/f32-gemm/gen-inc/1x8inc-minmax-neon-dup-ld64.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001385 "src/f32-gemm/gen-inc/1x8inc-minmax-neon-lane-ld64.c",
1386 "src/f32-gemm/gen-inc/1x8s4inc-minmax-neon.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001387 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-dup-ld64.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001388 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-dup-ld128.c",
1389 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-lane-ld64.c",
1390 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-lane-ld128.c",
1391 "src/f32-gemm/gen-inc/4x8s4inc-minmax-neon.c",
1392 "src/f32-gemm/gen-inc/5x8inc-minmax-neon-lane-ld64.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001393 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-dup-ld64.c",
1394 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-dup-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001395 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-lane-ld64.c",
1396 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-lane-ld128.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001397 "src/f32-gemm/gen-inc/6x8s4inc-minmax-neon.c",
1398 "src/f32-gemm/gen-inc/8x8s4inc-minmax-neon.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001399 "src/f32-gemm/gen/1x8-minmax-neon-dup-ld64.c",
1400 "src/f32-gemm/gen/1x8-minmax-neon-lane-ld64.c",
1401 "src/f32-gemm/gen/1x8s4-minmax-neon.c",
1402 "src/f32-gemm/gen/4x2-minmax-neon-lane-ld64.c",
1403 "src/f32-gemm/gen/4x8-minmax-neon-dup-ld64.c",
1404 "src/f32-gemm/gen/4x8-minmax-neon-dup-ld128.c",
1405 "src/f32-gemm/gen/4x8-minmax-neon-lane-ld64.c",
1406 "src/f32-gemm/gen/4x8-minmax-neon-lane-ld128.c",
1407 "src/f32-gemm/gen/4x8s4-minmax-neon.c",
1408 "src/f32-gemm/gen/5x8-minmax-neon-lane-ld64.c",
1409 "src/f32-gemm/gen/6x8-minmax-neon-dup-ld64.c",
1410 "src/f32-gemm/gen/6x8-minmax-neon-dup-ld128.c",
1411 "src/f32-gemm/gen/6x8-minmax-neon-lane-ld64.c",
1412 "src/f32-gemm/gen/6x8-minmax-neon-lane-ld128.c",
1413 "src/f32-gemm/gen/6x8s4-minmax-neon.c",
1414 "src/f32-gemm/gen/8x8s4-minmax-neon.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07001415 "src/f32-hswish/gen/hswish-neon-x4.c",
1416 "src/f32-hswish/gen/hswish-neon-x8.c",
1417 "src/f32-hswish/gen/hswish-neon-x16.c",
Marat Dukhan660fd192020-03-10 04:55:30 -07001418 "src/f32-ibilinear/gen/neon-c4.c",
1419 "src/f32-ibilinear/gen/neon-c8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001420 "src/f32-igemm/gen/1x8-minmax-neon-dup-ld64.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001421 "src/f32-igemm/gen/1x8-minmax-neon-lane-ld64.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001422 "src/f32-igemm/gen/1x8s4-minmax-neon.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001423 "src/f32-igemm/gen/4x2-minmax-neon-lane-ld64.c",
1424 "src/f32-igemm/gen/4x4-minmax-neon-lane-ld64.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001425 "src/f32-igemm/gen/4x8-minmax-neon-dup-ld64.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001426 "src/f32-igemm/gen/4x8-minmax-neon-dup-ld128.c",
1427 "src/f32-igemm/gen/4x8-minmax-neon-lane-ld64.c",
1428 "src/f32-igemm/gen/4x8-minmax-neon-lane-ld128.c",
1429 "src/f32-igemm/gen/4x8s4-minmax-neon.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001430 "src/f32-igemm/gen/6x8-minmax-neon-dup-ld64.c",
1431 "src/f32-igemm/gen/6x8-minmax-neon-dup-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001432 "src/f32-igemm/gen/6x8-minmax-neon-lane-ld64.c",
1433 "src/f32-igemm/gen/6x8-minmax-neon-lane-ld128.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001434 "src/f32-igemm/gen/6x8s4-minmax-neon.c",
1435 "src/f32-igemm/gen/8x8s4-minmax-neon.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001436 "src/f32-maxpool/9p8x-minmax-neon-c4.c",
1437 "src/f32-pavgpool/9p8x-minmax-neon-c4.c",
1438 "src/f32-pavgpool/9x-minmax-neon-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001439 "src/f32-ppmm/gen/4x8-minmax-neon.c",
1440 "src/f32-ppmm/gen/8x8-minmax-neon.c",
Frank Barcharda5316982020-07-23 13:19:28 -07001441 "src/f32-prelu/gen/neon-1x4.c",
1442 "src/f32-prelu/gen/neon-1x8.c",
1443 "src/f32-prelu/gen/neon-1x16.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001444 "src/f32-prelu/gen/neon-2x4.c",
1445 "src/f32-prelu/gen/neon-2x8.c",
Frank Barcharda5316982020-07-23 13:19:28 -07001446 "src/f32-prelu/gen/neon-2x16.c",
1447 "src/f32-prelu/gen/neon-4x4.c",
1448 "src/f32-prelu/gen/neon-4x8.c",
1449 "src/f32-prelu/gen/neon-4x16.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001450 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x4.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001451 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x8-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001452 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x8.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001453 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12-acc2.c",
1454 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001455 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001456 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16-acc2.c",
1457 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001458 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001459 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20-acc2.c",
1460 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001461 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20.c",
1462 "src/f32-raddstoreexpminusmax/gen/neon-p5-x4.c",
1463 "src/f32-raddstoreexpminusmax/gen/neon-p5-x8-acc2.c",
1464 "src/f32-raddstoreexpminusmax/gen/neon-p5-x8.c",
1465 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12-acc2.c",
1466 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12-acc3.c",
1467 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12.c",
1468 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16-acc2.c",
1469 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16-acc4.c",
1470 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16.c",
1471 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20-acc2.c",
1472 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20-acc5.c",
1473 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07001474 "src/f32-relu/gen/neon-x4.c",
1475 "src/f32-relu/gen/neon-x8.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001476 "src/f32-rmax/neon.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001477 "src/f32-sigmoid/gen/neon-frac-p9-p10-nr1recps-x16.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -08001478 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x4.c",
1479 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x8.c",
1480 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x12.c",
1481 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x16.c",
1482 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x20.c",
1483 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x24.c",
1484 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x4.c",
1485 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x8.c",
1486 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x12.c",
1487 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x16.c",
1488 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x20.c",
1489 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x24.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001490 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x4.c",
1491 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x8.c",
1492 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x12.c",
1493 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x16.c",
1494 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x20.c",
1495 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x24.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001496 "src/f32-vbinary/gen/vadd-minmax-neon-x4.c",
1497 "src/f32-vbinary/gen/vadd-minmax-neon-x8.c",
1498 "src/f32-vbinary/gen/vaddc-minmax-neon-x4.c",
1499 "src/f32-vbinary/gen/vaddc-minmax-neon-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -08001500 "src/f32-vbinary/gen/vmax-neon-x4.c",
1501 "src/f32-vbinary/gen/vmax-neon-x8.c",
1502 "src/f32-vbinary/gen/vmaxc-neon-x4.c",
1503 "src/f32-vbinary/gen/vmaxc-neon-x8.c",
1504 "src/f32-vbinary/gen/vmin-neon-x4.c",
1505 "src/f32-vbinary/gen/vmin-neon-x8.c",
1506 "src/f32-vbinary/gen/vminc-neon-x4.c",
1507 "src/f32-vbinary/gen/vminc-neon-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001508 "src/f32-vbinary/gen/vmul-minmax-neon-x4.c",
1509 "src/f32-vbinary/gen/vmul-minmax-neon-x8.c",
1510 "src/f32-vbinary/gen/vmulc-minmax-neon-x4.c",
1511 "src/f32-vbinary/gen/vmulc-minmax-neon-x8.c",
1512 "src/f32-vbinary/gen/vrsubc-minmax-neon-x4.c",
1513 "src/f32-vbinary/gen/vrsubc-minmax-neon-x8.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07001514 "src/f32-vbinary/gen/vsqrdiff-neon-x4.c",
1515 "src/f32-vbinary/gen/vsqrdiff-neon-x8.c",
1516 "src/f32-vbinary/gen/vsqrdiffc-neon-x4.c",
1517 "src/f32-vbinary/gen/vsqrdiffc-neon-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001518 "src/f32-vbinary/gen/vsub-minmax-neon-x4.c",
1519 "src/f32-vbinary/gen/vsub-minmax-neon-x8.c",
1520 "src/f32-vbinary/gen/vsubc-minmax-neon-x4.c",
1521 "src/f32-vbinary/gen/vsubc-minmax-neon-x8.c",
Marat Dukhaned6baaf2020-12-01 15:07:08 -08001522 "src/f32-velu/gen/velu-neon-rr2-lut16-p3-x4.c",
1523 "src/f32-velu/gen/velu-neon-rr2-lut16-p3-x8.c",
1524 "src/f32-velu/gen/velu-neon-rr2-lut16-p3-x12.c",
1525 "src/f32-velu/gen/velu-neon-rr2-lut16-p3-x16.c",
1526 "src/f32-velu/gen/velu-neon-rr2-lut16-p3-x20.c",
1527 "src/f32-velu/gen/velu-neon-rr2-lut16-p3-x24.c",
1528 "src/f32-velu/gen/velu-neon-rr2-p6-x4.c",
1529 "src/f32-velu/gen/velu-neon-rr2-p6-x8.c",
1530 "src/f32-velu/gen/velu-neon-rr2-p6-x12.c",
1531 "src/f32-velu/gen/velu-neon-rr2-p6-x16.c",
1532 "src/f32-velu/gen/velu-neon-rr2-p6-x20.c",
1533 "src/f32-velu/gen/velu-neon-rr2-p6-x24.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07001534 "src/f32-vlrelu/gen/vlrelu-neon-x4.c",
1535 "src/f32-vlrelu/gen/vlrelu-neon-x8.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001536 "src/f32-vmulcaddc/gen/c4-minmax-neon-2x.c",
1537 "src/f32-vmulcaddc/gen/c8-minmax-neon-2x.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07001538 "src/f32-vrnd/gen/vrndd-neon-x4.c",
1539 "src/f32-vrnd/gen/vrndd-neon-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001540 "src/f32-vrnd/gen/vrndne-neon-x4.c",
1541 "src/f32-vrnd/gen/vrndne-neon-x8.c",
1542 "src/f32-vrnd/gen/vrndu-neon-x4.c",
1543 "src/f32-vrnd/gen/vrndu-neon-x8.c",
1544 "src/f32-vrnd/gen/vrndz-neon-x4.c",
1545 "src/f32-vrnd/gen/vrndz-neon-x8.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07001546 "src/f32-vunary/gen/vabs-neon-x4.c",
1547 "src/f32-vunary/gen/vabs-neon-x8.c",
1548 "src/f32-vunary/gen/vneg-neon-x4.c",
1549 "src/f32-vunary/gen/vneg-neon-x8.c",
1550 "src/f32-vunary/gen/vsqr-neon-x4.c",
1551 "src/f32-vunary/gen/vsqr-neon-x8.c",
Marat Dukhande390d42020-11-29 19:32:18 -08001552 "src/math/expm1minus-neon-rr2-lut16-p3.c",
1553 "src/math/expm1minus-neon-rr2-p6.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001554 "src/math/roundd-neon-addsub.c",
1555 "src/math/roundd-neon-cvt.c",
1556 "src/math/roundne-neon-addsub.c",
1557 "src/math/roundu-neon-addsub.c",
1558 "src/math/roundu-neon-cvt.c",
1559 "src/math/roundz-neon-addsub.c",
1560 "src/math/roundz-neon-cvt.c",
1561 "src/math/sigmoid-neon-frac-p9-p10-nr1recps.c",
1562 "src/math/sigmoid-neon-rr1-lut64-p2-nr2recps.c",
1563 "src/math/sigmoid-neon-rr1-lut2048-p1-nr2recps.c",
1564 "src/math/sigmoid-neon-rr1-p5-nr2recps.c",
1565 "src/math/sigmoid-neon-rr2-lut64-p2-nr2recps.c",
1566 "src/math/sigmoid-neon-rr2-lut2048-p1-nr2recps.c",
1567 "src/math/sigmoid-neon-rr2-p5-nr2recps.c",
1568 "src/math/sqrt-neon-nr1rsqrts.c",
1569 "src/math/sqrt-neon-nr2rsqrts.c",
1570 "src/math/sqrt-neon-nr3rsqrts.c",
Marat Dukhan023bcf92020-08-10 12:40:50 -07001571 "src/qs8-dwconv/gen/up8x9-minmax-neon-mul16.c",
1572 "src/qs8-dwconv/gen/up16x9-minmax-neon-mul16.c",
1573 "src/qs8-dwconv/gen/up24x9-minmax-neon-mul16.c",
1574 "src/qs8-dwconv/gen/up32x9-minmax-neon-mul16.c",
Marat Dukhan281262d2020-08-10 13:23:21 -07001575 "src/qs8-gavgpool/gen/7p7x-minmax-neon-c8-acc2.c",
1576 "src/qs8-gavgpool/gen/7p7x-minmax-neon-c16-acc2.c",
1577 "src/qs8-gavgpool/gen/7p7x-minmax-neon-c24-acc2.c",
1578 "src/qs8-gavgpool/gen/7p7x-minmax-neon-c32-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001579 "src/qs8-gavgpool/gen/7x-minmax-neon-c8-acc2.c",
1580 "src/qs8-gavgpool/gen/7x-minmax-neon-c16-acc2.c",
1581 "src/qs8-gavgpool/gen/7x-minmax-neon-c24-acc2.c",
1582 "src/qs8-gavgpool/gen/7x-minmax-neon-c32-acc2.c",
Marat Dukhan40bbafe2020-08-04 02:04:22 -07001583 "src/qs8-gemm/gen/1x8-minmax-neon-mlal-lane.c",
Marat Dukhan40bbafe2020-08-04 02:04:22 -07001584 "src/qs8-gemm/gen/1x16-minmax-neon-mlal-lane.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001585 "src/qs8-gemm/gen/2x8-minmax-neon-mlal-lane.c",
Marat Dukhan40bbafe2020-08-04 02:04:22 -07001586 "src/qs8-gemm/gen/2x16-minmax-neon-mlal-lane.c",
1587 "src/qs8-igemm/gen/1x8-minmax-neon-mlal-lane.c",
Marat Dukhan40bbafe2020-08-04 02:04:22 -07001588 "src/qs8-igemm/gen/1x16-minmax-neon-mlal-lane.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001589 "src/qs8-igemm/gen/2x8-minmax-neon-mlal-lane.c",
Marat Dukhan40bbafe2020-08-04 02:04:22 -07001590 "src/qs8-igemm/gen/2x16-minmax-neon-mlal-lane.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -07001591 "src/qs8-requantization/fp32-neon.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001592 "src/qs8-requantization/precise-neon.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -07001593 "src/qs8-requantization/q31-neon.c",
Marat Dukhanba7b2792020-09-02 14:26:45 -07001594 "src/qs8-vadd/gen/minmax-neon-ld64-x8.c",
1595 "src/qs8-vadd/gen/minmax-neon-ld64-x16.c",
1596 "src/qs8-vadd/gen/minmax-neon-ld64-x24.c",
1597 "src/qs8-vadd/gen/minmax-neon-ld64-x32.c",
1598 "src/qs8-vaddc/gen/minmax-neon-ld64-x8.c",
1599 "src/qs8-vaddc/gen/minmax-neon-ld64-x16.c",
1600 "src/qs8-vaddc/gen/minmax-neon-ld64-x24.c",
1601 "src/qs8-vaddc/gen/minmax-neon-ld64-x32.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -07001602 "src/qu8-avgpool/9p8x-minmax-neon-c8.c",
1603 "src/qu8-avgpool/9x-minmax-neon-c8.c",
1604 "src/qu8-dwconv/up8x9-minmax-neon.c",
1605 "src/qu8-gavgpool/7p7x-minmax-neon-c8.c",
1606 "src/qu8-gavgpool/7x-minmax-neon-c8.c",
1607 "src/qu8-gemm/4x8-minmax-neon.c",
1608 "src/qu8-gemm/8x8-minmax-neon.c",
1609 "src/qu8-igemm/4x8-minmax-neon.c",
1610 "src/qu8-igemm/8x8-minmax-neon.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07001611 "src/qu8-requantization/fp32-neon.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001612 "src/qu8-requantization/precise-neon.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07001613 "src/qu8-requantization/q31-neon.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -07001614 "src/qu8-vadd/minmax-neon.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001615 "src/u8-clamp/neon-x64.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001616 "src/u8-maxpool/9p8x-minmax-neon-c16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001617 "src/u8-rmax/neon.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001618 "src/x8-zip/x2-neon.c",
1619 "src/x8-zip/x3-neon.c",
1620 "src/x8-zip/x4-neon.c",
1621 "src/x8-zip/xm-neon.c",
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -07001622 "src/x32-fill/neon.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001623 "src/x32-packx/x4-neon-st4.c",
Marat Dukhan63523d42020-05-22 17:07:33 -07001624 "src/x32-pad/neon.c",
Marat Dukhan57dccd82020-04-14 00:53:10 -07001625 "src/x32-unpool/neon.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001626 "src/x32-zip/x2-neon.c",
1627 "src/x32-zip/x3-neon.c",
1628 "src/x32-zip/x4-neon.c",
1629 "src/x32-zip/xm-neon.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001630]
1631
1632NEONFMA_UKERNELS = [
Frank Barchard04336c12020-10-22 16:48:55 -07001633 "src/f32-dwconv/gen/up4x4-minmax-neonfma-acc2.c",
1634 "src/f32-dwconv/gen/up4x4-minmax-neonfma.c",
1635 "src/f32-dwconv/gen/up4x9-minmax-neonfma-acc2.c",
1636 "src/f32-dwconv/gen/up4x9-minmax-neonfma.c",
1637 "src/f32-dwconv/gen/up4x25-minmax-neonfma-acc2.c",
1638 "src/f32-dwconv/gen/up4x25-minmax-neonfma.c",
1639 "src/f32-dwconv/gen/up8x4-minmax-neonfma-acc2.c",
1640 "src/f32-dwconv/gen/up8x4-minmax-neonfma.c",
1641 "src/f32-dwconv/gen/up8x9-minmax-neonfma-acc2.c",
1642 "src/f32-dwconv/gen/up8x9-minmax-neonfma.c",
1643 "src/f32-dwconv/gen/up8x25-minmax-neonfma-acc2.c",
1644 "src/f32-dwconv/gen/up8x25-minmax-neonfma.c",
1645 "src/f32-gemm/gen-inc/1x8inc-minmax-neonfma-dup-ld64.c",
1646 "src/f32-gemm/gen-inc/1x8s4inc-minmax-neonfma.c",
1647 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-dup-ld64.c",
1648 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-dup-ld128.c",
1649 "src/f32-gemm/gen-inc/4x8s4inc-minmax-neonfma.c",
1650 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-dup-ld64.c",
1651 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-dup-ld128.c",
1652 "src/f32-gemm/gen-inc/6x8s4inc-minmax-neonfma.c",
1653 "src/f32-gemm/gen-inc/8x8s4inc-minmax-neonfma.c",
1654 "src/f32-gemm/gen/1x8-minmax-neonfma-dup-ld64.c",
1655 "src/f32-gemm/gen/1x8s4-minmax-neonfma.c",
1656 "src/f32-gemm/gen/4x8-minmax-neonfma-dup-ld64.c",
1657 "src/f32-gemm/gen/4x8-minmax-neonfma-dup-ld128.c",
1658 "src/f32-gemm/gen/4x8s4-minmax-neonfma.c",
1659 "src/f32-gemm/gen/6x8-minmax-neonfma-dup-ld64.c",
1660 "src/f32-gemm/gen/6x8-minmax-neonfma-dup-ld128.c",
1661 "src/f32-gemm/gen/6x8s4-minmax-neonfma.c",
1662 "src/f32-gemm/gen/8x8s4-minmax-neonfma.c",
Marat Dukhan660fd192020-03-10 04:55:30 -07001663 "src/f32-ibilinear/gen/neonfma-c4.c",
1664 "src/f32-ibilinear/gen/neonfma-c8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001665 "src/f32-igemm/gen/1x8-minmax-neonfma-dup-ld64.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001666 "src/f32-igemm/gen/1x8s4-minmax-neonfma.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001667 "src/f32-igemm/gen/4x8-minmax-neonfma-dup-ld64.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001668 "src/f32-igemm/gen/4x8-minmax-neonfma-dup-ld128.c",
1669 "src/f32-igemm/gen/4x8s4-minmax-neonfma.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001670 "src/f32-igemm/gen/6x8-minmax-neonfma-dup-ld64.c",
1671 "src/f32-igemm/gen/6x8-minmax-neonfma-dup-ld128.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001672 "src/f32-igemm/gen/6x8s4-minmax-neonfma.c",
1673 "src/f32-igemm/gen/8x8s4-minmax-neonfma.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001674 "src/f32-ppmm/gen/4x8-minmax-neonfma.c",
1675 "src/f32-ppmm/gen/8x8-minmax-neonfma.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001676 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x4.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001677 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x8-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001678 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x8.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001679 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12-acc2.c",
1680 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001681 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001682 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16-acc2.c",
1683 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001684 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001685 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20-acc2.c",
1686 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001687 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20.c",
1688 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x4.c",
1689 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x8-acc2.c",
1690 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x8.c",
1691 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12-acc2.c",
1692 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12-acc3.c",
1693 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12.c",
1694 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16-acc2.c",
1695 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16-acc4.c",
1696 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16.c",
1697 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20-acc2.c",
1698 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20-acc5.c",
1699 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -08001700 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x4.c",
1701 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x8.c",
1702 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x12.c",
1703 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x16.c",
1704 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x20.c",
1705 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x24.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001706 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x4.c",
1707 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x8.c",
1708 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x12.c",
1709 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x16.c",
1710 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x20.c",
1711 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x24.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -08001712 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x4.c",
1713 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x8.c",
1714 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x12.c",
1715 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x16.c",
1716 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x20.c",
1717 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x24.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -08001718 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x4.c",
1719 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x8.c",
1720 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x12.c",
1721 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x16.c",
1722 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x20.c",
1723 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x24.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001724 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x4.c",
1725 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x8.c",
1726 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x12.c",
1727 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x16.c",
1728 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x20.c",
1729 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x24.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -08001730 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x4.c",
1731 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x8.c",
1732 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x12.c",
1733 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x16.c",
1734 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x20.c",
1735 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x24.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001736 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x4.c",
1737 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x8.c",
1738 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x12.c",
1739 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x16.c",
1740 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x20.c",
1741 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x24.c",
1742 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x4.c",
1743 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x8.c",
1744 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x12.c",
1745 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x16.c",
1746 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x20.c",
1747 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x24.c",
1748 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x4.c",
1749 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x8.c",
1750 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x12.c",
1751 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x16.c",
1752 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x20.c",
1753 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x24.c",
Marat Dukhaned6baaf2020-12-01 15:07:08 -08001754 "src/f32-velu/gen/velu-neonfma-rr1-lut16-p3-x4.c",
1755 "src/f32-velu/gen/velu-neonfma-rr1-lut16-p3-x8.c",
1756 "src/f32-velu/gen/velu-neonfma-rr1-lut16-p3-x12.c",
1757 "src/f32-velu/gen/velu-neonfma-rr1-lut16-p3-x16.c",
1758 "src/f32-velu/gen/velu-neonfma-rr1-lut16-p3-x20.c",
1759 "src/f32-velu/gen/velu-neonfma-rr1-lut16-p3-x24.c",
1760 "src/f32-velu/gen/velu-neonfma-rr1-p6-x4.c",
1761 "src/f32-velu/gen/velu-neonfma-rr1-p6-x8.c",
1762 "src/f32-velu/gen/velu-neonfma-rr1-p6-x12.c",
1763 "src/f32-velu/gen/velu-neonfma-rr1-p6-x16.c",
1764 "src/f32-velu/gen/velu-neonfma-rr1-p6-x20.c",
1765 "src/f32-velu/gen/velu-neonfma-rr1-p6-x24.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001766 "src/f32-vmulcaddc/gen/c4-minmax-neonfma-2x.c",
1767 "src/f32-vmulcaddc/gen/c8-minmax-neonfma-2x.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07001768 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x4.c",
1769 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x8.c",
1770 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x12.c",
1771 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x16.c",
1772 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x20.c",
1773 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x24.c",
1774 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x28.c",
1775 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x32.c",
1776 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x36.c",
1777 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x40.c",
1778 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x4.c",
1779 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x8.c",
1780 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x12.c",
1781 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x16.c",
1782 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x20.c",
1783 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x24.c",
1784 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x28.c",
1785 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x32.c",
1786 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x36.c",
1787 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x40.c",
Marat Dukhanb7633f22020-11-20 16:34:56 -08001788 "src/math/exp-neonfma-rr2-lut64-p2.c",
1789 "src/math/exp-neonfma-rr2-p5.c",
Marat Dukhan9dd119a2020-11-20 18:20:04 -08001790 "src/math/expminus-neonfma-rr2-lut64-p2.c",
1791 "src/math/expminus-neonfma-rr2-lut2048-p1.c",
1792 "src/math/expminus-neonfma-rr2-p5.c",
Marat Dukhande390d42020-11-29 19:32:18 -08001793 "src/math/expm1minus-neonfma-rr1-lut16-p3.c",
1794 "src/math/expm1minus-neonfma-rr1-p6.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001795 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr1recps1fma.c",
1796 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr2fma.c",
1797 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr2recps.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001798 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr1recps1fma.c",
1799 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr2fma.c",
1800 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr2recps.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001801 "src/math/sigmoid-neonfma-rr1-p5-nr1recps1fma.c",
1802 "src/math/sigmoid-neonfma-rr1-p5-nr2fma.c",
1803 "src/math/sigmoid-neonfma-rr1-p5-nr2recps.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001804 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr1recps1fma.c",
1805 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr2fma.c",
1806 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr2recps.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001807 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr1recps1fma.c",
1808 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr2fma.c",
1809 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr2recps.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001810 "src/math/sigmoid-neonfma-rr2-p5-nr1recps1fma.c",
1811 "src/math/sigmoid-neonfma-rr2-p5-nr2fma.c",
1812 "src/math/sigmoid-neonfma-rr2-p5-nr2recps.c",
Marat Dukhan84000762020-06-29 18:38:43 -07001813 "src/math/sqrt-neonfma-nr1fma.c",
Marat Dukhan84000762020-06-29 18:38:43 -07001814 "src/math/sqrt-neonfma-nr1rsqrts1fma1adj.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001815 "src/math/sqrt-neonfma-nr2fma.c",
1816 "src/math/sqrt-neonfma-nr2fma1adj.c",
1817 "src/math/sqrt-neonfma-nr3fma.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001818]
1819
1820AARCH64_NEONFMA_UKERNELS = [
Marat Dukhan56b10cd2020-05-18 09:35:49 -07001821 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neonfma-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001822 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neonfma-2x2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001823 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neonfma-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001824 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neonfma-2x2.c",
Marat Dukhan56b10cd2020-05-18 09:35:49 -07001825 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neonfma-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001826 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neonfma-2x2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001827 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neonfma-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001828 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neonfma-2x2.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -07001829 "src/f32-conv-hwc2chw/3x3s2p1c3x4-neonfma-2x2.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001830 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-1x4-acc2.c",
1831 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-1x4-acc3.c",
1832 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-1x4-acc4.c",
Marat Dukhan1268a242020-10-24 00:36:32 -07001833 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-1x4.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001834 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-2x4-acc2.c",
Marat Dukhan1268a242020-10-24 00:36:32 -07001835 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-2x4.c",
1836 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-3x4.c",
1837 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-4x4.c",
1838 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-5x4.c",
1839 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-6x4.c",
Marat Dukhan82f0c322020-10-25 19:17:35 -07001840 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neonfma-1x4-acc2.c",
1841 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neonfma-1x4-acc3.c",
1842 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neonfma-1x4-acc4.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001843 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neonfma-1x4.c",
Marat Dukhan82f0c322020-10-25 19:17:35 -07001844 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neonfma-2x4-acc2.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001845 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neonfma-2x4.c",
1846 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neonfma-3x4.c",
1847 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neonfma-4x4.c",
Marat Dukhan149f0ea2020-10-26 12:50:33 -07001848 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-1x4-acc2.c",
1849 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-1x4-acc3.c",
1850 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-1x4-acc4.c",
1851 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-1x4-acc5.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001852 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-1x4.c",
Marat Dukhan149f0ea2020-10-26 12:50:33 -07001853 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-2x4-acc2.c",
1854 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-2x4-acc3.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001855 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-2x4.c",
Marat Dukhan149f0ea2020-10-26 12:50:33 -07001856 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-3x4-acc2.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001857 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-3x4.c",
Marat Dukhan149f0ea2020-10-26 12:50:33 -07001858 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-4x4-acc2.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001859 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-4x4.c",
1860 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-5x4.c",
Marat Dukhan30d4b252020-10-29 16:33:22 -07001861 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neonfma-1x4-acc2.c",
1862 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neonfma-1x4-acc3.c",
1863 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neonfma-1x4-acc4.c",
1864 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neonfma-1x4-acc5.c",
1865 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neonfma-1x4.c",
1866 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neonfma-2x4-acc2.c",
1867 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neonfma-2x4-acc3.c",
1868 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neonfma-2x4.c",
Marat Dukhan30d4b252020-10-29 16:33:22 -07001869 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neonfma-3x4-acc2.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07001870 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neonfma-3x4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001871 "src/f32-gemm/gen-inc/1x8inc-minmax-neonfma-lane-ld64.c",
1872 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-lane-ld64.c",
1873 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-lane-ld128.c",
1874 "src/f32-gemm/gen-inc/5x8inc-minmax-neonfma-lane-ld64.c",
1875 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-lane-ld64.c",
1876 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-lane-ld128.c",
1877 "src/f32-gemm/gen/1x8-minmax-neonfma-lane-ld64.c",
1878 "src/f32-gemm/gen/4x2-minmax-neonfma-lane-ld64.c",
1879 "src/f32-gemm/gen/4x8-minmax-neonfma-lane-ld64.c",
1880 "src/f32-gemm/gen/4x8-minmax-neonfma-lane-ld128.c",
1881 "src/f32-gemm/gen/5x8-minmax-neonfma-lane-ld64.c",
1882 "src/f32-gemm/gen/6x8-minmax-neonfma-lane-ld64.c",
1883 "src/f32-gemm/gen/6x8-minmax-neonfma-lane-ld128.c",
1884 "src/f32-igemm/gen/1x8-minmax-neonfma-lane-ld64.c",
1885 "src/f32-igemm/gen/4x2-minmax-neonfma-lane-ld64.c",
1886 "src/f32-igemm/gen/4x4-minmax-neonfma-lane-ld64.c",
1887 "src/f32-igemm/gen/4x8-minmax-neonfma-lane-ld64.c",
1888 "src/f32-igemm/gen/4x8-minmax-neonfma-lane-ld128.c",
1889 "src/f32-igemm/gen/6x8-minmax-neonfma-lane-ld64.c",
1890 "src/f32-igemm/gen/6x8-minmax-neonfma-lane-ld128.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -08001891 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x4.c",
1892 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x8.c",
1893 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x12.c",
1894 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x16.c",
1895 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x20.c",
1896 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x24.c",
1897 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x4.c",
1898 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x8.c",
1899 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x12.c",
1900 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x16.c",
1901 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x20.c",
1902 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x24.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001903 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x4.c",
1904 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x8.c",
1905 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x12.c",
1906 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x16.c",
1907 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x20.c",
1908 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x24.c",
Marat Dukhan355ab432020-04-09 19:01:52 -07001909 "src/f32-spmm/gen/4x1-minmax-neonfma-pipelined.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07001910 "src/f32-spmm/gen/4x1-minmax-neonfma-x2.c",
Marat Dukhan355ab432020-04-09 19:01:52 -07001911 "src/f32-spmm/gen/4x1-minmax-neonfma.c",
1912 "src/f32-spmm/gen/4x2-minmax-neonfma.c",
1913 "src/f32-spmm/gen/4x4-minmax-neonfma.c",
1914 "src/f32-spmm/gen/8x1-minmax-neonfma-pipelined.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07001915 "src/f32-spmm/gen/8x1-minmax-neonfma-x2.c",
Marat Dukhan355ab432020-04-09 19:01:52 -07001916 "src/f32-spmm/gen/8x1-minmax-neonfma.c",
1917 "src/f32-spmm/gen/8x2-minmax-neonfma.c",
1918 "src/f32-spmm/gen/8x4-minmax-neonfma.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001919 "src/f32-spmm/gen/12x1-minmax-neonfma.c",
1920 "src/f32-spmm/gen/12x2-minmax-neonfma.c",
1921 "src/f32-spmm/gen/12x4-minmax-neonfma.c",
1922 "src/f32-spmm/gen/16x1-minmax-neonfma-pipelined.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07001923 "src/f32-spmm/gen/16x1-minmax-neonfma-x2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001924 "src/f32-spmm/gen/16x1-minmax-neonfma.c",
1925 "src/f32-spmm/gen/16x2-minmax-neonfma.c",
1926 "src/f32-spmm/gen/16x4-minmax-neonfma.c",
Frank Barchard846c0c62020-10-26 15:01:39 -07001927 "src/f32-spmm/gen/32x1-minmax-neonfma-pipelined.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07001928 "src/f32-spmm/gen/32x1-minmax-neonfma-x2.c",
Frank Barchard846c0c62020-10-26 15:01:39 -07001929 "src/f32-spmm/gen/32x1-minmax-neonfma.c",
1930 "src/f32-spmm/gen/32x2-minmax-neonfma.c",
1931 "src/f32-spmm/gen/32x4-minmax-neonfma.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001932 "src/f32-vbinary/gen/vdiv-minmax-neon-x4.c",
1933 "src/f32-vbinary/gen/vdiv-minmax-neon-x8.c",
1934 "src/f32-vbinary/gen/vdivc-minmax-neon-x4.c",
1935 "src/f32-vbinary/gen/vdivc-minmax-neon-x8.c",
1936 "src/f32-vbinary/gen/vrdivc-minmax-neon-x4.c",
1937 "src/f32-vbinary/gen/vrdivc-minmax-neon-x8.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07001938 "src/f32-vsqrt/gen/neon-sqrt-x4.c",
1939 "src/f32-vsqrt/gen/neon-sqrt-x8.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001940 "src/math/sigmoid-neonfma-rr1-lut64-p2-div.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001941 "src/math/sigmoid-neonfma-rr1-lut2048-p1-div.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001942 "src/math/sigmoid-neonfma-rr1-p5-div.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001943 "src/math/sigmoid-neonfma-rr2-lut64-p2-div.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001944 "src/math/sigmoid-neonfma-rr2-lut2048-p1-div.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001945 "src/math/sigmoid-neonfma-rr2-p5-div.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001946]
1947
Marat Dukhan8853b822020-05-07 12:19:01 -07001948NEONV8_UKERNELS = [
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07001949 "src/f32-vrnd/gen/vrndd-neonv8-x4.c",
1950 "src/f32-vrnd/gen/vrndd-neonv8-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001951 "src/f32-vrnd/gen/vrndne-neonv8-x4.c",
1952 "src/f32-vrnd/gen/vrndne-neonv8-x8.c",
1953 "src/f32-vrnd/gen/vrndu-neonv8-x4.c",
1954 "src/f32-vrnd/gen/vrndu-neonv8-x8.c",
1955 "src/f32-vrnd/gen/vrndz-neonv8-x4.c",
1956 "src/f32-vrnd/gen/vrndz-neonv8-x8.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001957 "src/math/roundd-neonv8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001958 "src/math/roundne-neonv8.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001959 "src/math/roundu-neonv8.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07001960 "src/math/roundz-neonv8.c",
Marat Dukhan8853b822020-05-07 12:19:01 -07001961]
1962
Marat Dukhan08c4a432019-10-03 09:29:21 -07001963AARCH64_NEONFP16ARITH_UKERNELS = [
Frank Barchardb1966592020-05-12 13:47:06 -07001964 "src/f16-clamp/gen/neonfp16arith-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001965 "src/f16-clamp/gen/neonfp16arith-x16.c",
Frank Barchard5a599a62020-06-04 20:12:44 -07001966 "src/f16-dwconv/gen/up8x4-minmax-neonfp16arith-acc2.c",
1967 "src/f16-dwconv/gen/up8x4-minmax-neonfp16arith.c",
1968 "src/f16-dwconv/gen/up8x9-minmax-neonfp16arith-acc2.c",
1969 "src/f16-dwconv/gen/up8x9-minmax-neonfp16arith.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001970 "src/f16-dwconv/gen/up8x25-minmax-neonfp16arith-acc2.c",
1971 "src/f16-dwconv/gen/up8x25-minmax-neonfp16arith.c",
1972 "src/f16-dwconv/gen/up16x4-minmax-neonfp16arith-acc2.c",
1973 "src/f16-dwconv/gen/up16x4-minmax-neonfp16arith.c",
1974 "src/f16-dwconv/gen/up16x9-minmax-neonfp16arith-acc2.c",
1975 "src/f16-dwconv/gen/up16x9-minmax-neonfp16arith.c",
1976 "src/f16-dwconv/gen/up16x25-minmax-neonfp16arith-acc2.c",
1977 "src/f16-dwconv/gen/up16x25-minmax-neonfp16arith.c",
Frank Barchard0bb49a72020-06-04 11:35:11 -07001978 "src/f16-gavgpool/7p7x-minmax-neonfp16arith-c8.c",
1979 "src/f16-gavgpool/7x-minmax-neonfp16arith-c8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001980 "src/f16-gemm/gen-inc/1x8inc-minmax-neonfp16arith-ld64.c",
1981 "src/f16-gemm/gen-inc/1x16inc-minmax-neonfp16arith-ld64.c",
1982 "src/f16-gemm/gen-inc/4x8inc-minmax-neonfp16arith-ld64.c",
1983 "src/f16-gemm/gen-inc/4x16inc-minmax-neonfp16arith-ld64.c",
1984 "src/f16-gemm/gen-inc/6x8inc-minmax-neonfp16arith-ld64.c",
1985 "src/f16-gemm/gen-inc/6x16inc-minmax-neonfp16arith-ld64.c",
1986 "src/f16-gemm/gen-inc/8x8inc-minmax-neonfp16arith-ld64.c",
1987 "src/f16-gemm/gen-inc/8x16inc-minmax-neonfp16arith-ld64.c",
1988 "src/f16-gemm/gen/1x8-minmax-neonfp16arith-ld64.c",
1989 "src/f16-gemm/gen/1x16-minmax-neonfp16arith-ld64.c",
1990 "src/f16-gemm/gen/4x8-minmax-neonfp16arith-ld64.c",
1991 "src/f16-gemm/gen/4x16-minmax-neonfp16arith-ld64.c",
1992 "src/f16-gemm/gen/6x8-minmax-neonfp16arith-ld64.c",
1993 "src/f16-gemm/gen/6x16-minmax-neonfp16arith-ld64.c",
1994 "src/f16-gemm/gen/8x8-minmax-neonfp16arith-ld64.c",
1995 "src/f16-gemm/gen/8x16-minmax-neonfp16arith-ld64.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07001996 "src/f16-hswish/gen/hswish-neonfp16arith-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001997 "src/f16-hswish/gen/hswish-neonfp16arith-x16.c",
1998 "src/f16-igemm/gen/1x8-minmax-neonfp16arith-ld64.c",
1999 "src/f16-igemm/gen/1x16-minmax-neonfp16arith-ld64.c",
2000 "src/f16-igemm/gen/4x8-minmax-neonfp16arith-ld64.c",
2001 "src/f16-igemm/gen/4x16-minmax-neonfp16arith-ld64.c",
2002 "src/f16-igemm/gen/6x8-minmax-neonfp16arith-ld64.c",
2003 "src/f16-igemm/gen/6x16-minmax-neonfp16arith-ld64.c",
2004 "src/f16-igemm/gen/8x8-minmax-neonfp16arith-ld64.c",
2005 "src/f16-igemm/gen/8x16-minmax-neonfp16arith-ld64.c",
Frank Barchardb1966592020-05-12 13:47:06 -07002006 "src/f16-prelu/gen/neonfp16arith-2x8.c",
Frank Barcharda5316982020-07-23 13:19:28 -07002007 "src/f16-prelu/gen/neonfp16arith-2x16.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07002008 "src/f16-relu/gen/neonfp16arith-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002009 "src/f16-relu/gen/neonfp16arith-x16.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07002010 "src/f16-spmm/gen/8x1-minmax-neonfp16arith-x2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002011 "src/f16-spmm/gen/8x1-minmax-neonfp16arith.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07002012 "src/f16-spmm/gen/16x1-minmax-neonfp16arith-x2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002013 "src/f16-spmm/gen/16x1-minmax-neonfp16arith.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07002014 "src/f16-spmm/gen/24x1-minmax-neonfp16arith-x2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002015 "src/f16-spmm/gen/24x1-minmax-neonfp16arith.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07002016 "src/f16-spmm/gen/32x1-minmax-neonfp16arith-x2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002017 "src/f16-spmm/gen/32x1-minmax-neonfp16arith.c",
2018 "src/f16-vbinary/gen/vadd-minmax-neonfp16arith-x8.c",
2019 "src/f16-vbinary/gen/vadd-minmax-neonfp16arith-x16.c",
2020 "src/f16-vbinary/gen/vaddc-minmax-neonfp16arith-x8.c",
2021 "src/f16-vbinary/gen/vaddc-minmax-neonfp16arith-x16.c",
2022 "src/f16-vbinary/gen/vdiv-minmax-neonfp16arith-x8.c",
2023 "src/f16-vbinary/gen/vdiv-minmax-neonfp16arith-x16.c",
2024 "src/f16-vbinary/gen/vdivc-minmax-neonfp16arith-x8.c",
2025 "src/f16-vbinary/gen/vdivc-minmax-neonfp16arith-x16.c",
2026 "src/f16-vbinary/gen/vmax-neonfp16arith-x8.c",
2027 "src/f16-vbinary/gen/vmax-neonfp16arith-x16.c",
2028 "src/f16-vbinary/gen/vmaxc-neonfp16arith-x8.c",
2029 "src/f16-vbinary/gen/vmaxc-neonfp16arith-x16.c",
2030 "src/f16-vbinary/gen/vmin-neonfp16arith-x8.c",
2031 "src/f16-vbinary/gen/vmin-neonfp16arith-x16.c",
2032 "src/f16-vbinary/gen/vminc-neonfp16arith-x8.c",
2033 "src/f16-vbinary/gen/vminc-neonfp16arith-x16.c",
2034 "src/f16-vbinary/gen/vmul-minmax-neonfp16arith-x8.c",
2035 "src/f16-vbinary/gen/vmul-minmax-neonfp16arith-x16.c",
2036 "src/f16-vbinary/gen/vmulc-minmax-neonfp16arith-x8.c",
2037 "src/f16-vbinary/gen/vmulc-minmax-neonfp16arith-x16.c",
2038 "src/f16-vbinary/gen/vrdivc-minmax-neonfp16arith-x8.c",
2039 "src/f16-vbinary/gen/vrdivc-minmax-neonfp16arith-x16.c",
2040 "src/f16-vbinary/gen/vrsubc-minmax-neonfp16arith-x8.c",
2041 "src/f16-vbinary/gen/vrsubc-minmax-neonfp16arith-x16.c",
2042 "src/f16-vbinary/gen/vsub-minmax-neonfp16arith-x8.c",
2043 "src/f16-vbinary/gen/vsub-minmax-neonfp16arith-x16.c",
2044 "src/f16-vbinary/gen/vsubc-minmax-neonfp16arith-x8.c",
2045 "src/f16-vbinary/gen/vsubc-minmax-neonfp16arith-x16.c",
2046 "src/f16-vmulcaddc/gen/c8-minmax-neonfp16arith-2x.c",
2047 "src/f16-vmulcaddc/gen/c16-minmax-neonfp16arith-2x.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002048]
2049
Benoit Jacoba9644732020-08-13 12:48:55 -07002050NEONDOT_UKERNELS = [
Marat Dukhand4c83032020-09-09 23:20:12 -07002051 "src/qs8-gemm/gen/1x8c4-minmax-neondot.c",
Marat Dukhand4c83032020-09-09 23:20:12 -07002052 "src/qs8-gemm/gen/1x16c4-minmax-neondot.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002053 "src/qs8-gemm/gen/4x8c4-minmax-neondot.c",
Benoit Jacoba9644732020-08-13 12:48:55 -07002054 "src/qs8-gemm/gen/4x16c4-minmax-neondot.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002055 "src/qs8-gemm/gen/6x8c4-minmax-neondot.c",
Frank Barchard2fa17452020-09-22 09:51:46 -07002056 "src/qs8-gemm/gen/6x16c4-minmax-neondot.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002057 "src/qs8-gemm/gen/8x8c4-minmax-neondot.c",
Frank Barcharda48848f2020-09-24 00:59:24 -07002058 "src/qs8-gemm/gen/8x16c4-minmax-neondot.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002059 "src/qs8-gemm/gen/12x8c4-minmax-neondot.c",
Marat Dukhand4c83032020-09-09 23:20:12 -07002060 "src/qs8-igemm/gen/1x8c4-minmax-neondot.c",
Marat Dukhand4c83032020-09-09 23:20:12 -07002061 "src/qs8-igemm/gen/1x16c4-minmax-neondot.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002062 "src/qs8-igemm/gen/4x8c4-minmax-neondot.c",
Benoit Jacoba05487f2020-08-14 13:37:21 -07002063 "src/qs8-igemm/gen/4x16c4-minmax-neondot.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002064 "src/qs8-igemm/gen/6x8c4-minmax-neondot.c",
Frank Barchard2fa17452020-09-22 09:51:46 -07002065 "src/qs8-igemm/gen/6x16c4-minmax-neondot.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002066 "src/qs8-igemm/gen/8x8c4-minmax-neondot.c",
Frank Barcharda48848f2020-09-24 00:59:24 -07002067 "src/qs8-igemm/gen/8x16c4-minmax-neondot.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002068 "src/qs8-igemm/gen/12x8c4-minmax-neondot.c",
Benoit Jacoba9644732020-08-13 12:48:55 -07002069]
2070
Marat Dukhan08c4a432019-10-03 09:29:21 -07002071SSE_UKERNELS = [
Marat Dukhan99936602020-04-11 16:47:01 -07002072 "src/f32-avgpool/9p8x-minmax-sse-c4.c",
2073 "src/f32-avgpool/9x-minmax-sse-c4.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -07002074 "src/f32-clamp/gen/sse-x4.c",
2075 "src/f32-clamp/gen/sse-x8.c",
Erich Elsenb1233402020-06-08 15:53:15 -07002076 "src/f32-conv-hwc2chw/3x3s2p1c3x4-sse-1x1.c",
2077 "src/f32-conv-hwc2chw/3x3s2p1c3x4-sse-2x2.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002078 "src/f32-dwconv/gen/up4x4-minmax-sse-acc2.c",
2079 "src/f32-dwconv/gen/up4x4-minmax-sse.c",
2080 "src/f32-dwconv/gen/up4x9-minmax-sse-acc2.c",
2081 "src/f32-dwconv/gen/up4x9-minmax-sse.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002082 "src/f32-dwconv/gen/up4x25-minmax-sse-acc2.c",
2083 "src/f32-dwconv/gen/up4x25-minmax-sse.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002084 "src/f32-dwconv/gen/up8x4-minmax-sse-acc2.c",
2085 "src/f32-dwconv/gen/up8x4-minmax-sse.c",
2086 "src/f32-dwconv/gen/up8x9-minmax-sse-acc2.c",
2087 "src/f32-dwconv/gen/up8x9-minmax-sse.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002088 "src/f32-dwconv/gen/up8x25-minmax-sse-acc2.c",
2089 "src/f32-dwconv/gen/up8x25-minmax-sse.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07002090 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-1x4-acc2.c",
2091 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-1x4-acc3.c",
2092 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-1x4-acc4.c",
Marat Dukhan470078a2020-10-23 22:36:52 -07002093 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-1x4.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07002094 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-2x4-acc2.c",
Marat Dukhan470078a2020-10-23 22:36:52 -07002095 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-2x4.c",
2096 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-3x4.c",
2097 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-4x4.c",
2098 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-5x4.c",
2099 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-6x4.c",
Marat Dukhan0ff97182020-10-25 19:14:03 -07002100 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-sse-1x4-acc2.c",
2101 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-sse-1x4-acc3.c",
2102 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-sse-1x4-acc4.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07002103 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-sse-1x4.c",
Marat Dukhan0ff97182020-10-25 19:14:03 -07002104 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-sse-2x4-acc2.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07002105 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-sse-2x4.c",
2106 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-sse-3x4.c",
2107 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-sse-4x4.c",
Marat Dukhand0503892020-10-30 08:22:04 -07002108 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-1x4-acc2.c",
2109 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-1x4-acc3.c",
2110 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-1x4-acc4.c",
2111 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-1x4-acc5.c",
2112 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-1x4.c",
2113 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-2x4-acc2.c",
2114 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-2x4-acc3.c",
2115 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-2x4.c",
2116 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-3x4-acc2.c",
2117 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-3x4.c",
2118 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-4x4-acc2.c",
2119 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-4x4.c",
2120 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-5x4.c",
Marat Dukhanccca2142020-10-30 17:32:45 -07002121 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-sse-1x4-acc2.c",
2122 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-sse-1x4-acc3.c",
2123 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-sse-1x4-acc4.c",
2124 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-sse-1x4-acc5.c",
2125 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-sse-1x4.c",
2126 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-sse-2x4-acc2.c",
2127 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-sse-2x4-acc3.c",
2128 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-sse-2x4.c",
Marat Dukhanccca2142020-10-30 17:32:45 -07002129 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-sse-3x4-acc2.c",
Frank Barchard8ef44cd2020-11-03 12:30:23 -08002130 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-sse-3x4.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -07002131 "src/f32-gavgpool-cw/sse-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -07002132 "src/f32-gavgpool/7p7x-minmax-sse-c4.c",
2133 "src/f32-gavgpool/7x-minmax-sse-c4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002134 "src/f32-gemm/gen-inc/1x8inc-minmax-sse-dup.c",
2135 "src/f32-gemm/gen-inc/1x8inc-minmax-sse-load1.c",
2136 "src/f32-gemm/gen-inc/1x8s4inc-minmax-sse.c",
2137 "src/f32-gemm/gen-inc/4x8inc-minmax-sse-dup.c",
2138 "src/f32-gemm/gen-inc/4x8inc-minmax-sse-load1.c",
2139 "src/f32-gemm/gen-inc/4x8s4inc-minmax-sse.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002140 "src/f32-gemm/gen/1x8-minmax-sse-dup.c",
2141 "src/f32-gemm/gen/1x8-minmax-sse-load1.c",
2142 "src/f32-gemm/gen/1x8s4-minmax-sse.c",
2143 "src/f32-gemm/gen/4x2c4-minmax-sse.c",
2144 "src/f32-gemm/gen/4x8-minmax-sse-dup.c",
2145 "src/f32-gemm/gen/4x8-minmax-sse-load1.c",
2146 "src/f32-gemm/gen/4x8s4-minmax-sse.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07002147 "src/f32-hswish/gen/hswish-sse-x4.c",
2148 "src/f32-hswish/gen/hswish-sse-x8.c",
Marat Dukhan660fd192020-03-10 04:55:30 -07002149 "src/f32-ibilinear/gen/sse-c4.c",
2150 "src/f32-ibilinear/gen/sse-c8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002151 "src/f32-igemm/gen/1x8-minmax-sse-dup.c",
2152 "src/f32-igemm/gen/1x8-minmax-sse-load1.c",
2153 "src/f32-igemm/gen/1x8s4-minmax-sse.c",
2154 "src/f32-igemm/gen/4x2c4-minmax-sse.c",
2155 "src/f32-igemm/gen/4x8-minmax-sse-dup.c",
2156 "src/f32-igemm/gen/4x8-minmax-sse-load1.c",
2157 "src/f32-igemm/gen/4x8s4-minmax-sse.c",
Marat Dukhan99936602020-04-11 16:47:01 -07002158 "src/f32-maxpool/9p8x-minmax-sse-c4.c",
2159 "src/f32-pavgpool/9p8x-minmax-sse-c4.c",
2160 "src/f32-pavgpool/9x-minmax-sse-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002161 "src/f32-ppmm/gen/4x8-minmax-sse.c",
Marat Dukhan39b5e942020-06-24 15:03:48 -07002162 "src/f32-prelu/gen/sse-2x4.c",
2163 "src/f32-prelu/gen/sse-2x8.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07002164 "src/f32-relu/gen/sse-x4.c",
2165 "src/f32-relu/gen/sse-x8.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002166 "src/f32-rmax/sse.c",
Marat Dukhan355ab432020-04-09 19:01:52 -07002167 "src/f32-spmm/gen/4x1-minmax-sse.c",
2168 "src/f32-spmm/gen/8x1-minmax-sse.c",
Erich Elsen6e80fdc2020-06-09 15:35:37 -07002169 "src/f32-spmm/gen/16x1-minmax-sse.c",
Frank Barchard846c0c62020-10-26 15:01:39 -07002170 "src/f32-spmm/gen/32x1-minmax-sse.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002171 "src/f32-vbinary/gen/vadd-minmax-sse-x4.c",
2172 "src/f32-vbinary/gen/vadd-minmax-sse-x8.c",
2173 "src/f32-vbinary/gen/vaddc-minmax-sse-x4.c",
2174 "src/f32-vbinary/gen/vaddc-minmax-sse-x8.c",
2175 "src/f32-vbinary/gen/vdiv-minmax-sse-x4.c",
2176 "src/f32-vbinary/gen/vdiv-minmax-sse-x8.c",
2177 "src/f32-vbinary/gen/vdivc-minmax-sse-x4.c",
2178 "src/f32-vbinary/gen/vdivc-minmax-sse-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -08002179 "src/f32-vbinary/gen/vmax-sse-x4.c",
2180 "src/f32-vbinary/gen/vmax-sse-x8.c",
2181 "src/f32-vbinary/gen/vmaxc-sse-x4.c",
2182 "src/f32-vbinary/gen/vmaxc-sse-x8.c",
2183 "src/f32-vbinary/gen/vmin-sse-x4.c",
2184 "src/f32-vbinary/gen/vmin-sse-x8.c",
2185 "src/f32-vbinary/gen/vminc-sse-x4.c",
2186 "src/f32-vbinary/gen/vminc-sse-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002187 "src/f32-vbinary/gen/vmul-minmax-sse-x4.c",
2188 "src/f32-vbinary/gen/vmul-minmax-sse-x8.c",
2189 "src/f32-vbinary/gen/vmulc-minmax-sse-x4.c",
2190 "src/f32-vbinary/gen/vmulc-minmax-sse-x8.c",
2191 "src/f32-vbinary/gen/vrdivc-minmax-sse-x4.c",
2192 "src/f32-vbinary/gen/vrdivc-minmax-sse-x8.c",
2193 "src/f32-vbinary/gen/vrsubc-minmax-sse-x4.c",
2194 "src/f32-vbinary/gen/vrsubc-minmax-sse-x8.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07002195 "src/f32-vbinary/gen/vsqrdiff-sse-x4.c",
2196 "src/f32-vbinary/gen/vsqrdiff-sse-x8.c",
2197 "src/f32-vbinary/gen/vsqrdiffc-sse-x4.c",
2198 "src/f32-vbinary/gen/vsqrdiffc-sse-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002199 "src/f32-vbinary/gen/vsub-minmax-sse-x4.c",
2200 "src/f32-vbinary/gen/vsub-minmax-sse-x8.c",
2201 "src/f32-vbinary/gen/vsubc-minmax-sse-x4.c",
2202 "src/f32-vbinary/gen/vsubc-minmax-sse-x8.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07002203 "src/f32-vlrelu/gen/vlrelu-sse-x4.c",
2204 "src/f32-vlrelu/gen/vlrelu-sse-x8.c",
Marat Dukhan99936602020-04-11 16:47:01 -07002205 "src/f32-vmulcaddc/gen/c4-minmax-sse-2x.c",
2206 "src/f32-vmulcaddc/gen/c8-minmax-sse-2x.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07002207 "src/f32-vsqrt/gen/sse-sqrt-x4.c",
2208 "src/f32-vsqrt/gen/sse-sqrt-x8.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07002209 "src/f32-vunary/gen/vabs-sse-x4.c",
2210 "src/f32-vunary/gen/vabs-sse-x8.c",
2211 "src/f32-vunary/gen/vneg-sse-x4.c",
2212 "src/f32-vunary/gen/vneg-sse-x8.c",
2213 "src/f32-vunary/gen/vsqr-sse-x4.c",
2214 "src/f32-vunary/gen/vsqr-sse-x8.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07002215 "src/math/roundd-sse-addsub.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002216 "src/math/roundne-sse-addsub.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07002217 "src/math/roundu-sse-addsub.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07002218 "src/math/roundz-sse-addsub.c",
Marat Dukhan84000762020-06-29 18:38:43 -07002219 "src/math/sqrt-sse-hh1mac.c",
2220 "src/math/sqrt-sse-nr1mac.c",
2221 "src/math/sqrt-sse-nr2mac.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002222 "src/x32-fill/sse.c",
2223 "src/x32-packx/x4-sse.c",
2224 "src/x32-pad/sse.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002225]
2226
2227SSE2_UKERNELS = [
Marat Dukhan329da642019-11-19 21:44:39 -08002228 "src/f32-argmaxpool/4x-sse2-c4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002229 "src/f32-argmaxpool/9p8x-sse2-c4.c",
Marat Dukhan329da642019-11-19 21:44:39 -08002230 "src/f32-argmaxpool/9x-sse2-c4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08002231 "src/f32-prelu/gen/sse2-2x4.c",
2232 "src/f32-prelu/gen/sse2-2x8.c",
Marat Dukhanb39689d2020-01-24 13:32:20 -08002233 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x4.c",
Marat Dukhanb39689d2020-01-24 13:32:20 -08002234 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x8-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002235 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x8.c",
Marat Dukhanb39689d2020-01-24 13:32:20 -08002236 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12-acc2.c",
2237 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002238 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12.c",
Marat Dukhanb39689d2020-01-24 13:32:20 -08002239 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16-acc2.c",
2240 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002241 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16.c",
Marat Dukhanb39689d2020-01-24 13:32:20 -08002242 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20-acc2.c",
2243 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002244 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20.c",
Marat Dukhand243c1a2020-09-17 18:48:10 -07002245 "src/f32-sigmoid/gen/sse2-lut64-p2-div-x4.c",
2246 "src/f32-sigmoid/gen/sse2-lut64-p2-div-x8.c",
2247 "src/f32-sigmoid/gen/sse2-lut64-p2-div-x12.c",
2248 "src/f32-sigmoid/gen/sse2-lut64-p2-div-x16.c",
2249 "src/f32-sigmoid/gen/sse2-lut64-p2-div-x20.c",
2250 "src/f32-sigmoid/gen/sse2-lut64-p2-div-x24.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002251 "src/f32-sigmoid/gen/sse2-p5-div-x4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08002252 "src/f32-sigmoid/gen/sse2-p5-div-x8.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002253 "src/f32-sigmoid/gen/sse2-p5-div-x12.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08002254 "src/f32-sigmoid/gen/sse2-p5-div-x16.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002255 "src/f32-sigmoid/gen/sse2-p5-div-x20.c",
2256 "src/f32-sigmoid/gen/sse2-p5-div-x24.c",
Marat Dukhaned6baaf2020-12-01 15:07:08 -08002257 "src/f32-velu/gen/velu-sse2-rr2-lut16-p3-x4.c",
2258 "src/f32-velu/gen/velu-sse2-rr2-lut16-p3-x8.c",
2259 "src/f32-velu/gen/velu-sse2-rr2-lut16-p3-x12.c",
2260 "src/f32-velu/gen/velu-sse2-rr2-lut16-p3-x16.c",
2261 "src/f32-velu/gen/velu-sse2-rr2-lut16-p3-x20.c",
2262 "src/f32-velu/gen/velu-sse2-rr2-lut16-p3-x24.c",
2263 "src/f32-velu/gen/velu-sse2-rr2-p6-x4.c",
2264 "src/f32-velu/gen/velu-sse2-rr2-p6-x8.c",
2265 "src/f32-velu/gen/velu-sse2-rr2-p6-x12.c",
2266 "src/f32-velu/gen/velu-sse2-rr2-p6-x16.c",
2267 "src/f32-velu/gen/velu-sse2-rr2-p6-x20.c",
2268 "src/f32-velu/gen/velu-sse2-rr2-p6-x24.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07002269 "src/f32-vlrelu/gen/vlrelu-sse2-x4.c",
2270 "src/f32-vlrelu/gen/vlrelu-sse2-x8.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07002271 "src/f32-vrnd/gen/vrndd-sse2-x4.c",
2272 "src/f32-vrnd/gen/vrndd-sse2-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002273 "src/f32-vrnd/gen/vrndne-sse2-x4.c",
2274 "src/f32-vrnd/gen/vrndne-sse2-x8.c",
2275 "src/f32-vrnd/gen/vrndu-sse2-x4.c",
2276 "src/f32-vrnd/gen/vrndu-sse2-x8.c",
2277 "src/f32-vrnd/gen/vrndz-sse2-x4.c",
2278 "src/f32-vrnd/gen/vrndz-sse2-x8.c",
Marat Dukhanb7633f22020-11-20 16:34:56 -08002279 "src/math/exp-sse2-rr2-lut64-p2.c",
2280 "src/math/exp-sse2-rr2-p5.c",
Marat Dukhande390d42020-11-29 19:32:18 -08002281 "src/math/expm1minus-sse2-rr2-lut16-p3.c",
Marat Dukhana438aca2020-11-20 15:45:01 -08002282 "src/math/expm1minus-sse2-rr2-p6.c",
Frank Barchard3b800452020-11-22 12:12:35 -08002283 "src/math/expminus-sse2-rr2-p5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002284 "src/math/roundd-sse2-cvt.c",
2285 "src/math/roundne-sse2-cvt.c",
2286 "src/math/roundu-sse2-cvt.c",
2287 "src/math/roundz-sse2-cvt.c",
2288 "src/math/sigmoid-sse2-rr2-lut64-p2-div.c",
2289 "src/math/sigmoid-sse2-rr2-lut64-p2-nr1.c",
2290 "src/math/sigmoid-sse2-rr2-lut64-p2-nr2.c",
2291 "src/math/sigmoid-sse2-rr2-p5-div.c",
2292 "src/math/sigmoid-sse2-rr2-p5-nr1.c",
2293 "src/math/sigmoid-sse2-rr2-p5-nr2.c",
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07002294 "src/qs8-dwconv/gen/up8x9-minmax-sse2-mul16.c",
2295 "src/qs8-dwconv/gen/up16x9-minmax-sse2-mul16.c",
2296 "src/qs8-dwconv/gen/up24x9-minmax-sse2-mul16.c",
Marat Dukhan159688f2020-08-06 10:34:29 -07002297 "src/qs8-gavgpool/gen/7p7x-minmax-sse2-c8-acc2.c",
2298 "src/qs8-gavgpool/gen/7p7x-minmax-sse2-c16-acc2.c",
2299 "src/qs8-gavgpool/gen/7p7x-minmax-sse2-c24-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002300 "src/qs8-gavgpool/gen/7x-minmax-sse2-c8-acc2.c",
2301 "src/qs8-gavgpool/gen/7x-minmax-sse2-c16-acc2.c",
2302 "src/qs8-gavgpool/gen/7x-minmax-sse2-c24-acc2.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002303 "src/qs8-gemm/gen/1x4c2-minmax-sse2-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002304 "src/qs8-gemm/gen/1x4c2-minmax-sse2-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002305 "src/qs8-gemm/gen/1x4c2-xw-minmax-sse2.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002306 "src/qs8-gemm/gen/1x4c8-minmax-sse2-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002307 "src/qs8-gemm/gen/1x4c8-minmax-sse2-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002308 "src/qs8-gemm/gen/1x4c8-xw-minmax-sse2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002309 "src/qs8-gemm/gen/2x4c8-minmax-sse2-ld64.c",
2310 "src/qs8-gemm/gen/2x4c8-minmax-sse2-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002311 "src/qs8-gemm/gen/2x4c8-xw-minmax-sse2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002312 "src/qs8-gemm/gen/3x4c8-minmax-sse2-ld64.c",
2313 "src/qs8-gemm/gen/3x4c8-minmax-sse2-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002314 "src/qs8-gemm/gen/3x4c8-xw-minmax-sse2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002315 "src/qs8-gemm/gen/4x4c2-minmax-sse2-ld64.c",
2316 "src/qs8-gemm/gen/4x4c2-minmax-sse2-ld128.c",
2317 "src/qs8-gemm/gen/4x4c2-xw-minmax-sse2.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002318 "src/qs8-igemm/gen/1x4c2-minmax-sse2-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002319 "src/qs8-igemm/gen/1x4c2-minmax-sse2-ld128.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002320 "src/qs8-igemm/gen/1x4c8-minmax-sse2-ld64.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002321 "src/qs8-igemm/gen/1x4c8-minmax-sse2-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002322 "src/qs8-igemm/gen/2x4c8-minmax-sse2-ld64.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002323 "src/qs8-igemm/gen/2x4c8-minmax-sse2-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002324 "src/qs8-igemm/gen/3x4c8-minmax-sse2-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002325 "src/qs8-igemm/gen/3x4c8-minmax-sse2-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002326 "src/qs8-igemm/gen/4x4c2-minmax-sse2-ld64.c",
2327 "src/qs8-igemm/gen/4x4c2-minmax-sse2-ld128.c",
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07002328 "src/qs8-requantization/fp32-sse2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002329 "src/qs8-requantization/precise-sse2.c",
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07002330 "src/qs8-requantization/q31-sse2.c",
Marat Dukhand9f3ad42020-08-10 12:30:58 -07002331 "src/qs8-vadd/gen/minmax-sse2-mul16-ld64-x8.c",
2332 "src/qs8-vadd/gen/minmax-sse2-mul16-ld64-x16.c",
2333 "src/qs8-vadd/gen/minmax-sse2-mul16-ld64-x24.c",
2334 "src/qs8-vadd/gen/minmax-sse2-mul16-ld64-x32.c",
Marat Dukhan0270d9f2020-08-11 00:56:46 -07002335 "src/qs8-vaddc/gen/minmax-sse2-mul16-ld64-x8.c",
2336 "src/qs8-vaddc/gen/minmax-sse2-mul16-ld64-x16.c",
2337 "src/qs8-vaddc/gen/minmax-sse2-mul16-ld64-x24.c",
2338 "src/qs8-vaddc/gen/minmax-sse2-mul16-ld64-x32.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -07002339 "src/qu8-avgpool/9p8x-minmax-sse2-c8.c",
2340 "src/qu8-avgpool/9x-minmax-sse2-c8.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -07002341 "src/qu8-dwconv/up8x9-minmax-sse2.c",
2342 "src/qu8-gavgpool/7p7x-minmax-sse2-c8.c",
2343 "src/qu8-gavgpool/7x-minmax-sse2-c8.c",
2344 "src/qu8-gemm/2x4c8-minmax-sse2.c",
2345 "src/qu8-gemm/4x4c2-minmax-sse2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002346 "src/qu8-igemm/4x4c2-minmax-sse2.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07002347 "src/qu8-requantization/fp32-sse2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002348 "src/qu8-requantization/precise-sse2.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07002349 "src/qu8-requantization/q31-sse2.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -07002350 "src/qu8-vadd/minmax-sse2.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -07002351 "src/u8-clamp/sse2-x64.c",
Marat Dukhan99936602020-04-11 16:47:01 -07002352 "src/u8-maxpool/9p8x-minmax-sse2-c16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002353 "src/u8-rmax/sse2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002354 "src/x8-zip/x2-sse2.c",
2355 "src/x8-zip/x3-sse2.c",
2356 "src/x8-zip/x4-sse2.c",
2357 "src/x8-zip/xm-sse2.c",
Marat Dukhan57dccd82020-04-14 00:53:10 -07002358 "src/x32-unpool/sse2.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002359 "src/x32-zip/x2-sse2.c",
2360 "src/x32-zip/x3-sse2.c",
2361 "src/x32-zip/x4-sse2.c",
2362 "src/x32-zip/xm-sse2.c",
Marat Dukhanfe7acb62020-03-09 19:30:05 -07002363]
2364
2365SSSE3_UKERNELS = [
Frank Barchard35db7d02020-10-26 13:37:34 -07002366 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-1x4-acc2.c",
2367 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-1x4-acc3.c",
2368 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-1x4-acc4.c",
Marat Dukhan98f2eeb2020-10-23 23:13:41 -07002369 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-1x4.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07002370 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-2x4-acc2.c",
Marat Dukhan98f2eeb2020-10-23 23:13:41 -07002371 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-2x4.c",
2372 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-3x4.c",
2373 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-4x4.c",
2374 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-5x4.c",
2375 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-6x4.c",
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07002376 "src/qs8-dwconv/gen/up8x9-minmax-ssse3-mul16.c",
2377 "src/qs8-dwconv/gen/up16x9-minmax-ssse3-mul16.c",
2378 "src/qs8-dwconv/gen/up24x9-minmax-ssse3-mul16.c",
Marat Dukhan159688f2020-08-06 10:34:29 -07002379 "src/qs8-gavgpool/gen/7p7x-minmax-ssse3-c8-acc2.c",
2380 "src/qs8-gavgpool/gen/7p7x-minmax-ssse3-c16-acc2.c",
2381 "src/qs8-gavgpool/gen/7p7x-minmax-ssse3-c24-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002382 "src/qs8-gavgpool/gen/7x-minmax-ssse3-c8-acc2.c",
2383 "src/qs8-gavgpool/gen/7x-minmax-ssse3-c16-acc2.c",
2384 "src/qs8-gavgpool/gen/7x-minmax-ssse3-c24-acc2.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002385 "src/qs8-gemm/gen/1x4c2-minmax-ssse3-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002386 "src/qs8-gemm/gen/1x4c2-minmax-ssse3-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002387 "src/qs8-gemm/gen/1x4c2-xw-minmax-ssse3.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002388 "src/qs8-gemm/gen/1x4c8-minmax-ssse3-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002389 "src/qs8-gemm/gen/1x4c8-minmax-ssse3-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002390 "src/qs8-gemm/gen/1x4c8-xw-minmax-ssse3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002391 "src/qs8-gemm/gen/2x4c8-minmax-ssse3-ld64.c",
2392 "src/qs8-gemm/gen/2x4c8-minmax-ssse3-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002393 "src/qs8-gemm/gen/2x4c8-xw-minmax-ssse3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002394 "src/qs8-gemm/gen/3x4c8-minmax-ssse3-ld64.c",
2395 "src/qs8-gemm/gen/3x4c8-minmax-ssse3-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002396 "src/qs8-gemm/gen/3x4c8-xw-minmax-ssse3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002397 "src/qs8-gemm/gen/4x4c2-minmax-ssse3-ld64.c",
2398 "src/qs8-gemm/gen/4x4c2-minmax-ssse3-ld128.c",
2399 "src/qs8-gemm/gen/4x4c2-xw-minmax-ssse3.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002400 "src/qs8-igemm/gen/1x4c2-minmax-ssse3-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002401 "src/qs8-igemm/gen/1x4c2-minmax-ssse3-ld128.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002402 "src/qs8-igemm/gen/1x4c8-minmax-ssse3-ld64.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002403 "src/qs8-igemm/gen/1x4c8-minmax-ssse3-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002404 "src/qs8-igemm/gen/2x4c8-minmax-ssse3-ld64.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002405 "src/qs8-igemm/gen/2x4c8-minmax-ssse3-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002406 "src/qs8-igemm/gen/3x4c8-minmax-ssse3-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002407 "src/qs8-igemm/gen/3x4c8-minmax-ssse3-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002408 "src/qs8-igemm/gen/4x4c2-minmax-ssse3-ld64.c",
2409 "src/qs8-igemm/gen/4x4c2-minmax-ssse3-ld128.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -07002410 "src/qs8-requantization/precise-ssse3.c",
2411 "src/qs8-requantization/q31-ssse3.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07002412 "src/qu8-requantization/precise-ssse3.c",
2413 "src/qu8-requantization/q31-ssse3.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002414]
2415
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08002416SSE41_UKERNELS = [
Marat Dukhan40a672f2019-11-25 03:08:22 -08002417 "src/f32-prelu/gen/sse41-2x4.c",
2418 "src/f32-prelu/gen/sse41-2x8.c",
Marat Dukhand243c1a2020-09-17 18:48:10 -07002419 "src/f32-sigmoid/gen/sse41-lut64-p2-div-x4.c",
2420 "src/f32-sigmoid/gen/sse41-lut64-p2-div-x8.c",
2421 "src/f32-sigmoid/gen/sse41-lut64-p2-div-x12.c",
2422 "src/f32-sigmoid/gen/sse41-lut64-p2-div-x16.c",
2423 "src/f32-sigmoid/gen/sse41-lut64-p2-div-x20.c",
2424 "src/f32-sigmoid/gen/sse41-lut64-p2-div-x24.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002425 "src/f32-sigmoid/gen/sse41-p5-div-x4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08002426 "src/f32-sigmoid/gen/sse41-p5-div-x8.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002427 "src/f32-sigmoid/gen/sse41-p5-div-x12.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08002428 "src/f32-sigmoid/gen/sse41-p5-div-x16.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002429 "src/f32-sigmoid/gen/sse41-p5-div-x20.c",
2430 "src/f32-sigmoid/gen/sse41-p5-div-x24.c",
Marat Dukhaned6baaf2020-12-01 15:07:08 -08002431 "src/f32-velu/gen/velu-sse41-rr2-lut16-p3-x4.c",
2432 "src/f32-velu/gen/velu-sse41-rr2-lut16-p3-x8.c",
2433 "src/f32-velu/gen/velu-sse41-rr2-lut16-p3-x12.c",
2434 "src/f32-velu/gen/velu-sse41-rr2-lut16-p3-x16.c",
2435 "src/f32-velu/gen/velu-sse41-rr2-lut16-p3-x20.c",
2436 "src/f32-velu/gen/velu-sse41-rr2-lut16-p3-x24.c",
2437 "src/f32-velu/gen/velu-sse41-rr2-p6-x4.c",
2438 "src/f32-velu/gen/velu-sse41-rr2-p6-x8.c",
2439 "src/f32-velu/gen/velu-sse41-rr2-p6-x12.c",
2440 "src/f32-velu/gen/velu-sse41-rr2-p6-x16.c",
2441 "src/f32-velu/gen/velu-sse41-rr2-p6-x20.c",
2442 "src/f32-velu/gen/velu-sse41-rr2-p6-x24.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07002443 "src/f32-vlrelu/gen/vlrelu-sse41-x4.c",
2444 "src/f32-vlrelu/gen/vlrelu-sse41-x8.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07002445 "src/f32-vrnd/gen/vrndd-sse41-x4.c",
2446 "src/f32-vrnd/gen/vrndd-sse41-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002447 "src/f32-vrnd/gen/vrndne-sse41-x4.c",
2448 "src/f32-vrnd/gen/vrndne-sse41-x8.c",
2449 "src/f32-vrnd/gen/vrndu-sse41-x4.c",
2450 "src/f32-vrnd/gen/vrndu-sse41-x8.c",
2451 "src/f32-vrnd/gen/vrndz-sse41-x4.c",
2452 "src/f32-vrnd/gen/vrndz-sse41-x8.c",
2453 "src/math/roundd-sse41.c",
2454 "src/math/roundne-sse41.c",
2455 "src/math/roundu-sse41.c",
2456 "src/math/roundz-sse41.c",
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07002457 "src/qs8-dwconv/gen/up8x9-minmax-sse41-mul16.c",
2458 "src/qs8-dwconv/gen/up16x9-minmax-sse41-mul16.c",
2459 "src/qs8-dwconv/gen/up24x9-minmax-sse41-mul16.c",
Marat Dukhan159688f2020-08-06 10:34:29 -07002460 "src/qs8-gavgpool/gen/7p7x-minmax-sse41-c8-acc2.c",
2461 "src/qs8-gavgpool/gen/7p7x-minmax-sse41-c16-acc2.c",
2462 "src/qs8-gavgpool/gen/7p7x-minmax-sse41-c24-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002463 "src/qs8-gavgpool/gen/7x-minmax-sse41-c8-acc2.c",
2464 "src/qs8-gavgpool/gen/7x-minmax-sse41-c16-acc2.c",
2465 "src/qs8-gavgpool/gen/7x-minmax-sse41-c24-acc2.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002466 "src/qs8-gemm/gen/1x4c2-minmax-sse41-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002467 "src/qs8-gemm/gen/1x4c2-minmax-sse41-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002468 "src/qs8-gemm/gen/1x4c2-xw-minmax-sse41.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002469 "src/qs8-gemm/gen/1x4c8-minmax-sse41-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002470 "src/qs8-gemm/gen/1x4c8-minmax-sse41-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002471 "src/qs8-gemm/gen/1x4c8-xw-minmax-sse41.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002472 "src/qs8-gemm/gen/2x4c8-minmax-sse41-ld64.c",
2473 "src/qs8-gemm/gen/2x4c8-minmax-sse41-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002474 "src/qs8-gemm/gen/2x4c8-xw-minmax-sse41.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002475 "src/qs8-gemm/gen/3x4c8-minmax-sse41-ld64.c",
2476 "src/qs8-gemm/gen/3x4c8-minmax-sse41-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002477 "src/qs8-gemm/gen/3x4c8-xw-minmax-sse41.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002478 "src/qs8-gemm/gen/4x4c2-minmax-sse41-ld64.c",
2479 "src/qs8-gemm/gen/4x4c2-minmax-sse41-ld128.c",
2480 "src/qs8-gemm/gen/4x4c2-xw-minmax-sse41.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002481 "src/qs8-igemm/gen/1x4c2-minmax-sse41-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002482 "src/qs8-igemm/gen/1x4c2-minmax-sse41-ld128.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002483 "src/qs8-igemm/gen/1x4c8-minmax-sse41-ld64.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002484 "src/qs8-igemm/gen/1x4c8-minmax-sse41-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002485 "src/qs8-igemm/gen/2x4c8-minmax-sse41-ld64.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002486 "src/qs8-igemm/gen/2x4c8-minmax-sse41-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002487 "src/qs8-igemm/gen/3x4c8-minmax-sse41-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002488 "src/qs8-igemm/gen/3x4c8-minmax-sse41-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002489 "src/qs8-igemm/gen/4x4c2-minmax-sse41-ld64.c",
2490 "src/qs8-igemm/gen/4x4c2-minmax-sse41-ld128.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -07002491 "src/qs8-requantization/fp32-sse4.c",
2492 "src/qs8-requantization/precise-sse4.c",
2493 "src/qs8-requantization/q31-sse4.c",
Marat Dukhand9f3ad42020-08-10 12:30:58 -07002494 "src/qs8-vadd/gen/minmax-sse41-mul16-ld64-x8.c",
2495 "src/qs8-vadd/gen/minmax-sse41-mul16-ld64-x16.c",
2496 "src/qs8-vadd/gen/minmax-sse41-mul16-ld64-x24.c",
2497 "src/qs8-vadd/gen/minmax-sse41-mul16-ld64-x32.c",
Marat Dukhanbb9225e2020-09-06 22:40:56 -07002498 "src/qs8-vadd/gen/minmax-sse41-mul32-ld32-x8.c",
2499 "src/qs8-vadd/gen/minmax-sse41-mul32-ld32-x16.c",
2500 "src/qs8-vadd/gen/minmax-sse41-mul32-ld32-x24.c",
2501 "src/qs8-vadd/gen/minmax-sse41-mul32-ld32-x32.c",
Marat Dukhan0270d9f2020-08-11 00:56:46 -07002502 "src/qs8-vaddc/gen/minmax-sse41-mul16-ld64-x8.c",
2503 "src/qs8-vaddc/gen/minmax-sse41-mul16-ld64-x16.c",
2504 "src/qs8-vaddc/gen/minmax-sse41-mul16-ld64-x24.c",
2505 "src/qs8-vaddc/gen/minmax-sse41-mul16-ld64-x32.c",
Marat Dukhanbb9225e2020-09-06 22:40:56 -07002506 "src/qs8-vaddc/gen/minmax-sse41-mul32-ld32-x8.c",
2507 "src/qs8-vaddc/gen/minmax-sse41-mul32-ld32-x16.c",
2508 "src/qs8-vaddc/gen/minmax-sse41-mul32-ld32-x24.c",
2509 "src/qs8-vaddc/gen/minmax-sse41-mul32-ld32-x32.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07002510 "src/qu8-requantization/precise-sse4.c",
2511 "src/qu8-requantization/q31-sse4.c",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08002512]
2513
Marat Dukhan08c4a432019-10-03 09:29:21 -07002514AVX_UKERNELS = [
Marat Dukhan5c5fa962020-03-10 18:38:33 -07002515 "src/f32-clamp/gen/avx-x8.c",
2516 "src/f32-clamp/gen/avx-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002517 "src/f32-dwconv/gen/up8x4-minmax-avx-acc2.c",
2518 "src/f32-dwconv/gen/up8x4-minmax-avx.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002519 "src/f32-dwconv/gen/up8x9-minmax-avx-acc2.c",
2520 "src/f32-dwconv/gen/up8x9-minmax-avx.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002521 "src/f32-dwconv/gen/up8x25-minmax-avx-acc2.c",
2522 "src/f32-dwconv/gen/up8x25-minmax-avx.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002523 "src/f32-dwconv/gen/up16x4-minmax-avx-acc2.c",
2524 "src/f32-dwconv/gen/up16x4-minmax-avx.c",
2525 "src/f32-dwconv/gen/up16x9-minmax-avx-acc2.c",
2526 "src/f32-dwconv/gen/up16x9-minmax-avx.c",
2527 "src/f32-dwconv/gen/up16x25-minmax-avx-acc2.c",
2528 "src/f32-dwconv/gen/up16x25-minmax-avx.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002529 "src/f32-gemm/gen-inc/1x8inc-minmax-avx-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002530 "src/f32-gemm/gen-inc/1x16inc-minmax-avx-broadcast.c",
2531 "src/f32-gemm/gen-inc/3x16inc-minmax-avx-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002532 "src/f32-gemm/gen-inc/4x8inc-minmax-avx-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002533 "src/f32-gemm/gen-inc/4x16inc-minmax-avx-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002534 "src/f32-gemm/gen-inc/5x8inc-minmax-avx-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002535 "src/f32-gemm/gen-inc/5x16inc-minmax-avx-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002536 "src/f32-gemm/gen-inc/6x8inc-minmax-avx-broadcast.c",
2537 "src/f32-gemm/gen-inc/7x8inc-minmax-avx-broadcast.c",
2538 "src/f32-gemm/gen/1x8-minmax-avx-broadcast.c",
2539 "src/f32-gemm/gen/1x16-minmax-avx-broadcast.c",
2540 "src/f32-gemm/gen/3x16-minmax-avx-broadcast.c",
2541 "src/f32-gemm/gen/4x8-minmax-avx-broadcast.c",
2542 "src/f32-gemm/gen/4x16-minmax-avx-broadcast.c",
2543 "src/f32-gemm/gen/5x8-minmax-avx-broadcast.c",
2544 "src/f32-gemm/gen/5x16-minmax-avx-broadcast.c",
2545 "src/f32-gemm/gen/6x8-minmax-avx-broadcast.c",
2546 "src/f32-gemm/gen/7x8-minmax-avx-broadcast.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07002547 "src/f32-hswish/gen/hswish-avx-x8.c",
2548 "src/f32-hswish/gen/hswish-avx-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002549 "src/f32-igemm/gen/1x8-minmax-avx-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002550 "src/f32-igemm/gen/1x16-minmax-avx-broadcast.c",
2551 "src/f32-igemm/gen/3x16-minmax-avx-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002552 "src/f32-igemm/gen/4x8-minmax-avx-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002553 "src/f32-igemm/gen/4x16-minmax-avx-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002554 "src/f32-igemm/gen/5x8-minmax-avx-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002555 "src/f32-igemm/gen/5x16-minmax-avx-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002556 "src/f32-igemm/gen/6x8-minmax-avx-broadcast.c",
2557 "src/f32-igemm/gen/7x8-minmax-avx-broadcast.c",
Marat Dukhan90eca0a2020-03-11 00:52:23 -07002558 "src/f32-prelu/gen/avx-2x8.c",
2559 "src/f32-prelu/gen/avx-2x16.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07002560 "src/f32-relu/gen/avx-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002561 "src/f32-relu/gen/avx-x16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002562 "src/f32-rmax/avx.c",
T.J. Alumbaughdc2b29c2020-10-14 13:56:08 -07002563 "src/f32-sigmoid/gen/avx-rr2-p5-div-x8.c",
2564 "src/f32-sigmoid/gen/avx-rr2-p5-div-x16.c",
2565 "src/f32-sigmoid/gen/avx-rr2-p5-div-x24.c",
2566 "src/f32-sigmoid/gen/avx-rr2-p5-div-x32.c",
2567 "src/f32-sigmoid/gen/avx-rr2-p5-div-x40.c",
2568 "src/f32-sigmoid/gen/avx-rr2-p5-div-x48.c",
2569 "src/f32-sigmoid/gen/avx-rr2-p5-div-x56.c",
2570 "src/f32-sigmoid/gen/avx-rr2-p5-div-x64.c",
2571 "src/f32-sigmoid/gen/avx-rr2-p5-div-x72.c",
2572 "src/f32-sigmoid/gen/avx-rr2-p5-div-x80.c",
2573 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x8.c",
2574 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x16.c",
2575 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x24.c",
2576 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x32.c",
2577 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x40.c",
2578 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x48.c",
2579 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x56.c",
2580 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x64.c",
2581 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x72.c",
2582 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x80.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002583 "src/f32-vbinary/gen/vadd-minmax-avx-x8.c",
2584 "src/f32-vbinary/gen/vadd-minmax-avx-x16.c",
2585 "src/f32-vbinary/gen/vaddc-minmax-avx-x8.c",
2586 "src/f32-vbinary/gen/vaddc-minmax-avx-x16.c",
2587 "src/f32-vbinary/gen/vdiv-minmax-avx-x8.c",
2588 "src/f32-vbinary/gen/vdiv-minmax-avx-x16.c",
2589 "src/f32-vbinary/gen/vdivc-minmax-avx-x8.c",
2590 "src/f32-vbinary/gen/vdivc-minmax-avx-x16.c",
Marat Dukhan9a88efe2019-12-10 15:54:24 -08002591 "src/f32-vbinary/gen/vmax-avx-x8.c",
2592 "src/f32-vbinary/gen/vmax-avx-x16.c",
2593 "src/f32-vbinary/gen/vmaxc-avx-x8.c",
2594 "src/f32-vbinary/gen/vmaxc-avx-x16.c",
2595 "src/f32-vbinary/gen/vmin-avx-x8.c",
2596 "src/f32-vbinary/gen/vmin-avx-x16.c",
2597 "src/f32-vbinary/gen/vminc-avx-x8.c",
2598 "src/f32-vbinary/gen/vminc-avx-x16.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002599 "src/f32-vbinary/gen/vmul-minmax-avx-x8.c",
2600 "src/f32-vbinary/gen/vmul-minmax-avx-x16.c",
2601 "src/f32-vbinary/gen/vmulc-minmax-avx-x8.c",
2602 "src/f32-vbinary/gen/vmulc-minmax-avx-x16.c",
2603 "src/f32-vbinary/gen/vrdivc-minmax-avx-x8.c",
2604 "src/f32-vbinary/gen/vrdivc-minmax-avx-x16.c",
2605 "src/f32-vbinary/gen/vrsubc-minmax-avx-x8.c",
2606 "src/f32-vbinary/gen/vrsubc-minmax-avx-x16.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07002607 "src/f32-vbinary/gen/vsqrdiff-avx-x8.c",
2608 "src/f32-vbinary/gen/vsqrdiff-avx-x16.c",
2609 "src/f32-vbinary/gen/vsqrdiffc-avx-x8.c",
2610 "src/f32-vbinary/gen/vsqrdiffc-avx-x16.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002611 "src/f32-vbinary/gen/vsub-minmax-avx-x8.c",
2612 "src/f32-vbinary/gen/vsub-minmax-avx-x16.c",
2613 "src/f32-vbinary/gen/vsubc-minmax-avx-x8.c",
2614 "src/f32-vbinary/gen/vsubc-minmax-avx-x16.c",
Marat Dukhaned6baaf2020-12-01 15:07:08 -08002615 "src/f32-velu/gen/velu-avx-rr2-lut4-p4-perm-x8.c",
2616 "src/f32-velu/gen/velu-avx-rr2-lut4-p4-perm-x16.c",
2617 "src/f32-velu/gen/velu-avx-rr2-lut4-p4-perm-x24.c",
2618 "src/f32-velu/gen/velu-avx-rr2-lut4-p4-perm-x32.c",
2619 "src/f32-velu/gen/velu-avx-rr2-lut4-p4-perm-x40.c",
2620 "src/f32-velu/gen/velu-avx-rr2-lut4-p4-perm-x48.c",
2621 "src/f32-velu/gen/velu-avx-rr2-lut16-p3-x8.c",
2622 "src/f32-velu/gen/velu-avx-rr2-lut16-p3-x16.c",
2623 "src/f32-velu/gen/velu-avx-rr2-lut16-p3-x24.c",
2624 "src/f32-velu/gen/velu-avx-rr2-lut16-p3-x32.c",
2625 "src/f32-velu/gen/velu-avx-rr2-lut16-p3-x40.c",
2626 "src/f32-velu/gen/velu-avx-rr2-lut16-p3-x48.c",
2627 "src/f32-velu/gen/velu-avx-rr2-p6-x8.c",
2628 "src/f32-velu/gen/velu-avx-rr2-p6-x16.c",
2629 "src/f32-velu/gen/velu-avx-rr2-p6-x24.c",
2630 "src/f32-velu/gen/velu-avx-rr2-p6-x32.c",
2631 "src/f32-velu/gen/velu-avx-rr2-p6-x40.c",
2632 "src/f32-velu/gen/velu-avx-rr2-p6-x48.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07002633 "src/f32-vlrelu/gen/vlrelu-avx-x8.c",
2634 "src/f32-vlrelu/gen/vlrelu-avx-x16.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07002635 "src/f32-vrnd/gen/vrndd-avx-x8.c",
2636 "src/f32-vrnd/gen/vrndd-avx-x16.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002637 "src/f32-vrnd/gen/vrndne-avx-x8.c",
2638 "src/f32-vrnd/gen/vrndne-avx-x16.c",
2639 "src/f32-vrnd/gen/vrndu-avx-x8.c",
2640 "src/f32-vrnd/gen/vrndu-avx-x16.c",
2641 "src/f32-vrnd/gen/vrndz-avx-x8.c",
2642 "src/f32-vrnd/gen/vrndz-avx-x16.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07002643 "src/f32-vscale/avx-x32.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07002644 "src/f32-vsqrt/gen/avx-sqrt-x8.c",
2645 "src/f32-vsqrt/gen/avx-sqrt-x16.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07002646 "src/f32-vunary/gen/vabs-avx-x8.c",
2647 "src/f32-vunary/gen/vabs-avx-x16.c",
2648 "src/f32-vunary/gen/vneg-avx-x8.c",
2649 "src/f32-vunary/gen/vneg-avx-x16.c",
2650 "src/f32-vunary/gen/vsqr-avx-x8.c",
2651 "src/f32-vunary/gen/vsqr-avx-x16.c",
Marat Dukhanb7633f22020-11-20 16:34:56 -08002652 "src/math/exp-avx-rr2-p5.c",
Marat Dukhande390d42020-11-29 19:32:18 -08002653 "src/math/expm1minus-avx-rr2-lut4-p4-perm.c",
2654 "src/math/expm1minus-avx-rr2-lut16-p3.c",
2655 "src/math/expm1minus-avx-rr2-p6.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002656 "src/math/sigmoid-avx-rr2-lut64-p2-div.c",
Marat Dukhan2e4727b2020-09-24 00:48:39 -07002657 "src/math/sigmoid-avx-rr2-p5-div.c",
Marat Dukhan39ab16f2020-10-16 14:30:35 -07002658 "src/math/sigmoid-avx-rr2-p5-nr1.c",
2659 "src/math/sigmoid-avx-rr2-p5-nr2.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002660]
2661
Marat Dukhan1566fee2020-08-02 21:55:41 -07002662XOP_UKERNELS = [
2663 "src/qs8-gemm/gen/1x4c2-minmax-xop-ld64.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002664 "src/qs8-gemm/gen/1x4c2-minmax-xop-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002665 "src/qs8-gemm/gen/1x4c2-xw-minmax-xop.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002666 "src/qs8-gemm/gen/1x4c8-minmax-xop-ld64.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002667 "src/qs8-gemm/gen/1x4c8-minmax-xop-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002668 "src/qs8-gemm/gen/1x4c8-xw-minmax-xop.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002669 "src/qs8-gemm/gen/2x4c8-minmax-xop-ld64.c",
2670 "src/qs8-gemm/gen/2x4c8-minmax-xop-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002671 "src/qs8-gemm/gen/2x4c8-xw-minmax-xop.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002672 "src/qs8-gemm/gen/3x4c8-minmax-xop-ld64.c",
2673 "src/qs8-gemm/gen/3x4c8-minmax-xop-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002674 "src/qs8-gemm/gen/3x4c8-xw-minmax-xop.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002675 "src/qs8-gemm/gen/4x4c2-minmax-xop-ld64.c",
2676 "src/qs8-gemm/gen/4x4c2-minmax-xop-ld128.c",
2677 "src/qs8-gemm/gen/4x4c2-xw-minmax-xop.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002678 "src/qs8-igemm/gen/1x4c2-minmax-xop-ld64.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002679 "src/qs8-igemm/gen/1x4c2-minmax-xop-ld128.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002680 "src/qs8-igemm/gen/1x4c8-minmax-xop-ld64.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002681 "src/qs8-igemm/gen/1x4c8-minmax-xop-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002682 "src/qs8-igemm/gen/2x4c8-minmax-xop-ld64.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002683 "src/qs8-igemm/gen/2x4c8-minmax-xop-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002684 "src/qs8-igemm/gen/3x4c8-minmax-xop-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002685 "src/qs8-igemm/gen/3x4c8-minmax-xop-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002686 "src/qs8-igemm/gen/4x4c2-minmax-xop-ld64.c",
2687 "src/qs8-igemm/gen/4x4c2-minmax-xop-ld128.c",
Marat Dukhanbb9225e2020-09-06 22:40:56 -07002688 "src/qs8-vadd/gen/minmax-xop-mul32-ld32-x8.c",
2689 "src/qs8-vadd/gen/minmax-xop-mul32-ld32-x16.c",
2690 "src/qs8-vadd/gen/minmax-xop-mul32-ld32-x24.c",
2691 "src/qs8-vadd/gen/minmax-xop-mul32-ld32-x32.c",
2692 "src/qs8-vaddc/gen/minmax-xop-mul32-ld32-x8.c",
2693 "src/qs8-vaddc/gen/minmax-xop-mul32-ld32-x16.c",
2694 "src/qs8-vaddc/gen/minmax-xop-mul32-ld32-x24.c",
2695 "src/qs8-vaddc/gen/minmax-xop-mul32-ld32-x32.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002696]
2697
Marat Dukhanfda12b82019-11-21 12:27:59 -08002698FMA3_UKERNELS = [
Marat Dukhan1c587112020-04-08 20:04:28 -07002699 "src/f32-dwconv/gen/up8x4-minmax-fma3-acc2.c",
2700 "src/f32-dwconv/gen/up8x4-minmax-fma3.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002701 "src/f32-dwconv/gen/up8x9-minmax-fma3-acc2.c",
2702 "src/f32-dwconv/gen/up8x9-minmax-fma3.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002703 "src/f32-dwconv/gen/up8x25-minmax-fma3-acc2.c",
2704 "src/f32-dwconv/gen/up8x25-minmax-fma3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002705 "src/f32-dwconv/gen/up16x4-minmax-fma3-acc2.c",
2706 "src/f32-dwconv/gen/up16x4-minmax-fma3.c",
2707 "src/f32-dwconv/gen/up16x9-minmax-fma3-acc2.c",
2708 "src/f32-dwconv/gen/up16x9-minmax-fma3.c",
2709 "src/f32-dwconv/gen/up16x25-minmax-fma3-acc2.c",
2710 "src/f32-dwconv/gen/up16x25-minmax-fma3.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002711 "src/f32-gemm/gen-inc/1x8inc-minmax-fma3-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002712 "src/f32-gemm/gen-inc/1x16inc-minmax-fma3-broadcast.c",
2713 "src/f32-gemm/gen-inc/1x16s4inc-minmax-fma3-broadcast.c",
2714 "src/f32-gemm/gen-inc/3x16inc-minmax-fma3-broadcast.c",
2715 "src/f32-gemm/gen-inc/3x16s4inc-minmax-fma3-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002716 "src/f32-gemm/gen-inc/4x8inc-minmax-fma3-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002717 "src/f32-gemm/gen-inc/4x16inc-minmax-fma3-broadcast.c",
2718 "src/f32-gemm/gen-inc/4x16s4inc-minmax-fma3-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002719 "src/f32-gemm/gen-inc/5x8inc-minmax-fma3-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002720 "src/f32-gemm/gen-inc/5x16inc-minmax-fma3-broadcast.c",
2721 "src/f32-gemm/gen-inc/5x16s4inc-minmax-fma3-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002722 "src/f32-gemm/gen-inc/6x8inc-minmax-fma3-broadcast.c",
2723 "src/f32-gemm/gen-inc/7x8inc-minmax-fma3-broadcast.c",
2724 "src/f32-gemm/gen-inc/8x8inc-minmax-fma3-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002725 "src/f32-gemm/gen/1x8-minmax-fma3-broadcast.c",
2726 "src/f32-gemm/gen/1x16-minmax-fma3-broadcast.c",
2727 "src/f32-gemm/gen/1x16s4-minmax-fma3-broadcast.c",
2728 "src/f32-gemm/gen/3x16-minmax-fma3-broadcast.c",
2729 "src/f32-gemm/gen/3x16s4-minmax-fma3-broadcast.c",
2730 "src/f32-gemm/gen/4x8-minmax-fma3-broadcast.c",
2731 "src/f32-gemm/gen/4x16-minmax-fma3-broadcast.c",
2732 "src/f32-gemm/gen/4x16s4-minmax-fma3-broadcast.c",
2733 "src/f32-gemm/gen/5x8-minmax-fma3-broadcast.c",
2734 "src/f32-gemm/gen/5x16-minmax-fma3-broadcast.c",
2735 "src/f32-gemm/gen/5x16s4-minmax-fma3-broadcast.c",
2736 "src/f32-gemm/gen/6x8-minmax-fma3-broadcast.c",
2737 "src/f32-gemm/gen/7x8-minmax-fma3-broadcast.c",
2738 "src/f32-gemm/gen/8x8-minmax-fma3-broadcast.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07002739 "src/f32-hswish/gen/hswish-fma3-x8.c",
2740 "src/f32-hswish/gen/hswish-fma3-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002741 "src/f32-igemm/gen/1x8-minmax-fma3-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002742 "src/f32-igemm/gen/1x16-minmax-fma3-broadcast.c",
2743 "src/f32-igemm/gen/1x16s4-minmax-fma3-broadcast.c",
2744 "src/f32-igemm/gen/3x16-minmax-fma3-broadcast.c",
2745 "src/f32-igemm/gen/3x16s4-minmax-fma3-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002746 "src/f32-igemm/gen/4x8-minmax-fma3-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002747 "src/f32-igemm/gen/4x16-minmax-fma3-broadcast.c",
2748 "src/f32-igemm/gen/4x16s4-minmax-fma3-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002749 "src/f32-igemm/gen/5x8-minmax-fma3-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002750 "src/f32-igemm/gen/5x16-minmax-fma3-broadcast.c",
2751 "src/f32-igemm/gen/5x16s4-minmax-fma3-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002752 "src/f32-igemm/gen/6x8-minmax-fma3-broadcast.c",
2753 "src/f32-igemm/gen/7x8-minmax-fma3-broadcast.c",
2754 "src/f32-igemm/gen/8x8-minmax-fma3-broadcast.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07002755 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x8.c",
2756 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x16.c",
2757 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x24.c",
2758 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x32.c",
2759 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x40.c",
2760 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x48.c",
2761 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x56.c",
2762 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x64.c",
Marat Dukhan84000762020-06-29 18:38:43 -07002763 "src/math/sqrt-fma3-nr1fma.c",
Marat Dukhan84000762020-06-29 18:38:43 -07002764 "src/math/sqrt-fma3-nr1fma1adj.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002765 "src/math/sqrt-fma3-nr2fma.c",
Marat Dukhanfda12b82019-11-21 12:27:59 -08002766]
2767
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002768AVX2_UKERNELS = [
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002769 "src/f32-raddexpminusmax/gen/avx2-p5-x64-acc2.c",
2770 "src/f32-raddexpminusmax/gen/avx2-p5-x64-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002771 "src/f32-raddexpminusmax/gen/avx2-p5-x64.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002772 "src/f32-raddexpminusmax/gen/avx2-p5-x72-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002773 "src/f32-raddexpminusmax/gen/avx2-p5-x72.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002774 "src/f32-raddexpminusmax/gen/avx2-p5-x80-acc2.c",
2775 "src/f32-raddexpminusmax/gen/avx2-p5-x80-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002776 "src/f32-raddexpminusmax/gen/avx2-p5-x80.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002777 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc2.c",
2778 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc3.c",
2779 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc6.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002780 "src/f32-raddexpminusmax/gen/avx2-p5-x96.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002781 "src/f32-raddextexp/gen/avx2-p5-x64-acc2.c",
2782 "src/f32-raddextexp/gen/avx2-p5-x64-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002783 "src/f32-raddextexp/gen/avx2-p5-x64.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002784 "src/f32-raddextexp/gen/avx2-p5-x72-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002785 "src/f32-raddextexp/gen/avx2-p5-x72.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002786 "src/f32-raddextexp/gen/avx2-p5-x80-acc2.c",
2787 "src/f32-raddextexp/gen/avx2-p5-x80-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002788 "src/f32-raddextexp/gen/avx2-p5-x80.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002789 "src/f32-raddextexp/gen/avx2-p5-x96-acc2.c",
2790 "src/f32-raddextexp/gen/avx2-p5-x96-acc3.c",
2791 "src/f32-raddextexp/gen/avx2-p5-x96-acc6.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002792 "src/f32-raddextexp/gen/avx2-p5-x96.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002793 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64-acc2.c",
2794 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002795 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002796 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x72-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002797 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x72.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002798 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80-acc2.c",
2799 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002800 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002801 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc2.c",
2802 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc3.c",
2803 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc6.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002804 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96.c",
Marat Dukhanfa0a4322020-01-06 16:14:29 -08002805 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x8.c",
2806 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x16.c",
2807 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x24.c",
2808 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x32.c",
2809 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x40.c",
2810 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x48.c",
2811 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x56.c",
2812 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x64.c",
2813 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x72.c",
2814 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x80.c",
2815 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x8.c",
2816 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x16.c",
2817 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x24.c",
2818 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x32.c",
2819 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x40.c",
2820 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x48.c",
2821 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x56.c",
2822 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x64.c",
2823 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x72.c",
2824 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x80.c",
2825 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x8.c",
2826 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x16.c",
2827 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x24.c",
2828 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x32.c",
2829 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x40.c",
2830 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x48.c",
2831 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x56.c",
2832 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x64.c",
2833 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x72.c",
2834 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x80.c",
Marat Dukhaned6baaf2020-12-01 15:07:08 -08002835 "src/f32-velu/gen/velu-avx2-rr1-lut4-p4-perm-x8.c",
2836 "src/f32-velu/gen/velu-avx2-rr1-lut4-p4-perm-x16.c",
2837 "src/f32-velu/gen/velu-avx2-rr1-lut4-p4-perm-x24.c",
2838 "src/f32-velu/gen/velu-avx2-rr1-lut4-p4-perm-x32.c",
2839 "src/f32-velu/gen/velu-avx2-rr1-lut4-p4-perm-x40.c",
2840 "src/f32-velu/gen/velu-avx2-rr1-lut4-p4-perm-x48.c",
2841 "src/f32-velu/gen/velu-avx2-rr1-lut4-p4-perm-x56.c",
2842 "src/f32-velu/gen/velu-avx2-rr1-lut4-p4-perm-x64.c",
2843 "src/f32-velu/gen/velu-avx2-rr1-lut4-p4-perm-x72.c",
2844 "src/f32-velu/gen/velu-avx2-rr1-lut4-p4-perm-x80.c",
2845 "src/f32-velu/gen/velu-avx2-rr1-lut8-p4-perm-x8.c",
2846 "src/f32-velu/gen/velu-avx2-rr1-lut8-p4-perm-x16.c",
2847 "src/f32-velu/gen/velu-avx2-rr1-lut8-p4-perm-x24.c",
2848 "src/f32-velu/gen/velu-avx2-rr1-lut8-p4-perm-x32.c",
2849 "src/f32-velu/gen/velu-avx2-rr1-lut8-p4-perm-x40.c",
2850 "src/f32-velu/gen/velu-avx2-rr1-lut8-p4-perm-x48.c",
2851 "src/f32-velu/gen/velu-avx2-rr1-lut8-p4-perm-x56.c",
2852 "src/f32-velu/gen/velu-avx2-rr1-lut8-p4-perm-x64.c",
2853 "src/f32-velu/gen/velu-avx2-rr1-lut8-p4-perm-x72.c",
2854 "src/f32-velu/gen/velu-avx2-rr1-lut8-p4-perm-x80.c",
2855 "src/f32-velu/gen/velu-avx2-rr1-lut16-p3-gather-x8.c",
2856 "src/f32-velu/gen/velu-avx2-rr1-lut16-p3-gather-x16.c",
2857 "src/f32-velu/gen/velu-avx2-rr1-lut16-p3-gather-x24.c",
2858 "src/f32-velu/gen/velu-avx2-rr1-lut16-p3-gather-x32.c",
2859 "src/f32-velu/gen/velu-avx2-rr1-lut16-p3-gather-x40.c",
2860 "src/f32-velu/gen/velu-avx2-rr1-lut16-p3-gather-x48.c",
2861 "src/f32-velu/gen/velu-avx2-rr1-lut16-p3-gather-x56.c",
2862 "src/f32-velu/gen/velu-avx2-rr1-lut16-p3-gather-x64.c",
2863 "src/f32-velu/gen/velu-avx2-rr1-lut16-p3-gather-x72.c",
2864 "src/f32-velu/gen/velu-avx2-rr1-lut16-p3-gather-x80.c",
2865 "src/f32-velu/gen/velu-avx2-rr1-p6-x8.c",
2866 "src/f32-velu/gen/velu-avx2-rr1-p6-x16.c",
2867 "src/f32-velu/gen/velu-avx2-rr1-p6-x24.c",
2868 "src/f32-velu/gen/velu-avx2-rr1-p6-x32.c",
2869 "src/f32-velu/gen/velu-avx2-rr1-p6-x40.c",
2870 "src/f32-velu/gen/velu-avx2-rr1-p6-x48.c",
2871 "src/f32-velu/gen/velu-avx2-rr1-p6-x56.c",
2872 "src/f32-velu/gen/velu-avx2-rr1-p6-x64.c",
2873 "src/f32-velu/gen/velu-avx2-rr1-p6-x72.c",
2874 "src/f32-velu/gen/velu-avx2-rr1-p6-x80.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002875 "src/f32-vscaleexpminusmax/gen/avx2-p5-x8.c",
2876 "src/f32-vscaleexpminusmax/gen/avx2-p5-x16.c",
2877 "src/f32-vscaleexpminusmax/gen/avx2-p5-x24.c",
2878 "src/f32-vscaleexpminusmax/gen/avx2-p5-x32.c",
2879 "src/f32-vscaleexpminusmax/gen/avx2-p5-x40.c",
2880 "src/f32-vscaleexpminusmax/gen/avx2-p5-x48.c",
2881 "src/f32-vscaleexpminusmax/gen/avx2-p5-x56.c",
2882 "src/f32-vscaleexpminusmax/gen/avx2-p5-x64.c",
2883 "src/f32-vscaleexpminusmax/gen/avx2-p5-x72.c",
2884 "src/f32-vscaleexpminusmax/gen/avx2-p5-x80.c",
2885 "src/f32-vscaleexpminusmax/gen/avx2-p5-x88.c",
2886 "src/f32-vscaleexpminusmax/gen/avx2-p5-x96.c",
2887 "src/f32-vscaleextexp/gen/avx2-p5-x8.c",
2888 "src/f32-vscaleextexp/gen/avx2-p5-x16.c",
2889 "src/f32-vscaleextexp/gen/avx2-p5-x24.c",
2890 "src/f32-vscaleextexp/gen/avx2-p5-x32.c",
2891 "src/f32-vscaleextexp/gen/avx2-p5-x40.c",
2892 "src/f32-vscaleextexp/gen/avx2-p5-x48.c",
2893 "src/f32-vscaleextexp/gen/avx2-p5-x56.c",
2894 "src/f32-vscaleextexp/gen/avx2-p5-x64.c",
2895 "src/f32-vscaleextexp/gen/avx2-p5-x72.c",
2896 "src/f32-vscaleextexp/gen/avx2-p5-x80.c",
2897 "src/f32-vscaleextexp/gen/avx2-p5-x88.c",
2898 "src/f32-vscaleextexp/gen/avx2-p5-x96.c",
Marat Dukhanb7633f22020-11-20 16:34:56 -08002899 "src/math/exp-avx2-rr2-lut8-p3-perm.c",
2900 "src/math/exp-avx2-rr2-lut8-p4-perm.c",
2901 "src/math/exp-avx2-rr2-p5.c",
Marat Dukhan9dd119a2020-11-20 18:20:04 -08002902 "src/math/expminus-avx2-rr2-p5.c",
Marat Dukhande390d42020-11-29 19:32:18 -08002903 "src/math/expm1minus-avx2-rr1-lut4-p4-perm.c",
2904 "src/math/expm1minus-avx2-rr1-lut8-p4-perm.c",
2905 "src/math/expm1minus-avx2-rr1-lut16-p3-gather.c",
2906 "src/math/expm1minus-avx2-rr1-p6.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002907 "src/math/extexp-avx2-p5.c",
2908 "src/math/sigmoid-avx2-rr1-lut64-p2-gather-div.c",
2909 "src/math/sigmoid-avx2-rr1-lut64-p2-gather-nr1fma.c",
2910 "src/math/sigmoid-avx2-rr1-lut64-p2-gather-nr2fma.c",
2911 "src/math/sigmoid-avx2-rr1-lut64-p2-gather-nr2fma1adj.c",
2912 "src/math/sigmoid-avx2-rr1-p5-div.c",
2913 "src/math/sigmoid-avx2-rr1-p5-nr1fma.c",
2914 "src/math/sigmoid-avx2-rr1-p5-nr2fma.c",
2915 "src/math/sigmoid-avx2-rr2-lut64-p2-gather-div.c",
2916 "src/math/sigmoid-avx2-rr2-lut64-p2-gather-nr1fma.c",
2917 "src/math/sigmoid-avx2-rr2-lut64-p2-gather-nr2fma.c",
2918 "src/math/sigmoid-avx2-rr2-lut64-p2-gather-nr2fma1adj.c",
2919 "src/math/sigmoid-avx2-rr2-p5-div.c",
2920 "src/math/sigmoid-avx2-rr2-p5-nr1fma.c",
2921 "src/math/sigmoid-avx2-rr2-p5-nr2fma.c",
Marat Dukhand65a1522020-08-04 19:28:18 -07002922 "src/qs8-dwconv/gen/up8x9-minmax-avx2-mul32.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002923 "src/qs8-dwconv/gen/up16x9-minmax-avx2-mul16.c",
Marat Dukhand65a1522020-08-04 19:28:18 -07002924 "src/qs8-dwconv/gen/up16x9-minmax-avx2-mul32.c",
2925 "src/qs8-dwconv/gen/up24x9-minmax-avx2-mul32.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002926 "src/qs8-dwconv/gen/up32x9-minmax-avx2-mul16.c",
Marat Dukhand65a1522020-08-04 19:28:18 -07002927 "src/qs8-dwconv/gen/up32x9-minmax-avx2-mul32.c",
Marat Dukhan12809522020-08-02 22:23:51 -07002928 "src/qs8-gemm/gen/1x8c8-minmax-avx2.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002929 "src/qs8-gemm/gen/1x8c8-xw-minmax-avx2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002930 "src/qs8-gemm/gen/2x8c8-minmax-avx2.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002931 "src/qs8-gemm/gen/2x8c8-xw-minmax-avx2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002932 "src/qs8-gemm/gen/3x8c8-minmax-avx2.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002933 "src/qs8-gemm/gen/3x8c8-xw-minmax-avx2.c",
Marat Dukhan12809522020-08-02 22:23:51 -07002934 "src/qs8-igemm/gen/1x8c8-minmax-avx2.c",
2935 "src/qs8-igemm/gen/2x8c8-minmax-avx2.c",
2936 "src/qs8-igemm/gen/3x8c8-minmax-avx2.c",
Marat Dukhane6dc0b62020-09-08 23:57:14 -07002937 "src/qs8-vadd/gen/minmax-avx2-mul32-ld64-x8.c",
2938 "src/qs8-vadd/gen/minmax-avx2-mul32-ld64-x16.c",
2939 "src/qs8-vadd/gen/minmax-avx2-mul32-ld64-x24.c",
2940 "src/qs8-vadd/gen/minmax-avx2-mul32-ld64-x32.c",
2941 "src/qs8-vaddc/gen/minmax-avx2-mul32-ld64-x8.c",
2942 "src/qs8-vaddc/gen/minmax-avx2-mul32-ld64-x16.c",
2943 "src/qs8-vaddc/gen/minmax-avx2-mul32-ld64-x24.c",
2944 "src/qs8-vaddc/gen/minmax-avx2-mul32-ld64-x32.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002945]
2946
Marat Dukhan08c4a432019-10-03 09:29:21 -07002947AVX512F_UKERNELS = [
Marat Dukhan5c5fa962020-03-10 18:38:33 -07002948 "src/f32-clamp/gen/avx512f-x16.c",
2949 "src/f32-clamp/gen/avx512f-x32.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002950 "src/f32-dwconv/gen/up16x4-minmax-avx512f-acc2.c",
2951 "src/f32-dwconv/gen/up16x4-minmax-avx512f.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002952 "src/f32-dwconv/gen/up16x9-minmax-avx512f-acc2.c",
2953 "src/f32-dwconv/gen/up16x9-minmax-avx512f.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002954 "src/f32-dwconv/gen/up16x25-minmax-avx512f-acc2.c",
2955 "src/f32-dwconv/gen/up16x25-minmax-avx512f.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002956 "src/f32-dwconv/gen/up32x4-minmax-avx512f-acc2.c",
2957 "src/f32-dwconv/gen/up32x4-minmax-avx512f.c",
2958 "src/f32-dwconv/gen/up32x9-minmax-avx512f-acc2.c",
2959 "src/f32-dwconv/gen/up32x9-minmax-avx512f.c",
2960 "src/f32-dwconv/gen/up32x25-minmax-avx512f-acc2.c",
2961 "src/f32-dwconv/gen/up32x25-minmax-avx512f.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002962 "src/f32-gemm/gen-inc/1x16inc-minmax-avx512f-broadcast.c",
2963 "src/f32-gemm/gen-inc/4x16inc-minmax-avx512f-broadcast.c",
2964 "src/f32-gemm/gen-inc/5x16inc-minmax-avx512f-broadcast.c",
2965 "src/f32-gemm/gen-inc/6x16inc-minmax-avx512f-broadcast.c",
2966 "src/f32-gemm/gen-inc/7x16inc-minmax-avx512f-broadcast.c",
2967 "src/f32-gemm/gen-inc/8x16inc-minmax-avx512f-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002968 "src/f32-gemm/gen/1x16-minmax-avx512f-broadcast.c",
2969 "src/f32-gemm/gen/4x16-minmax-avx512f-broadcast.c",
2970 "src/f32-gemm/gen/5x16-minmax-avx512f-broadcast.c",
2971 "src/f32-gemm/gen/6x16-minmax-avx512f-broadcast.c",
2972 "src/f32-gemm/gen/7x16-minmax-avx512f-broadcast.c",
2973 "src/f32-gemm/gen/8x16-minmax-avx512f-broadcast.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07002974 "src/f32-hswish/gen/hswish-avx512f-x16.c",
2975 "src/f32-hswish/gen/hswish-avx512f-x32.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002976 "src/f32-igemm/gen/1x16-minmax-avx512f-broadcast.c",
2977 "src/f32-igemm/gen/4x16-minmax-avx512f-broadcast.c",
2978 "src/f32-igemm/gen/5x16-minmax-avx512f-broadcast.c",
2979 "src/f32-igemm/gen/6x16-minmax-avx512f-broadcast.c",
2980 "src/f32-igemm/gen/7x16-minmax-avx512f-broadcast.c",
2981 "src/f32-igemm/gen/8x16-minmax-avx512f-broadcast.c",
Marat Dukhan90eca0a2020-03-11 00:52:23 -07002982 "src/f32-prelu/gen/avx512f-2x16.c",
2983 "src/f32-prelu/gen/avx512f-2x32.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002984 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128-acc2.c",
2985 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002986 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002987 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x144-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002988 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x144.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002989 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160-acc2.c",
2990 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002991 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002992 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc2.c",
2993 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc3.c",
2994 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc6.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002995 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002996 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128-acc2.c",
2997 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002998 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002999 "src/f32-raddextexp/gen/avx512f-p5-scalef-x144-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003000 "src/f32-raddextexp/gen/avx512f-p5-scalef-x144.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003001 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160-acc2.c",
3002 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003003 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003004 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc2.c",
3005 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc3.c",
3006 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc6.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003007 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003008 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128-acc2.c",
3009 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003010 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003011 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x144-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003012 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x144.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003013 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160-acc2.c",
3014 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003015 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003016 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc2.c",
3017 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc3.c",
3018 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc6.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003019 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07003020 "src/f32-relu/gen/avx512f-x16.c",
3021 "src/f32-relu/gen/avx512f-x32.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003022 "src/f32-rmax/avx512f.c",
Marat Dukhand9ca7e62020-09-23 23:45:29 -07003023 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x16.c",
3024 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x32.c",
3025 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x48.c",
3026 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x64.c",
3027 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x80.c",
3028 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x96.c",
3029 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x112.c",
3030 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x128.c",
3031 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x16.c",
3032 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x32.c",
3033 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x48.c",
3034 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x64.c",
3035 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x80.c",
3036 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x96.c",
3037 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x112.c",
3038 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003039 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x16.c",
3040 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x32.c",
3041 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x48.c",
3042 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x64.c",
3043 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x80.c",
3044 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x96.c",
3045 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x112.c",
3046 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x128.c",
3047 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x16.c",
3048 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x32.c",
3049 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x48.c",
3050 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x64.c",
3051 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x80.c",
3052 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x96.c",
3053 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x112.c",
3054 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x128.c",
Marat Dukhand9ca7e62020-09-23 23:45:29 -07003055 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x16.c",
3056 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x32.c",
3057 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x48.c",
3058 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x64.c",
3059 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x80.c",
3060 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x96.c",
3061 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x112.c",
3062 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x128.c",
3063 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x16.c",
3064 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x32.c",
3065 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x48.c",
3066 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x64.c",
3067 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x80.c",
3068 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x96.c",
3069 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x112.c",
3070 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x128.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07003071 "src/f32-vbinary/gen/vadd-minmax-avx512f-x16.c",
3072 "src/f32-vbinary/gen/vadd-minmax-avx512f-x32.c",
3073 "src/f32-vbinary/gen/vaddc-minmax-avx512f-x16.c",
3074 "src/f32-vbinary/gen/vaddc-minmax-avx512f-x32.c",
3075 "src/f32-vbinary/gen/vdiv-minmax-avx512f-x16.c",
3076 "src/f32-vbinary/gen/vdiv-minmax-avx512f-x32.c",
3077 "src/f32-vbinary/gen/vdivc-minmax-avx512f-x16.c",
3078 "src/f32-vbinary/gen/vdivc-minmax-avx512f-x32.c",
Marat Dukhan9a88efe2019-12-10 15:54:24 -08003079 "src/f32-vbinary/gen/vmax-avx512f-x16.c",
3080 "src/f32-vbinary/gen/vmax-avx512f-x32.c",
3081 "src/f32-vbinary/gen/vmaxc-avx512f-x16.c",
3082 "src/f32-vbinary/gen/vmaxc-avx512f-x32.c",
3083 "src/f32-vbinary/gen/vmin-avx512f-x16.c",
3084 "src/f32-vbinary/gen/vmin-avx512f-x32.c",
3085 "src/f32-vbinary/gen/vminc-avx512f-x16.c",
3086 "src/f32-vbinary/gen/vminc-avx512f-x32.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07003087 "src/f32-vbinary/gen/vmul-minmax-avx512f-x16.c",
3088 "src/f32-vbinary/gen/vmul-minmax-avx512f-x32.c",
3089 "src/f32-vbinary/gen/vmulc-minmax-avx512f-x16.c",
3090 "src/f32-vbinary/gen/vmulc-minmax-avx512f-x32.c",
3091 "src/f32-vbinary/gen/vrdivc-minmax-avx512f-x16.c",
3092 "src/f32-vbinary/gen/vrdivc-minmax-avx512f-x32.c",
3093 "src/f32-vbinary/gen/vrsubc-minmax-avx512f-x16.c",
3094 "src/f32-vbinary/gen/vrsubc-minmax-avx512f-x32.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07003095 "src/f32-vbinary/gen/vsqrdiff-avx512f-x16.c",
3096 "src/f32-vbinary/gen/vsqrdiff-avx512f-x32.c",
3097 "src/f32-vbinary/gen/vsqrdiffc-avx512f-x16.c",
3098 "src/f32-vbinary/gen/vsqrdiffc-avx512f-x32.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07003099 "src/f32-vbinary/gen/vsub-minmax-avx512f-x16.c",
3100 "src/f32-vbinary/gen/vsub-minmax-avx512f-x32.c",
3101 "src/f32-vbinary/gen/vsubc-minmax-avx512f-x16.c",
3102 "src/f32-vbinary/gen/vsubc-minmax-avx512f-x32.c",
Marat Dukhaned6baaf2020-12-01 15:07:08 -08003103 "src/f32-velu/gen/velu-avx512f-rr1-lut16-p3-perm-x16.c",
3104 "src/f32-velu/gen/velu-avx512f-rr1-lut16-p3-perm-x32.c",
3105 "src/f32-velu/gen/velu-avx512f-rr1-lut16-p3-perm-x48.c",
3106 "src/f32-velu/gen/velu-avx512f-rr1-lut16-p3-perm-x64.c",
3107 "src/f32-velu/gen/velu-avx512f-rr1-lut16-p3-perm-x80.c",
3108 "src/f32-velu/gen/velu-avx512f-rr1-lut16-p3-perm-x96.c",
3109 "src/f32-velu/gen/velu-avx512f-rr1-lut16-p3-perm-x112.c",
3110 "src/f32-velu/gen/velu-avx512f-rr1-lut16-p3-perm-x128.c",
3111 "src/f32-velu/gen/velu-avx512f-rr1-p6-x16.c",
3112 "src/f32-velu/gen/velu-avx512f-rr1-p6-x32.c",
3113 "src/f32-velu/gen/velu-avx512f-rr1-p6-x48.c",
3114 "src/f32-velu/gen/velu-avx512f-rr1-p6-x64.c",
3115 "src/f32-velu/gen/velu-avx512f-rr1-p6-x80.c",
3116 "src/f32-velu/gen/velu-avx512f-rr1-p6-x96.c",
3117 "src/f32-velu/gen/velu-avx512f-rr1-p6-x112.c",
3118 "src/f32-velu/gen/velu-avx512f-rr1-p6-x128.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07003119 "src/f32-vlrelu/gen/vlrelu-avx512f-x16.c",
3120 "src/f32-vlrelu/gen/vlrelu-avx512f-x32.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003121 "src/f32-vrnd/gen/vrndd-avx512f-x16.c",
3122 "src/f32-vrnd/gen/vrndd-avx512f-x32.c",
3123 "src/f32-vrnd/gen/vrndne-avx512f-x16.c",
3124 "src/f32-vrnd/gen/vrndne-avx512f-x32.c",
3125 "src/f32-vrnd/gen/vrndu-avx512f-x16.c",
3126 "src/f32-vrnd/gen/vrndu-avx512f-x32.c",
3127 "src/f32-vrnd/gen/vrndz-avx512f-x16.c",
3128 "src/f32-vrnd/gen/vrndz-avx512f-x32.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07003129 "src/f32-vscale/avx512f-x64.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003130 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x16.c",
3131 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x32.c",
3132 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x48.c",
3133 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x64.c",
3134 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x80.c",
3135 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x96.c",
3136 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x112.c",
3137 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x128.c",
3138 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x144.c",
3139 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x160.c",
3140 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x176.c",
3141 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x192.c",
3142 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x16.c",
3143 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x32.c",
3144 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x48.c",
3145 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x64.c",
3146 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x80.c",
3147 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x96.c",
3148 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x112.c",
3149 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x128.c",
3150 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x144.c",
3151 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x160.c",
3152 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x176.c",
3153 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x192.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07003154 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x16.c",
3155 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x32.c",
3156 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x48.c",
3157 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x64.c",
3158 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x80.c",
3159 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x96.c",
3160 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x112.c",
3161 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x128.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07003162 "src/f32-vunary/gen/vabs-avx512f-x16.c",
3163 "src/f32-vunary/gen/vabs-avx512f-x32.c",
3164 "src/f32-vunary/gen/vneg-avx512f-x16.c",
3165 "src/f32-vunary/gen/vneg-avx512f-x32.c",
3166 "src/f32-vunary/gen/vsqr-avx512f-x16.c",
3167 "src/f32-vunary/gen/vsqr-avx512f-x32.c",
Marat Dukhanb7633f22020-11-20 16:34:56 -08003168 "src/math/exp-avx512f-rr2-lut16-p3-perm-scalef.c",
3169 "src/math/exp-avx512f-rr2-lut16-p3-perm.c",
3170 "src/math/exp-avx512f-rr2-lut32-p2-perm2-scalef.c",
3171 "src/math/exp-avx512f-rr2-lut32-p2-perm2.c",
3172 "src/math/exp-avx512f-rr2-p5-scalef.c",
3173 "src/math/exp-avx512f-rr2-p5.c",
Marat Dukhande390d42020-11-29 19:32:18 -08003174 "src/math/expm1minus-avx512f-rr1-lut16-p3-perm.c",
3175 "src/math/expm1minus-avx512f-rr1-p6.c",
Marat Dukhan98ba4412019-10-23 02:14:28 -07003176 "src/math/extexp-avx512f-p5.c",
Marat Dukhan6a34c5f2020-09-22 21:44:15 -07003177 "src/math/sigmoid-avx512f-rr1-lut16-p3-perm-scalef-div.c",
Marat Dukhan6a34c5f2020-09-22 21:44:15 -07003178 "src/math/sigmoid-avx512f-rr1-lut16-p3-perm-scalef-nr1fma.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003179 "src/math/sigmoid-avx512f-rr1-lut16-p3-perm-scalef-nr1fma1adj.c",
Marat Dukhan6a34c5f2020-09-22 21:44:15 -07003180 "src/math/sigmoid-avx512f-rr1-lut32-p2-perm2-scalef-div.c",
Marat Dukhan6a34c5f2020-09-22 21:44:15 -07003181 "src/math/sigmoid-avx512f-rr1-lut32-p2-perm2-scalef-nr1fma.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003182 "src/math/sigmoid-avx512f-rr1-lut32-p2-perm2-scalef-nr1fma1adj.c",
Marat Dukhan36173d22020-10-15 17:14:26 -07003183 "src/math/sigmoid-avx512f-rr1-lut64-p2-gather-scalef-div.c",
Marat Dukhan36173d22020-10-15 17:14:26 -07003184 "src/math/sigmoid-avx512f-rr1-lut64-p2-gather-scalef-nr1fma.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003185 "src/math/sigmoid-avx512f-rr1-lut64-p2-gather-scalef-nr1fma1adj.c",
3186 "src/math/sigmoid-avx512f-rr1-p5-scalef-div.c",
3187 "src/math/sigmoid-avx512f-rr1-p5-scalef-nr1fma.c",
3188 "src/math/sigmoid-avx512f-rr1-p5-scalef-nr1fma1adj.c",
3189 "src/math/sigmoid-avx512f-rr2-lut16-p3-perm-scalef-div.c",
3190 "src/math/sigmoid-avx512f-rr2-lut16-p3-perm-scalef-nr1fma.c",
3191 "src/math/sigmoid-avx512f-rr2-lut16-p3-perm-scalef-nr1fma1adj.c",
3192 "src/math/sigmoid-avx512f-rr2-lut32-p2-perm2-scalef-div.c",
3193 "src/math/sigmoid-avx512f-rr2-lut32-p2-perm2-scalef-nr1fma.c",
3194 "src/math/sigmoid-avx512f-rr2-lut32-p2-perm2-scalef-nr1fma1adj.c",
Marat Dukhan36173d22020-10-15 17:14:26 -07003195 "src/math/sigmoid-avx512f-rr2-lut64-p2-gather-scalef-div.c",
Marat Dukhan36173d22020-10-15 17:14:26 -07003196 "src/math/sigmoid-avx512f-rr2-lut64-p2-gather-scalef-nr1fma.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003197 "src/math/sigmoid-avx512f-rr2-lut64-p2-gather-scalef-nr1fma1adj.c",
3198 "src/math/sigmoid-avx512f-rr2-p5-scalef-div.c",
3199 "src/math/sigmoid-avx512f-rr2-p5-scalef-nr1fma.c",
3200 "src/math/sigmoid-avx512f-rr2-p5-scalef-nr1fma1adj.c",
Marat Dukhan84000762020-06-29 18:38:43 -07003201 "src/math/sqrt-avx512f-nr1fma.c",
Marat Dukhan84000762020-06-29 18:38:43 -07003202 "src/math/sqrt-avx512f-nr1fma1adj.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003203 "src/math/sqrt-avx512f-nr2fma.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003204]
3205
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07003206AVX512SKX_UKERNELS = [
Marat Dukhan2ffc5e62020-09-06 22:33:38 -07003207 "src/qs8-dwconv/gen/up16x9-minmax-avx512skx-mul32.c",
3208 "src/qs8-dwconv/gen/up32x9-minmax-avx512skx-mul32.c",
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07003209 "src/qs8-gemm/gen/1x16c8-minmax-avx512skx.c",
3210 "src/qs8-gemm/gen/2x16c8-minmax-avx512skx.c",
3211 "src/qs8-gemm/gen/3x16c8-minmax-avx512skx.c",
3212 "src/qs8-gemm/gen/4x16c8-minmax-avx512skx.c",
3213 "src/qs8-igemm/gen/1x16c8-minmax-avx512skx.c",
3214 "src/qs8-igemm/gen/2x16c8-minmax-avx512skx.c",
3215 "src/qs8-igemm/gen/3x16c8-minmax-avx512skx.c",
3216 "src/qs8-igemm/gen/4x16c8-minmax-avx512skx.c",
3217]
3218
Frank Barchardbcedc082020-08-17 18:00:51 -07003219WASM32_ASM_UKERNELS = [
3220 "src/f32-relu/wasm_shr_x1.S",
3221 "src/f32-relu/wasm_shr_x2.S",
3222 "src/f32-relu/wasm_shr_x4.S",
3223]
3224
Marat Dukhan08c4a432019-10-03 09:29:21 -07003225AARCH32_ASM_UKERNELS = [
Marat Dukhan32f93812020-05-17 20:31:21 -07003226 "src/f32-gemm/4x4-aarch32-vfp-ld64.S",
Marat Dukhan3b98f6b2020-05-17 10:09:22 -07003227 "src/f32-gemm/4x4-minmax-aarch32-vfp-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003228 "src/f32-gemm/4x8-minmax-aarch32-neon-cortex-a53.S",
3229 "src/f32-gemm/4x8-minmax-aarch32-neon-cortex-a55.S",
Frank Barchard490febe2020-07-16 18:42:17 -07003230 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-cortex-a7.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003231 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-cortex-a75.S",
Frank Barchard569561d2020-06-17 13:11:12 -07003232 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-ld64.S",
Frank Barchard490febe2020-07-16 18:42:17 -07003233 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-pld-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003234 "src/f32-igemm/4x8-minmax-aarch32-neon-cortex-a53.S",
3235 "src/f32-igemm/4x8-minmax-aarch32-neon-cortex-a55.S",
Frank Barchard490febe2020-07-16 18:42:17 -07003236 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-cortex-a7.S",
3237 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-cortex-a75.S",
3238 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-ld64.S",
3239 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-pld-cortex-a75.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003240]
3241
3242AARCH64_ASM_UKERNELS = [
Frank Barchardbddfbcd2020-04-15 12:32:41 -07003243 "src/f16-gemm/gen-inc/1x8inc-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchard04336c12020-10-22 16:48:55 -07003244 "src/f16-gemm/gen-inc/1x16inc-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchardbddfbcd2020-04-15 12:32:41 -07003245 "src/f16-gemm/gen-inc/4x8inc-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchard04336c12020-10-22 16:48:55 -07003246 "src/f16-gemm/gen-inc/4x16inc-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchardbddfbcd2020-04-15 12:32:41 -07003247 "src/f16-gemm/gen-inc/6x8inc-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchard04336c12020-10-22 16:48:55 -07003248 "src/f16-gemm/gen-inc/6x16inc-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchard3b8e5662020-04-20 12:12:53 -07003249 "src/f16-gemm/gen-inc/8x8inc-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchard04336c12020-10-22 16:48:55 -07003250 "src/f16-gemm/gen/1x8-minmax-aarch64-neonfp16arith-ld64.S",
3251 "src/f16-gemm/gen/1x16-minmax-aarch64-neonfp16arith-ld32.S",
3252 "src/f16-gemm/gen/4x8-minmax-aarch64-neonfp16arith-ld64.S",
3253 "src/f16-gemm/gen/4x16-minmax-aarch64-neonfp16arith-ld32.S",
3254 "src/f16-gemm/gen/6x8-minmax-aarch64-neonfp16arith-ld64.S",
3255 "src/f16-gemm/gen/6x16-minmax-aarch64-neonfp16arith-ld32.S",
3256 "src/f16-gemm/gen/8x8-minmax-aarch64-neonfp16arith-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003257 "src/f32-dwconv/up4x9-minmax-aarch64-neonfma-cortex-a55.S",
3258 "src/f32-dwconv/up4x9-minmax-aarch64-neonfma.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003259 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a53.S",
3260 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a57.S",
3261 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a75.S",
Frank Barchard04336c12020-10-22 16:48:55 -07003262 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-ld64.S",
3263 "src/f32-gemm/gen-inc/1x12inc-minmax-aarch64-neonfma-cortex-a53.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003264 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a53.S",
3265 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a55.S",
3266 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a57.S",
3267 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003268 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-ld64.S",
Frank Barchard04336c12020-10-22 16:48:55 -07003269 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-ld128.S",
3270 "src/f32-gemm/gen-inc/4x12inc-minmax-aarch64-neonfma-cortex-a53.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003271 "src/f32-gemm/gen-inc/5x8inc-minmax-aarch64-neonfma-cortex-a57.S",
3272 "src/f32-gemm/gen-inc/5x8inc-minmax-aarch64-neonfma-cortex-a75.S",
3273 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a53.S",
3274 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a55.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003275 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a57.S",
Frank Barchard04336c12020-10-22 16:48:55 -07003276 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a73.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003277 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003278 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-ld64.S",
Frank Barchard04336c12020-10-22 16:48:55 -07003279 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-ld128.S",
3280 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a53.S",
3281 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a57.S",
3282 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a75.S",
3283 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-ld64.S",
3284 "src/f32-gemm/gen/1x12-minmax-aarch64-neonfma-cortex-a53.S",
3285 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a53.S",
3286 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a55.S",
3287 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a57.S",
3288 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a75.S",
3289 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-ld64.S",
3290 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-ld128.S",
3291 "src/f32-gemm/gen/4x12-minmax-aarch64-neonfma-cortex-a53.S",
3292 "src/f32-gemm/gen/5x8-minmax-aarch64-neonfma-cortex-a57.S",
3293 "src/f32-gemm/gen/5x8-minmax-aarch64-neonfma-cortex-a75.S",
3294 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a53.S",
3295 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a55.S",
3296 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a57.S",
3297 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a73.S",
3298 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a75.S",
3299 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-ld64.S",
3300 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-ld128.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003301 "src/f32-igemm/1x8-minmax-aarch64-neonfma-cortex-a53.S",
Frank Barchard04336c12020-10-22 16:48:55 -07003302 "src/f32-igemm/1x12-minmax-aarch64-neonfma-cortex-a53.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003303 "src/f32-igemm/4x8-minmax-aarch64-neonfma-cortex-a53.S",
3304 "src/f32-igemm/4x8-minmax-aarch64-neonfma-cortex-a55.S",
Frank Barchard04336c12020-10-22 16:48:55 -07003305 "src/f32-igemm/4x12-minmax-aarch64-neonfma-cortex-a53.S",
3306 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a53.S",
3307 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a55.S",
3308 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a73.S",
3309 "src/f32-igemm/gen/1x8-minmax-aarch64-neonfma-cortex-a57.S",
3310 "src/f32-igemm/gen/1x8-minmax-aarch64-neonfma-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003311 "src/f32-igemm/gen/4x8-minmax-aarch64-neonfma-cortex-a57.S",
3312 "src/f32-igemm/gen/4x8-minmax-aarch64-neonfma-cortex-a75.S",
3313 "src/f32-igemm/gen/5x8-minmax-aarch64-neonfma-cortex-a57.S",
3314 "src/f32-igemm/gen/5x8-minmax-aarch64-neonfma-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003315 "src/f32-igemm/gen/6x8-minmax-aarch64-neonfma-cortex-a57.S",
3316 "src/f32-igemm/gen/6x8-minmax-aarch64-neonfma-cortex-a75.S",
Frank Barchardf1fd89e2020-10-06 23:27:11 -07003317 "src/qs8-gemm/1x16c4-aarch64-neondot-ld64.S",
Frank Barchardf2742c42020-10-12 21:16:46 -07003318 "src/qs8-gemm/4x16c4-aarch64-neondot-cortex-a55.S",
Frank Barcharda5237a52020-10-06 16:18:46 -07003319 "src/qs8-gemm/4x16c4-aarch64-neondot-ld64.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003320]
3321
Marat Dukhan1b354632020-03-23 12:50:22 -07003322INTERNAL_MICROKERNEL_HDRS = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003323 "src/xnnpack/argmaxpool.h",
3324 "src/xnnpack/avgpool.h",
3325 "src/xnnpack/clamp.h",
3326 "src/xnnpack/common.h",
3327 "src/xnnpack/conv.h",
Yury Kartynnike7841862020-11-04 18:22:18 -08003328 "src/xnnpack/depthtospace.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003329 "src/xnnpack/dwconv.h",
Frank Barchard04336c12020-10-22 16:48:55 -07003330 "src/xnnpack/fill.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003331 "src/xnnpack/gavgpool.h",
3332 "src/xnnpack/gemm.h",
3333 "src/xnnpack/hswish.h",
Marat Dukhan660fd192020-03-10 04:55:30 -07003334 "src/xnnpack/ibilinear.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003335 "src/xnnpack/igemm.h",
Marat Dukhancfb31342019-12-05 10:42:57 -08003336 "src/xnnpack/intrinsics-polyfill.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003337 "src/xnnpack/lut.h",
3338 "src/xnnpack/math.h",
3339 "src/xnnpack/maxpool.h",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003340 "src/xnnpack/memory.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003341 "src/xnnpack/packx.h",
3342 "src/xnnpack/pad.h",
3343 "src/xnnpack/params.h",
3344 "src/xnnpack/pavgpool.h",
3345 "src/xnnpack/ppmm.h",
3346 "src/xnnpack/prelu.h",
Marat Dukhan97579532019-10-18 16:40:39 -07003347 "src/xnnpack/raddexpminusmax.h",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07003348 "src/xnnpack/raddextexp.h",
Marat Dukhan97579532019-10-18 16:40:39 -07003349 "src/xnnpack/raddstoreexpminusmax.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003350 "src/xnnpack/rmax.h",
3351 "src/xnnpack/scalar-utils.h",
3352 "src/xnnpack/spmm.h",
3353 "src/xnnpack/unpool.h",
3354 "src/xnnpack/vadd.h",
Marat Dukhan1e782c42019-11-21 17:02:40 -08003355 "src/xnnpack/vbinary.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003356 "src/xnnpack/vmulcaddc.h",
Marat Dukhan05ac8e32019-10-21 15:39:33 -07003357 "src/xnnpack/vscale.h",
Marat Dukhan97579532019-10-18 16:40:39 -07003358 "src/xnnpack/vscaleexpminusmax.h",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07003359 "src/xnnpack/vscaleextexp.h",
Marat Dukhan1e782c42019-11-21 17:02:40 -08003360 "src/xnnpack/vunary.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003361 "src/xnnpack/zip.h",
Marat Dukhan1b354632020-03-23 12:50:22 -07003362]
3363
3364INTERNAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003365 "include/xnnpack.h",
3366 "src/xnnpack/allocator.h",
3367 "src/xnnpack/compute.h",
3368 "src/xnnpack/im2col.h",
3369 "src/xnnpack/indirection.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003370 "src/xnnpack/math-stubs.h",
Chao Mei6ddfc602020-05-13 22:29:36 -07003371 "src/xnnpack/memory-planner.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003372 "src/xnnpack/operator.h",
3373 "src/xnnpack/pack.h",
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07003374 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003375 "src/xnnpack/requantization-stubs.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003376 "src/xnnpack/requantization.h",
Marat Dukhan1d75a542020-02-03 12:23:01 -08003377 "src/xnnpack/subgraph.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003378]
3379
Marat Dukhan1b354632020-03-23 12:50:22 -07003380ACCURACY_EVAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003381 "src/xnnpack/math-stubs.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003382]
3383
Marat Dukhan1b354632020-03-23 12:50:22 -07003384MICROKERNEL_BENCHMARK_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003385 "include/xnnpack.h",
Frank Barchard35db7d02020-10-26 13:37:34 -07003386 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003387]
3388
Marat Dukhan1b354632020-03-23 12:50:22 -07003389MICROKERNEL_TEST_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Frank Barchard35db7d02020-10-26 13:37:34 -07003390 "include/xnnpack.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003391 "src/xnnpack/isa-checks.h",
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07003392 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003393 "src/xnnpack/requantization.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003394]
3395
3396OPERATOR_TEST_PARAMS_HDRS = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003397 "src/xnnpack/common.h",
Frank Barchard04336c12020-10-22 16:48:55 -07003398 "src/xnnpack/params.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003399]
3400
3401WEIGHTS_PACK_HDRS = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003402 "src/xnnpack/compute.h",
Frank Barchard04336c12020-10-22 16:48:55 -07003403 "src/xnnpack/operator.h",
3404 "src/xnnpack/pack.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003405]
3406
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07003407LOGGING_COPTS = select({
3408 # No logging in optimized mode
3409 ":optimized_build": ["-DXNN_LOG_LEVEL=0"],
3410 # Full logging in debug mode
3411 ":debug_build": ["-DXNN_LOG_LEVEL=5"],
3412 # Error-only logging in default (fastbuild) mode
3413 "//conditions:default": ["-DXNN_LOG_LEVEL=2"],
3414})
3415
Marat Dukhan3b59de22020-06-03 20:15:19 -07003416LOGGING_SRCS = select({
3417 # No logging in optimized mode
3418 ":optimized_build": [],
3419 "//conditions:default": [
3420 "src/operator-strings.c",
3421 "src/subgraph-strings.c",
3422 ],
3423})
3424
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07003425LOGGING_HDRS = [
3426 "src/xnnpack/log.h",
3427]
3428
Marat Dukhan08c4a432019-10-03 09:29:21 -07003429xnnpack_cc_library(
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003430 name = "tables",
3431 srcs = TABLE_SRCS,
3432 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003433 gcc_copts = xnnpack_gcc_std_copts(),
3434 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003435)
3436
3437xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003438 name = "scalar_ukernels",
3439 srcs = SCALAR_UKERNELS,
3440 hdrs = INTERNAL_HDRS,
3441 aarch32_copts = ["-marm"],
Marat Dukhan10a38082020-04-17 03:58:35 -07003442 gcc_copts = xnnpack_gcc_std_copts(),
3443 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003444 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003445 ":tables",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003446 "@FP16",
3447 "@FXdiv",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003448 "@pthreadpool",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003449 ],
3450)
3451
3452xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003453 name = "scalar_ukernels_test_mode",
3454 srcs = SCALAR_UKERNELS,
3455 hdrs = INTERNAL_HDRS,
3456 aarch32_copts = ["-marm"],
3457 copts = [
3458 "-UNDEBUG",
3459 "-DXNN_TEST_MODE=1",
3460 ],
3461 gcc_copts = xnnpack_gcc_std_copts(),
3462 msvc_copts = xnnpack_msvc_std_copts(),
3463 deps = [
3464 ":tables",
3465 "@FP16",
3466 "@FXdiv",
3467 "@pthreadpool",
3468 ],
3469)
3470
3471xnnpack_cc_library(
Marat Dukhan436ebe62019-12-04 15:10:12 -08003472 name = "wasm_ukernels",
3473 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003474 gcc_copts = xnnpack_gcc_std_copts(),
3475 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan436ebe62019-12-04 15:10:12 -08003476 wasm_srcs = WASM_UKERNELS,
Marat Dukhan1c6e3892020-06-25 23:56:10 -07003477 wasmsimd_srcs = WASM_UKERNELS + WASMSIMD_UKERNELS,
Marat Dukhan436ebe62019-12-04 15:10:12 -08003478 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003479 ":tables",
Marat Dukhan436ebe62019-12-04 15:10:12 -08003480 "@FP16",
3481 "@FXdiv",
3482 "@pthreadpool",
3483 ],
3484)
3485
3486xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003487 name = "wasm_ukernels_test_mode",
3488 hdrs = INTERNAL_HDRS,
3489 copts = [
3490 "-UNDEBUG",
3491 "-DXNN_TEST_MODE=1",
3492 ],
3493 gcc_copts = xnnpack_gcc_std_copts(),
3494 msvc_copts = xnnpack_msvc_std_copts(),
3495 wasm_srcs = WASM_UKERNELS,
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07003496 wasmsimd_srcs = WASM_UKERNELS + WASMSIMD_UKERNELS,
Marat Dukhan33fcf782020-05-24 14:27:15 -07003497 deps = [
3498 ":tables",
3499 "@FP16",
3500 "@FXdiv",
3501 "@pthreadpool",
3502 ],
3503)
3504
3505xnnpack_cc_library(
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08003506 name = "psimd_accmath_ukernels",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08003507 hdrs = INTERNAL_HDRS,
3508 aarch32_copts = [
3509 "-marm",
3510 "-mfpu=neon",
3511 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07003512 gcc_copts = xnnpack_gcc_std_copts(),
3513 gcc_x86_copts = ["-msse2"],
3514 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08003515 optimized_copts = [
3516 "-O3",
3517 ],
Marat Dukhan500b8892020-04-15 17:09:50 -07003518 psimd_srcs = PSIMD_ACCMATH_UKERNELS,
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08003519 deps = [
3520 ":tables",
3521 "@FP16",
3522 "@psimd",
3523 "@pthreadpool",
3524 ],
3525)
3526
3527xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003528 name = "psimd_accmath_ukernels_test_mode",
3529 hdrs = INTERNAL_HDRS,
3530 aarch32_copts = [
3531 "-marm",
3532 "-mfpu=neon",
3533 ],
3534 copts = [
3535 "-UNDEBUG",
3536 "-DXNN_TEST_MODE=1",
3537 ],
3538 gcc_copts = xnnpack_gcc_std_copts(),
3539 gcc_x86_copts = ["-msse2"],
3540 msvc_copts = xnnpack_msvc_std_copts(),
3541 optimized_copts = [
3542 "-O3",
3543 ],
3544 psimd_srcs = PSIMD_ACCMATH_UKERNELS,
3545 deps = [
3546 ":tables",
3547 "@FP16",
3548 "@psimd",
3549 "@pthreadpool",
3550 ],
3551)
3552
3553xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003554 name = "neon_ukernels",
3555 hdrs = INTERNAL_HDRS,
3556 aarch32_copts = [
3557 "-marm",
Marat Dukhan8853b822020-05-07 12:19:01 -07003558 "-march=armv7-a",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003559 "-mfpu=neon",
3560 ],
3561 aarch32_srcs = NEON_UKERNELS,
3562 aarch64_srcs = NEON_UKERNELS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003563 gcc_copts = xnnpack_gcc_std_copts(),
3564 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08003565 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003566 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003567 "@FP16",
3568 "@pthreadpool",
3569 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003570)
3571
3572xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003573 name = "neon_ukernels_test_mode",
3574 hdrs = INTERNAL_HDRS,
3575 aarch32_copts = [
3576 "-marm",
3577 "-march=armv7-a",
3578 "-mfpu=neon",
3579 ],
3580 aarch32_srcs = NEON_UKERNELS,
3581 aarch64_srcs = NEON_UKERNELS,
3582 copts = [
3583 "-UNDEBUG",
3584 "-DXNN_TEST_MODE=1",
3585 ],
3586 gcc_copts = xnnpack_gcc_std_copts(),
3587 msvc_copts = xnnpack_msvc_std_copts(),
3588 deps = [
3589 ":tables",
3590 "@FP16",
3591 "@pthreadpool",
3592 ],
3593)
3594
3595xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003596 name = "neonfma_ukernels",
3597 hdrs = INTERNAL_HDRS,
3598 aarch32_copts = [
3599 "-marm",
Marat Dukhan8853b822020-05-07 12:19:01 -07003600 "-march=armv7-a",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003601 "-mfpu=neon-vfpv4",
3602 ],
3603 aarch32_srcs = NEONFMA_UKERNELS,
3604 aarch64_srcs = NEONFMA_UKERNELS + AARCH64_NEONFMA_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07003605 apple_aarch32_copts = [
3606 "-mcpu=swift",
3607 "-mtune=generic",
3608 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07003609 gcc_copts = xnnpack_gcc_std_copts(),
3610 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08003611 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003612 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003613 "@FP16",
3614 "@pthreadpool",
3615 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003616)
3617
3618xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003619 name = "neonfma_ukernels_test_mode",
3620 hdrs = INTERNAL_HDRS,
3621 aarch32_copts = [
3622 "-marm",
3623 "-march=armv7-a",
3624 "-mfpu=neon-vfpv4",
3625 ],
3626 aarch32_srcs = NEONFMA_UKERNELS,
3627 aarch64_srcs = NEONFMA_UKERNELS + AARCH64_NEONFMA_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07003628 apple_aarch32_copts = [
3629 "-mcpu=swift",
3630 "-mtune=generic",
3631 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07003632 copts = [
3633 "-UNDEBUG",
3634 "-DXNN_TEST_MODE=1",
3635 ],
3636 gcc_copts = xnnpack_gcc_std_copts(),
3637 msvc_copts = xnnpack_msvc_std_copts(),
3638 deps = [
3639 ":tables",
3640 "@FP16",
3641 "@pthreadpool",
3642 ],
3643)
3644
3645xnnpack_cc_library(
Marat Dukhan8853b822020-05-07 12:19:01 -07003646 name = "neonv8_ukernels",
3647 hdrs = INTERNAL_HDRS,
3648 aarch32_copts = [
3649 "-marm",
3650 "-march=armv8-a",
3651 "-mfpu=neon-fp-armv8",
3652 ],
3653 aarch32_srcs = NEONV8_UKERNELS,
3654 aarch64_srcs = NEONV8_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07003655 apple_aarch32_copts = [
3656 "-mcpu=cyclone",
3657 "-mtune=generic",
3658 ],
Marat Dukhan8853b822020-05-07 12:19:01 -07003659 gcc_copts = xnnpack_gcc_std_copts(),
3660 msvc_copts = xnnpack_msvc_std_copts(),
3661 deps = [
3662 ":tables",
3663 "@FP16",
3664 "@pthreadpool",
3665 ],
3666)
3667
3668xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003669 name = "neonv8_ukernels_test_mode",
3670 hdrs = INTERNAL_HDRS,
3671 aarch32_copts = [
3672 "-marm",
3673 "-march=armv8-a",
3674 "-mfpu=neon-fp-armv8",
3675 ],
3676 aarch32_srcs = NEONV8_UKERNELS,
3677 aarch64_srcs = NEONV8_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07003678 apple_aarch32_copts = [
3679 "-mcpu=cyclone",
3680 "-mtune=generic",
3681 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07003682 copts = [
3683 "-UNDEBUG",
3684 "-DXNN_TEST_MODE=1",
3685 ],
3686 gcc_copts = xnnpack_gcc_std_copts(),
3687 msvc_copts = xnnpack_msvc_std_copts(),
3688 deps = [
3689 ":tables",
3690 "@FP16",
3691 "@pthreadpool",
3692 ],
3693)
3694
3695xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003696 name = "neonfp16arith_ukernels",
3697 hdrs = INTERNAL_HDRS,
3698 aarch64_copts = ["-march=armv8.2-a+fp16"],
3699 aarch64_srcs = AARCH64_NEONFP16ARITH_UKERNELS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003700 gcc_copts = xnnpack_gcc_std_copts(),
3701 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08003702 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003703 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003704 "@FP16",
3705 "@pthreadpool",
3706 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003707)
3708
3709xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003710 name = "neonfp16arith_ukernels_test_mode",
3711 hdrs = INTERNAL_HDRS,
3712 aarch64_copts = ["-march=armv8.2-a+fp16"],
3713 aarch64_srcs = AARCH64_NEONFP16ARITH_UKERNELS,
3714 copts = [
3715 "-UNDEBUG",
3716 "-DXNN_TEST_MODE=1",
3717 ],
3718 gcc_copts = xnnpack_gcc_std_copts(),
3719 msvc_copts = xnnpack_msvc_std_copts(),
3720 deps = [
3721 ":tables",
3722 "@FP16",
3723 "@pthreadpool",
3724 ],
3725)
3726
3727xnnpack_cc_library(
Benoit Jacoba9644732020-08-13 12:48:55 -07003728 name = "neondot_ukernels",
3729 hdrs = INTERNAL_HDRS,
Marat Dukhan799ac752020-08-13 16:08:03 -07003730 aarch32_copts = [
3731 "-marm",
3732 "-march=armv8.2-a+dotprod",
3733 "-mfpu=neon-fp-armv8",
3734 ],
Benoit Jacoba9644732020-08-13 12:48:55 -07003735 aarch32_srcs = NEONDOT_UKERNELS,
3736 aarch64_copts = ["-march=armv8.2-a+dotprod"],
3737 aarch64_srcs = NEONDOT_UKERNELS,
3738 gcc_copts = xnnpack_gcc_std_copts(),
3739 msvc_copts = xnnpack_msvc_std_copts(),
3740 deps = [
3741 ":tables",
3742 "@FP16",
3743 "@pthreadpool",
3744 ],
3745)
3746
3747xnnpack_cc_library(
3748 name = "neondot_ukernels_test_mode",
3749 hdrs = INTERNAL_HDRS,
Marat Dukhan799ac752020-08-13 16:08:03 -07003750 aarch32_copts = [
3751 "-marm",
3752 "-march=armv8.2-a+dotprod",
3753 "-mfpu=neon-fp-armv8",
3754 ],
Benoit Jacoba9644732020-08-13 12:48:55 -07003755 aarch32_srcs = NEONDOT_UKERNELS,
3756 aarch64_copts = ["-march=armv8.2-a+dotprod"],
3757 aarch64_srcs = NEONDOT_UKERNELS,
3758 copts = [
3759 "-UNDEBUG",
3760 "-DXNN_TEST_MODE=1",
3761 ],
3762 gcc_copts = xnnpack_gcc_std_copts(),
3763 msvc_copts = xnnpack_msvc_std_copts(),
3764 deps = [
3765 ":tables",
3766 "@FP16",
3767 "@pthreadpool",
3768 ],
3769)
3770
3771xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003772 name = "sse2_ukernels",
3773 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003774 gcc_copts = xnnpack_gcc_std_copts(),
3775 gcc_x86_copts = ["-msse2"],
3776 msvc_copts = xnnpack_msvc_std_copts(),
3777 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003778 x86_srcs = SSE_UKERNELS + SSE2_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08003779 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003780 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003781 "@FP16",
3782 "@pthreadpool",
3783 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003784)
3785
3786xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003787 name = "sse2_ukernels_test_mode",
3788 hdrs = INTERNAL_HDRS,
3789 copts = [
3790 "-UNDEBUG",
3791 "-DXNN_TEST_MODE=1",
3792 ],
3793 gcc_copts = xnnpack_gcc_std_copts(),
3794 gcc_x86_copts = ["-msse2"],
3795 msvc_copts = xnnpack_msvc_std_copts(),
3796 msvc_x86_32_copts = ["/arch:SSE2"],
3797 x86_srcs = SSE_UKERNELS + SSE2_UKERNELS,
3798 deps = [
3799 ":tables",
3800 "@FP16",
3801 "@pthreadpool",
3802 ],
3803)
3804
3805xnnpack_cc_library(
Marat Dukhanfe7acb62020-03-09 19:30:05 -07003806 name = "ssse3_ukernels",
3807 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003808 gcc_copts = xnnpack_gcc_std_copts(),
3809 gcc_x86_copts = ["-mssse3"],
3810 msvc_copts = xnnpack_msvc_std_copts(),
3811 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhanfe7acb62020-03-09 19:30:05 -07003812 x86_srcs = SSSE3_UKERNELS,
3813 deps = [
3814 ":tables",
3815 "@FP16",
3816 "@pthreadpool",
3817 ],
3818)
3819
3820xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003821 name = "ssse3_ukernels_test_mode",
3822 hdrs = INTERNAL_HDRS,
3823 copts = [
3824 "-UNDEBUG",
3825 "-DXNN_TEST_MODE=1",
3826 ],
3827 gcc_copts = xnnpack_gcc_std_copts(),
3828 gcc_x86_copts = ["-mssse3"],
3829 msvc_copts = xnnpack_msvc_std_copts(),
3830 msvc_x86_32_copts = ["/arch:SSE2"],
3831 x86_srcs = SSSE3_UKERNELS,
3832 deps = [
3833 ":tables",
3834 "@FP16",
3835 "@pthreadpool",
3836 ],
3837)
3838
3839xnnpack_cc_library(
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08003840 name = "sse41_ukernels",
3841 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003842 gcc_copts = xnnpack_gcc_std_copts(),
3843 gcc_x86_copts = ["-msse4.1"],
3844 msvc_copts = xnnpack_msvc_std_copts(),
3845 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08003846 x86_srcs = SSE41_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08003847 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003848 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003849 "@FP16",
3850 "@pthreadpool",
3851 ],
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08003852)
3853
3854xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003855 name = "sse41_ukernels_test_mode",
3856 hdrs = INTERNAL_HDRS,
3857 copts = [
3858 "-UNDEBUG",
3859 "-DXNN_TEST_MODE=1",
3860 ],
3861 gcc_copts = xnnpack_gcc_std_copts(),
3862 gcc_x86_copts = ["-msse4.1"],
3863 msvc_copts = xnnpack_msvc_std_copts(),
3864 msvc_x86_32_copts = ["/arch:SSE2"],
3865 x86_srcs = SSE41_UKERNELS,
3866 deps = [
3867 ":tables",
3868 "@FP16",
3869 "@pthreadpool",
3870 ],
3871)
3872
3873xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003874 name = "avx_ukernels",
3875 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003876 gcc_copts = xnnpack_gcc_std_copts(),
3877 gcc_x86_copts = ["-mavx"],
3878 msvc_copts = xnnpack_msvc_std_copts(),
3879 msvc_x86_32_copts = ["/arch:AVX"],
3880 msvc_x86_64_copts = ["/arch:AVX"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003881 x86_srcs = AVX_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08003882 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003883 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003884 "@FP16",
3885 "@pthreadpool",
3886 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003887)
3888
3889xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003890 name = "avx_ukernels_test_mode",
3891 hdrs = INTERNAL_HDRS,
3892 copts = [
3893 "-UNDEBUG",
3894 "-DXNN_TEST_MODE=1",
3895 ],
3896 gcc_copts = xnnpack_gcc_std_copts(),
3897 gcc_x86_copts = ["-mavx"],
3898 msvc_copts = xnnpack_msvc_std_copts(),
3899 msvc_x86_32_copts = ["/arch:AVX"],
3900 msvc_x86_64_copts = ["/arch:AVX"],
3901 x86_srcs = AVX_UKERNELS,
3902 deps = [
3903 ":tables",
3904 "@FP16",
3905 "@pthreadpool",
3906 ],
3907)
3908
3909xnnpack_cc_library(
Marat Dukhan1566fee2020-08-02 21:55:41 -07003910 name = "xop_ukernels",
3911 hdrs = INTERNAL_HDRS,
3912 gcc_copts = xnnpack_gcc_std_copts(),
3913 gcc_x86_copts = ["-mxop"],
3914 msvc_copts = xnnpack_msvc_std_copts(),
3915 msvc_x86_32_copts = ["/arch:AVX"],
3916 msvc_x86_64_copts = ["/arch:AVX"],
3917 x86_srcs = XOP_UKERNELS,
3918 deps = [
3919 ":tables",
3920 "@FP16",
3921 "@pthreadpool",
3922 ],
3923)
3924
3925xnnpack_cc_library(
3926 name = "xop_ukernels_test_mode",
3927 hdrs = INTERNAL_HDRS,
3928 copts = [
3929 "-UNDEBUG",
3930 "-DXNN_TEST_MODE=1",
3931 ],
3932 gcc_copts = xnnpack_gcc_std_copts(),
3933 gcc_x86_copts = ["-mxop"],
3934 msvc_copts = xnnpack_msvc_std_copts(),
3935 msvc_x86_32_copts = ["/arch:AVX"],
3936 msvc_x86_64_copts = ["/arch:AVX"],
3937 x86_srcs = XOP_UKERNELS,
3938 deps = [
3939 ":tables",
3940 "@FP16",
3941 "@pthreadpool",
3942 ],
3943)
3944
3945xnnpack_cc_library(
Marat Dukhanfda12b82019-11-21 12:27:59 -08003946 name = "fma3_ukernels",
3947 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003948 gcc_copts = xnnpack_gcc_std_copts(),
3949 gcc_x86_copts = ["-mfma"],
3950 msvc_copts = xnnpack_msvc_std_copts(),
3951 msvc_x86_32_copts = ["/arch:AVX"],
3952 msvc_x86_64_copts = ["/arch:AVX"],
Marat Dukhanfda12b82019-11-21 12:27:59 -08003953 x86_srcs = FMA3_UKERNELS,
3954 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003955 ":tables",
Marat Dukhanfda12b82019-11-21 12:27:59 -08003956 "@FP16",
3957 "@pthreadpool",
3958 ],
3959)
3960
3961xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003962 name = "fma3_ukernels_test_mode",
3963 hdrs = INTERNAL_HDRS,
3964 copts = [
3965 "-UNDEBUG",
3966 "-DXNN_TEST_MODE=1",
3967 ],
3968 gcc_copts = xnnpack_gcc_std_copts(),
3969 gcc_x86_copts = ["-mfma"],
3970 msvc_copts = xnnpack_msvc_std_copts(),
3971 msvc_x86_32_copts = ["/arch:AVX"],
3972 msvc_x86_64_copts = ["/arch:AVX"],
3973 x86_srcs = FMA3_UKERNELS,
3974 deps = [
3975 ":tables",
3976 "@FP16",
3977 "@pthreadpool",
3978 ],
3979)
3980
3981xnnpack_cc_library(
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003982 name = "avx2_ukernels",
3983 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003984 gcc_copts = xnnpack_gcc_std_copts(),
3985 gcc_x86_copts = [
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003986 "-mfma",
3987 "-mavx2",
3988 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07003989 msvc_copts = xnnpack_msvc_std_copts(),
3990 msvc_x86_32_copts = ["/arch:AVX2"],
3991 msvc_x86_64_copts = ["/arch:AVX2"],
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003992 x86_srcs = AVX2_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08003993 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003994 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003995 "@FP16",
3996 "@pthreadpool",
3997 ],
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003998)
3999
4000xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07004001 name = "avx2_ukernels_test_mode",
4002 hdrs = INTERNAL_HDRS,
4003 copts = [
4004 "-UNDEBUG",
4005 "-DXNN_TEST_MODE=1",
4006 ],
4007 gcc_copts = xnnpack_gcc_std_copts(),
4008 gcc_x86_copts = [
4009 "-mfma",
4010 "-mavx2",
4011 ],
4012 msvc_copts = xnnpack_msvc_std_copts(),
4013 msvc_x86_32_copts = ["/arch:AVX2"],
4014 msvc_x86_64_copts = ["/arch:AVX2"],
4015 x86_srcs = AVX2_UKERNELS,
4016 deps = [
4017 ":tables",
4018 "@FP16",
4019 "@pthreadpool",
4020 ],
4021)
4022
4023xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004024 name = "avx512f_ukernels",
4025 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004026 gcc_copts = xnnpack_gcc_std_copts(),
4027 gcc_x86_copts = ["-mavx512f"],
4028 mingw_copts = ["-fno-asynchronous-unwind-tables"],
4029 msvc_copts = xnnpack_msvc_std_copts(),
4030 msvc_x86_32_copts = ["/arch:AVX512"],
4031 msvc_x86_64_copts = ["/arch:AVX512"],
4032 msys_copts = ["-fno-asynchronous-unwind-tables"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004033 x86_srcs = AVX512F_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08004034 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08004035 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08004036 "@FP16",
4037 "@pthreadpool",
4038 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004039)
4040
4041xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07004042 name = "avx512f_ukernels_test_mode",
4043 hdrs = INTERNAL_HDRS,
4044 copts = [
4045 "-UNDEBUG",
4046 "-DXNN_TEST_MODE=1",
4047 ],
4048 gcc_copts = xnnpack_gcc_std_copts(),
4049 gcc_x86_copts = ["-mavx512f"],
4050 mingw_copts = ["-fno-asynchronous-unwind-tables"],
4051 msvc_copts = xnnpack_msvc_std_copts(),
4052 msvc_x86_32_copts = ["/arch:AVX512"],
4053 msvc_x86_64_copts = ["/arch:AVX512"],
4054 msys_copts = ["-fno-asynchronous-unwind-tables"],
4055 x86_srcs = AVX512F_UKERNELS,
4056 deps = [
4057 ":tables",
4058 "@FP16",
4059 "@pthreadpool",
4060 ],
4061)
4062
4063xnnpack_cc_library(
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07004064 name = "avx512skx_ukernels",
4065 hdrs = INTERNAL_HDRS,
4066 gcc_copts = xnnpack_gcc_std_copts(),
4067 gcc_x86_copts = [
4068 "-mavx512f",
4069 "-mavx512cd",
4070 "-mavx512bw",
4071 "-mavx512dq",
4072 "-mavx512vl",
4073 ],
4074 mingw_copts = ["-fno-asynchronous-unwind-tables"],
4075 msvc_copts = xnnpack_msvc_std_copts(),
4076 msvc_x86_32_copts = ["/arch:AVX512"],
4077 msvc_x86_64_copts = ["/arch:AVX512"],
4078 msys_copts = ["-fno-asynchronous-unwind-tables"],
4079 x86_srcs = AVX512SKX_UKERNELS,
4080 deps = [
4081 ":tables",
4082 "@FP16",
4083 "@pthreadpool",
4084 ],
4085)
4086
4087xnnpack_cc_library(
4088 name = "avx512skx_ukernels_test_mode",
4089 hdrs = INTERNAL_HDRS,
4090 copts = [
4091 "-UNDEBUG",
4092 "-DXNN_TEST_MODE=1",
4093 ],
4094 gcc_copts = xnnpack_gcc_std_copts(),
4095 gcc_x86_copts = [
4096 "-mavx512f",
4097 "-mavx512cd",
4098 "-mavx512bw",
4099 "-mavx512dq",
4100 "-mavx512vl",
4101 ],
4102 mingw_copts = ["-fno-asynchronous-unwind-tables"],
4103 msvc_copts = xnnpack_msvc_std_copts(),
4104 msvc_x86_32_copts = ["/arch:AVX512"],
4105 msvc_x86_64_copts = ["/arch:AVX512"],
4106 msys_copts = ["-fno-asynchronous-unwind-tables"],
4107 x86_srcs = AVX512SKX_UKERNELS,
4108 deps = [
4109 ":tables",
4110 "@FP16",
4111 "@pthreadpool",
4112 ],
4113)
4114
4115xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004116 name = "asm_ukernels",
4117 hdrs = ["src/xnnpack/assembly.h"],
4118 aarch32_srcs = AARCH32_ASM_UKERNELS,
Frank Barchard31bb45b2020-10-06 00:26:33 -07004119 aarch64_copts = ["-march=armv8.2-a+fp16+dotprod"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004120 aarch64_srcs = AARCH64_ASM_UKERNELS,
Frank Barchardbcedc082020-08-17 18:00:51 -07004121 wasm_srcs = WASM32_ASM_UKERNELS,
4122 wasmsimd_srcs = WASM32_ASM_UKERNELS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07004123)
4124
Marat Dukhan3b59de22020-06-03 20:15:19 -07004125xnnpack_cc_library(
4126 name = "logging_utils",
4127 srcs = LOGGING_SRCS,
4128 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
4129 copts = LOGGING_COPTS + [
4130 "-Isrc",
4131 "-Iinclude",
4132 ] + select({
4133 ":debug_build": [],
4134 "//conditions:default": xnnpack_min_size_copts(),
4135 }),
4136 gcc_copts = xnnpack_gcc_std_copts(),
4137 msvc_copts = xnnpack_msvc_std_copts(),
4138 visibility = xnnpack_visibility(),
4139 deps = [
4140 "@FP16",
4141 "@clog",
4142 "@pthreadpool",
4143 ],
4144)
4145
Marat Dukhan08c4a432019-10-03 09:29:21 -07004146xnnpack_aggregate_library(
4147 name = "ukernels",
4148 aarch32_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07004149 ":neon_ukernels",
4150 ":neonfma_ukernels",
Marat Dukhan8853b822020-05-07 12:19:01 -07004151 ":neonv8_ukernels",
Benoit Jacoba9644732020-08-13 12:48:55 -07004152 ":neondot_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004153 ":asm_ukernels",
4154 ],
4155 aarch64_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07004156 ":neon_ukernels",
4157 ":neonfma_ukernels",
Marat Dukhan8853b822020-05-07 12:19:01 -07004158 ":neonv8_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004159 ":neonfp16arith_ukernels",
Benoit Jacoba9644732020-08-13 12:48:55 -07004160 ":neondot_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004161 ":asm_ukernels",
4162 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07004163 generic_deps = [
4164 ":scalar_ukernels",
4165 ],
Marat Dukhan500b8892020-04-15 17:09:50 -07004166 psimd_deps = [
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08004167 ":psimd_accmath_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004168 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07004169 wasm_deps = [
4170 ":wasm_ukernels",
Frank Barchardbcedc082020-08-17 18:00:51 -07004171 ":asm_ukernels",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004172 ],
4173 wasmsimd_deps = [
4174 ":wasm_ukernels",
Frank Barchardbcedc082020-08-17 18:00:51 -07004175 ":asm_ukernels",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004176 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004177 x86_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07004178 ":sse2_ukernels",
Marat Dukhanfe7acb62020-03-09 19:30:05 -07004179 ":ssse3_ukernels",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08004180 ":sse41_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004181 ":avx_ukernels",
Marat Dukhan1566fee2020-08-02 21:55:41 -07004182 ":xop_ukernels",
Marat Dukhanfda12b82019-11-21 12:27:59 -08004183 ":fma3_ukernels",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07004184 ":avx2_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004185 ":avx512f_ukernels",
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07004186 ":avx512skx_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004187 ],
4188)
4189
Marat Dukhan33fcf782020-05-24 14:27:15 -07004190xnnpack_aggregate_library(
4191 name = "ukernels_test_mode",
4192 aarch32_deps = [
4193 ":neon_ukernels_test_mode",
4194 ":neonfma_ukernels_test_mode",
4195 ":neonv8_ukernels_test_mode",
Benoit Jacoba9644732020-08-13 12:48:55 -07004196 ":neondot_ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004197 ":asm_ukernels",
4198 ],
4199 aarch64_deps = [
4200 ":neon_ukernels_test_mode",
4201 ":neonfma_ukernels_test_mode",
4202 ":neonv8_ukernels_test_mode",
4203 ":neonfp16arith_ukernels_test_mode",
Benoit Jacoba9644732020-08-13 12:48:55 -07004204 ":neondot_ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004205 ":asm_ukernels",
4206 ],
4207 generic_deps = [
4208 ":scalar_ukernels_test_mode",
4209 ],
4210 psimd_deps = [
Marat Dukhan33fcf782020-05-24 14:27:15 -07004211 ":psimd_accmath_ukernels_test_mode",
4212 ],
4213 wasm_deps = [
4214 ":wasm_ukernels_test_mode",
Frank Barchardbcedc082020-08-17 18:00:51 -07004215 ":asm_ukernels",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004216 ],
4217 wasmsimd_deps = [
4218 ":wasm_ukernels_test_mode",
Frank Barchardbcedc082020-08-17 18:00:51 -07004219 ":asm_ukernels",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004220 ],
4221 x86_deps = [
4222 ":sse2_ukernels_test_mode",
4223 ":ssse3_ukernels_test_mode",
4224 ":sse41_ukernels_test_mode",
4225 ":avx_ukernels_test_mode",
Marat Dukhan1566fee2020-08-02 21:55:41 -07004226 ":xop_ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004227 ":fma3_ukernels_test_mode",
4228 ":avx2_ukernels_test_mode",
4229 ":avx512f_ukernels_test_mode",
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07004230 ":avx512skx_ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004231 ],
4232)
4233
Marat Dukhan86df3842020-06-17 14:03:31 -07004234# This target must be compatible with Android NDK r17 and compile with Clang 6.
4235# It is similar to :ukernels target, but excludes microkernels which require
4236# NEON FP16 arithmetics, NEON Dot Product, and newer intrinsics.
Marat Dukhanae046f52020-06-15 13:16:14 -07004237xnnpack_aggregate_library(
Marat Dukhan86df3842020-06-17 14:03:31 -07004238 name = "ukernels_max_compatibility",
Marat Dukhanae046f52020-06-15 13:16:14 -07004239 aarch32_deps = [
4240 ":neon_ukernels",
4241 ":neonfma_ukernels",
4242 ":neonv8_ukernels",
4243 ":asm_ukernels",
4244 ],
4245 aarch64_deps = [
4246 ":neon_ukernels",
4247 ":neonfma_ukernels",
4248 ":neonv8_ukernels",
4249 ":asm_ukernels",
4250 ],
4251 generic_deps = [
4252 ":scalar_ukernels",
4253 ],
4254 psimd_deps = [
Marat Dukhanae046f52020-06-15 13:16:14 -07004255 ":psimd_accmath_ukernels",
4256 ],
4257 wasm_deps = [
4258 ":wasm_ukernels",
4259 ],
4260 wasmsimd_deps = [
4261 ":wasm_ukernels",
4262 ],
4263 x86_deps = [
4264 ":sse2_ukernels",
4265 ":ssse3_ukernels",
4266 ":sse41_ukernels",
4267 ":avx_ukernels",
Marat Dukhan1566fee2020-08-02 21:55:41 -07004268 ":xop_ukernels",
Marat Dukhanae046f52020-06-15 13:16:14 -07004269 ":fma3_ukernels",
4270 ":avx2_ukernels",
4271 ":avx512f_ukernels",
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07004272 ":avx512skx_ukernels",
Marat Dukhanae046f52020-06-15 13:16:14 -07004273 ],
4274)
4275
Marat Dukhan08c4a432019-10-03 09:29:21 -07004276xnnpack_cc_library(
4277 name = "im2col",
4278 srcs = ["src/im2col.c"],
4279 hdrs = [
4280 "src/xnnpack/common.h",
4281 "src/xnnpack/im2col.h",
4282 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07004283 gcc_copts = xnnpack_gcc_std_copts(),
4284 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004285)
4286
4287xnnpack_cc_library(
4288 name = "indirection",
4289 srcs = ["src/indirection.c"],
4290 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004291 gcc_copts = xnnpack_gcc_std_copts(),
4292 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004293 deps = [
4294 "@FP16",
4295 "@FXdiv",
4296 "@pthreadpool",
4297 ],
4298)
4299
4300xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07004301 name = "indirection_test_mode",
4302 srcs = ["src/indirection.c"],
4303 hdrs = INTERNAL_HDRS,
4304 copts = [
4305 "-UNDEBUG",
4306 "-DXNN_TEST_MODE=1",
4307 ],
4308 gcc_copts = xnnpack_gcc_std_copts(),
4309 msvc_copts = xnnpack_msvc_std_copts(),
4310 deps = [
4311 "@FP16",
4312 "@FXdiv",
4313 "@pthreadpool",
4314 ],
4315)
4316
4317xnnpack_cc_library(
Marat Dukhanab582382020-07-06 13:32:08 -07004318 name = "packing",
4319 srcs = ["src/packing.c"],
4320 hdrs = INTERNAL_HDRS,
4321 gcc_copts = xnnpack_gcc_std_copts(),
4322 msvc_copts = xnnpack_msvc_std_copts(),
4323 deps = [
4324 "@FP16",
4325 "@FXdiv",
4326 "@pthreadpool",
4327 ],
4328)
4329
4330xnnpack_cc_library(
4331 name = "packing_test_mode",
4332 srcs = ["src/packing.c"],
4333 hdrs = INTERNAL_HDRS,
4334 copts = [
4335 "-UNDEBUG",
4336 "-DXNN_TEST_MODE=1",
4337 ],
4338 gcc_copts = xnnpack_gcc_std_copts(),
4339 msvc_copts = xnnpack_msvc_std_copts(),
4340 deps = [
4341 "@FP16",
4342 "@FXdiv",
4343 "@pthreadpool",
4344 ],
4345)
4346
4347xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004348 name = "operator_run",
4349 srcs = ["src/operator-run.c"],
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07004350 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004351 copts = LOGGING_COPTS + select({
Marat Dukhan05702cf2020-03-26 15:41:33 -07004352 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4353 "//conditions:default": [],
4354 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07004355 gcc_copts = xnnpack_gcc_std_copts(),
4356 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004357 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07004358 ":logging_utils",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004359 "@FP16",
4360 "@FXdiv",
4361 "@clog",
4362 "@pthreadpool",
4363 ],
4364)
4365
Chao Mei6ddfc602020-05-13 22:29:36 -07004366xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07004367 name = "operator_run_test_mode",
4368 srcs = ["src/operator-run.c"],
4369 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
4370 copts = LOGGING_COPTS + [
4371 "-UNDEBUG",
4372 "-DXNN_TEST_MODE=1",
4373 ] + select({
4374 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4375 "//conditions:default": [],
4376 }),
4377 gcc_copts = xnnpack_gcc_std_copts(),
4378 msvc_copts = xnnpack_msvc_std_copts(),
4379 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07004380 ":logging_utils",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004381 "@FP16",
4382 "@FXdiv",
4383 "@clog",
4384 "@pthreadpool",
4385 ],
4386)
4387
4388xnnpack_cc_library(
Chao Mei6ddfc602020-05-13 22:29:36 -07004389 name = "memory_planner",
4390 srcs = ["src/memory-planner.c"],
4391 hdrs = INTERNAL_HDRS,
4392 defines = select({
4393 ":xnn_enable_memopt_explicit_true": ["XNN_ENABLE_MEMOPT=1"],
4394 ":xnn_enable_memopt_explicit_false": ["XNN_ENABLE_MEMOPT=0"],
4395 "//conditions:default": ["XNN_ENABLE_MEMOPT=1"],
4396 }),
4397 gcc_copts = xnnpack_gcc_std_copts(),
4398 msvc_copts = xnnpack_msvc_std_copts(),
4399 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07004400 ":logging_utils",
Chao Mei6ddfc602020-05-13 22:29:36 -07004401 "@pthreadpool",
4402 ],
4403)
4404
Marat Dukhan33fcf782020-05-24 14:27:15 -07004405xnnpack_cc_library(
4406 name = "memory_planner_test_mode",
4407 srcs = ["src/memory-planner.c"],
4408 hdrs = INTERNAL_HDRS,
4409 copts = [
4410 "-UNDEBUG",
4411 "-DXNN_TEST_MODE=1",
4412 ],
4413 defines = select({
4414 ":xnn_enable_memopt_explicit_true": ["XNN_ENABLE_MEMOPT=1"],
4415 ":xnn_enable_memopt_explicit_false": ["XNN_ENABLE_MEMOPT=0"],
4416 "//conditions:default": ["XNN_ENABLE_MEMOPT=1"],
4417 }),
4418 gcc_copts = xnnpack_gcc_std_copts(),
4419 msvc_copts = xnnpack_msvc_std_copts(),
4420 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07004421 ":logging_utils",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004422 "@pthreadpool",
4423 ],
4424)
4425
Marat Dukhan08c4a432019-10-03 09:29:21 -07004426cc_library(
4427 name = "enable_assembly",
4428 defines = select({
4429 ":xnn_enable_assembly_explicit_true": ["XNN_ENABLE_ASSEMBLY=1"],
4430 ":xnn_enable_assembly_explicit_false": ["XNN_ENABLE_ASSEMBLY=0"],
Frank Barchard810171d2019-10-10 10:34:51 -07004431 "//conditions:default": ["XNN_ENABLE_ASSEMBLY=1"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004432 }),
4433)
4434
Marat Dukhan9de90e02020-06-18 16:04:12 -07004435cc_library(
4436 name = "enable_sparse",
4437 defines = select({
4438 ":xnn_enable_sparse_explicit_true": ["XNN_ENABLE_SPARSE=1"],
4439 ":xnn_enable_sparse_explicit_false": ["XNN_ENABLE_SPARSE=0"],
Marat Dukhan8b283aa2020-06-18 16:48:57 -07004440 "//conditions:default": ["XNN_ENABLE_SPARSE=0"],
Marat Dukhan9de90e02020-06-18 16:04:12 -07004441 }),
4442)
4443
Marat Dukhancf056b22019-10-07 10:26:29 -07004444xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004445 name = "operators",
4446 srcs = OPERATOR_SRCS + [
Marat Dukhan04f03be2019-11-19 12:36:47 -08004447 "src/memory.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004448 "src/operator-delete.c",
Marat Dukhancf056b22019-10-07 10:26:29 -07004449 ],
4450 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004451 copts = LOGGING_COPTS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07004452 "-Isrc",
4453 "-Iinclude",
4454 ] + select({
4455 ":debug_build": [],
4456 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07004457 }) + select({
4458 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4459 "//conditions:default": [],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004460 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07004461 gcc_copts = xnnpack_gcc_std_copts(),
4462 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004463 deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07004464 ":indirection",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004465 ":logging_utils",
Marat Dukhanab582382020-07-06 13:32:08 -07004466 ":packing",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004467 "@FP16",
4468 "@FXdiv",
4469 "@clog",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004470 "@pthreadpool",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004471 ],
4472)
4473
Marat Dukhan10a38082020-04-17 03:58:35 -07004474xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07004475 name = "operators_test_mode",
4476 srcs = OPERATOR_SRCS + [
4477 "src/memory.c",
4478 "src/operator-delete.c",
4479 ],
4480 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
4481 copts = LOGGING_COPTS + [
4482 "-Isrc",
4483 "-Iinclude",
4484 "-UNDEBUG",
4485 "-DXNN_TEST_MODE=1",
4486 ] + select({
4487 ":debug_build": [],
4488 "//conditions:default": xnnpack_min_size_copts(),
4489 }) + select({
4490 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4491 "//conditions:default": [],
4492 }),
4493 gcc_copts = xnnpack_gcc_std_copts(),
4494 msvc_copts = xnnpack_msvc_std_copts(),
4495 deps = [
4496 ":indirection_test_mode",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004497 ":logging_utils",
Marat Dukhanab582382020-07-06 13:32:08 -07004498 ":packing_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004499 "@FP16",
4500 "@FXdiv",
4501 "@clog",
4502 "@pthreadpool",
4503 ],
4504)
4505
4506xnnpack_cc_library(
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004507 name = "XNNPACK",
4508 srcs = [
4509 "src/init.c",
Marat Dukhanccfdbd12020-02-03 14:27:45 -08004510 "src/runtime.c",
4511 "src/subgraph.c",
4512 "src/tensor.c",
Marat Dukhanf03da0d2020-06-10 16:00:20 -07004513 ] + SUBGRAPH_SRCS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004514 hdrs = ["include/xnnpack.h"],
4515 copts = LOGGING_COPTS + [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004516 "-Isrc",
4517 "-Iinclude",
4518 ] + select({
4519 ":debug_build": [],
4520 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07004521 }) + select({
4522 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4523 "//conditions:default": [],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004524 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07004525 gcc_copts = xnnpack_gcc_std_copts(),
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004526 includes = ["include"],
Marat Dukhan10a38082020-04-17 03:58:35 -07004527 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004528 visibility = xnnpack_visibility(),
4529 deps = [
4530 ":enable_assembly",
Marat Dukhan9de90e02020-06-18 16:04:12 -07004531 ":enable_sparse",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004532 ":logging_utils",
Chao Mei6ddfc602020-05-13 22:29:36 -07004533 ":memory_planner",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004534 ":operator_run",
4535 ":operators",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004536 ":ukernels",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004537 "@clog",
Marat Dukhanab2946c2020-05-21 20:04:13 -07004538 "@FP16",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004539 "@pthreadpool",
Marat Dukhand343c222019-10-07 09:22:14 -07004540 ] + select({
4541 ":emscripten": [],
4542 "//conditions:default": ["@cpuinfo"],
4543 }),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004544)
4545
Marat Dukhan10a38082020-04-17 03:58:35 -07004546xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07004547 name = "XNNPACK_test_mode",
4548 srcs = [
4549 "src/init.c",
4550 "src/runtime.c",
4551 "src/subgraph.c",
4552 "src/tensor.c",
Marat Dukhanf03da0d2020-06-10 16:00:20 -07004553 ] + SUBGRAPH_SRCS,
Marat Dukhan33fcf782020-05-24 14:27:15 -07004554 hdrs = ["include/xnnpack.h"],
4555 copts = LOGGING_COPTS + [
4556 "-Isrc",
4557 "-Iinclude",
4558 "-UNDEBUG",
4559 "-DXNN_TEST_MODE=1",
4560 ] + select({
4561 ":debug_build": [],
4562 "//conditions:default": xnnpack_min_size_copts(),
4563 }) + select({
4564 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4565 "//conditions:default": [],
4566 }),
4567 gcc_copts = xnnpack_gcc_std_copts(),
4568 includes = ["include"],
4569 msvc_copts = xnnpack_msvc_std_copts(),
4570 visibility = xnnpack_visibility(),
4571 deps = [
4572 ":enable_assembly",
Marat Dukhan9de90e02020-06-18 16:04:12 -07004573 ":enable_sparse",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004574 ":logging_utils",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004575 ":memory_planner_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004576 ":operator_run_test_mode",
4577 ":operators_test_mode",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004578 ":ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004579 "@clog",
4580 "@FP16",
4581 "@pthreadpool",
4582 ] + select({
4583 ":emscripten": [],
4584 "//conditions:default": ["@cpuinfo"],
4585 }),
4586)
4587
4588xnnpack_cc_library(
Marat Dukhanae046f52020-06-15 13:16:14 -07004589 name = "xnnpack_f32",
4590 srcs = [
4591 "src/init.c",
4592 "src/runtime.c",
4593 "src/subgraph.c",
4594 "src/tensor.c",
4595 ] + SUBGRAPH_SRCS,
4596 hdrs = ["include/xnnpack.h"],
4597 copts = LOGGING_COPTS + [
4598 "-Isrc",
4599 "-Iinclude",
4600 ] + select({
4601 ":debug_build": [],
4602 "//conditions:default": xnnpack_min_size_copts(),
4603 }) + select({
4604 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4605 "//conditions:default": [],
4606 }),
4607 defines = [
Marat Dukhan16f1e1a2020-08-04 16:38:22 -07004608 "XNN_NO_QS8_OPERATORS",
Marat Dukhan08b7a972020-07-14 18:17:29 -07004609 "XNN_NO_QU8_OPERATORS",
Marat Dukhanae046f52020-06-15 13:16:14 -07004610 "XNN_NO_U8_OPERATORS",
4611 "XNN_NO_X8_OPERATORS",
4612 "XNN_NO_F16_OPERATORS",
4613 "XNN_NO_X16_OPERATORS",
4614 ],
4615 gcc_copts = xnnpack_gcc_std_copts(),
4616 includes = ["include"],
4617 msvc_copts = xnnpack_msvc_std_copts(),
4618 visibility = xnnpack_visibility(),
4619 deps = [
4620 ":enable_assembly",
Marat Dukhan9de90e02020-06-18 16:04:12 -07004621 ":enable_sparse",
Marat Dukhanae046f52020-06-15 13:16:14 -07004622 ":logging_utils",
4623 ":memory_planner",
4624 ":operator_run",
4625 ":operators",
Marat Dukhan86df3842020-06-17 14:03:31 -07004626 ":ukernels_max_compatibility",
Marat Dukhanae046f52020-06-15 13:16:14 -07004627 "@clog",
4628 "@FP16",
4629 "@pthreadpool",
4630 ] + select({
4631 ":emscripten": [],
4632 "//conditions:default": ["@cpuinfo"],
4633 }),
4634)
4635
4636xnnpack_cc_library(
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004637 name = "xnnpack_operators_nhwc_f32",
4638 srcs = [
4639 "src/init.c",
4640 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07004641 hdrs = ["include/xnnpack.h"],
4642 copts = LOGGING_COPTS + [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004643 "-Isrc",
4644 "-Iinclude",
4645 ] + select({
4646 ":debug_build": [],
4647 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07004648 }) + select({
4649 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4650 "//conditions:default": [],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004651 }),
4652 defines = [
Marat Dukhan16f1e1a2020-08-04 16:38:22 -07004653 "XNN_NO_QS8_OPERATORS",
Marat Dukhan08b7a972020-07-14 18:17:29 -07004654 "XNN_NO_QU8_OPERATORS",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004655 "XNN_NO_U8_OPERATORS",
4656 "XNN_NO_X8_OPERATORS",
Marat Dukhanefc47b82019-11-18 09:25:38 -08004657 "XNN_NO_NCHW_OPERATORS",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004658 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07004659 gcc_copts = xnnpack_gcc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004660 includes = ["include"],
Marat Dukhan10a38082020-04-17 03:58:35 -07004661 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004662 visibility = xnnpack_visibility(),
4663 deps = [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004664 ":enable_assembly",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004665 ":logging_utils",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004666 ":operator_run",
4667 ":operators",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004668 ":ukernels",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004669 "@clog",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004670 "@pthreadpool",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004671 ] + select({
4672 ":emscripten": [],
4673 "//conditions:default": ["@cpuinfo"],
4674 }),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004675)
4676
Marat Dukhancf056b22019-10-07 10:26:29 -07004677xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004678 name = "bench_utils",
4679 srcs = ["bench/utils.cc"],
4680 hdrs = ["bench/utils.h"],
Marat Dukhanbad48fe2019-11-04 10:35:22 -08004681 deps = [
4682 "@com_google_benchmark//:benchmark",
4683 "@cpuinfo",
4684 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004685)
4686
Frank Barchard7e955972019-10-11 10:34:25 -07004687######################### Benchmarks for micro-kernels #########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07004688
4689xnnpack_benchmark(
Marat Dukhan595e1702020-07-31 10:12:52 -07004690 name = "qs8_gemm_bench",
4691 srcs = [
4692 "bench/gemm.h",
4693 "bench/qs8-gemm.cc",
4694 "src/xnnpack/AlignedAllocator.h",
4695 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Frank Barchard31328cb2020-10-12 11:55:18 -07004696 copts = xnnpack_optional_ruy_copts() + xnnpack_optional_gemmlowp_copts(),
4697 deps = MICROKERNEL_BENCHMARK_DEPS + [":packing"] + xnnpack_optional_ruy_deps() + xnnpack_optional_gemmlowp_deps(),
Marat Dukhan595e1702020-07-31 10:12:52 -07004698)
4699
4700xnnpack_benchmark(
Marat Dukhan56bdd4a2020-08-03 19:47:04 -07004701 name = "qs8_requantization_bench",
4702 srcs = [
4703 "bench/qs8-requantization.cc",
Marat Dukhan56bdd4a2020-08-03 19:47:04 -07004704 "src/xnnpack/AlignedAllocator.h",
Frank Barchard04336c12020-10-22 16:48:55 -07004705 "src/xnnpack/requantization-stubs.h",
Marat Dukhan56bdd4a2020-08-03 19:47:04 -07004706 ] + MICROKERNEL_BENCHMARK_HDRS,
4707 deps = MICROKERNEL_BENCHMARK_DEPS,
4708)
4709
4710xnnpack_benchmark(
Marat Dukhan08b7a972020-07-14 18:17:29 -07004711 name = "qu8_gemm_bench",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004712 srcs = [
4713 "bench/gemm.h",
Marat Dukhan08b7a972020-07-14 18:17:29 -07004714 "bench/qu8-gemm.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004715 "src/xnnpack/AlignedAllocator.h",
4716 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004717 copts = xnnpack_optional_ruy_copts() + xnnpack_optional_gemmlowp_copts(),
Marat Dukhanab582382020-07-06 13:32:08 -07004718 deps = MICROKERNEL_BENCHMARK_DEPS + [":packing"] + xnnpack_optional_ruy_deps() + xnnpack_optional_gemmlowp_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004719)
4720
4721xnnpack_benchmark(
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07004722 name = "qu8_requantization_bench",
4723 srcs = [
4724 "bench/qu8-requantization.cc",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07004725 "src/xnnpack/AlignedAllocator.h",
Frank Barchard04336c12020-10-22 16:48:55 -07004726 "src/xnnpack/requantization-stubs.h",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07004727 ] + MICROKERNEL_BENCHMARK_HDRS,
4728 deps = MICROKERNEL_BENCHMARK_DEPS,
4729)
4730
4731xnnpack_benchmark(
Frank Barchard40d20fe2020-05-05 00:37:45 -07004732 name = "f16_igemm_bench",
4733 srcs = [
4734 "bench/f16-igemm.cc",
4735 "bench/conv.h",
4736 "bench/google/conv.h",
4737 "src/xnnpack/AlignedAllocator.h",
4738 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004739 deps = MICROKERNEL_BENCHMARK_DEPS + [
4740 ":indirection",
4741 ":packing",
4742 ],
Frank Barchard40d20fe2020-05-05 00:37:45 -07004743)
4744
4745xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004746 name = "f16_gemm_bench",
4747 srcs = [
4748 "bench/f16-gemm.cc",
4749 "bench/gemm.h",
4750 "src/xnnpack/AlignedAllocator.h",
4751 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004752 deps = MICROKERNEL_BENCHMARK_DEPS + [
4753 ":packing",
4754 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004755)
4756
4757xnnpack_benchmark(
Marat Dukhanbdb56f52020-02-05 21:42:49 -08004758 name = "f16_spmm_bench",
4759 srcs = [
4760 "bench/f16-spmm.cc",
4761 "bench/gemm.h",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08004762 "src/xnnpack/AlignedAllocator.h",
4763 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanbdb56f52020-02-05 21:42:49 -08004764 deps = MICROKERNEL_BENCHMARK_DEPS,
4765)
4766
4767xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004768 name = "f32_igemm_bench",
4769 srcs = [
4770 "bench/f32-igemm.cc",
4771 "bench/conv.h",
4772 "src/xnnpack/AlignedAllocator.h",
4773 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004774 deps = MICROKERNEL_BENCHMARK_DEPS + [
4775 ":indirection",
4776 ":packing",
4777 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004778)
4779
4780xnnpack_benchmark(
Frank Barchardc2010062020-07-19 10:04:12 -07004781 name = "f16_relu_bench",
4782 srcs = [
4783 "bench/f16-relu.cc",
4784 "src/xnnpack/AlignedAllocator.h",
4785 ] + MICROKERNEL_BENCHMARK_HDRS,
4786 deps = MICROKERNEL_BENCHMARK_DEPS,
4787)
4788
4789xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004790 name = "f32_conv_hwc_bench",
4791 srcs = [
4792 "bench/f32-conv-hwc.cc",
4793 "bench/dconv.h",
4794 "src/xnnpack/AlignedAllocator.h",
4795 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004796 deps = MICROKERNEL_BENCHMARK_DEPS + [
4797 ":packing",
4798 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004799)
4800
4801xnnpack_benchmark(
Marat Dukhan1f29b802020-05-15 23:46:39 -07004802 name = "f32_conv_hwc2chw_bench",
Erich Elsen563df5f2019-10-23 08:02:21 -07004803 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07004804 "bench/f32-conv-hwc2chw.cc",
Erich Elsen563df5f2019-10-23 08:02:21 -07004805 "bench/dconv.h",
4806 "src/xnnpack/AlignedAllocator.h",
4807 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004808 deps = MICROKERNEL_BENCHMARK_DEPS + [
4809 ":packing",
4810 ],
Erich Elsen563df5f2019-10-23 08:02:21 -07004811)
4812
4813xnnpack_benchmark(
Frank Barchard5a599a62020-06-04 20:12:44 -07004814 name = "f16_dwconv_bench",
4815 srcs = [
4816 "bench/f16-dwconv.cc",
4817 "bench/dwconv.h",
4818 "bench/google/dwconv.h",
4819 "src/xnnpack/AlignedAllocator.h",
4820 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004821 deps = MICROKERNEL_BENCHMARK_DEPS + [
4822 ":indirection",
4823 ":packing",
4824 ],
Frank Barchard5a599a62020-06-04 20:12:44 -07004825)
4826
4827xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004828 name = "f32_dwconv_bench",
4829 srcs = [
4830 "bench/f32-dwconv.cc",
4831 "bench/dwconv.h",
4832 "src/xnnpack/AlignedAllocator.h",
4833 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004834 deps = MICROKERNEL_BENCHMARK_DEPS + [
4835 ":indirection",
4836 ":packing",
4837 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004838)
4839
4840xnnpack_benchmark(
Marat Dukhanbf715f92020-10-23 20:17:00 -07004841 name = "f32_dwconv2d_chw_bench",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004842 srcs = [
Marat Dukhanbf715f92020-10-23 20:17:00 -07004843 "bench/f32-dwconv2d-chw.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004844 "bench/dwconv.h",
4845 "src/xnnpack/AlignedAllocator.h",
4846 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004847 deps = MICROKERNEL_BENCHMARK_DEPS + [
4848 ":indirection",
4849 ":packing",
4850 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004851)
4852
4853xnnpack_benchmark(
4854 name = "f32_gemm_bench",
4855 srcs = [
4856 "bench/f32-gemm.cc",
4857 "bench/gemm.h",
4858 "src/xnnpack/AlignedAllocator.h",
4859 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004860 copts = xnnpack_optional_ruy_copts(),
Marat Dukhanab582382020-07-06 13:32:08 -07004861 deps = MICROKERNEL_BENCHMARK_DEPS + [":packing"] + xnnpack_optional_ruy_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004862)
4863
4864xnnpack_benchmark(
Marat Dukhana11ca342020-06-25 23:45:07 -07004865 name = "f32_hswish_bench",
4866 srcs = [
4867 "bench/f32-hswish.cc",
4868 "src/xnnpack/AlignedAllocator.h",
4869 ] + MICROKERNEL_BENCHMARK_HDRS,
4870 deps = MICROKERNEL_BENCHMARK_DEPS,
4871)
4872
4873xnnpack_benchmark(
Marat Dukhan4c4eb002019-12-08 21:27:49 -08004874 name = "f32_raddexpminusmax_bench",
4875 srcs = [
4876 "bench/f32-raddexpminusmax.cc",
4877 "src/xnnpack/AlignedAllocator.h",
4878 ] + MICROKERNEL_BENCHMARK_HDRS,
4879 deps = MICROKERNEL_BENCHMARK_DEPS,
4880)
4881
4882xnnpack_benchmark(
4883 name = "f32_raddextexp_bench",
4884 srcs = [
4885 "bench/f32-raddextexp.cc",
4886 "src/xnnpack/AlignedAllocator.h",
4887 ] + MICROKERNEL_BENCHMARK_HDRS,
4888 deps = MICROKERNEL_BENCHMARK_DEPS,
4889)
4890
4891xnnpack_benchmark(
4892 name = "f32_raddstoreexpminusmax_bench",
4893 srcs = [
4894 "bench/f32-raddstoreexpminusmax.cc",
4895 "src/xnnpack/AlignedAllocator.h",
4896 ] + MICROKERNEL_BENCHMARK_HDRS,
4897 deps = MICROKERNEL_BENCHMARK_DEPS,
4898)
4899
4900xnnpack_benchmark(
Frank Barchard3c4a9522020-07-18 16:11:22 -07004901 name = "f32_relu_bench",
4902 srcs = [
4903 "bench/f32-relu.cc",
4904 "src/xnnpack/AlignedAllocator.h",
4905 ] + MICROKERNEL_BENCHMARK_HDRS,
4906 deps = MICROKERNEL_BENCHMARK_DEPS,
4907)
4908
4909xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004910 name = "f32_rmax_bench",
4911 srcs = [
4912 "bench/f32-rmax.cc",
4913 "src/xnnpack/AlignedAllocator.h",
4914 ] + MICROKERNEL_BENCHMARK_HDRS,
4915 deps = MICROKERNEL_BENCHMARK_DEPS,
4916)
4917
4918xnnpack_benchmark(
Marat Dukhan14bec502019-11-18 11:35:31 -08004919 name = "f32_sigmoid_bench",
4920 srcs = [
4921 "bench/f32-sigmoid.cc",
4922 "src/xnnpack/AlignedAllocator.h",
4923 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan14bec502019-11-18 11:35:31 -08004924 deps = MICROKERNEL_BENCHMARK_DEPS,
4925)
4926
4927xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004928 name = "f32_spmm_bench",
4929 srcs = [
4930 "bench/f32-spmm.cc",
4931 "bench/gemm.h",
4932 "src/xnnpack/AlignedAllocator.h",
4933 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07004934 deps = MICROKERNEL_BENCHMARK_DEPS,
4935)
4936
4937xnnpack_benchmark(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08004938 name = "f32_softmax_bench",
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07004939 srcs = [
Marat Dukhanfd8e6892020-01-27 15:25:25 -08004940 "bench/f32-softmax.cc",
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07004941 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004942 copts = xnnpack_optional_dnnl_copts(),
Marat Dukhan8d3c6932020-03-06 20:27:27 -08004943 deps = MICROKERNEL_BENCHMARK_DEPS + xnnpack_optional_dnnl_deps(),
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07004944)
4945
4946xnnpack_benchmark(
Marat Dukhaned6baaf2020-12-01 15:07:08 -08004947 name = "f32_velu_bench",
4948 srcs = [
4949 "bench/f32-velu.cc",
4950 "src/xnnpack/AlignedAllocator.h",
4951 ] + MICROKERNEL_BENCHMARK_HDRS,
4952 deps = MICROKERNEL_BENCHMARK_DEPS,
4953)
4954
4955xnnpack_benchmark(
Marat Dukhan4c4eb002019-12-08 21:27:49 -08004956 name = "f32_vscaleexpminusmax_bench",
4957 srcs = [
4958 "bench/f32-vscaleexpminusmax.cc",
4959 "src/xnnpack/AlignedAllocator.h",
4960 ] + MICROKERNEL_BENCHMARK_HDRS,
4961 deps = MICROKERNEL_BENCHMARK_DEPS,
4962)
4963
4964xnnpack_benchmark(
4965 name = "f32_vscaleextexp_bench",
4966 srcs = [
4967 "bench/f32-vscaleextexp.cc",
4968 "src/xnnpack/AlignedAllocator.h",
4969 ] + MICROKERNEL_BENCHMARK_HDRS,
4970 deps = MICROKERNEL_BENCHMARK_DEPS,
4971)
4972
4973xnnpack_benchmark(
Marat Dukhanf4db2f32020-06-30 10:55:30 -07004974 name = "f32_vsqrt_bench",
4975 srcs = [
4976 "bench/f32-vsqrt.cc",
4977 "src/xnnpack/AlignedAllocator.h",
4978 ] + MICROKERNEL_BENCHMARK_HDRS,
4979 deps = MICROKERNEL_BENCHMARK_DEPS,
4980)
4981
4982xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004983 name = "f32_im2col_gemm_bench",
4984 srcs = [
4985 "bench/f32-im2col-gemm.cc",
4986 "bench/conv.h",
4987 "src/xnnpack/AlignedAllocator.h",
4988 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004989 deps = MICROKERNEL_BENCHMARK_DEPS + [
4990 ":im2col",
4991 ":packing",
4992 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004993)
4994
Marat Dukhanfe7acb62020-03-09 19:30:05 -07004995xnnpack_benchmark(
Marat Dukhanffbf96a2020-05-14 02:59:08 -07004996 name = "rounding_bench",
4997 srcs = [
4998 "bench/rounding.cc",
Marat Dukhanffbf96a2020-05-14 02:59:08 -07004999 "src/xnnpack/AlignedAllocator.h",
Frank Barchard04336c12020-10-22 16:48:55 -07005000 "src/xnnpack/math-stubs.h",
Marat Dukhanffbf96a2020-05-14 02:59:08 -07005001 ] + MICROKERNEL_BENCHMARK_HDRS,
5002 deps = MICROKERNEL_BENCHMARK_DEPS,
5003)
5004
Marat Dukhan08c4a432019-10-03 09:29:21 -07005005########################### Benchmarks for operators ###########################
5006
5007xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005008 name = "average_pooling_bench",
5009 srcs = ["bench/average-pooling.cc"],
Marat Dukhan7a16d8b2020-03-11 04:22:44 -07005010 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07005011 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07005012 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07005013)
5014
5015xnnpack_benchmark(
Marat Dukhan87727142020-06-24 15:24:10 -07005016 name = "bankers_rounding_bench",
5017 srcs = ["bench/bankers-rounding.cc"],
5018 copts = xnnpack_optional_tflite_copts(),
5019 tags = ["nowin32"],
5020 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
5021)
5022
5023xnnpack_benchmark(
5024 name = "ceiling_bench",
5025 srcs = ["bench/ceiling.cc"],
5026 copts = xnnpack_optional_tflite_copts(),
5027 tags = ["nowin32"],
5028 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
5029)
5030
5031xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005032 name = "channel_shuffle_bench",
5033 srcs = ["bench/channel-shuffle.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07005034 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005035)
5036
5037xnnpack_benchmark(
5038 name = "convolution_bench",
5039 srcs = ["bench/convolution.cc"],
5040 copts = xnnpack_optional_tflite_copts() + xnnpack_optional_armcl_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07005041 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07005042 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps() + xnnpack_optional_armcl_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07005043)
5044
5045xnnpack_benchmark(
5046 name = "deconvolution_bench",
5047 srcs = ["bench/deconvolution.cc"],
5048 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07005049 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07005050 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07005051)
5052
5053xnnpack_benchmark(
Marat Dukhanb6bd4bc2020-12-01 17:01:40 -08005054 name = "elu_bench",
5055 srcs = ["bench/elu.cc"],
5056 copts = xnnpack_optional_tflite_copts(),
5057 tags = ["nowin32"],
5058 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
5059)
5060
5061xnnpack_benchmark(
Marat Dukhan87727142020-06-24 15:24:10 -07005062 name = "floor_bench",
5063 srcs = ["bench/floor.cc"],
5064 copts = xnnpack_optional_tflite_copts(),
5065 tags = ["nowin32"],
5066 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
5067)
5068
5069xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005070 name = "global_average_pooling_bench",
5071 srcs = ["bench/global-average-pooling.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07005072 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005073)
5074
5075xnnpack_benchmark(
Marat Dukhanad352602020-06-25 21:50:54 -07005076 name = "hardswish_bench",
5077 srcs = ["bench/hardswish.cc"],
5078 copts = xnnpack_optional_tflite_copts(),
5079 tags = ["nowin32"],
5080 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
5081)
5082
5083xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005084 name = "max_pooling_bench",
5085 srcs = ["bench/max-pooling.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07005086 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005087)
5088
5089xnnpack_benchmark(
5090 name = "sigmoid_bench",
5091 srcs = ["bench/sigmoid.cc"],
Marat Dukhanc3b9e862019-11-17 13:18:54 -08005092 copts = xnnpack_optional_tflite_copts(),
Marat Dukhanca2ba702020-04-24 01:31:47 -07005093 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07005094 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07005095)
5096
5097xnnpack_benchmark(
Marat Dukhan95b22432019-10-30 16:30:14 -07005098 name = "prelu_bench",
5099 srcs = ["bench/prelu.cc"],
5100 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07005101 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07005102 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan95b22432019-10-30 16:30:14 -07005103)
5104
5105xnnpack_benchmark(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08005106 name = "softmax_bench",
5107 srcs = ["bench/softmax.cc"],
Marat Dukhan9c0db962020-01-28 12:30:14 -08005108 copts = xnnpack_optional_tflite_copts(),
Marat Dukhanca2ba702020-04-24 01:31:47 -07005109 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07005110 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07005111)
5112
Marat Dukhan87727142020-06-24 15:24:10 -07005113xnnpack_benchmark(
Marat Dukhan6804bbd2020-06-30 19:26:11 -07005114 name = "square_root_bench",
5115 srcs = ["bench/square-root.cc"],
5116 copts = xnnpack_optional_tflite_copts(),
5117 tags = ["nowin32"],
5118 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
5119)
5120
5121xnnpack_benchmark(
Marat Dukhan87727142020-06-24 15:24:10 -07005122 name = "truncation_bench",
5123 srcs = ["bench/truncation.cc"],
5124 deps = OPERATOR_BENCHMARK_DEPS,
5125)
5126
Marat Dukhanc068bb62019-10-04 13:24:39 -07005127############################# End-to-end benchmarks ############################
5128
5129cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07005130 name = "fp32_mobilenet_v1",
5131 srcs = ["models/fp32-mobilenet-v1.cc"],
Marat Dukhanc068bb62019-10-04 13:24:39 -07005132 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08005133 copts = xnnpack_std_cxxopts(),
Marat Dukhanc068bb62019-10-04 13:24:39 -07005134 linkstatic = True,
5135 deps = [
5136 ":XNNPACK",
5137 "@pthreadpool",
5138 ],
5139)
5140
5141cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07005142 name = "fp16_mobilenet_v1",
5143 srcs = ["models/fp16-mobilenet-v1.cc"],
5144 hdrs = ["models/models.h"],
5145 copts = xnnpack_std_cxxopts(),
5146 linkstatic = True,
5147 deps = [
5148 ":XNNPACK",
5149 "@FP16",
5150 "@pthreadpool",
5151 ],
5152)
5153
5154cc_library(
Marat Dukhan0743cdf2020-08-04 18:52:07 -07005155 name = "qs8_mobilenet_v1",
5156 srcs = ["models/qs8-mobilenet-v1.cc"],
5157 hdrs = ["models/models.h"],
5158 copts = xnnpack_std_cxxopts(),
5159 linkstatic = True,
5160 deps = [
5161 ":XNNPACK",
5162 "@pthreadpool",
5163 ],
5164)
5165
5166cc_library(
Marat Dukhan70a96182020-09-03 17:13:58 -07005167 name = "qs8_mobilenet_v2",
5168 srcs = ["models/qs8-mobilenet-v2.cc"],
5169 hdrs = ["models/models.h"],
5170 copts = xnnpack_std_cxxopts(),
5171 linkstatic = True,
5172 deps = [
5173 ":XNNPACK",
5174 "@pthreadpool",
5175 ],
5176)
5177
5178cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07005179 name = "fp32_mobilenet_v2",
5180 srcs = ["models/fp32-mobilenet-v2.cc"],
Marat Dukhanc068bb62019-10-04 13:24:39 -07005181 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08005182 copts = xnnpack_std_cxxopts(),
Marat Dukhanc068bb62019-10-04 13:24:39 -07005183 linkstatic = True,
5184 deps = [
5185 ":XNNPACK",
5186 "@pthreadpool",
5187 ],
5188)
5189
Marat Dukhanc08cdf52019-12-09 09:17:51 -08005190cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07005191 name = "fp16_mobilenet_v2",
5192 srcs = ["models/fp16-mobilenet-v2.cc"],
5193 hdrs = ["models/models.h"],
5194 copts = xnnpack_std_cxxopts(),
5195 linkstatic = True,
5196 deps = [
5197 ":XNNPACK",
5198 "@FP16",
5199 "@pthreadpool",
5200 ],
5201)
5202
5203cc_library(
5204 name = "fp32_mobilenet_v3_large",
5205 srcs = ["models/fp32-mobilenet-v3-large.cc"],
Marat Dukhanc08cdf52019-12-09 09:17:51 -08005206 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08005207 copts = xnnpack_std_cxxopts(),
Marat Dukhanc08cdf52019-12-09 09:17:51 -08005208 linkstatic = True,
5209 deps = [
5210 ":XNNPACK",
5211 "@pthreadpool",
5212 ],
5213)
5214
5215cc_library(
Marat Dukhan66f3ccd2020-09-14 16:09:38 -07005216 name = "fp16_mobilenet_v3_large",
5217 srcs = ["models/fp16-mobilenet-v3-large.cc"],
5218 hdrs = ["models/models.h"],
5219 copts = xnnpack_std_cxxopts(),
5220 linkstatic = True,
5221 deps = [
5222 ":XNNPACK",
5223 "@FP16",
5224 "@pthreadpool",
5225 ],
5226)
5227
5228cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07005229 name = "fp32_mobilenet_v3_small",
5230 srcs = ["models/fp32-mobilenet-v3-small.cc"],
Marat Dukhanc08cdf52019-12-09 09:17:51 -08005231 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08005232 copts = xnnpack_std_cxxopts(),
Marat Dukhanc08cdf52019-12-09 09:17:51 -08005233 linkstatic = True,
5234 deps = [
5235 ":XNNPACK",
5236 "@pthreadpool",
5237 ],
5238)
5239
Marat Dukhan66f3ccd2020-09-14 16:09:38 -07005240cc_library(
5241 name = "fp16_mobilenet_v3_small",
5242 srcs = ["models/fp16-mobilenet-v3-small.cc"],
5243 hdrs = ["models/models.h"],
5244 copts = xnnpack_std_cxxopts(),
5245 linkstatic = True,
5246 deps = [
5247 ":XNNPACK",
5248 "@FP16",
5249 "@pthreadpool",
5250 ],
5251)
5252
Marat Dukhanc068bb62019-10-04 13:24:39 -07005253xnnpack_benchmark(
Marat Dukhanef4416e2019-10-31 13:44:40 -07005254 name = "f32_dwconv_e2e_bench",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08005255 srcs = [
5256 "bench/f32-dwconv-e2e.cc",
5257 "bench/end2end.h",
5258 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanef4416e2019-10-31 13:44:40 -07005259 deps = MICROKERNEL_BENCHMARK_DEPS + [
5260 ":XNNPACK",
Marat Dukhan270a2c42020-06-26 16:45:52 -07005261 ":fp32_mobilenet_v1",
5262 ":fp32_mobilenet_v2",
5263 ":fp32_mobilenet_v3_large",
5264 ":fp32_mobilenet_v3_small",
Marat Dukhanef4416e2019-10-31 13:44:40 -07005265 ],
5266)
5267
5268xnnpack_benchmark(
Marat Dukhan5f18d262019-10-31 10:24:14 -07005269 name = "f32_gemm_e2e_bench",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08005270 srcs = [
5271 "bench/f32-gemm-e2e.cc",
5272 "bench/end2end.h",
5273 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan5f18d262019-10-31 10:24:14 -07005274 deps = MICROKERNEL_BENCHMARK_DEPS + [
5275 ":XNNPACK",
Marat Dukhan270a2c42020-06-26 16:45:52 -07005276 ":fp32_mobilenet_v1",
5277 ":fp32_mobilenet_v2",
5278 ":fp32_mobilenet_v3_large",
5279 ":fp32_mobilenet_v3_small",
Marat Dukhan5f18d262019-10-31 10:24:14 -07005280 ],
5281)
5282
5283xnnpack_benchmark(
Marat Dukhanc068bb62019-10-04 13:24:39 -07005284 name = "end2end_bench",
5285 srcs = ["bench/end2end.cc"],
5286 deps = [
5287 ":XNNPACK",
Frank Barchardc712fa42019-10-31 14:00:21 -07005288 ":bench_utils",
Marat Dukhan270a2c42020-06-26 16:45:52 -07005289 ":fp16_mobilenet_v1",
5290 ":fp16_mobilenet_v2",
Marat Dukhan66f3ccd2020-09-14 16:09:38 -07005291 ":fp16_mobilenet_v3_large",
5292 ":fp16_mobilenet_v3_small",
Marat Dukhan270a2c42020-06-26 16:45:52 -07005293 ":fp32_mobilenet_v1",
5294 ":fp32_mobilenet_v2",
5295 ":fp32_mobilenet_v3_large",
5296 ":fp32_mobilenet_v3_small",
Marat Dukhan0743cdf2020-08-04 18:52:07 -07005297 ":qs8_mobilenet_v1",
Marat Dukhan70a96182020-09-03 17:13:58 -07005298 ":qs8_mobilenet_v2",
Marat Dukhanc068bb62019-10-04 13:24:39 -07005299 "@pthreadpool",
5300 ],
5301)
5302
Marat Dukhan6adff4e2019-10-14 18:32:07 -07005303#################### Accuracy evaluation for math functions ####################
5304
5305xnnpack_benchmark(
5306 name = "f32_exp_eval",
5307 srcs = [
5308 "eval/f32-exp.cc",
5309 "src/xnnpack/AlignedAllocator.h",
5310 ] + ACCURACY_EVAL_HDRS,
5311 deps = ACCURACY_EVAL_DEPS,
5312)
5313
Marat Dukhan515c9772019-10-17 18:07:57 -07005314xnnpack_benchmark(
5315 name = "f32_expminus_eval",
5316 srcs = [
5317 "eval/f32-expminus.cc",
5318 "src/xnnpack/AlignedAllocator.h",
5319 ] + ACCURACY_EVAL_HDRS,
5320 deps = ACCURACY_EVAL_DEPS,
5321)
5322
Marat Dukhan98ba4412019-10-23 02:14:28 -07005323xnnpack_benchmark(
Marat Dukhana438aca2020-11-20 15:45:01 -08005324 name = "f32_expm1minus_eval",
5325 srcs = [
5326 "eval/f32-expm1minus.cc",
5327 "src/xnnpack/AlignedAllocator.h",
5328 ] + ACCURACY_EVAL_HDRS,
Marat Dukhande390d42020-11-29 19:32:18 -08005329 deps = ACCURACY_EVAL_DEPS + [
5330 ":bench_utils",
5331 ],
Marat Dukhana438aca2020-11-20 15:45:01 -08005332)
5333
5334xnnpack_benchmark(
Marat Dukhan98ba4412019-10-23 02:14:28 -07005335 name = "f32_extexp_eval",
5336 srcs = [
5337 "eval/f32-extexp.cc",
5338 "src/xnnpack/AlignedAllocator.h",
5339 ] + ACCURACY_EVAL_HDRS,
5340 deps = ACCURACY_EVAL_DEPS,
5341)
5342
Marat Dukhan8853b822020-05-07 12:19:01 -07005343xnnpack_unit_test(
5344 name = "f32_roundne_eval",
5345 srcs = [
5346 "eval/f32-roundne.cc",
5347 "src/xnnpack/AlignedAllocator.h",
5348 "src/xnnpack/math-stubs.h",
5349 ] + MICROKERNEL_TEST_HDRS,
Marat Dukhan22eed3d2020-05-11 20:13:37 -07005350 automatic = False,
Marat Dukhan8853b822020-05-07 12:19:01 -07005351 deps = MICROKERNEL_TEST_DEPS,
5352)
5353
Marat Dukhan2dbb9442020-05-12 20:43:43 -07005354xnnpack_unit_test(
Marat Dukhanc9852ba2020-05-13 17:21:29 -07005355 name = "f32_roundd_eval",
5356 srcs = [
5357 "eval/f32-roundd.cc",
5358 "src/xnnpack/AlignedAllocator.h",
5359 "src/xnnpack/math-stubs.h",
5360 ] + MICROKERNEL_TEST_HDRS,
5361 automatic = False,
5362 deps = MICROKERNEL_TEST_DEPS,
5363)
5364
5365xnnpack_unit_test(
5366 name = "f32_roundu_eval",
5367 srcs = [
5368 "eval/f32-roundu.cc",
5369 "src/xnnpack/AlignedAllocator.h",
5370 "src/xnnpack/math-stubs.h",
5371 ] + MICROKERNEL_TEST_HDRS,
5372 automatic = False,
5373 deps = MICROKERNEL_TEST_DEPS,
5374)
5375
5376xnnpack_unit_test(
Marat Dukhan2dbb9442020-05-12 20:43:43 -07005377 name = "f32_roundz_eval",
5378 srcs = [
5379 "eval/f32-roundz.cc",
5380 "src/xnnpack/AlignedAllocator.h",
5381 "src/xnnpack/math-stubs.h",
5382 ] + MICROKERNEL_TEST_HDRS,
Marat Dukhanc9852ba2020-05-13 17:21:29 -07005383 automatic = False,
Marat Dukhan2dbb9442020-05-12 20:43:43 -07005384 deps = MICROKERNEL_TEST_DEPS,
5385)
5386
Marat Dukhan346a9e52019-11-15 09:06:30 -08005387xnnpack_benchmark(
5388 name = "f32_sigmoid_eval",
5389 srcs = [
5390 "eval/f32-sigmoid.cc",
5391 "src/xnnpack/AlignedAllocator.h",
5392 ] + ACCURACY_EVAL_HDRS,
5393 deps = ACCURACY_EVAL_DEPS,
5394)
5395
Marat Dukhan84000762020-06-29 18:38:43 -07005396xnnpack_benchmark(
5397 name = "f32_sqrt_eval",
5398 srcs = [
5399 "eval/f32-sqrt.cc",
5400 "src/xnnpack/AlignedAllocator.h",
5401 ] + ACCURACY_EVAL_HDRS,
5402 deps = ACCURACY_EVAL_DEPS + [
5403 ":bench_utils",
5404 ],
5405)
5406
Marat Dukhan08c4a432019-10-03 09:29:21 -07005407######################### Unit tests for micro-kernels #########################
5408
5409xnnpack_unit_test(
Marat Dukhande06f492020-04-09 00:19:31 -07005410 name = "f16_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005411 srcs = [
Marat Dukhande06f492020-04-09 00:19:31 -07005412 "test/f16-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005413 "test/gemm-microkernel-tester.h",
5414 "src/xnnpack/AlignedAllocator.h",
5415 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005416 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005417)
5418
5419xnnpack_unit_test(
Marat Dukhan355ab432020-04-09 19:01:52 -07005420 name = "f16_spmm_minmax_test",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08005421 srcs = [
Marat Dukhan355ab432020-04-09 19:01:52 -07005422 "test/f16-spmm-minmax.cc",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08005423 "test/spmm-microkernel-tester.h",
5424 "src/xnnpack/AlignedAllocator.h",
5425 ] + MICROKERNEL_TEST_HDRS,
5426 deps = MICROKERNEL_TEST_DEPS,
5427)
5428
5429xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005430 name = "f32_argmaxpool_test",
5431 srcs = [
5432 "test/f32-argmaxpool.cc",
5433 "test/argmaxpool-microkernel-tester.h",
5434 "src/xnnpack/AlignedAllocator.h",
5435 ] + MICROKERNEL_TEST_HDRS,
5436 deps = MICROKERNEL_TEST_DEPS,
5437)
5438
5439xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07005440 name = "f32_avgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005441 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07005442 "test/f32-avgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005443 "test/avgpool-microkernel-tester.h",
5444 "src/xnnpack/AlignedAllocator.h",
5445 ] + MICROKERNEL_TEST_HDRS,
5446 deps = MICROKERNEL_TEST_DEPS,
5447)
5448
5449xnnpack_unit_test(
Marat Dukhan660fd192020-03-10 04:55:30 -07005450 name = "f32_ibilinear_test",
Marat Dukhan35dacfb2019-11-07 19:18:16 -08005451 srcs = [
Marat Dukhan660fd192020-03-10 04:55:30 -07005452 "test/f32-ibilinear.cc",
5453 "test/ibilinear-microkernel-tester.h",
Marat Dukhan35dacfb2019-11-07 19:18:16 -08005454 "src/xnnpack/AlignedAllocator.h",
5455 ] + MICROKERNEL_TEST_HDRS,
5456 deps = MICROKERNEL_TEST_DEPS,
5457)
5458
5459xnnpack_unit_test(
XNNPACK Team21432672020-10-19 19:58:48 -07005460 name = "f32_ibilinear_chw_test",
5461 srcs = [
5462 "test/f32-ibilinear-chw.cc",
XNNPACK Team6be46b22020-10-22 23:34:54 -07005463 "test/ibilinear-microkernel-tester.h",
XNNPACK Team21432672020-10-19 19:58:48 -07005464 "src/xnnpack/AlignedAllocator.h",
5465 ] + MICROKERNEL_TEST_HDRS,
5466 deps = MICROKERNEL_TEST_DEPS,
5467)
5468
5469xnnpack_unit_test(
Frank Barchardb1966592020-05-12 13:47:06 -07005470 name = "f16_clamp_test",
5471 srcs = [
5472 "test/f16-clamp.cc",
5473 "test/clamp-microkernel-tester.h",
5474 ] + MICROKERNEL_TEST_HDRS,
5475 deps = MICROKERNEL_TEST_DEPS,
5476)
5477
5478xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005479 name = "f32_clamp_test",
5480 srcs = [
5481 "test/f32-clamp.cc",
5482 "test/clamp-microkernel-tester.h",
5483 ] + MICROKERNEL_TEST_HDRS,
5484 deps = MICROKERNEL_TEST_DEPS,
5485)
5486
5487xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07005488 name = "f32_igemm_test",
5489 srcs = [
5490 "test/f32-igemm.cc",
5491 "test/gemm-microkernel-tester.h",
5492 "src/xnnpack/AlignedAllocator.h",
5493 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005494 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan163a7e62020-04-09 04:19:26 -07005495)
5496
5497xnnpack_unit_test(
Marat Dukhan467f6362020-05-22 23:21:55 -07005498 name = "f32_igemm_relu_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005499 srcs = [
Marat Dukhan467f6362020-05-22 23:21:55 -07005500 "test/f32-igemm-relu.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005501 "test/gemm-microkernel-tester.h",
5502 "src/xnnpack/AlignedAllocator.h",
5503 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005504 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005505)
5506
5507xnnpack_unit_test(
Marat Dukhane207b7b2020-05-28 16:27:42 -07005508 name = "f32_igemm_minmax_test",
5509 srcs = [
5510 "test/f32-igemm-minmax.cc",
5511 "test/gemm-microkernel-tester.h",
5512 "src/xnnpack/AlignedAllocator.h",
5513 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005514 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhane207b7b2020-05-28 16:27:42 -07005515)
5516
5517xnnpack_unit_test(
Frank Barchardb0e4fae2020-05-04 15:27:51 -07005518 name = "f16_igemm_minmax_test",
5519 srcs = [
5520 "test/f16-igemm-minmax.cc",
5521 "test/gemm-microkernel-tester.h",
5522 "src/xnnpack/AlignedAllocator.h",
5523 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005524 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Frank Barchardb0e4fae2020-05-04 15:27:51 -07005525)
5526
5527xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005528 name = "f32_conv_hwc_test",
5529 srcs = [
5530 "test/f32-conv-hwc.cc",
5531 "test/conv-hwc-microkernel-tester.h",
5532 "src/xnnpack/AlignedAllocator.h",
5533 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005534 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005535)
5536
5537xnnpack_unit_test(
Marat Dukhan1f29b802020-05-15 23:46:39 -07005538 name = "f32_conv_hwc2chw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005539 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07005540 "test/f32-conv-hwc2chw.cc",
5541 "test/conv-hwc2chw-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005542 "src/xnnpack/AlignedAllocator.h",
5543 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005544 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005545)
5546
5547xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07005548 name = "f32_dwconv_test",
5549 srcs = [
5550 "test/f32-dwconv.cc",
5551 "test/dwconv-microkernel-tester.h",
5552 "src/xnnpack/AlignedAllocator.h",
5553 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005554 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan163a7e62020-04-09 04:19:26 -07005555)
5556
5557xnnpack_unit_test(
Frank Barchard5a599a62020-06-04 20:12:44 -07005558 name = "f16_dwconv_minmax_test",
5559 srcs = [
5560 "test/f16-dwconv-minmax.cc",
5561 "test/dwconv-microkernel-tester.h",
5562 "src/xnnpack/AlignedAllocator.h",
5563 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005564 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Frank Barchard5a599a62020-06-04 20:12:44 -07005565)
5566
5567xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07005568 name = "f32_dwconv_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005569 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07005570 "test/f32-dwconv-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005571 "test/dwconv-microkernel-tester.h",
5572 "src/xnnpack/AlignedAllocator.h",
5573 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005574 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005575)
5576
5577xnnpack_unit_test(
Marat Dukhanbf715f92020-10-23 20:17:00 -07005578 name = "f32_dwconv2d_chw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005579 srcs = [
Marat Dukhanbf715f92020-10-23 20:17:00 -07005580 "test/f32-dwconv2d-chw.cc",
5581 "test/dwconv2d-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005582 "src/xnnpack/AlignedAllocator.h",
5583 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005584 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005585)
5586
5587xnnpack_unit_test(
Frank Barchard0bb49a72020-06-04 11:35:11 -07005588 name = "f16_gavgpool_minmax_test",
5589 srcs = [
5590 "test/f16-gavgpool-minmax.cc",
5591 "test/gavgpool-microkernel-tester.h",
5592 "src/xnnpack/AlignedAllocator.h",
5593 ] + MICROKERNEL_TEST_HDRS,
5594 deps = MICROKERNEL_TEST_DEPS,
5595)
5596
5597xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07005598 name = "f32_gavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005599 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07005600 "test/f32-gavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005601 "test/gavgpool-microkernel-tester.h",
5602 "src/xnnpack/AlignedAllocator.h",
5603 ] + MICROKERNEL_TEST_HDRS,
5604 deps = MICROKERNEL_TEST_DEPS,
5605)
5606
5607xnnpack_unit_test(
Marat Dukhan1f29b802020-05-15 23:46:39 -07005608 name = "f32_gavgpool_cw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005609 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07005610 "test/f32-gavgpool-cw.cc",
5611 "test/gavgpool-cw-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005612 "src/xnnpack/AlignedAllocator.h",
5613 ] + MICROKERNEL_TEST_HDRS,
5614 deps = MICROKERNEL_TEST_DEPS,
5615)
5616
5617xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07005618 name = "f32_gemm_test",
5619 srcs = [
5620 "test/f32-gemm.cc",
5621 "test/gemm-microkernel-tester.h",
5622 "src/xnnpack/AlignedAllocator.h",
5623 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005624 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan163a7e62020-04-09 04:19:26 -07005625)
5626
5627xnnpack_unit_test(
Marat Dukhan467f6362020-05-22 23:21:55 -07005628 name = "f32_gemm_relu_test",
5629 srcs = [
5630 "test/f32-gemm-relu.cc",
5631 "test/gemm-microkernel-tester.h",
5632 "src/xnnpack/AlignedAllocator.h",
5633 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005634 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan467f6362020-05-22 23:21:55 -07005635)
5636
5637xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07005638 name = "f32_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005639 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07005640 "test/f32-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005641 "test/gemm-microkernel-tester.h",
5642 "src/xnnpack/AlignedAllocator.h",
5643 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005644 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005645)
5646
5647xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07005648 name = "f32_gemminc_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005649 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07005650 "test/f32-gemminc-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005651 "test/gemm-microkernel-tester.h",
5652 "src/xnnpack/AlignedAllocator.h",
5653 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005654 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005655)
5656
5657xnnpack_unit_test(
Frank Barchardb1966592020-05-12 13:47:06 -07005658 name = "f16_hswish_test",
5659 srcs = [
5660 "test/f16-hswish.cc",
5661 "test/hswish-microkernel-tester.h",
5662 ] + MICROKERNEL_TEST_HDRS,
5663 deps = MICROKERNEL_TEST_DEPS,
5664)
5665
5666xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005667 name = "f32_hswish_test",
5668 srcs = [
5669 "test/f32-hswish.cc",
5670 "test/hswish-microkernel-tester.h",
5671 ] + MICROKERNEL_TEST_HDRS,
5672 deps = MICROKERNEL_TEST_DEPS,
5673)
5674
5675xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07005676 name = "f32_maxpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005677 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07005678 "test/f32-maxpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005679 "test/maxpool-microkernel-tester.h",
5680 ] + MICROKERNEL_TEST_HDRS,
5681 deps = MICROKERNEL_TEST_DEPS,
5682)
5683
5684xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07005685 name = "f32_pavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005686 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07005687 "test/f32-pavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005688 "test/avgpool-microkernel-tester.h",
5689 "src/xnnpack/AlignedAllocator.h",
5690 ] + MICROKERNEL_TEST_HDRS,
5691 deps = MICROKERNEL_TEST_DEPS,
5692)
5693
5694xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07005695 name = "f32_ppmm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005696 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07005697 "test/f32-ppmm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005698 "test/gemm-microkernel-tester.h",
5699 "src/xnnpack/AlignedAllocator.h",
5700 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005701 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005702)
5703
5704xnnpack_unit_test(
Frank Barchardb1966592020-05-12 13:47:06 -07005705 name = "f16_prelu_test",
5706 srcs = [
5707 "test/f16-prelu.cc",
5708 "test/prelu-microkernel-tester.h",
5709 "src/xnnpack/AlignedAllocator.h",
5710 ] + MICROKERNEL_TEST_HDRS,
5711 deps = MICROKERNEL_TEST_DEPS,
5712)
5713
5714xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005715 name = "f32_prelu_test",
5716 srcs = [
5717 "test/f32-prelu.cc",
5718 "test/prelu-microkernel-tester.h",
5719 "src/xnnpack/AlignedAllocator.h",
5720 ] + MICROKERNEL_TEST_HDRS,
5721 deps = MICROKERNEL_TEST_DEPS,
5722)
5723
5724xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07005725 name = "f32_raddexpminusmax_test",
5726 srcs = [
5727 "test/f32-raddexpminusmax.cc",
5728 "test/raddexpminusmax-microkernel-tester.h",
5729 ] + MICROKERNEL_TEST_HDRS,
5730 deps = MICROKERNEL_TEST_DEPS,
5731)
5732
5733xnnpack_unit_test(
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07005734 name = "f32_raddextexp_test",
5735 srcs = [
5736 "test/f32-raddextexp.cc",
5737 "test/raddextexp-microkernel-tester.h",
5738 ] + MICROKERNEL_TEST_HDRS,
5739 deps = MICROKERNEL_TEST_DEPS,
5740)
5741
5742xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07005743 name = "f32_raddstoreexpminusmax_test",
5744 srcs = [
5745 "test/f32-raddstoreexpminusmax.cc",
5746 "test/raddstoreexpminusmax-microkernel-tester.h",
5747 ] + MICROKERNEL_TEST_HDRS,
5748 deps = MICROKERNEL_TEST_DEPS,
5749)
5750
5751xnnpack_unit_test(
Frank Barchardfb158e22020-07-15 16:10:10 -07005752 name = "f32_relu_test",
5753 srcs = [
5754 "test/f32-relu.cc",
5755 "test/vunary-microkernel-tester.h",
5756 ] + MICROKERNEL_TEST_HDRS,
5757 deps = MICROKERNEL_TEST_DEPS,
5758)
5759
5760xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005761 name = "f32_rmax_test",
5762 srcs = [
5763 "test/f32-rmax.cc",
5764 "test/rmax-microkernel-tester.h",
5765 ] + MICROKERNEL_TEST_HDRS,
5766 deps = MICROKERNEL_TEST_DEPS,
5767)
5768
5769xnnpack_unit_test(
Marat Dukhan346a9e52019-11-15 09:06:30 -08005770 name = "f32_sigmoid_test",
5771 srcs = [
5772 "test/f32-sigmoid.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005773 "test/vunary-microkernel-tester.h",
Marat Dukhan346a9e52019-11-15 09:06:30 -08005774 ] + MICROKERNEL_TEST_HDRS,
5775 deps = MICROKERNEL_TEST_DEPS,
5776)
5777
5778xnnpack_unit_test(
Marat Dukhan355ab432020-04-09 19:01:52 -07005779 name = "f32_spmm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005780 srcs = [
Marat Dukhan355ab432020-04-09 19:01:52 -07005781 "test/f32-spmm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005782 "test/spmm-microkernel-tester.h",
5783 "src/xnnpack/AlignedAllocator.h",
5784 ] + MICROKERNEL_TEST_HDRS,
5785 deps = MICROKERNEL_TEST_DEPS,
5786)
5787
5788xnnpack_unit_test(
Frank Barchardd793f6c2020-05-08 13:37:43 -07005789 name = "f16_vadd_minmax_test",
5790 srcs = [
5791 "test/f16-vadd-minmax.cc",
5792 "test/vbinary-microkernel-tester.h",
5793 ] + MICROKERNEL_TEST_HDRS,
5794 deps = MICROKERNEL_TEST_DEPS,
5795)
5796
5797xnnpack_unit_test(
5798 name = "f16_vaddc_minmax_test",
5799 srcs = [
5800 "test/f16-vaddc-minmax.cc",
5801 "test/vbinaryc-microkernel-tester.h",
5802 ] + MICROKERNEL_TEST_HDRS,
5803 deps = MICROKERNEL_TEST_DEPS,
5804)
5805
5806xnnpack_unit_test(
5807 name = "f16_vdiv_minmax_test",
5808 srcs = [
5809 "test/f16-vdiv-minmax.cc",
5810 "test/vbinary-microkernel-tester.h",
5811 ] + MICROKERNEL_TEST_HDRS,
5812 deps = MICROKERNEL_TEST_DEPS,
5813)
5814
5815xnnpack_unit_test(
5816 name = "f16_vdivc_minmax_test",
5817 srcs = [
5818 "test/f16-vdivc-minmax.cc",
5819 "test/vbinaryc-microkernel-tester.h",
5820 ] + MICROKERNEL_TEST_HDRS,
5821 deps = MICROKERNEL_TEST_DEPS,
5822)
5823
5824xnnpack_unit_test(
5825 name = "f16_vrdivc_minmax_test",
5826 srcs = [
5827 "test/f16-vrdivc-minmax.cc",
5828 "test/vbinaryc-microkernel-tester.h",
5829 ] + MICROKERNEL_TEST_HDRS,
5830 deps = MICROKERNEL_TEST_DEPS,
5831)
5832
5833xnnpack_unit_test(
5834 name = "f16_vmax_test",
5835 srcs = [
5836 "test/f16-vmax.cc",
5837 "test/vbinary-microkernel-tester.h",
5838 ] + MICROKERNEL_TEST_HDRS,
5839 deps = MICROKERNEL_TEST_DEPS,
5840)
5841
5842xnnpack_unit_test(
5843 name = "f16_vmaxc_test",
5844 srcs = [
5845 "test/f16-vmaxc.cc",
5846 "test/vbinaryc-microkernel-tester.h",
5847 ] + MICROKERNEL_TEST_HDRS,
5848 deps = MICROKERNEL_TEST_DEPS,
5849)
5850
5851xnnpack_unit_test(
5852 name = "f16_vmin_test",
5853 srcs = [
5854 "test/f16-vmin.cc",
5855 "test/vbinary-microkernel-tester.h",
5856 ] + MICROKERNEL_TEST_HDRS,
5857 deps = MICROKERNEL_TEST_DEPS,
5858)
5859
5860xnnpack_unit_test(
5861 name = "f16_vminc_test",
5862 srcs = [
5863 "test/f16-vminc.cc",
5864 "test/vbinaryc-microkernel-tester.h",
5865 ] + MICROKERNEL_TEST_HDRS,
5866 deps = MICROKERNEL_TEST_DEPS,
5867)
5868
5869xnnpack_unit_test(
5870 name = "f16_vmul_minmax_test",
5871 srcs = [
5872 "test/f16-vmul-minmax.cc",
5873 "test/vbinary-microkernel-tester.h",
5874 ] + MICROKERNEL_TEST_HDRS,
5875 deps = MICROKERNEL_TEST_DEPS,
5876)
5877
5878xnnpack_unit_test(
5879 name = "f16_vmulc_minmax_test",
5880 srcs = [
5881 "test/f16-vmulc-minmax.cc",
5882 "test/vbinaryc-microkernel-tester.h",
5883 ] + MICROKERNEL_TEST_HDRS,
5884 deps = MICROKERNEL_TEST_DEPS,
5885)
5886
5887xnnpack_unit_test(
5888 name = "f16_vsub_minmax_test",
5889 srcs = [
5890 "test/f16-vsub-minmax.cc",
5891 "test/vbinary-microkernel-tester.h",
5892 ] + MICROKERNEL_TEST_HDRS,
5893 deps = MICROKERNEL_TEST_DEPS,
5894)
5895
5896xnnpack_unit_test(
5897 name = "f16_vsubc_minmax_test",
5898 srcs = [
5899 "test/f16-vsubc-minmax.cc",
5900 "test/vbinaryc-microkernel-tester.h",
5901 ] + MICROKERNEL_TEST_HDRS,
5902 deps = MICROKERNEL_TEST_DEPS,
5903)
5904
5905xnnpack_unit_test(
5906 name = "f16_vrsubc_minmax_test",
5907 srcs = [
5908 "test/f16-vrsubc-minmax.cc",
5909 "test/vbinaryc-microkernel-tester.h",
5910 ] + MICROKERNEL_TEST_HDRS,
5911 deps = MICROKERNEL_TEST_DEPS,
5912)
5913
5914xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07005915 name = "f32_vabs_test",
5916 srcs = [
5917 "test/f32-vabs.cc",
5918 "test/vunary-microkernel-tester.h",
5919 ] + MICROKERNEL_TEST_HDRS,
5920 deps = MICROKERNEL_TEST_DEPS,
5921)
5922
5923xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005924 name = "f32_vadd_test",
5925 srcs = [
5926 "test/f32-vadd.cc",
5927 "test/vbinary-microkernel-tester.h",
5928 ] + MICROKERNEL_TEST_HDRS,
5929 deps = MICROKERNEL_TEST_DEPS,
5930)
5931
5932xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005933 name = "f32_vadd_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005934 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005935 "test/f32-vadd-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005936 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005937 ] + MICROKERNEL_TEST_HDRS,
5938 deps = MICROKERNEL_TEST_DEPS,
5939)
5940
5941xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005942 name = "f32_vadd_relu_test",
5943 srcs = [
5944 "test/f32-vadd-relu.cc",
5945 "test/vbinary-microkernel-tester.h",
5946 ] + MICROKERNEL_TEST_HDRS,
5947 deps = MICROKERNEL_TEST_DEPS,
5948)
5949
5950xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005951 name = "f32_vaddc_test",
5952 srcs = [
5953 "test/f32-vaddc.cc",
5954 "test/vbinaryc-microkernel-tester.h",
5955 ] + MICROKERNEL_TEST_HDRS,
5956 deps = MICROKERNEL_TEST_DEPS,
5957)
5958
5959xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005960 name = "f32_vaddc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005961 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005962 "test/f32-vaddc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005963 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005964 ] + MICROKERNEL_TEST_HDRS,
5965 deps = MICROKERNEL_TEST_DEPS,
5966)
5967
5968xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005969 name = "f32_vaddc_relu_test",
5970 srcs = [
5971 "test/f32-vaddc-relu.cc",
5972 "test/vbinaryc-microkernel-tester.h",
5973 ] + MICROKERNEL_TEST_HDRS,
5974 deps = MICROKERNEL_TEST_DEPS,
5975)
5976
5977xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005978 name = "f32_vdiv_test",
5979 srcs = [
5980 "test/f32-vdiv.cc",
5981 "test/vbinary-microkernel-tester.h",
5982 ] + MICROKERNEL_TEST_HDRS,
5983 deps = MICROKERNEL_TEST_DEPS,
5984)
5985
5986xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005987 name = "f32_vdiv_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08005988 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005989 "test/f32-vdiv-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08005990 "test/vbinary-microkernel-tester.h",
5991 ] + MICROKERNEL_TEST_HDRS,
5992 deps = MICROKERNEL_TEST_DEPS,
5993)
5994
5995xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005996 name = "f32_vdiv_relu_test",
5997 srcs = [
5998 "test/f32-vdiv-relu.cc",
5999 "test/vbinary-microkernel-tester.h",
6000 ] + MICROKERNEL_TEST_HDRS,
6001 deps = MICROKERNEL_TEST_DEPS,
6002)
6003
6004xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07006005 name = "f32_vdivc_test",
6006 srcs = [
6007 "test/f32-vdivc.cc",
6008 "test/vbinaryc-microkernel-tester.h",
6009 ] + MICROKERNEL_TEST_HDRS,
6010 deps = MICROKERNEL_TEST_DEPS,
6011)
6012
6013xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006014 name = "f32_vdivc_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08006015 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006016 "test/f32-vdivc-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08006017 "test/vbinaryc-microkernel-tester.h",
6018 ] + MICROKERNEL_TEST_HDRS,
6019 deps = MICROKERNEL_TEST_DEPS,
6020)
6021
6022xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07006023 name = "f32_vdivc_relu_test",
6024 srcs = [
6025 "test/f32-vdivc-relu.cc",
6026 "test/vbinaryc-microkernel-tester.h",
6027 ] + MICROKERNEL_TEST_HDRS,
6028 deps = MICROKERNEL_TEST_DEPS,
6029)
6030
6031xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07006032 name = "f32_vrdivc_test",
6033 srcs = [
6034 "test/f32-vrdivc.cc",
6035 "test/vbinaryc-microkernel-tester.h",
6036 ] + MICROKERNEL_TEST_HDRS,
6037 deps = MICROKERNEL_TEST_DEPS,
6038)
6039
6040xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006041 name = "f32_vrdivc_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08006042 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006043 "test/f32-vrdivc-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08006044 "test/vbinaryc-microkernel-tester.h",
6045 ] + MICROKERNEL_TEST_HDRS,
6046 deps = MICROKERNEL_TEST_DEPS,
6047)
6048
6049xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07006050 name = "f32_vrdivc_relu_test",
6051 srcs = [
6052 "test/f32-vrdivc-relu.cc",
6053 "test/vbinaryc-microkernel-tester.h",
6054 ] + MICROKERNEL_TEST_HDRS,
6055 deps = MICROKERNEL_TEST_DEPS,
6056)
6057
6058xnnpack_unit_test(
Marat Dukhaned6baaf2020-12-01 15:07:08 -08006059 name = "f32_velu_test",
6060 srcs = [
6061 "test/f32-velu.cc",
6062 "test/vunary-microkernel-tester.h",
6063 ] + MICROKERNEL_TEST_HDRS,
6064 deps = MICROKERNEL_TEST_DEPS,
6065)
6066
6067xnnpack_unit_test(
Marat Dukhan403b7d42019-12-05 12:49:11 -08006068 name = "f32_vmax_test",
6069 srcs = [
6070 "test/f32-vmax.cc",
6071 "test/vbinary-microkernel-tester.h",
6072 ] + MICROKERNEL_TEST_HDRS,
6073 deps = MICROKERNEL_TEST_DEPS,
6074)
6075
6076xnnpack_unit_test(
6077 name = "f32_vmaxc_test",
6078 srcs = [
6079 "test/f32-vmaxc.cc",
6080 "test/vbinaryc-microkernel-tester.h",
6081 ] + MICROKERNEL_TEST_HDRS,
6082 deps = MICROKERNEL_TEST_DEPS,
6083)
6084
6085xnnpack_unit_test(
6086 name = "f32_vmin_test",
6087 srcs = [
6088 "test/f32-vmin.cc",
6089 "test/vbinary-microkernel-tester.h",
6090 ] + MICROKERNEL_TEST_HDRS,
6091 deps = MICROKERNEL_TEST_DEPS,
6092)
6093
6094xnnpack_unit_test(
6095 name = "f32_vminc_test",
6096 srcs = [
6097 "test/f32-vminc.cc",
6098 "test/vbinaryc-microkernel-tester.h",
6099 ] + MICROKERNEL_TEST_HDRS,
6100 deps = MICROKERNEL_TEST_DEPS,
6101)
6102
6103xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07006104 name = "f32_vmul_test",
6105 srcs = [
6106 "test/f32-vmul.cc",
6107 "test/vbinary-microkernel-tester.h",
6108 ] + MICROKERNEL_TEST_HDRS,
6109 deps = MICROKERNEL_TEST_DEPS,
6110)
6111
6112xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006113 name = "f32_vmul_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006114 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006115 "test/f32-vmul-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08006116 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08006117 ] + MICROKERNEL_TEST_HDRS,
6118 deps = MICROKERNEL_TEST_DEPS,
6119)
6120
6121xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07006122 name = "f32_vmul_relu_test",
6123 srcs = [
6124 "test/f32-vmul-relu.cc",
6125 "test/vbinary-microkernel-tester.h",
6126 ] + MICROKERNEL_TEST_HDRS,
6127 deps = MICROKERNEL_TEST_DEPS,
6128)
6129
6130xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07006131 name = "f32_vmulc_test",
6132 srcs = [
6133 "test/f32-vmulc.cc",
6134 "test/vbinaryc-microkernel-tester.h",
6135 ] + MICROKERNEL_TEST_HDRS,
6136 deps = MICROKERNEL_TEST_DEPS,
6137)
6138
6139xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006140 name = "f32_vmulc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08006141 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006142 "test/f32-vmulc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08006143 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006144 ] + MICROKERNEL_TEST_HDRS,
6145 deps = MICROKERNEL_TEST_DEPS,
6146)
6147
6148xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07006149 name = "f32_vmulc_relu_test",
6150 srcs = [
6151 "test/f32-vmulc-relu.cc",
6152 "test/vbinaryc-microkernel-tester.h",
6153 ] + MICROKERNEL_TEST_HDRS,
6154 deps = MICROKERNEL_TEST_DEPS,
6155)
6156
6157xnnpack_unit_test(
Frank Barchard2a1049c2020-06-03 02:31:27 -07006158 name = "f16_vmulcaddc_minmax_test",
6159 srcs = [
6160 "test/f16-vmulcaddc-minmax.cc",
6161 "test/vmulcaddc-microkernel-tester.h",
6162 "src/xnnpack/AlignedAllocator.h",
6163 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07006164 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Frank Barchard2a1049c2020-06-03 02:31:27 -07006165)
6166
6167xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07006168 name = "f32_vmulcaddc_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006169 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07006170 "test/f32-vmulcaddc-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006171 "test/vmulcaddc-microkernel-tester.h",
6172 "src/xnnpack/AlignedAllocator.h",
6173 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07006174 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07006175)
6176
6177xnnpack_unit_test(
Marat Dukhan8cc7efe2020-06-10 16:24:27 -07006178 name = "f32_vlrelu_test",
6179 srcs = [
6180 "test/f32-vlrelu.cc",
6181 "test/vunary-microkernel-tester.h",
6182 ] + MICROKERNEL_TEST_HDRS,
6183 deps = MICROKERNEL_TEST_DEPS,
6184)
6185
6186xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07006187 name = "f32_vneg_test",
6188 srcs = [
6189 "test/f32-vneg.cc",
6190 "test/vunary-microkernel-tester.h",
6191 ] + MICROKERNEL_TEST_HDRS,
6192 deps = MICROKERNEL_TEST_DEPS,
6193)
6194
6195xnnpack_unit_test(
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07006196 name = "f32_vrndne_test",
6197 srcs = [
6198 "test/f32-vrndne.cc",
6199 "test/vunary-microkernel-tester.h",
6200 ] + MICROKERNEL_TEST_HDRS,
6201 deps = MICROKERNEL_TEST_DEPS,
6202)
6203
6204xnnpack_unit_test(
6205 name = "f32_vrndz_test",
6206 srcs = [
6207 "test/f32-vrndz.cc",
6208 "test/vunary-microkernel-tester.h",
6209 ] + MICROKERNEL_TEST_HDRS,
6210 deps = MICROKERNEL_TEST_DEPS,
6211)
6212
6213xnnpack_unit_test(
6214 name = "f32_vrndu_test",
6215 srcs = [
6216 "test/f32-vrndu.cc",
6217 "test/vunary-microkernel-tester.h",
6218 ] + MICROKERNEL_TEST_HDRS,
6219 deps = MICROKERNEL_TEST_DEPS,
6220)
6221
6222xnnpack_unit_test(
6223 name = "f32_vrndd_test",
6224 srcs = [
6225 "test/f32-vrndd.cc",
6226 "test/vunary-microkernel-tester.h",
6227 ] + MICROKERNEL_TEST_HDRS,
6228 deps = MICROKERNEL_TEST_DEPS,
6229)
6230
6231xnnpack_unit_test(
Marat Dukhan05ac8e32019-10-21 15:39:33 -07006232 name = "f32_vscale_test",
6233 srcs = [
6234 "test/f32-vscale.cc",
6235 "test/vscale-microkernel-tester.h",
6236 ] + MICROKERNEL_TEST_HDRS,
6237 deps = MICROKERNEL_TEST_DEPS,
6238)
6239
6240xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07006241 name = "f32_vscaleexpminusmax_test",
6242 srcs = [
6243 "test/f32-vscaleexpminusmax.cc",
6244 "test/vscaleexpminusmax-microkernel-tester.h",
6245 ] + MICROKERNEL_TEST_HDRS,
6246 deps = MICROKERNEL_TEST_DEPS,
6247)
6248
6249xnnpack_unit_test(
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07006250 name = "f32_vscaleextexp_test",
6251 srcs = [
6252 "test/f32-vscaleextexp.cc",
6253 "test/vscaleextexp-microkernel-tester.h",
6254 ] + MICROKERNEL_TEST_HDRS,
6255 deps = MICROKERNEL_TEST_DEPS,
6256)
6257
6258xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07006259 name = "f32_vsqr_test",
6260 srcs = [
6261 "test/f32-vsqr.cc",
6262 "test/vunary-microkernel-tester.h",
6263 ] + MICROKERNEL_TEST_HDRS,
6264 deps = MICROKERNEL_TEST_DEPS,
6265)
6266
6267xnnpack_unit_test(
Marat Dukhan13bafb02020-06-05 00:43:11 -07006268 name = "f32_vsqrdiff_test",
6269 srcs = [
6270 "test/f32-vsqrdiff.cc",
6271 "test/vbinary-microkernel-tester.h",
6272 ] + MICROKERNEL_TEST_HDRS,
6273 deps = MICROKERNEL_TEST_DEPS,
6274)
6275
6276xnnpack_unit_test(
6277 name = "f32_vsqrdiffc_test",
6278 srcs = [
6279 "test/f32-vsqrdiffc.cc",
6280 "test/vbinaryc-microkernel-tester.h",
6281 ] + MICROKERNEL_TEST_HDRS,
6282 deps = MICROKERNEL_TEST_DEPS,
6283)
6284
6285xnnpack_unit_test(
Marat Dukhanf4db2f32020-06-30 10:55:30 -07006286 name = "f32_vsqrt_test",
6287 srcs = [
6288 "test/f32-vsqrt.cc",
6289 "test/vunary-microkernel-tester.h",
6290 ] + MICROKERNEL_TEST_HDRS,
6291 deps = MICROKERNEL_TEST_DEPS,
6292)
6293
6294xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07006295 name = "f32_vsub_test",
6296 srcs = [
6297 "test/f32-vsub.cc",
6298 "test/vbinary-microkernel-tester.h",
6299 ] + MICROKERNEL_TEST_HDRS,
6300 deps = MICROKERNEL_TEST_DEPS,
6301)
6302
6303xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006304 name = "f32_vsub_minmax_test",
Marat Dukhan97579532019-10-18 16:40:39 -07006305 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006306 "test/f32-vsub-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08006307 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08006308 ] + MICROKERNEL_TEST_HDRS,
6309 deps = MICROKERNEL_TEST_DEPS,
6310)
6311
6312xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07006313 name = "f32_vsub_relu_test",
6314 srcs = [
6315 "test/f32-vsub-relu.cc",
6316 "test/vbinary-microkernel-tester.h",
6317 ] + MICROKERNEL_TEST_HDRS,
6318 deps = MICROKERNEL_TEST_DEPS,
6319)
6320
6321xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07006322 name = "f32_vsubc_test",
6323 srcs = [
6324 "test/f32-vsubc.cc",
6325 "test/vbinaryc-microkernel-tester.h",
6326 ] + MICROKERNEL_TEST_HDRS,
6327 deps = MICROKERNEL_TEST_DEPS,
6328)
6329
6330xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006331 name = "f32_vsubc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08006332 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006333 "test/f32-vsubc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08006334 "test/vbinaryc-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08006335 ] + MICROKERNEL_TEST_HDRS,
6336 deps = MICROKERNEL_TEST_DEPS,
6337)
6338
6339xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07006340 name = "f32_vsubc_relu_test",
6341 srcs = [
6342 "test/f32-vsubc-relu.cc",
6343 "test/vbinaryc-microkernel-tester.h",
6344 ] + MICROKERNEL_TEST_HDRS,
6345 deps = MICROKERNEL_TEST_DEPS,
6346)
6347
6348xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07006349 name = "f32_vrsubc_test",
6350 srcs = [
6351 "test/f32-vrsubc.cc",
6352 "test/vbinaryc-microkernel-tester.h",
6353 ] + MICROKERNEL_TEST_HDRS,
6354 deps = MICROKERNEL_TEST_DEPS,
6355)
6356
6357xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006358 name = "f32_vrsubc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08006359 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006360 "test/f32-vrsubc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08006361 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan97579532019-10-18 16:40:39 -07006362 ] + MICROKERNEL_TEST_HDRS,
6363 deps = MICROKERNEL_TEST_DEPS,
6364)
6365
6366xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07006367 name = "f32_vrsubc_relu_test",
6368 srcs = [
6369 "test/f32-vrsubc-relu.cc",
6370 "test/vbinaryc-microkernel-tester.h",
6371 ] + MICROKERNEL_TEST_HDRS,
6372 deps = MICROKERNEL_TEST_DEPS,
6373)
6374
6375xnnpack_unit_test(
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07006376 name = "qs8_dwconv_minmax_test",
6377 srcs = [
6378 "test/qs8-dwconv-minmax.cc",
6379 "test/dwconv-microkernel-tester.h",
6380 "src/xnnpack/AlignedAllocator.h",
6381 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
6382 deps = MICROKERNEL_TEST_DEPS + [":packing"],
6383)
6384
6385xnnpack_unit_test(
Marat Dukhan4ed53f42020-08-06 01:12:55 -07006386 name = "qs8_gavgpool_minmax_test",
6387 srcs = [
6388 "test/qs8-gavgpool-minmax.cc",
6389 "test/gavgpool-microkernel-tester.h",
6390 "src/xnnpack/AlignedAllocator.h",
6391 ] + MICROKERNEL_TEST_HDRS,
6392 deps = MICROKERNEL_TEST_DEPS,
6393)
6394
6395xnnpack_unit_test(
Marat Dukhan595e1702020-07-31 10:12:52 -07006396 name = "qs8_gemm_minmax_test",
6397 srcs = [
6398 "test/qs8-gemm-minmax.cc",
6399 "test/gemm-microkernel-tester.h",
6400 "src/xnnpack/AlignedAllocator.h",
6401 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
6402 deps = MICROKERNEL_TEST_DEPS + [":packing"],
6403)
6404
6405xnnpack_unit_test(
Marat Dukhanf9480682020-07-31 14:50:24 -07006406 name = "qs8_igemm_minmax_test",
6407 srcs = [
6408 "test/qs8-igemm-minmax.cc",
6409 "test/gemm-microkernel-tester.h",
6410 "src/xnnpack/AlignedAllocator.h",
6411 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
6412 deps = MICROKERNEL_TEST_DEPS + [":packing"],
6413)
6414
6415xnnpack_unit_test(
6416 name = "qs8_requantization_test",
6417 srcs = [
6418 "src/xnnpack/requantization-stubs.h",
6419 "test/qs8-requantization.cc",
6420 "test/requantization-tester.h",
6421 ] + MICROKERNEL_TEST_HDRS,
6422 deps = MICROKERNEL_TEST_DEPS,
6423)
6424
6425xnnpack_unit_test(
Marat Dukhand9f3ad42020-08-10 12:30:58 -07006426 name = "qs8_vadd_minmax_test",
6427 srcs = [
6428 "test/qs8-vadd-minmax.cc",
6429 "test/vadd-microkernel-tester.h",
6430 ] + MICROKERNEL_TEST_HDRS,
6431 deps = MICROKERNEL_TEST_DEPS,
6432)
6433
6434xnnpack_unit_test(
Marat Dukhan0270d9f2020-08-11 00:56:46 -07006435 name = "qs8_vaddc_minmax_test",
6436 srcs = [
6437 "test/qs8-vaddc-minmax.cc",
6438 "test/vaddc-microkernel-tester.h",
6439 ] + MICROKERNEL_TEST_HDRS,
6440 deps = MICROKERNEL_TEST_DEPS,
6441)
6442
6443xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07006444 name = "qu8_avgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006445 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07006446 "test/qu8-avgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006447 "test/avgpool-microkernel-tester.h",
6448 "src/xnnpack/AlignedAllocator.h",
6449 ] + MICROKERNEL_TEST_HDRS,
6450 deps = MICROKERNEL_TEST_DEPS,
6451)
6452
6453xnnpack_unit_test(
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07006454 name = "qu8_dwconv_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006455 srcs = [
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07006456 "test/qu8-dwconv-minmax.cc",
6457 "test/dwconv-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006458 "src/xnnpack/AlignedAllocator.h",
6459 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07006460 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07006461)
6462
6463xnnpack_unit_test(
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07006464 name = "qu8_igemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006465 srcs = [
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07006466 "test/qu8-igemm-minmax.cc",
6467 "test/gemm-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006468 "src/xnnpack/AlignedAllocator.h",
6469 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07006470 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07006471)
6472
6473xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07006474 name = "qu8_gavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006475 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07006476 "test/qu8-gavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006477 "test/gavgpool-microkernel-tester.h",
6478 "src/xnnpack/AlignedAllocator.h",
6479 ] + MICROKERNEL_TEST_HDRS,
6480 deps = MICROKERNEL_TEST_DEPS,
6481)
6482
6483xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07006484 name = "qu8_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006485 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07006486 "test/qu8-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006487 "test/gemm-microkernel-tester.h",
6488 "src/xnnpack/AlignedAllocator.h",
6489 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07006490 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07006491)
6492
6493xnnpack_unit_test(
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07006494 name = "qu8_requantization_test",
6495 srcs = [
6496 "src/xnnpack/requantization-stubs.h",
6497 "test/qu8-requantization.cc",
6498 "test/requantization-tester.h",
6499 ] + MICROKERNEL_TEST_HDRS,
6500 deps = MICROKERNEL_TEST_DEPS,
6501)
6502
6503xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07006504 name = "qu8_vadd_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006505 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07006506 "test/qu8-vadd-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006507 "test/vadd-microkernel-tester.h",
6508 ] + MICROKERNEL_TEST_HDRS,
6509 deps = MICROKERNEL_TEST_DEPS,
6510)
6511
6512xnnpack_unit_test(
6513 name = "u8_clamp_test",
6514 srcs = [
6515 "test/u8-clamp.cc",
6516 "test/clamp-microkernel-tester.h",
6517 ] + MICROKERNEL_TEST_HDRS,
6518 deps = MICROKERNEL_TEST_DEPS,
6519)
6520
6521xnnpack_unit_test(
6522 name = "u8_lut32norm_test",
6523 srcs = [
6524 "test/u8-lut32norm.cc",
6525 "test/lut-norm-microkernel-tester.h",
6526 ] + MICROKERNEL_TEST_HDRS,
6527 deps = MICROKERNEL_TEST_DEPS,
6528)
6529
6530xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07006531 name = "u8_maxpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006532 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07006533 "test/u8-maxpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006534 "test/maxpool-microkernel-tester.h",
6535 ] + MICROKERNEL_TEST_HDRS,
6536 deps = MICROKERNEL_TEST_DEPS,
6537)
6538
6539xnnpack_unit_test(
6540 name = "u8_rmax_test",
6541 srcs = [
6542 "test/u8-rmax.cc",
6543 "test/rmax-microkernel-tester.h",
6544 ] + MICROKERNEL_TEST_HDRS,
6545 deps = MICROKERNEL_TEST_DEPS,
6546)
6547
6548xnnpack_unit_test(
Marat Dukhanad71b9a2020-11-20 00:01:51 -08006549 name = "x32_depthtospace2d_chw2hwc_test",
Yury Kartynnike7841862020-11-04 18:22:18 -08006550 srcs = [
Marat Dukhanad71b9a2020-11-20 00:01:51 -08006551 "test/x32-depthtospace2d-chw2hwc.cc",
6552 "test/depthtospace-microkernel-tester.h",
Yury Kartynnike7841862020-11-04 18:22:18 -08006553 ] + MICROKERNEL_TEST_HDRS,
6554 deps = MICROKERNEL_TEST_DEPS,
6555)
6556
6557xnnpack_unit_test(
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -07006558 name = "x32_fill_test",
6559 srcs = [
6560 "test/x32-fill.cc",
6561 "test/fill-microkernel-tester.h",
6562 ] + MICROKERNEL_TEST_HDRS,
6563 deps = MICROKERNEL_TEST_DEPS,
6564)
6565
6566xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07006567 name = "x32_packx_test",
6568 srcs = [
6569 "test/x32-packx.cc",
6570 "test/pack-microkernel-tester.h",
6571 "src/xnnpack/AlignedAllocator.h",
6572 ] + MICROKERNEL_TEST_HDRS,
6573 deps = MICROKERNEL_TEST_DEPS,
6574)
6575
6576xnnpack_unit_test(
6577 name = "x32_pad_test",
6578 srcs = [
6579 "test/x32-pad.cc",
6580 "test/pad-microkernel-tester.h",
6581 ] + MICROKERNEL_TEST_HDRS,
6582 deps = MICROKERNEL_TEST_DEPS,
6583)
6584
6585xnnpack_unit_test(
6586 name = "x32_unpool_test",
6587 srcs = [
6588 "test/x32-unpool.cc",
6589 "test/unpool-microkernel-tester.h",
6590 ] + MICROKERNEL_TEST_HDRS,
6591 deps = MICROKERNEL_TEST_DEPS,
6592)
6593
6594xnnpack_unit_test(
6595 name = "x32_zip_test",
6596 srcs = [
6597 "test/x32-zip.cc",
6598 "test/zip-microkernel-tester.h",
6599 ] + MICROKERNEL_TEST_HDRS,
6600 deps = MICROKERNEL_TEST_DEPS,
6601)
6602
6603xnnpack_unit_test(
6604 name = "x8_lut_test",
6605 srcs = [
6606 "test/x8-lut.cc",
6607 "test/lut-microkernel-tester.h",
6608 ] + MICROKERNEL_TEST_HDRS,
6609 deps = MICROKERNEL_TEST_DEPS,
6610)
6611
6612xnnpack_unit_test(
6613 name = "x8_zip_test",
6614 srcs = [
6615 "test/x8-zip.cc",
6616 "test/zip-microkernel-tester.h",
6617 ] + MICROKERNEL_TEST_HDRS,
6618 deps = MICROKERNEL_TEST_DEPS,
6619)
6620
Marat Dukhan20c3b922020-03-10 03:45:06 -07006621########################## Size tests for the library #########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07006622
6623xnnpack_binary(
Marat Dukhan20c3b922020-03-10 03:45:06 -07006624 name = "operator_size_test",
6625 srcs = ["test/operator-size.c"],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07006626 deps = [":xnnpack_operators_nhwc_f32"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07006627)
6628
Marat Dukhan20c3b922020-03-10 03:45:06 -07006629xnnpack_binary(
6630 name = "subgraph_size_test",
6631 srcs = ["test/subgraph-size.c"],
6632 deps = [":XNNPACK"],
6633)
6634
6635########################### Unit tests for operators ##########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07006636
6637xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07006638 name = "abs_nc_test",
6639 srcs = [
6640 "test/abs-nc.cc",
6641 "test/abs-operator-tester.h",
6642 ],
6643 deps = OPERATOR_TEST_DEPS,
6644)
6645
6646xnnpack_unit_test(
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08006647 name = "add_nd_test",
6648 srcs = [
6649 "test/add-nd.cc",
6650 "test/binary-elementwise-operator-tester.h",
6651 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006652 deps = OPERATOR_TEST_DEPS,
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08006653)
6654
6655xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006656 name = "argmax_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006657 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006658 "test/argmax-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006659 "test/argmax-pooling-operator-tester.h",
6660 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07006661 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006662)
6663
6664xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006665 name = "average_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006666 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006667 "test/average-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006668 "test/average-pooling-operator-tester.h",
6669 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07006670 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006671)
6672
6673xnnpack_unit_test(
Marat Dukhan64e52512020-06-09 13:41:16 -07006674 name = "bankers_rounding_nc_test",
6675 srcs = [
6676 "test/bankers-rounding-nc.cc",
6677 "test/bankers-rounding-operator-tester.h",
6678 ],
6679 deps = OPERATOR_TEST_DEPS,
6680)
6681
6682xnnpack_unit_test(
6683 name = "ceiling_nc_test",
6684 srcs = [
6685 "test/ceiling-nc.cc",
6686 "test/ceiling-operator-tester.h",
6687 ],
6688 deps = OPERATOR_TEST_DEPS,
6689)
6690
6691xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006692 name = "channel_shuffle_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006693 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006694 "test/channel-shuffle-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006695 "test/channel-shuffle-operator-tester.h",
6696 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006697 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006698)
6699
6700xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006701 name = "clamp_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006702 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006703 "test/clamp-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006704 "test/clamp-operator-tester.h",
6705 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006706 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006707)
6708
6709xnnpack_unit_test(
Marat Dukhan065b11e2020-05-22 09:49:41 -07006710 name = "constant_pad_nd_test",
6711 srcs = [
6712 "test/constant-pad-nd.cc",
6713 "test/constant-pad-operator-tester.h",
6714 ],
6715 deps = OPERATOR_TEST_DEPS,
6716)
6717
6718xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006719 name = "convolution_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006720 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006721 "test/convolution-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006722 "test/convolution-operator-tester.h",
6723 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006724 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006725)
6726
6727xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006728 name = "convolution_nchw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006729 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006730 "test/convolution-nchw.cc",
6731 "test/convolution-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006732 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006733 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006734)
6735
6736xnnpack_unit_test(
Marat Dukhan4e21b272020-06-04 18:45:01 -07006737 name = "copy_nc_test",
6738 srcs = [
6739 "test/copy-nc.cc",
6740 "test/copy-operator-tester.h",
6741 ],
6742 deps = OPERATOR_TEST_DEPS,
6743)
6744
6745xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006746 name = "deconvolution_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006747 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006748 "test/deconvolution-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006749 "test/deconvolution-operator-tester.h",
6750 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07006751 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006752)
6753
6754xnnpack_unit_test(
Marat Dukhan188d1042020-11-24 23:39:40 -08006755 name = "depth_to_space_nchw2nhwc_test",
Artsiom Ablavatski0f1dc182020-11-05 19:21:50 -08006756 srcs = [
Marat Dukhan188d1042020-11-24 23:39:40 -08006757 "test/depth-to-space-nchw2nhwc.cc",
Artsiom Ablavatski0f1dc182020-11-05 19:21:50 -08006758 "test/depth-to-space-operator-tester.h",
6759 ] + OPERATOR_TEST_PARAMS_HDRS,
6760 deps = OPERATOR_TEST_DEPS,
6761)
6762
6763xnnpack_unit_test(
Marat Dukhan0e521172020-11-25 13:10:04 -08006764 name = "depth_to_space_nhwc_test",
6765 srcs = [
6766 "test/depth-to-space-nhwc.cc",
6767 "test/depth-to-space-operator-tester.h",
6768 ] + OPERATOR_TEST_PARAMS_HDRS,
6769 deps = OPERATOR_TEST_DEPS,
6770)
6771
6772xnnpack_unit_test(
Marat Dukhan69180502019-12-06 15:00:31 -08006773 name = "divide_nd_test",
6774 srcs = [
6775 "test/binary-elementwise-operator-tester.h",
6776 "test/divide-nd.cc",
6777 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006778 deps = OPERATOR_TEST_DEPS,
Marat Dukhan69180502019-12-06 15:00:31 -08006779)
6780
6781xnnpack_unit_test(
Marat Dukhanb6bd4bc2020-12-01 17:01:40 -08006782 name = "elu_nc_test",
6783 srcs = [
6784 "test/elu-nc.cc",
6785 "test/elu-operator-tester.h",
6786 ],
6787 deps = OPERATOR_TEST_DEPS,
6788)
6789
6790xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006791 name = "fully_connected_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006792 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006793 "test/fully-connected-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006794 "test/fully-connected-operator-tester.h",
6795 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006796 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006797)
6798
6799xnnpack_unit_test(
Marat Dukhan64e52512020-06-09 13:41:16 -07006800 name = "floor_nc_test",
6801 srcs = [
6802 "test/floor-nc.cc",
6803 "test/floor-operator-tester.h",
6804 ],
6805 deps = OPERATOR_TEST_DEPS,
6806)
6807
6808xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006809 name = "global_average_pooling_nwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006810 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006811 "test/global-average-pooling-nwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006812 "test/global-average-pooling-operator-tester.h",
Marat Dukhanef61d022020-06-19 13:54:49 -07006813 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07006814 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006815)
6816
6817xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006818 name = "global_average_pooling_ncw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006819 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006820 "test/global-average-pooling-ncw.cc",
6821 "test/global-average-pooling-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006822 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006823 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006824)
6825
6826xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006827 name = "hardswish_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006828 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006829 "test/hardswish-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006830 "test/hardswish-operator-tester.h",
6831 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006832 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006833)
6834
6835xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006836 name = "leaky_relu_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006837 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006838 "test/leaky-relu-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006839 "test/leaky-relu-operator-tester.h",
6840 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006841 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006842)
6843
6844xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006845 name = "max_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006846 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006847 "test/max-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006848 "test/max-pooling-operator-tester.h",
6849 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07006850 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006851)
6852
6853xnnpack_unit_test(
Marat Dukhan79e7f842019-12-05 14:35:50 -08006854 name = "maximum_nd_test",
6855 srcs = [
6856 "test/binary-elementwise-operator-tester.h",
6857 "test/maximum-nd.cc",
6858 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006859 deps = OPERATOR_TEST_DEPS,
Marat Dukhan79e7f842019-12-05 14:35:50 -08006860)
6861
6862xnnpack_unit_test(
6863 name = "minimum_nd_test",
6864 srcs = [
6865 "test/binary-elementwise-operator-tester.h",
6866 "test/minimum-nd.cc",
6867 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006868 deps = OPERATOR_TEST_DEPS,
Marat Dukhan79e7f842019-12-05 14:35:50 -08006869)
6870
6871xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006872 name = "multiply_nd_test",
Marat Dukhanca2733c2019-11-15 23:21:17 -08006873 srcs = [
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08006874 "test/binary-elementwise-operator-tester.h",
Marat Dukhanefc47b82019-11-18 09:25:38 -08006875 "test/multiply-nd.cc",
Marat Dukhanca2733c2019-11-15 23:21:17 -08006876 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006877 deps = OPERATOR_TEST_DEPS,
Marat Dukhanca2733c2019-11-15 23:21:17 -08006878)
6879
6880xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07006881 name = "negate_nc_test",
6882 srcs = [
6883 "test/negate-nc.cc",
6884 "test/negate-operator-tester.h",
6885 ],
6886 deps = OPERATOR_TEST_DEPS,
6887)
6888
6889xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006890 name = "prelu_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006891 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006892 "test/prelu-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006893 "test/prelu-operator-tester.h",
6894 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07006895 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006896)
6897
6898xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006899 name = "resize_bilinear_nhwc_test",
Marat Dukhan69722492019-11-11 19:55:50 -08006900 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006901 "test/resize-bilinear-nhwc.cc",
Marat Dukhan69722492019-11-11 19:55:50 -08006902 "test/resize-bilinear-operator-tester.h",
6903 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07006904 deps = OPERATOR_TEST_DEPS,
Marat Dukhan69722492019-11-11 19:55:50 -08006905)
6906
6907xnnpack_unit_test(
Artsiom Ablavatski97918102020-10-27 15:52:59 -07006908 name = "resize_bilinear_nchw_test",
6909 srcs = [
6910 "test/resize-bilinear-nchw.cc",
6911 "test/resize-bilinear-operator-tester.h",
6912 ] + OPERATOR_TEST_PARAMS_HDRS,
6913 deps = OPERATOR_TEST_DEPS,
6914)
6915
6916xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006917 name = "sigmoid_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006918 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006919 "test/sigmoid-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006920 "test/sigmoid-operator-tester.h",
6921 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006922 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006923)
6924
6925xnnpack_unit_test(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08006926 name = "softmax_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006927 srcs = [
Marat Dukhanfd8e6892020-01-27 15:25:25 -08006928 "test/softmax-nc.cc",
6929 "test/softmax-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006930 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006931 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006932)
6933
6934xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07006935 name = "square_nc_test",
6936 srcs = [
6937 "test/square-nc.cc",
6938 "test/square-operator-tester.h",
6939 ],
6940 deps = OPERATOR_TEST_DEPS,
6941)
6942
6943xnnpack_unit_test(
Marat Dukhan6804bbd2020-06-30 19:26:11 -07006944 name = "square_root_nc_test",
6945 srcs = [
6946 "test/square-root-nc.cc",
6947 "test/square-root-operator-tester.h",
6948 ],
6949 deps = OPERATOR_TEST_DEPS,
6950)
6951
6952xnnpack_unit_test(
Marat Dukhanf7399262020-06-05 10:58:44 -07006953 name = "squared_difference_nd_test",
6954 srcs = [
6955 "test/binary-elementwise-operator-tester.h",
6956 "test/squared-difference-nd.cc",
6957 ],
6958 deps = OPERATOR_TEST_DEPS,
6959)
6960
6961xnnpack_unit_test(
Marat Dukhan05f3f6d2019-12-03 15:13:53 -08006962 name = "subtract_nd_test",
6963 srcs = [
6964 "test/binary-elementwise-operator-tester.h",
6965 "test/subtract-nd.cc",
6966 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006967 deps = OPERATOR_TEST_DEPS,
Marat Dukhan05f3f6d2019-12-03 15:13:53 -08006968)
6969
6970xnnpack_unit_test(
Marat Dukhan64e52512020-06-09 13:41:16 -07006971 name = "truncation_nc_test",
6972 srcs = [
6973 "test/truncation-nc.cc",
6974 "test/truncation-operator-tester.h",
6975 ],
6976 deps = OPERATOR_TEST_DEPS,
6977)
6978
6979xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006980 name = "unpooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006981 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006982 "test/unpooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006983 "test/unpooling-operator-tester.h",
6984 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006985 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006986)
6987
Chao Mei6ddfc602020-05-13 22:29:36 -07006988############################### Misc unit tests ###############################
6989
6990xnnpack_unit_test(
6991 name = "memory_planner_test",
6992 srcs = [
6993 "test/memory-planner-test.cc",
6994 ],
6995 deps = [
6996 ":XNNPACK",
6997 ":memory_planner",
6998 ],
6999)
7000
XNNPACK Teamab8c4c82020-10-09 08:05:51 -07007001xnnpack_unit_test(
7002 name = "subgraph_nchw_test",
7003 srcs = [
7004 "src/xnnpack/subgraph.h",
7005 "test/subgraph-nchw.cc",
7006 "test/subgraph-tester.h",
7007 ],
7008 deps = [
7009 ":XNNPACK",
7010 ],
7011)
7012
Marat Dukhan08c4a432019-10-03 09:29:21 -07007013############################# Build configurations #############################
7014
Marat Dukhanb8642352019-10-30 15:43:02 -07007015# Enables usage of assembly kernels.
Marat Dukhan08c4a432019-10-03 09:29:21 -07007016config_setting(
Marat Dukhanb8642352019-10-30 15:43:02 -07007017 name = "xnn_enable_assembly_explicit_true",
7018 define_values = {"xnn_enable_assembly": "true"},
7019)
7020
7021# Disables usage of assembly kernels.
7022config_setting(
7023 name = "xnn_enable_assembly_explicit_false",
7024 define_values = {"xnn_enable_assembly": "false"},
7025)
7026
Marat Dukhan9de90e02020-06-18 16:04:12 -07007027# Enables usage of sparse inference.
7028config_setting(
7029 name = "xnn_enable_sparse_explicit_true",
7030 define_values = {"xnn_enable_sparse": "true"},
7031)
7032
7033# Disables usage of sparse inference.
7034config_setting(
7035 name = "xnn_enable_sparse_explicit_false",
7036 define_values = {"xnn_enable_sparse": "false"},
7037)
7038
Marat Dukhan05702cf2020-03-26 15:41:33 -07007039# Disables usage of HMP-aware optimizations.
7040config_setting(
7041 name = "xnn_enable_hmp_explicit_false",
7042 define_values = {"xnn_enable_hmp": "false"},
7043)
7044
Chao Mei6ddfc602020-05-13 22:29:36 -07007045# Enable usage of optimized memory allocation
7046config_setting(
7047 name = "xnn_enable_memopt_explicit_true",
7048 define_values = {"xnnpack_enable_memopt": "true"},
7049)
7050
7051# Disable usage of optimized memory allocation
7052config_setting(
7053 name = "xnn_enable_memopt_explicit_false",
7054 define_values = {"xnnpack_enable_memopt": "false"},
7055)
7056
Marat Dukhanb8642352019-10-30 15:43:02 -07007057# Builds with -c dbg
7058config_setting(
7059 name = "debug_build",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007060 values = {
Marat Dukhanb8642352019-10-30 15:43:02 -07007061 "compilation_mode": "dbg",
7062 },
7063)
7064
7065# Builds with -c opt
7066config_setting(
7067 name = "optimized_build",
7068 values = {
7069 "compilation_mode": "opt",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007070 },
7071)
7072
7073config_setting(
Marat Dukhanb8642352019-10-30 15:43:02 -07007074 name = "linux_k8",
7075 values = {"cpu": "k8"},
7076)
7077
7078config_setting(
Marat Dukhan582094e2020-04-30 17:21:25 -07007079 name = "linux_arm",
7080 values = {"cpu": "arm"},
Marat Dukhan4e45e662019-10-03 15:40:24 -07007081)
7082
7083config_setting(
Marat Dukhanf0bd4de2020-06-15 15:53:19 -07007084 name = "linux_armeabi",
7085 values = {"cpu": "armeabi"},
7086)
7087
7088config_setting(
Terry Heo68eef3f2020-04-13 22:53:52 -07007089 name = "linux_armhf",
7090 values = {"cpu": "armhf"},
7091)
7092
7093config_setting(
Marat Dukhana720e932020-06-10 13:01:11 -07007094 name = "linux_armv7a",
7095 values = {"cpu": "armv7a"},
7096)
7097
7098config_setting(
Marat Dukhan582094e2020-04-30 17:21:25 -07007099 name = "linux_aarch64",
7100 values = {"cpu": "aarch64"},
7101)
7102
7103config_setting(
Marat Dukhan08c4a432019-10-03 09:29:21 -07007104 name = "android",
7105 values = {"crosstool_top": "//external:android/crosstool"},
7106)
7107
7108config_setting(
7109 name = "android_armv7",
7110 values = {
7111 "crosstool_top": "//external:android/crosstool",
7112 "cpu": "armeabi-v7a",
7113 },
7114)
7115
7116config_setting(
7117 name = "android_arm64",
7118 values = {
7119 "crosstool_top": "//external:android/crosstool",
7120 "cpu": "arm64-v8a",
7121 },
7122)
7123
7124config_setting(
7125 name = "android_x86",
7126 values = {
7127 "crosstool_top": "//external:android/crosstool",
7128 "cpu": "x86",
7129 },
7130)
7131
7132config_setting(
7133 name = "android_x86_64",
7134 values = {
7135 "crosstool_top": "//external:android/crosstool",
7136 "cpu": "x86_64",
7137 },
7138)
7139
7140config_setting(
Marat Dukhan10a38082020-04-17 03:58:35 -07007141 name = "windows_x86_64",
7142 values = {"cpu": "x64_windows"},
Marat Dukhan9fe932e2020-04-11 17:14:15 -07007143)
7144
7145config_setting(
Marat Dukhan10a38082020-04-17 03:58:35 -07007146 name = "windows_x86_64_clang",
7147 values = {
7148 "compiler": "clang-cl",
7149 "cpu": "x64_windows",
7150 },
7151)
7152
7153config_setting(
7154 name = "windows_x86_64_mingw",
7155 values = {
7156 "compiler": "mingw-gcc",
7157 "cpu": "x64_windows",
7158 },
7159)
7160
7161config_setting(
7162 name = "windows_x86_64_msys",
7163 values = {
7164 "compiler": "msys-gcc",
7165 "cpu": "x64_windows",
7166 },
Marat Dukhan9fe932e2020-04-11 17:14:15 -07007167)
7168
7169config_setting(
Marat Dukhan885ca242019-10-07 09:17:32 -07007170 name = "macos_x86_64",
7171 values = {
7172 "apple_platform_type": "macos",
7173 "cpu": "darwin",
7174 },
7175)
7176
7177config_setting(
Marat Dukhan08c4a432019-10-03 09:29:21 -07007178 name = "emscripten",
Marat Dukhan1a729ec2019-10-07 09:31:44 -07007179 values = {"crosstool_top": "//toolchain:emscripten"},
Marat Dukhan08c4a432019-10-03 09:29:21 -07007180)
7181
7182config_setting(
7183 name = "emscripten_wasm",
7184 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07007185 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007186 "cpu": "wasm",
7187 },
7188)
7189
7190config_setting(
7191 name = "emscripten_wasmsimd",
7192 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07007193 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007194 "cpu": "wasm",
Marat Dukhan81c62602020-05-29 13:22:49 -07007195 "copt": "-msimd128",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007196 },
7197)
7198
7199config_setting(
Marat Dukhan1498d1d2020-02-11 20:00:05 -08007200 name = "ios_armv7",
7201 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08007202 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08007203 "cpu": "ios_armv7",
7204 },
7205)
7206
7207config_setting(
7208 name = "ios_arm64",
7209 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08007210 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08007211 "cpu": "ios_arm64",
7212 },
7213)
7214
7215config_setting(
7216 name = "ios_arm64e",
7217 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08007218 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08007219 "cpu": "ios_arm64e",
7220 },
7221)
7222
7223config_setting(
7224 name = "ios_x86",
7225 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08007226 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08007227 "cpu": "ios_i386",
7228 },
7229)
7230
7231config_setting(
7232 name = "ios_x86_64",
7233 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08007234 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08007235 "cpu": "ios_x86_64",
7236 },
7237)
7238
7239config_setting(
7240 name = "watchos_armv7k",
7241 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08007242 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08007243 "cpu": "watchos_armv7k",
7244 },
7245)
7246
7247config_setting(
7248 name = "watchos_arm64_32",
7249 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08007250 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08007251 "cpu": "watchos_arm64_32",
7252 },
7253)
7254
7255config_setting(
7256 name = "watchos_x86",
7257 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08007258 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08007259 "cpu": "watchos_i386",
7260 },
7261)
7262
7263config_setting(
7264 name = "watchos_x86_64",
7265 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08007266 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08007267 "cpu": "watchos_x86_64",
7268 },
7269)
7270
7271config_setting(
7272 name = "tvos_arm64",
7273 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08007274 "apple_platform_type": "tvos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08007275 "cpu": "tvos_arm64",
7276 },
7277)
7278
7279config_setting(
7280 name = "tvos_x86_64",
7281 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08007282 "apple_platform_type": "tvos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08007283 "cpu": "tvos_x86_64",
7284 },
7285)