blob: 6188d256ef472a442b2fefe9aa78ce2758b8b138 [file] [log] [blame]
Frank Barchard1a953052020-11-16 18:44:58 -08001# Copyright 2020 Google LLC
Marat Dukhan08c4a432019-10-03 09:29:21 -07002#
3# This source code is licensed under the BSD-style license found in the
4# LICENSE file in the root directory of this source tree.
5#
6# Description:
7# XNNPACK - optimized floating-point neural network operators library
8
Marat Dukhan10a38082020-04-17 03:58:35 -07009load(":build_defs.bzl", "xnnpack_aggregate_library", "xnnpack_benchmark", "xnnpack_binary", "xnnpack_cc_library", "xnnpack_gcc_std_copts", "xnnpack_min_size_copts", "xnnpack_msvc_std_copts", "xnnpack_optional_armcl_copts", "xnnpack_optional_armcl_deps", "xnnpack_optional_dnnl_copts", "xnnpack_optional_dnnl_deps", "xnnpack_optional_gemmlowp_copts", "xnnpack_optional_gemmlowp_deps", "xnnpack_optional_ruy_copts", "xnnpack_optional_ruy_deps", "xnnpack_optional_tflite_copts", "xnnpack_optional_tflite_deps", "xnnpack_std_cxxopts", "xnnpack_unit_test", "xnnpack_visibility")
Marat Dukhan69c3f2c2019-11-06 12:30:01 -080010
Marat Dukhan08c4a432019-10-03 09:29:21 -070011licenses(["notice"])
12
13exports_files(["LICENSE"])
14
Marat Dukhan1b354632020-03-23 12:50:22 -070015OPERATOR_BENCHMARK_DEPS = [
16 ":XNNPACK",
17 ":bench_utils",
18 "@cpuinfo",
Frank Barchard0c849732020-06-12 13:31:32 -070019 "@FP16",
Marat Dukhan1b354632020-03-23 12:50:22 -070020 "@pthreadpool",
21]
22
Marat Dukhan08c4a432019-10-03 09:29:21 -070023MICROKERNEL_BENCHMARK_DEPS = [
24 ":ukernels",
25 ":bench_utils",
Frank Barchard7e955972019-10-11 10:34:25 -070026 ":enable_assembly",
Marat Dukhan08c4a432019-10-03 09:29:21 -070027 "@cpuinfo",
28 "@FP16",
29 "@pthreadpool",
30]
31
Marat Dukhan6adff4e2019-10-14 18:32:07 -070032ACCURACY_EVAL_DEPS = [
33 ":XNNPACK",
34 ":ukernels",
35 "@FP16",
36 "@pthreadpool",
37]
38
Marat Dukhan08c4a432019-10-03 09:29:21 -070039MICROKERNEL_TEST_DEPS = [
Marat Dukhan33fcf782020-05-24 14:27:15 -070040 ":ukernels_test_mode",
Frank Barchard7e955972019-10-11 10:34:25 -070041 ":enable_assembly",
Marat Dukhan08c4a432019-10-03 09:29:21 -070042 "@cpuinfo",
43 "@FP16",
44 "@pthreadpool",
45]
46
Marat Dukhan1b354632020-03-23 12:50:22 -070047OPERATOR_TEST_DEPS = [
Marat Dukhan33fcf782020-05-24 14:27:15 -070048 ":XNNPACK_test_mode",
Marat Dukhan1b354632020-03-23 12:50:22 -070049 "@pthreadpool",
50 "@FP16",
51]
52
Marat Dukhan08c4a432019-10-03 09:29:21 -070053OPERATOR_SRCS = [
Marat Dukhane8265432020-04-28 18:42:59 -070054 "src/operators/argmax-pooling-nhwc.c",
55 "src/operators/average-pooling-nhwc.c",
56 "src/operators/binary-elementwise-nd.c",
Marat Dukhane8265432020-04-28 18:42:59 -070057 "src/operators/channel-shuffle-nc.c",
Marat Dukhan065b11e2020-05-22 09:49:41 -070058 "src/operators/constant-pad-nd.c",
Marat Dukhane8265432020-04-28 18:42:59 -070059 "src/operators/convolution-nchw.c",
60 "src/operators/convolution-nhwc.c",
61 "src/operators/deconvolution-nhwc.c",
Marat Dukhan13b68f22020-11-12 11:55:19 -080062 "src/operators/depth-to-space-nchw2nhwc.c",
Marat Dukhan0e521172020-11-25 13:10:04 -080063 "src/operators/depth-to-space-nhwc.c",
Marat Dukhane8265432020-04-28 18:42:59 -070064 "src/operators/fully-connected-nc.c",
65 "src/operators/global-average-pooling-ncw.c",
66 "src/operators/global-average-pooling-nwc.c",
Marat Dukhane8265432020-04-28 18:42:59 -070067 "src/operators/leaky-relu-nc.c",
68 "src/operators/max-pooling-nhwc.c",
69 "src/operators/prelu-nc.c",
Artsiom Ablavatski97918102020-10-27 15:52:59 -070070 "src/operators/resize-bilinear-nchw.c",
Marat Dukhane8265432020-04-28 18:42:59 -070071 "src/operators/resize-bilinear-nhwc.c",
72 "src/operators/sigmoid-nc.c",
73 "src/operators/softmax-nc.c",
Marat Dukhanc3065f52020-06-04 13:33:32 -070074 "src/operators/unary-elementwise-nc.c",
Marat Dukhane8265432020-04-28 18:42:59 -070075 "src/operators/unpooling-nhwc.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -070076]
77
Marat Dukhan95e8b7a2020-06-03 12:46:26 -070078SUBGRAPH_SRCS = [
Marat Dukhan5fab4092020-06-10 01:28:28 -070079 "src/subgraph/abs.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -070080 "src/subgraph/add2.c",
81 "src/subgraph/argmax-pooling-2d.c",
82 "src/subgraph/average-pooling-2d.c",
Marat Dukhan5fab4092020-06-10 01:28:28 -070083 "src/subgraph/bankers-rounding.c",
84 "src/subgraph/ceiling.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -070085 "src/subgraph/clamp.c",
86 "src/subgraph/convolution-2d.c",
87 "src/subgraph/deconvolution-2d.c",
Artsiom Ablavatskibbe85062020-11-05 14:07:37 -080088 "src/subgraph/depth-to-space.c",
Frank Barchard9cef5ea2020-11-18 14:52:08 -080089 "src/subgraph/depthwise-convolution-2d.c",
Marat Dukhan9d3a4592020-06-05 16:52:42 -070090 "src/subgraph/divide.c",
Marat Dukhana1600202020-12-01 22:17:16 -080091 "src/subgraph/elu.c",
Marat Dukhan5fab4092020-06-10 01:28:28 -070092 "src/subgraph/floor.c",
Frank Barchard04336c12020-10-22 16:48:55 -070093 "src/subgraph/fully-connected.c",
Marat Dukhana059b7d2020-06-11 11:41:27 -070094 "src/subgraph/global-average-pooling-2d.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -070095 "src/subgraph/hardswish.c",
Marat Dukhan5bbebac2020-06-10 19:42:15 -070096 "src/subgraph/leaky-relu.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -070097 "src/subgraph/max-pooling-2d.c",
Marat Dukhan9d3a4592020-06-05 16:52:42 -070098 "src/subgraph/maximum2.c",
99 "src/subgraph/minimum2.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -0700100 "src/subgraph/multiply2.c",
Marat Dukhan5fab4092020-06-10 01:28:28 -0700101 "src/subgraph/negate.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -0700102 "src/subgraph/prelu.c",
103 "src/subgraph/sigmoid.c",
104 "src/subgraph/softmax.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700105 "src/subgraph/square-root.c",
106 "src/subgraph/square.c",
107 "src/subgraph/squared-difference.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -0700108 "src/subgraph/static-constant-pad.c",
Marat Dukhand27202d2020-07-09 23:43:40 -0700109 "src/subgraph/static-reshape.c",
Marat Dukhanaff24e22020-07-23 01:43:58 -0700110 "src/subgraph/static-resize-bilinear-2d.c",
Marat Dukhan9d3a4592020-06-05 16:52:42 -0700111 "src/subgraph/subtract.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -0700112 "src/subgraph/unpooling-2d.c",
113]
114
Marat Dukhan3a77ea72019-12-23 12:10:24 -0800115TABLE_SRCS = [
116 "src/tables/exp2-k-over-64.c",
117 "src/tables/exp2-k-over-2048.c",
Marat Dukhande390d42020-11-29 19:32:18 -0800118 "src/tables/exp2minus-k-over-4.c",
119 "src/tables/exp2minus-k-over-8.c",
Marat Dukhanc60742b2020-11-23 12:33:27 -0800120 "src/tables/exp2minus-k-over-16.c",
Marat Dukhan1f256fc2020-09-24 21:27:14 -0700121 "src/tables/exp2minus-k-over-64.c",
122 "src/tables/exp2minus-k-over-2048.c",
Marat Dukhan3a77ea72019-12-23 12:10:24 -0800123]
124
Marat Dukhan08c4a432019-10-03 09:29:21 -0700125SCALAR_UKERNELS = [
Marat Dukhan329da642019-11-19 21:44:39 -0800126 "src/f32-argmaxpool/4x-scalar-c1.c",
Marat Dukhan1e782c42019-11-21 17:02:40 -0800127 "src/f32-argmaxpool/9p8x-scalar-c1.c",
Marat Dukhan329da642019-11-19 21:44:39 -0800128 "src/f32-argmaxpool/9x-scalar-c1.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700129 "src/f32-avgpool/9p8x-minmax-scalar-c1.c",
130 "src/f32-avgpool/9x-minmax-scalar-c1.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -0700131 "src/f32-clamp/gen/scalar-x1.c",
132 "src/f32-clamp/gen/scalar-x2.c",
133 "src/f32-clamp/gen/scalar-x4.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700134 "src/f32-conv-hwc/3x3s2p0p1c3x4-scalar-1x1.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700135 "src/f32-conv-hwc/3x3s2p1c3x4-scalar-1x1.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -0700136 "src/f32-conv-hwc2chw/3x3s2p1c3x4-scalar-1x1.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700137 "src/f32-dwconv/gen/up1x4-minmax-scalar-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700138 "src/f32-dwconv/gen/up1x4-minmax-scalar.c",
139 "src/f32-dwconv/gen/up1x4-scalar-acc2.c",
140 "src/f32-dwconv/gen/up1x4-scalar.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700141 "src/f32-dwconv/gen/up1x9-minmax-scalar-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700142 "src/f32-dwconv/gen/up1x9-minmax-scalar.c",
143 "src/f32-dwconv/gen/up1x9-scalar-acc2.c",
144 "src/f32-dwconv/gen/up1x9-scalar.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700145 "src/f32-dwconv/gen/up1x25-minmax-scalar-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700146 "src/f32-dwconv/gen/up1x25-minmax-scalar.c",
147 "src/f32-dwconv/gen/up1x25-scalar-acc2.c",
148 "src/f32-dwconv/gen/up1x25-scalar.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700149 "src/f32-dwconv/gen/up2x4-minmax-scalar-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700150 "src/f32-dwconv/gen/up2x4-minmax-scalar.c",
151 "src/f32-dwconv/gen/up2x4-scalar-acc2.c",
152 "src/f32-dwconv/gen/up2x4-scalar.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700153 "src/f32-dwconv/gen/up2x9-minmax-scalar-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700154 "src/f32-dwconv/gen/up2x9-minmax-scalar.c",
155 "src/f32-dwconv/gen/up2x9-scalar-acc2.c",
156 "src/f32-dwconv/gen/up2x9-scalar.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700157 "src/f32-dwconv/gen/up2x25-minmax-scalar-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700158 "src/f32-dwconv/gen/up2x25-minmax-scalar.c",
159 "src/f32-dwconv/gen/up2x25-scalar-acc2.c",
160 "src/f32-dwconv/gen/up2x25-scalar.c",
Frank Barchard35db7d02020-10-26 13:37:34 -0700161 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-scalar-1x1-acc2.c",
162 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-scalar-1x1-acc3.c",
163 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-scalar-1x1-acc4.c",
Marat Dukhan91249d22020-10-24 12:02:51 -0700164 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-scalar-1x1.c",
Frank Barchard35db7d02020-10-26 13:37:34 -0700165 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-scalar-2x1-acc2.c",
Marat Dukhan91249d22020-10-24 12:02:51 -0700166 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-scalar-2x1.c",
167 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-scalar-3x1.c",
168 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-scalar-4x1.c",
169 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-scalar-5x1.c",
170 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-scalar-6x1.c",
Marat Dukhancf5b3c32020-10-25 19:21:10 -0700171 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-scalar-1x1-acc2.c",
172 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-scalar-1x1-acc3.c",
173 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-scalar-1x1-acc4.c",
Frank Barchard35db7d02020-10-26 13:37:34 -0700174 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-scalar-1x1.c",
Marat Dukhancf5b3c32020-10-25 19:21:10 -0700175 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-scalar-2x1-acc2.c",
Frank Barchard35db7d02020-10-26 13:37:34 -0700176 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-scalar-2x1.c",
177 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-scalar-3x1.c",
178 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-scalar-4x1.c",
Marat Dukhanc4efb002020-10-25 23:14:47 -0700179 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-scalar-1x1-acc2.c",
180 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-scalar-1x1-acc3.c",
181 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-scalar-1x1-acc4.c",
182 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-scalar-1x1-acc5.c",
Frank Barchard35db7d02020-10-26 13:37:34 -0700183 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-scalar-1x1.c",
Marat Dukhanc4efb002020-10-25 23:14:47 -0700184 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-scalar-2x1-acc2.c",
185 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-scalar-2x1-acc3.c",
Frank Barchard35db7d02020-10-26 13:37:34 -0700186 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-scalar-2x1.c",
Marat Dukhanc4efb002020-10-25 23:14:47 -0700187 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-scalar-3x1-acc2.c",
Frank Barchard35db7d02020-10-26 13:37:34 -0700188 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-scalar-3x1.c",
Marat Dukhan29c0c332020-10-28 22:11:00 -0700189 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-scalar-1x1-acc2.c",
190 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-scalar-1x1-acc3.c",
191 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-scalar-1x1-acc4.c",
192 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-scalar-1x1-acc5.c",
193 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-scalar-1x1.c",
194 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-scalar-2x1-acc2.c",
195 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-scalar-2x1-acc3.c",
196 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-scalar-2x1.c",
197 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-scalar-3x1-acc2.c",
198 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-scalar-3x1.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -0700199 "src/f32-gavgpool-cw/scalar-x1.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700200 "src/f32-gavgpool/7p7x-minmax-scalar-c1.c",
201 "src/f32-gavgpool/7x-minmax-scalar-c1.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700202 "src/f32-gemm/gen-inc/1x4inc-minmax-scalar.c",
203 "src/f32-gemm/gen-inc/2x4inc-minmax-scalar.c",
204 "src/f32-gemm/gen-inc/4x4inc-minmax-scalar.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700205 "src/f32-gemm/gen/1x4-minmax-scalar.c",
206 "src/f32-gemm/gen/1x4-relu-scalar.c",
207 "src/f32-gemm/gen/1x4-scalar.c",
208 "src/f32-gemm/gen/2x4-minmax-scalar.c",
209 "src/f32-gemm/gen/2x4-relu-scalar.c",
210 "src/f32-gemm/gen/2x4-scalar.c",
211 "src/f32-gemm/gen/4x2-minmax-scalar.c",
212 "src/f32-gemm/gen/4x2-relu-scalar.c",
213 "src/f32-gemm/gen/4x2-scalar.c",
214 "src/f32-gemm/gen/4x4-minmax-scalar.c",
215 "src/f32-gemm/gen/4x4-relu-scalar.c",
216 "src/f32-gemm/gen/4x4-scalar.c",
Marat Dukhanefaac272020-07-16 16:58:07 -0700217 "src/f32-hswish/gen/hswish-scalar-x1.c",
218 "src/f32-hswish/gen/hswish-scalar-x2.c",
219 "src/f32-hswish/gen/hswish-scalar-x4.c",
XNNPACK Team21432672020-10-19 19:58:48 -0700220 "src/f32-ibilinear-chw/gen/scalar-p1.c",
221 "src/f32-ibilinear-chw/gen/scalar-p2.c",
222 "src/f32-ibilinear-chw/gen/scalar-p4.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700223 "src/f32-ibilinear/gen/scalar-c1.c",
224 "src/f32-ibilinear/gen/scalar-c2.c",
225 "src/f32-ibilinear/gen/scalar-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700226 "src/f32-igemm/gen/1x4-minmax-scalar.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700227 "src/f32-igemm/gen/1x4-relu-scalar.c",
228 "src/f32-igemm/gen/1x4-scalar.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700229 "src/f32-igemm/gen/2x4-minmax-scalar.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700230 "src/f32-igemm/gen/2x4-relu-scalar.c",
231 "src/f32-igemm/gen/2x4-scalar.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700232 "src/f32-igemm/gen/4x2-minmax-scalar.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700233 "src/f32-igemm/gen/4x2-relu-scalar.c",
234 "src/f32-igemm/gen/4x2-scalar.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700235 "src/f32-igemm/gen/4x4-minmax-scalar.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700236 "src/f32-igemm/gen/4x4-relu-scalar.c",
237 "src/f32-igemm/gen/4x4-scalar.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700238 "src/f32-maxpool/9p8x-minmax-scalar-c1.c",
239 "src/f32-pavgpool/9p8x-minmax-scalar-c1.c",
240 "src/f32-pavgpool/9x-minmax-scalar-c1.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700241 "src/f32-ppmm/gen/2x4-minmax-scalar.c",
242 "src/f32-ppmm/gen/3x3-minmax-scalar.c",
243 "src/f32-ppmm/gen/4x2-minmax-scalar.c",
244 "src/f32-ppmm/gen/4x4-minmax-scalar.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800245 "src/f32-prelu/gen/scalar-2x1.c",
246 "src/f32-prelu/gen/scalar-2x4.c",
Marat Dukhanf46f6752020-01-21 11:03:49 -0800247 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x1.c",
Marat Dukhanf46f6752020-01-21 11:03:49 -0800248 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x2-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700249 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x2.c",
Marat Dukhanf46f6752020-01-21 11:03:49 -0800250 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x4-acc2.c",
251 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x4-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700252 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x4.c",
Marat Dukhanf46f6752020-01-21 11:03:49 -0800253 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x1.c",
Marat Dukhanf46f6752020-01-21 11:03:49 -0800254 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x2-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700255 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x2.c",
Marat Dukhanf46f6752020-01-21 11:03:49 -0800256 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x4-acc2.c",
257 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x4-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700258 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x4.c",
Frank Barchardfb158e22020-07-15 16:10:10 -0700259 "src/f32-relu/gen/scalar-x1.c",
260 "src/f32-relu/gen/scalar-x2.c",
261 "src/f32-relu/gen/scalar-x4.c",
Frank Barchard37297a62020-08-27 12:15:10 -0700262 "src/f32-relu/gen/scalar-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700263 "src/f32-rmax/scalar.c",
Marat Dukhan3a77ea72019-12-23 12:10:24 -0800264 "src/f32-sigmoid/gen/scalar-lut64-p2-div-x1.c",
265 "src/f32-sigmoid/gen/scalar-lut64-p2-div-x2.c",
266 "src/f32-sigmoid/gen/scalar-lut64-p2-div-x4.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700267 "src/f32-sigmoid/gen/scalar-lut2048-p1-div-x1.c",
268 "src/f32-sigmoid/gen/scalar-lut2048-p1-div-x2.c",
269 "src/f32-sigmoid/gen/scalar-lut2048-p1-div-x4.c",
Marat Dukhan3a77ea72019-12-23 12:10:24 -0800270 "src/f32-sigmoid/gen/scalar-p5-div-x1.c",
271 "src/f32-sigmoid/gen/scalar-p5-div-x2.c",
272 "src/f32-sigmoid/gen/scalar-p5-div-x4.c",
Marat Dukhan355ab432020-04-09 19:01:52 -0700273 "src/f32-spmm/gen/1x1-minmax-scalar-pipelined.c",
274 "src/f32-spmm/gen/1x1-minmax-scalar.c",
275 "src/f32-spmm/gen/2x1-minmax-scalar-pipelined.c",
276 "src/f32-spmm/gen/2x1-minmax-scalar.c",
277 "src/f32-spmm/gen/4x1-minmax-scalar-pipelined.c",
278 "src/f32-spmm/gen/4x1-minmax-scalar.c",
279 "src/f32-spmm/gen/8x1-minmax-scalar-pipelined.c",
280 "src/f32-spmm/gen/8x1-minmax-scalar.c",
281 "src/f32-spmm/gen/8x2-minmax-scalar.c",
282 "src/f32-spmm/gen/8x4-minmax-scalar.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700283 "src/f32-vbinary/gen/vadd-minmax-scalar-x1.c",
284 "src/f32-vbinary/gen/vadd-minmax-scalar-x2.c",
285 "src/f32-vbinary/gen/vadd-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700286 "src/f32-vbinary/gen/vadd-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700287 "src/f32-vbinary/gen/vadd-relu-scalar-x1.c",
288 "src/f32-vbinary/gen/vadd-relu-scalar-x2.c",
289 "src/f32-vbinary/gen/vadd-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700290 "src/f32-vbinary/gen/vadd-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700291 "src/f32-vbinary/gen/vadd-scalar-x1.c",
292 "src/f32-vbinary/gen/vadd-scalar-x2.c",
293 "src/f32-vbinary/gen/vadd-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700294 "src/f32-vbinary/gen/vadd-scalar-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700295 "src/f32-vbinary/gen/vaddc-minmax-scalar-x1.c",
296 "src/f32-vbinary/gen/vaddc-minmax-scalar-x2.c",
297 "src/f32-vbinary/gen/vaddc-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700298 "src/f32-vbinary/gen/vaddc-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700299 "src/f32-vbinary/gen/vaddc-relu-scalar-x1.c",
300 "src/f32-vbinary/gen/vaddc-relu-scalar-x2.c",
301 "src/f32-vbinary/gen/vaddc-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700302 "src/f32-vbinary/gen/vaddc-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700303 "src/f32-vbinary/gen/vaddc-scalar-x1.c",
304 "src/f32-vbinary/gen/vaddc-scalar-x2.c",
305 "src/f32-vbinary/gen/vaddc-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700306 "src/f32-vbinary/gen/vaddc-scalar-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700307 "src/f32-vbinary/gen/vdiv-minmax-scalar-x1.c",
308 "src/f32-vbinary/gen/vdiv-minmax-scalar-x2.c",
309 "src/f32-vbinary/gen/vdiv-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700310 "src/f32-vbinary/gen/vdiv-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700311 "src/f32-vbinary/gen/vdiv-relu-scalar-x1.c",
312 "src/f32-vbinary/gen/vdiv-relu-scalar-x2.c",
313 "src/f32-vbinary/gen/vdiv-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700314 "src/f32-vbinary/gen/vdiv-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700315 "src/f32-vbinary/gen/vdiv-scalar-x1.c",
316 "src/f32-vbinary/gen/vdiv-scalar-x2.c",
317 "src/f32-vbinary/gen/vdiv-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700318 "src/f32-vbinary/gen/vdiv-scalar-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700319 "src/f32-vbinary/gen/vdivc-minmax-scalar-x1.c",
320 "src/f32-vbinary/gen/vdivc-minmax-scalar-x2.c",
321 "src/f32-vbinary/gen/vdivc-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700322 "src/f32-vbinary/gen/vdivc-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700323 "src/f32-vbinary/gen/vdivc-relu-scalar-x1.c",
324 "src/f32-vbinary/gen/vdivc-relu-scalar-x2.c",
325 "src/f32-vbinary/gen/vdivc-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700326 "src/f32-vbinary/gen/vdivc-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700327 "src/f32-vbinary/gen/vdivc-scalar-x1.c",
328 "src/f32-vbinary/gen/vdivc-scalar-x2.c",
329 "src/f32-vbinary/gen/vdivc-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700330 "src/f32-vbinary/gen/vdivc-scalar-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800331 "src/f32-vbinary/gen/vmax-scalar-x1.c",
332 "src/f32-vbinary/gen/vmax-scalar-x2.c",
333 "src/f32-vbinary/gen/vmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700334 "src/f32-vbinary/gen/vmax-scalar-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800335 "src/f32-vbinary/gen/vmaxc-scalar-x1.c",
336 "src/f32-vbinary/gen/vmaxc-scalar-x2.c",
337 "src/f32-vbinary/gen/vmaxc-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700338 "src/f32-vbinary/gen/vmaxc-scalar-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800339 "src/f32-vbinary/gen/vmin-scalar-x1.c",
340 "src/f32-vbinary/gen/vmin-scalar-x2.c",
341 "src/f32-vbinary/gen/vmin-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700342 "src/f32-vbinary/gen/vmin-scalar-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800343 "src/f32-vbinary/gen/vminc-scalar-x1.c",
344 "src/f32-vbinary/gen/vminc-scalar-x2.c",
345 "src/f32-vbinary/gen/vminc-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700346 "src/f32-vbinary/gen/vminc-scalar-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700347 "src/f32-vbinary/gen/vmul-minmax-scalar-x1.c",
348 "src/f32-vbinary/gen/vmul-minmax-scalar-x2.c",
349 "src/f32-vbinary/gen/vmul-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700350 "src/f32-vbinary/gen/vmul-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700351 "src/f32-vbinary/gen/vmul-relu-scalar-x1.c",
352 "src/f32-vbinary/gen/vmul-relu-scalar-x2.c",
353 "src/f32-vbinary/gen/vmul-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700354 "src/f32-vbinary/gen/vmul-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700355 "src/f32-vbinary/gen/vmul-scalar-x1.c",
356 "src/f32-vbinary/gen/vmul-scalar-x2.c",
357 "src/f32-vbinary/gen/vmul-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700358 "src/f32-vbinary/gen/vmul-scalar-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700359 "src/f32-vbinary/gen/vmulc-minmax-scalar-x1.c",
360 "src/f32-vbinary/gen/vmulc-minmax-scalar-x2.c",
361 "src/f32-vbinary/gen/vmulc-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700362 "src/f32-vbinary/gen/vmulc-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700363 "src/f32-vbinary/gen/vmulc-relu-scalar-x1.c",
364 "src/f32-vbinary/gen/vmulc-relu-scalar-x2.c",
365 "src/f32-vbinary/gen/vmulc-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700366 "src/f32-vbinary/gen/vmulc-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700367 "src/f32-vbinary/gen/vmulc-scalar-x1.c",
368 "src/f32-vbinary/gen/vmulc-scalar-x2.c",
369 "src/f32-vbinary/gen/vmulc-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700370 "src/f32-vbinary/gen/vmulc-scalar-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700371 "src/f32-vbinary/gen/vrdivc-minmax-scalar-x1.c",
372 "src/f32-vbinary/gen/vrdivc-minmax-scalar-x2.c",
373 "src/f32-vbinary/gen/vrdivc-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700374 "src/f32-vbinary/gen/vrdivc-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700375 "src/f32-vbinary/gen/vrdivc-relu-scalar-x1.c",
376 "src/f32-vbinary/gen/vrdivc-relu-scalar-x2.c",
377 "src/f32-vbinary/gen/vrdivc-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700378 "src/f32-vbinary/gen/vrdivc-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700379 "src/f32-vbinary/gen/vrdivc-scalar-x1.c",
380 "src/f32-vbinary/gen/vrdivc-scalar-x2.c",
381 "src/f32-vbinary/gen/vrdivc-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700382 "src/f32-vbinary/gen/vrdivc-scalar-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700383 "src/f32-vbinary/gen/vrsubc-minmax-scalar-x1.c",
384 "src/f32-vbinary/gen/vrsubc-minmax-scalar-x2.c",
385 "src/f32-vbinary/gen/vrsubc-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700386 "src/f32-vbinary/gen/vrsubc-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700387 "src/f32-vbinary/gen/vrsubc-relu-scalar-x1.c",
388 "src/f32-vbinary/gen/vrsubc-relu-scalar-x2.c",
389 "src/f32-vbinary/gen/vrsubc-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700390 "src/f32-vbinary/gen/vrsubc-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700391 "src/f32-vbinary/gen/vrsubc-scalar-x1.c",
392 "src/f32-vbinary/gen/vrsubc-scalar-x2.c",
393 "src/f32-vbinary/gen/vrsubc-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700394 "src/f32-vbinary/gen/vrsubc-scalar-x8.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -0700395 "src/f32-vbinary/gen/vsqrdiff-scalar-x1.c",
396 "src/f32-vbinary/gen/vsqrdiff-scalar-x2.c",
397 "src/f32-vbinary/gen/vsqrdiff-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700398 "src/f32-vbinary/gen/vsqrdiff-scalar-x8.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -0700399 "src/f32-vbinary/gen/vsqrdiffc-scalar-x1.c",
400 "src/f32-vbinary/gen/vsqrdiffc-scalar-x2.c",
401 "src/f32-vbinary/gen/vsqrdiffc-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700402 "src/f32-vbinary/gen/vsqrdiffc-scalar-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700403 "src/f32-vbinary/gen/vsub-minmax-scalar-x1.c",
404 "src/f32-vbinary/gen/vsub-minmax-scalar-x2.c",
405 "src/f32-vbinary/gen/vsub-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700406 "src/f32-vbinary/gen/vsub-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700407 "src/f32-vbinary/gen/vsub-relu-scalar-x1.c",
408 "src/f32-vbinary/gen/vsub-relu-scalar-x2.c",
409 "src/f32-vbinary/gen/vsub-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700410 "src/f32-vbinary/gen/vsub-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700411 "src/f32-vbinary/gen/vsub-scalar-x1.c",
412 "src/f32-vbinary/gen/vsub-scalar-x2.c",
413 "src/f32-vbinary/gen/vsub-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700414 "src/f32-vbinary/gen/vsub-scalar-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700415 "src/f32-vbinary/gen/vsubc-minmax-scalar-x1.c",
416 "src/f32-vbinary/gen/vsubc-minmax-scalar-x2.c",
417 "src/f32-vbinary/gen/vsubc-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700418 "src/f32-vbinary/gen/vsubc-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700419 "src/f32-vbinary/gen/vsubc-relu-scalar-x1.c",
420 "src/f32-vbinary/gen/vsubc-relu-scalar-x2.c",
421 "src/f32-vbinary/gen/vsubc-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700422 "src/f32-vbinary/gen/vsubc-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700423 "src/f32-vbinary/gen/vsubc-scalar-x1.c",
424 "src/f32-vbinary/gen/vsubc-scalar-x2.c",
425 "src/f32-vbinary/gen/vsubc-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700426 "src/f32-vbinary/gen/vsubc-scalar-x8.c",
Marat Dukhaned6baaf2020-12-01 15:07:08 -0800427 "src/f32-velu/gen/velu-scalar-rr2-lut16-p3-x1.c",
428 "src/f32-velu/gen/velu-scalar-rr2-lut16-p3-x2.c",
429 "src/f32-velu/gen/velu-scalar-rr2-lut16-p3-x3.c",
430 "src/f32-velu/gen/velu-scalar-rr2-lut16-p3-x4.c",
431 "src/f32-velu/gen/velu-scalar-rr2-lut16-p3-x5.c",
432 "src/f32-velu/gen/velu-scalar-rr2-lut16-p3-x6.c",
433 "src/f32-velu/gen/velu-scalar-rr2-p6-x1.c",
434 "src/f32-velu/gen/velu-scalar-rr2-p6-x2.c",
435 "src/f32-velu/gen/velu-scalar-rr2-p6-x3.c",
436 "src/f32-velu/gen/velu-scalar-rr2-p6-x4.c",
437 "src/f32-velu/gen/velu-scalar-rr2-p6-x5.c",
438 "src/f32-velu/gen/velu-scalar-rr2-p6-x6.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700439 "src/f32-vlrelu/gen/vlrelu-scalar-x1.c",
440 "src/f32-vlrelu/gen/vlrelu-scalar-x2.c",
441 "src/f32-vlrelu/gen/vlrelu-scalar-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700442 "src/f32-vmulcaddc/gen/c1-minmax-scalar-2x.c",
443 "src/f32-vmulcaddc/gen/c2-minmax-scalar-2x.c",
444 "src/f32-vmulcaddc/gen/c4-minmax-scalar-2x.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -0700445 "src/f32-vrnd/gen/vrndd-scalar-libm-x1.c",
446 "src/f32-vrnd/gen/vrndd-scalar-libm-x2.c",
447 "src/f32-vrnd/gen/vrndd-scalar-libm-x4.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700448 "src/f32-vrnd/gen/vrndne-scalar-libm-x1.c",
449 "src/f32-vrnd/gen/vrndne-scalar-libm-x2.c",
450 "src/f32-vrnd/gen/vrndne-scalar-libm-x4.c",
451 "src/f32-vrnd/gen/vrndu-scalar-libm-x1.c",
452 "src/f32-vrnd/gen/vrndu-scalar-libm-x2.c",
453 "src/f32-vrnd/gen/vrndu-scalar-libm-x4.c",
454 "src/f32-vrnd/gen/vrndz-scalar-libm-x1.c",
455 "src/f32-vrnd/gen/vrndz-scalar-libm-x2.c",
456 "src/f32-vrnd/gen/vrndz-scalar-libm-x4.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -0700457 "src/f32-vsqrt/gen/scalar-sqrt-x1.c",
458 "src/f32-vsqrt/gen/scalar-sqrt-x2.c",
459 "src/f32-vsqrt/gen/scalar-sqrt-x4.c",
Marat Dukhan5020b962020-06-08 13:30:10 -0700460 "src/f32-vunary/gen/vabs-scalar-x1.c",
461 "src/f32-vunary/gen/vabs-scalar-x2.c",
462 "src/f32-vunary/gen/vabs-scalar-x4.c",
463 "src/f32-vunary/gen/vneg-scalar-x1.c",
464 "src/f32-vunary/gen/vneg-scalar-x2.c",
465 "src/f32-vunary/gen/vneg-scalar-x4.c",
466 "src/f32-vunary/gen/vsqr-scalar-x1.c",
467 "src/f32-vunary/gen/vsqr-scalar-x2.c",
468 "src/f32-vunary/gen/vsqr-scalar-x4.c",
Marat Dukhande390d42020-11-29 19:32:18 -0800469 "src/math/expm1minus-scalar-rr2-lut4-p4.c",
470 "src/math/expm1minus-scalar-rr2-lut8-p3.c",
471 "src/math/expm1minus-scalar-rr2-lut8-p4.c",
Marat Dukhanc60742b2020-11-23 12:33:27 -0800472 "src/math/expm1minus-scalar-rr2-lut16-p3.c",
473 "src/math/expm1minus-scalar-rr2-lut16-p4.c",
474 "src/math/expm1minus-scalar-rr2-p5.c",
475 "src/math/expm1minus-scalar-rr2-p6.c",
Frank Barchard22136062020-11-24 18:44:46 -0800476 "src/math/expminus-scalar-rr2-lut64-p2.c",
477 "src/math/expminus-scalar-rr2-lut2048-p1.c",
478 "src/math/expminus-scalar-rr2-p5.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -0700479 "src/math/roundd-scalar-addsub.c",
480 "src/math/roundd-scalar-cvt.c",
Marat Dukhanffbf96a2020-05-14 02:59:08 -0700481 "src/math/roundd-scalar-floor.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700482 "src/math/roundne-scalar-addsub.c",
483 "src/math/roundne-scalar-nearbyint.c",
484 "src/math/roundne-scalar-rint.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -0700485 "src/math/roundu-scalar-addsub.c",
Marat Dukhanffbf96a2020-05-14 02:59:08 -0700486 "src/math/roundu-scalar-ceil.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -0700487 "src/math/roundu-scalar-cvt.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -0700488 "src/math/roundz-scalar-addsub.c",
489 "src/math/roundz-scalar-cvt.c",
Marat Dukhanffbf96a2020-05-14 02:59:08 -0700490 "src/math/roundz-scalar-trunc.c",
Marat Dukhanf8475d62020-09-17 15:01:43 -0700491 "src/math/sigmoid-scalar-rr2-lut64-p2-div.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700492 "src/math/sigmoid-scalar-rr2-lut2048-p1-div.c",
Marat Dukhanf8475d62020-09-17 15:01:43 -0700493 "src/math/sigmoid-scalar-rr2-p5-div.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -0700494 "src/qs8-requantization/fp32-scalar-lrintf.c",
495 "src/qs8-requantization/fp32-scalar-magic.c",
496 "src/qs8-requantization/precise-scalar-signed64.c",
497 "src/qs8-requantization/precise-scalar-unsigned32.c",
498 "src/qs8-requantization/precise-scalar-unsigned64.c",
499 "src/qs8-requantization/q31-scalar.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -0700500 "src/qu8-avgpool/9p8x-minmax-scalar-c1.c",
501 "src/qu8-avgpool/9x-minmax-scalar-c1.c",
502 "src/qu8-dwconv/up1x9-minmax-scalar.c",
503 "src/qu8-gavgpool/7p7x-minmax-scalar-c1.c",
504 "src/qu8-gavgpool/7x-minmax-scalar-c1.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700505 "src/qu8-gemm/2x2-minmax-scalar.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -0700506 "src/qu8-igemm/2x2-minmax-scalar.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -0700507 "src/qu8-requantization/fp32-scalar-lrintf.c",
508 "src/qu8-requantization/fp32-scalar-magic.c",
509 "src/qu8-requantization/precise-scalar-signed64.c",
510 "src/qu8-requantization/precise-scalar-unsigned32.c",
511 "src/qu8-requantization/precise-scalar-unsigned64.c",
512 "src/qu8-requantization/q31-scalar.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -0700513 "src/qu8-vadd/minmax-scalar.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -0700514 "src/u8-clamp/scalar-x4.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700515 "src/u8-lut32norm/scalar.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700516 "src/u8-maxpool/9p8x-minmax-scalar-c1.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700517 "src/u8-rmax/scalar.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700518 "src/x8-lut/scalar.c",
519 "src/x8-zip/x2-scalar.c",
520 "src/x8-zip/x3-scalar.c",
521 "src/x8-zip/x4-scalar.c",
522 "src/x8-zip/xm-scalar.c",
Marat Dukhanad71b9a2020-11-20 00:01:51 -0800523 "src/x32-depthtospace2d-chw2hwc/scalar.c",
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -0700524 "src/x32-fill/scalar-float.c",
525 "src/x32-fill/scalar-int.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700526 "src/x32-packx/x2-scalar.c",
527 "src/x32-packx/x3-scalar.c",
528 "src/x32-packx/x4-scalar.c",
Marat Dukhan63523d42020-05-22 17:07:33 -0700529 "src/x32-pad/scalar-float.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700530 "src/x32-pad/scalar-int.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700531 "src/x32-unpool/scalar.c",
532 "src/x32-zip/x2-scalar.c",
533 "src/x32-zip/x3-scalar.c",
534 "src/x32-zip/x4-scalar.c",
535 "src/x32-zip/xm-scalar.c",
Marat Dukhan048931b2020-11-24 20:53:54 -0800536 "src/xx-copy/memcpy.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700537]
538
Marat Dukhan436ebe62019-12-04 15:10:12 -0800539WASM_UKERNELS = [
Marat Dukhan99936602020-04-11 16:47:01 -0700540 "src/f32-avgpool/9p8x-minmax-wasm-c1.c",
541 "src/f32-avgpool/9x-minmax-wasm-c1.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -0700542 "src/f32-clamp/gen/wasm-x1.c",
543 "src/f32-clamp/gen/wasm-x2.c",
544 "src/f32-clamp/gen/wasm-x4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700545 "src/f32-dwconv/gen/up1x4-minmax-wasm-acc2.c",
546 "src/f32-dwconv/gen/up1x4-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700547 "src/f32-dwconv/gen/up1x4-wasm-acc2.c",
548 "src/f32-dwconv/gen/up1x4-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700549 "src/f32-dwconv/gen/up1x9-minmax-wasm-acc2.c",
550 "src/f32-dwconv/gen/up1x9-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700551 "src/f32-dwconv/gen/up1x9-wasm-acc2.c",
552 "src/f32-dwconv/gen/up1x9-wasm.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700553 "src/f32-dwconv/gen/up1x25-minmax-wasm-acc2.c",
554 "src/f32-dwconv/gen/up1x25-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700555 "src/f32-dwconv/gen/up1x25-wasm-acc2.c",
556 "src/f32-dwconv/gen/up1x25-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700557 "src/f32-dwconv/gen/up2x4-minmax-wasm-acc2.c",
558 "src/f32-dwconv/gen/up2x4-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700559 "src/f32-dwconv/gen/up2x4-wasm-acc2.c",
560 "src/f32-dwconv/gen/up2x4-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700561 "src/f32-dwconv/gen/up2x9-minmax-wasm-acc2.c",
562 "src/f32-dwconv/gen/up2x9-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700563 "src/f32-dwconv/gen/up2x9-wasm-acc2.c",
564 "src/f32-dwconv/gen/up2x9-wasm.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700565 "src/f32-dwconv/gen/up2x25-minmax-wasm-acc2.c",
566 "src/f32-dwconv/gen/up2x25-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700567 "src/f32-dwconv/gen/up2x25-wasm-acc2.c",
568 "src/f32-dwconv/gen/up2x25-wasm.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700569 "src/f32-gavgpool/7p7x-minmax-wasm-c1.c",
570 "src/f32-gavgpool/7x-minmax-wasm-c1.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700571 "src/f32-gemm/gen-inc/1x4inc-minmax-wasm.c",
572 "src/f32-gemm/gen-inc/2x4inc-minmax-wasm.c",
573 "src/f32-gemm/gen-inc/4x4inc-minmax-wasm.c",
574 "src/f32-gemm/gen/1x4-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700575 "src/f32-gemm/gen/1x4-relu-wasm.c",
576 "src/f32-gemm/gen/1x4-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700577 "src/f32-gemm/gen/2x4-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700578 "src/f32-gemm/gen/2x4-relu-wasm.c",
579 "src/f32-gemm/gen/2x4-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700580 "src/f32-gemm/gen/4x2-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700581 "src/f32-gemm/gen/4x2-relu-wasm.c",
582 "src/f32-gemm/gen/4x2-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700583 "src/f32-gemm/gen/4x4-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700584 "src/f32-gemm/gen/4x4-relu-wasm.c",
585 "src/f32-gemm/gen/4x4-wasm.c",
Marat Dukhanefaac272020-07-16 16:58:07 -0700586 "src/f32-hswish/gen/hswish-wasm-x1.c",
587 "src/f32-hswish/gen/hswish-wasm-x2.c",
588 "src/f32-hswish/gen/hswish-wasm-x4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700589 "src/f32-igemm/gen/1x4-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700590 "src/f32-igemm/gen/1x4-relu-wasm.c",
591 "src/f32-igemm/gen/1x4-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700592 "src/f32-igemm/gen/2x4-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700593 "src/f32-igemm/gen/2x4-relu-wasm.c",
594 "src/f32-igemm/gen/2x4-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700595 "src/f32-igemm/gen/4x2-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700596 "src/f32-igemm/gen/4x2-relu-wasm.c",
597 "src/f32-igemm/gen/4x2-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700598 "src/f32-igemm/gen/4x4-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700599 "src/f32-igemm/gen/4x4-relu-wasm.c",
600 "src/f32-igemm/gen/4x4-wasm.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700601 "src/f32-maxpool/9p8x-minmax-wasm-c1.c",
602 "src/f32-pavgpool/9p8x-minmax-wasm-c1.c",
603 "src/f32-pavgpool/9x-minmax-wasm-c1.c",
Marat Dukhan7c1f8082020-06-25 13:26:20 -0700604 "src/f32-prelu/gen/wasm-2x1.c",
605 "src/f32-prelu/gen/wasm-2x4.c",
Frank Barchardfb158e22020-07-15 16:10:10 -0700606 "src/f32-relu/gen/wasm-x1.c",
607 "src/f32-relu/gen/wasm-x2.c",
608 "src/f32-relu/gen/wasm-x4.c",
Frank Barchard37297a62020-08-27 12:15:10 -0700609 "src/f32-relu/gen/wasm-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700610 "src/f32-vbinary/gen/vadd-minmax-wasm-x1.c",
611 "src/f32-vbinary/gen/vadd-minmax-wasm-x2.c",
612 "src/f32-vbinary/gen/vadd-minmax-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700613 "src/f32-vbinary/gen/vadd-minmax-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700614 "src/f32-vbinary/gen/vadd-relu-wasm-x1.c",
615 "src/f32-vbinary/gen/vadd-relu-wasm-x2.c",
616 "src/f32-vbinary/gen/vadd-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700617 "src/f32-vbinary/gen/vadd-relu-wasm-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700618 "src/f32-vbinary/gen/vaddc-minmax-wasm-x1.c",
619 "src/f32-vbinary/gen/vaddc-minmax-wasm-x2.c",
620 "src/f32-vbinary/gen/vaddc-minmax-wasm-x4.c",
621 "src/f32-vbinary/gen/vaddc-minmax-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700622 "src/f32-vbinary/gen/vaddc-relu-wasm-x1.c",
623 "src/f32-vbinary/gen/vaddc-relu-wasm-x2.c",
624 "src/f32-vbinary/gen/vaddc-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700625 "src/f32-vbinary/gen/vaddc-relu-wasm-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700626 "src/f32-vbinary/gen/vdiv-minmax-wasm-x1.c",
627 "src/f32-vbinary/gen/vdiv-minmax-wasm-x2.c",
628 "src/f32-vbinary/gen/vdiv-minmax-wasm-x4.c",
629 "src/f32-vbinary/gen/vdiv-minmax-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700630 "src/f32-vbinary/gen/vdiv-relu-wasm-x1.c",
631 "src/f32-vbinary/gen/vdiv-relu-wasm-x2.c",
632 "src/f32-vbinary/gen/vdiv-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700633 "src/f32-vbinary/gen/vdiv-relu-wasm-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700634 "src/f32-vbinary/gen/vdivc-minmax-wasm-x1.c",
635 "src/f32-vbinary/gen/vdivc-minmax-wasm-x2.c",
636 "src/f32-vbinary/gen/vdivc-minmax-wasm-x4.c",
637 "src/f32-vbinary/gen/vdivc-minmax-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700638 "src/f32-vbinary/gen/vdivc-relu-wasm-x1.c",
639 "src/f32-vbinary/gen/vdivc-relu-wasm-x2.c",
640 "src/f32-vbinary/gen/vdivc-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700641 "src/f32-vbinary/gen/vdivc-relu-wasm-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800642 "src/f32-vbinary/gen/vmax-wasm-x1.c",
643 "src/f32-vbinary/gen/vmax-wasm-x2.c",
644 "src/f32-vbinary/gen/vmax-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700645 "src/f32-vbinary/gen/vmax-wasm-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800646 "src/f32-vbinary/gen/vmaxc-wasm-x1.c",
647 "src/f32-vbinary/gen/vmaxc-wasm-x2.c",
648 "src/f32-vbinary/gen/vmaxc-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700649 "src/f32-vbinary/gen/vmaxc-wasm-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800650 "src/f32-vbinary/gen/vmin-wasm-x1.c",
651 "src/f32-vbinary/gen/vmin-wasm-x2.c",
652 "src/f32-vbinary/gen/vmin-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700653 "src/f32-vbinary/gen/vmin-wasm-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800654 "src/f32-vbinary/gen/vminc-wasm-x1.c",
655 "src/f32-vbinary/gen/vminc-wasm-x2.c",
656 "src/f32-vbinary/gen/vminc-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700657 "src/f32-vbinary/gen/vminc-wasm-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700658 "src/f32-vbinary/gen/vmul-minmax-wasm-x1.c",
659 "src/f32-vbinary/gen/vmul-minmax-wasm-x2.c",
660 "src/f32-vbinary/gen/vmul-minmax-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700661 "src/f32-vbinary/gen/vmul-minmax-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700662 "src/f32-vbinary/gen/vmul-relu-wasm-x1.c",
663 "src/f32-vbinary/gen/vmul-relu-wasm-x2.c",
664 "src/f32-vbinary/gen/vmul-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700665 "src/f32-vbinary/gen/vmul-relu-wasm-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700666 "src/f32-vbinary/gen/vmulc-minmax-wasm-x1.c",
667 "src/f32-vbinary/gen/vmulc-minmax-wasm-x2.c",
668 "src/f32-vbinary/gen/vmulc-minmax-wasm-x4.c",
669 "src/f32-vbinary/gen/vmulc-minmax-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700670 "src/f32-vbinary/gen/vmulc-relu-wasm-x1.c",
671 "src/f32-vbinary/gen/vmulc-relu-wasm-x2.c",
672 "src/f32-vbinary/gen/vmulc-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700673 "src/f32-vbinary/gen/vmulc-relu-wasm-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700674 "src/f32-vbinary/gen/vrdivc-minmax-wasm-x1.c",
675 "src/f32-vbinary/gen/vrdivc-minmax-wasm-x2.c",
676 "src/f32-vbinary/gen/vrdivc-minmax-wasm-x4.c",
677 "src/f32-vbinary/gen/vrdivc-minmax-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700678 "src/f32-vbinary/gen/vrdivc-relu-wasm-x1.c",
679 "src/f32-vbinary/gen/vrdivc-relu-wasm-x2.c",
680 "src/f32-vbinary/gen/vrdivc-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700681 "src/f32-vbinary/gen/vrdivc-relu-wasm-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700682 "src/f32-vbinary/gen/vrsubc-minmax-wasm-x1.c",
683 "src/f32-vbinary/gen/vrsubc-minmax-wasm-x2.c",
684 "src/f32-vbinary/gen/vrsubc-minmax-wasm-x4.c",
685 "src/f32-vbinary/gen/vrsubc-minmax-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700686 "src/f32-vbinary/gen/vrsubc-relu-wasm-x1.c",
687 "src/f32-vbinary/gen/vrsubc-relu-wasm-x2.c",
688 "src/f32-vbinary/gen/vrsubc-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700689 "src/f32-vbinary/gen/vrsubc-relu-wasm-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700690 "src/f32-vbinary/gen/vsub-minmax-wasm-x1.c",
691 "src/f32-vbinary/gen/vsub-minmax-wasm-x2.c",
692 "src/f32-vbinary/gen/vsub-minmax-wasm-x4.c",
693 "src/f32-vbinary/gen/vsub-minmax-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700694 "src/f32-vbinary/gen/vsub-relu-wasm-x1.c",
695 "src/f32-vbinary/gen/vsub-relu-wasm-x2.c",
696 "src/f32-vbinary/gen/vsub-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700697 "src/f32-vbinary/gen/vsub-relu-wasm-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700698 "src/f32-vbinary/gen/vsubc-minmax-wasm-x1.c",
699 "src/f32-vbinary/gen/vsubc-minmax-wasm-x2.c",
700 "src/f32-vbinary/gen/vsubc-minmax-wasm-x4.c",
701 "src/f32-vbinary/gen/vsubc-minmax-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700702 "src/f32-vbinary/gen/vsubc-relu-wasm-x1.c",
703 "src/f32-vbinary/gen/vsubc-relu-wasm-x2.c",
704 "src/f32-vbinary/gen/vsubc-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700705 "src/f32-vbinary/gen/vsubc-relu-wasm-x8.c",
Marat Dukhaned6baaf2020-12-01 15:07:08 -0800706 "src/f32-velu/gen/velu-wasm-rr2-lut16-p3-x1.c",
707 "src/f32-velu/gen/velu-wasm-rr2-lut16-p3-x2.c",
708 "src/f32-velu/gen/velu-wasm-rr2-lut16-p3-x3.c",
709 "src/f32-velu/gen/velu-wasm-rr2-lut16-p3-x4.c",
710 "src/f32-velu/gen/velu-wasm-rr2-lut16-p3-x5.c",
711 "src/f32-velu/gen/velu-wasm-rr2-lut16-p3-x6.c",
712 "src/f32-velu/gen/velu-wasm-rr2-p6-x1.c",
713 "src/f32-velu/gen/velu-wasm-rr2-p6-x2.c",
714 "src/f32-velu/gen/velu-wasm-rr2-p6-x3.c",
715 "src/f32-velu/gen/velu-wasm-rr2-p6-x4.c",
716 "src/f32-velu/gen/velu-wasm-rr2-p6-x5.c",
717 "src/f32-velu/gen/velu-wasm-rr2-p6-x6.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -0700718 "src/f32-vlrelu/gen/vlrelu-wasm-x1.c",
719 "src/f32-vlrelu/gen/vlrelu-wasm-x2.c",
720 "src/f32-vlrelu/gen/vlrelu-wasm-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700721 "src/f32-vmulcaddc/gen/c1-minmax-wasm-2x.c",
722 "src/f32-vmulcaddc/gen/c2-minmax-wasm-2x.c",
723 "src/f32-vmulcaddc/gen/c4-minmax-wasm-2x.c",
Marat Dukhan436ebe62019-12-04 15:10:12 -0800724]
725
Marat Dukhan290055c2020-06-09 12:24:29 -0700726WASMSIMD_UKERNELS = [
Marat Dukhan40f05522020-07-16 22:33:12 -0700727 "src/f32-argmaxpool/4x-wasmsimd-c4.c",
728 "src/f32-argmaxpool/9p8x-wasmsimd-c4.c",
729 "src/f32-argmaxpool/9x-wasmsimd-c4.c",
Marat Dukhan3b7432d2020-07-16 17:46:32 -0700730 "src/f32-avgpool/9p8x-minmax-wasmsimd-arm-c4.c",
731 "src/f32-avgpool/9p8x-minmax-wasmsimd-x86-c4.c",
732 "src/f32-avgpool/9x-minmax-wasmsimd-arm-c4.c",
733 "src/f32-avgpool/9x-minmax-wasmsimd-x86-c4.c",
Marat Dukhan3fa52c82020-07-08 12:54:33 -0700734 "src/f32-clamp/gen/wasmsimd-arm-x4.c",
735 "src/f32-clamp/gen/wasmsimd-arm-x8.c",
736 "src/f32-clamp/gen/wasmsimd-x86-x4.c",
737 "src/f32-clamp/gen/wasmsimd-x86-x8.c",
Frank Barchard22136062020-11-24 18:44:46 -0800738 "src/f32-conv-hwc2chw/3x3s2p1c3x4-wasmsimd-2x2.c",
Frank Barchard0725b8d2020-12-07 11:07:35 -0800739 "src/f32-dwconv/gen/up4x4-minmax-wasmsimd-arm-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700740 "src/f32-dwconv/gen/up4x4-minmax-wasmsimd-arm.c",
Frank Barchard0725b8d2020-12-07 11:07:35 -0800741 "src/f32-dwconv/gen/up4x4-minmax-wasmsimd-x86-acc2.c",
Marat Dukhanac014d72020-06-16 08:36:47 -0700742 "src/f32-dwconv/gen/up4x4-minmax-wasmsimd-x86.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700743 "src/f32-dwconv/gen/up4x4-wasmsimd.c",
Frank Barchard0725b8d2020-12-07 11:07:35 -0800744 "src/f32-dwconv/gen/up4x9-minmax-wasmsimd-arm-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700745 "src/f32-dwconv/gen/up4x9-minmax-wasmsimd-arm.c",
Frank Barchard0725b8d2020-12-07 11:07:35 -0800746 "src/f32-dwconv/gen/up4x9-minmax-wasmsimd-x86-acc2.c",
Marat Dukhanac014d72020-06-16 08:36:47 -0700747 "src/f32-dwconv/gen/up4x9-minmax-wasmsimd-x86.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700748 "src/f32-dwconv/gen/up4x9-wasmsimd.c",
Frank Barchard0725b8d2020-12-07 11:07:35 -0800749 "src/f32-dwconv/gen/up4x25-minmax-wasmsimd-arm-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700750 "src/f32-dwconv/gen/up4x25-minmax-wasmsimd-arm.c",
Frank Barchard0725b8d2020-12-07 11:07:35 -0800751 "src/f32-dwconv/gen/up4x25-minmax-wasmsimd-x86-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700752 "src/f32-dwconv/gen/up4x25-minmax-wasmsimd-x86.c",
753 "src/f32-dwconv/gen/up4x25-wasmsimd.c",
Frank Barchard0725b8d2020-12-07 11:07:35 -0800754 "src/f32-dwconv/gen/up8x4-minmax-wasmsimd-arm-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700755 "src/f32-dwconv/gen/up8x4-minmax-wasmsimd-arm.c",
Frank Barchard0725b8d2020-12-07 11:07:35 -0800756 "src/f32-dwconv/gen/up8x4-minmax-wasmsimd-x86-acc2.c",
Marat Dukhanac014d72020-06-16 08:36:47 -0700757 "src/f32-dwconv/gen/up8x4-minmax-wasmsimd-x86.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700758 "src/f32-dwconv/gen/up8x4-wasmsimd.c",
Frank Barchard0725b8d2020-12-07 11:07:35 -0800759 "src/f32-dwconv/gen/up8x9-minmax-wasmsimd-arm-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700760 "src/f32-dwconv/gen/up8x9-minmax-wasmsimd-arm.c",
Frank Barchard0725b8d2020-12-07 11:07:35 -0800761 "src/f32-dwconv/gen/up8x9-minmax-wasmsimd-x86-acc2.c",
Marat Dukhanac014d72020-06-16 08:36:47 -0700762 "src/f32-dwconv/gen/up8x9-minmax-wasmsimd-x86.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700763 "src/f32-dwconv/gen/up8x9-wasmsimd.c",
Frank Barchard0725b8d2020-12-07 11:07:35 -0800764 "src/f32-dwconv/gen/up8x25-minmax-wasmsimd-arm-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700765 "src/f32-dwconv/gen/up8x25-minmax-wasmsimd-arm.c",
Frank Barchard0725b8d2020-12-07 11:07:35 -0800766 "src/f32-dwconv/gen/up8x25-minmax-wasmsimd-x86-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700767 "src/f32-dwconv/gen/up8x25-minmax-wasmsimd-x86.c",
768 "src/f32-dwconv/gen/up8x25-wasmsimd.c",
Frank Barchard412e2f42020-12-11 11:40:50 -0800769 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-loadsplat-1x4-acc2.c",
770 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-loadsplat-1x4-acc3.c",
771 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-loadsplat-1x4-acc4.c",
772 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-loadsplat-1x4.c",
773 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-loadsplat-2x4-acc2.c",
774 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-loadsplat-2x4.c",
775 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-loadsplat-3x4.c",
776 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-loadsplat-4x4.c",
777 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-loadsplat-5x4.c",
778 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-loadsplat-6x4.c",
Frank Barchard02bb4292020-12-15 18:25:32 -0800779 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-splat-1x4-acc2.c",
780 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-splat-1x4-acc3.c",
781 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-splat-1x4-acc4.c",
782 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-splat-1x4.c",
783 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-splat-2x4-acc2.c",
784 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-splat-2x4.c",
785 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-splat-3x4.c",
786 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-splat-4x4.c",
787 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-splat-5x4.c",
788 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-splat-6x4.c",
Frank Barchard412e2f42020-12-11 11:40:50 -0800789 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-loadsplat-1x4-acc2.c",
790 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-loadsplat-1x4-acc3.c",
791 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-loadsplat-1x4-acc4.c",
792 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-loadsplat-1x4.c",
793 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-loadsplat-2x4-acc2.c",
794 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-loadsplat-2x4.c",
795 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-loadsplat-3x4.c",
796 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-loadsplat-4x4.c",
797 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-loadsplat-5x4.c",
798 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-loadsplat-6x4.c",
Frank Barchard02bb4292020-12-15 18:25:32 -0800799 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-splat-1x4-acc2.c",
800 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-splat-1x4-acc3.c",
801 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-splat-1x4-acc4.c",
802 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-splat-1x4.c",
803 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-splat-2x4-acc2.c",
804 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-splat-2x4.c",
805 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-splat-3x4.c",
806 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-splat-4x4.c",
807 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-splat-5x4.c",
808 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-splat-6x4.c",
Frank Barchardc5704bf2020-12-21 23:09:00 -0800809 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-arm-loadsplat-1x4-acc2.c",
810 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-arm-loadsplat-1x4-acc3.c",
811 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-arm-loadsplat-1x4-acc4.c",
812 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-arm-loadsplat-1x4.c",
813 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-arm-loadsplat-2x4-acc2.c",
814 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-arm-loadsplat-2x4.c",
815 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-arm-loadsplat-3x4.c",
816 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-arm-loadsplat-4x4.c",
Frank Barchard412e2f42020-12-11 11:40:50 -0800817 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-arm-splat-1x4-acc2.c",
818 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-arm-splat-1x4-acc3.c",
819 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-arm-splat-1x4-acc4.c",
820 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-arm-splat-1x4.c",
821 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-arm-splat-2x4-acc2.c",
822 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-arm-splat-2x4.c",
823 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-arm-splat-3x4.c",
824 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-arm-splat-4x4.c",
Frank Barchardcadd4222021-01-20 16:27:25 -0800825 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-x86-loadsplat-1x4-acc2.c",
826 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-x86-loadsplat-1x4-acc3.c",
827 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-x86-loadsplat-1x4-acc4.c",
828 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-x86-loadsplat-1x4.c",
829 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-x86-loadsplat-2x4-acc2.c",
830 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-x86-loadsplat-2x4.c",
831 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-x86-loadsplat-3x4.c",
832 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-x86-loadsplat-4x4.c",
Frank Barchard412e2f42020-12-11 11:40:50 -0800833 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-x86-splat-1x4-acc2.c",
834 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-x86-splat-1x4-acc3.c",
835 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-x86-splat-1x4-acc4.c",
836 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-x86-splat-1x4.c",
837 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-x86-splat-2x4-acc2.c",
838 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-x86-splat-2x4.c",
839 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-x86-splat-3x4.c",
840 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-x86-splat-4x4.c",
Frank Barchardb20dcd62020-12-15 16:46:14 -0800841 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-loadsplat-1x4-acc2.c",
842 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-loadsplat-1x4-acc3.c",
843 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-loadsplat-1x4-acc4.c",
844 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-loadsplat-1x4-acc5.c",
845 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-loadsplat-1x4.c",
846 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-loadsplat-2x4-acc2.c",
847 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-loadsplat-2x4-acc3.c",
848 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-loadsplat-2x4.c",
849 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-loadsplat-3x4-acc2.c",
850 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-loadsplat-3x4.c",
851 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-loadsplat-4x4-acc2.c",
852 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-loadsplat-4x4.c",
853 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-loadsplat-5x4.c",
Frank Barchard412e2f42020-12-11 11:40:50 -0800854 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-splat-1x4-acc2.c",
855 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-splat-1x4-acc3.c",
856 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-splat-1x4-acc4.c",
857 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-splat-1x4-acc5.c",
858 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-splat-1x4.c",
859 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-splat-2x4-acc2.c",
860 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-splat-2x4-acc3.c",
861 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-splat-2x4.c",
862 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-splat-3x4-acc2.c",
863 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-splat-3x4.c",
864 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-splat-4x4-acc2.c",
865 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-splat-4x4.c",
866 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-splat-5x4.c",
Frank Barchardb20dcd62020-12-15 16:46:14 -0800867 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-loadsplat-1x4-acc2.c",
868 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-loadsplat-1x4-acc3.c",
869 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-loadsplat-1x4-acc4.c",
870 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-loadsplat-1x4-acc5.c",
871 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-loadsplat-1x4.c",
872 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-loadsplat-2x4-acc2.c",
873 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-loadsplat-2x4-acc3.c",
874 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-loadsplat-2x4.c",
875 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-loadsplat-3x4-acc2.c",
876 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-loadsplat-3x4.c",
877 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-loadsplat-4x4-acc2.c",
878 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-loadsplat-4x4.c",
879 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-loadsplat-5x4.c",
Frank Barchard412e2f42020-12-11 11:40:50 -0800880 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-splat-1x4-acc2.c",
881 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-splat-1x4-acc3.c",
882 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-splat-1x4-acc4.c",
883 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-splat-1x4-acc5.c",
884 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-splat-1x4.c",
885 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-splat-2x4-acc2.c",
886 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-splat-2x4-acc3.c",
887 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-splat-2x4.c",
888 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-splat-3x4-acc2.c",
889 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-splat-3x4.c",
890 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-splat-4x4-acc2.c",
891 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-splat-4x4.c",
892 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-splat-5x4.c",
Frank Barchardc6889b32020-12-21 11:27:22 -0800893 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-arm-loadsplat-1x4-acc2.c",
894 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-arm-loadsplat-1x4-acc3.c",
895 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-arm-loadsplat-1x4-acc4.c",
896 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-arm-loadsplat-1x4-acc5.c",
897 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-arm-loadsplat-1x4.c",
898 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-arm-loadsplat-2x4-acc2.c",
899 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-arm-loadsplat-2x4-acc3.c",
900 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-arm-loadsplat-2x4.c",
901 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-arm-loadsplat-3x4-acc2.c",
902 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-arm-loadsplat-3x4.c",
Frank Barchard412e2f42020-12-11 11:40:50 -0800903 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-arm-splat-1x4-acc2.c",
904 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-arm-splat-1x4-acc3.c",
905 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-arm-splat-1x4-acc4.c",
906 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-arm-splat-1x4-acc5.c",
907 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-arm-splat-1x4.c",
908 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-arm-splat-2x4-acc2.c",
909 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-arm-splat-2x4-acc3.c",
910 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-arm-splat-2x4.c",
911 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-arm-splat-3x4-acc2.c",
912 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-arm-splat-3x4.c",
Frank Barchardc6889b32020-12-21 11:27:22 -0800913 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-x86-loadsplat-1x4-acc2.c",
914 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-x86-loadsplat-1x4-acc3.c",
915 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-x86-loadsplat-1x4-acc4.c",
916 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-x86-loadsplat-1x4-acc5.c",
917 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-x86-loadsplat-1x4.c",
918 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-x86-loadsplat-2x4-acc2.c",
919 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-x86-loadsplat-2x4-acc3.c",
920 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-x86-loadsplat-2x4.c",
921 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-x86-loadsplat-3x4-acc2.c",
922 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-x86-loadsplat-3x4.c",
Frank Barchard412e2f42020-12-11 11:40:50 -0800923 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-x86-splat-1x4-acc2.c",
924 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-x86-splat-1x4-acc3.c",
925 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-x86-splat-1x4-acc4.c",
926 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-x86-splat-1x4-acc5.c",
927 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-x86-splat-1x4.c",
928 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-x86-splat-2x4-acc2.c",
929 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-x86-splat-2x4-acc3.c",
930 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-x86-splat-2x4.c",
931 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-x86-splat-3x4-acc2.c",
932 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-x86-splat-3x4.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700933 "src/f32-gavgpool-cw/wasmsimd-arm-x4.c",
934 "src/f32-gavgpool-cw/wasmsimd-x86-x4.c",
Marat Dukhanc6016802020-07-16 18:51:28 -0700935 "src/f32-gavgpool/7p7x-minmax-wasmsimd-arm-c4.c",
936 "src/f32-gavgpool/7p7x-minmax-wasmsimd-x86-c4.c",
937 "src/f32-gavgpool/7x-minmax-wasmsimd-arm-c4.c",
938 "src/f32-gavgpool/7x-minmax-wasmsimd-x86-c4.c",
Frank Barchard0725b8d2020-12-07 11:07:35 -0800939 "src/f32-gemm/gen-inc/1x8inc-minmax-wasmsimd-arm-loadsplat.c",
940 "src/f32-gemm/gen-inc/1x8inc-minmax-wasmsimd-arm-splat.c",
941 "src/f32-gemm/gen-inc/1x8inc-minmax-wasmsimd-x86-loadsplat.c",
942 "src/f32-gemm/gen-inc/1x8inc-minmax-wasmsimd-x86-splat.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700943 "src/f32-gemm/gen-inc/1x8s4inc-minmax-wasmsimd-arm.c",
944 "src/f32-gemm/gen-inc/1x8s4inc-minmax-wasmsimd-x86.c",
Frank Barchard0725b8d2020-12-07 11:07:35 -0800945 "src/f32-gemm/gen-inc/3x8inc-minmax-wasmsimd-arm-loadsplat.c",
946 "src/f32-gemm/gen-inc/3x8inc-minmax-wasmsimd-arm-splat.c",
947 "src/f32-gemm/gen-inc/3x8inc-minmax-wasmsimd-x86-loadsplat.c",
948 "src/f32-gemm/gen-inc/3x8inc-minmax-wasmsimd-x86-splat.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700949 "src/f32-gemm/gen-inc/3x8s4inc-minmax-wasmsimd-arm.c",
950 "src/f32-gemm/gen-inc/3x8s4inc-minmax-wasmsimd-x86.c",
Frank Barchard0725b8d2020-12-07 11:07:35 -0800951 "src/f32-gemm/gen-inc/4x8inc-minmax-wasmsimd-arm-loadsplat.c",
952 "src/f32-gemm/gen-inc/4x8inc-minmax-wasmsimd-arm-splat.c",
953 "src/f32-gemm/gen-inc/4x8inc-minmax-wasmsimd-x86-loadsplat.c",
954 "src/f32-gemm/gen-inc/4x8inc-minmax-wasmsimd-x86-splat.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700955 "src/f32-gemm/gen-inc/4x8s4inc-minmax-wasmsimd-arm.c",
956 "src/f32-gemm/gen-inc/4x8s4inc-minmax-wasmsimd-x86.c",
Frank Barchard0725b8d2020-12-07 11:07:35 -0800957 "src/f32-gemm/gen-inc/5x8inc-minmax-wasmsimd-arm-loadsplat.c",
958 "src/f32-gemm/gen-inc/5x8inc-minmax-wasmsimd-arm-splat.c",
959 "src/f32-gemm/gen-inc/5x8inc-minmax-wasmsimd-x86-loadsplat.c",
960 "src/f32-gemm/gen-inc/5x8inc-minmax-wasmsimd-x86-splat.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700961 "src/f32-gemm/gen-inc/5x8s4inc-minmax-wasmsimd-arm.c",
962 "src/f32-gemm/gen-inc/5x8s4inc-minmax-wasmsimd-x86.c",
Frank Barchard0725b8d2020-12-07 11:07:35 -0800963 "src/f32-gemm/gen-inc/6x8inc-minmax-wasmsimd-arm-loadsplat.c",
964 "src/f32-gemm/gen-inc/6x8inc-minmax-wasmsimd-arm-splat.c",
965 "src/f32-gemm/gen-inc/6x8inc-minmax-wasmsimd-x86-loadsplat.c",
966 "src/f32-gemm/gen-inc/6x8inc-minmax-wasmsimd-x86-splat.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700967 "src/f32-gemm/gen-inc/6x8s4inc-minmax-wasmsimd-arm.c",
968 "src/f32-gemm/gen-inc/6x8s4inc-minmax-wasmsimd-x86.c",
Frank Barchard0725b8d2020-12-07 11:07:35 -0800969 "src/f32-gemm/gen/1x8-minmax-wasmsimd-arm-loadsplat.c",
970 "src/f32-gemm/gen/1x8-minmax-wasmsimd-arm-splat.c",
971 "src/f32-gemm/gen/1x8-minmax-wasmsimd-x86-loadsplat.c",
972 "src/f32-gemm/gen/1x8-minmax-wasmsimd-x86-splat.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700973 "src/f32-gemm/gen/1x8-relu-wasmsimd-splat.c",
974 "src/f32-gemm/gen/1x8-wasmsimd-splat.c",
975 "src/f32-gemm/gen/1x8s4-minmax-wasmsimd-arm.c",
976 "src/f32-gemm/gen/1x8s4-minmax-wasmsimd-x86.c",
Frank Barchard0725b8d2020-12-07 11:07:35 -0800977 "src/f32-gemm/gen/3x8-minmax-wasmsimd-arm-loadsplat.c",
978 "src/f32-gemm/gen/3x8-minmax-wasmsimd-arm-splat.c",
979 "src/f32-gemm/gen/3x8-minmax-wasmsimd-x86-loadsplat.c",
980 "src/f32-gemm/gen/3x8-minmax-wasmsimd-x86-splat.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700981 "src/f32-gemm/gen/3x8s4-minmax-wasmsimd-arm.c",
982 "src/f32-gemm/gen/3x8s4-minmax-wasmsimd-x86.c",
983 "src/f32-gemm/gen/4x2c4-minmax-wasmsimd-arm.c",
984 "src/f32-gemm/gen/4x2c4-minmax-wasmsimd-x86.c",
985 "src/f32-gemm/gen/4x2c4-relu-wasmsimd.c",
986 "src/f32-gemm/gen/4x2c4-wasmsimd.c",
Frank Barchard0725b8d2020-12-07 11:07:35 -0800987 "src/f32-gemm/gen/4x8-minmax-wasmsimd-arm-loadsplat.c",
988 "src/f32-gemm/gen/4x8-minmax-wasmsimd-arm-splat.c",
989 "src/f32-gemm/gen/4x8-minmax-wasmsimd-x86-loadsplat.c",
990 "src/f32-gemm/gen/4x8-minmax-wasmsimd-x86-splat.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700991 "src/f32-gemm/gen/4x8-relu-wasmsimd-splat.c",
992 "src/f32-gemm/gen/4x8-wasmsimd-splat.c",
993 "src/f32-gemm/gen/4x8s4-minmax-wasmsimd-arm.c",
994 "src/f32-gemm/gen/4x8s4-minmax-wasmsimd-x86.c",
Frank Barchard0725b8d2020-12-07 11:07:35 -0800995 "src/f32-gemm/gen/5x8-minmax-wasmsimd-arm-loadsplat.c",
996 "src/f32-gemm/gen/5x8-minmax-wasmsimd-arm-splat.c",
997 "src/f32-gemm/gen/5x8-minmax-wasmsimd-x86-loadsplat.c",
998 "src/f32-gemm/gen/5x8-minmax-wasmsimd-x86-splat.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700999 "src/f32-gemm/gen/5x8-relu-wasmsimd-splat.c",
1000 "src/f32-gemm/gen/5x8-wasmsimd-splat.c",
1001 "src/f32-gemm/gen/5x8s4-minmax-wasmsimd-arm.c",
1002 "src/f32-gemm/gen/5x8s4-minmax-wasmsimd-x86.c",
Frank Barchard0725b8d2020-12-07 11:07:35 -08001003 "src/f32-gemm/gen/6x8-minmax-wasmsimd-arm-loadsplat.c",
1004 "src/f32-gemm/gen/6x8-minmax-wasmsimd-arm-splat.c",
1005 "src/f32-gemm/gen/6x8-minmax-wasmsimd-x86-loadsplat.c",
1006 "src/f32-gemm/gen/6x8-minmax-wasmsimd-x86-splat.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001007 "src/f32-gemm/gen/6x8s4-minmax-wasmsimd-arm.c",
1008 "src/f32-gemm/gen/6x8s4-minmax-wasmsimd-x86.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07001009 "src/f32-hswish/gen/hswish-wasmsimd-x4.c",
1010 "src/f32-hswish/gen/hswish-wasmsimd-x8.c",
1011 "src/f32-hswish/gen/hswish-wasmsimd-x16.c",
XNNPACK Team965272b2020-10-23 21:10:15 -07001012 "src/f32-ibilinear-chw/gen/wasmsimd-p4.c",
1013 "src/f32-ibilinear-chw/gen/wasmsimd-p8.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001014 "src/f32-ibilinear/gen/wasmsimd-c4.c",
1015 "src/f32-ibilinear/gen/wasmsimd-c8.c",
Frank Barchard0725b8d2020-12-07 11:07:35 -08001016 "src/f32-igemm/gen/1x8-minmax-wasmsimd-arm-loadsplat.c",
1017 "src/f32-igemm/gen/1x8-minmax-wasmsimd-arm-splat.c",
1018 "src/f32-igemm/gen/1x8-minmax-wasmsimd-x86-loadsplat.c",
1019 "src/f32-igemm/gen/1x8-minmax-wasmsimd-x86-splat.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001020 "src/f32-igemm/gen/1x8-relu-wasmsimd-splat.c",
1021 "src/f32-igemm/gen/1x8-wasmsimd-splat.c",
1022 "src/f32-igemm/gen/1x8s4-minmax-wasmsimd-arm.c",
1023 "src/f32-igemm/gen/1x8s4-minmax-wasmsimd-x86.c",
Frank Barchard0725b8d2020-12-07 11:07:35 -08001024 "src/f32-igemm/gen/3x8-minmax-wasmsimd-arm-loadsplat.c",
1025 "src/f32-igemm/gen/3x8-minmax-wasmsimd-arm-splat.c",
1026 "src/f32-igemm/gen/3x8-minmax-wasmsimd-x86-loadsplat.c",
1027 "src/f32-igemm/gen/3x8-minmax-wasmsimd-x86-splat.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001028 "src/f32-igemm/gen/3x8s4-minmax-wasmsimd-arm.c",
1029 "src/f32-igemm/gen/3x8s4-minmax-wasmsimd-x86.c",
1030 "src/f32-igemm/gen/4x2c4-minmax-wasmsimd-arm.c",
1031 "src/f32-igemm/gen/4x2c4-minmax-wasmsimd-x86.c",
1032 "src/f32-igemm/gen/4x2c4-relu-wasmsimd.c",
1033 "src/f32-igemm/gen/4x2c4-wasmsimd.c",
Frank Barchard0725b8d2020-12-07 11:07:35 -08001034 "src/f32-igemm/gen/4x8-minmax-wasmsimd-arm-loadsplat.c",
1035 "src/f32-igemm/gen/4x8-minmax-wasmsimd-arm-splat.c",
1036 "src/f32-igemm/gen/4x8-minmax-wasmsimd-x86-loadsplat.c",
1037 "src/f32-igemm/gen/4x8-minmax-wasmsimd-x86-splat.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001038 "src/f32-igemm/gen/4x8-relu-wasmsimd-splat.c",
1039 "src/f32-igemm/gen/4x8-wasmsimd-splat.c",
1040 "src/f32-igemm/gen/4x8s4-minmax-wasmsimd-arm.c",
1041 "src/f32-igemm/gen/4x8s4-minmax-wasmsimd-x86.c",
Frank Barchard0725b8d2020-12-07 11:07:35 -08001042 "src/f32-igemm/gen/5x8-minmax-wasmsimd-arm-loadsplat.c",
1043 "src/f32-igemm/gen/5x8-minmax-wasmsimd-arm-splat.c",
1044 "src/f32-igemm/gen/5x8-minmax-wasmsimd-x86-loadsplat.c",
1045 "src/f32-igemm/gen/5x8-minmax-wasmsimd-x86-splat.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001046 "src/f32-igemm/gen/5x8-relu-wasmsimd-splat.c",
1047 "src/f32-igemm/gen/5x8-wasmsimd-splat.c",
1048 "src/f32-igemm/gen/5x8s4-minmax-wasmsimd-arm.c",
1049 "src/f32-igemm/gen/5x8s4-minmax-wasmsimd-x86.c",
Frank Barchard0725b8d2020-12-07 11:07:35 -08001050 "src/f32-igemm/gen/6x8-minmax-wasmsimd-arm-loadsplat.c",
1051 "src/f32-igemm/gen/6x8-minmax-wasmsimd-arm-splat.c",
1052 "src/f32-igemm/gen/6x8-minmax-wasmsimd-x86-loadsplat.c",
1053 "src/f32-igemm/gen/6x8-minmax-wasmsimd-x86-splat.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001054 "src/f32-igemm/gen/6x8s4-minmax-wasmsimd-arm.c",
1055 "src/f32-igemm/gen/6x8s4-minmax-wasmsimd-x86.c",
Marat Dukhanf6e24802020-07-08 22:20:40 -07001056 "src/f32-maxpool/9p8x-minmax-wasmsimd-arm-c4.c",
1057 "src/f32-maxpool/9p8x-minmax-wasmsimd-x86-c4.c",
Marat Dukhan1483c532020-07-16 18:08:19 -07001058 "src/f32-pavgpool/9p8x-minmax-wasmsimd-arm-c4.c",
1059 "src/f32-pavgpool/9p8x-minmax-wasmsimd-x86-c4.c",
1060 "src/f32-pavgpool/9x-minmax-wasmsimd-arm-c4.c",
1061 "src/f32-pavgpool/9x-minmax-wasmsimd-x86-c4.c",
Frank Barchard0725b8d2020-12-07 11:07:35 -08001062 "src/f32-ppmm/gen/4x8-minmax-wasmsimd-arm-splat.c",
1063 "src/f32-ppmm/gen/4x8-minmax-wasmsimd-x86-splat.c",
Frank Barcharda5316982020-07-23 13:19:28 -07001064 "src/f32-prelu/gen/wasmsimd-bitselect-1x4.c",
1065 "src/f32-prelu/gen/wasmsimd-bitselect-1x8.c",
1066 "src/f32-prelu/gen/wasmsimd-bitselect-1x16.c",
Marat Dukhan195f8eb2020-06-25 12:50:57 -07001067 "src/f32-prelu/gen/wasmsimd-bitselect-2x4.c",
1068 "src/f32-prelu/gen/wasmsimd-bitselect-2x8.c",
Frank Barcharda5316982020-07-23 13:19:28 -07001069 "src/f32-prelu/gen/wasmsimd-bitselect-2x16.c",
1070 "src/f32-prelu/gen/wasmsimd-bitselect-4x4.c",
1071 "src/f32-prelu/gen/wasmsimd-bitselect-4x8.c",
1072 "src/f32-prelu/gen/wasmsimd-bitselect-4x16.c",
1073 "src/f32-prelu/gen/wasmsimd-minmax-1x4.c",
1074 "src/f32-prelu/gen/wasmsimd-minmax-1x8.c",
1075 "src/f32-prelu/gen/wasmsimd-minmax-1x16.c",
Marat Dukhan195f8eb2020-06-25 12:50:57 -07001076 "src/f32-prelu/gen/wasmsimd-minmax-2x4.c",
1077 "src/f32-prelu/gen/wasmsimd-minmax-2x8.c",
Frank Barcharda5316982020-07-23 13:19:28 -07001078 "src/f32-prelu/gen/wasmsimd-minmax-2x16.c",
1079 "src/f32-prelu/gen/wasmsimd-minmax-4x4.c",
1080 "src/f32-prelu/gen/wasmsimd-minmax-4x8.c",
1081 "src/f32-prelu/gen/wasmsimd-minmax-4x16.c",
Marat Dukhan52238f02020-07-16 15:30:28 -07001082 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x4.c",
Marat Dukhan52238f02020-07-16 15:30:28 -07001083 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x8-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001084 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x8.c",
Marat Dukhan52238f02020-07-16 15:30:28 -07001085 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x12-acc2.c",
1086 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x12-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001087 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x12.c",
Marat Dukhan52238f02020-07-16 15:30:28 -07001088 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x16-acc2.c",
1089 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x16-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001090 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x16.c",
Marat Dukhan52238f02020-07-16 15:30:28 -07001091 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x20-acc2.c",
1092 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x20-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001093 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x20.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07001094 "src/f32-relu/gen/wasmsimd-x4.c",
1095 "src/f32-relu/gen/wasmsimd-x8.c",
Frank Barchard37297a62020-08-27 12:15:10 -07001096 "src/f32-relu/gen/wasmsimd-x16.c",
Marat Dukhan8c417962020-07-08 12:27:50 -07001097 "src/f32-rmax/wasmsimd-arm.c",
1098 "src/f32-rmax/wasmsimd-x86.c",
Marat Dukhand187a5b2020-07-20 01:07:17 -07001099 "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x4.c",
1100 "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x8.c",
1101 "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x12.c",
1102 "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x16.c",
1103 "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x20.c",
1104 "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x24.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001105 "src/f32-sigmoid/gen/wasmsimd-p5-div-x4.c",
1106 "src/f32-sigmoid/gen/wasmsimd-p5-div-x8.c",
1107 "src/f32-sigmoid/gen/wasmsimd-p5-div-x12.c",
1108 "src/f32-sigmoid/gen/wasmsimd-p5-div-x16.c",
1109 "src/f32-sigmoid/gen/wasmsimd-p5-div-x20.c",
1110 "src/f32-sigmoid/gen/wasmsimd-p5-div-x24.c",
Frank Barchard8ef44cd2020-11-03 12:30:23 -08001111 "src/f32-spmm/gen/4x1-minmax-wasmsimd-arm-pipelined-x2.c",
1112 "src/f32-spmm/gen/4x1-minmax-wasmsimd-arm-pipelined.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07001113 "src/f32-spmm/gen/4x1-minmax-wasmsimd-arm-x2.c",
1114 "src/f32-spmm/gen/4x1-minmax-wasmsimd-arm-x4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001115 "src/f32-spmm/gen/4x1-minmax-wasmsimd-arm.c",
Frank Barchard8ef44cd2020-11-03 12:30:23 -08001116 "src/f32-spmm/gen/4x1-minmax-wasmsimd-x86-pipelined-x2.c",
1117 "src/f32-spmm/gen/4x1-minmax-wasmsimd-x86-pipelined.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07001118 "src/f32-spmm/gen/4x1-minmax-wasmsimd-x86-x2.c",
1119 "src/f32-spmm/gen/4x1-minmax-wasmsimd-x86-x4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001120 "src/f32-spmm/gen/4x1-minmax-wasmsimd-x86.c",
Frank Barchard8ef44cd2020-11-03 12:30:23 -08001121 "src/f32-spmm/gen/8x1-minmax-wasmsimd-arm-pipelined-x2.c",
1122 "src/f32-spmm/gen/8x1-minmax-wasmsimd-arm-pipelined.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07001123 "src/f32-spmm/gen/8x1-minmax-wasmsimd-arm-x2.c",
1124 "src/f32-spmm/gen/8x1-minmax-wasmsimd-arm-x4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001125 "src/f32-spmm/gen/8x1-minmax-wasmsimd-arm.c",
Frank Barchard8ef44cd2020-11-03 12:30:23 -08001126 "src/f32-spmm/gen/8x1-minmax-wasmsimd-x86-pipelined-x2.c",
1127 "src/f32-spmm/gen/8x1-minmax-wasmsimd-x86-pipelined.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07001128 "src/f32-spmm/gen/8x1-minmax-wasmsimd-x86-x2.c",
1129 "src/f32-spmm/gen/8x1-minmax-wasmsimd-x86-x4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001130 "src/f32-spmm/gen/8x1-minmax-wasmsimd-x86.c",
Frank Barchard8ef44cd2020-11-03 12:30:23 -08001131 "src/f32-spmm/gen/16x1-minmax-wasmsimd-arm-pipelined-x2.c",
1132 "src/f32-spmm/gen/16x1-minmax-wasmsimd-arm-pipelined.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07001133 "src/f32-spmm/gen/16x1-minmax-wasmsimd-arm-x2.c",
1134 "src/f32-spmm/gen/16x1-minmax-wasmsimd-arm-x4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001135 "src/f32-spmm/gen/16x1-minmax-wasmsimd-arm.c",
Frank Barchard8ef44cd2020-11-03 12:30:23 -08001136 "src/f32-spmm/gen/16x1-minmax-wasmsimd-x86-pipelined-x2.c",
1137 "src/f32-spmm/gen/16x1-minmax-wasmsimd-x86-pipelined.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07001138 "src/f32-spmm/gen/16x1-minmax-wasmsimd-x86-x2.c",
1139 "src/f32-spmm/gen/16x1-minmax-wasmsimd-x86-x4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001140 "src/f32-spmm/gen/16x1-minmax-wasmsimd-x86.c",
Frank Barchard8ef44cd2020-11-03 12:30:23 -08001141 "src/f32-spmm/gen/32x1-minmax-wasmsimd-arm-pipelined-x2.c",
1142 "src/f32-spmm/gen/32x1-minmax-wasmsimd-arm-pipelined.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07001143 "src/f32-spmm/gen/32x1-minmax-wasmsimd-arm-x2.c",
1144 "src/f32-spmm/gen/32x1-minmax-wasmsimd-arm-x4.c",
Frank Barchard846c0c62020-10-26 15:01:39 -07001145 "src/f32-spmm/gen/32x1-minmax-wasmsimd-arm.c",
Frank Barchard8ef44cd2020-11-03 12:30:23 -08001146 "src/f32-spmm/gen/32x1-minmax-wasmsimd-x86-pipelined-x2.c",
1147 "src/f32-spmm/gen/32x1-minmax-wasmsimd-x86-pipelined.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07001148 "src/f32-spmm/gen/32x1-minmax-wasmsimd-x86-x2.c",
1149 "src/f32-spmm/gen/32x1-minmax-wasmsimd-x86-x4.c",
Frank Barchard846c0c62020-10-26 15:01:39 -07001150 "src/f32-spmm/gen/32x1-minmax-wasmsimd-x86.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -07001151 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-arm-x4.c",
1152 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001153 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-arm-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -07001154 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-x86-x4.c",
1155 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001156 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -07001157 "src/f32-vbinary/gen/vadd-relu-wasmsimd-x4.c",
1158 "src/f32-vbinary/gen/vadd-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001159 "src/f32-vbinary/gen/vadd-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07001160 "src/f32-vbinary/gen/vadd-wasmsimd-x4.c",
1161 "src/f32-vbinary/gen/vadd-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001162 "src/f32-vbinary/gen/vadd-wasmsimd-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001163 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-arm-x4.c",
1164 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001165 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-arm-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001166 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-x86-x4.c",
1167 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001168 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -07001169 "src/f32-vbinary/gen/vaddc-relu-wasmsimd-x4.c",
1170 "src/f32-vbinary/gen/vaddc-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001171 "src/f32-vbinary/gen/vaddc-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07001172 "src/f32-vbinary/gen/vaddc-wasmsimd-x4.c",
1173 "src/f32-vbinary/gen/vaddc-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001174 "src/f32-vbinary/gen/vaddc-wasmsimd-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -07001175 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-arm-x4.c",
1176 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001177 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-arm-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -07001178 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-x86-x4.c",
1179 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001180 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -07001181 "src/f32-vbinary/gen/vdiv-relu-wasmsimd-x4.c",
1182 "src/f32-vbinary/gen/vdiv-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001183 "src/f32-vbinary/gen/vdiv-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07001184 "src/f32-vbinary/gen/vdiv-wasmsimd-x4.c",
1185 "src/f32-vbinary/gen/vdiv-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001186 "src/f32-vbinary/gen/vdiv-wasmsimd-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001187 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-arm-x4.c",
1188 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001189 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-arm-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001190 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-x86-x4.c",
1191 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001192 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -07001193 "src/f32-vbinary/gen/vdivc-relu-wasmsimd-x4.c",
1194 "src/f32-vbinary/gen/vdivc-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001195 "src/f32-vbinary/gen/vdivc-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07001196 "src/f32-vbinary/gen/vdivc-wasmsimd-x4.c",
1197 "src/f32-vbinary/gen/vdivc-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001198 "src/f32-vbinary/gen/vdivc-wasmsimd-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -07001199 "src/f32-vbinary/gen/vmax-wasmsimd-arm-x4.c",
1200 "src/f32-vbinary/gen/vmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001201 "src/f32-vbinary/gen/vmax-wasmsimd-arm-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -07001202 "src/f32-vbinary/gen/vmax-wasmsimd-x86-x4.c",
1203 "src/f32-vbinary/gen/vmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001204 "src/f32-vbinary/gen/vmax-wasmsimd-x86-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001205 "src/f32-vbinary/gen/vmaxc-wasmsimd-arm-x4.c",
1206 "src/f32-vbinary/gen/vmaxc-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001207 "src/f32-vbinary/gen/vmaxc-wasmsimd-arm-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001208 "src/f32-vbinary/gen/vmaxc-wasmsimd-x86-x4.c",
1209 "src/f32-vbinary/gen/vmaxc-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001210 "src/f32-vbinary/gen/vmaxc-wasmsimd-x86-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -07001211 "src/f32-vbinary/gen/vmin-wasmsimd-arm-x4.c",
1212 "src/f32-vbinary/gen/vmin-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001213 "src/f32-vbinary/gen/vmin-wasmsimd-arm-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -07001214 "src/f32-vbinary/gen/vmin-wasmsimd-x86-x4.c",
1215 "src/f32-vbinary/gen/vmin-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001216 "src/f32-vbinary/gen/vmin-wasmsimd-x86-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001217 "src/f32-vbinary/gen/vminc-wasmsimd-arm-x4.c",
1218 "src/f32-vbinary/gen/vminc-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001219 "src/f32-vbinary/gen/vminc-wasmsimd-arm-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001220 "src/f32-vbinary/gen/vminc-wasmsimd-x86-x4.c",
1221 "src/f32-vbinary/gen/vminc-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001222 "src/f32-vbinary/gen/vminc-wasmsimd-x86-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -07001223 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-arm-x4.c",
1224 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001225 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-arm-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -07001226 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-x86-x4.c",
1227 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001228 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -07001229 "src/f32-vbinary/gen/vmul-relu-wasmsimd-x4.c",
1230 "src/f32-vbinary/gen/vmul-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001231 "src/f32-vbinary/gen/vmul-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07001232 "src/f32-vbinary/gen/vmul-wasmsimd-x4.c",
1233 "src/f32-vbinary/gen/vmul-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001234 "src/f32-vbinary/gen/vmul-wasmsimd-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001235 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-arm-x4.c",
1236 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001237 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-arm-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001238 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-x86-x4.c",
1239 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001240 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -07001241 "src/f32-vbinary/gen/vmulc-relu-wasmsimd-x4.c",
1242 "src/f32-vbinary/gen/vmulc-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001243 "src/f32-vbinary/gen/vmulc-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07001244 "src/f32-vbinary/gen/vmulc-wasmsimd-x4.c",
1245 "src/f32-vbinary/gen/vmulc-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001246 "src/f32-vbinary/gen/vmulc-wasmsimd-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001247 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-arm-x4.c",
1248 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001249 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-arm-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001250 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-x86-x4.c",
1251 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001252 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -07001253 "src/f32-vbinary/gen/vrdivc-relu-wasmsimd-x4.c",
1254 "src/f32-vbinary/gen/vrdivc-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001255 "src/f32-vbinary/gen/vrdivc-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07001256 "src/f32-vbinary/gen/vrdivc-wasmsimd-x4.c",
1257 "src/f32-vbinary/gen/vrdivc-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001258 "src/f32-vbinary/gen/vrdivc-wasmsimd-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001259 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-arm-x4.c",
1260 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001261 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-arm-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001262 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-x86-x4.c",
1263 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001264 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -07001265 "src/f32-vbinary/gen/vrsubc-relu-wasmsimd-x4.c",
1266 "src/f32-vbinary/gen/vrsubc-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001267 "src/f32-vbinary/gen/vrsubc-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07001268 "src/f32-vbinary/gen/vrsubc-wasmsimd-x4.c",
1269 "src/f32-vbinary/gen/vrsubc-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001270 "src/f32-vbinary/gen/vrsubc-wasmsimd-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -07001271 "src/f32-vbinary/gen/vsqrdiff-wasmsimd-x4.c",
1272 "src/f32-vbinary/gen/vsqrdiff-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001273 "src/f32-vbinary/gen/vsqrdiff-wasmsimd-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001274 "src/f32-vbinary/gen/vsqrdiffc-wasmsimd-x4.c",
1275 "src/f32-vbinary/gen/vsqrdiffc-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001276 "src/f32-vbinary/gen/vsqrdiffc-wasmsimd-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -07001277 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-arm-x4.c",
1278 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001279 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-arm-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -07001280 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-x86-x4.c",
1281 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001282 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -07001283 "src/f32-vbinary/gen/vsub-relu-wasmsimd-x4.c",
1284 "src/f32-vbinary/gen/vsub-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001285 "src/f32-vbinary/gen/vsub-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07001286 "src/f32-vbinary/gen/vsub-wasmsimd-x4.c",
1287 "src/f32-vbinary/gen/vsub-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001288 "src/f32-vbinary/gen/vsub-wasmsimd-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001289 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-arm-x4.c",
1290 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001291 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-arm-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001292 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-x86-x4.c",
1293 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001294 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -07001295 "src/f32-vbinary/gen/vsubc-relu-wasmsimd-x4.c",
1296 "src/f32-vbinary/gen/vsubc-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001297 "src/f32-vbinary/gen/vsubc-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07001298 "src/f32-vbinary/gen/vsubc-wasmsimd-x4.c",
1299 "src/f32-vbinary/gen/vsubc-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001300 "src/f32-vbinary/gen/vsubc-wasmsimd-x16.c",
Marat Dukhaned6baaf2020-12-01 15:07:08 -08001301 "src/f32-velu/gen/velu-wasmsimd-arm-rr2-lut16-p3-x4.c",
1302 "src/f32-velu/gen/velu-wasmsimd-arm-rr2-lut16-p3-x8.c",
1303 "src/f32-velu/gen/velu-wasmsimd-arm-rr2-lut16-p3-x12.c",
1304 "src/f32-velu/gen/velu-wasmsimd-arm-rr2-lut16-p3-x16.c",
1305 "src/f32-velu/gen/velu-wasmsimd-arm-rr2-lut16-p3-x20.c",
1306 "src/f32-velu/gen/velu-wasmsimd-arm-rr2-lut16-p3-x24.c",
Marat Dukhaned6baaf2020-12-01 15:07:08 -08001307 "src/f32-velu/gen/velu-wasmsimd-arm-rr2-p6-x4.c",
1308 "src/f32-velu/gen/velu-wasmsimd-arm-rr2-p6-x8.c",
1309 "src/f32-velu/gen/velu-wasmsimd-arm-rr2-p6-x12.c",
1310 "src/f32-velu/gen/velu-wasmsimd-arm-rr2-p6-x16.c",
1311 "src/f32-velu/gen/velu-wasmsimd-arm-rr2-p6-x20.c",
1312 "src/f32-velu/gen/velu-wasmsimd-arm-rr2-p6-x24.c",
Frank Barcharde7223ee2020-12-04 19:04:01 -08001313 "src/f32-velu/gen/velu-wasmsimd-x86-rr2-lut16-p3-x4.c",
1314 "src/f32-velu/gen/velu-wasmsimd-x86-rr2-lut16-p3-x8.c",
1315 "src/f32-velu/gen/velu-wasmsimd-x86-rr2-lut16-p3-x12.c",
1316 "src/f32-velu/gen/velu-wasmsimd-x86-rr2-lut16-p3-x16.c",
1317 "src/f32-velu/gen/velu-wasmsimd-x86-rr2-lut16-p3-x20.c",
1318 "src/f32-velu/gen/velu-wasmsimd-x86-rr2-lut16-p3-x24.c",
Marat Dukhaned6baaf2020-12-01 15:07:08 -08001319 "src/f32-velu/gen/velu-wasmsimd-x86-rr2-p6-x4.c",
1320 "src/f32-velu/gen/velu-wasmsimd-x86-rr2-p6-x8.c",
1321 "src/f32-velu/gen/velu-wasmsimd-x86-rr2-p6-x12.c",
1322 "src/f32-velu/gen/velu-wasmsimd-x86-rr2-p6-x16.c",
1323 "src/f32-velu/gen/velu-wasmsimd-x86-rr2-p6-x20.c",
1324 "src/f32-velu/gen/velu-wasmsimd-x86-rr2-p6-x24.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07001325 "src/f32-vlrelu/gen/vlrelu-wasmsimd-bitselect-x4.c",
1326 "src/f32-vlrelu/gen/vlrelu-wasmsimd-bitselect-x8.c",
1327 "src/f32-vlrelu/gen/vlrelu-wasmsimd-minmax-x4.c",
1328 "src/f32-vlrelu/gen/vlrelu-wasmsimd-minmax-x8.c",
Marat Dukhand816f622020-07-15 10:14:39 -07001329 "src/f32-vmulcaddc/gen/c4-minmax-wasmsimd-arm-2x.c",
Marat Dukhand816f622020-07-15 10:14:39 -07001330 "src/f32-vmulcaddc/gen/c4-minmax-wasmsimd-x86-2x.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001331 "src/f32-vmulcaddc/gen/c8-minmax-wasmsimd-arm-2x.c",
Marat Dukhand816f622020-07-15 10:14:39 -07001332 "src/f32-vmulcaddc/gen/c8-minmax-wasmsimd-x86-2x.c",
Marat Dukhanb82b2cd2020-07-16 02:23:42 -07001333 "src/f32-vrnd/gen/vrndd-wasmsimd-addsub-x4.c",
1334 "src/f32-vrnd/gen/vrndd-wasmsimd-addsub-x8.c",
1335 "src/f32-vrnd/gen/vrndd-wasmsimd-cvt-x4.c",
1336 "src/f32-vrnd/gen/vrndd-wasmsimd-cvt-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001337 "src/f32-vrnd/gen/vrndne-wasmsimd-addsub-x4.c",
1338 "src/f32-vrnd/gen/vrndne-wasmsimd-addsub-x8.c",
1339 "src/f32-vrnd/gen/vrndu-wasmsimd-addsub-x4.c",
1340 "src/f32-vrnd/gen/vrndu-wasmsimd-addsub-x8.c",
1341 "src/f32-vrnd/gen/vrndu-wasmsimd-cvt-x4.c",
1342 "src/f32-vrnd/gen/vrndu-wasmsimd-cvt-x8.c",
1343 "src/f32-vrnd/gen/vrndz-wasmsimd-addsub-x4.c",
1344 "src/f32-vrnd/gen/vrndz-wasmsimd-addsub-x8.c",
1345 "src/f32-vrnd/gen/vrndz-wasmsimd-cvt-x4.c",
1346 "src/f32-vrnd/gen/vrndz-wasmsimd-cvt-x8.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07001347 "src/f32-vsqrt/gen/wasmsimd-sqrt-x4.c",
1348 "src/f32-vsqrt/gen/wasmsimd-sqrt-x8.c",
Marat Dukhan37c83512020-06-29 13:25:53 -07001349 "src/f32-vunary/gen/vabs-wasmsimd-x4.c",
1350 "src/f32-vunary/gen/vabs-wasmsimd-x8.c",
1351 "src/f32-vunary/gen/vneg-wasmsimd-x4.c",
1352 "src/f32-vunary/gen/vneg-wasmsimd-x8.c",
1353 "src/f32-vunary/gen/vsqr-wasmsimd-x4.c",
1354 "src/f32-vunary/gen/vsqr-wasmsimd-x8.c",
Marat Dukhande390d42020-11-29 19:32:18 -08001355 "src/math/expm1minus-wasmsimd-rr2-lut16-p3-andnot.c",
1356 "src/math/expm1minus-wasmsimd-rr2-lut16-p3-max.c",
1357 "src/math/expm1minus-wasmsimd-rr2-p6-andnot.c",
1358 "src/math/expm1minus-wasmsimd-rr2-p6-max.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001359 "src/math/roundd-wasmsimd-addsub.c",
1360 "src/math/roundd-wasmsimd-cvt.c",
1361 "src/math/roundne-wasmsimd-addsub.c",
1362 "src/math/roundu-wasmsimd-addsub.c",
1363 "src/math/roundu-wasmsimd-cvt.c",
1364 "src/math/roundz-wasmsimd-addsub.c",
1365 "src/math/roundz-wasmsimd-cvt.c",
1366 "src/math/sigmoid-wasmsimd-rr2-lut64-p2-div.c",
1367 "src/math/sigmoid-wasmsimd-rr2-p5-div.c",
Marat Dukhancc8f34c2020-08-05 16:36:38 -07001368 "src/qs8-dwconv/gen/up8x9-minmax-wasmsimd-mul16.c",
1369 "src/qs8-dwconv/gen/up16x9-minmax-wasmsimd-mul16.c",
1370 "src/qs8-dwconv/gen/up24x9-minmax-wasmsimd-mul16.c",
Marat Dukhanb5e3d172020-08-06 13:29:53 -07001371 "src/qs8-gavgpool/gen/7p7x-minmax-wasmsimd-c8-acc2.c",
1372 "src/qs8-gavgpool/gen/7p7x-minmax-wasmsimd-c16-acc2.c",
1373 "src/qs8-gavgpool/gen/7p7x-minmax-wasmsimd-c24-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001374 "src/qs8-gavgpool/gen/7x-minmax-wasmsimd-c8-acc2.c",
1375 "src/qs8-gavgpool/gen/7x-minmax-wasmsimd-c16-acc2.c",
1376 "src/qs8-gavgpool/gen/7x-minmax-wasmsimd-c24-acc2.c",
Marat Dukhan27203da2020-08-05 15:19:03 -07001377 "src/qs8-gemm/gen/1x4c8-minmax-wasmsimd-ld64.c",
Marat Dukhan27203da2020-08-05 15:19:03 -07001378 "src/qs8-gemm/gen/1x4c8-minmax-wasmsimd-ld128.c",
Marat Dukhan27203da2020-08-05 15:19:03 -07001379 "src/qs8-gemm/gen/1x4c8-xw-minmax-wasmsimd.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001380 "src/qs8-gemm/gen/2x4c8-minmax-wasmsimd-ld64.c",
1381 "src/qs8-gemm/gen/2x4c8-minmax-wasmsimd-ld128.c",
Marat Dukhan27203da2020-08-05 15:19:03 -07001382 "src/qs8-gemm/gen/2x4c8-xw-minmax-wasmsimd.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001383 "src/qs8-gemm/gen/3x4c8-minmax-wasmsimd-ld64.c",
1384 "src/qs8-gemm/gen/3x4c8-minmax-wasmsimd-ld128.c",
Marat Dukhan27203da2020-08-05 15:19:03 -07001385 "src/qs8-gemm/gen/3x4c8-xw-minmax-wasmsimd.c",
1386 "src/qs8-igemm/gen/1x4c8-minmax-wasmsimd-ld64.c",
Marat Dukhan27203da2020-08-05 15:19:03 -07001387 "src/qs8-igemm/gen/1x4c8-minmax-wasmsimd-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001388 "src/qs8-igemm/gen/2x4c8-minmax-wasmsimd-ld64.c",
Marat Dukhan27203da2020-08-05 15:19:03 -07001389 "src/qs8-igemm/gen/2x4c8-minmax-wasmsimd-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001390 "src/qs8-igemm/gen/3x4c8-minmax-wasmsimd-ld64.c",
Marat Dukhan27203da2020-08-05 15:19:03 -07001391 "src/qs8-igemm/gen/3x4c8-minmax-wasmsimd-ld128.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -07001392 "src/qs8-requantization/fp32-wasmsimd.c",
Marat Dukhan138560c2020-08-03 18:57:34 -07001393 "src/qs8-requantization/q31-wasmsimd.c",
Marat Dukhan5df27f82020-09-02 23:59:21 -07001394 "src/qs8-vadd/gen/minmax-wasmsimd-x8.c",
1395 "src/qs8-vadd/gen/minmax-wasmsimd-x16.c",
1396 "src/qs8-vadd/gen/minmax-wasmsimd-x24.c",
1397 "src/qs8-vadd/gen/minmax-wasmsimd-x32.c",
1398 "src/qs8-vaddc/gen/minmax-wasmsimd-x8.c",
1399 "src/qs8-vaddc/gen/minmax-wasmsimd-x16.c",
1400 "src/qs8-vaddc/gen/minmax-wasmsimd-x24.c",
1401 "src/qs8-vaddc/gen/minmax-wasmsimd-x32.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07001402 "src/qu8-requantization/fp32-wasmsimd.c",
Marat Dukhan22de5e72020-08-03 21:53:18 -07001403 "src/qu8-requantization/q31-wasmsimd.c",
Marat Dukhan8ee37012020-07-16 13:17:13 -07001404 "src/x32-fill/wasmsimd.c",
Marat Dukhan66d99e92020-07-16 12:56:21 -07001405 "src/x32-packx/x4-wasmsimd.c",
Marat Dukhan9306ae02020-07-16 15:51:13 -07001406 "src/x32-pad/wasmsimd.c",
Marat Dukhan9d4bfa22020-07-16 19:07:04 -07001407 "src/x32-unpool/wasmsimd.c",
Marat Dukhane3b78762020-07-16 20:02:58 -07001408 "src/x32-zip/x2-wasmsimd.c",
1409 "src/x32-zip/x3-wasmsimd.c",
1410 "src/x32-zip/x4-wasmsimd.c",
1411 "src/x32-zip/xm-wasmsimd.c",
Marat Dukhan290055c2020-06-09 12:24:29 -07001412]
1413
Marat Dukhan08c4a432019-10-03 09:29:21 -07001414# ISA-specific micro-kernels
1415NEON_UKERNELS = [
Marat Dukhanef25c6d2020-07-24 00:59:40 -07001416 "src/f32-argmaxpool/4x-neon-c4.c",
1417 "src/f32-argmaxpool/9p8x-neon-c4.c",
1418 "src/f32-argmaxpool/9x-neon-c4.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001419 "src/f32-avgpool/9p8x-minmax-neon-c4.c",
1420 "src/f32-avgpool/9x-minmax-neon-c4.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001421 "src/f32-clamp/gen/neon-x4.c",
1422 "src/f32-clamp/gen/neon-x8.c",
Marat Dukhan56b10cd2020-05-18 09:35:49 -07001423 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neon-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001424 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neon-2x2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001425 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neon-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001426 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neon-2x2.c",
Marat Dukhan56b10cd2020-05-18 09:35:49 -07001427 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neon-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001428 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neon-2x2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001429 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neon-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001430 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neon-2x2.c",
Marat Dukhanc7634882020-12-07 15:11:12 -08001431 "src/f32-conv-hwc2chw/3x3s2p1c3x4-neon-2x2.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -07001432 "src/f32-dwconv/gen/up4x4-minmax-neon-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001433 "src/f32-dwconv/gen/up4x4-minmax-neon.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001434 "src/f32-dwconv/gen/up4x9-minmax-neon-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001435 "src/f32-dwconv/gen/up4x9-minmax-neon.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -07001436 "src/f32-dwconv/gen/up4x25-minmax-neon-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001437 "src/f32-dwconv/gen/up4x25-minmax-neon.c",
1438 "src/f32-dwconv/gen/up8x4-minmax-neon-acc2.c",
1439 "src/f32-dwconv/gen/up8x4-minmax-neon.c",
1440 "src/f32-dwconv/gen/up8x9-minmax-neon-acc2.c",
1441 "src/f32-dwconv/gen/up8x9-minmax-neon.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -07001442 "src/f32-dwconv/gen/up8x25-minmax-neon-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001443 "src/f32-dwconv/gen/up8x25-minmax-neon.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001444 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-1x4-acc2.c",
1445 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-1x4-acc3.c",
1446 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-1x4-acc4.c",
Marat Dukhanc581e482020-10-24 01:28:11 -07001447 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-1x4.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001448 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-2x4-acc2.c",
Marat Dukhanc581e482020-10-24 01:28:11 -07001449 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-2x4.c",
1450 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-3x4.c",
1451 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-4x4.c",
1452 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-5x4.c",
1453 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-6x4.c",
Marat Dukhan82f0c322020-10-25 19:17:35 -07001454 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neon-1x4-acc2.c",
1455 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neon-1x4-acc3.c",
1456 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neon-1x4-acc4.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001457 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neon-1x4.c",
Marat Dukhan82f0c322020-10-25 19:17:35 -07001458 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neon-2x4-acc2.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001459 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neon-2x4.c",
1460 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neon-3x4.c",
1461 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neon-4x4.c",
Marat Dukhan149f0ea2020-10-26 12:50:33 -07001462 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-1x4-acc2.c",
1463 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-1x4-acc3.c",
1464 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-1x4-acc4.c",
1465 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-1x4-acc5.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001466 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-1x4.c",
Marat Dukhan149f0ea2020-10-26 12:50:33 -07001467 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-2x4-acc2.c",
1468 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-2x4-acc3.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001469 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-2x4.c",
Marat Dukhan149f0ea2020-10-26 12:50:33 -07001470 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-3x4-acc2.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001471 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-3x4.c",
Marat Dukhan149f0ea2020-10-26 12:50:33 -07001472 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-4x4-acc2.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001473 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-4x4.c",
1474 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-5x4.c",
Marat Dukhan30d4b252020-10-29 16:33:22 -07001475 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neon-1x4-acc2.c",
1476 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neon-1x4-acc3.c",
1477 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neon-1x4-acc4.c",
1478 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neon-1x4-acc5.c",
1479 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neon-1x4.c",
1480 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neon-2x4-acc2.c",
1481 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neon-2x4-acc3.c",
1482 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neon-2x4.c",
Marat Dukhan30d4b252020-10-29 16:33:22 -07001483 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neon-3x4-acc2.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07001484 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neon-3x4.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -07001485 "src/f32-gavgpool-cw/neon-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001486 "src/f32-gavgpool/7p7x-minmax-neon-c4.c",
1487 "src/f32-gavgpool/7x-minmax-neon-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001488 "src/f32-gemm/gen-inc/1x8inc-minmax-neon-dup-ld64.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001489 "src/f32-gemm/gen-inc/1x8inc-minmax-neon-lane-ld64.c",
1490 "src/f32-gemm/gen-inc/1x8s4inc-minmax-neon.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001491 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-dup-ld64.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001492 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-dup-ld128.c",
1493 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-lane-ld64.c",
1494 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-lane-ld128.c",
1495 "src/f32-gemm/gen-inc/4x8s4inc-minmax-neon.c",
1496 "src/f32-gemm/gen-inc/5x8inc-minmax-neon-lane-ld64.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001497 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-dup-ld64.c",
1498 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-dup-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001499 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-lane-ld64.c",
1500 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-lane-ld128.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001501 "src/f32-gemm/gen-inc/6x8s4inc-minmax-neon.c",
1502 "src/f32-gemm/gen-inc/8x8s4inc-minmax-neon.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001503 "src/f32-gemm/gen/1x8-minmax-neon-dup-ld64.c",
1504 "src/f32-gemm/gen/1x8-minmax-neon-lane-ld64.c",
1505 "src/f32-gemm/gen/1x8s4-minmax-neon.c",
1506 "src/f32-gemm/gen/4x2-minmax-neon-lane-ld64.c",
1507 "src/f32-gemm/gen/4x8-minmax-neon-dup-ld64.c",
1508 "src/f32-gemm/gen/4x8-minmax-neon-dup-ld128.c",
1509 "src/f32-gemm/gen/4x8-minmax-neon-lane-ld64.c",
1510 "src/f32-gemm/gen/4x8-minmax-neon-lane-ld128.c",
1511 "src/f32-gemm/gen/4x8s4-minmax-neon.c",
1512 "src/f32-gemm/gen/5x8-minmax-neon-lane-ld64.c",
1513 "src/f32-gemm/gen/6x8-minmax-neon-dup-ld64.c",
1514 "src/f32-gemm/gen/6x8-minmax-neon-dup-ld128.c",
1515 "src/f32-gemm/gen/6x8-minmax-neon-lane-ld64.c",
1516 "src/f32-gemm/gen/6x8-minmax-neon-lane-ld128.c",
1517 "src/f32-gemm/gen/6x8s4-minmax-neon.c",
1518 "src/f32-gemm/gen/8x8s4-minmax-neon.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07001519 "src/f32-hswish/gen/hswish-neon-x4.c",
1520 "src/f32-hswish/gen/hswish-neon-x8.c",
1521 "src/f32-hswish/gen/hswish-neon-x16.c",
Artsiom Ablavatski2202c812021-01-22 14:16:43 -08001522 "src/f32-ibilinear-chw/gen/neon-p4.c",
1523 "src/f32-ibilinear-chw/gen/neon-p8.c",
Frank Barchard8247e212021-02-03 18:12:33 -08001524 "src/f32-ibilinear/gen/neon-c4.c",
1525 "src/f32-ibilinear/gen/neon-c8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001526 "src/f32-igemm/gen/1x8-minmax-neon-dup-ld64.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001527 "src/f32-igemm/gen/1x8-minmax-neon-lane-ld64.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001528 "src/f32-igemm/gen/1x8s4-minmax-neon.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001529 "src/f32-igemm/gen/4x2-minmax-neon-lane-ld64.c",
1530 "src/f32-igemm/gen/4x4-minmax-neon-lane-ld64.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001531 "src/f32-igemm/gen/4x8-minmax-neon-dup-ld64.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001532 "src/f32-igemm/gen/4x8-minmax-neon-dup-ld128.c",
1533 "src/f32-igemm/gen/4x8-minmax-neon-lane-ld64.c",
1534 "src/f32-igemm/gen/4x8-minmax-neon-lane-ld128.c",
1535 "src/f32-igemm/gen/4x8s4-minmax-neon.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001536 "src/f32-igemm/gen/6x8-minmax-neon-dup-ld64.c",
1537 "src/f32-igemm/gen/6x8-minmax-neon-dup-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001538 "src/f32-igemm/gen/6x8-minmax-neon-lane-ld64.c",
1539 "src/f32-igemm/gen/6x8-minmax-neon-lane-ld128.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001540 "src/f32-igemm/gen/6x8s4-minmax-neon.c",
1541 "src/f32-igemm/gen/8x8s4-minmax-neon.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001542 "src/f32-maxpool/9p8x-minmax-neon-c4.c",
1543 "src/f32-pavgpool/9p8x-minmax-neon-c4.c",
1544 "src/f32-pavgpool/9x-minmax-neon-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001545 "src/f32-ppmm/gen/4x8-minmax-neon.c",
1546 "src/f32-ppmm/gen/8x8-minmax-neon.c",
Frank Barcharda5316982020-07-23 13:19:28 -07001547 "src/f32-prelu/gen/neon-1x4.c",
1548 "src/f32-prelu/gen/neon-1x8.c",
1549 "src/f32-prelu/gen/neon-1x16.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001550 "src/f32-prelu/gen/neon-2x4.c",
1551 "src/f32-prelu/gen/neon-2x8.c",
Frank Barcharda5316982020-07-23 13:19:28 -07001552 "src/f32-prelu/gen/neon-2x16.c",
1553 "src/f32-prelu/gen/neon-4x4.c",
1554 "src/f32-prelu/gen/neon-4x8.c",
1555 "src/f32-prelu/gen/neon-4x16.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001556 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x4.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001557 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x8-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001558 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x8.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001559 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12-acc2.c",
1560 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001561 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001562 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16-acc2.c",
1563 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001564 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001565 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20-acc2.c",
1566 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001567 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20.c",
1568 "src/f32-raddstoreexpminusmax/gen/neon-p5-x4.c",
1569 "src/f32-raddstoreexpminusmax/gen/neon-p5-x8-acc2.c",
1570 "src/f32-raddstoreexpminusmax/gen/neon-p5-x8.c",
1571 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12-acc2.c",
1572 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12-acc3.c",
1573 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12.c",
1574 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16-acc2.c",
1575 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16-acc4.c",
1576 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16.c",
1577 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20-acc2.c",
1578 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20-acc5.c",
1579 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07001580 "src/f32-relu/gen/neon-x4.c",
1581 "src/f32-relu/gen/neon-x8.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001582 "src/f32-rmax/neon.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -08001583 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x4.c",
1584 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x8.c",
1585 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x12.c",
1586 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x16.c",
1587 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x20.c",
1588 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x24.c",
1589 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x4.c",
1590 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x8.c",
1591 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x12.c",
1592 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x16.c",
1593 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x20.c",
1594 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x24.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001595 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x4.c",
1596 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x8.c",
1597 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x12.c",
1598 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x16.c",
1599 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x20.c",
1600 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x24.c",
Marat Dukhan5b86c432020-12-06 19:15:03 -08001601 "src/f32-spmm/gen/4x1-minmax-neon-pipelined.c",
1602 "src/f32-spmm/gen/4x1-minmax-neon-x2.c",
1603 "src/f32-spmm/gen/4x1-minmax-neon.c",
1604 "src/f32-spmm/gen/8x1-minmax-neon-pipelined.c",
1605 "src/f32-spmm/gen/8x1-minmax-neon-x2.c",
1606 "src/f32-spmm/gen/8x1-minmax-neon.c",
1607 "src/f32-spmm/gen/12x1-minmax-neon.c",
1608 "src/f32-spmm/gen/16x1-minmax-neon-pipelined.c",
1609 "src/f32-spmm/gen/16x1-minmax-neon-x2.c",
1610 "src/f32-spmm/gen/16x1-minmax-neon.c",
1611 "src/f32-spmm/gen/32x1-minmax-neon-pipelined.c",
1612 "src/f32-spmm/gen/32x1-minmax-neon-x2.c",
1613 "src/f32-spmm/gen/32x1-minmax-neon.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001614 "src/f32-vbinary/gen/vadd-minmax-neon-x4.c",
1615 "src/f32-vbinary/gen/vadd-minmax-neon-x8.c",
1616 "src/f32-vbinary/gen/vaddc-minmax-neon-x4.c",
1617 "src/f32-vbinary/gen/vaddc-minmax-neon-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -08001618 "src/f32-vbinary/gen/vmax-neon-x4.c",
1619 "src/f32-vbinary/gen/vmax-neon-x8.c",
1620 "src/f32-vbinary/gen/vmaxc-neon-x4.c",
1621 "src/f32-vbinary/gen/vmaxc-neon-x8.c",
1622 "src/f32-vbinary/gen/vmin-neon-x4.c",
1623 "src/f32-vbinary/gen/vmin-neon-x8.c",
1624 "src/f32-vbinary/gen/vminc-neon-x4.c",
1625 "src/f32-vbinary/gen/vminc-neon-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001626 "src/f32-vbinary/gen/vmul-minmax-neon-x4.c",
1627 "src/f32-vbinary/gen/vmul-minmax-neon-x8.c",
1628 "src/f32-vbinary/gen/vmulc-minmax-neon-x4.c",
1629 "src/f32-vbinary/gen/vmulc-minmax-neon-x8.c",
1630 "src/f32-vbinary/gen/vrsubc-minmax-neon-x4.c",
1631 "src/f32-vbinary/gen/vrsubc-minmax-neon-x8.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07001632 "src/f32-vbinary/gen/vsqrdiff-neon-x4.c",
1633 "src/f32-vbinary/gen/vsqrdiff-neon-x8.c",
1634 "src/f32-vbinary/gen/vsqrdiffc-neon-x4.c",
1635 "src/f32-vbinary/gen/vsqrdiffc-neon-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001636 "src/f32-vbinary/gen/vsub-minmax-neon-x4.c",
1637 "src/f32-vbinary/gen/vsub-minmax-neon-x8.c",
1638 "src/f32-vbinary/gen/vsubc-minmax-neon-x4.c",
1639 "src/f32-vbinary/gen/vsubc-minmax-neon-x8.c",
Marat Dukhaned6baaf2020-12-01 15:07:08 -08001640 "src/f32-velu/gen/velu-neon-rr2-lut16-p3-x4.c",
1641 "src/f32-velu/gen/velu-neon-rr2-lut16-p3-x8.c",
1642 "src/f32-velu/gen/velu-neon-rr2-lut16-p3-x12.c",
1643 "src/f32-velu/gen/velu-neon-rr2-lut16-p3-x16.c",
1644 "src/f32-velu/gen/velu-neon-rr2-lut16-p3-x20.c",
1645 "src/f32-velu/gen/velu-neon-rr2-lut16-p3-x24.c",
1646 "src/f32-velu/gen/velu-neon-rr2-p6-x4.c",
1647 "src/f32-velu/gen/velu-neon-rr2-p6-x8.c",
1648 "src/f32-velu/gen/velu-neon-rr2-p6-x12.c",
1649 "src/f32-velu/gen/velu-neon-rr2-p6-x16.c",
1650 "src/f32-velu/gen/velu-neon-rr2-p6-x20.c",
1651 "src/f32-velu/gen/velu-neon-rr2-p6-x24.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07001652 "src/f32-vlrelu/gen/vlrelu-neon-x4.c",
1653 "src/f32-vlrelu/gen/vlrelu-neon-x8.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001654 "src/f32-vmulcaddc/gen/c4-minmax-neon-2x.c",
1655 "src/f32-vmulcaddc/gen/c8-minmax-neon-2x.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07001656 "src/f32-vrnd/gen/vrndd-neon-x4.c",
1657 "src/f32-vrnd/gen/vrndd-neon-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001658 "src/f32-vrnd/gen/vrndne-neon-x4.c",
1659 "src/f32-vrnd/gen/vrndne-neon-x8.c",
1660 "src/f32-vrnd/gen/vrndu-neon-x4.c",
1661 "src/f32-vrnd/gen/vrndu-neon-x8.c",
1662 "src/f32-vrnd/gen/vrndz-neon-x4.c",
1663 "src/f32-vrnd/gen/vrndz-neon-x8.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07001664 "src/f32-vunary/gen/vabs-neon-x4.c",
1665 "src/f32-vunary/gen/vabs-neon-x8.c",
1666 "src/f32-vunary/gen/vneg-neon-x4.c",
1667 "src/f32-vunary/gen/vneg-neon-x8.c",
1668 "src/f32-vunary/gen/vsqr-neon-x4.c",
1669 "src/f32-vunary/gen/vsqr-neon-x8.c",
Marat Dukhande390d42020-11-29 19:32:18 -08001670 "src/math/expm1minus-neon-rr2-lut16-p3.c",
1671 "src/math/expm1minus-neon-rr2-p6.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001672 "src/math/roundd-neon-addsub.c",
1673 "src/math/roundd-neon-cvt.c",
1674 "src/math/roundne-neon-addsub.c",
1675 "src/math/roundu-neon-addsub.c",
1676 "src/math/roundu-neon-cvt.c",
1677 "src/math/roundz-neon-addsub.c",
1678 "src/math/roundz-neon-cvt.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001679 "src/math/sigmoid-neon-rr2-lut64-p2-nr2recps.c",
1680 "src/math/sigmoid-neon-rr2-lut2048-p1-nr2recps.c",
1681 "src/math/sigmoid-neon-rr2-p5-nr2recps.c",
1682 "src/math/sqrt-neon-nr1rsqrts.c",
1683 "src/math/sqrt-neon-nr2rsqrts.c",
1684 "src/math/sqrt-neon-nr3rsqrts.c",
Marat Dukhan023bcf92020-08-10 12:40:50 -07001685 "src/qs8-dwconv/gen/up8x9-minmax-neon-mul16.c",
1686 "src/qs8-dwconv/gen/up16x9-minmax-neon-mul16.c",
1687 "src/qs8-dwconv/gen/up24x9-minmax-neon-mul16.c",
1688 "src/qs8-dwconv/gen/up32x9-minmax-neon-mul16.c",
Marat Dukhan281262d2020-08-10 13:23:21 -07001689 "src/qs8-gavgpool/gen/7p7x-minmax-neon-c8-acc2.c",
1690 "src/qs8-gavgpool/gen/7p7x-minmax-neon-c16-acc2.c",
1691 "src/qs8-gavgpool/gen/7p7x-minmax-neon-c24-acc2.c",
1692 "src/qs8-gavgpool/gen/7p7x-minmax-neon-c32-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001693 "src/qs8-gavgpool/gen/7x-minmax-neon-c8-acc2.c",
1694 "src/qs8-gavgpool/gen/7x-minmax-neon-c16-acc2.c",
1695 "src/qs8-gavgpool/gen/7x-minmax-neon-c24-acc2.c",
1696 "src/qs8-gavgpool/gen/7x-minmax-neon-c32-acc2.c",
Frank Barchardec0bf142021-01-15 10:14:05 -08001697 "src/qs8-gemm/gen/1x8-minmax-neon-mlal-lane.c",
Frank Barchardec0bf142021-01-15 10:14:05 -08001698 "src/qs8-gemm/gen/1x8-minmax-neon-mull-addw-dup.c",
Frank Barchard8247e212021-02-03 18:12:33 -08001699 "src/qs8-gemm/gen/1x8c2-minmax-neon-mlal-padal-dup.c",
Frank Barchard2302ffd2021-01-22 14:08:02 -08001700 "src/qs8-gemm/gen/1x8c2-minmax-neon-mull-padal-dup.c",
Frank Barchardda78da12021-03-02 14:28:00 -08001701 "src/qs8-gemm/gen/1x8c8-minmax-neon-mlal-padal.c",
Frank Barchard8247e212021-02-03 18:12:33 -08001702 "src/qs8-gemm/gen/1x8c8-minmax-neon-mull-padal.c",
Frank Barchardcadd4222021-01-20 16:27:25 -08001703 "src/qs8-gemm/gen/1x16-minmax-neon-mlal-lane.c",
1704 "src/qs8-gemm/gen/1x16-minmax-neon-mull-addw-dup.c",
Frank Barchard8247e212021-02-03 18:12:33 -08001705 "src/qs8-gemm/gen/1x16c2-minmax-neon-mlal-padal-dup.c",
Frank Barchard2302ffd2021-01-22 14:08:02 -08001706 "src/qs8-gemm/gen/1x16c2-minmax-neon-mull-padal-dup.c",
Frank Barchardda78da12021-03-02 14:28:00 -08001707 "src/qs8-gemm/gen/1x16c8-minmax-neon-mlal-padal.c",
Frank Barchard8247e212021-02-03 18:12:33 -08001708 "src/qs8-gemm/gen/1x16c8-minmax-neon-mull-padal.c",
Frank Barchardcadd4222021-01-20 16:27:25 -08001709 "src/qs8-gemm/gen/2x8-minmax-neon-mlal-lane.c",
Frank Barchardec0bf142021-01-15 10:14:05 -08001710 "src/qs8-gemm/gen/2x8-minmax-neon-mull-addw-dup.c",
Frank Barchard8247e212021-02-03 18:12:33 -08001711 "src/qs8-gemm/gen/2x8c2-minmax-neon-mlal-padal-dup.c",
Frank Barchard2302ffd2021-01-22 14:08:02 -08001712 "src/qs8-gemm/gen/2x8c2-minmax-neon-mull-padal-dup.c",
Frank Barchardda78da12021-03-02 14:28:00 -08001713 "src/qs8-gemm/gen/2x8c8-minmax-neon-mlal-padal.c",
Frank Barchard8247e212021-02-03 18:12:33 -08001714 "src/qs8-gemm/gen/2x8c8-minmax-neon-mull-padal.c",
Frank Barchardcadd4222021-01-20 16:27:25 -08001715 "src/qs8-gemm/gen/2x16-minmax-neon-mlal-lane.c",
1716 "src/qs8-gemm/gen/2x16-minmax-neon-mull-addw-dup.c",
Frank Barchard8247e212021-02-03 18:12:33 -08001717 "src/qs8-gemm/gen/2x16c2-minmax-neon-mlal-padal-dup.c",
Frank Barchard2302ffd2021-01-22 14:08:02 -08001718 "src/qs8-gemm/gen/2x16c2-minmax-neon-mull-padal-dup.c",
Frank Barchardda78da12021-03-02 14:28:00 -08001719 "src/qs8-gemm/gen/2x16c8-minmax-neon-mlal-padal.c",
Frank Barchard8247e212021-02-03 18:12:33 -08001720 "src/qs8-gemm/gen/2x16c8-minmax-neon-mull-padal.c",
Frank Barchard2302ffd2021-01-22 14:08:02 -08001721 "src/qs8-gemm/gen/3x8-minmax-neon-mlal-lane.c",
1722 "src/qs8-gemm/gen/3x8-minmax-neon-mull-addw-dup.c",
Frank Barchard8247e212021-02-03 18:12:33 -08001723 "src/qs8-gemm/gen/3x8c2-minmax-neon-mlal-padal-dup.c",
Frank Barchard2302ffd2021-01-22 14:08:02 -08001724 "src/qs8-gemm/gen/3x8c2-minmax-neon-mull-padal-dup.c",
Frank Barchardda78da12021-03-02 14:28:00 -08001725 "src/qs8-gemm/gen/3x8c8-minmax-neon-mlal-padal.c",
Frank Barchard8247e212021-02-03 18:12:33 -08001726 "src/qs8-gemm/gen/3x8c8-minmax-neon-mull-padal.c",
1727 "src/qs8-gemm/gen/3x16-minmax-neon-mlal-lane.c",
1728 "src/qs8-gemm/gen/3x16-minmax-neon-mull-addw-dup.c",
1729 "src/qs8-gemm/gen/3x16c2-minmax-neon-mlal-padal-dup.c",
1730 "src/qs8-gemm/gen/3x16c2-minmax-neon-mull-padal-dup.c",
Frank Barchardda78da12021-03-02 14:28:00 -08001731 "src/qs8-gemm/gen/3x16c8-minmax-neon-mlal-padal.c",
Frank Barchard8247e212021-02-03 18:12:33 -08001732 "src/qs8-gemm/gen/3x16c8-minmax-neon-mull-padal.c",
Frank Barchardcadd4222021-01-20 16:27:25 -08001733 "src/qs8-gemm/gen/4x8-minmax-neon-mlal-lane.c",
Frank Barchardec0bf142021-01-15 10:14:05 -08001734 "src/qs8-gemm/gen/4x8-minmax-neon-mull-addw-dup.c",
Frank Barchard8247e212021-02-03 18:12:33 -08001735 "src/qs8-gemm/gen/4x8c2-minmax-neon-mlal-padal-dup.c",
Frank Barchard2302ffd2021-01-22 14:08:02 -08001736 "src/qs8-gemm/gen/4x8c2-minmax-neon-mull-padal-dup.c",
Frank Barchardda78da12021-03-02 14:28:00 -08001737 "src/qs8-gemm/gen/4x8c8-minmax-neon-mlal-padal.c",
Frank Barchard8247e212021-02-03 18:12:33 -08001738 "src/qs8-gemm/gen/4x8c8-minmax-neon-mull-padal.c",
Frank Barchardcadd4222021-01-20 16:27:25 -08001739 "src/qs8-gemm/gen/4x16-minmax-neon-mlal-lane.c",
1740 "src/qs8-gemm/gen/4x16-minmax-neon-mull-addw-dup.c",
Frank Barchard8247e212021-02-03 18:12:33 -08001741 "src/qs8-gemm/gen/4x16c2-minmax-neon-mlal-padal-dup.c",
Frank Barchard2302ffd2021-01-22 14:08:02 -08001742 "src/qs8-gemm/gen/4x16c2-minmax-neon-mull-padal-dup.c",
Frank Barchardda78da12021-03-02 14:28:00 -08001743 "src/qs8-gemm/gen/4x16c8-minmax-neon-mlal-padal.c",
Frank Barchard58990122021-01-29 16:09:56 -08001744 "src/qs8-gemm/gen/4x16c8-minmax-neon-mull-padal.c",
Frank Barchardda78da12021-03-02 14:28:00 -08001745 "src/qs8-gemm/gen/1x8c16-minmax-neon-mlal-padal.c",
1746 "src/qs8-gemm/gen/2x8c16-minmax-neon-mlal-padal.c",
1747 "src/qs8-gemm/gen/3x8c16-minmax-neon-mlal-padal.c",
1748 "src/qs8-gemm/gen/4x8c16-minmax-neon-mlal-padal.c",
1749 "src/qs8-gemm/gen/1x16c16-minmax-neon-mlal-padal.c",
1750 "src/qs8-gemm/gen/2x16c16-minmax-neon-mlal-padal.c",
1751 "src/qs8-gemm/gen/3x16c16-minmax-neon-mlal-padal.c",
Frank Barchard71c4d1a2021-02-19 10:07:23 -08001752 "src/qs8-gemm/gen/4x16c16-minmax-neon-mlal-padal.c",
Frank Barchardec0bf142021-01-15 10:14:05 -08001753 "src/qs8-igemm/gen/1x8-minmax-neon-mlal-lane.c",
Frank Barchardd76a37b2021-03-02 11:36:16 -08001754 "src/qs8-igemm/gen/1x16-minmax-neon-mlal-lane.c",
Frank Barchardd76a37b2021-03-02 11:36:16 -08001755 "src/qs8-igemm/gen/2x8-minmax-neon-mlal-lane.c",
Frank Barchardd76a37b2021-03-02 11:36:16 -08001756 "src/qs8-igemm/gen/2x16-minmax-neon-mlal-lane.c",
Frank Barchardd76a37b2021-03-02 11:36:16 -08001757 "src/qs8-igemm/gen/3x8-minmax-neon-mlal-lane.c",
Frank Barchardd76a37b2021-03-02 11:36:16 -08001758 "src/qs8-igemm/gen/3x16-minmax-neon-mlal-lane.c",
Frank Barchardd76a37b2021-03-02 11:36:16 -08001759 "src/qs8-igemm/gen/4x8-minmax-neon-mlal-lane.c",
Frank Barchardd76a37b2021-03-02 11:36:16 -08001760 "src/qs8-igemm/gen/4x16-minmax-neon-mlal-lane.c",
Frank Barchardda78da12021-03-02 14:28:00 -08001761 "src/qs8-igemm/gen/1x16c8-minmax-neon-mlal-padal.c",
1762 "src/qs8-igemm/gen/1x16c8-minmax-neon-mull-padal.c",
1763 "src/qs8-igemm/gen/1x8c8-minmax-neon-mlal-padal.c",
1764 "src/qs8-igemm/gen/1x8c8-minmax-neon-mull-padal.c",
1765 "src/qs8-igemm/gen/2x16c8-minmax-neon-mlal-padal.c",
1766 "src/qs8-igemm/gen/2x16c8-minmax-neon-mull-padal.c",
1767 "src/qs8-igemm/gen/2x8c8-minmax-neon-mlal-padal.c",
1768 "src/qs8-igemm/gen/2x8c8-minmax-neon-mull-padal.c",
1769 "src/qs8-igemm/gen/3x16c8-minmax-neon-mlal-padal.c",
1770 "src/qs8-igemm/gen/3x16c8-minmax-neon-mull-padal.c",
1771 "src/qs8-igemm/gen/3x8c8-minmax-neon-mlal-padal.c",
1772 "src/qs8-igemm/gen/3x8c8-minmax-neon-mull-padal.c",
1773 "src/qs8-igemm/gen/4x16c8-minmax-neon-mlal-padal.c",
Frank Barchardd76a37b2021-03-02 11:36:16 -08001774 "src/qs8-igemm/gen/4x16c8-minmax-neon-mull-padal.c",
Frank Barchardda78da12021-03-02 14:28:00 -08001775 "src/qs8-igemm/gen/4x8c8-minmax-neon-mlal-padal.c",
1776 "src/qs8-igemm/gen/4x8c8-minmax-neon-mull-padal.c",
1777 "src/qs8-igemm/gen/1x16c16-minmax-neon-mlal-padal.c",
1778 "src/qs8-igemm/gen/1x8c16-minmax-neon-mlal-padal.c",
1779 "src/qs8-igemm/gen/2x16c16-minmax-neon-mlal-padal.c",
1780 "src/qs8-igemm/gen/2x8c16-minmax-neon-mlal-padal.c",
1781 "src/qs8-igemm/gen/3x16c16-minmax-neon-mlal-padal.c",
1782 "src/qs8-igemm/gen/3x8c16-minmax-neon-mlal-padal.c",
Frank Barchardd76a37b2021-03-02 11:36:16 -08001783 "src/qs8-igemm/gen/4x16c16-minmax-neon-mlal-padal.c",
Frank Barchardda78da12021-03-02 14:28:00 -08001784 "src/qs8-igemm/gen/4x8c16-minmax-neon-mlal-padal.c",
1785 "src/qs8-igemm/gen/1x16c2-minmax-neon-mlal-padal-dup.c",
1786 "src/qs8-igemm/gen/1x8c2-minmax-neon-mlal-padal-dup.c",
1787 "src/qs8-igemm/gen/2x16c2-minmax-neon-mlal-padal-dup.c",
1788 "src/qs8-igemm/gen/2x8c2-minmax-neon-mlal-padal-dup.c",
1789 "src/qs8-igemm/gen/3x16c2-minmax-neon-mlal-padal-dup.c",
1790 "src/qs8-igemm/gen/3x8c2-minmax-neon-mlal-padal-dup.c",
1791 "src/qs8-igemm/gen/4x16c2-minmax-neon-mlal-padal-dup.c",
1792 "src/qs8-igemm/gen/4x8c2-minmax-neon-mlal-padal-dup.c",
1793 "src/qs8-igemm/gen/1x16c2-minmax-neon-mull-padal-dup.c",
1794 "src/qs8-igemm/gen/1x8c2-minmax-neon-mull-padal-dup.c",
1795 "src/qs8-igemm/gen/2x16c2-minmax-neon-mull-padal-dup.c",
1796 "src/qs8-igemm/gen/2x8c2-minmax-neon-mull-padal-dup.c",
1797 "src/qs8-igemm/gen/3x16c2-minmax-neon-mull-padal-dup.c",
1798 "src/qs8-igemm/gen/3x8c2-minmax-neon-mull-padal-dup.c",
1799 "src/qs8-igemm/gen/4x16c2-minmax-neon-mull-padal-dup.c",
1800 "src/qs8-igemm/gen/4x8c2-minmax-neon-mull-padal-dup.c",
1801 "src/qs8-igemm/gen/1x16-minmax-neon-mull-addw-dup.c",
1802 "src/qs8-igemm/gen/1x8-minmax-neon-mull-addw-dup.c",
1803 "src/qs8-igemm/gen/2x16-minmax-neon-mull-addw-dup.c",
1804 "src/qs8-igemm/gen/2x8-minmax-neon-mull-addw-dup.c",
1805 "src/qs8-igemm/gen/3x16-minmax-neon-mull-addw-dup.c",
1806 "src/qs8-igemm/gen/3x8-minmax-neon-mull-addw-dup.c",
1807 "src/qs8-igemm/gen/4x16-minmax-neon-mull-addw-dup.c",
1808 "src/qs8-igemm/gen/4x8-minmax-neon-mull-addw-dup.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -07001809 "src/qs8-requantization/fp32-neon.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001810 "src/qs8-requantization/precise-neon.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -07001811 "src/qs8-requantization/q31-neon.c",
Marat Dukhanba7b2792020-09-02 14:26:45 -07001812 "src/qs8-vadd/gen/minmax-neon-ld64-x8.c",
1813 "src/qs8-vadd/gen/minmax-neon-ld64-x16.c",
1814 "src/qs8-vadd/gen/minmax-neon-ld64-x24.c",
1815 "src/qs8-vadd/gen/minmax-neon-ld64-x32.c",
1816 "src/qs8-vaddc/gen/minmax-neon-ld64-x8.c",
1817 "src/qs8-vaddc/gen/minmax-neon-ld64-x16.c",
1818 "src/qs8-vaddc/gen/minmax-neon-ld64-x24.c",
1819 "src/qs8-vaddc/gen/minmax-neon-ld64-x32.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -07001820 "src/qu8-avgpool/9p8x-minmax-neon-c8.c",
1821 "src/qu8-avgpool/9x-minmax-neon-c8.c",
1822 "src/qu8-dwconv/up8x9-minmax-neon.c",
1823 "src/qu8-gavgpool/7p7x-minmax-neon-c8.c",
1824 "src/qu8-gavgpool/7x-minmax-neon-c8.c",
1825 "src/qu8-gemm/4x8-minmax-neon.c",
1826 "src/qu8-gemm/8x8-minmax-neon.c",
1827 "src/qu8-igemm/4x8-minmax-neon.c",
1828 "src/qu8-igemm/8x8-minmax-neon.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07001829 "src/qu8-requantization/fp32-neon.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001830 "src/qu8-requantization/precise-neon.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07001831 "src/qu8-requantization/q31-neon.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -07001832 "src/qu8-vadd/minmax-neon.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001833 "src/u8-clamp/neon-x64.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001834 "src/u8-maxpool/9p8x-minmax-neon-c16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001835 "src/u8-rmax/neon.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001836 "src/x8-zip/x2-neon.c",
1837 "src/x8-zip/x3-neon.c",
1838 "src/x8-zip/x4-neon.c",
1839 "src/x8-zip/xm-neon.c",
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -07001840 "src/x32-fill/neon.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001841 "src/x32-packx/x4-neon-st4.c",
Marat Dukhan63523d42020-05-22 17:07:33 -07001842 "src/x32-pad/neon.c",
Marat Dukhan57dccd82020-04-14 00:53:10 -07001843 "src/x32-unpool/neon.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001844 "src/x32-zip/x2-neon.c",
1845 "src/x32-zip/x3-neon.c",
1846 "src/x32-zip/x4-neon.c",
1847 "src/x32-zip/xm-neon.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001848]
1849
1850NEONFMA_UKERNELS = [
Frank Barchard04336c12020-10-22 16:48:55 -07001851 "src/f32-dwconv/gen/up4x4-minmax-neonfma-acc2.c",
1852 "src/f32-dwconv/gen/up4x4-minmax-neonfma.c",
1853 "src/f32-dwconv/gen/up4x9-minmax-neonfma-acc2.c",
1854 "src/f32-dwconv/gen/up4x9-minmax-neonfma.c",
1855 "src/f32-dwconv/gen/up4x25-minmax-neonfma-acc2.c",
1856 "src/f32-dwconv/gen/up4x25-minmax-neonfma.c",
1857 "src/f32-dwconv/gen/up8x4-minmax-neonfma-acc2.c",
1858 "src/f32-dwconv/gen/up8x4-minmax-neonfma.c",
1859 "src/f32-dwconv/gen/up8x9-minmax-neonfma-acc2.c",
1860 "src/f32-dwconv/gen/up8x9-minmax-neonfma.c",
1861 "src/f32-dwconv/gen/up8x25-minmax-neonfma-acc2.c",
1862 "src/f32-dwconv/gen/up8x25-minmax-neonfma.c",
1863 "src/f32-gemm/gen-inc/1x8inc-minmax-neonfma-dup-ld64.c",
1864 "src/f32-gemm/gen-inc/1x8s4inc-minmax-neonfma.c",
1865 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-dup-ld64.c",
1866 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-dup-ld128.c",
1867 "src/f32-gemm/gen-inc/4x8s4inc-minmax-neonfma.c",
1868 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-dup-ld64.c",
1869 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-dup-ld128.c",
1870 "src/f32-gemm/gen-inc/6x8s4inc-minmax-neonfma.c",
1871 "src/f32-gemm/gen-inc/8x8s4inc-minmax-neonfma.c",
1872 "src/f32-gemm/gen/1x8-minmax-neonfma-dup-ld64.c",
1873 "src/f32-gemm/gen/1x8s4-minmax-neonfma.c",
1874 "src/f32-gemm/gen/4x8-minmax-neonfma-dup-ld64.c",
1875 "src/f32-gemm/gen/4x8-minmax-neonfma-dup-ld128.c",
1876 "src/f32-gemm/gen/4x8s4-minmax-neonfma.c",
1877 "src/f32-gemm/gen/6x8-minmax-neonfma-dup-ld64.c",
1878 "src/f32-gemm/gen/6x8-minmax-neonfma-dup-ld128.c",
1879 "src/f32-gemm/gen/6x8s4-minmax-neonfma.c",
1880 "src/f32-gemm/gen/8x8s4-minmax-neonfma.c",
Artsiom Ablavatski2202c812021-01-22 14:16:43 -08001881 "src/f32-ibilinear-chw/gen/neonfma-p4.c",
1882 "src/f32-ibilinear-chw/gen/neonfma-p8.c",
Frank Barchard8247e212021-02-03 18:12:33 -08001883 "src/f32-ibilinear/gen/neonfma-c4.c",
1884 "src/f32-ibilinear/gen/neonfma-c8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001885 "src/f32-igemm/gen/1x8-minmax-neonfma-dup-ld64.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001886 "src/f32-igemm/gen/1x8s4-minmax-neonfma.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001887 "src/f32-igemm/gen/4x8-minmax-neonfma-dup-ld64.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001888 "src/f32-igemm/gen/4x8-minmax-neonfma-dup-ld128.c",
1889 "src/f32-igemm/gen/4x8s4-minmax-neonfma.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001890 "src/f32-igemm/gen/6x8-minmax-neonfma-dup-ld64.c",
1891 "src/f32-igemm/gen/6x8-minmax-neonfma-dup-ld128.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001892 "src/f32-igemm/gen/6x8s4-minmax-neonfma.c",
1893 "src/f32-igemm/gen/8x8s4-minmax-neonfma.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001894 "src/f32-ppmm/gen/4x8-minmax-neonfma.c",
1895 "src/f32-ppmm/gen/8x8-minmax-neonfma.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001896 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x4.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001897 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x8-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001898 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x8.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001899 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12-acc2.c",
1900 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001901 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001902 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16-acc2.c",
1903 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001904 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001905 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20-acc2.c",
1906 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001907 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20.c",
1908 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x4.c",
1909 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x8-acc2.c",
1910 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x8.c",
1911 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12-acc2.c",
1912 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12-acc3.c",
1913 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12.c",
1914 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16-acc2.c",
1915 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16-acc4.c",
1916 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16.c",
1917 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20-acc2.c",
1918 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20-acc5.c",
1919 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -08001920 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x4.c",
1921 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x8.c",
1922 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x12.c",
1923 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x16.c",
1924 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x20.c",
1925 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x24.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001926 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x4.c",
1927 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x8.c",
1928 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x12.c",
1929 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x16.c",
1930 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x20.c",
1931 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x24.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -08001932 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x4.c",
1933 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x8.c",
1934 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x12.c",
1935 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x16.c",
1936 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x20.c",
1937 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x24.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -08001938 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x4.c",
1939 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x8.c",
1940 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x12.c",
1941 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x16.c",
1942 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x20.c",
1943 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x24.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001944 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x4.c",
1945 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x8.c",
1946 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x12.c",
1947 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x16.c",
1948 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x20.c",
1949 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x24.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -08001950 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x4.c",
1951 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x8.c",
1952 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x12.c",
1953 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x16.c",
1954 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x20.c",
1955 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x24.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001956 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x4.c",
1957 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x8.c",
1958 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x12.c",
1959 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x16.c",
1960 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x20.c",
1961 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x24.c",
1962 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x4.c",
1963 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x8.c",
1964 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x12.c",
1965 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x16.c",
1966 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x20.c",
1967 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x24.c",
1968 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x4.c",
1969 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x8.c",
1970 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x12.c",
1971 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x16.c",
1972 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x20.c",
1973 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x24.c",
Marat Dukhan2fa7a0c2020-12-06 19:09:02 -08001974 "src/f32-spmm/gen/4x1-minmax-neonfma-pipelined.c",
1975 "src/f32-spmm/gen/4x1-minmax-neonfma-x2.c",
1976 "src/f32-spmm/gen/4x1-minmax-neonfma.c",
1977 "src/f32-spmm/gen/8x1-minmax-neonfma-pipelined.c",
1978 "src/f32-spmm/gen/8x1-minmax-neonfma-x2.c",
1979 "src/f32-spmm/gen/8x1-minmax-neonfma.c",
1980 "src/f32-spmm/gen/12x1-minmax-neonfma.c",
1981 "src/f32-spmm/gen/16x1-minmax-neonfma-pipelined.c",
1982 "src/f32-spmm/gen/16x1-minmax-neonfma-x2.c",
1983 "src/f32-spmm/gen/16x1-minmax-neonfma.c",
1984 "src/f32-spmm/gen/32x1-minmax-neonfma-pipelined.c",
1985 "src/f32-spmm/gen/32x1-minmax-neonfma-x2.c",
1986 "src/f32-spmm/gen/32x1-minmax-neonfma.c",
Marat Dukhaned6baaf2020-12-01 15:07:08 -08001987 "src/f32-velu/gen/velu-neonfma-rr1-lut16-p3-x4.c",
1988 "src/f32-velu/gen/velu-neonfma-rr1-lut16-p3-x8.c",
1989 "src/f32-velu/gen/velu-neonfma-rr1-lut16-p3-x12.c",
1990 "src/f32-velu/gen/velu-neonfma-rr1-lut16-p3-x16.c",
1991 "src/f32-velu/gen/velu-neonfma-rr1-lut16-p3-x20.c",
1992 "src/f32-velu/gen/velu-neonfma-rr1-lut16-p3-x24.c",
1993 "src/f32-velu/gen/velu-neonfma-rr1-p6-x4.c",
1994 "src/f32-velu/gen/velu-neonfma-rr1-p6-x8.c",
1995 "src/f32-velu/gen/velu-neonfma-rr1-p6-x12.c",
1996 "src/f32-velu/gen/velu-neonfma-rr1-p6-x16.c",
1997 "src/f32-velu/gen/velu-neonfma-rr1-p6-x20.c",
1998 "src/f32-velu/gen/velu-neonfma-rr1-p6-x24.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001999 "src/f32-vmulcaddc/gen/c4-minmax-neonfma-2x.c",
2000 "src/f32-vmulcaddc/gen/c8-minmax-neonfma-2x.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07002001 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x4.c",
2002 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x8.c",
2003 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x12.c",
2004 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x16.c",
2005 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x20.c",
2006 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x24.c",
2007 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x28.c",
2008 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x32.c",
2009 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x36.c",
2010 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x40.c",
2011 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x4.c",
2012 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x8.c",
2013 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x12.c",
2014 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x16.c",
2015 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x20.c",
2016 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x24.c",
2017 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x28.c",
2018 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x32.c",
2019 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x36.c",
2020 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x40.c",
Marat Dukhanb7633f22020-11-20 16:34:56 -08002021 "src/math/exp-neonfma-rr2-lut64-p2.c",
2022 "src/math/exp-neonfma-rr2-p5.c",
Frank Barcharde7223ee2020-12-04 19:04:01 -08002023 "src/math/expm1minus-neonfma-rr1-lut16-p3.c",
2024 "src/math/expm1minus-neonfma-rr1-p6.c",
Marat Dukhan9dd119a2020-11-20 18:20:04 -08002025 "src/math/expminus-neonfma-rr2-lut64-p2.c",
2026 "src/math/expminus-neonfma-rr2-lut2048-p1.c",
2027 "src/math/expminus-neonfma-rr2-p5.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08002028 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr1recps1fma.c",
2029 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr2fma.c",
2030 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr2recps.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002031 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr1recps1fma.c",
2032 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr2fma.c",
2033 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr2recps.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08002034 "src/math/sigmoid-neonfma-rr1-p5-nr1recps1fma.c",
2035 "src/math/sigmoid-neonfma-rr1-p5-nr2fma.c",
2036 "src/math/sigmoid-neonfma-rr1-p5-nr2recps.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08002037 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr1recps1fma.c",
2038 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr2fma.c",
2039 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr2recps.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002040 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr1recps1fma.c",
2041 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr2fma.c",
2042 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr2recps.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08002043 "src/math/sigmoid-neonfma-rr2-p5-nr1recps1fma.c",
2044 "src/math/sigmoid-neonfma-rr2-p5-nr2fma.c",
2045 "src/math/sigmoid-neonfma-rr2-p5-nr2recps.c",
Marat Dukhan84000762020-06-29 18:38:43 -07002046 "src/math/sqrt-neonfma-nr1fma.c",
Marat Dukhan84000762020-06-29 18:38:43 -07002047 "src/math/sqrt-neonfma-nr1rsqrts1fma1adj.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002048 "src/math/sqrt-neonfma-nr2fma.c",
2049 "src/math/sqrt-neonfma-nr2fma1adj.c",
2050 "src/math/sqrt-neonfma-nr3fma.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002051]
2052
2053AARCH64_NEONFMA_UKERNELS = [
Marat Dukhan56b10cd2020-05-18 09:35:49 -07002054 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neonfma-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07002055 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neonfma-2x2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002056 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neonfma-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07002057 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neonfma-2x2.c",
Marat Dukhan56b10cd2020-05-18 09:35:49 -07002058 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neonfma-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07002059 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neonfma-2x2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002060 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neonfma-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07002061 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neonfma-2x2.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -07002062 "src/f32-conv-hwc2chw/3x3s2p1c3x4-neonfma-2x2.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07002063 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-1x4-acc2.c",
2064 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-1x4-acc3.c",
2065 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-1x4-acc4.c",
Marat Dukhan1268a242020-10-24 00:36:32 -07002066 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-1x4.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07002067 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-2x4-acc2.c",
Marat Dukhan1268a242020-10-24 00:36:32 -07002068 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-2x4.c",
2069 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-3x4.c",
2070 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-4x4.c",
2071 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-5x4.c",
2072 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-6x4.c",
Marat Dukhan82f0c322020-10-25 19:17:35 -07002073 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neonfma-1x4-acc2.c",
2074 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neonfma-1x4-acc3.c",
2075 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neonfma-1x4-acc4.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07002076 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neonfma-1x4.c",
Marat Dukhan82f0c322020-10-25 19:17:35 -07002077 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neonfma-2x4-acc2.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07002078 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neonfma-2x4.c",
2079 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neonfma-3x4.c",
2080 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neonfma-4x4.c",
Marat Dukhan149f0ea2020-10-26 12:50:33 -07002081 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-1x4-acc2.c",
2082 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-1x4-acc3.c",
2083 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-1x4-acc4.c",
2084 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-1x4-acc5.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07002085 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-1x4.c",
Marat Dukhan149f0ea2020-10-26 12:50:33 -07002086 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-2x4-acc2.c",
2087 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-2x4-acc3.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07002088 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-2x4.c",
Marat Dukhan149f0ea2020-10-26 12:50:33 -07002089 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-3x4-acc2.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07002090 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-3x4.c",
Marat Dukhan149f0ea2020-10-26 12:50:33 -07002091 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-4x4-acc2.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07002092 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-4x4.c",
2093 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-5x4.c",
Marat Dukhan30d4b252020-10-29 16:33:22 -07002094 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neonfma-1x4-acc2.c",
2095 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neonfma-1x4-acc3.c",
2096 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neonfma-1x4-acc4.c",
2097 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neonfma-1x4-acc5.c",
2098 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neonfma-1x4.c",
2099 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neonfma-2x4-acc2.c",
2100 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neonfma-2x4-acc3.c",
2101 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neonfma-2x4.c",
Marat Dukhan30d4b252020-10-29 16:33:22 -07002102 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neonfma-3x4-acc2.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07002103 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neonfma-3x4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002104 "src/f32-gemm/gen-inc/1x8inc-minmax-neonfma-lane-ld64.c",
2105 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-lane-ld64.c",
2106 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-lane-ld128.c",
2107 "src/f32-gemm/gen-inc/5x8inc-minmax-neonfma-lane-ld64.c",
2108 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-lane-ld64.c",
2109 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-lane-ld128.c",
2110 "src/f32-gemm/gen/1x8-minmax-neonfma-lane-ld64.c",
2111 "src/f32-gemm/gen/4x2-minmax-neonfma-lane-ld64.c",
2112 "src/f32-gemm/gen/4x8-minmax-neonfma-lane-ld64.c",
2113 "src/f32-gemm/gen/4x8-minmax-neonfma-lane-ld128.c",
2114 "src/f32-gemm/gen/5x8-minmax-neonfma-lane-ld64.c",
2115 "src/f32-gemm/gen/6x8-minmax-neonfma-lane-ld64.c",
2116 "src/f32-gemm/gen/6x8-minmax-neonfma-lane-ld128.c",
2117 "src/f32-igemm/gen/1x8-minmax-neonfma-lane-ld64.c",
2118 "src/f32-igemm/gen/4x2-minmax-neonfma-lane-ld64.c",
2119 "src/f32-igemm/gen/4x4-minmax-neonfma-lane-ld64.c",
2120 "src/f32-igemm/gen/4x8-minmax-neonfma-lane-ld64.c",
2121 "src/f32-igemm/gen/4x8-minmax-neonfma-lane-ld128.c",
2122 "src/f32-igemm/gen/6x8-minmax-neonfma-lane-ld64.c",
2123 "src/f32-igemm/gen/6x8-minmax-neonfma-lane-ld128.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -08002124 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x4.c",
2125 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x8.c",
2126 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x12.c",
2127 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x16.c",
2128 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x20.c",
2129 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x24.c",
2130 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x4.c",
2131 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x8.c",
2132 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x12.c",
2133 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x16.c",
2134 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x20.c",
2135 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x24.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002136 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x4.c",
2137 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x8.c",
2138 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x12.c",
2139 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x16.c",
2140 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x20.c",
2141 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x24.c",
Marat Dukhan355ab432020-04-09 19:01:52 -07002142 "src/f32-spmm/gen/4x2-minmax-neonfma.c",
2143 "src/f32-spmm/gen/4x4-minmax-neonfma.c",
Marat Dukhan355ab432020-04-09 19:01:52 -07002144 "src/f32-spmm/gen/8x2-minmax-neonfma.c",
2145 "src/f32-spmm/gen/8x4-minmax-neonfma.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002146 "src/f32-spmm/gen/12x2-minmax-neonfma.c",
2147 "src/f32-spmm/gen/12x4-minmax-neonfma.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002148 "src/f32-spmm/gen/16x2-minmax-neonfma.c",
2149 "src/f32-spmm/gen/16x4-minmax-neonfma.c",
Frank Barchard846c0c62020-10-26 15:01:39 -07002150 "src/f32-spmm/gen/32x2-minmax-neonfma.c",
2151 "src/f32-spmm/gen/32x4-minmax-neonfma.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002152 "src/f32-vbinary/gen/vdiv-minmax-neon-x4.c",
2153 "src/f32-vbinary/gen/vdiv-minmax-neon-x8.c",
2154 "src/f32-vbinary/gen/vdivc-minmax-neon-x4.c",
2155 "src/f32-vbinary/gen/vdivc-minmax-neon-x8.c",
2156 "src/f32-vbinary/gen/vrdivc-minmax-neon-x4.c",
2157 "src/f32-vbinary/gen/vrdivc-minmax-neon-x8.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07002158 "src/f32-vsqrt/gen/neon-sqrt-x4.c",
2159 "src/f32-vsqrt/gen/neon-sqrt-x8.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08002160 "src/math/sigmoid-neonfma-rr1-lut64-p2-div.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002161 "src/math/sigmoid-neonfma-rr1-lut2048-p1-div.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08002162 "src/math/sigmoid-neonfma-rr1-p5-div.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08002163 "src/math/sigmoid-neonfma-rr2-lut64-p2-div.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002164 "src/math/sigmoid-neonfma-rr2-lut2048-p1-div.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08002165 "src/math/sigmoid-neonfma-rr2-p5-div.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002166]
2167
Marat Dukhan8853b822020-05-07 12:19:01 -07002168NEONV8_UKERNELS = [
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07002169 "src/f32-vrnd/gen/vrndd-neonv8-x4.c",
2170 "src/f32-vrnd/gen/vrndd-neonv8-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002171 "src/f32-vrnd/gen/vrndne-neonv8-x4.c",
2172 "src/f32-vrnd/gen/vrndne-neonv8-x8.c",
2173 "src/f32-vrnd/gen/vrndu-neonv8-x4.c",
2174 "src/f32-vrnd/gen/vrndu-neonv8-x8.c",
2175 "src/f32-vrnd/gen/vrndz-neonv8-x4.c",
2176 "src/f32-vrnd/gen/vrndz-neonv8-x8.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07002177 "src/math/roundd-neonv8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002178 "src/math/roundne-neonv8.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07002179 "src/math/roundu-neonv8.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07002180 "src/math/roundz-neonv8.c",
Marat Dukhan8853b822020-05-07 12:19:01 -07002181]
2182
Marat Dukhan08c4a432019-10-03 09:29:21 -07002183AARCH64_NEONFP16ARITH_UKERNELS = [
Frank Barchardb1966592020-05-12 13:47:06 -07002184 "src/f16-clamp/gen/neonfp16arith-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002185 "src/f16-clamp/gen/neonfp16arith-x16.c",
Frank Barchard5a599a62020-06-04 20:12:44 -07002186 "src/f16-dwconv/gen/up8x4-minmax-neonfp16arith-acc2.c",
2187 "src/f16-dwconv/gen/up8x4-minmax-neonfp16arith.c",
2188 "src/f16-dwconv/gen/up8x9-minmax-neonfp16arith-acc2.c",
2189 "src/f16-dwconv/gen/up8x9-minmax-neonfp16arith.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002190 "src/f16-dwconv/gen/up8x25-minmax-neonfp16arith-acc2.c",
2191 "src/f16-dwconv/gen/up8x25-minmax-neonfp16arith.c",
2192 "src/f16-dwconv/gen/up16x4-minmax-neonfp16arith-acc2.c",
2193 "src/f16-dwconv/gen/up16x4-minmax-neonfp16arith.c",
2194 "src/f16-dwconv/gen/up16x9-minmax-neonfp16arith-acc2.c",
2195 "src/f16-dwconv/gen/up16x9-minmax-neonfp16arith.c",
2196 "src/f16-dwconv/gen/up16x25-minmax-neonfp16arith-acc2.c",
2197 "src/f16-dwconv/gen/up16x25-minmax-neonfp16arith.c",
Frank Barchard0bb49a72020-06-04 11:35:11 -07002198 "src/f16-gavgpool/7p7x-minmax-neonfp16arith-c8.c",
2199 "src/f16-gavgpool/7x-minmax-neonfp16arith-c8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002200 "src/f16-gemm/gen-inc/1x8inc-minmax-neonfp16arith-ld64.c",
2201 "src/f16-gemm/gen-inc/1x16inc-minmax-neonfp16arith-ld64.c",
2202 "src/f16-gemm/gen-inc/4x8inc-minmax-neonfp16arith-ld64.c",
2203 "src/f16-gemm/gen-inc/4x16inc-minmax-neonfp16arith-ld64.c",
2204 "src/f16-gemm/gen-inc/6x8inc-minmax-neonfp16arith-ld64.c",
2205 "src/f16-gemm/gen-inc/6x16inc-minmax-neonfp16arith-ld64.c",
2206 "src/f16-gemm/gen-inc/8x8inc-minmax-neonfp16arith-ld64.c",
2207 "src/f16-gemm/gen-inc/8x16inc-minmax-neonfp16arith-ld64.c",
2208 "src/f16-gemm/gen/1x8-minmax-neonfp16arith-ld64.c",
2209 "src/f16-gemm/gen/1x16-minmax-neonfp16arith-ld64.c",
2210 "src/f16-gemm/gen/4x8-minmax-neonfp16arith-ld64.c",
2211 "src/f16-gemm/gen/4x16-minmax-neonfp16arith-ld64.c",
2212 "src/f16-gemm/gen/6x8-minmax-neonfp16arith-ld64.c",
2213 "src/f16-gemm/gen/6x16-minmax-neonfp16arith-ld64.c",
2214 "src/f16-gemm/gen/8x8-minmax-neonfp16arith-ld64.c",
2215 "src/f16-gemm/gen/8x16-minmax-neonfp16arith-ld64.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07002216 "src/f16-hswish/gen/hswish-neonfp16arith-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002217 "src/f16-hswish/gen/hswish-neonfp16arith-x16.c",
2218 "src/f16-igemm/gen/1x8-minmax-neonfp16arith-ld64.c",
2219 "src/f16-igemm/gen/1x16-minmax-neonfp16arith-ld64.c",
2220 "src/f16-igemm/gen/4x8-minmax-neonfp16arith-ld64.c",
2221 "src/f16-igemm/gen/4x16-minmax-neonfp16arith-ld64.c",
2222 "src/f16-igemm/gen/6x8-minmax-neonfp16arith-ld64.c",
2223 "src/f16-igemm/gen/6x16-minmax-neonfp16arith-ld64.c",
2224 "src/f16-igemm/gen/8x8-minmax-neonfp16arith-ld64.c",
2225 "src/f16-igemm/gen/8x16-minmax-neonfp16arith-ld64.c",
Frank Barchardb1966592020-05-12 13:47:06 -07002226 "src/f16-prelu/gen/neonfp16arith-2x8.c",
Frank Barcharda5316982020-07-23 13:19:28 -07002227 "src/f16-prelu/gen/neonfp16arith-2x16.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07002228 "src/f16-relu/gen/neonfp16arith-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002229 "src/f16-relu/gen/neonfp16arith-x16.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07002230 "src/f16-spmm/gen/8x1-minmax-neonfp16arith-x2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002231 "src/f16-spmm/gen/8x1-minmax-neonfp16arith.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07002232 "src/f16-spmm/gen/16x1-minmax-neonfp16arith-x2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002233 "src/f16-spmm/gen/16x1-minmax-neonfp16arith.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07002234 "src/f16-spmm/gen/24x1-minmax-neonfp16arith-x2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002235 "src/f16-spmm/gen/24x1-minmax-neonfp16arith.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07002236 "src/f16-spmm/gen/32x1-minmax-neonfp16arith-x2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002237 "src/f16-spmm/gen/32x1-minmax-neonfp16arith.c",
2238 "src/f16-vbinary/gen/vadd-minmax-neonfp16arith-x8.c",
2239 "src/f16-vbinary/gen/vadd-minmax-neonfp16arith-x16.c",
2240 "src/f16-vbinary/gen/vaddc-minmax-neonfp16arith-x8.c",
2241 "src/f16-vbinary/gen/vaddc-minmax-neonfp16arith-x16.c",
2242 "src/f16-vbinary/gen/vdiv-minmax-neonfp16arith-x8.c",
2243 "src/f16-vbinary/gen/vdiv-minmax-neonfp16arith-x16.c",
2244 "src/f16-vbinary/gen/vdivc-minmax-neonfp16arith-x8.c",
2245 "src/f16-vbinary/gen/vdivc-minmax-neonfp16arith-x16.c",
2246 "src/f16-vbinary/gen/vmax-neonfp16arith-x8.c",
2247 "src/f16-vbinary/gen/vmax-neonfp16arith-x16.c",
2248 "src/f16-vbinary/gen/vmaxc-neonfp16arith-x8.c",
2249 "src/f16-vbinary/gen/vmaxc-neonfp16arith-x16.c",
2250 "src/f16-vbinary/gen/vmin-neonfp16arith-x8.c",
2251 "src/f16-vbinary/gen/vmin-neonfp16arith-x16.c",
2252 "src/f16-vbinary/gen/vminc-neonfp16arith-x8.c",
2253 "src/f16-vbinary/gen/vminc-neonfp16arith-x16.c",
2254 "src/f16-vbinary/gen/vmul-minmax-neonfp16arith-x8.c",
2255 "src/f16-vbinary/gen/vmul-minmax-neonfp16arith-x16.c",
2256 "src/f16-vbinary/gen/vmulc-minmax-neonfp16arith-x8.c",
2257 "src/f16-vbinary/gen/vmulc-minmax-neonfp16arith-x16.c",
2258 "src/f16-vbinary/gen/vrdivc-minmax-neonfp16arith-x8.c",
2259 "src/f16-vbinary/gen/vrdivc-minmax-neonfp16arith-x16.c",
2260 "src/f16-vbinary/gen/vrsubc-minmax-neonfp16arith-x8.c",
2261 "src/f16-vbinary/gen/vrsubc-minmax-neonfp16arith-x16.c",
2262 "src/f16-vbinary/gen/vsub-minmax-neonfp16arith-x8.c",
2263 "src/f16-vbinary/gen/vsub-minmax-neonfp16arith-x16.c",
2264 "src/f16-vbinary/gen/vsubc-minmax-neonfp16arith-x8.c",
2265 "src/f16-vbinary/gen/vsubc-minmax-neonfp16arith-x16.c",
2266 "src/f16-vmulcaddc/gen/c8-minmax-neonfp16arith-2x.c",
2267 "src/f16-vmulcaddc/gen/c16-minmax-neonfp16arith-2x.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002268]
2269
Benoit Jacoba9644732020-08-13 12:48:55 -07002270NEONDOT_UKERNELS = [
Marat Dukhand4c83032020-09-09 23:20:12 -07002271 "src/qs8-gemm/gen/1x8c4-minmax-neondot.c",
Marat Dukhand4c83032020-09-09 23:20:12 -07002272 "src/qs8-gemm/gen/1x16c4-minmax-neondot.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002273 "src/qs8-gemm/gen/4x8c4-minmax-neondot.c",
Benoit Jacoba9644732020-08-13 12:48:55 -07002274 "src/qs8-gemm/gen/4x16c4-minmax-neondot.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002275 "src/qs8-gemm/gen/6x8c4-minmax-neondot.c",
Frank Barchard2fa17452020-09-22 09:51:46 -07002276 "src/qs8-gemm/gen/6x16c4-minmax-neondot.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002277 "src/qs8-gemm/gen/8x8c4-minmax-neondot.c",
Frank Barcharda48848f2020-09-24 00:59:24 -07002278 "src/qs8-gemm/gen/8x16c4-minmax-neondot.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002279 "src/qs8-gemm/gen/12x8c4-minmax-neondot.c",
Marat Dukhand4c83032020-09-09 23:20:12 -07002280 "src/qs8-igemm/gen/1x8c4-minmax-neondot.c",
Marat Dukhand4c83032020-09-09 23:20:12 -07002281 "src/qs8-igemm/gen/1x16c4-minmax-neondot.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002282 "src/qs8-igemm/gen/4x8c4-minmax-neondot.c",
Benoit Jacoba05487f2020-08-14 13:37:21 -07002283 "src/qs8-igemm/gen/4x16c4-minmax-neondot.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002284 "src/qs8-igemm/gen/6x8c4-minmax-neondot.c",
Frank Barchard2fa17452020-09-22 09:51:46 -07002285 "src/qs8-igemm/gen/6x16c4-minmax-neondot.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002286 "src/qs8-igemm/gen/8x8c4-minmax-neondot.c",
Frank Barcharda48848f2020-09-24 00:59:24 -07002287 "src/qs8-igemm/gen/8x16c4-minmax-neondot.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002288 "src/qs8-igemm/gen/12x8c4-minmax-neondot.c",
Benoit Jacoba9644732020-08-13 12:48:55 -07002289]
2290
Marat Dukhan08c4a432019-10-03 09:29:21 -07002291SSE_UKERNELS = [
Marat Dukhan99936602020-04-11 16:47:01 -07002292 "src/f32-avgpool/9p8x-minmax-sse-c4.c",
2293 "src/f32-avgpool/9x-minmax-sse-c4.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -07002294 "src/f32-clamp/gen/sse-x4.c",
2295 "src/f32-clamp/gen/sse-x8.c",
Erich Elsenb1233402020-06-08 15:53:15 -07002296 "src/f32-conv-hwc2chw/3x3s2p1c3x4-sse-1x1.c",
2297 "src/f32-conv-hwc2chw/3x3s2p1c3x4-sse-2x2.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002298 "src/f32-dwconv/gen/up4x4-minmax-sse-acc2.c",
2299 "src/f32-dwconv/gen/up4x4-minmax-sse.c",
2300 "src/f32-dwconv/gen/up4x9-minmax-sse-acc2.c",
2301 "src/f32-dwconv/gen/up4x9-minmax-sse.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002302 "src/f32-dwconv/gen/up4x25-minmax-sse-acc2.c",
2303 "src/f32-dwconv/gen/up4x25-minmax-sse.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002304 "src/f32-dwconv/gen/up8x4-minmax-sse-acc2.c",
2305 "src/f32-dwconv/gen/up8x4-minmax-sse.c",
2306 "src/f32-dwconv/gen/up8x9-minmax-sse-acc2.c",
2307 "src/f32-dwconv/gen/up8x9-minmax-sse.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002308 "src/f32-dwconv/gen/up8x25-minmax-sse-acc2.c",
2309 "src/f32-dwconv/gen/up8x25-minmax-sse.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07002310 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-1x4-acc2.c",
2311 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-1x4-acc3.c",
2312 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-1x4-acc4.c",
Marat Dukhan470078a2020-10-23 22:36:52 -07002313 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-1x4.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07002314 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-2x4-acc2.c",
Marat Dukhan470078a2020-10-23 22:36:52 -07002315 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-2x4.c",
2316 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-3x4.c",
2317 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-4x4.c",
2318 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-5x4.c",
2319 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-6x4.c",
Marat Dukhan0ff97182020-10-25 19:14:03 -07002320 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-sse-1x4-acc2.c",
2321 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-sse-1x4-acc3.c",
2322 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-sse-1x4-acc4.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07002323 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-sse-1x4.c",
Marat Dukhan0ff97182020-10-25 19:14:03 -07002324 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-sse-2x4-acc2.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07002325 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-sse-2x4.c",
2326 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-sse-3x4.c",
2327 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-sse-4x4.c",
Marat Dukhand0503892020-10-30 08:22:04 -07002328 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-1x4-acc2.c",
2329 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-1x4-acc3.c",
2330 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-1x4-acc4.c",
2331 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-1x4-acc5.c",
2332 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-1x4.c",
2333 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-2x4-acc2.c",
2334 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-2x4-acc3.c",
2335 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-2x4.c",
2336 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-3x4-acc2.c",
2337 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-3x4.c",
2338 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-4x4-acc2.c",
2339 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-4x4.c",
2340 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-5x4.c",
Marat Dukhanccca2142020-10-30 17:32:45 -07002341 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-sse-1x4-acc2.c",
2342 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-sse-1x4-acc3.c",
2343 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-sse-1x4-acc4.c",
2344 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-sse-1x4-acc5.c",
2345 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-sse-1x4.c",
2346 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-sse-2x4-acc2.c",
2347 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-sse-2x4-acc3.c",
2348 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-sse-2x4.c",
Marat Dukhanccca2142020-10-30 17:32:45 -07002349 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-sse-3x4-acc2.c",
Frank Barchard8ef44cd2020-11-03 12:30:23 -08002350 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-sse-3x4.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -07002351 "src/f32-gavgpool-cw/sse-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -07002352 "src/f32-gavgpool/7p7x-minmax-sse-c4.c",
2353 "src/f32-gavgpool/7x-minmax-sse-c4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002354 "src/f32-gemm/gen-inc/1x8inc-minmax-sse-dup.c",
2355 "src/f32-gemm/gen-inc/1x8inc-minmax-sse-load1.c",
2356 "src/f32-gemm/gen-inc/1x8s4inc-minmax-sse.c",
Marat Dukhan802fcae2020-12-11 14:37:25 -08002357 "src/f32-gemm/gen-inc/3x8inc-minmax-sse-dup.c",
2358 "src/f32-gemm/gen-inc/3x8inc-minmax-sse-load1.c",
2359 "src/f32-gemm/gen-inc/3x8s4inc-minmax-sse.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002360 "src/f32-gemm/gen-inc/4x8inc-minmax-sse-dup.c",
2361 "src/f32-gemm/gen-inc/4x8inc-minmax-sse-load1.c",
2362 "src/f32-gemm/gen-inc/4x8s4inc-minmax-sse.c",
Marat Dukhan802fcae2020-12-11 14:37:25 -08002363 "src/f32-gemm/gen-inc/5x8inc-minmax-sse-dup.c",
2364 "src/f32-gemm/gen-inc/5x8inc-minmax-sse-load1.c",
2365 "src/f32-gemm/gen-inc/5x8s4inc-minmax-sse.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002366 "src/f32-gemm/gen/1x8-minmax-sse-dup.c",
2367 "src/f32-gemm/gen/1x8-minmax-sse-load1.c",
2368 "src/f32-gemm/gen/1x8s4-minmax-sse.c",
Marat Dukhan802fcae2020-12-11 14:37:25 -08002369 "src/f32-gemm/gen/3x8-minmax-sse-dup.c",
2370 "src/f32-gemm/gen/3x8-minmax-sse-load1.c",
2371 "src/f32-gemm/gen/3x8s4-minmax-sse.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002372 "src/f32-gemm/gen/4x2c4-minmax-sse.c",
2373 "src/f32-gemm/gen/4x8-minmax-sse-dup.c",
2374 "src/f32-gemm/gen/4x8-minmax-sse-load1.c",
2375 "src/f32-gemm/gen/4x8s4-minmax-sse.c",
Marat Dukhan802fcae2020-12-11 14:37:25 -08002376 "src/f32-gemm/gen/5x8-minmax-sse-dup.c",
2377 "src/f32-gemm/gen/5x8-minmax-sse-load1.c",
2378 "src/f32-gemm/gen/5x8s4-minmax-sse.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07002379 "src/f32-hswish/gen/hswish-sse-x4.c",
2380 "src/f32-hswish/gen/hswish-sse-x8.c",
Marat Dukhan660fd192020-03-10 04:55:30 -07002381 "src/f32-ibilinear/gen/sse-c4.c",
2382 "src/f32-ibilinear/gen/sse-c8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002383 "src/f32-igemm/gen/1x8-minmax-sse-dup.c",
2384 "src/f32-igemm/gen/1x8-minmax-sse-load1.c",
2385 "src/f32-igemm/gen/1x8s4-minmax-sse.c",
Marat Dukhan802fcae2020-12-11 14:37:25 -08002386 "src/f32-igemm/gen/3x8-minmax-sse-dup.c",
2387 "src/f32-igemm/gen/3x8-minmax-sse-load1.c",
2388 "src/f32-igemm/gen/3x8s4-minmax-sse.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002389 "src/f32-igemm/gen/4x2c4-minmax-sse.c",
2390 "src/f32-igemm/gen/4x8-minmax-sse-dup.c",
2391 "src/f32-igemm/gen/4x8-minmax-sse-load1.c",
2392 "src/f32-igemm/gen/4x8s4-minmax-sse.c",
Marat Dukhan802fcae2020-12-11 14:37:25 -08002393 "src/f32-igemm/gen/5x8-minmax-sse-dup.c",
2394 "src/f32-igemm/gen/5x8-minmax-sse-load1.c",
2395 "src/f32-igemm/gen/5x8s4-minmax-sse.c",
Marat Dukhan99936602020-04-11 16:47:01 -07002396 "src/f32-maxpool/9p8x-minmax-sse-c4.c",
2397 "src/f32-pavgpool/9p8x-minmax-sse-c4.c",
2398 "src/f32-pavgpool/9x-minmax-sse-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002399 "src/f32-ppmm/gen/4x8-minmax-sse.c",
Marat Dukhan39b5e942020-06-24 15:03:48 -07002400 "src/f32-prelu/gen/sse-2x4.c",
2401 "src/f32-prelu/gen/sse-2x8.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07002402 "src/f32-relu/gen/sse-x4.c",
2403 "src/f32-relu/gen/sse-x8.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002404 "src/f32-rmax/sse.c",
Marat Dukhan355ab432020-04-09 19:01:52 -07002405 "src/f32-spmm/gen/4x1-minmax-sse.c",
2406 "src/f32-spmm/gen/8x1-minmax-sse.c",
Erich Elsen6e80fdc2020-06-09 15:35:37 -07002407 "src/f32-spmm/gen/16x1-minmax-sse.c",
Frank Barchard846c0c62020-10-26 15:01:39 -07002408 "src/f32-spmm/gen/32x1-minmax-sse.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002409 "src/f32-vbinary/gen/vadd-minmax-sse-x4.c",
2410 "src/f32-vbinary/gen/vadd-minmax-sse-x8.c",
2411 "src/f32-vbinary/gen/vaddc-minmax-sse-x4.c",
2412 "src/f32-vbinary/gen/vaddc-minmax-sse-x8.c",
2413 "src/f32-vbinary/gen/vdiv-minmax-sse-x4.c",
2414 "src/f32-vbinary/gen/vdiv-minmax-sse-x8.c",
2415 "src/f32-vbinary/gen/vdivc-minmax-sse-x4.c",
2416 "src/f32-vbinary/gen/vdivc-minmax-sse-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -08002417 "src/f32-vbinary/gen/vmax-sse-x4.c",
2418 "src/f32-vbinary/gen/vmax-sse-x8.c",
2419 "src/f32-vbinary/gen/vmaxc-sse-x4.c",
2420 "src/f32-vbinary/gen/vmaxc-sse-x8.c",
2421 "src/f32-vbinary/gen/vmin-sse-x4.c",
2422 "src/f32-vbinary/gen/vmin-sse-x8.c",
2423 "src/f32-vbinary/gen/vminc-sse-x4.c",
2424 "src/f32-vbinary/gen/vminc-sse-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002425 "src/f32-vbinary/gen/vmul-minmax-sse-x4.c",
2426 "src/f32-vbinary/gen/vmul-minmax-sse-x8.c",
2427 "src/f32-vbinary/gen/vmulc-minmax-sse-x4.c",
2428 "src/f32-vbinary/gen/vmulc-minmax-sse-x8.c",
2429 "src/f32-vbinary/gen/vrdivc-minmax-sse-x4.c",
2430 "src/f32-vbinary/gen/vrdivc-minmax-sse-x8.c",
2431 "src/f32-vbinary/gen/vrsubc-minmax-sse-x4.c",
2432 "src/f32-vbinary/gen/vrsubc-minmax-sse-x8.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07002433 "src/f32-vbinary/gen/vsqrdiff-sse-x4.c",
2434 "src/f32-vbinary/gen/vsqrdiff-sse-x8.c",
2435 "src/f32-vbinary/gen/vsqrdiffc-sse-x4.c",
2436 "src/f32-vbinary/gen/vsqrdiffc-sse-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002437 "src/f32-vbinary/gen/vsub-minmax-sse-x4.c",
2438 "src/f32-vbinary/gen/vsub-minmax-sse-x8.c",
2439 "src/f32-vbinary/gen/vsubc-minmax-sse-x4.c",
2440 "src/f32-vbinary/gen/vsubc-minmax-sse-x8.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07002441 "src/f32-vlrelu/gen/vlrelu-sse-x4.c",
2442 "src/f32-vlrelu/gen/vlrelu-sse-x8.c",
Marat Dukhan99936602020-04-11 16:47:01 -07002443 "src/f32-vmulcaddc/gen/c4-minmax-sse-2x.c",
2444 "src/f32-vmulcaddc/gen/c8-minmax-sse-2x.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07002445 "src/f32-vsqrt/gen/sse-sqrt-x4.c",
2446 "src/f32-vsqrt/gen/sse-sqrt-x8.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07002447 "src/f32-vunary/gen/vabs-sse-x4.c",
2448 "src/f32-vunary/gen/vabs-sse-x8.c",
2449 "src/f32-vunary/gen/vneg-sse-x4.c",
2450 "src/f32-vunary/gen/vneg-sse-x8.c",
2451 "src/f32-vunary/gen/vsqr-sse-x4.c",
2452 "src/f32-vunary/gen/vsqr-sse-x8.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07002453 "src/math/roundd-sse-addsub.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002454 "src/math/roundne-sse-addsub.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07002455 "src/math/roundu-sse-addsub.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07002456 "src/math/roundz-sse-addsub.c",
Marat Dukhan84000762020-06-29 18:38:43 -07002457 "src/math/sqrt-sse-hh1mac.c",
2458 "src/math/sqrt-sse-nr1mac.c",
2459 "src/math/sqrt-sse-nr2mac.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002460 "src/x32-fill/sse.c",
2461 "src/x32-packx/x4-sse.c",
2462 "src/x32-pad/sse.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002463]
2464
2465SSE2_UKERNELS = [
Marat Dukhan329da642019-11-19 21:44:39 -08002466 "src/f32-argmaxpool/4x-sse2-c4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002467 "src/f32-argmaxpool/9p8x-sse2-c4.c",
Marat Dukhan329da642019-11-19 21:44:39 -08002468 "src/f32-argmaxpool/9x-sse2-c4.c",
Marat Dukhan802fcae2020-12-11 14:37:25 -08002469 "src/f32-gemm/gen-inc/1x8inc-minmax-sse2-dup.c",
2470 "src/f32-gemm/gen-inc/3x8inc-minmax-sse2-dup.c",
2471 "src/f32-gemm/gen-inc/4x8inc-minmax-sse2-dup.c",
2472 "src/f32-gemm/gen-inc/5x8inc-minmax-sse2-dup.c",
2473 "src/f32-gemm/gen/1x8-minmax-sse2-dup.c",
2474 "src/f32-gemm/gen/3x8-minmax-sse2-dup.c",
2475 "src/f32-gemm/gen/4x8-minmax-sse2-dup.c",
2476 "src/f32-gemm/gen/5x8-minmax-sse2-dup.c",
2477 "src/f32-igemm/gen/1x8-minmax-sse2-dup.c",
2478 "src/f32-igemm/gen/3x8-minmax-sse2-dup.c",
2479 "src/f32-igemm/gen/4x8-minmax-sse2-dup.c",
2480 "src/f32-igemm/gen/5x8-minmax-sse2-dup.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08002481 "src/f32-prelu/gen/sse2-2x4.c",
2482 "src/f32-prelu/gen/sse2-2x8.c",
Marat Dukhanb39689d2020-01-24 13:32:20 -08002483 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x4.c",
Marat Dukhanb39689d2020-01-24 13:32:20 -08002484 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x8-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002485 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x8.c",
Marat Dukhanb39689d2020-01-24 13:32:20 -08002486 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12-acc2.c",
2487 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002488 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12.c",
Marat Dukhanb39689d2020-01-24 13:32:20 -08002489 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16-acc2.c",
2490 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002491 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16.c",
Marat Dukhanb39689d2020-01-24 13:32:20 -08002492 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20-acc2.c",
2493 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002494 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20.c",
Marat Dukhand243c1a2020-09-17 18:48:10 -07002495 "src/f32-sigmoid/gen/sse2-lut64-p2-div-x4.c",
2496 "src/f32-sigmoid/gen/sse2-lut64-p2-div-x8.c",
2497 "src/f32-sigmoid/gen/sse2-lut64-p2-div-x12.c",
2498 "src/f32-sigmoid/gen/sse2-lut64-p2-div-x16.c",
2499 "src/f32-sigmoid/gen/sse2-lut64-p2-div-x20.c",
2500 "src/f32-sigmoid/gen/sse2-lut64-p2-div-x24.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002501 "src/f32-sigmoid/gen/sse2-p5-div-x4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08002502 "src/f32-sigmoid/gen/sse2-p5-div-x8.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002503 "src/f32-sigmoid/gen/sse2-p5-div-x12.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08002504 "src/f32-sigmoid/gen/sse2-p5-div-x16.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002505 "src/f32-sigmoid/gen/sse2-p5-div-x20.c",
2506 "src/f32-sigmoid/gen/sse2-p5-div-x24.c",
Marat Dukhaned6baaf2020-12-01 15:07:08 -08002507 "src/f32-velu/gen/velu-sse2-rr2-lut16-p3-x4.c",
2508 "src/f32-velu/gen/velu-sse2-rr2-lut16-p3-x8.c",
2509 "src/f32-velu/gen/velu-sse2-rr2-lut16-p3-x12.c",
2510 "src/f32-velu/gen/velu-sse2-rr2-lut16-p3-x16.c",
2511 "src/f32-velu/gen/velu-sse2-rr2-lut16-p3-x20.c",
2512 "src/f32-velu/gen/velu-sse2-rr2-lut16-p3-x24.c",
2513 "src/f32-velu/gen/velu-sse2-rr2-p6-x4.c",
2514 "src/f32-velu/gen/velu-sse2-rr2-p6-x8.c",
2515 "src/f32-velu/gen/velu-sse2-rr2-p6-x12.c",
2516 "src/f32-velu/gen/velu-sse2-rr2-p6-x16.c",
2517 "src/f32-velu/gen/velu-sse2-rr2-p6-x20.c",
2518 "src/f32-velu/gen/velu-sse2-rr2-p6-x24.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07002519 "src/f32-vlrelu/gen/vlrelu-sse2-x4.c",
2520 "src/f32-vlrelu/gen/vlrelu-sse2-x8.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07002521 "src/f32-vrnd/gen/vrndd-sse2-x4.c",
2522 "src/f32-vrnd/gen/vrndd-sse2-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002523 "src/f32-vrnd/gen/vrndne-sse2-x4.c",
2524 "src/f32-vrnd/gen/vrndne-sse2-x8.c",
2525 "src/f32-vrnd/gen/vrndu-sse2-x4.c",
2526 "src/f32-vrnd/gen/vrndu-sse2-x8.c",
2527 "src/f32-vrnd/gen/vrndz-sse2-x4.c",
2528 "src/f32-vrnd/gen/vrndz-sse2-x8.c",
Marat Dukhanb7633f22020-11-20 16:34:56 -08002529 "src/math/exp-sse2-rr2-lut64-p2.c",
2530 "src/math/exp-sse2-rr2-p5.c",
Marat Dukhande390d42020-11-29 19:32:18 -08002531 "src/math/expm1minus-sse2-rr2-lut16-p3.c",
Marat Dukhana438aca2020-11-20 15:45:01 -08002532 "src/math/expm1minus-sse2-rr2-p6.c",
Frank Barchard3b800452020-11-22 12:12:35 -08002533 "src/math/expminus-sse2-rr2-p5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002534 "src/math/roundd-sse2-cvt.c",
2535 "src/math/roundne-sse2-cvt.c",
2536 "src/math/roundu-sse2-cvt.c",
2537 "src/math/roundz-sse2-cvt.c",
2538 "src/math/sigmoid-sse2-rr2-lut64-p2-div.c",
2539 "src/math/sigmoid-sse2-rr2-lut64-p2-nr1.c",
2540 "src/math/sigmoid-sse2-rr2-lut64-p2-nr2.c",
2541 "src/math/sigmoid-sse2-rr2-p5-div.c",
2542 "src/math/sigmoid-sse2-rr2-p5-nr1.c",
2543 "src/math/sigmoid-sse2-rr2-p5-nr2.c",
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07002544 "src/qs8-dwconv/gen/up8x9-minmax-sse2-mul16.c",
2545 "src/qs8-dwconv/gen/up16x9-minmax-sse2-mul16.c",
2546 "src/qs8-dwconv/gen/up24x9-minmax-sse2-mul16.c",
Marat Dukhan159688f2020-08-06 10:34:29 -07002547 "src/qs8-gavgpool/gen/7p7x-minmax-sse2-c8-acc2.c",
2548 "src/qs8-gavgpool/gen/7p7x-minmax-sse2-c16-acc2.c",
2549 "src/qs8-gavgpool/gen/7p7x-minmax-sse2-c24-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002550 "src/qs8-gavgpool/gen/7x-minmax-sse2-c8-acc2.c",
2551 "src/qs8-gavgpool/gen/7x-minmax-sse2-c16-acc2.c",
2552 "src/qs8-gavgpool/gen/7x-minmax-sse2-c24-acc2.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002553 "src/qs8-gemm/gen/1x4c2-minmax-sse2-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002554 "src/qs8-gemm/gen/1x4c2-minmax-sse2-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002555 "src/qs8-gemm/gen/1x4c2-xw-minmax-sse2.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002556 "src/qs8-gemm/gen/1x4c8-minmax-sse2-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002557 "src/qs8-gemm/gen/1x4c8-minmax-sse2-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002558 "src/qs8-gemm/gen/1x4c8-xw-minmax-sse2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002559 "src/qs8-gemm/gen/2x4c8-minmax-sse2-ld64.c",
2560 "src/qs8-gemm/gen/2x4c8-minmax-sse2-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002561 "src/qs8-gemm/gen/2x4c8-xw-minmax-sse2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002562 "src/qs8-gemm/gen/3x4c8-minmax-sse2-ld64.c",
2563 "src/qs8-gemm/gen/3x4c8-minmax-sse2-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002564 "src/qs8-gemm/gen/3x4c8-xw-minmax-sse2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002565 "src/qs8-gemm/gen/4x4c2-minmax-sse2-ld64.c",
2566 "src/qs8-gemm/gen/4x4c2-minmax-sse2-ld128.c",
2567 "src/qs8-gemm/gen/4x4c2-xw-minmax-sse2.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002568 "src/qs8-igemm/gen/1x4c2-minmax-sse2-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002569 "src/qs8-igemm/gen/1x4c2-minmax-sse2-ld128.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002570 "src/qs8-igemm/gen/1x4c8-minmax-sse2-ld64.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002571 "src/qs8-igemm/gen/1x4c8-minmax-sse2-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002572 "src/qs8-igemm/gen/2x4c8-minmax-sse2-ld64.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002573 "src/qs8-igemm/gen/2x4c8-minmax-sse2-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002574 "src/qs8-igemm/gen/3x4c8-minmax-sse2-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002575 "src/qs8-igemm/gen/3x4c8-minmax-sse2-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002576 "src/qs8-igemm/gen/4x4c2-minmax-sse2-ld64.c",
2577 "src/qs8-igemm/gen/4x4c2-minmax-sse2-ld128.c",
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07002578 "src/qs8-requantization/fp32-sse2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002579 "src/qs8-requantization/precise-sse2.c",
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07002580 "src/qs8-requantization/q31-sse2.c",
Marat Dukhand9f3ad42020-08-10 12:30:58 -07002581 "src/qs8-vadd/gen/minmax-sse2-mul16-ld64-x8.c",
2582 "src/qs8-vadd/gen/minmax-sse2-mul16-ld64-x16.c",
2583 "src/qs8-vadd/gen/minmax-sse2-mul16-ld64-x24.c",
2584 "src/qs8-vadd/gen/minmax-sse2-mul16-ld64-x32.c",
Marat Dukhan0270d9f2020-08-11 00:56:46 -07002585 "src/qs8-vaddc/gen/minmax-sse2-mul16-ld64-x8.c",
2586 "src/qs8-vaddc/gen/minmax-sse2-mul16-ld64-x16.c",
2587 "src/qs8-vaddc/gen/minmax-sse2-mul16-ld64-x24.c",
2588 "src/qs8-vaddc/gen/minmax-sse2-mul16-ld64-x32.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -07002589 "src/qu8-avgpool/9p8x-minmax-sse2-c8.c",
2590 "src/qu8-avgpool/9x-minmax-sse2-c8.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -07002591 "src/qu8-dwconv/up8x9-minmax-sse2.c",
2592 "src/qu8-gavgpool/7p7x-minmax-sse2-c8.c",
2593 "src/qu8-gavgpool/7x-minmax-sse2-c8.c",
2594 "src/qu8-gemm/2x4c8-minmax-sse2.c",
2595 "src/qu8-gemm/4x4c2-minmax-sse2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002596 "src/qu8-igemm/4x4c2-minmax-sse2.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07002597 "src/qu8-requantization/fp32-sse2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002598 "src/qu8-requantization/precise-sse2.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07002599 "src/qu8-requantization/q31-sse2.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -07002600 "src/qu8-vadd/minmax-sse2.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -07002601 "src/u8-clamp/sse2-x64.c",
Marat Dukhan99936602020-04-11 16:47:01 -07002602 "src/u8-maxpool/9p8x-minmax-sse2-c16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002603 "src/u8-rmax/sse2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002604 "src/x8-zip/x2-sse2.c",
2605 "src/x8-zip/x3-sse2.c",
2606 "src/x8-zip/x4-sse2.c",
2607 "src/x8-zip/xm-sse2.c",
Marat Dukhan57dccd82020-04-14 00:53:10 -07002608 "src/x32-unpool/sse2.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002609 "src/x32-zip/x2-sse2.c",
2610 "src/x32-zip/x3-sse2.c",
2611 "src/x32-zip/x4-sse2.c",
2612 "src/x32-zip/xm-sse2.c",
Marat Dukhanfe7acb62020-03-09 19:30:05 -07002613]
2614
2615SSSE3_UKERNELS = [
Frank Barchard35db7d02020-10-26 13:37:34 -07002616 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-1x4-acc2.c",
2617 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-1x4-acc3.c",
2618 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-1x4-acc4.c",
Marat Dukhan98f2eeb2020-10-23 23:13:41 -07002619 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-1x4.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07002620 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-2x4-acc2.c",
Marat Dukhan98f2eeb2020-10-23 23:13:41 -07002621 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-2x4.c",
2622 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-3x4.c",
2623 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-4x4.c",
2624 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-5x4.c",
2625 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-6x4.c",
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07002626 "src/qs8-dwconv/gen/up8x9-minmax-ssse3-mul16.c",
2627 "src/qs8-dwconv/gen/up16x9-minmax-ssse3-mul16.c",
2628 "src/qs8-dwconv/gen/up24x9-minmax-ssse3-mul16.c",
Marat Dukhan159688f2020-08-06 10:34:29 -07002629 "src/qs8-gavgpool/gen/7p7x-minmax-ssse3-c8-acc2.c",
2630 "src/qs8-gavgpool/gen/7p7x-minmax-ssse3-c16-acc2.c",
2631 "src/qs8-gavgpool/gen/7p7x-minmax-ssse3-c24-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002632 "src/qs8-gavgpool/gen/7x-minmax-ssse3-c8-acc2.c",
2633 "src/qs8-gavgpool/gen/7x-minmax-ssse3-c16-acc2.c",
2634 "src/qs8-gavgpool/gen/7x-minmax-ssse3-c24-acc2.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002635 "src/qs8-gemm/gen/1x4c2-minmax-ssse3-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002636 "src/qs8-gemm/gen/1x4c2-minmax-ssse3-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002637 "src/qs8-gemm/gen/1x4c2-xw-minmax-ssse3.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002638 "src/qs8-gemm/gen/1x4c8-minmax-ssse3-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002639 "src/qs8-gemm/gen/1x4c8-minmax-ssse3-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002640 "src/qs8-gemm/gen/1x4c8-xw-minmax-ssse3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002641 "src/qs8-gemm/gen/2x4c8-minmax-ssse3-ld64.c",
2642 "src/qs8-gemm/gen/2x4c8-minmax-ssse3-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002643 "src/qs8-gemm/gen/2x4c8-xw-minmax-ssse3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002644 "src/qs8-gemm/gen/3x4c8-minmax-ssse3-ld64.c",
2645 "src/qs8-gemm/gen/3x4c8-minmax-ssse3-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002646 "src/qs8-gemm/gen/3x4c8-xw-minmax-ssse3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002647 "src/qs8-gemm/gen/4x4c2-minmax-ssse3-ld64.c",
2648 "src/qs8-gemm/gen/4x4c2-minmax-ssse3-ld128.c",
2649 "src/qs8-gemm/gen/4x4c2-xw-minmax-ssse3.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002650 "src/qs8-igemm/gen/1x4c2-minmax-ssse3-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002651 "src/qs8-igemm/gen/1x4c2-minmax-ssse3-ld128.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002652 "src/qs8-igemm/gen/1x4c8-minmax-ssse3-ld64.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002653 "src/qs8-igemm/gen/1x4c8-minmax-ssse3-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002654 "src/qs8-igemm/gen/2x4c8-minmax-ssse3-ld64.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002655 "src/qs8-igemm/gen/2x4c8-minmax-ssse3-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002656 "src/qs8-igemm/gen/3x4c8-minmax-ssse3-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002657 "src/qs8-igemm/gen/3x4c8-minmax-ssse3-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002658 "src/qs8-igemm/gen/4x4c2-minmax-ssse3-ld64.c",
2659 "src/qs8-igemm/gen/4x4c2-minmax-ssse3-ld128.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -07002660 "src/qs8-requantization/precise-ssse3.c",
2661 "src/qs8-requantization/q31-ssse3.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07002662 "src/qu8-requantization/precise-ssse3.c",
2663 "src/qu8-requantization/q31-ssse3.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002664]
2665
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08002666SSE41_UKERNELS = [
Marat Dukhan40a672f2019-11-25 03:08:22 -08002667 "src/f32-prelu/gen/sse41-2x4.c",
2668 "src/f32-prelu/gen/sse41-2x8.c",
Marat Dukhand243c1a2020-09-17 18:48:10 -07002669 "src/f32-sigmoid/gen/sse41-lut64-p2-div-x4.c",
2670 "src/f32-sigmoid/gen/sse41-lut64-p2-div-x8.c",
2671 "src/f32-sigmoid/gen/sse41-lut64-p2-div-x12.c",
2672 "src/f32-sigmoid/gen/sse41-lut64-p2-div-x16.c",
2673 "src/f32-sigmoid/gen/sse41-lut64-p2-div-x20.c",
2674 "src/f32-sigmoid/gen/sse41-lut64-p2-div-x24.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002675 "src/f32-sigmoid/gen/sse41-p5-div-x4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08002676 "src/f32-sigmoid/gen/sse41-p5-div-x8.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002677 "src/f32-sigmoid/gen/sse41-p5-div-x12.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08002678 "src/f32-sigmoid/gen/sse41-p5-div-x16.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002679 "src/f32-sigmoid/gen/sse41-p5-div-x20.c",
2680 "src/f32-sigmoid/gen/sse41-p5-div-x24.c",
Marat Dukhaned6baaf2020-12-01 15:07:08 -08002681 "src/f32-velu/gen/velu-sse41-rr2-lut16-p3-x4.c",
2682 "src/f32-velu/gen/velu-sse41-rr2-lut16-p3-x8.c",
2683 "src/f32-velu/gen/velu-sse41-rr2-lut16-p3-x12.c",
2684 "src/f32-velu/gen/velu-sse41-rr2-lut16-p3-x16.c",
2685 "src/f32-velu/gen/velu-sse41-rr2-lut16-p3-x20.c",
2686 "src/f32-velu/gen/velu-sse41-rr2-lut16-p3-x24.c",
2687 "src/f32-velu/gen/velu-sse41-rr2-p6-x4.c",
2688 "src/f32-velu/gen/velu-sse41-rr2-p6-x8.c",
2689 "src/f32-velu/gen/velu-sse41-rr2-p6-x12.c",
2690 "src/f32-velu/gen/velu-sse41-rr2-p6-x16.c",
2691 "src/f32-velu/gen/velu-sse41-rr2-p6-x20.c",
2692 "src/f32-velu/gen/velu-sse41-rr2-p6-x24.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07002693 "src/f32-vlrelu/gen/vlrelu-sse41-x4.c",
2694 "src/f32-vlrelu/gen/vlrelu-sse41-x8.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07002695 "src/f32-vrnd/gen/vrndd-sse41-x4.c",
2696 "src/f32-vrnd/gen/vrndd-sse41-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002697 "src/f32-vrnd/gen/vrndne-sse41-x4.c",
2698 "src/f32-vrnd/gen/vrndne-sse41-x8.c",
2699 "src/f32-vrnd/gen/vrndu-sse41-x4.c",
2700 "src/f32-vrnd/gen/vrndu-sse41-x8.c",
2701 "src/f32-vrnd/gen/vrndz-sse41-x4.c",
2702 "src/f32-vrnd/gen/vrndz-sse41-x8.c",
2703 "src/math/roundd-sse41.c",
2704 "src/math/roundne-sse41.c",
2705 "src/math/roundu-sse41.c",
2706 "src/math/roundz-sse41.c",
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07002707 "src/qs8-dwconv/gen/up8x9-minmax-sse41-mul16.c",
2708 "src/qs8-dwconv/gen/up16x9-minmax-sse41-mul16.c",
2709 "src/qs8-dwconv/gen/up24x9-minmax-sse41-mul16.c",
Marat Dukhan159688f2020-08-06 10:34:29 -07002710 "src/qs8-gavgpool/gen/7p7x-minmax-sse41-c8-acc2.c",
2711 "src/qs8-gavgpool/gen/7p7x-minmax-sse41-c16-acc2.c",
2712 "src/qs8-gavgpool/gen/7p7x-minmax-sse41-c24-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002713 "src/qs8-gavgpool/gen/7x-minmax-sse41-c8-acc2.c",
2714 "src/qs8-gavgpool/gen/7x-minmax-sse41-c16-acc2.c",
2715 "src/qs8-gavgpool/gen/7x-minmax-sse41-c24-acc2.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002716 "src/qs8-gemm/gen/1x4c2-minmax-sse41-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002717 "src/qs8-gemm/gen/1x4c2-minmax-sse41-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002718 "src/qs8-gemm/gen/1x4c2-xw-minmax-sse41.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002719 "src/qs8-gemm/gen/1x4c8-minmax-sse41-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002720 "src/qs8-gemm/gen/1x4c8-minmax-sse41-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002721 "src/qs8-gemm/gen/1x4c8-xw-minmax-sse41.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002722 "src/qs8-gemm/gen/2x4c8-minmax-sse41-ld64.c",
2723 "src/qs8-gemm/gen/2x4c8-minmax-sse41-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002724 "src/qs8-gemm/gen/2x4c8-xw-minmax-sse41.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002725 "src/qs8-gemm/gen/3x4c8-minmax-sse41-ld64.c",
2726 "src/qs8-gemm/gen/3x4c8-minmax-sse41-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002727 "src/qs8-gemm/gen/3x4c8-xw-minmax-sse41.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002728 "src/qs8-gemm/gen/4x4c2-minmax-sse41-ld64.c",
2729 "src/qs8-gemm/gen/4x4c2-minmax-sse41-ld128.c",
2730 "src/qs8-gemm/gen/4x4c2-xw-minmax-sse41.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002731 "src/qs8-igemm/gen/1x4c2-minmax-sse41-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002732 "src/qs8-igemm/gen/1x4c2-minmax-sse41-ld128.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002733 "src/qs8-igemm/gen/1x4c8-minmax-sse41-ld64.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002734 "src/qs8-igemm/gen/1x4c8-minmax-sse41-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002735 "src/qs8-igemm/gen/2x4c8-minmax-sse41-ld64.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002736 "src/qs8-igemm/gen/2x4c8-minmax-sse41-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002737 "src/qs8-igemm/gen/3x4c8-minmax-sse41-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002738 "src/qs8-igemm/gen/3x4c8-minmax-sse41-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002739 "src/qs8-igemm/gen/4x4c2-minmax-sse41-ld64.c",
2740 "src/qs8-igemm/gen/4x4c2-minmax-sse41-ld128.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -07002741 "src/qs8-requantization/fp32-sse4.c",
2742 "src/qs8-requantization/precise-sse4.c",
2743 "src/qs8-requantization/q31-sse4.c",
Marat Dukhand9f3ad42020-08-10 12:30:58 -07002744 "src/qs8-vadd/gen/minmax-sse41-mul16-ld64-x8.c",
2745 "src/qs8-vadd/gen/minmax-sse41-mul16-ld64-x16.c",
2746 "src/qs8-vadd/gen/minmax-sse41-mul16-ld64-x24.c",
2747 "src/qs8-vadd/gen/minmax-sse41-mul16-ld64-x32.c",
Marat Dukhanbb9225e2020-09-06 22:40:56 -07002748 "src/qs8-vadd/gen/minmax-sse41-mul32-ld32-x8.c",
2749 "src/qs8-vadd/gen/minmax-sse41-mul32-ld32-x16.c",
2750 "src/qs8-vadd/gen/minmax-sse41-mul32-ld32-x24.c",
2751 "src/qs8-vadd/gen/minmax-sse41-mul32-ld32-x32.c",
Marat Dukhan0270d9f2020-08-11 00:56:46 -07002752 "src/qs8-vaddc/gen/minmax-sse41-mul16-ld64-x8.c",
2753 "src/qs8-vaddc/gen/minmax-sse41-mul16-ld64-x16.c",
2754 "src/qs8-vaddc/gen/minmax-sse41-mul16-ld64-x24.c",
2755 "src/qs8-vaddc/gen/minmax-sse41-mul16-ld64-x32.c",
Marat Dukhanbb9225e2020-09-06 22:40:56 -07002756 "src/qs8-vaddc/gen/minmax-sse41-mul32-ld32-x8.c",
2757 "src/qs8-vaddc/gen/minmax-sse41-mul32-ld32-x16.c",
2758 "src/qs8-vaddc/gen/minmax-sse41-mul32-ld32-x24.c",
2759 "src/qs8-vaddc/gen/minmax-sse41-mul32-ld32-x32.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07002760 "src/qu8-requantization/precise-sse4.c",
2761 "src/qu8-requantization/q31-sse4.c",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08002762]
2763
Marat Dukhan08c4a432019-10-03 09:29:21 -07002764AVX_UKERNELS = [
Marat Dukhan5c5fa962020-03-10 18:38:33 -07002765 "src/f32-clamp/gen/avx-x8.c",
2766 "src/f32-clamp/gen/avx-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002767 "src/f32-dwconv/gen/up8x4-minmax-avx-acc2.c",
2768 "src/f32-dwconv/gen/up8x4-minmax-avx.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002769 "src/f32-dwconv/gen/up8x9-minmax-avx-acc2.c",
2770 "src/f32-dwconv/gen/up8x9-minmax-avx.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002771 "src/f32-dwconv/gen/up8x25-minmax-avx-acc2.c",
2772 "src/f32-dwconv/gen/up8x25-minmax-avx.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002773 "src/f32-dwconv/gen/up16x4-minmax-avx-acc2.c",
2774 "src/f32-dwconv/gen/up16x4-minmax-avx.c",
2775 "src/f32-dwconv/gen/up16x9-minmax-avx-acc2.c",
2776 "src/f32-dwconv/gen/up16x9-minmax-avx.c",
2777 "src/f32-dwconv/gen/up16x25-minmax-avx-acc2.c",
2778 "src/f32-dwconv/gen/up16x25-minmax-avx.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002779 "src/f32-gemm/gen-inc/1x8inc-minmax-avx-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002780 "src/f32-gemm/gen-inc/1x16inc-minmax-avx-broadcast.c",
2781 "src/f32-gemm/gen-inc/3x16inc-minmax-avx-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002782 "src/f32-gemm/gen-inc/4x8inc-minmax-avx-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002783 "src/f32-gemm/gen-inc/4x16inc-minmax-avx-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002784 "src/f32-gemm/gen-inc/5x8inc-minmax-avx-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002785 "src/f32-gemm/gen-inc/5x16inc-minmax-avx-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002786 "src/f32-gemm/gen-inc/6x8inc-minmax-avx-broadcast.c",
2787 "src/f32-gemm/gen-inc/7x8inc-minmax-avx-broadcast.c",
2788 "src/f32-gemm/gen/1x8-minmax-avx-broadcast.c",
2789 "src/f32-gemm/gen/1x16-minmax-avx-broadcast.c",
2790 "src/f32-gemm/gen/3x16-minmax-avx-broadcast.c",
2791 "src/f32-gemm/gen/4x8-minmax-avx-broadcast.c",
2792 "src/f32-gemm/gen/4x16-minmax-avx-broadcast.c",
2793 "src/f32-gemm/gen/5x8-minmax-avx-broadcast.c",
2794 "src/f32-gemm/gen/5x16-minmax-avx-broadcast.c",
2795 "src/f32-gemm/gen/6x8-minmax-avx-broadcast.c",
2796 "src/f32-gemm/gen/7x8-minmax-avx-broadcast.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07002797 "src/f32-hswish/gen/hswish-avx-x8.c",
2798 "src/f32-hswish/gen/hswish-avx-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002799 "src/f32-igemm/gen/1x8-minmax-avx-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002800 "src/f32-igemm/gen/1x16-minmax-avx-broadcast.c",
2801 "src/f32-igemm/gen/3x16-minmax-avx-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002802 "src/f32-igemm/gen/4x8-minmax-avx-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002803 "src/f32-igemm/gen/4x16-minmax-avx-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002804 "src/f32-igemm/gen/5x8-minmax-avx-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002805 "src/f32-igemm/gen/5x16-minmax-avx-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002806 "src/f32-igemm/gen/6x8-minmax-avx-broadcast.c",
2807 "src/f32-igemm/gen/7x8-minmax-avx-broadcast.c",
Marat Dukhan90eca0a2020-03-11 00:52:23 -07002808 "src/f32-prelu/gen/avx-2x8.c",
2809 "src/f32-prelu/gen/avx-2x16.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07002810 "src/f32-relu/gen/avx-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002811 "src/f32-relu/gen/avx-x16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002812 "src/f32-rmax/avx.c",
T.J. Alumbaughdc2b29c2020-10-14 13:56:08 -07002813 "src/f32-sigmoid/gen/avx-rr2-p5-div-x8.c",
2814 "src/f32-sigmoid/gen/avx-rr2-p5-div-x16.c",
2815 "src/f32-sigmoid/gen/avx-rr2-p5-div-x24.c",
2816 "src/f32-sigmoid/gen/avx-rr2-p5-div-x32.c",
2817 "src/f32-sigmoid/gen/avx-rr2-p5-div-x40.c",
2818 "src/f32-sigmoid/gen/avx-rr2-p5-div-x48.c",
2819 "src/f32-sigmoid/gen/avx-rr2-p5-div-x56.c",
2820 "src/f32-sigmoid/gen/avx-rr2-p5-div-x64.c",
2821 "src/f32-sigmoid/gen/avx-rr2-p5-div-x72.c",
2822 "src/f32-sigmoid/gen/avx-rr2-p5-div-x80.c",
2823 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x8.c",
2824 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x16.c",
2825 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x24.c",
2826 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x32.c",
2827 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x40.c",
2828 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x48.c",
2829 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x56.c",
2830 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x64.c",
2831 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x72.c",
2832 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x80.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002833 "src/f32-vbinary/gen/vadd-minmax-avx-x8.c",
2834 "src/f32-vbinary/gen/vadd-minmax-avx-x16.c",
2835 "src/f32-vbinary/gen/vaddc-minmax-avx-x8.c",
2836 "src/f32-vbinary/gen/vaddc-minmax-avx-x16.c",
2837 "src/f32-vbinary/gen/vdiv-minmax-avx-x8.c",
2838 "src/f32-vbinary/gen/vdiv-minmax-avx-x16.c",
2839 "src/f32-vbinary/gen/vdivc-minmax-avx-x8.c",
2840 "src/f32-vbinary/gen/vdivc-minmax-avx-x16.c",
Marat Dukhan9a88efe2019-12-10 15:54:24 -08002841 "src/f32-vbinary/gen/vmax-avx-x8.c",
2842 "src/f32-vbinary/gen/vmax-avx-x16.c",
2843 "src/f32-vbinary/gen/vmaxc-avx-x8.c",
2844 "src/f32-vbinary/gen/vmaxc-avx-x16.c",
2845 "src/f32-vbinary/gen/vmin-avx-x8.c",
2846 "src/f32-vbinary/gen/vmin-avx-x16.c",
2847 "src/f32-vbinary/gen/vminc-avx-x8.c",
2848 "src/f32-vbinary/gen/vminc-avx-x16.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002849 "src/f32-vbinary/gen/vmul-minmax-avx-x8.c",
2850 "src/f32-vbinary/gen/vmul-minmax-avx-x16.c",
2851 "src/f32-vbinary/gen/vmulc-minmax-avx-x8.c",
2852 "src/f32-vbinary/gen/vmulc-minmax-avx-x16.c",
2853 "src/f32-vbinary/gen/vrdivc-minmax-avx-x8.c",
2854 "src/f32-vbinary/gen/vrdivc-minmax-avx-x16.c",
2855 "src/f32-vbinary/gen/vrsubc-minmax-avx-x8.c",
2856 "src/f32-vbinary/gen/vrsubc-minmax-avx-x16.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07002857 "src/f32-vbinary/gen/vsqrdiff-avx-x8.c",
2858 "src/f32-vbinary/gen/vsqrdiff-avx-x16.c",
2859 "src/f32-vbinary/gen/vsqrdiffc-avx-x8.c",
2860 "src/f32-vbinary/gen/vsqrdiffc-avx-x16.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002861 "src/f32-vbinary/gen/vsub-minmax-avx-x8.c",
2862 "src/f32-vbinary/gen/vsub-minmax-avx-x16.c",
2863 "src/f32-vbinary/gen/vsubc-minmax-avx-x8.c",
2864 "src/f32-vbinary/gen/vsubc-minmax-avx-x16.c",
Marat Dukhaned6baaf2020-12-01 15:07:08 -08002865 "src/f32-velu/gen/velu-avx-rr2-lut4-p4-perm-x8.c",
2866 "src/f32-velu/gen/velu-avx-rr2-lut4-p4-perm-x16.c",
2867 "src/f32-velu/gen/velu-avx-rr2-lut4-p4-perm-x24.c",
2868 "src/f32-velu/gen/velu-avx-rr2-lut4-p4-perm-x32.c",
2869 "src/f32-velu/gen/velu-avx-rr2-lut4-p4-perm-x40.c",
2870 "src/f32-velu/gen/velu-avx-rr2-lut4-p4-perm-x48.c",
2871 "src/f32-velu/gen/velu-avx-rr2-lut16-p3-x8.c",
2872 "src/f32-velu/gen/velu-avx-rr2-lut16-p3-x16.c",
2873 "src/f32-velu/gen/velu-avx-rr2-lut16-p3-x24.c",
2874 "src/f32-velu/gen/velu-avx-rr2-lut16-p3-x32.c",
2875 "src/f32-velu/gen/velu-avx-rr2-lut16-p3-x40.c",
2876 "src/f32-velu/gen/velu-avx-rr2-lut16-p3-x48.c",
2877 "src/f32-velu/gen/velu-avx-rr2-p6-x8.c",
2878 "src/f32-velu/gen/velu-avx-rr2-p6-x16.c",
2879 "src/f32-velu/gen/velu-avx-rr2-p6-x24.c",
2880 "src/f32-velu/gen/velu-avx-rr2-p6-x32.c",
2881 "src/f32-velu/gen/velu-avx-rr2-p6-x40.c",
2882 "src/f32-velu/gen/velu-avx-rr2-p6-x48.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07002883 "src/f32-vlrelu/gen/vlrelu-avx-x8.c",
2884 "src/f32-vlrelu/gen/vlrelu-avx-x16.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07002885 "src/f32-vrnd/gen/vrndd-avx-x8.c",
2886 "src/f32-vrnd/gen/vrndd-avx-x16.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002887 "src/f32-vrnd/gen/vrndne-avx-x8.c",
2888 "src/f32-vrnd/gen/vrndne-avx-x16.c",
2889 "src/f32-vrnd/gen/vrndu-avx-x8.c",
2890 "src/f32-vrnd/gen/vrndu-avx-x16.c",
2891 "src/f32-vrnd/gen/vrndz-avx-x8.c",
2892 "src/f32-vrnd/gen/vrndz-avx-x16.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07002893 "src/f32-vscale/avx-x32.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07002894 "src/f32-vsqrt/gen/avx-sqrt-x8.c",
2895 "src/f32-vsqrt/gen/avx-sqrt-x16.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07002896 "src/f32-vunary/gen/vabs-avx-x8.c",
2897 "src/f32-vunary/gen/vabs-avx-x16.c",
2898 "src/f32-vunary/gen/vneg-avx-x8.c",
2899 "src/f32-vunary/gen/vneg-avx-x16.c",
2900 "src/f32-vunary/gen/vsqr-avx-x8.c",
2901 "src/f32-vunary/gen/vsqr-avx-x16.c",
Marat Dukhanb7633f22020-11-20 16:34:56 -08002902 "src/math/exp-avx-rr2-p5.c",
Marat Dukhande390d42020-11-29 19:32:18 -08002903 "src/math/expm1minus-avx-rr2-lut4-p4-perm.c",
2904 "src/math/expm1minus-avx-rr2-lut16-p3.c",
2905 "src/math/expm1minus-avx-rr2-p6.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002906 "src/math/sigmoid-avx-rr2-lut64-p2-div.c",
Marat Dukhan2e4727b2020-09-24 00:48:39 -07002907 "src/math/sigmoid-avx-rr2-p5-div.c",
Marat Dukhan39ab16f2020-10-16 14:30:35 -07002908 "src/math/sigmoid-avx-rr2-p5-nr1.c",
2909 "src/math/sigmoid-avx-rr2-p5-nr2.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002910]
2911
Marat Dukhan1566fee2020-08-02 21:55:41 -07002912XOP_UKERNELS = [
2913 "src/qs8-gemm/gen/1x4c2-minmax-xop-ld64.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002914 "src/qs8-gemm/gen/1x4c2-minmax-xop-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002915 "src/qs8-gemm/gen/1x4c2-xw-minmax-xop.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002916 "src/qs8-gemm/gen/1x4c8-minmax-xop-ld64.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002917 "src/qs8-gemm/gen/1x4c8-minmax-xop-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002918 "src/qs8-gemm/gen/1x4c8-xw-minmax-xop.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002919 "src/qs8-gemm/gen/2x4c8-minmax-xop-ld64.c",
2920 "src/qs8-gemm/gen/2x4c8-minmax-xop-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002921 "src/qs8-gemm/gen/2x4c8-xw-minmax-xop.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002922 "src/qs8-gemm/gen/3x4c8-minmax-xop-ld64.c",
2923 "src/qs8-gemm/gen/3x4c8-minmax-xop-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002924 "src/qs8-gemm/gen/3x4c8-xw-minmax-xop.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002925 "src/qs8-gemm/gen/4x4c2-minmax-xop-ld64.c",
2926 "src/qs8-gemm/gen/4x4c2-minmax-xop-ld128.c",
2927 "src/qs8-gemm/gen/4x4c2-xw-minmax-xop.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002928 "src/qs8-igemm/gen/1x4c2-minmax-xop-ld64.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002929 "src/qs8-igemm/gen/1x4c2-minmax-xop-ld128.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002930 "src/qs8-igemm/gen/1x4c8-minmax-xop-ld64.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002931 "src/qs8-igemm/gen/1x4c8-minmax-xop-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002932 "src/qs8-igemm/gen/2x4c8-minmax-xop-ld64.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002933 "src/qs8-igemm/gen/2x4c8-minmax-xop-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002934 "src/qs8-igemm/gen/3x4c8-minmax-xop-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002935 "src/qs8-igemm/gen/3x4c8-minmax-xop-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002936 "src/qs8-igemm/gen/4x4c2-minmax-xop-ld64.c",
2937 "src/qs8-igemm/gen/4x4c2-minmax-xop-ld128.c",
Marat Dukhanbb9225e2020-09-06 22:40:56 -07002938 "src/qs8-vadd/gen/minmax-xop-mul32-ld32-x8.c",
2939 "src/qs8-vadd/gen/minmax-xop-mul32-ld32-x16.c",
2940 "src/qs8-vadd/gen/minmax-xop-mul32-ld32-x24.c",
2941 "src/qs8-vadd/gen/minmax-xop-mul32-ld32-x32.c",
2942 "src/qs8-vaddc/gen/minmax-xop-mul32-ld32-x8.c",
2943 "src/qs8-vaddc/gen/minmax-xop-mul32-ld32-x16.c",
2944 "src/qs8-vaddc/gen/minmax-xop-mul32-ld32-x24.c",
2945 "src/qs8-vaddc/gen/minmax-xop-mul32-ld32-x32.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002946]
2947
Marat Dukhanfda12b82019-11-21 12:27:59 -08002948FMA3_UKERNELS = [
Marat Dukhan1c587112020-04-08 20:04:28 -07002949 "src/f32-dwconv/gen/up8x4-minmax-fma3-acc2.c",
2950 "src/f32-dwconv/gen/up8x4-minmax-fma3.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002951 "src/f32-dwconv/gen/up8x9-minmax-fma3-acc2.c",
2952 "src/f32-dwconv/gen/up8x9-minmax-fma3.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002953 "src/f32-dwconv/gen/up8x25-minmax-fma3-acc2.c",
2954 "src/f32-dwconv/gen/up8x25-minmax-fma3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002955 "src/f32-dwconv/gen/up16x4-minmax-fma3-acc2.c",
2956 "src/f32-dwconv/gen/up16x4-minmax-fma3.c",
2957 "src/f32-dwconv/gen/up16x9-minmax-fma3-acc2.c",
2958 "src/f32-dwconv/gen/up16x9-minmax-fma3.c",
2959 "src/f32-dwconv/gen/up16x25-minmax-fma3-acc2.c",
2960 "src/f32-dwconv/gen/up16x25-minmax-fma3.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002961 "src/f32-gemm/gen-inc/1x8inc-minmax-fma3-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002962 "src/f32-gemm/gen-inc/1x16inc-minmax-fma3-broadcast.c",
2963 "src/f32-gemm/gen-inc/1x16s4inc-minmax-fma3-broadcast.c",
2964 "src/f32-gemm/gen-inc/3x16inc-minmax-fma3-broadcast.c",
2965 "src/f32-gemm/gen-inc/3x16s4inc-minmax-fma3-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002966 "src/f32-gemm/gen-inc/4x8inc-minmax-fma3-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002967 "src/f32-gemm/gen-inc/4x16inc-minmax-fma3-broadcast.c",
2968 "src/f32-gemm/gen-inc/4x16s4inc-minmax-fma3-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002969 "src/f32-gemm/gen-inc/5x8inc-minmax-fma3-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002970 "src/f32-gemm/gen-inc/5x16inc-minmax-fma3-broadcast.c",
2971 "src/f32-gemm/gen-inc/5x16s4inc-minmax-fma3-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002972 "src/f32-gemm/gen-inc/6x8inc-minmax-fma3-broadcast.c",
2973 "src/f32-gemm/gen-inc/7x8inc-minmax-fma3-broadcast.c",
2974 "src/f32-gemm/gen-inc/8x8inc-minmax-fma3-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002975 "src/f32-gemm/gen/1x8-minmax-fma3-broadcast.c",
2976 "src/f32-gemm/gen/1x16-minmax-fma3-broadcast.c",
2977 "src/f32-gemm/gen/1x16s4-minmax-fma3-broadcast.c",
2978 "src/f32-gemm/gen/3x16-minmax-fma3-broadcast.c",
2979 "src/f32-gemm/gen/3x16s4-minmax-fma3-broadcast.c",
2980 "src/f32-gemm/gen/4x8-minmax-fma3-broadcast.c",
2981 "src/f32-gemm/gen/4x16-minmax-fma3-broadcast.c",
2982 "src/f32-gemm/gen/4x16s4-minmax-fma3-broadcast.c",
2983 "src/f32-gemm/gen/5x8-minmax-fma3-broadcast.c",
2984 "src/f32-gemm/gen/5x16-minmax-fma3-broadcast.c",
2985 "src/f32-gemm/gen/5x16s4-minmax-fma3-broadcast.c",
2986 "src/f32-gemm/gen/6x8-minmax-fma3-broadcast.c",
2987 "src/f32-gemm/gen/7x8-minmax-fma3-broadcast.c",
2988 "src/f32-gemm/gen/8x8-minmax-fma3-broadcast.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07002989 "src/f32-hswish/gen/hswish-fma3-x8.c",
2990 "src/f32-hswish/gen/hswish-fma3-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002991 "src/f32-igemm/gen/1x8-minmax-fma3-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002992 "src/f32-igemm/gen/1x16-minmax-fma3-broadcast.c",
2993 "src/f32-igemm/gen/1x16s4-minmax-fma3-broadcast.c",
2994 "src/f32-igemm/gen/3x16-minmax-fma3-broadcast.c",
2995 "src/f32-igemm/gen/3x16s4-minmax-fma3-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002996 "src/f32-igemm/gen/4x8-minmax-fma3-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002997 "src/f32-igemm/gen/4x16-minmax-fma3-broadcast.c",
2998 "src/f32-igemm/gen/4x16s4-minmax-fma3-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002999 "src/f32-igemm/gen/5x8-minmax-fma3-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003000 "src/f32-igemm/gen/5x16-minmax-fma3-broadcast.c",
3001 "src/f32-igemm/gen/5x16s4-minmax-fma3-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07003002 "src/f32-igemm/gen/6x8-minmax-fma3-broadcast.c",
3003 "src/f32-igemm/gen/7x8-minmax-fma3-broadcast.c",
3004 "src/f32-igemm/gen/8x8-minmax-fma3-broadcast.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07003005 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x8.c",
3006 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x16.c",
3007 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x24.c",
3008 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x32.c",
3009 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x40.c",
3010 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x48.c",
3011 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x56.c",
3012 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x64.c",
Marat Dukhan84000762020-06-29 18:38:43 -07003013 "src/math/sqrt-fma3-nr1fma.c",
Marat Dukhan84000762020-06-29 18:38:43 -07003014 "src/math/sqrt-fma3-nr1fma1adj.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003015 "src/math/sqrt-fma3-nr2fma.c",
Marat Dukhanfda12b82019-11-21 12:27:59 -08003016]
3017
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003018AVX2_UKERNELS = [
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003019 "src/f32-raddexpminusmax/gen/avx2-p5-x64-acc2.c",
3020 "src/f32-raddexpminusmax/gen/avx2-p5-x64-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003021 "src/f32-raddexpminusmax/gen/avx2-p5-x64.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003022 "src/f32-raddexpminusmax/gen/avx2-p5-x72-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003023 "src/f32-raddexpminusmax/gen/avx2-p5-x72.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003024 "src/f32-raddexpminusmax/gen/avx2-p5-x80-acc2.c",
3025 "src/f32-raddexpminusmax/gen/avx2-p5-x80-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003026 "src/f32-raddexpminusmax/gen/avx2-p5-x80.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003027 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc2.c",
3028 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc3.c",
3029 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc6.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003030 "src/f32-raddexpminusmax/gen/avx2-p5-x96.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003031 "src/f32-raddextexp/gen/avx2-p5-x64-acc2.c",
3032 "src/f32-raddextexp/gen/avx2-p5-x64-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003033 "src/f32-raddextexp/gen/avx2-p5-x64.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003034 "src/f32-raddextexp/gen/avx2-p5-x72-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003035 "src/f32-raddextexp/gen/avx2-p5-x72.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003036 "src/f32-raddextexp/gen/avx2-p5-x80-acc2.c",
3037 "src/f32-raddextexp/gen/avx2-p5-x80-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003038 "src/f32-raddextexp/gen/avx2-p5-x80.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003039 "src/f32-raddextexp/gen/avx2-p5-x96-acc2.c",
3040 "src/f32-raddextexp/gen/avx2-p5-x96-acc3.c",
3041 "src/f32-raddextexp/gen/avx2-p5-x96-acc6.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003042 "src/f32-raddextexp/gen/avx2-p5-x96.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003043 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64-acc2.c",
3044 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003045 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003046 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x72-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003047 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x72.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003048 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80-acc2.c",
3049 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003050 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003051 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc2.c",
3052 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc3.c",
3053 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc6.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003054 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96.c",
Marat Dukhanfa0a4322020-01-06 16:14:29 -08003055 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x8.c",
3056 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x16.c",
3057 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x24.c",
3058 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x32.c",
3059 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x40.c",
3060 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x48.c",
3061 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x56.c",
3062 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x64.c",
3063 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x72.c",
3064 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x80.c",
3065 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x8.c",
3066 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x16.c",
3067 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x24.c",
3068 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x32.c",
3069 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x40.c",
3070 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x48.c",
3071 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x56.c",
3072 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x64.c",
3073 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x72.c",
3074 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x80.c",
3075 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x8.c",
3076 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x16.c",
3077 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x24.c",
3078 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x32.c",
3079 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x40.c",
3080 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x48.c",
3081 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x56.c",
3082 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x64.c",
3083 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x72.c",
3084 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x80.c",
Marat Dukhaned6baaf2020-12-01 15:07:08 -08003085 "src/f32-velu/gen/velu-avx2-rr1-lut4-p4-perm-x8.c",
3086 "src/f32-velu/gen/velu-avx2-rr1-lut4-p4-perm-x16.c",
3087 "src/f32-velu/gen/velu-avx2-rr1-lut4-p4-perm-x24.c",
3088 "src/f32-velu/gen/velu-avx2-rr1-lut4-p4-perm-x32.c",
3089 "src/f32-velu/gen/velu-avx2-rr1-lut4-p4-perm-x40.c",
3090 "src/f32-velu/gen/velu-avx2-rr1-lut4-p4-perm-x48.c",
3091 "src/f32-velu/gen/velu-avx2-rr1-lut4-p4-perm-x56.c",
3092 "src/f32-velu/gen/velu-avx2-rr1-lut4-p4-perm-x64.c",
3093 "src/f32-velu/gen/velu-avx2-rr1-lut4-p4-perm-x72.c",
3094 "src/f32-velu/gen/velu-avx2-rr1-lut4-p4-perm-x80.c",
3095 "src/f32-velu/gen/velu-avx2-rr1-lut8-p4-perm-x8.c",
3096 "src/f32-velu/gen/velu-avx2-rr1-lut8-p4-perm-x16.c",
3097 "src/f32-velu/gen/velu-avx2-rr1-lut8-p4-perm-x24.c",
3098 "src/f32-velu/gen/velu-avx2-rr1-lut8-p4-perm-x32.c",
3099 "src/f32-velu/gen/velu-avx2-rr1-lut8-p4-perm-x40.c",
3100 "src/f32-velu/gen/velu-avx2-rr1-lut8-p4-perm-x48.c",
3101 "src/f32-velu/gen/velu-avx2-rr1-lut8-p4-perm-x56.c",
3102 "src/f32-velu/gen/velu-avx2-rr1-lut8-p4-perm-x64.c",
3103 "src/f32-velu/gen/velu-avx2-rr1-lut8-p4-perm-x72.c",
3104 "src/f32-velu/gen/velu-avx2-rr1-lut8-p4-perm-x80.c",
3105 "src/f32-velu/gen/velu-avx2-rr1-lut16-p3-gather-x8.c",
3106 "src/f32-velu/gen/velu-avx2-rr1-lut16-p3-gather-x16.c",
3107 "src/f32-velu/gen/velu-avx2-rr1-lut16-p3-gather-x24.c",
3108 "src/f32-velu/gen/velu-avx2-rr1-lut16-p3-gather-x32.c",
3109 "src/f32-velu/gen/velu-avx2-rr1-lut16-p3-gather-x40.c",
3110 "src/f32-velu/gen/velu-avx2-rr1-lut16-p3-gather-x48.c",
3111 "src/f32-velu/gen/velu-avx2-rr1-lut16-p3-gather-x56.c",
3112 "src/f32-velu/gen/velu-avx2-rr1-lut16-p3-gather-x64.c",
3113 "src/f32-velu/gen/velu-avx2-rr1-lut16-p3-gather-x72.c",
3114 "src/f32-velu/gen/velu-avx2-rr1-lut16-p3-gather-x80.c",
3115 "src/f32-velu/gen/velu-avx2-rr1-p6-x8.c",
3116 "src/f32-velu/gen/velu-avx2-rr1-p6-x16.c",
3117 "src/f32-velu/gen/velu-avx2-rr1-p6-x24.c",
3118 "src/f32-velu/gen/velu-avx2-rr1-p6-x32.c",
3119 "src/f32-velu/gen/velu-avx2-rr1-p6-x40.c",
3120 "src/f32-velu/gen/velu-avx2-rr1-p6-x48.c",
3121 "src/f32-velu/gen/velu-avx2-rr1-p6-x56.c",
3122 "src/f32-velu/gen/velu-avx2-rr1-p6-x64.c",
3123 "src/f32-velu/gen/velu-avx2-rr1-p6-x72.c",
3124 "src/f32-velu/gen/velu-avx2-rr1-p6-x80.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003125 "src/f32-vscaleexpminusmax/gen/avx2-p5-x8.c",
3126 "src/f32-vscaleexpminusmax/gen/avx2-p5-x16.c",
3127 "src/f32-vscaleexpminusmax/gen/avx2-p5-x24.c",
3128 "src/f32-vscaleexpminusmax/gen/avx2-p5-x32.c",
3129 "src/f32-vscaleexpminusmax/gen/avx2-p5-x40.c",
3130 "src/f32-vscaleexpminusmax/gen/avx2-p5-x48.c",
3131 "src/f32-vscaleexpminusmax/gen/avx2-p5-x56.c",
3132 "src/f32-vscaleexpminusmax/gen/avx2-p5-x64.c",
3133 "src/f32-vscaleexpminusmax/gen/avx2-p5-x72.c",
3134 "src/f32-vscaleexpminusmax/gen/avx2-p5-x80.c",
3135 "src/f32-vscaleexpminusmax/gen/avx2-p5-x88.c",
3136 "src/f32-vscaleexpminusmax/gen/avx2-p5-x96.c",
3137 "src/f32-vscaleextexp/gen/avx2-p5-x8.c",
3138 "src/f32-vscaleextexp/gen/avx2-p5-x16.c",
3139 "src/f32-vscaleextexp/gen/avx2-p5-x24.c",
3140 "src/f32-vscaleextexp/gen/avx2-p5-x32.c",
3141 "src/f32-vscaleextexp/gen/avx2-p5-x40.c",
3142 "src/f32-vscaleextexp/gen/avx2-p5-x48.c",
3143 "src/f32-vscaleextexp/gen/avx2-p5-x56.c",
3144 "src/f32-vscaleextexp/gen/avx2-p5-x64.c",
3145 "src/f32-vscaleextexp/gen/avx2-p5-x72.c",
3146 "src/f32-vscaleextexp/gen/avx2-p5-x80.c",
3147 "src/f32-vscaleextexp/gen/avx2-p5-x88.c",
3148 "src/f32-vscaleextexp/gen/avx2-p5-x96.c",
Marat Dukhanb7633f22020-11-20 16:34:56 -08003149 "src/math/exp-avx2-rr2-lut8-p3-perm.c",
3150 "src/math/exp-avx2-rr2-lut8-p4-perm.c",
3151 "src/math/exp-avx2-rr2-p5.c",
Marat Dukhande390d42020-11-29 19:32:18 -08003152 "src/math/expm1minus-avx2-rr1-lut4-p4-perm.c",
3153 "src/math/expm1minus-avx2-rr1-lut8-p4-perm.c",
3154 "src/math/expm1minus-avx2-rr1-lut16-p3-gather.c",
3155 "src/math/expm1minus-avx2-rr1-p6.c",
Frank Barcharde7223ee2020-12-04 19:04:01 -08003156 "src/math/expminus-avx2-rr2-p5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003157 "src/math/extexp-avx2-p5.c",
3158 "src/math/sigmoid-avx2-rr1-lut64-p2-gather-div.c",
3159 "src/math/sigmoid-avx2-rr1-lut64-p2-gather-nr1fma.c",
3160 "src/math/sigmoid-avx2-rr1-lut64-p2-gather-nr2fma.c",
3161 "src/math/sigmoid-avx2-rr1-lut64-p2-gather-nr2fma1adj.c",
3162 "src/math/sigmoid-avx2-rr1-p5-div.c",
3163 "src/math/sigmoid-avx2-rr1-p5-nr1fma.c",
3164 "src/math/sigmoid-avx2-rr1-p5-nr2fma.c",
3165 "src/math/sigmoid-avx2-rr2-lut64-p2-gather-div.c",
3166 "src/math/sigmoid-avx2-rr2-lut64-p2-gather-nr1fma.c",
3167 "src/math/sigmoid-avx2-rr2-lut64-p2-gather-nr2fma.c",
3168 "src/math/sigmoid-avx2-rr2-lut64-p2-gather-nr2fma1adj.c",
3169 "src/math/sigmoid-avx2-rr2-p5-div.c",
3170 "src/math/sigmoid-avx2-rr2-p5-nr1fma.c",
3171 "src/math/sigmoid-avx2-rr2-p5-nr2fma.c",
Marat Dukhand65a1522020-08-04 19:28:18 -07003172 "src/qs8-dwconv/gen/up8x9-minmax-avx2-mul32.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003173 "src/qs8-dwconv/gen/up16x9-minmax-avx2-mul16.c",
Marat Dukhand65a1522020-08-04 19:28:18 -07003174 "src/qs8-dwconv/gen/up16x9-minmax-avx2-mul32.c",
3175 "src/qs8-dwconv/gen/up24x9-minmax-avx2-mul32.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003176 "src/qs8-dwconv/gen/up32x9-minmax-avx2-mul16.c",
Marat Dukhand65a1522020-08-04 19:28:18 -07003177 "src/qs8-dwconv/gen/up32x9-minmax-avx2-mul32.c",
Marat Dukhan12809522020-08-02 22:23:51 -07003178 "src/qs8-gemm/gen/1x8c8-minmax-avx2.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07003179 "src/qs8-gemm/gen/1x8c8-xw-minmax-avx2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003180 "src/qs8-gemm/gen/2x8c8-minmax-avx2.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07003181 "src/qs8-gemm/gen/2x8c8-xw-minmax-avx2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003182 "src/qs8-gemm/gen/3x8c8-minmax-avx2.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07003183 "src/qs8-gemm/gen/3x8c8-xw-minmax-avx2.c",
Marat Dukhan12809522020-08-02 22:23:51 -07003184 "src/qs8-igemm/gen/1x8c8-minmax-avx2.c",
3185 "src/qs8-igemm/gen/2x8c8-minmax-avx2.c",
3186 "src/qs8-igemm/gen/3x8c8-minmax-avx2.c",
Marat Dukhane6dc0b62020-09-08 23:57:14 -07003187 "src/qs8-vadd/gen/minmax-avx2-mul32-ld64-x8.c",
3188 "src/qs8-vadd/gen/minmax-avx2-mul32-ld64-x16.c",
3189 "src/qs8-vadd/gen/minmax-avx2-mul32-ld64-x24.c",
3190 "src/qs8-vadd/gen/minmax-avx2-mul32-ld64-x32.c",
3191 "src/qs8-vaddc/gen/minmax-avx2-mul32-ld64-x8.c",
3192 "src/qs8-vaddc/gen/minmax-avx2-mul32-ld64-x16.c",
3193 "src/qs8-vaddc/gen/minmax-avx2-mul32-ld64-x24.c",
3194 "src/qs8-vaddc/gen/minmax-avx2-mul32-ld64-x32.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003195]
3196
Marat Dukhan08c4a432019-10-03 09:29:21 -07003197AVX512F_UKERNELS = [
Marat Dukhan5c5fa962020-03-10 18:38:33 -07003198 "src/f32-clamp/gen/avx512f-x16.c",
3199 "src/f32-clamp/gen/avx512f-x32.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07003200 "src/f32-dwconv/gen/up16x4-minmax-avx512f-acc2.c",
3201 "src/f32-dwconv/gen/up16x4-minmax-avx512f.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07003202 "src/f32-dwconv/gen/up16x9-minmax-avx512f-acc2.c",
3203 "src/f32-dwconv/gen/up16x9-minmax-avx512f.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07003204 "src/f32-dwconv/gen/up16x25-minmax-avx512f-acc2.c",
3205 "src/f32-dwconv/gen/up16x25-minmax-avx512f.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003206 "src/f32-dwconv/gen/up32x4-minmax-avx512f-acc2.c",
3207 "src/f32-dwconv/gen/up32x4-minmax-avx512f.c",
3208 "src/f32-dwconv/gen/up32x9-minmax-avx512f-acc2.c",
3209 "src/f32-dwconv/gen/up32x9-minmax-avx512f.c",
3210 "src/f32-dwconv/gen/up32x25-minmax-avx512f-acc2.c",
3211 "src/f32-dwconv/gen/up32x25-minmax-avx512f.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07003212 "src/f32-gemm/gen-inc/1x16inc-minmax-avx512f-broadcast.c",
3213 "src/f32-gemm/gen-inc/4x16inc-minmax-avx512f-broadcast.c",
3214 "src/f32-gemm/gen-inc/5x16inc-minmax-avx512f-broadcast.c",
3215 "src/f32-gemm/gen-inc/6x16inc-minmax-avx512f-broadcast.c",
3216 "src/f32-gemm/gen-inc/7x16inc-minmax-avx512f-broadcast.c",
3217 "src/f32-gemm/gen-inc/8x16inc-minmax-avx512f-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003218 "src/f32-gemm/gen/1x16-minmax-avx512f-broadcast.c",
3219 "src/f32-gemm/gen/4x16-minmax-avx512f-broadcast.c",
3220 "src/f32-gemm/gen/5x16-minmax-avx512f-broadcast.c",
3221 "src/f32-gemm/gen/6x16-minmax-avx512f-broadcast.c",
3222 "src/f32-gemm/gen/7x16-minmax-avx512f-broadcast.c",
3223 "src/f32-gemm/gen/8x16-minmax-avx512f-broadcast.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07003224 "src/f32-hswish/gen/hswish-avx512f-x16.c",
3225 "src/f32-hswish/gen/hswish-avx512f-x32.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07003226 "src/f32-igemm/gen/1x16-minmax-avx512f-broadcast.c",
3227 "src/f32-igemm/gen/4x16-minmax-avx512f-broadcast.c",
3228 "src/f32-igemm/gen/5x16-minmax-avx512f-broadcast.c",
3229 "src/f32-igemm/gen/6x16-minmax-avx512f-broadcast.c",
3230 "src/f32-igemm/gen/7x16-minmax-avx512f-broadcast.c",
3231 "src/f32-igemm/gen/8x16-minmax-avx512f-broadcast.c",
Marat Dukhan90eca0a2020-03-11 00:52:23 -07003232 "src/f32-prelu/gen/avx512f-2x16.c",
3233 "src/f32-prelu/gen/avx512f-2x32.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003234 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128-acc2.c",
3235 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003236 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003237 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x144-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003238 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x144.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003239 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160-acc2.c",
3240 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003241 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003242 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc2.c",
3243 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc3.c",
3244 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc6.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003245 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003246 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128-acc2.c",
3247 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003248 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003249 "src/f32-raddextexp/gen/avx512f-p5-scalef-x144-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003250 "src/f32-raddextexp/gen/avx512f-p5-scalef-x144.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003251 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160-acc2.c",
3252 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003253 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003254 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc2.c",
3255 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc3.c",
3256 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc6.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003257 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003258 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128-acc2.c",
3259 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003260 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003261 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x144-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003262 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x144.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003263 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160-acc2.c",
3264 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003265 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003266 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc2.c",
3267 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc3.c",
3268 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc6.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003269 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07003270 "src/f32-relu/gen/avx512f-x16.c",
3271 "src/f32-relu/gen/avx512f-x32.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003272 "src/f32-rmax/avx512f.c",
Marat Dukhand9ca7e62020-09-23 23:45:29 -07003273 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x16.c",
3274 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x32.c",
3275 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x48.c",
3276 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x64.c",
3277 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x80.c",
3278 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x96.c",
3279 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x112.c",
3280 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x128.c",
3281 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x16.c",
3282 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x32.c",
3283 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x48.c",
3284 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x64.c",
3285 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x80.c",
3286 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x96.c",
3287 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x112.c",
3288 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003289 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x16.c",
3290 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x32.c",
3291 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x48.c",
3292 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x64.c",
3293 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x80.c",
3294 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x96.c",
3295 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x112.c",
3296 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x128.c",
3297 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x16.c",
3298 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x32.c",
3299 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x48.c",
3300 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x64.c",
3301 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x80.c",
3302 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x96.c",
3303 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x112.c",
3304 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x128.c",
Marat Dukhand9ca7e62020-09-23 23:45:29 -07003305 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x16.c",
3306 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x32.c",
3307 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x48.c",
3308 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x64.c",
3309 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x80.c",
3310 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x96.c",
3311 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x112.c",
3312 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x128.c",
3313 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x16.c",
3314 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x32.c",
3315 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x48.c",
3316 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x64.c",
3317 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x80.c",
3318 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x96.c",
3319 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x112.c",
3320 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x128.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07003321 "src/f32-vbinary/gen/vadd-minmax-avx512f-x16.c",
3322 "src/f32-vbinary/gen/vadd-minmax-avx512f-x32.c",
3323 "src/f32-vbinary/gen/vaddc-minmax-avx512f-x16.c",
3324 "src/f32-vbinary/gen/vaddc-minmax-avx512f-x32.c",
3325 "src/f32-vbinary/gen/vdiv-minmax-avx512f-x16.c",
3326 "src/f32-vbinary/gen/vdiv-minmax-avx512f-x32.c",
3327 "src/f32-vbinary/gen/vdivc-minmax-avx512f-x16.c",
3328 "src/f32-vbinary/gen/vdivc-minmax-avx512f-x32.c",
Marat Dukhan9a88efe2019-12-10 15:54:24 -08003329 "src/f32-vbinary/gen/vmax-avx512f-x16.c",
3330 "src/f32-vbinary/gen/vmax-avx512f-x32.c",
3331 "src/f32-vbinary/gen/vmaxc-avx512f-x16.c",
3332 "src/f32-vbinary/gen/vmaxc-avx512f-x32.c",
3333 "src/f32-vbinary/gen/vmin-avx512f-x16.c",
3334 "src/f32-vbinary/gen/vmin-avx512f-x32.c",
3335 "src/f32-vbinary/gen/vminc-avx512f-x16.c",
3336 "src/f32-vbinary/gen/vminc-avx512f-x32.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07003337 "src/f32-vbinary/gen/vmul-minmax-avx512f-x16.c",
3338 "src/f32-vbinary/gen/vmul-minmax-avx512f-x32.c",
3339 "src/f32-vbinary/gen/vmulc-minmax-avx512f-x16.c",
3340 "src/f32-vbinary/gen/vmulc-minmax-avx512f-x32.c",
3341 "src/f32-vbinary/gen/vrdivc-minmax-avx512f-x16.c",
3342 "src/f32-vbinary/gen/vrdivc-minmax-avx512f-x32.c",
3343 "src/f32-vbinary/gen/vrsubc-minmax-avx512f-x16.c",
3344 "src/f32-vbinary/gen/vrsubc-minmax-avx512f-x32.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07003345 "src/f32-vbinary/gen/vsqrdiff-avx512f-x16.c",
3346 "src/f32-vbinary/gen/vsqrdiff-avx512f-x32.c",
3347 "src/f32-vbinary/gen/vsqrdiffc-avx512f-x16.c",
3348 "src/f32-vbinary/gen/vsqrdiffc-avx512f-x32.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07003349 "src/f32-vbinary/gen/vsub-minmax-avx512f-x16.c",
3350 "src/f32-vbinary/gen/vsub-minmax-avx512f-x32.c",
3351 "src/f32-vbinary/gen/vsubc-minmax-avx512f-x16.c",
3352 "src/f32-vbinary/gen/vsubc-minmax-avx512f-x32.c",
Marat Dukhaned6baaf2020-12-01 15:07:08 -08003353 "src/f32-velu/gen/velu-avx512f-rr1-lut16-p3-perm-x16.c",
3354 "src/f32-velu/gen/velu-avx512f-rr1-lut16-p3-perm-x32.c",
3355 "src/f32-velu/gen/velu-avx512f-rr1-lut16-p3-perm-x48.c",
3356 "src/f32-velu/gen/velu-avx512f-rr1-lut16-p3-perm-x64.c",
3357 "src/f32-velu/gen/velu-avx512f-rr1-lut16-p3-perm-x80.c",
3358 "src/f32-velu/gen/velu-avx512f-rr1-lut16-p3-perm-x96.c",
3359 "src/f32-velu/gen/velu-avx512f-rr1-lut16-p3-perm-x112.c",
3360 "src/f32-velu/gen/velu-avx512f-rr1-lut16-p3-perm-x128.c",
3361 "src/f32-velu/gen/velu-avx512f-rr1-p6-x16.c",
3362 "src/f32-velu/gen/velu-avx512f-rr1-p6-x32.c",
3363 "src/f32-velu/gen/velu-avx512f-rr1-p6-x48.c",
3364 "src/f32-velu/gen/velu-avx512f-rr1-p6-x64.c",
3365 "src/f32-velu/gen/velu-avx512f-rr1-p6-x80.c",
3366 "src/f32-velu/gen/velu-avx512f-rr1-p6-x96.c",
3367 "src/f32-velu/gen/velu-avx512f-rr1-p6-x112.c",
3368 "src/f32-velu/gen/velu-avx512f-rr1-p6-x128.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07003369 "src/f32-vlrelu/gen/vlrelu-avx512f-x16.c",
3370 "src/f32-vlrelu/gen/vlrelu-avx512f-x32.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003371 "src/f32-vrnd/gen/vrndd-avx512f-x16.c",
3372 "src/f32-vrnd/gen/vrndd-avx512f-x32.c",
3373 "src/f32-vrnd/gen/vrndne-avx512f-x16.c",
3374 "src/f32-vrnd/gen/vrndne-avx512f-x32.c",
3375 "src/f32-vrnd/gen/vrndu-avx512f-x16.c",
3376 "src/f32-vrnd/gen/vrndu-avx512f-x32.c",
3377 "src/f32-vrnd/gen/vrndz-avx512f-x16.c",
3378 "src/f32-vrnd/gen/vrndz-avx512f-x32.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07003379 "src/f32-vscale/avx512f-x64.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003380 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x16.c",
3381 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x32.c",
3382 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x48.c",
3383 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x64.c",
3384 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x80.c",
3385 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x96.c",
3386 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x112.c",
3387 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x128.c",
3388 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x144.c",
3389 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x160.c",
3390 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x176.c",
3391 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x192.c",
3392 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x16.c",
3393 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x32.c",
3394 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x48.c",
3395 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x64.c",
3396 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x80.c",
3397 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x96.c",
3398 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x112.c",
3399 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x128.c",
3400 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x144.c",
3401 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x160.c",
3402 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x176.c",
3403 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x192.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07003404 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x16.c",
3405 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x32.c",
3406 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x48.c",
3407 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x64.c",
3408 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x80.c",
3409 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x96.c",
3410 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x112.c",
3411 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x128.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07003412 "src/f32-vunary/gen/vabs-avx512f-x16.c",
3413 "src/f32-vunary/gen/vabs-avx512f-x32.c",
3414 "src/f32-vunary/gen/vneg-avx512f-x16.c",
3415 "src/f32-vunary/gen/vneg-avx512f-x32.c",
3416 "src/f32-vunary/gen/vsqr-avx512f-x16.c",
3417 "src/f32-vunary/gen/vsqr-avx512f-x32.c",
Marat Dukhanb7633f22020-11-20 16:34:56 -08003418 "src/math/exp-avx512f-rr2-lut16-p3-perm-scalef.c",
3419 "src/math/exp-avx512f-rr2-lut16-p3-perm.c",
3420 "src/math/exp-avx512f-rr2-lut32-p2-perm2-scalef.c",
3421 "src/math/exp-avx512f-rr2-lut32-p2-perm2.c",
3422 "src/math/exp-avx512f-rr2-p5-scalef.c",
3423 "src/math/exp-avx512f-rr2-p5.c",
Marat Dukhande390d42020-11-29 19:32:18 -08003424 "src/math/expm1minus-avx512f-rr1-lut16-p3-perm.c",
3425 "src/math/expm1minus-avx512f-rr1-p6.c",
Marat Dukhan98ba4412019-10-23 02:14:28 -07003426 "src/math/extexp-avx512f-p5.c",
Marat Dukhan6a34c5f2020-09-22 21:44:15 -07003427 "src/math/sigmoid-avx512f-rr1-lut16-p3-perm-scalef-div.c",
Marat Dukhan6a34c5f2020-09-22 21:44:15 -07003428 "src/math/sigmoid-avx512f-rr1-lut16-p3-perm-scalef-nr1fma.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003429 "src/math/sigmoid-avx512f-rr1-lut16-p3-perm-scalef-nr1fma1adj.c",
Marat Dukhan6a34c5f2020-09-22 21:44:15 -07003430 "src/math/sigmoid-avx512f-rr1-lut32-p2-perm2-scalef-div.c",
Marat Dukhan6a34c5f2020-09-22 21:44:15 -07003431 "src/math/sigmoid-avx512f-rr1-lut32-p2-perm2-scalef-nr1fma.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003432 "src/math/sigmoid-avx512f-rr1-lut32-p2-perm2-scalef-nr1fma1adj.c",
Marat Dukhan36173d22020-10-15 17:14:26 -07003433 "src/math/sigmoid-avx512f-rr1-lut64-p2-gather-scalef-div.c",
Marat Dukhan36173d22020-10-15 17:14:26 -07003434 "src/math/sigmoid-avx512f-rr1-lut64-p2-gather-scalef-nr1fma.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003435 "src/math/sigmoid-avx512f-rr1-lut64-p2-gather-scalef-nr1fma1adj.c",
3436 "src/math/sigmoid-avx512f-rr1-p5-scalef-div.c",
3437 "src/math/sigmoid-avx512f-rr1-p5-scalef-nr1fma.c",
3438 "src/math/sigmoid-avx512f-rr1-p5-scalef-nr1fma1adj.c",
3439 "src/math/sigmoid-avx512f-rr2-lut16-p3-perm-scalef-div.c",
3440 "src/math/sigmoid-avx512f-rr2-lut16-p3-perm-scalef-nr1fma.c",
3441 "src/math/sigmoid-avx512f-rr2-lut16-p3-perm-scalef-nr1fma1adj.c",
3442 "src/math/sigmoid-avx512f-rr2-lut32-p2-perm2-scalef-div.c",
3443 "src/math/sigmoid-avx512f-rr2-lut32-p2-perm2-scalef-nr1fma.c",
3444 "src/math/sigmoid-avx512f-rr2-lut32-p2-perm2-scalef-nr1fma1adj.c",
Marat Dukhan36173d22020-10-15 17:14:26 -07003445 "src/math/sigmoid-avx512f-rr2-lut64-p2-gather-scalef-div.c",
Marat Dukhan36173d22020-10-15 17:14:26 -07003446 "src/math/sigmoid-avx512f-rr2-lut64-p2-gather-scalef-nr1fma.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003447 "src/math/sigmoid-avx512f-rr2-lut64-p2-gather-scalef-nr1fma1adj.c",
3448 "src/math/sigmoid-avx512f-rr2-p5-scalef-div.c",
3449 "src/math/sigmoid-avx512f-rr2-p5-scalef-nr1fma.c",
3450 "src/math/sigmoid-avx512f-rr2-p5-scalef-nr1fma1adj.c",
Marat Dukhan84000762020-06-29 18:38:43 -07003451 "src/math/sqrt-avx512f-nr1fma.c",
Marat Dukhan84000762020-06-29 18:38:43 -07003452 "src/math/sqrt-avx512f-nr1fma1adj.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003453 "src/math/sqrt-avx512f-nr2fma.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003454]
3455
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07003456AVX512SKX_UKERNELS = [
Marat Dukhan2ffc5e62020-09-06 22:33:38 -07003457 "src/qs8-dwconv/gen/up16x9-minmax-avx512skx-mul32.c",
3458 "src/qs8-dwconv/gen/up32x9-minmax-avx512skx-mul32.c",
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07003459 "src/qs8-gemm/gen/1x16c8-minmax-avx512skx.c",
3460 "src/qs8-gemm/gen/2x16c8-minmax-avx512skx.c",
3461 "src/qs8-gemm/gen/3x16c8-minmax-avx512skx.c",
3462 "src/qs8-gemm/gen/4x16c8-minmax-avx512skx.c",
3463 "src/qs8-igemm/gen/1x16c8-minmax-avx512skx.c",
3464 "src/qs8-igemm/gen/2x16c8-minmax-avx512skx.c",
3465 "src/qs8-igemm/gen/3x16c8-minmax-avx512skx.c",
3466 "src/qs8-igemm/gen/4x16c8-minmax-avx512skx.c",
3467]
3468
Frank Barchardbcedc082020-08-17 18:00:51 -07003469WASM32_ASM_UKERNELS = [
3470 "src/f32-relu/wasm_shr_x1.S",
3471 "src/f32-relu/wasm_shr_x2.S",
3472 "src/f32-relu/wasm_shr_x4.S",
3473]
3474
Marat Dukhan08c4a432019-10-03 09:29:21 -07003475AARCH32_ASM_UKERNELS = [
Marat Dukhan32f93812020-05-17 20:31:21 -07003476 "src/f32-gemm/4x4-aarch32-vfp-ld64.S",
Marat Dukhan3b98f6b2020-05-17 10:09:22 -07003477 "src/f32-gemm/4x4-minmax-aarch32-vfp-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003478 "src/f32-gemm/4x8-minmax-aarch32-neon-cortex-a53.S",
3479 "src/f32-gemm/4x8-minmax-aarch32-neon-cortex-a55.S",
Frank Barchard490febe2020-07-16 18:42:17 -07003480 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-cortex-a7.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003481 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-cortex-a75.S",
Frank Barchard569561d2020-06-17 13:11:12 -07003482 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-ld64.S",
Frank Barchard490febe2020-07-16 18:42:17 -07003483 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-pld-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003484 "src/f32-igemm/4x8-minmax-aarch32-neon-cortex-a53.S",
3485 "src/f32-igemm/4x8-minmax-aarch32-neon-cortex-a55.S",
Frank Barchard490febe2020-07-16 18:42:17 -07003486 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-cortex-a7.S",
3487 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-cortex-a75.S",
3488 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-ld64.S",
3489 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-pld-cortex-a75.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003490]
3491
3492AARCH64_ASM_UKERNELS = [
Frank Barchardbddfbcd2020-04-15 12:32:41 -07003493 "src/f16-gemm/gen-inc/1x8inc-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchard04336c12020-10-22 16:48:55 -07003494 "src/f16-gemm/gen-inc/1x16inc-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchardbddfbcd2020-04-15 12:32:41 -07003495 "src/f16-gemm/gen-inc/4x8inc-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchard04336c12020-10-22 16:48:55 -07003496 "src/f16-gemm/gen-inc/4x16inc-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchardbddfbcd2020-04-15 12:32:41 -07003497 "src/f16-gemm/gen-inc/6x8inc-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchard04336c12020-10-22 16:48:55 -07003498 "src/f16-gemm/gen-inc/6x16inc-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchard3b8e5662020-04-20 12:12:53 -07003499 "src/f16-gemm/gen-inc/8x8inc-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchard04336c12020-10-22 16:48:55 -07003500 "src/f16-gemm/gen/1x8-minmax-aarch64-neonfp16arith-ld64.S",
3501 "src/f16-gemm/gen/1x16-minmax-aarch64-neonfp16arith-ld32.S",
3502 "src/f16-gemm/gen/4x8-minmax-aarch64-neonfp16arith-ld64.S",
3503 "src/f16-gemm/gen/4x16-minmax-aarch64-neonfp16arith-ld32.S",
3504 "src/f16-gemm/gen/6x8-minmax-aarch64-neonfp16arith-ld64.S",
3505 "src/f16-gemm/gen/6x16-minmax-aarch64-neonfp16arith-ld32.S",
3506 "src/f16-gemm/gen/8x8-minmax-aarch64-neonfp16arith-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003507 "src/f32-dwconv/up4x9-minmax-aarch64-neonfma-cortex-a55.S",
3508 "src/f32-dwconv/up4x9-minmax-aarch64-neonfma.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003509 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a53.S",
3510 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a57.S",
3511 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a75.S",
Frank Barchard04336c12020-10-22 16:48:55 -07003512 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-ld64.S",
3513 "src/f32-gemm/gen-inc/1x12inc-minmax-aarch64-neonfma-cortex-a53.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003514 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a53.S",
3515 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a55.S",
3516 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a57.S",
3517 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003518 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-ld64.S",
Frank Barchard04336c12020-10-22 16:48:55 -07003519 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-ld128.S",
3520 "src/f32-gemm/gen-inc/4x12inc-minmax-aarch64-neonfma-cortex-a53.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003521 "src/f32-gemm/gen-inc/5x8inc-minmax-aarch64-neonfma-cortex-a57.S",
3522 "src/f32-gemm/gen-inc/5x8inc-minmax-aarch64-neonfma-cortex-a75.S",
3523 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a53.S",
3524 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a55.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003525 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a57.S",
Frank Barchard04336c12020-10-22 16:48:55 -07003526 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a73.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003527 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003528 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-ld64.S",
Frank Barchard04336c12020-10-22 16:48:55 -07003529 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-ld128.S",
3530 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a53.S",
3531 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a57.S",
3532 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a75.S",
3533 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-ld64.S",
3534 "src/f32-gemm/gen/1x12-minmax-aarch64-neonfma-cortex-a53.S",
3535 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a53.S",
3536 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a55.S",
3537 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a57.S",
3538 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a75.S",
3539 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-ld64.S",
3540 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-ld128.S",
3541 "src/f32-gemm/gen/4x12-minmax-aarch64-neonfma-cortex-a53.S",
3542 "src/f32-gemm/gen/5x8-minmax-aarch64-neonfma-cortex-a57.S",
3543 "src/f32-gemm/gen/5x8-minmax-aarch64-neonfma-cortex-a75.S",
3544 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a53.S",
3545 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a55.S",
3546 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a57.S",
3547 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a73.S",
3548 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a75.S",
3549 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-ld64.S",
3550 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-ld128.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003551 "src/f32-igemm/1x8-minmax-aarch64-neonfma-cortex-a53.S",
Frank Barchard04336c12020-10-22 16:48:55 -07003552 "src/f32-igemm/1x12-minmax-aarch64-neonfma-cortex-a53.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003553 "src/f32-igemm/4x8-minmax-aarch64-neonfma-cortex-a53.S",
3554 "src/f32-igemm/4x8-minmax-aarch64-neonfma-cortex-a55.S",
Frank Barchard04336c12020-10-22 16:48:55 -07003555 "src/f32-igemm/4x12-minmax-aarch64-neonfma-cortex-a53.S",
3556 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a53.S",
3557 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a55.S",
3558 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a73.S",
3559 "src/f32-igemm/gen/1x8-minmax-aarch64-neonfma-cortex-a57.S",
3560 "src/f32-igemm/gen/1x8-minmax-aarch64-neonfma-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003561 "src/f32-igemm/gen/4x8-minmax-aarch64-neonfma-cortex-a57.S",
3562 "src/f32-igemm/gen/4x8-minmax-aarch64-neonfma-cortex-a75.S",
3563 "src/f32-igemm/gen/5x8-minmax-aarch64-neonfma-cortex-a57.S",
3564 "src/f32-igemm/gen/5x8-minmax-aarch64-neonfma-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003565 "src/f32-igemm/gen/6x8-minmax-aarch64-neonfma-cortex-a57.S",
3566 "src/f32-igemm/gen/6x8-minmax-aarch64-neonfma-cortex-a75.S",
Frank Barchard4a4be4e2021-03-01 15:00:28 -08003567 "src/qs8-gemm/1x16c4-aarch64-neondot-ld32.S",
Frank Barchardf1fd89e2020-10-06 23:27:11 -07003568 "src/qs8-gemm/1x16c4-aarch64-neondot-ld64.S",
Frank Barchardf2742c42020-10-12 21:16:46 -07003569 "src/qs8-gemm/4x16c4-aarch64-neondot-cortex-a55.S",
Frank Barchardd76a37b2021-03-02 11:36:16 -08003570 "src/qs8-gemm/4x16c4-aarch64-neondot-ld64.S",
Frank Barchardda78da12021-03-02 14:28:00 -08003571 "src/qs8-gemm/4x16c4-aarch64-neondot-ld32.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003572]
3573
Marat Dukhan1b354632020-03-23 12:50:22 -07003574INTERNAL_MICROKERNEL_HDRS = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003575 "src/xnnpack/argmaxpool.h",
3576 "src/xnnpack/avgpool.h",
3577 "src/xnnpack/clamp.h",
3578 "src/xnnpack/common.h",
3579 "src/xnnpack/conv.h",
Yury Kartynnike7841862020-11-04 18:22:18 -08003580 "src/xnnpack/depthtospace.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003581 "src/xnnpack/dwconv.h",
Frank Barchard04336c12020-10-22 16:48:55 -07003582 "src/xnnpack/fill.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003583 "src/xnnpack/gavgpool.h",
3584 "src/xnnpack/gemm.h",
3585 "src/xnnpack/hswish.h",
Marat Dukhan660fd192020-03-10 04:55:30 -07003586 "src/xnnpack/ibilinear.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003587 "src/xnnpack/igemm.h",
Marat Dukhancfb31342019-12-05 10:42:57 -08003588 "src/xnnpack/intrinsics-polyfill.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003589 "src/xnnpack/lut.h",
3590 "src/xnnpack/math.h",
3591 "src/xnnpack/maxpool.h",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003592 "src/xnnpack/memory.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003593 "src/xnnpack/packx.h",
3594 "src/xnnpack/pad.h",
3595 "src/xnnpack/params.h",
3596 "src/xnnpack/pavgpool.h",
3597 "src/xnnpack/ppmm.h",
3598 "src/xnnpack/prelu.h",
Marat Dukhan97579532019-10-18 16:40:39 -07003599 "src/xnnpack/raddexpminusmax.h",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07003600 "src/xnnpack/raddextexp.h",
Marat Dukhan97579532019-10-18 16:40:39 -07003601 "src/xnnpack/raddstoreexpminusmax.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003602 "src/xnnpack/rmax.h",
3603 "src/xnnpack/scalar-utils.h",
3604 "src/xnnpack/spmm.h",
3605 "src/xnnpack/unpool.h",
3606 "src/xnnpack/vadd.h",
Marat Dukhan1e782c42019-11-21 17:02:40 -08003607 "src/xnnpack/vbinary.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003608 "src/xnnpack/vmulcaddc.h",
Marat Dukhan05ac8e32019-10-21 15:39:33 -07003609 "src/xnnpack/vscale.h",
Marat Dukhan97579532019-10-18 16:40:39 -07003610 "src/xnnpack/vscaleexpminusmax.h",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07003611 "src/xnnpack/vscaleextexp.h",
Marat Dukhan1e782c42019-11-21 17:02:40 -08003612 "src/xnnpack/vunary.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003613 "src/xnnpack/zip.h",
Marat Dukhan1b354632020-03-23 12:50:22 -07003614]
3615
3616INTERNAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003617 "include/xnnpack.h",
3618 "src/xnnpack/allocator.h",
3619 "src/xnnpack/compute.h",
3620 "src/xnnpack/im2col.h",
3621 "src/xnnpack/indirection.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003622 "src/xnnpack/math-stubs.h",
Chao Mei6ddfc602020-05-13 22:29:36 -07003623 "src/xnnpack/memory-planner.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003624 "src/xnnpack/operator.h",
3625 "src/xnnpack/pack.h",
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07003626 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003627 "src/xnnpack/requantization-stubs.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003628 "src/xnnpack/requantization.h",
Marat Dukhan1d75a542020-02-03 12:23:01 -08003629 "src/xnnpack/subgraph.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003630]
3631
Marat Dukhan1b354632020-03-23 12:50:22 -07003632ACCURACY_EVAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003633 "src/xnnpack/math-stubs.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003634]
3635
Marat Dukhan1b354632020-03-23 12:50:22 -07003636MICROKERNEL_BENCHMARK_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003637 "include/xnnpack.h",
Frank Barchard35db7d02020-10-26 13:37:34 -07003638 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003639]
3640
Marat Dukhan1b354632020-03-23 12:50:22 -07003641MICROKERNEL_TEST_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Frank Barchard35db7d02020-10-26 13:37:34 -07003642 "include/xnnpack.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003643 "src/xnnpack/isa-checks.h",
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07003644 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003645 "src/xnnpack/requantization.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003646]
3647
3648OPERATOR_TEST_PARAMS_HDRS = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003649 "src/xnnpack/common.h",
Frank Barchard04336c12020-10-22 16:48:55 -07003650 "src/xnnpack/params.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003651]
3652
3653WEIGHTS_PACK_HDRS = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003654 "src/xnnpack/compute.h",
Frank Barchard04336c12020-10-22 16:48:55 -07003655 "src/xnnpack/operator.h",
3656 "src/xnnpack/pack.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003657]
3658
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07003659LOGGING_COPTS = select({
3660 # No logging in optimized mode
3661 ":optimized_build": ["-DXNN_LOG_LEVEL=0"],
3662 # Full logging in debug mode
3663 ":debug_build": ["-DXNN_LOG_LEVEL=5"],
3664 # Error-only logging in default (fastbuild) mode
3665 "//conditions:default": ["-DXNN_LOG_LEVEL=2"],
3666})
3667
Marat Dukhan3b59de22020-06-03 20:15:19 -07003668LOGGING_SRCS = select({
3669 # No logging in optimized mode
3670 ":optimized_build": [],
3671 "//conditions:default": [
3672 "src/operator-strings.c",
3673 "src/subgraph-strings.c",
3674 ],
3675})
3676
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07003677LOGGING_HDRS = [
3678 "src/xnnpack/log.h",
3679]
3680
Marat Dukhan08c4a432019-10-03 09:29:21 -07003681xnnpack_cc_library(
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003682 name = "tables",
3683 srcs = TABLE_SRCS,
3684 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003685 gcc_copts = xnnpack_gcc_std_copts(),
3686 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003687)
3688
3689xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003690 name = "scalar_ukernels",
3691 srcs = SCALAR_UKERNELS,
3692 hdrs = INTERNAL_HDRS,
3693 aarch32_copts = ["-marm"],
Marat Dukhan10a38082020-04-17 03:58:35 -07003694 gcc_copts = xnnpack_gcc_std_copts(),
3695 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003696 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003697 ":tables",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003698 "@FP16",
3699 "@FXdiv",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003700 "@pthreadpool",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003701 ],
3702)
3703
3704xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003705 name = "scalar_ukernels_test_mode",
3706 srcs = SCALAR_UKERNELS,
3707 hdrs = INTERNAL_HDRS,
3708 aarch32_copts = ["-marm"],
3709 copts = [
3710 "-UNDEBUG",
3711 "-DXNN_TEST_MODE=1",
3712 ],
3713 gcc_copts = xnnpack_gcc_std_copts(),
3714 msvc_copts = xnnpack_msvc_std_copts(),
3715 deps = [
3716 ":tables",
3717 "@FP16",
3718 "@FXdiv",
3719 "@pthreadpool",
3720 ],
3721)
3722
3723xnnpack_cc_library(
Marat Dukhan436ebe62019-12-04 15:10:12 -08003724 name = "wasm_ukernels",
3725 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003726 gcc_copts = xnnpack_gcc_std_copts(),
3727 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan436ebe62019-12-04 15:10:12 -08003728 wasm_srcs = WASM_UKERNELS,
Marat Dukhan1c6e3892020-06-25 23:56:10 -07003729 wasmsimd_srcs = WASM_UKERNELS + WASMSIMD_UKERNELS,
Marat Dukhan436ebe62019-12-04 15:10:12 -08003730 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003731 ":tables",
Marat Dukhan436ebe62019-12-04 15:10:12 -08003732 "@FP16",
3733 "@FXdiv",
3734 "@pthreadpool",
3735 ],
3736)
3737
3738xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003739 name = "wasm_ukernels_test_mode",
3740 hdrs = INTERNAL_HDRS,
3741 copts = [
3742 "-UNDEBUG",
3743 "-DXNN_TEST_MODE=1",
3744 ],
3745 gcc_copts = xnnpack_gcc_std_copts(),
3746 msvc_copts = xnnpack_msvc_std_copts(),
3747 wasm_srcs = WASM_UKERNELS,
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07003748 wasmsimd_srcs = WASM_UKERNELS + WASMSIMD_UKERNELS,
Marat Dukhan33fcf782020-05-24 14:27:15 -07003749 deps = [
3750 ":tables",
3751 "@FP16",
3752 "@FXdiv",
3753 "@pthreadpool",
3754 ],
3755)
3756
3757xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003758 name = "neon_ukernels",
3759 hdrs = INTERNAL_HDRS,
3760 aarch32_copts = [
3761 "-marm",
Marat Dukhan8853b822020-05-07 12:19:01 -07003762 "-march=armv7-a",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003763 "-mfpu=neon",
3764 ],
3765 aarch32_srcs = NEON_UKERNELS,
3766 aarch64_srcs = NEON_UKERNELS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003767 gcc_copts = xnnpack_gcc_std_copts(),
3768 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08003769 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003770 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003771 "@FP16",
3772 "@pthreadpool",
3773 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003774)
3775
3776xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003777 name = "neon_ukernels_test_mode",
3778 hdrs = INTERNAL_HDRS,
3779 aarch32_copts = [
3780 "-marm",
3781 "-march=armv7-a",
3782 "-mfpu=neon",
3783 ],
3784 aarch32_srcs = NEON_UKERNELS,
3785 aarch64_srcs = NEON_UKERNELS,
3786 copts = [
3787 "-UNDEBUG",
3788 "-DXNN_TEST_MODE=1",
3789 ],
3790 gcc_copts = xnnpack_gcc_std_copts(),
3791 msvc_copts = xnnpack_msvc_std_copts(),
3792 deps = [
3793 ":tables",
3794 "@FP16",
3795 "@pthreadpool",
3796 ],
3797)
3798
3799xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003800 name = "neonfma_ukernels",
3801 hdrs = INTERNAL_HDRS,
3802 aarch32_copts = [
3803 "-marm",
Marat Dukhan8853b822020-05-07 12:19:01 -07003804 "-march=armv7-a",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003805 "-mfpu=neon-vfpv4",
3806 ],
3807 aarch32_srcs = NEONFMA_UKERNELS,
3808 aarch64_srcs = NEONFMA_UKERNELS + AARCH64_NEONFMA_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07003809 apple_aarch32_copts = [
3810 "-mcpu=swift",
3811 "-mtune=generic",
3812 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07003813 gcc_copts = xnnpack_gcc_std_copts(),
3814 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08003815 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003816 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003817 "@FP16",
3818 "@pthreadpool",
3819 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003820)
3821
3822xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003823 name = "neonfma_ukernels_test_mode",
3824 hdrs = INTERNAL_HDRS,
3825 aarch32_copts = [
3826 "-marm",
3827 "-march=armv7-a",
3828 "-mfpu=neon-vfpv4",
3829 ],
3830 aarch32_srcs = NEONFMA_UKERNELS,
3831 aarch64_srcs = NEONFMA_UKERNELS + AARCH64_NEONFMA_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07003832 apple_aarch32_copts = [
3833 "-mcpu=swift",
3834 "-mtune=generic",
3835 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07003836 copts = [
3837 "-UNDEBUG",
3838 "-DXNN_TEST_MODE=1",
3839 ],
3840 gcc_copts = xnnpack_gcc_std_copts(),
3841 msvc_copts = xnnpack_msvc_std_copts(),
3842 deps = [
3843 ":tables",
3844 "@FP16",
3845 "@pthreadpool",
3846 ],
3847)
3848
3849xnnpack_cc_library(
Marat Dukhan8853b822020-05-07 12:19:01 -07003850 name = "neonv8_ukernels",
3851 hdrs = INTERNAL_HDRS,
3852 aarch32_copts = [
3853 "-marm",
3854 "-march=armv8-a",
3855 "-mfpu=neon-fp-armv8",
3856 ],
3857 aarch32_srcs = NEONV8_UKERNELS,
3858 aarch64_srcs = NEONV8_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07003859 apple_aarch32_copts = [
3860 "-mcpu=cyclone",
3861 "-mtune=generic",
3862 ],
Marat Dukhan8853b822020-05-07 12:19:01 -07003863 gcc_copts = xnnpack_gcc_std_copts(),
3864 msvc_copts = xnnpack_msvc_std_copts(),
3865 deps = [
3866 ":tables",
3867 "@FP16",
3868 "@pthreadpool",
3869 ],
3870)
3871
3872xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003873 name = "neonv8_ukernels_test_mode",
3874 hdrs = INTERNAL_HDRS,
3875 aarch32_copts = [
3876 "-marm",
3877 "-march=armv8-a",
3878 "-mfpu=neon-fp-armv8",
3879 ],
3880 aarch32_srcs = NEONV8_UKERNELS,
3881 aarch64_srcs = NEONV8_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07003882 apple_aarch32_copts = [
3883 "-mcpu=cyclone",
3884 "-mtune=generic",
3885 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07003886 copts = [
3887 "-UNDEBUG",
3888 "-DXNN_TEST_MODE=1",
3889 ],
3890 gcc_copts = xnnpack_gcc_std_copts(),
3891 msvc_copts = xnnpack_msvc_std_copts(),
3892 deps = [
3893 ":tables",
3894 "@FP16",
3895 "@pthreadpool",
3896 ],
3897)
3898
3899xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003900 name = "neonfp16arith_ukernels",
3901 hdrs = INTERNAL_HDRS,
3902 aarch64_copts = ["-march=armv8.2-a+fp16"],
3903 aarch64_srcs = AARCH64_NEONFP16ARITH_UKERNELS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003904 gcc_copts = xnnpack_gcc_std_copts(),
3905 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08003906 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003907 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003908 "@FP16",
3909 "@pthreadpool",
3910 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003911)
3912
3913xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003914 name = "neonfp16arith_ukernels_test_mode",
3915 hdrs = INTERNAL_HDRS,
3916 aarch64_copts = ["-march=armv8.2-a+fp16"],
3917 aarch64_srcs = AARCH64_NEONFP16ARITH_UKERNELS,
3918 copts = [
3919 "-UNDEBUG",
3920 "-DXNN_TEST_MODE=1",
3921 ],
3922 gcc_copts = xnnpack_gcc_std_copts(),
3923 msvc_copts = xnnpack_msvc_std_copts(),
3924 deps = [
3925 ":tables",
3926 "@FP16",
3927 "@pthreadpool",
3928 ],
3929)
3930
3931xnnpack_cc_library(
Benoit Jacoba9644732020-08-13 12:48:55 -07003932 name = "neondot_ukernels",
3933 hdrs = INTERNAL_HDRS,
Marat Dukhan799ac752020-08-13 16:08:03 -07003934 aarch32_copts = [
3935 "-marm",
3936 "-march=armv8.2-a+dotprod",
3937 "-mfpu=neon-fp-armv8",
3938 ],
Benoit Jacoba9644732020-08-13 12:48:55 -07003939 aarch32_srcs = NEONDOT_UKERNELS,
3940 aarch64_copts = ["-march=armv8.2-a+dotprod"],
3941 aarch64_srcs = NEONDOT_UKERNELS,
3942 gcc_copts = xnnpack_gcc_std_copts(),
3943 msvc_copts = xnnpack_msvc_std_copts(),
3944 deps = [
3945 ":tables",
3946 "@FP16",
3947 "@pthreadpool",
3948 ],
3949)
3950
3951xnnpack_cc_library(
3952 name = "neondot_ukernels_test_mode",
3953 hdrs = INTERNAL_HDRS,
Marat Dukhan799ac752020-08-13 16:08:03 -07003954 aarch32_copts = [
3955 "-marm",
3956 "-march=armv8.2-a+dotprod",
3957 "-mfpu=neon-fp-armv8",
3958 ],
Benoit Jacoba9644732020-08-13 12:48:55 -07003959 aarch32_srcs = NEONDOT_UKERNELS,
3960 aarch64_copts = ["-march=armv8.2-a+dotprod"],
3961 aarch64_srcs = NEONDOT_UKERNELS,
3962 copts = [
3963 "-UNDEBUG",
3964 "-DXNN_TEST_MODE=1",
3965 ],
3966 gcc_copts = xnnpack_gcc_std_copts(),
3967 msvc_copts = xnnpack_msvc_std_copts(),
3968 deps = [
3969 ":tables",
3970 "@FP16",
3971 "@pthreadpool",
3972 ],
3973)
3974
3975xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003976 name = "sse2_ukernels",
3977 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003978 gcc_copts = xnnpack_gcc_std_copts(),
3979 gcc_x86_copts = ["-msse2"],
3980 msvc_copts = xnnpack_msvc_std_copts(),
3981 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003982 x86_srcs = SSE_UKERNELS + SSE2_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08003983 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003984 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003985 "@FP16",
3986 "@pthreadpool",
3987 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003988)
3989
3990xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003991 name = "sse2_ukernels_test_mode",
3992 hdrs = INTERNAL_HDRS,
3993 copts = [
3994 "-UNDEBUG",
3995 "-DXNN_TEST_MODE=1",
3996 ],
3997 gcc_copts = xnnpack_gcc_std_copts(),
3998 gcc_x86_copts = ["-msse2"],
3999 msvc_copts = xnnpack_msvc_std_copts(),
4000 msvc_x86_32_copts = ["/arch:SSE2"],
4001 x86_srcs = SSE_UKERNELS + SSE2_UKERNELS,
4002 deps = [
4003 ":tables",
4004 "@FP16",
4005 "@pthreadpool",
4006 ],
4007)
4008
4009xnnpack_cc_library(
Marat Dukhanfe7acb62020-03-09 19:30:05 -07004010 name = "ssse3_ukernels",
4011 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004012 gcc_copts = xnnpack_gcc_std_copts(),
4013 gcc_x86_copts = ["-mssse3"],
4014 msvc_copts = xnnpack_msvc_std_copts(),
4015 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhanfe7acb62020-03-09 19:30:05 -07004016 x86_srcs = SSSE3_UKERNELS,
4017 deps = [
4018 ":tables",
4019 "@FP16",
4020 "@pthreadpool",
4021 ],
4022)
4023
4024xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07004025 name = "ssse3_ukernels_test_mode",
4026 hdrs = INTERNAL_HDRS,
4027 copts = [
4028 "-UNDEBUG",
4029 "-DXNN_TEST_MODE=1",
4030 ],
4031 gcc_copts = xnnpack_gcc_std_copts(),
4032 gcc_x86_copts = ["-mssse3"],
4033 msvc_copts = xnnpack_msvc_std_copts(),
4034 msvc_x86_32_copts = ["/arch:SSE2"],
4035 x86_srcs = SSSE3_UKERNELS,
4036 deps = [
4037 ":tables",
4038 "@FP16",
4039 "@pthreadpool",
4040 ],
4041)
4042
4043xnnpack_cc_library(
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08004044 name = "sse41_ukernels",
4045 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004046 gcc_copts = xnnpack_gcc_std_copts(),
4047 gcc_x86_copts = ["-msse4.1"],
4048 msvc_copts = xnnpack_msvc_std_copts(),
4049 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08004050 x86_srcs = SSE41_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08004051 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08004052 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08004053 "@FP16",
4054 "@pthreadpool",
4055 ],
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08004056)
4057
4058xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07004059 name = "sse41_ukernels_test_mode",
4060 hdrs = INTERNAL_HDRS,
4061 copts = [
4062 "-UNDEBUG",
4063 "-DXNN_TEST_MODE=1",
4064 ],
4065 gcc_copts = xnnpack_gcc_std_copts(),
4066 gcc_x86_copts = ["-msse4.1"],
4067 msvc_copts = xnnpack_msvc_std_copts(),
4068 msvc_x86_32_copts = ["/arch:SSE2"],
4069 x86_srcs = SSE41_UKERNELS,
4070 deps = [
4071 ":tables",
4072 "@FP16",
4073 "@pthreadpool",
4074 ],
4075)
4076
4077xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004078 name = "avx_ukernels",
4079 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004080 gcc_copts = xnnpack_gcc_std_copts(),
4081 gcc_x86_copts = ["-mavx"],
4082 msvc_copts = xnnpack_msvc_std_copts(),
4083 msvc_x86_32_copts = ["/arch:AVX"],
4084 msvc_x86_64_copts = ["/arch:AVX"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004085 x86_srcs = AVX_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08004086 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08004087 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08004088 "@FP16",
4089 "@pthreadpool",
4090 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004091)
4092
4093xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07004094 name = "avx_ukernels_test_mode",
4095 hdrs = INTERNAL_HDRS,
4096 copts = [
4097 "-UNDEBUG",
4098 "-DXNN_TEST_MODE=1",
4099 ],
4100 gcc_copts = xnnpack_gcc_std_copts(),
4101 gcc_x86_copts = ["-mavx"],
4102 msvc_copts = xnnpack_msvc_std_copts(),
4103 msvc_x86_32_copts = ["/arch:AVX"],
4104 msvc_x86_64_copts = ["/arch:AVX"],
4105 x86_srcs = AVX_UKERNELS,
4106 deps = [
4107 ":tables",
4108 "@FP16",
4109 "@pthreadpool",
4110 ],
4111)
4112
4113xnnpack_cc_library(
Marat Dukhan1566fee2020-08-02 21:55:41 -07004114 name = "xop_ukernels",
4115 hdrs = INTERNAL_HDRS,
4116 gcc_copts = xnnpack_gcc_std_copts(),
4117 gcc_x86_copts = ["-mxop"],
4118 msvc_copts = xnnpack_msvc_std_copts(),
4119 msvc_x86_32_copts = ["/arch:AVX"],
4120 msvc_x86_64_copts = ["/arch:AVX"],
4121 x86_srcs = XOP_UKERNELS,
4122 deps = [
4123 ":tables",
4124 "@FP16",
4125 "@pthreadpool",
4126 ],
4127)
4128
4129xnnpack_cc_library(
4130 name = "xop_ukernels_test_mode",
4131 hdrs = INTERNAL_HDRS,
4132 copts = [
4133 "-UNDEBUG",
4134 "-DXNN_TEST_MODE=1",
4135 ],
4136 gcc_copts = xnnpack_gcc_std_copts(),
4137 gcc_x86_copts = ["-mxop"],
4138 msvc_copts = xnnpack_msvc_std_copts(),
4139 msvc_x86_32_copts = ["/arch:AVX"],
4140 msvc_x86_64_copts = ["/arch:AVX"],
4141 x86_srcs = XOP_UKERNELS,
4142 deps = [
4143 ":tables",
4144 "@FP16",
4145 "@pthreadpool",
4146 ],
4147)
4148
4149xnnpack_cc_library(
Marat Dukhanfda12b82019-11-21 12:27:59 -08004150 name = "fma3_ukernels",
4151 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004152 gcc_copts = xnnpack_gcc_std_copts(),
4153 gcc_x86_copts = ["-mfma"],
4154 msvc_copts = xnnpack_msvc_std_copts(),
4155 msvc_x86_32_copts = ["/arch:AVX"],
4156 msvc_x86_64_copts = ["/arch:AVX"],
Marat Dukhanfda12b82019-11-21 12:27:59 -08004157 x86_srcs = FMA3_UKERNELS,
4158 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08004159 ":tables",
Marat Dukhanfda12b82019-11-21 12:27:59 -08004160 "@FP16",
4161 "@pthreadpool",
4162 ],
4163)
4164
4165xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07004166 name = "fma3_ukernels_test_mode",
4167 hdrs = INTERNAL_HDRS,
4168 copts = [
4169 "-UNDEBUG",
4170 "-DXNN_TEST_MODE=1",
4171 ],
4172 gcc_copts = xnnpack_gcc_std_copts(),
4173 gcc_x86_copts = ["-mfma"],
4174 msvc_copts = xnnpack_msvc_std_copts(),
4175 msvc_x86_32_copts = ["/arch:AVX"],
4176 msvc_x86_64_copts = ["/arch:AVX"],
4177 x86_srcs = FMA3_UKERNELS,
4178 deps = [
4179 ":tables",
4180 "@FP16",
4181 "@pthreadpool",
4182 ],
4183)
4184
4185xnnpack_cc_library(
Marat Dukhan6adff4e2019-10-14 18:32:07 -07004186 name = "avx2_ukernels",
4187 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004188 gcc_copts = xnnpack_gcc_std_copts(),
4189 gcc_x86_copts = [
Marat Dukhan6adff4e2019-10-14 18:32:07 -07004190 "-mfma",
4191 "-mavx2",
4192 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07004193 msvc_copts = xnnpack_msvc_std_copts(),
4194 msvc_x86_32_copts = ["/arch:AVX2"],
4195 msvc_x86_64_copts = ["/arch:AVX2"],
Marat Dukhan6adff4e2019-10-14 18:32:07 -07004196 x86_srcs = AVX2_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08004197 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08004198 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08004199 "@FP16",
4200 "@pthreadpool",
4201 ],
Marat Dukhan6adff4e2019-10-14 18:32:07 -07004202)
4203
4204xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07004205 name = "avx2_ukernels_test_mode",
4206 hdrs = INTERNAL_HDRS,
4207 copts = [
4208 "-UNDEBUG",
4209 "-DXNN_TEST_MODE=1",
4210 ],
4211 gcc_copts = xnnpack_gcc_std_copts(),
4212 gcc_x86_copts = [
4213 "-mfma",
4214 "-mavx2",
4215 ],
4216 msvc_copts = xnnpack_msvc_std_copts(),
4217 msvc_x86_32_copts = ["/arch:AVX2"],
4218 msvc_x86_64_copts = ["/arch:AVX2"],
4219 x86_srcs = AVX2_UKERNELS,
4220 deps = [
4221 ":tables",
4222 "@FP16",
4223 "@pthreadpool",
4224 ],
4225)
4226
4227xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004228 name = "avx512f_ukernels",
4229 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004230 gcc_copts = xnnpack_gcc_std_copts(),
4231 gcc_x86_copts = ["-mavx512f"],
4232 mingw_copts = ["-fno-asynchronous-unwind-tables"],
4233 msvc_copts = xnnpack_msvc_std_copts(),
4234 msvc_x86_32_copts = ["/arch:AVX512"],
4235 msvc_x86_64_copts = ["/arch:AVX512"],
4236 msys_copts = ["-fno-asynchronous-unwind-tables"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004237 x86_srcs = AVX512F_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08004238 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08004239 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08004240 "@FP16",
4241 "@pthreadpool",
4242 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004243)
4244
4245xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07004246 name = "avx512f_ukernels_test_mode",
4247 hdrs = INTERNAL_HDRS,
4248 copts = [
4249 "-UNDEBUG",
4250 "-DXNN_TEST_MODE=1",
4251 ],
4252 gcc_copts = xnnpack_gcc_std_copts(),
4253 gcc_x86_copts = ["-mavx512f"],
4254 mingw_copts = ["-fno-asynchronous-unwind-tables"],
4255 msvc_copts = xnnpack_msvc_std_copts(),
4256 msvc_x86_32_copts = ["/arch:AVX512"],
4257 msvc_x86_64_copts = ["/arch:AVX512"],
4258 msys_copts = ["-fno-asynchronous-unwind-tables"],
4259 x86_srcs = AVX512F_UKERNELS,
4260 deps = [
4261 ":tables",
4262 "@FP16",
4263 "@pthreadpool",
4264 ],
4265)
4266
4267xnnpack_cc_library(
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07004268 name = "avx512skx_ukernels",
4269 hdrs = INTERNAL_HDRS,
4270 gcc_copts = xnnpack_gcc_std_copts(),
4271 gcc_x86_copts = [
4272 "-mavx512f",
4273 "-mavx512cd",
4274 "-mavx512bw",
4275 "-mavx512dq",
4276 "-mavx512vl",
4277 ],
4278 mingw_copts = ["-fno-asynchronous-unwind-tables"],
4279 msvc_copts = xnnpack_msvc_std_copts(),
4280 msvc_x86_32_copts = ["/arch:AVX512"],
4281 msvc_x86_64_copts = ["/arch:AVX512"],
4282 msys_copts = ["-fno-asynchronous-unwind-tables"],
4283 x86_srcs = AVX512SKX_UKERNELS,
4284 deps = [
4285 ":tables",
4286 "@FP16",
4287 "@pthreadpool",
4288 ],
4289)
4290
4291xnnpack_cc_library(
4292 name = "avx512skx_ukernels_test_mode",
4293 hdrs = INTERNAL_HDRS,
4294 copts = [
4295 "-UNDEBUG",
4296 "-DXNN_TEST_MODE=1",
4297 ],
4298 gcc_copts = xnnpack_gcc_std_copts(),
4299 gcc_x86_copts = [
4300 "-mavx512f",
4301 "-mavx512cd",
4302 "-mavx512bw",
4303 "-mavx512dq",
4304 "-mavx512vl",
4305 ],
4306 mingw_copts = ["-fno-asynchronous-unwind-tables"],
4307 msvc_copts = xnnpack_msvc_std_copts(),
4308 msvc_x86_32_copts = ["/arch:AVX512"],
4309 msvc_x86_64_copts = ["/arch:AVX512"],
4310 msys_copts = ["-fno-asynchronous-unwind-tables"],
4311 x86_srcs = AVX512SKX_UKERNELS,
4312 deps = [
4313 ":tables",
4314 "@FP16",
4315 "@pthreadpool",
4316 ],
4317)
4318
4319xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004320 name = "asm_ukernels",
4321 hdrs = ["src/xnnpack/assembly.h"],
4322 aarch32_srcs = AARCH32_ASM_UKERNELS,
Frank Barchard31bb45b2020-10-06 00:26:33 -07004323 aarch64_copts = ["-march=armv8.2-a+fp16+dotprod"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004324 aarch64_srcs = AARCH64_ASM_UKERNELS,
Frank Barchardbcedc082020-08-17 18:00:51 -07004325 wasm_srcs = WASM32_ASM_UKERNELS,
4326 wasmsimd_srcs = WASM32_ASM_UKERNELS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07004327)
4328
Marat Dukhan3b59de22020-06-03 20:15:19 -07004329xnnpack_cc_library(
4330 name = "logging_utils",
4331 srcs = LOGGING_SRCS,
4332 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
4333 copts = LOGGING_COPTS + [
4334 "-Isrc",
4335 "-Iinclude",
4336 ] + select({
4337 ":debug_build": [],
4338 "//conditions:default": xnnpack_min_size_copts(),
4339 }),
4340 gcc_copts = xnnpack_gcc_std_copts(),
4341 msvc_copts = xnnpack_msvc_std_copts(),
4342 visibility = xnnpack_visibility(),
4343 deps = [
4344 "@FP16",
4345 "@clog",
4346 "@pthreadpool",
4347 ],
4348)
4349
Marat Dukhan08c4a432019-10-03 09:29:21 -07004350xnnpack_aggregate_library(
4351 name = "ukernels",
4352 aarch32_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07004353 ":neon_ukernels",
4354 ":neonfma_ukernels",
Marat Dukhan8853b822020-05-07 12:19:01 -07004355 ":neonv8_ukernels",
Benoit Jacoba9644732020-08-13 12:48:55 -07004356 ":neondot_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004357 ":asm_ukernels",
4358 ],
4359 aarch64_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07004360 ":neon_ukernels",
4361 ":neonfma_ukernels",
Marat Dukhan8853b822020-05-07 12:19:01 -07004362 ":neonv8_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004363 ":neonfp16arith_ukernels",
Benoit Jacoba9644732020-08-13 12:48:55 -07004364 ":neondot_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004365 ":asm_ukernels",
4366 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07004367 generic_deps = [
4368 ":scalar_ukernels",
4369 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07004370 wasm_deps = [
4371 ":wasm_ukernels",
Frank Barchardbcedc082020-08-17 18:00:51 -07004372 ":asm_ukernels",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004373 ],
4374 wasmsimd_deps = [
4375 ":wasm_ukernels",
Frank Barchardbcedc082020-08-17 18:00:51 -07004376 ":asm_ukernels",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004377 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004378 x86_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07004379 ":sse2_ukernels",
Marat Dukhanfe7acb62020-03-09 19:30:05 -07004380 ":ssse3_ukernels",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08004381 ":sse41_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004382 ":avx_ukernels",
Marat Dukhan1566fee2020-08-02 21:55:41 -07004383 ":xop_ukernels",
Marat Dukhanfda12b82019-11-21 12:27:59 -08004384 ":fma3_ukernels",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07004385 ":avx2_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004386 ":avx512f_ukernels",
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07004387 ":avx512skx_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004388 ],
4389)
4390
Marat Dukhan33fcf782020-05-24 14:27:15 -07004391xnnpack_aggregate_library(
4392 name = "ukernels_test_mode",
4393 aarch32_deps = [
4394 ":neon_ukernels_test_mode",
4395 ":neonfma_ukernels_test_mode",
4396 ":neonv8_ukernels_test_mode",
Benoit Jacoba9644732020-08-13 12:48:55 -07004397 ":neondot_ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004398 ":asm_ukernels",
4399 ],
4400 aarch64_deps = [
4401 ":neon_ukernels_test_mode",
4402 ":neonfma_ukernels_test_mode",
4403 ":neonv8_ukernels_test_mode",
4404 ":neonfp16arith_ukernels_test_mode",
Benoit Jacoba9644732020-08-13 12:48:55 -07004405 ":neondot_ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004406 ":asm_ukernels",
4407 ],
4408 generic_deps = [
4409 ":scalar_ukernels_test_mode",
4410 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07004411 wasm_deps = [
4412 ":wasm_ukernels_test_mode",
Frank Barchardbcedc082020-08-17 18:00:51 -07004413 ":asm_ukernels",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004414 ],
4415 wasmsimd_deps = [
4416 ":wasm_ukernels_test_mode",
Frank Barchardbcedc082020-08-17 18:00:51 -07004417 ":asm_ukernels",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004418 ],
4419 x86_deps = [
4420 ":sse2_ukernels_test_mode",
4421 ":ssse3_ukernels_test_mode",
4422 ":sse41_ukernels_test_mode",
4423 ":avx_ukernels_test_mode",
Marat Dukhan1566fee2020-08-02 21:55:41 -07004424 ":xop_ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004425 ":fma3_ukernels_test_mode",
4426 ":avx2_ukernels_test_mode",
4427 ":avx512f_ukernels_test_mode",
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07004428 ":avx512skx_ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004429 ],
4430)
4431
Marat Dukhan86df3842020-06-17 14:03:31 -07004432# This target must be compatible with Android NDK r17 and compile with Clang 6.
4433# It is similar to :ukernels target, but excludes microkernels which require
4434# NEON FP16 arithmetics, NEON Dot Product, and newer intrinsics.
Marat Dukhanae046f52020-06-15 13:16:14 -07004435xnnpack_aggregate_library(
Marat Dukhan86df3842020-06-17 14:03:31 -07004436 name = "ukernels_max_compatibility",
Marat Dukhanae046f52020-06-15 13:16:14 -07004437 aarch32_deps = [
4438 ":neon_ukernels",
4439 ":neonfma_ukernels",
4440 ":neonv8_ukernels",
4441 ":asm_ukernels",
4442 ],
4443 aarch64_deps = [
4444 ":neon_ukernels",
4445 ":neonfma_ukernels",
4446 ":neonv8_ukernels",
4447 ":asm_ukernels",
4448 ],
4449 generic_deps = [
4450 ":scalar_ukernels",
4451 ],
Marat Dukhanae046f52020-06-15 13:16:14 -07004452 wasm_deps = [
4453 ":wasm_ukernels",
4454 ],
4455 wasmsimd_deps = [
4456 ":wasm_ukernels",
4457 ],
4458 x86_deps = [
4459 ":sse2_ukernels",
4460 ":ssse3_ukernels",
4461 ":sse41_ukernels",
4462 ":avx_ukernels",
Marat Dukhan1566fee2020-08-02 21:55:41 -07004463 ":xop_ukernels",
Marat Dukhanae046f52020-06-15 13:16:14 -07004464 ":fma3_ukernels",
4465 ":avx2_ukernels",
4466 ":avx512f_ukernels",
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07004467 ":avx512skx_ukernels",
Marat Dukhanae046f52020-06-15 13:16:14 -07004468 ],
4469)
4470
Marat Dukhan08c4a432019-10-03 09:29:21 -07004471xnnpack_cc_library(
4472 name = "im2col",
4473 srcs = ["src/im2col.c"],
4474 hdrs = [
4475 "src/xnnpack/common.h",
4476 "src/xnnpack/im2col.h",
4477 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07004478 gcc_copts = xnnpack_gcc_std_copts(),
4479 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004480)
4481
4482xnnpack_cc_library(
4483 name = "indirection",
4484 srcs = ["src/indirection.c"],
4485 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004486 gcc_copts = xnnpack_gcc_std_copts(),
4487 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004488 deps = [
4489 "@FP16",
4490 "@FXdiv",
4491 "@pthreadpool",
4492 ],
4493)
4494
4495xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07004496 name = "indirection_test_mode",
4497 srcs = ["src/indirection.c"],
4498 hdrs = INTERNAL_HDRS,
4499 copts = [
4500 "-UNDEBUG",
4501 "-DXNN_TEST_MODE=1",
4502 ],
4503 gcc_copts = xnnpack_gcc_std_copts(),
4504 msvc_copts = xnnpack_msvc_std_copts(),
4505 deps = [
4506 "@FP16",
4507 "@FXdiv",
4508 "@pthreadpool",
4509 ],
4510)
4511
4512xnnpack_cc_library(
Marat Dukhanab582382020-07-06 13:32:08 -07004513 name = "packing",
4514 srcs = ["src/packing.c"],
4515 hdrs = INTERNAL_HDRS,
4516 gcc_copts = xnnpack_gcc_std_copts(),
4517 msvc_copts = xnnpack_msvc_std_copts(),
4518 deps = [
4519 "@FP16",
4520 "@FXdiv",
4521 "@pthreadpool",
4522 ],
4523)
4524
4525xnnpack_cc_library(
4526 name = "packing_test_mode",
4527 srcs = ["src/packing.c"],
4528 hdrs = INTERNAL_HDRS,
4529 copts = [
4530 "-UNDEBUG",
4531 "-DXNN_TEST_MODE=1",
4532 ],
4533 gcc_copts = xnnpack_gcc_std_copts(),
4534 msvc_copts = xnnpack_msvc_std_copts(),
4535 deps = [
4536 "@FP16",
4537 "@FXdiv",
4538 "@pthreadpool",
4539 ],
4540)
4541
4542xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004543 name = "operator_run",
4544 srcs = ["src/operator-run.c"],
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07004545 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004546 copts = LOGGING_COPTS + select({
Marat Dukhan05702cf2020-03-26 15:41:33 -07004547 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4548 "//conditions:default": [],
4549 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07004550 gcc_copts = xnnpack_gcc_std_copts(),
4551 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004552 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07004553 ":logging_utils",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004554 "@FP16",
4555 "@FXdiv",
4556 "@clog",
4557 "@pthreadpool",
4558 ],
4559)
4560
Chao Mei6ddfc602020-05-13 22:29:36 -07004561xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07004562 name = "operator_run_test_mode",
4563 srcs = ["src/operator-run.c"],
4564 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
4565 copts = LOGGING_COPTS + [
4566 "-UNDEBUG",
4567 "-DXNN_TEST_MODE=1",
4568 ] + select({
4569 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4570 "//conditions:default": [],
4571 }),
4572 gcc_copts = xnnpack_gcc_std_copts(),
4573 msvc_copts = xnnpack_msvc_std_copts(),
4574 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07004575 ":logging_utils",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004576 "@FP16",
4577 "@FXdiv",
4578 "@clog",
4579 "@pthreadpool",
4580 ],
4581)
4582
4583xnnpack_cc_library(
Chao Mei6ddfc602020-05-13 22:29:36 -07004584 name = "memory_planner",
4585 srcs = ["src/memory-planner.c"],
4586 hdrs = INTERNAL_HDRS,
4587 defines = select({
4588 ":xnn_enable_memopt_explicit_true": ["XNN_ENABLE_MEMOPT=1"],
4589 ":xnn_enable_memopt_explicit_false": ["XNN_ENABLE_MEMOPT=0"],
4590 "//conditions:default": ["XNN_ENABLE_MEMOPT=1"],
4591 }),
4592 gcc_copts = xnnpack_gcc_std_copts(),
4593 msvc_copts = xnnpack_msvc_std_copts(),
4594 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07004595 ":logging_utils",
Chao Mei6ddfc602020-05-13 22:29:36 -07004596 "@pthreadpool",
4597 ],
4598)
4599
Marat Dukhan33fcf782020-05-24 14:27:15 -07004600xnnpack_cc_library(
4601 name = "memory_planner_test_mode",
4602 srcs = ["src/memory-planner.c"],
4603 hdrs = INTERNAL_HDRS,
4604 copts = [
4605 "-UNDEBUG",
4606 "-DXNN_TEST_MODE=1",
4607 ],
4608 defines = select({
4609 ":xnn_enable_memopt_explicit_true": ["XNN_ENABLE_MEMOPT=1"],
4610 ":xnn_enable_memopt_explicit_false": ["XNN_ENABLE_MEMOPT=0"],
4611 "//conditions:default": ["XNN_ENABLE_MEMOPT=1"],
4612 }),
4613 gcc_copts = xnnpack_gcc_std_copts(),
4614 msvc_copts = xnnpack_msvc_std_copts(),
4615 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07004616 ":logging_utils",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004617 "@pthreadpool",
4618 ],
4619)
4620
Marat Dukhan08c4a432019-10-03 09:29:21 -07004621cc_library(
4622 name = "enable_assembly",
4623 defines = select({
4624 ":xnn_enable_assembly_explicit_true": ["XNN_ENABLE_ASSEMBLY=1"],
4625 ":xnn_enable_assembly_explicit_false": ["XNN_ENABLE_ASSEMBLY=0"],
Frank Barchard810171d2019-10-10 10:34:51 -07004626 "//conditions:default": ["XNN_ENABLE_ASSEMBLY=1"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004627 }),
4628)
4629
Marat Dukhan9de90e02020-06-18 16:04:12 -07004630cc_library(
4631 name = "enable_sparse",
4632 defines = select({
4633 ":xnn_enable_sparse_explicit_true": ["XNN_ENABLE_SPARSE=1"],
4634 ":xnn_enable_sparse_explicit_false": ["XNN_ENABLE_SPARSE=0"],
Marat Dukhanb36582b2020-12-08 11:16:28 -08004635 "//conditions:default": ["XNN_ENABLE_SPARSE=1"],
Marat Dukhan9de90e02020-06-18 16:04:12 -07004636 }),
4637)
4638
Marat Dukhancf056b22019-10-07 10:26:29 -07004639xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004640 name = "operators",
4641 srcs = OPERATOR_SRCS + [
Marat Dukhan04f03be2019-11-19 12:36:47 -08004642 "src/memory.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004643 "src/operator-delete.c",
Marat Dukhancf056b22019-10-07 10:26:29 -07004644 ],
4645 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004646 copts = LOGGING_COPTS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07004647 "-Isrc",
4648 "-Iinclude",
4649 ] + select({
4650 ":debug_build": [],
4651 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07004652 }) + select({
4653 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4654 "//conditions:default": [],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004655 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07004656 gcc_copts = xnnpack_gcc_std_copts(),
4657 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004658 deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07004659 ":indirection",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004660 ":logging_utils",
Marat Dukhanab582382020-07-06 13:32:08 -07004661 ":packing",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004662 "@FP16",
4663 "@FXdiv",
4664 "@clog",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004665 "@pthreadpool",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004666 ],
4667)
4668
Marat Dukhan10a38082020-04-17 03:58:35 -07004669xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07004670 name = "operators_test_mode",
4671 srcs = OPERATOR_SRCS + [
4672 "src/memory.c",
4673 "src/operator-delete.c",
4674 ],
4675 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
4676 copts = LOGGING_COPTS + [
4677 "-Isrc",
4678 "-Iinclude",
4679 "-UNDEBUG",
4680 "-DXNN_TEST_MODE=1",
4681 ] + select({
4682 ":debug_build": [],
4683 "//conditions:default": xnnpack_min_size_copts(),
4684 }) + select({
4685 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4686 "//conditions:default": [],
4687 }),
4688 gcc_copts = xnnpack_gcc_std_copts(),
4689 msvc_copts = xnnpack_msvc_std_copts(),
4690 deps = [
4691 ":indirection_test_mode",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004692 ":logging_utils",
Marat Dukhanab582382020-07-06 13:32:08 -07004693 ":packing_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004694 "@FP16",
4695 "@FXdiv",
4696 "@clog",
4697 "@pthreadpool",
4698 ],
4699)
4700
4701xnnpack_cc_library(
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004702 name = "XNNPACK",
4703 srcs = [
4704 "src/init.c",
Marat Dukhanccfdbd12020-02-03 14:27:45 -08004705 "src/runtime.c",
4706 "src/subgraph.c",
4707 "src/tensor.c",
Marat Dukhanf03da0d2020-06-10 16:00:20 -07004708 ] + SUBGRAPH_SRCS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004709 hdrs = ["include/xnnpack.h"],
4710 copts = LOGGING_COPTS + [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004711 "-Isrc",
4712 "-Iinclude",
4713 ] + select({
4714 ":debug_build": [],
4715 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07004716 }) + select({
4717 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4718 "//conditions:default": [],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004719 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07004720 gcc_copts = xnnpack_gcc_std_copts(),
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004721 includes = ["include"],
Marat Dukhan10a38082020-04-17 03:58:35 -07004722 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004723 visibility = xnnpack_visibility(),
4724 deps = [
4725 ":enable_assembly",
Marat Dukhan9de90e02020-06-18 16:04:12 -07004726 ":enable_sparse",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004727 ":logging_utils",
Chao Mei6ddfc602020-05-13 22:29:36 -07004728 ":memory_planner",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004729 ":operator_run",
4730 ":operators",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004731 ":ukernels",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004732 "@clog",
Marat Dukhanab2946c2020-05-21 20:04:13 -07004733 "@FP16",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004734 "@pthreadpool",
Marat Dukhand343c222019-10-07 09:22:14 -07004735 ] + select({
4736 ":emscripten": [],
4737 "//conditions:default": ["@cpuinfo"],
4738 }),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004739)
4740
Marat Dukhan10a38082020-04-17 03:58:35 -07004741xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07004742 name = "XNNPACK_test_mode",
4743 srcs = [
4744 "src/init.c",
4745 "src/runtime.c",
4746 "src/subgraph.c",
4747 "src/tensor.c",
Marat Dukhanf03da0d2020-06-10 16:00:20 -07004748 ] + SUBGRAPH_SRCS,
Marat Dukhan33fcf782020-05-24 14:27:15 -07004749 hdrs = ["include/xnnpack.h"],
4750 copts = LOGGING_COPTS + [
4751 "-Isrc",
4752 "-Iinclude",
4753 "-UNDEBUG",
4754 "-DXNN_TEST_MODE=1",
4755 ] + select({
4756 ":debug_build": [],
4757 "//conditions:default": xnnpack_min_size_copts(),
4758 }) + select({
4759 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4760 "//conditions:default": [],
4761 }),
4762 gcc_copts = xnnpack_gcc_std_copts(),
4763 includes = ["include"],
4764 msvc_copts = xnnpack_msvc_std_copts(),
4765 visibility = xnnpack_visibility(),
4766 deps = [
4767 ":enable_assembly",
Marat Dukhan9de90e02020-06-18 16:04:12 -07004768 ":enable_sparse",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004769 ":logging_utils",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004770 ":memory_planner_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004771 ":operator_run_test_mode",
4772 ":operators_test_mode",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004773 ":ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004774 "@clog",
4775 "@FP16",
4776 "@pthreadpool",
4777 ] + select({
4778 ":emscripten": [],
4779 "//conditions:default": ["@cpuinfo"],
4780 }),
4781)
4782
4783xnnpack_cc_library(
Marat Dukhanae046f52020-06-15 13:16:14 -07004784 name = "xnnpack_f32",
4785 srcs = [
4786 "src/init.c",
4787 "src/runtime.c",
4788 "src/subgraph.c",
4789 "src/tensor.c",
4790 ] + SUBGRAPH_SRCS,
4791 hdrs = ["include/xnnpack.h"],
4792 copts = LOGGING_COPTS + [
4793 "-Isrc",
4794 "-Iinclude",
4795 ] + select({
4796 ":debug_build": [],
4797 "//conditions:default": xnnpack_min_size_copts(),
4798 }) + select({
4799 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4800 "//conditions:default": [],
4801 }),
4802 defines = [
Marat Dukhan16f1e1a2020-08-04 16:38:22 -07004803 "XNN_NO_QS8_OPERATORS",
Marat Dukhan08b7a972020-07-14 18:17:29 -07004804 "XNN_NO_QU8_OPERATORS",
Marat Dukhanae046f52020-06-15 13:16:14 -07004805 "XNN_NO_U8_OPERATORS",
4806 "XNN_NO_X8_OPERATORS",
4807 "XNN_NO_F16_OPERATORS",
4808 "XNN_NO_X16_OPERATORS",
4809 ],
4810 gcc_copts = xnnpack_gcc_std_copts(),
4811 includes = ["include"],
4812 msvc_copts = xnnpack_msvc_std_copts(),
4813 visibility = xnnpack_visibility(),
4814 deps = [
4815 ":enable_assembly",
Marat Dukhan9de90e02020-06-18 16:04:12 -07004816 ":enable_sparse",
Marat Dukhanae046f52020-06-15 13:16:14 -07004817 ":logging_utils",
4818 ":memory_planner",
4819 ":operator_run",
4820 ":operators",
Marat Dukhan86df3842020-06-17 14:03:31 -07004821 ":ukernels_max_compatibility",
Marat Dukhanae046f52020-06-15 13:16:14 -07004822 "@clog",
4823 "@FP16",
4824 "@pthreadpool",
4825 ] + select({
4826 ":emscripten": [],
4827 "//conditions:default": ["@cpuinfo"],
4828 }),
4829)
4830
4831xnnpack_cc_library(
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004832 name = "xnnpack_operators_nhwc_f32",
4833 srcs = [
4834 "src/init.c",
4835 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07004836 hdrs = ["include/xnnpack.h"],
4837 copts = LOGGING_COPTS + [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004838 "-Isrc",
4839 "-Iinclude",
4840 ] + select({
4841 ":debug_build": [],
4842 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07004843 }) + select({
4844 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4845 "//conditions:default": [],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004846 }),
4847 defines = [
Marat Dukhan16f1e1a2020-08-04 16:38:22 -07004848 "XNN_NO_QS8_OPERATORS",
Marat Dukhan08b7a972020-07-14 18:17:29 -07004849 "XNN_NO_QU8_OPERATORS",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004850 "XNN_NO_U8_OPERATORS",
4851 "XNN_NO_X8_OPERATORS",
Marat Dukhanefc47b82019-11-18 09:25:38 -08004852 "XNN_NO_NCHW_OPERATORS",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004853 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07004854 gcc_copts = xnnpack_gcc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004855 includes = ["include"],
Marat Dukhan10a38082020-04-17 03:58:35 -07004856 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004857 visibility = xnnpack_visibility(),
4858 deps = [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004859 ":enable_assembly",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004860 ":logging_utils",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004861 ":operator_run",
4862 ":operators",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004863 ":ukernels",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004864 "@clog",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004865 "@pthreadpool",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004866 ] + select({
4867 ":emscripten": [],
4868 "//conditions:default": ["@cpuinfo"],
4869 }),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004870)
4871
Marat Dukhancf056b22019-10-07 10:26:29 -07004872xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004873 name = "bench_utils",
4874 srcs = ["bench/utils.cc"],
4875 hdrs = ["bench/utils.h"],
Marat Dukhanbad48fe2019-11-04 10:35:22 -08004876 deps = [
4877 "@com_google_benchmark//:benchmark",
4878 "@cpuinfo",
4879 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004880)
4881
Frank Barchard7e955972019-10-11 10:34:25 -07004882######################### Benchmarks for micro-kernels #########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07004883
4884xnnpack_benchmark(
Marat Dukhan595e1702020-07-31 10:12:52 -07004885 name = "qs8_gemm_bench",
4886 srcs = [
4887 "bench/gemm.h",
4888 "bench/qs8-gemm.cc",
4889 "src/xnnpack/AlignedAllocator.h",
4890 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Frank Barchard31328cb2020-10-12 11:55:18 -07004891 copts = xnnpack_optional_ruy_copts() + xnnpack_optional_gemmlowp_copts(),
4892 deps = MICROKERNEL_BENCHMARK_DEPS + [":packing"] + xnnpack_optional_ruy_deps() + xnnpack_optional_gemmlowp_deps(),
Marat Dukhan595e1702020-07-31 10:12:52 -07004893)
4894
4895xnnpack_benchmark(
Marat Dukhan56bdd4a2020-08-03 19:47:04 -07004896 name = "qs8_requantization_bench",
4897 srcs = [
4898 "bench/qs8-requantization.cc",
Marat Dukhan56bdd4a2020-08-03 19:47:04 -07004899 "src/xnnpack/AlignedAllocator.h",
Frank Barchard04336c12020-10-22 16:48:55 -07004900 "src/xnnpack/requantization-stubs.h",
Marat Dukhan56bdd4a2020-08-03 19:47:04 -07004901 ] + MICROKERNEL_BENCHMARK_HDRS,
4902 deps = MICROKERNEL_BENCHMARK_DEPS,
4903)
4904
4905xnnpack_benchmark(
Marat Dukhan08b7a972020-07-14 18:17:29 -07004906 name = "qu8_gemm_bench",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004907 srcs = [
4908 "bench/gemm.h",
Marat Dukhan08b7a972020-07-14 18:17:29 -07004909 "bench/qu8-gemm.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004910 "src/xnnpack/AlignedAllocator.h",
4911 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004912 copts = xnnpack_optional_ruy_copts() + xnnpack_optional_gemmlowp_copts(),
Marat Dukhanab582382020-07-06 13:32:08 -07004913 deps = MICROKERNEL_BENCHMARK_DEPS + [":packing"] + xnnpack_optional_ruy_deps() + xnnpack_optional_gemmlowp_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004914)
4915
4916xnnpack_benchmark(
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07004917 name = "qu8_requantization_bench",
4918 srcs = [
4919 "bench/qu8-requantization.cc",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07004920 "src/xnnpack/AlignedAllocator.h",
Frank Barchard04336c12020-10-22 16:48:55 -07004921 "src/xnnpack/requantization-stubs.h",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07004922 ] + MICROKERNEL_BENCHMARK_HDRS,
4923 deps = MICROKERNEL_BENCHMARK_DEPS,
4924)
4925
4926xnnpack_benchmark(
Frank Barchard40d20fe2020-05-05 00:37:45 -07004927 name = "f16_igemm_bench",
4928 srcs = [
4929 "bench/f16-igemm.cc",
4930 "bench/conv.h",
4931 "bench/google/conv.h",
4932 "src/xnnpack/AlignedAllocator.h",
4933 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004934 deps = MICROKERNEL_BENCHMARK_DEPS + [
4935 ":indirection",
4936 ":packing",
4937 ],
Frank Barchard40d20fe2020-05-05 00:37:45 -07004938)
4939
4940xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004941 name = "f16_gemm_bench",
4942 srcs = [
4943 "bench/f16-gemm.cc",
4944 "bench/gemm.h",
4945 "src/xnnpack/AlignedAllocator.h",
4946 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004947 deps = MICROKERNEL_BENCHMARK_DEPS + [
4948 ":packing",
4949 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004950)
4951
4952xnnpack_benchmark(
Marat Dukhanbdb56f52020-02-05 21:42:49 -08004953 name = "f16_spmm_bench",
4954 srcs = [
4955 "bench/f16-spmm.cc",
Marat Dukhan1631e3e2020-12-06 19:29:31 -08004956 "bench/spmm.h",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08004957 "src/xnnpack/AlignedAllocator.h",
4958 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanbdb56f52020-02-05 21:42:49 -08004959 deps = MICROKERNEL_BENCHMARK_DEPS,
4960)
4961
4962xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004963 name = "f32_igemm_bench",
4964 srcs = [
4965 "bench/f32-igemm.cc",
4966 "bench/conv.h",
4967 "src/xnnpack/AlignedAllocator.h",
4968 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004969 deps = MICROKERNEL_BENCHMARK_DEPS + [
4970 ":indirection",
4971 ":packing",
4972 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004973)
4974
4975xnnpack_benchmark(
Frank Barchardc2010062020-07-19 10:04:12 -07004976 name = "f16_relu_bench",
4977 srcs = [
4978 "bench/f16-relu.cc",
4979 "src/xnnpack/AlignedAllocator.h",
4980 ] + MICROKERNEL_BENCHMARK_HDRS,
4981 deps = MICROKERNEL_BENCHMARK_DEPS,
4982)
4983
4984xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004985 name = "f32_conv_hwc_bench",
4986 srcs = [
4987 "bench/f32-conv-hwc.cc",
4988 "bench/dconv.h",
4989 "src/xnnpack/AlignedAllocator.h",
4990 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004991 deps = MICROKERNEL_BENCHMARK_DEPS + [
4992 ":packing",
4993 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004994)
4995
4996xnnpack_benchmark(
Marat Dukhan1f29b802020-05-15 23:46:39 -07004997 name = "f32_conv_hwc2chw_bench",
Erich Elsen563df5f2019-10-23 08:02:21 -07004998 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07004999 "bench/f32-conv-hwc2chw.cc",
Erich Elsen563df5f2019-10-23 08:02:21 -07005000 "bench/dconv.h",
5001 "src/xnnpack/AlignedAllocator.h",
5002 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005003 deps = MICROKERNEL_BENCHMARK_DEPS + [
5004 ":packing",
5005 ],
Erich Elsen563df5f2019-10-23 08:02:21 -07005006)
5007
5008xnnpack_benchmark(
Frank Barchard5a599a62020-06-04 20:12:44 -07005009 name = "f16_dwconv_bench",
5010 srcs = [
5011 "bench/f16-dwconv.cc",
5012 "bench/dwconv.h",
5013 "bench/google/dwconv.h",
5014 "src/xnnpack/AlignedAllocator.h",
5015 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005016 deps = MICROKERNEL_BENCHMARK_DEPS + [
5017 ":indirection",
5018 ":packing",
5019 ],
Frank Barchard5a599a62020-06-04 20:12:44 -07005020)
5021
5022xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005023 name = "f32_dwconv_bench",
5024 srcs = [
5025 "bench/f32-dwconv.cc",
5026 "bench/dwconv.h",
5027 "src/xnnpack/AlignedAllocator.h",
5028 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005029 deps = MICROKERNEL_BENCHMARK_DEPS + [
5030 ":indirection",
5031 ":packing",
5032 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005033)
5034
5035xnnpack_benchmark(
Marat Dukhanbf715f92020-10-23 20:17:00 -07005036 name = "f32_dwconv2d_chw_bench",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005037 srcs = [
Marat Dukhanbf715f92020-10-23 20:17:00 -07005038 "bench/f32-dwconv2d-chw.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005039 "bench/dwconv.h",
5040 "src/xnnpack/AlignedAllocator.h",
5041 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005042 deps = MICROKERNEL_BENCHMARK_DEPS + [
5043 ":indirection",
5044 ":packing",
5045 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005046)
5047
5048xnnpack_benchmark(
5049 name = "f32_gemm_bench",
5050 srcs = [
5051 "bench/f32-gemm.cc",
5052 "bench/gemm.h",
5053 "src/xnnpack/AlignedAllocator.h",
5054 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07005055 copts = xnnpack_optional_ruy_copts(),
Marat Dukhanab582382020-07-06 13:32:08 -07005056 deps = MICROKERNEL_BENCHMARK_DEPS + [":packing"] + xnnpack_optional_ruy_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07005057)
5058
5059xnnpack_benchmark(
Marat Dukhana11ca342020-06-25 23:45:07 -07005060 name = "f32_hswish_bench",
5061 srcs = [
5062 "bench/f32-hswish.cc",
5063 "src/xnnpack/AlignedAllocator.h",
5064 ] + MICROKERNEL_BENCHMARK_HDRS,
5065 deps = MICROKERNEL_BENCHMARK_DEPS,
5066)
5067
5068xnnpack_benchmark(
Marat Dukhan4c4eb002019-12-08 21:27:49 -08005069 name = "f32_raddexpminusmax_bench",
5070 srcs = [
5071 "bench/f32-raddexpminusmax.cc",
5072 "src/xnnpack/AlignedAllocator.h",
5073 ] + MICROKERNEL_BENCHMARK_HDRS,
5074 deps = MICROKERNEL_BENCHMARK_DEPS,
5075)
5076
5077xnnpack_benchmark(
5078 name = "f32_raddextexp_bench",
5079 srcs = [
5080 "bench/f32-raddextexp.cc",
5081 "src/xnnpack/AlignedAllocator.h",
5082 ] + MICROKERNEL_BENCHMARK_HDRS,
5083 deps = MICROKERNEL_BENCHMARK_DEPS,
5084)
5085
5086xnnpack_benchmark(
5087 name = "f32_raddstoreexpminusmax_bench",
5088 srcs = [
5089 "bench/f32-raddstoreexpminusmax.cc",
5090 "src/xnnpack/AlignedAllocator.h",
5091 ] + MICROKERNEL_BENCHMARK_HDRS,
5092 deps = MICROKERNEL_BENCHMARK_DEPS,
5093)
5094
5095xnnpack_benchmark(
Frank Barchard3c4a9522020-07-18 16:11:22 -07005096 name = "f32_relu_bench",
5097 srcs = [
5098 "bench/f32-relu.cc",
5099 "src/xnnpack/AlignedAllocator.h",
5100 ] + MICROKERNEL_BENCHMARK_HDRS,
5101 deps = MICROKERNEL_BENCHMARK_DEPS,
5102)
5103
5104xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005105 name = "f32_rmax_bench",
5106 srcs = [
5107 "bench/f32-rmax.cc",
5108 "src/xnnpack/AlignedAllocator.h",
5109 ] + MICROKERNEL_BENCHMARK_HDRS,
5110 deps = MICROKERNEL_BENCHMARK_DEPS,
5111)
5112
5113xnnpack_benchmark(
Marat Dukhan14bec502019-11-18 11:35:31 -08005114 name = "f32_sigmoid_bench",
5115 srcs = [
5116 "bench/f32-sigmoid.cc",
5117 "src/xnnpack/AlignedAllocator.h",
5118 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan14bec502019-11-18 11:35:31 -08005119 deps = MICROKERNEL_BENCHMARK_DEPS,
5120)
5121
5122xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005123 name = "f32_spmm_bench",
5124 srcs = [
5125 "bench/f32-spmm.cc",
Marat Dukhan1631e3e2020-12-06 19:29:31 -08005126 "bench/spmm.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005127 "src/xnnpack/AlignedAllocator.h",
5128 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005129 deps = MICROKERNEL_BENCHMARK_DEPS,
5130)
5131
5132xnnpack_benchmark(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08005133 name = "f32_softmax_bench",
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07005134 srcs = [
Marat Dukhanfd8e6892020-01-27 15:25:25 -08005135 "bench/f32-softmax.cc",
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07005136 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07005137 copts = xnnpack_optional_dnnl_copts(),
Marat Dukhan8d3c6932020-03-06 20:27:27 -08005138 deps = MICROKERNEL_BENCHMARK_DEPS + xnnpack_optional_dnnl_deps(),
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07005139)
5140
5141xnnpack_benchmark(
Marat Dukhaned6baaf2020-12-01 15:07:08 -08005142 name = "f32_velu_bench",
5143 srcs = [
5144 "bench/f32-velu.cc",
5145 "src/xnnpack/AlignedAllocator.h",
5146 ] + MICROKERNEL_BENCHMARK_HDRS,
5147 deps = MICROKERNEL_BENCHMARK_DEPS,
5148)
5149
5150xnnpack_benchmark(
Marat Dukhan4c4eb002019-12-08 21:27:49 -08005151 name = "f32_vscaleexpminusmax_bench",
5152 srcs = [
5153 "bench/f32-vscaleexpminusmax.cc",
5154 "src/xnnpack/AlignedAllocator.h",
5155 ] + MICROKERNEL_BENCHMARK_HDRS,
5156 deps = MICROKERNEL_BENCHMARK_DEPS,
5157)
5158
5159xnnpack_benchmark(
5160 name = "f32_vscaleextexp_bench",
5161 srcs = [
5162 "bench/f32-vscaleextexp.cc",
5163 "src/xnnpack/AlignedAllocator.h",
5164 ] + MICROKERNEL_BENCHMARK_HDRS,
5165 deps = MICROKERNEL_BENCHMARK_DEPS,
5166)
5167
5168xnnpack_benchmark(
Marat Dukhanf4db2f32020-06-30 10:55:30 -07005169 name = "f32_vsqrt_bench",
5170 srcs = [
5171 "bench/f32-vsqrt.cc",
5172 "src/xnnpack/AlignedAllocator.h",
5173 ] + MICROKERNEL_BENCHMARK_HDRS,
5174 deps = MICROKERNEL_BENCHMARK_DEPS,
5175)
5176
5177xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005178 name = "f32_im2col_gemm_bench",
5179 srcs = [
5180 "bench/f32-im2col-gemm.cc",
5181 "bench/conv.h",
5182 "src/xnnpack/AlignedAllocator.h",
5183 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005184 deps = MICROKERNEL_BENCHMARK_DEPS + [
5185 ":im2col",
5186 ":packing",
5187 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005188)
5189
Marat Dukhanfe7acb62020-03-09 19:30:05 -07005190xnnpack_benchmark(
Marat Dukhanffbf96a2020-05-14 02:59:08 -07005191 name = "rounding_bench",
5192 srcs = [
5193 "bench/rounding.cc",
Marat Dukhanffbf96a2020-05-14 02:59:08 -07005194 "src/xnnpack/AlignedAllocator.h",
Frank Barchard04336c12020-10-22 16:48:55 -07005195 "src/xnnpack/math-stubs.h",
Marat Dukhanffbf96a2020-05-14 02:59:08 -07005196 ] + MICROKERNEL_BENCHMARK_HDRS,
5197 deps = MICROKERNEL_BENCHMARK_DEPS,
5198)
5199
Marat Dukhan08c4a432019-10-03 09:29:21 -07005200########################### Benchmarks for operators ###########################
5201
5202xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005203 name = "average_pooling_bench",
5204 srcs = ["bench/average-pooling.cc"],
Marat Dukhan7a16d8b2020-03-11 04:22:44 -07005205 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07005206 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07005207 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07005208)
5209
5210xnnpack_benchmark(
Marat Dukhan87727142020-06-24 15:24:10 -07005211 name = "bankers_rounding_bench",
5212 srcs = ["bench/bankers-rounding.cc"],
5213 copts = xnnpack_optional_tflite_copts(),
5214 tags = ["nowin32"],
5215 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
5216)
5217
5218xnnpack_benchmark(
5219 name = "ceiling_bench",
5220 srcs = ["bench/ceiling.cc"],
5221 copts = xnnpack_optional_tflite_copts(),
5222 tags = ["nowin32"],
5223 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
5224)
5225
5226xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005227 name = "channel_shuffle_bench",
5228 srcs = ["bench/channel-shuffle.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07005229 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005230)
5231
5232xnnpack_benchmark(
5233 name = "convolution_bench",
5234 srcs = ["bench/convolution.cc"],
5235 copts = xnnpack_optional_tflite_copts() + xnnpack_optional_armcl_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07005236 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07005237 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps() + xnnpack_optional_armcl_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07005238)
5239
5240xnnpack_benchmark(
5241 name = "deconvolution_bench",
5242 srcs = ["bench/deconvolution.cc"],
5243 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07005244 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07005245 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07005246)
5247
5248xnnpack_benchmark(
Marat Dukhanb6bd4bc2020-12-01 17:01:40 -08005249 name = "elu_bench",
5250 srcs = ["bench/elu.cc"],
5251 copts = xnnpack_optional_tflite_copts(),
5252 tags = ["nowin32"],
5253 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
5254)
5255
5256xnnpack_benchmark(
Marat Dukhan87727142020-06-24 15:24:10 -07005257 name = "floor_bench",
5258 srcs = ["bench/floor.cc"],
5259 copts = xnnpack_optional_tflite_copts(),
5260 tags = ["nowin32"],
5261 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
5262)
5263
5264xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005265 name = "global_average_pooling_bench",
5266 srcs = ["bench/global-average-pooling.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07005267 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005268)
5269
5270xnnpack_benchmark(
Marat Dukhanad352602020-06-25 21:50:54 -07005271 name = "hardswish_bench",
5272 srcs = ["bench/hardswish.cc"],
5273 copts = xnnpack_optional_tflite_copts(),
5274 tags = ["nowin32"],
5275 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
5276)
5277
5278xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005279 name = "max_pooling_bench",
5280 srcs = ["bench/max-pooling.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07005281 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005282)
5283
5284xnnpack_benchmark(
5285 name = "sigmoid_bench",
5286 srcs = ["bench/sigmoid.cc"],
Marat Dukhanc3b9e862019-11-17 13:18:54 -08005287 copts = xnnpack_optional_tflite_copts(),
Marat Dukhanca2ba702020-04-24 01:31:47 -07005288 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07005289 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07005290)
5291
5292xnnpack_benchmark(
Marat Dukhan95b22432019-10-30 16:30:14 -07005293 name = "prelu_bench",
5294 srcs = ["bench/prelu.cc"],
5295 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07005296 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07005297 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan95b22432019-10-30 16:30:14 -07005298)
5299
5300xnnpack_benchmark(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08005301 name = "softmax_bench",
5302 srcs = ["bench/softmax.cc"],
Marat Dukhan9c0db962020-01-28 12:30:14 -08005303 copts = xnnpack_optional_tflite_copts(),
Marat Dukhanca2ba702020-04-24 01:31:47 -07005304 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07005305 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07005306)
5307
Marat Dukhan87727142020-06-24 15:24:10 -07005308xnnpack_benchmark(
Marat Dukhan6804bbd2020-06-30 19:26:11 -07005309 name = "square_root_bench",
5310 srcs = ["bench/square-root.cc"],
5311 copts = xnnpack_optional_tflite_copts(),
5312 tags = ["nowin32"],
5313 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
5314)
5315
5316xnnpack_benchmark(
Marat Dukhan87727142020-06-24 15:24:10 -07005317 name = "truncation_bench",
5318 srcs = ["bench/truncation.cc"],
5319 deps = OPERATOR_BENCHMARK_DEPS,
5320)
5321
Marat Dukhanc068bb62019-10-04 13:24:39 -07005322############################# End-to-end benchmarks ############################
5323
5324cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07005325 name = "fp32_mobilenet_v1",
5326 srcs = ["models/fp32-mobilenet-v1.cc"],
Marat Dukhanc068bb62019-10-04 13:24:39 -07005327 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08005328 copts = xnnpack_std_cxxopts(),
Marat Dukhanc068bb62019-10-04 13:24:39 -07005329 linkstatic = True,
5330 deps = [
5331 ":XNNPACK",
5332 "@pthreadpool",
5333 ],
5334)
5335
5336cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07005337 name = "fp16_mobilenet_v1",
5338 srcs = ["models/fp16-mobilenet-v1.cc"],
5339 hdrs = ["models/models.h"],
5340 copts = xnnpack_std_cxxopts(),
5341 linkstatic = True,
5342 deps = [
5343 ":XNNPACK",
5344 "@FP16",
5345 "@pthreadpool",
5346 ],
5347)
5348
5349cc_library(
Marat Dukhan0743cdf2020-08-04 18:52:07 -07005350 name = "qs8_mobilenet_v1",
5351 srcs = ["models/qs8-mobilenet-v1.cc"],
5352 hdrs = ["models/models.h"],
5353 copts = xnnpack_std_cxxopts(),
5354 linkstatic = True,
5355 deps = [
5356 ":XNNPACK",
5357 "@pthreadpool",
5358 ],
5359)
5360
5361cc_library(
Marat Dukhan70a96182020-09-03 17:13:58 -07005362 name = "qs8_mobilenet_v2",
5363 srcs = ["models/qs8-mobilenet-v2.cc"],
5364 hdrs = ["models/models.h"],
5365 copts = xnnpack_std_cxxopts(),
5366 linkstatic = True,
5367 deps = [
5368 ":XNNPACK",
5369 "@pthreadpool",
5370 ],
5371)
5372
5373cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07005374 name = "fp32_mobilenet_v2",
5375 srcs = ["models/fp32-mobilenet-v2.cc"],
Marat Dukhanc068bb62019-10-04 13:24:39 -07005376 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08005377 copts = xnnpack_std_cxxopts(),
Marat Dukhanc068bb62019-10-04 13:24:39 -07005378 linkstatic = True,
5379 deps = [
5380 ":XNNPACK",
5381 "@pthreadpool",
5382 ],
5383)
5384
Marat Dukhanc08cdf52019-12-09 09:17:51 -08005385cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07005386 name = "fp16_mobilenet_v2",
5387 srcs = ["models/fp16-mobilenet-v2.cc"],
5388 hdrs = ["models/models.h"],
5389 copts = xnnpack_std_cxxopts(),
5390 linkstatic = True,
5391 deps = [
5392 ":XNNPACK",
5393 "@FP16",
5394 "@pthreadpool",
5395 ],
5396)
5397
5398cc_library(
5399 name = "fp32_mobilenet_v3_large",
5400 srcs = ["models/fp32-mobilenet-v3-large.cc"],
Marat Dukhanc08cdf52019-12-09 09:17:51 -08005401 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08005402 copts = xnnpack_std_cxxopts(),
Marat Dukhanc08cdf52019-12-09 09:17:51 -08005403 linkstatic = True,
5404 deps = [
5405 ":XNNPACK",
5406 "@pthreadpool",
5407 ],
5408)
5409
5410cc_library(
Marat Dukhan66f3ccd2020-09-14 16:09:38 -07005411 name = "fp16_mobilenet_v3_large",
5412 srcs = ["models/fp16-mobilenet-v3-large.cc"],
5413 hdrs = ["models/models.h"],
5414 copts = xnnpack_std_cxxopts(),
5415 linkstatic = True,
5416 deps = [
5417 ":XNNPACK",
5418 "@FP16",
5419 "@pthreadpool",
5420 ],
5421)
5422
5423cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07005424 name = "fp32_mobilenet_v3_small",
5425 srcs = ["models/fp32-mobilenet-v3-small.cc"],
Marat Dukhanc08cdf52019-12-09 09:17:51 -08005426 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08005427 copts = xnnpack_std_cxxopts(),
Marat Dukhanc08cdf52019-12-09 09:17:51 -08005428 linkstatic = True,
5429 deps = [
5430 ":XNNPACK",
5431 "@pthreadpool",
5432 ],
5433)
5434
Marat Dukhan66f3ccd2020-09-14 16:09:38 -07005435cc_library(
5436 name = "fp16_mobilenet_v3_small",
5437 srcs = ["models/fp16-mobilenet-v3-small.cc"],
5438 hdrs = ["models/models.h"],
5439 copts = xnnpack_std_cxxopts(),
5440 linkstatic = True,
5441 deps = [
5442 ":XNNPACK",
5443 "@FP16",
5444 "@pthreadpool",
5445 ],
5446)
5447
Marat Dukhanc068bb62019-10-04 13:24:39 -07005448xnnpack_benchmark(
Marat Dukhanef4416e2019-10-31 13:44:40 -07005449 name = "f32_dwconv_e2e_bench",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08005450 srcs = [
5451 "bench/f32-dwconv-e2e.cc",
5452 "bench/end2end.h",
5453 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanef4416e2019-10-31 13:44:40 -07005454 deps = MICROKERNEL_BENCHMARK_DEPS + [
5455 ":XNNPACK",
Marat Dukhan270a2c42020-06-26 16:45:52 -07005456 ":fp32_mobilenet_v1",
5457 ":fp32_mobilenet_v2",
5458 ":fp32_mobilenet_v3_large",
5459 ":fp32_mobilenet_v3_small",
Marat Dukhanef4416e2019-10-31 13:44:40 -07005460 ],
5461)
5462
5463xnnpack_benchmark(
Marat Dukhan5f18d262019-10-31 10:24:14 -07005464 name = "f32_gemm_e2e_bench",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08005465 srcs = [
5466 "bench/f32-gemm-e2e.cc",
5467 "bench/end2end.h",
5468 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan5f18d262019-10-31 10:24:14 -07005469 deps = MICROKERNEL_BENCHMARK_DEPS + [
5470 ":XNNPACK",
Marat Dukhan270a2c42020-06-26 16:45:52 -07005471 ":fp32_mobilenet_v1",
5472 ":fp32_mobilenet_v2",
5473 ":fp32_mobilenet_v3_large",
5474 ":fp32_mobilenet_v3_small",
Marat Dukhan5f18d262019-10-31 10:24:14 -07005475 ],
5476)
5477
5478xnnpack_benchmark(
Frank Barcharddc909cb2021-02-08 13:59:31 -08005479 name = "qs8_gemm_e2e_bench",
5480 srcs = [
5481 "bench/qs8-gemm-e2e.cc",
5482 "bench/end2end.h",
5483 ] + MICROKERNEL_BENCHMARK_HDRS,
5484 deps = MICROKERNEL_BENCHMARK_DEPS + [
5485 ":XNNPACK",
5486 ":qs8_mobilenet_v1",
5487 ":qs8_mobilenet_v2",
5488 ],
5489)
5490
5491xnnpack_benchmark(
Marat Dukhanc068bb62019-10-04 13:24:39 -07005492 name = "end2end_bench",
5493 srcs = ["bench/end2end.cc"],
5494 deps = [
5495 ":XNNPACK",
Frank Barchardc712fa42019-10-31 14:00:21 -07005496 ":bench_utils",
Marat Dukhan270a2c42020-06-26 16:45:52 -07005497 ":fp16_mobilenet_v1",
5498 ":fp16_mobilenet_v2",
Marat Dukhan66f3ccd2020-09-14 16:09:38 -07005499 ":fp16_mobilenet_v3_large",
5500 ":fp16_mobilenet_v3_small",
Marat Dukhan270a2c42020-06-26 16:45:52 -07005501 ":fp32_mobilenet_v1",
5502 ":fp32_mobilenet_v2",
5503 ":fp32_mobilenet_v3_large",
5504 ":fp32_mobilenet_v3_small",
Marat Dukhan0743cdf2020-08-04 18:52:07 -07005505 ":qs8_mobilenet_v1",
Marat Dukhan70a96182020-09-03 17:13:58 -07005506 ":qs8_mobilenet_v2",
Marat Dukhanc068bb62019-10-04 13:24:39 -07005507 "@pthreadpool",
5508 ],
5509)
5510
Marat Dukhan6adff4e2019-10-14 18:32:07 -07005511#################### Accuracy evaluation for math functions ####################
5512
5513xnnpack_benchmark(
Marat Dukhancf67ec62020-12-14 09:56:37 -08005514 name = "f32_exp_ulp_eval",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07005515 srcs = [
Marat Dukhancf67ec62020-12-14 09:56:37 -08005516 "eval/f32-exp-ulp.cc",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07005517 "src/xnnpack/AlignedAllocator.h",
5518 ] + ACCURACY_EVAL_HDRS,
Marat Dukhan3a305212020-12-06 19:24:27 -08005519 deps = ACCURACY_EVAL_DEPS + [
5520 ":bench_utils",
5521 "@cpuinfo",
5522 ],
Marat Dukhan6adff4e2019-10-14 18:32:07 -07005523)
5524
Marat Dukhan515c9772019-10-17 18:07:57 -07005525xnnpack_benchmark(
Marat Dukhancf67ec62020-12-14 09:56:37 -08005526 name = "f32_expminus_ulp_eval",
Marat Dukhan515c9772019-10-17 18:07:57 -07005527 srcs = [
Marat Dukhancf67ec62020-12-14 09:56:37 -08005528 "eval/f32-expminus-ulp.cc",
Marat Dukhan515c9772019-10-17 18:07:57 -07005529 "src/xnnpack/AlignedAllocator.h",
5530 ] + ACCURACY_EVAL_HDRS,
Marat Dukhan3a305212020-12-06 19:24:27 -08005531 deps = ACCURACY_EVAL_DEPS + [
5532 ":bench_utils",
5533 "@cpuinfo",
5534 ],
Marat Dukhan515c9772019-10-17 18:07:57 -07005535)
5536
Marat Dukhan98ba4412019-10-23 02:14:28 -07005537xnnpack_benchmark(
Marat Dukhancf67ec62020-12-14 09:56:37 -08005538 name = "f32_expm1minus_ulp_eval",
Marat Dukhana438aca2020-11-20 15:45:01 -08005539 srcs = [
Marat Dukhancf67ec62020-12-14 09:56:37 -08005540 "eval/f32-expm1minus-ulp.cc",
Marat Dukhana438aca2020-11-20 15:45:01 -08005541 "src/xnnpack/AlignedAllocator.h",
5542 ] + ACCURACY_EVAL_HDRS,
Marat Dukhande390d42020-11-29 19:32:18 -08005543 deps = ACCURACY_EVAL_DEPS + [
5544 ":bench_utils",
Marat Dukhan3a305212020-12-06 19:24:27 -08005545 "@cpuinfo",
Marat Dukhande390d42020-11-29 19:32:18 -08005546 ],
Marat Dukhana438aca2020-11-20 15:45:01 -08005547)
5548
5549xnnpack_benchmark(
Marat Dukhancf67ec62020-12-14 09:56:37 -08005550 name = "f32_extexp_ulp_eval",
Marat Dukhan98ba4412019-10-23 02:14:28 -07005551 srcs = [
Marat Dukhancf67ec62020-12-14 09:56:37 -08005552 "eval/f32-extexp-ulp.cc",
Marat Dukhan98ba4412019-10-23 02:14:28 -07005553 "src/xnnpack/AlignedAllocator.h",
5554 ] + ACCURACY_EVAL_HDRS,
Marat Dukhan3a305212020-12-06 19:24:27 -08005555 deps = ACCURACY_EVAL_DEPS + [
5556 ":bench_utils",
5557 "@cpuinfo",
5558 ],
Marat Dukhan98ba4412019-10-23 02:14:28 -07005559)
5560
Marat Dukhanf44f0222020-12-14 11:53:27 -08005561xnnpack_benchmark(
5562 name = "f32_sigmoid_ulp_eval",
5563 srcs = [
5564 "eval/f32-sigmoid-ulp.cc",
5565 "src/xnnpack/AlignedAllocator.h",
5566 ] + ACCURACY_EVAL_HDRS,
5567 deps = ACCURACY_EVAL_DEPS + [
5568 ":bench_utils",
5569 "@cpuinfo",
5570 ],
5571)
5572
5573xnnpack_benchmark(
5574 name = "f32_sqrt_ulp_eval",
5575 srcs = [
5576 "eval/f32-sqrt-ulp.cc",
5577 "src/xnnpack/AlignedAllocator.h",
5578 ] + ACCURACY_EVAL_HDRS,
5579 deps = ACCURACY_EVAL_DEPS + [
5580 ":bench_utils",
5581 "@cpuinfo",
5582 ],
5583)
5584
5585################### Accuracy verification for math functions ##################
5586
5587xnnpack_unit_test(
Marat Dukhanf7291fc2020-12-15 11:02:50 -08005588 name = "f32_exp_eval",
5589 srcs = [
5590 "eval/f32-exp.cc",
5591 "src/xnnpack/AlignedAllocator.h",
5592 "src/xnnpack/math-stubs.h",
5593 ] + MICROKERNEL_TEST_HDRS,
5594 automatic = False,
5595 deps = MICROKERNEL_TEST_DEPS,
5596)
5597
5598xnnpack_unit_test(
Marat Dukhanf44f0222020-12-14 11:53:27 -08005599 name = "f32_expm1minus_eval",
5600 srcs = [
5601 "eval/f32-expm1minus.cc",
5602 "src/xnnpack/AlignedAllocator.h",
5603 "src/xnnpack/math-stubs.h",
5604 ] + MICROKERNEL_TEST_HDRS,
5605 automatic = False,
5606 deps = MICROKERNEL_TEST_DEPS,
5607)
5608
Marat Dukhan8853b822020-05-07 12:19:01 -07005609xnnpack_unit_test(
Marat Dukhand28a5a22020-12-14 15:27:22 -08005610 name = "f32_expminus_eval",
5611 srcs = [
5612 "eval/f32-expminus.cc",
5613 "src/xnnpack/AlignedAllocator.h",
5614 "src/xnnpack/math-stubs.h",
5615 ] + MICROKERNEL_TEST_HDRS,
5616 automatic = False,
5617 deps = MICROKERNEL_TEST_DEPS,
5618)
5619
5620xnnpack_unit_test(
Marat Dukhan8853b822020-05-07 12:19:01 -07005621 name = "f32_roundne_eval",
5622 srcs = [
5623 "eval/f32-roundne.cc",
5624 "src/xnnpack/AlignedAllocator.h",
5625 "src/xnnpack/math-stubs.h",
5626 ] + MICROKERNEL_TEST_HDRS,
Marat Dukhan22eed3d2020-05-11 20:13:37 -07005627 automatic = False,
Marat Dukhan8853b822020-05-07 12:19:01 -07005628 deps = MICROKERNEL_TEST_DEPS,
5629)
5630
Marat Dukhan2dbb9442020-05-12 20:43:43 -07005631xnnpack_unit_test(
Marat Dukhanc9852ba2020-05-13 17:21:29 -07005632 name = "f32_roundd_eval",
5633 srcs = [
5634 "eval/f32-roundd.cc",
5635 "src/xnnpack/AlignedAllocator.h",
5636 "src/xnnpack/math-stubs.h",
5637 ] + MICROKERNEL_TEST_HDRS,
5638 automatic = False,
5639 deps = MICROKERNEL_TEST_DEPS,
5640)
5641
5642xnnpack_unit_test(
5643 name = "f32_roundu_eval",
5644 srcs = [
5645 "eval/f32-roundu.cc",
5646 "src/xnnpack/AlignedAllocator.h",
5647 "src/xnnpack/math-stubs.h",
5648 ] + MICROKERNEL_TEST_HDRS,
5649 automatic = False,
5650 deps = MICROKERNEL_TEST_DEPS,
5651)
5652
5653xnnpack_unit_test(
Marat Dukhan2dbb9442020-05-12 20:43:43 -07005654 name = "f32_roundz_eval",
5655 srcs = [
5656 "eval/f32-roundz.cc",
5657 "src/xnnpack/AlignedAllocator.h",
5658 "src/xnnpack/math-stubs.h",
5659 ] + MICROKERNEL_TEST_HDRS,
Marat Dukhanc9852ba2020-05-13 17:21:29 -07005660 automatic = False,
Marat Dukhan2dbb9442020-05-12 20:43:43 -07005661 deps = MICROKERNEL_TEST_DEPS,
5662)
5663
Marat Dukhan08c4a432019-10-03 09:29:21 -07005664######################### Unit tests for micro-kernels #########################
5665
5666xnnpack_unit_test(
Marat Dukhande06f492020-04-09 00:19:31 -07005667 name = "f16_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005668 srcs = [
Marat Dukhande06f492020-04-09 00:19:31 -07005669 "test/f16-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005670 "test/gemm-microkernel-tester.h",
5671 "src/xnnpack/AlignedAllocator.h",
5672 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005673 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005674)
5675
5676xnnpack_unit_test(
Marat Dukhan355ab432020-04-09 19:01:52 -07005677 name = "f16_spmm_minmax_test",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08005678 srcs = [
Marat Dukhan355ab432020-04-09 19:01:52 -07005679 "test/f16-spmm-minmax.cc",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08005680 "test/spmm-microkernel-tester.h",
5681 "src/xnnpack/AlignedAllocator.h",
5682 ] + MICROKERNEL_TEST_HDRS,
5683 deps = MICROKERNEL_TEST_DEPS,
5684)
5685
5686xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005687 name = "f32_argmaxpool_test",
5688 srcs = [
5689 "test/f32-argmaxpool.cc",
5690 "test/argmaxpool-microkernel-tester.h",
5691 "src/xnnpack/AlignedAllocator.h",
5692 ] + MICROKERNEL_TEST_HDRS,
5693 deps = MICROKERNEL_TEST_DEPS,
5694)
5695
5696xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07005697 name = "f32_avgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005698 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07005699 "test/f32-avgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005700 "test/avgpool-microkernel-tester.h",
5701 "src/xnnpack/AlignedAllocator.h",
5702 ] + MICROKERNEL_TEST_HDRS,
5703 deps = MICROKERNEL_TEST_DEPS,
5704)
5705
5706xnnpack_unit_test(
Marat Dukhan660fd192020-03-10 04:55:30 -07005707 name = "f32_ibilinear_test",
Marat Dukhan35dacfb2019-11-07 19:18:16 -08005708 srcs = [
Marat Dukhan660fd192020-03-10 04:55:30 -07005709 "test/f32-ibilinear.cc",
5710 "test/ibilinear-microkernel-tester.h",
Marat Dukhan35dacfb2019-11-07 19:18:16 -08005711 "src/xnnpack/AlignedAllocator.h",
5712 ] + MICROKERNEL_TEST_HDRS,
5713 deps = MICROKERNEL_TEST_DEPS,
5714)
5715
5716xnnpack_unit_test(
XNNPACK Team21432672020-10-19 19:58:48 -07005717 name = "f32_ibilinear_chw_test",
5718 srcs = [
5719 "test/f32-ibilinear-chw.cc",
XNNPACK Team6be46b22020-10-22 23:34:54 -07005720 "test/ibilinear-microkernel-tester.h",
XNNPACK Team21432672020-10-19 19:58:48 -07005721 "src/xnnpack/AlignedAllocator.h",
5722 ] + MICROKERNEL_TEST_HDRS,
5723 deps = MICROKERNEL_TEST_DEPS,
5724)
5725
5726xnnpack_unit_test(
Frank Barchardb1966592020-05-12 13:47:06 -07005727 name = "f16_clamp_test",
5728 srcs = [
5729 "test/f16-clamp.cc",
5730 "test/clamp-microkernel-tester.h",
5731 ] + MICROKERNEL_TEST_HDRS,
5732 deps = MICROKERNEL_TEST_DEPS,
5733)
5734
5735xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005736 name = "f32_clamp_test",
5737 srcs = [
5738 "test/f32-clamp.cc",
5739 "test/clamp-microkernel-tester.h",
5740 ] + MICROKERNEL_TEST_HDRS,
5741 deps = MICROKERNEL_TEST_DEPS,
5742)
5743
5744xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07005745 name = "f32_igemm_test",
5746 srcs = [
5747 "test/f32-igemm.cc",
5748 "test/gemm-microkernel-tester.h",
5749 "src/xnnpack/AlignedAllocator.h",
5750 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005751 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan163a7e62020-04-09 04:19:26 -07005752)
5753
5754xnnpack_unit_test(
Marat Dukhan467f6362020-05-22 23:21:55 -07005755 name = "f32_igemm_relu_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005756 srcs = [
Marat Dukhan467f6362020-05-22 23:21:55 -07005757 "test/f32-igemm-relu.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005758 "test/gemm-microkernel-tester.h",
5759 "src/xnnpack/AlignedAllocator.h",
5760 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005761 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005762)
5763
5764xnnpack_unit_test(
Marat Dukhane207b7b2020-05-28 16:27:42 -07005765 name = "f32_igemm_minmax_test",
5766 srcs = [
5767 "test/f32-igemm-minmax.cc",
5768 "test/gemm-microkernel-tester.h",
5769 "src/xnnpack/AlignedAllocator.h",
5770 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005771 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhane207b7b2020-05-28 16:27:42 -07005772)
5773
5774xnnpack_unit_test(
Frank Barchardb0e4fae2020-05-04 15:27:51 -07005775 name = "f16_igemm_minmax_test",
5776 srcs = [
5777 "test/f16-igemm-minmax.cc",
5778 "test/gemm-microkernel-tester.h",
5779 "src/xnnpack/AlignedAllocator.h",
5780 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005781 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Frank Barchardb0e4fae2020-05-04 15:27:51 -07005782)
5783
5784xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005785 name = "f32_conv_hwc_test",
5786 srcs = [
5787 "test/f32-conv-hwc.cc",
5788 "test/conv-hwc-microkernel-tester.h",
5789 "src/xnnpack/AlignedAllocator.h",
5790 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005791 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005792)
5793
5794xnnpack_unit_test(
Marat Dukhan1f29b802020-05-15 23:46:39 -07005795 name = "f32_conv_hwc2chw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005796 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07005797 "test/f32-conv-hwc2chw.cc",
5798 "test/conv-hwc2chw-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005799 "src/xnnpack/AlignedAllocator.h",
5800 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005801 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005802)
5803
5804xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07005805 name = "f32_dwconv_test",
5806 srcs = [
5807 "test/f32-dwconv.cc",
5808 "test/dwconv-microkernel-tester.h",
5809 "src/xnnpack/AlignedAllocator.h",
5810 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005811 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan163a7e62020-04-09 04:19:26 -07005812)
5813
5814xnnpack_unit_test(
Frank Barchard5a599a62020-06-04 20:12:44 -07005815 name = "f16_dwconv_minmax_test",
5816 srcs = [
5817 "test/f16-dwconv-minmax.cc",
5818 "test/dwconv-microkernel-tester.h",
5819 "src/xnnpack/AlignedAllocator.h",
5820 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005821 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Frank Barchard5a599a62020-06-04 20:12:44 -07005822)
5823
5824xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07005825 name = "f32_dwconv_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005826 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07005827 "test/f32-dwconv-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005828 "test/dwconv-microkernel-tester.h",
5829 "src/xnnpack/AlignedAllocator.h",
5830 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005831 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005832)
5833
5834xnnpack_unit_test(
Marat Dukhanbf715f92020-10-23 20:17:00 -07005835 name = "f32_dwconv2d_chw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005836 srcs = [
Marat Dukhanbf715f92020-10-23 20:17:00 -07005837 "test/f32-dwconv2d-chw.cc",
5838 "test/dwconv2d-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005839 "src/xnnpack/AlignedAllocator.h",
5840 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005841 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005842)
5843
5844xnnpack_unit_test(
Frank Barchard0bb49a72020-06-04 11:35:11 -07005845 name = "f16_gavgpool_minmax_test",
5846 srcs = [
5847 "test/f16-gavgpool-minmax.cc",
5848 "test/gavgpool-microkernel-tester.h",
5849 "src/xnnpack/AlignedAllocator.h",
5850 ] + MICROKERNEL_TEST_HDRS,
5851 deps = MICROKERNEL_TEST_DEPS,
5852)
5853
5854xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07005855 name = "f32_gavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005856 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07005857 "test/f32-gavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005858 "test/gavgpool-microkernel-tester.h",
5859 "src/xnnpack/AlignedAllocator.h",
5860 ] + MICROKERNEL_TEST_HDRS,
5861 deps = MICROKERNEL_TEST_DEPS,
5862)
5863
5864xnnpack_unit_test(
Marat Dukhan1f29b802020-05-15 23:46:39 -07005865 name = "f32_gavgpool_cw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005866 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07005867 "test/f32-gavgpool-cw.cc",
5868 "test/gavgpool-cw-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005869 "src/xnnpack/AlignedAllocator.h",
5870 ] + MICROKERNEL_TEST_HDRS,
5871 deps = MICROKERNEL_TEST_DEPS,
5872)
5873
5874xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07005875 name = "f32_gemm_test",
5876 srcs = [
5877 "test/f32-gemm.cc",
5878 "test/gemm-microkernel-tester.h",
5879 "src/xnnpack/AlignedAllocator.h",
5880 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005881 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan163a7e62020-04-09 04:19:26 -07005882)
5883
5884xnnpack_unit_test(
Marat Dukhan467f6362020-05-22 23:21:55 -07005885 name = "f32_gemm_relu_test",
5886 srcs = [
5887 "test/f32-gemm-relu.cc",
5888 "test/gemm-microkernel-tester.h",
5889 "src/xnnpack/AlignedAllocator.h",
5890 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005891 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan467f6362020-05-22 23:21:55 -07005892)
5893
5894xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07005895 name = "f32_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005896 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07005897 "test/f32-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005898 "test/gemm-microkernel-tester.h",
5899 "src/xnnpack/AlignedAllocator.h",
5900 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005901 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005902)
5903
5904xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07005905 name = "f32_gemminc_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005906 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07005907 "test/f32-gemminc-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005908 "test/gemm-microkernel-tester.h",
5909 "src/xnnpack/AlignedAllocator.h",
5910 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005911 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005912)
5913
5914xnnpack_unit_test(
Frank Barchardb1966592020-05-12 13:47:06 -07005915 name = "f16_hswish_test",
5916 srcs = [
5917 "test/f16-hswish.cc",
5918 "test/hswish-microkernel-tester.h",
5919 ] + MICROKERNEL_TEST_HDRS,
5920 deps = MICROKERNEL_TEST_DEPS,
5921)
5922
5923xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005924 name = "f32_hswish_test",
5925 srcs = [
5926 "test/f32-hswish.cc",
5927 "test/hswish-microkernel-tester.h",
5928 ] + MICROKERNEL_TEST_HDRS,
5929 deps = MICROKERNEL_TEST_DEPS,
5930)
5931
5932xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07005933 name = "f32_maxpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005934 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07005935 "test/f32-maxpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005936 "test/maxpool-microkernel-tester.h",
5937 ] + MICROKERNEL_TEST_HDRS,
5938 deps = MICROKERNEL_TEST_DEPS,
5939)
5940
5941xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07005942 name = "f32_pavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005943 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07005944 "test/f32-pavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005945 "test/avgpool-microkernel-tester.h",
5946 "src/xnnpack/AlignedAllocator.h",
5947 ] + MICROKERNEL_TEST_HDRS,
5948 deps = MICROKERNEL_TEST_DEPS,
5949)
5950
5951xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07005952 name = "f32_ppmm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005953 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07005954 "test/f32-ppmm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005955 "test/gemm-microkernel-tester.h",
5956 "src/xnnpack/AlignedAllocator.h",
5957 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005958 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005959)
5960
5961xnnpack_unit_test(
Frank Barchardb1966592020-05-12 13:47:06 -07005962 name = "f16_prelu_test",
5963 srcs = [
5964 "test/f16-prelu.cc",
5965 "test/prelu-microkernel-tester.h",
5966 "src/xnnpack/AlignedAllocator.h",
5967 ] + MICROKERNEL_TEST_HDRS,
5968 deps = MICROKERNEL_TEST_DEPS,
5969)
5970
5971xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005972 name = "f32_prelu_test",
5973 srcs = [
5974 "test/f32-prelu.cc",
5975 "test/prelu-microkernel-tester.h",
5976 "src/xnnpack/AlignedAllocator.h",
5977 ] + MICROKERNEL_TEST_HDRS,
5978 deps = MICROKERNEL_TEST_DEPS,
5979)
5980
5981xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07005982 name = "f32_raddexpminusmax_test",
5983 srcs = [
5984 "test/f32-raddexpminusmax.cc",
5985 "test/raddexpminusmax-microkernel-tester.h",
5986 ] + MICROKERNEL_TEST_HDRS,
5987 deps = MICROKERNEL_TEST_DEPS,
5988)
5989
5990xnnpack_unit_test(
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07005991 name = "f32_raddextexp_test",
5992 srcs = [
5993 "test/f32-raddextexp.cc",
5994 "test/raddextexp-microkernel-tester.h",
5995 ] + MICROKERNEL_TEST_HDRS,
5996 deps = MICROKERNEL_TEST_DEPS,
5997)
5998
5999xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07006000 name = "f32_raddstoreexpminusmax_test",
6001 srcs = [
6002 "test/f32-raddstoreexpminusmax.cc",
6003 "test/raddstoreexpminusmax-microkernel-tester.h",
6004 ] + MICROKERNEL_TEST_HDRS,
6005 deps = MICROKERNEL_TEST_DEPS,
6006)
6007
6008xnnpack_unit_test(
Frank Barchardfb158e22020-07-15 16:10:10 -07006009 name = "f32_relu_test",
6010 srcs = [
6011 "test/f32-relu.cc",
6012 "test/vunary-microkernel-tester.h",
6013 ] + MICROKERNEL_TEST_HDRS,
6014 deps = MICROKERNEL_TEST_DEPS,
6015)
6016
6017xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07006018 name = "f32_rmax_test",
6019 srcs = [
6020 "test/f32-rmax.cc",
6021 "test/rmax-microkernel-tester.h",
6022 ] + MICROKERNEL_TEST_HDRS,
6023 deps = MICROKERNEL_TEST_DEPS,
6024)
6025
6026xnnpack_unit_test(
Marat Dukhan346a9e52019-11-15 09:06:30 -08006027 name = "f32_sigmoid_test",
6028 srcs = [
6029 "test/f32-sigmoid.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08006030 "test/vunary-microkernel-tester.h",
Marat Dukhan346a9e52019-11-15 09:06:30 -08006031 ] + MICROKERNEL_TEST_HDRS,
6032 deps = MICROKERNEL_TEST_DEPS,
6033)
6034
6035xnnpack_unit_test(
Marat Dukhan355ab432020-04-09 19:01:52 -07006036 name = "f32_spmm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006037 srcs = [
Marat Dukhan355ab432020-04-09 19:01:52 -07006038 "test/f32-spmm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006039 "test/spmm-microkernel-tester.h",
6040 "src/xnnpack/AlignedAllocator.h",
6041 ] + MICROKERNEL_TEST_HDRS,
6042 deps = MICROKERNEL_TEST_DEPS,
6043)
6044
6045xnnpack_unit_test(
Frank Barchardd793f6c2020-05-08 13:37:43 -07006046 name = "f16_vadd_minmax_test",
6047 srcs = [
6048 "test/f16-vadd-minmax.cc",
6049 "test/vbinary-microkernel-tester.h",
6050 ] + MICROKERNEL_TEST_HDRS,
6051 deps = MICROKERNEL_TEST_DEPS,
6052)
6053
6054xnnpack_unit_test(
6055 name = "f16_vaddc_minmax_test",
6056 srcs = [
6057 "test/f16-vaddc-minmax.cc",
6058 "test/vbinaryc-microkernel-tester.h",
6059 ] + MICROKERNEL_TEST_HDRS,
6060 deps = MICROKERNEL_TEST_DEPS,
6061)
6062
6063xnnpack_unit_test(
6064 name = "f16_vdiv_minmax_test",
6065 srcs = [
6066 "test/f16-vdiv-minmax.cc",
6067 "test/vbinary-microkernel-tester.h",
6068 ] + MICROKERNEL_TEST_HDRS,
6069 deps = MICROKERNEL_TEST_DEPS,
6070)
6071
6072xnnpack_unit_test(
6073 name = "f16_vdivc_minmax_test",
6074 srcs = [
6075 "test/f16-vdivc-minmax.cc",
6076 "test/vbinaryc-microkernel-tester.h",
6077 ] + MICROKERNEL_TEST_HDRS,
6078 deps = MICROKERNEL_TEST_DEPS,
6079)
6080
6081xnnpack_unit_test(
6082 name = "f16_vrdivc_minmax_test",
6083 srcs = [
6084 "test/f16-vrdivc-minmax.cc",
6085 "test/vbinaryc-microkernel-tester.h",
6086 ] + MICROKERNEL_TEST_HDRS,
6087 deps = MICROKERNEL_TEST_DEPS,
6088)
6089
6090xnnpack_unit_test(
6091 name = "f16_vmax_test",
6092 srcs = [
6093 "test/f16-vmax.cc",
6094 "test/vbinary-microkernel-tester.h",
6095 ] + MICROKERNEL_TEST_HDRS,
6096 deps = MICROKERNEL_TEST_DEPS,
6097)
6098
6099xnnpack_unit_test(
6100 name = "f16_vmaxc_test",
6101 srcs = [
6102 "test/f16-vmaxc.cc",
6103 "test/vbinaryc-microkernel-tester.h",
6104 ] + MICROKERNEL_TEST_HDRS,
6105 deps = MICROKERNEL_TEST_DEPS,
6106)
6107
6108xnnpack_unit_test(
6109 name = "f16_vmin_test",
6110 srcs = [
6111 "test/f16-vmin.cc",
6112 "test/vbinary-microkernel-tester.h",
6113 ] + MICROKERNEL_TEST_HDRS,
6114 deps = MICROKERNEL_TEST_DEPS,
6115)
6116
6117xnnpack_unit_test(
6118 name = "f16_vminc_test",
6119 srcs = [
6120 "test/f16-vminc.cc",
6121 "test/vbinaryc-microkernel-tester.h",
6122 ] + MICROKERNEL_TEST_HDRS,
6123 deps = MICROKERNEL_TEST_DEPS,
6124)
6125
6126xnnpack_unit_test(
6127 name = "f16_vmul_minmax_test",
6128 srcs = [
6129 "test/f16-vmul-minmax.cc",
6130 "test/vbinary-microkernel-tester.h",
6131 ] + MICROKERNEL_TEST_HDRS,
6132 deps = MICROKERNEL_TEST_DEPS,
6133)
6134
6135xnnpack_unit_test(
6136 name = "f16_vmulc_minmax_test",
6137 srcs = [
6138 "test/f16-vmulc-minmax.cc",
6139 "test/vbinaryc-microkernel-tester.h",
6140 ] + MICROKERNEL_TEST_HDRS,
6141 deps = MICROKERNEL_TEST_DEPS,
6142)
6143
6144xnnpack_unit_test(
6145 name = "f16_vsub_minmax_test",
6146 srcs = [
6147 "test/f16-vsub-minmax.cc",
6148 "test/vbinary-microkernel-tester.h",
6149 ] + MICROKERNEL_TEST_HDRS,
6150 deps = MICROKERNEL_TEST_DEPS,
6151)
6152
6153xnnpack_unit_test(
6154 name = "f16_vsubc_minmax_test",
6155 srcs = [
6156 "test/f16-vsubc-minmax.cc",
6157 "test/vbinaryc-microkernel-tester.h",
6158 ] + MICROKERNEL_TEST_HDRS,
6159 deps = MICROKERNEL_TEST_DEPS,
6160)
6161
6162xnnpack_unit_test(
6163 name = "f16_vrsubc_minmax_test",
6164 srcs = [
6165 "test/f16-vrsubc-minmax.cc",
6166 "test/vbinaryc-microkernel-tester.h",
6167 ] + MICROKERNEL_TEST_HDRS,
6168 deps = MICROKERNEL_TEST_DEPS,
6169)
6170
6171xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07006172 name = "f32_vabs_test",
6173 srcs = [
6174 "test/f32-vabs.cc",
6175 "test/vunary-microkernel-tester.h",
6176 ] + MICROKERNEL_TEST_HDRS,
6177 deps = MICROKERNEL_TEST_DEPS,
6178)
6179
6180xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07006181 name = "f32_vadd_test",
6182 srcs = [
6183 "test/f32-vadd.cc",
6184 "test/vbinary-microkernel-tester.h",
6185 ] + MICROKERNEL_TEST_HDRS,
6186 deps = MICROKERNEL_TEST_DEPS,
6187)
6188
6189xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006190 name = "f32_vadd_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006191 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006192 "test/f32-vadd-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08006193 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08006194 ] + MICROKERNEL_TEST_HDRS,
6195 deps = MICROKERNEL_TEST_DEPS,
6196)
6197
6198xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07006199 name = "f32_vadd_relu_test",
6200 srcs = [
6201 "test/f32-vadd-relu.cc",
6202 "test/vbinary-microkernel-tester.h",
6203 ] + MICROKERNEL_TEST_HDRS,
6204 deps = MICROKERNEL_TEST_DEPS,
6205)
6206
6207xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07006208 name = "f32_vaddc_test",
6209 srcs = [
6210 "test/f32-vaddc.cc",
6211 "test/vbinaryc-microkernel-tester.h",
6212 ] + MICROKERNEL_TEST_HDRS,
6213 deps = MICROKERNEL_TEST_DEPS,
6214)
6215
6216xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006217 name = "f32_vaddc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08006218 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006219 "test/f32-vaddc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08006220 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006221 ] + MICROKERNEL_TEST_HDRS,
6222 deps = MICROKERNEL_TEST_DEPS,
6223)
6224
6225xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07006226 name = "f32_vaddc_relu_test",
6227 srcs = [
6228 "test/f32-vaddc-relu.cc",
6229 "test/vbinaryc-microkernel-tester.h",
6230 ] + MICROKERNEL_TEST_HDRS,
6231 deps = MICROKERNEL_TEST_DEPS,
6232)
6233
6234xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07006235 name = "f32_vdiv_test",
6236 srcs = [
6237 "test/f32-vdiv.cc",
6238 "test/vbinary-microkernel-tester.h",
6239 ] + MICROKERNEL_TEST_HDRS,
6240 deps = MICROKERNEL_TEST_DEPS,
6241)
6242
6243xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006244 name = "f32_vdiv_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08006245 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006246 "test/f32-vdiv-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08006247 "test/vbinary-microkernel-tester.h",
6248 ] + MICROKERNEL_TEST_HDRS,
6249 deps = MICROKERNEL_TEST_DEPS,
6250)
6251
6252xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07006253 name = "f32_vdiv_relu_test",
6254 srcs = [
6255 "test/f32-vdiv-relu.cc",
6256 "test/vbinary-microkernel-tester.h",
6257 ] + MICROKERNEL_TEST_HDRS,
6258 deps = MICROKERNEL_TEST_DEPS,
6259)
6260
6261xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07006262 name = "f32_vdivc_test",
6263 srcs = [
6264 "test/f32-vdivc.cc",
6265 "test/vbinaryc-microkernel-tester.h",
6266 ] + MICROKERNEL_TEST_HDRS,
6267 deps = MICROKERNEL_TEST_DEPS,
6268)
6269
6270xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006271 name = "f32_vdivc_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08006272 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006273 "test/f32-vdivc-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08006274 "test/vbinaryc-microkernel-tester.h",
6275 ] + MICROKERNEL_TEST_HDRS,
6276 deps = MICROKERNEL_TEST_DEPS,
6277)
6278
6279xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07006280 name = "f32_vdivc_relu_test",
6281 srcs = [
6282 "test/f32-vdivc-relu.cc",
6283 "test/vbinaryc-microkernel-tester.h",
6284 ] + MICROKERNEL_TEST_HDRS,
6285 deps = MICROKERNEL_TEST_DEPS,
6286)
6287
6288xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07006289 name = "f32_vrdivc_test",
6290 srcs = [
6291 "test/f32-vrdivc.cc",
6292 "test/vbinaryc-microkernel-tester.h",
6293 ] + MICROKERNEL_TEST_HDRS,
6294 deps = MICROKERNEL_TEST_DEPS,
6295)
6296
6297xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006298 name = "f32_vrdivc_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08006299 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006300 "test/f32-vrdivc-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08006301 "test/vbinaryc-microkernel-tester.h",
6302 ] + MICROKERNEL_TEST_HDRS,
6303 deps = MICROKERNEL_TEST_DEPS,
6304)
6305
6306xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07006307 name = "f32_vrdivc_relu_test",
6308 srcs = [
6309 "test/f32-vrdivc-relu.cc",
6310 "test/vbinaryc-microkernel-tester.h",
6311 ] + MICROKERNEL_TEST_HDRS,
6312 deps = MICROKERNEL_TEST_DEPS,
6313)
6314
6315xnnpack_unit_test(
Marat Dukhaned6baaf2020-12-01 15:07:08 -08006316 name = "f32_velu_test",
6317 srcs = [
6318 "test/f32-velu.cc",
6319 "test/vunary-microkernel-tester.h",
6320 ] + MICROKERNEL_TEST_HDRS,
6321 deps = MICROKERNEL_TEST_DEPS,
6322)
6323
6324xnnpack_unit_test(
Marat Dukhan403b7d42019-12-05 12:49:11 -08006325 name = "f32_vmax_test",
6326 srcs = [
6327 "test/f32-vmax.cc",
6328 "test/vbinary-microkernel-tester.h",
6329 ] + MICROKERNEL_TEST_HDRS,
6330 deps = MICROKERNEL_TEST_DEPS,
6331)
6332
6333xnnpack_unit_test(
6334 name = "f32_vmaxc_test",
6335 srcs = [
6336 "test/f32-vmaxc.cc",
6337 "test/vbinaryc-microkernel-tester.h",
6338 ] + MICROKERNEL_TEST_HDRS,
6339 deps = MICROKERNEL_TEST_DEPS,
6340)
6341
6342xnnpack_unit_test(
6343 name = "f32_vmin_test",
6344 srcs = [
6345 "test/f32-vmin.cc",
6346 "test/vbinary-microkernel-tester.h",
6347 ] + MICROKERNEL_TEST_HDRS,
6348 deps = MICROKERNEL_TEST_DEPS,
6349)
6350
6351xnnpack_unit_test(
6352 name = "f32_vminc_test",
6353 srcs = [
6354 "test/f32-vminc.cc",
6355 "test/vbinaryc-microkernel-tester.h",
6356 ] + MICROKERNEL_TEST_HDRS,
6357 deps = MICROKERNEL_TEST_DEPS,
6358)
6359
6360xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07006361 name = "f32_vmul_test",
6362 srcs = [
6363 "test/f32-vmul.cc",
6364 "test/vbinary-microkernel-tester.h",
6365 ] + MICROKERNEL_TEST_HDRS,
6366 deps = MICROKERNEL_TEST_DEPS,
6367)
6368
6369xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006370 name = "f32_vmul_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006371 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006372 "test/f32-vmul-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08006373 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08006374 ] + MICROKERNEL_TEST_HDRS,
6375 deps = MICROKERNEL_TEST_DEPS,
6376)
6377
6378xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07006379 name = "f32_vmul_relu_test",
6380 srcs = [
6381 "test/f32-vmul-relu.cc",
6382 "test/vbinary-microkernel-tester.h",
6383 ] + MICROKERNEL_TEST_HDRS,
6384 deps = MICROKERNEL_TEST_DEPS,
6385)
6386
6387xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07006388 name = "f32_vmulc_test",
6389 srcs = [
6390 "test/f32-vmulc.cc",
6391 "test/vbinaryc-microkernel-tester.h",
6392 ] + MICROKERNEL_TEST_HDRS,
6393 deps = MICROKERNEL_TEST_DEPS,
6394)
6395
6396xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006397 name = "f32_vmulc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08006398 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006399 "test/f32-vmulc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08006400 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006401 ] + MICROKERNEL_TEST_HDRS,
6402 deps = MICROKERNEL_TEST_DEPS,
6403)
6404
6405xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07006406 name = "f32_vmulc_relu_test",
6407 srcs = [
6408 "test/f32-vmulc-relu.cc",
6409 "test/vbinaryc-microkernel-tester.h",
6410 ] + MICROKERNEL_TEST_HDRS,
6411 deps = MICROKERNEL_TEST_DEPS,
6412)
6413
6414xnnpack_unit_test(
Frank Barchard2a1049c2020-06-03 02:31:27 -07006415 name = "f16_vmulcaddc_minmax_test",
6416 srcs = [
6417 "test/f16-vmulcaddc-minmax.cc",
6418 "test/vmulcaddc-microkernel-tester.h",
6419 "src/xnnpack/AlignedAllocator.h",
6420 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07006421 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Frank Barchard2a1049c2020-06-03 02:31:27 -07006422)
6423
6424xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07006425 name = "f32_vmulcaddc_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006426 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07006427 "test/f32-vmulcaddc-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006428 "test/vmulcaddc-microkernel-tester.h",
6429 "src/xnnpack/AlignedAllocator.h",
6430 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07006431 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07006432)
6433
6434xnnpack_unit_test(
Marat Dukhan8cc7efe2020-06-10 16:24:27 -07006435 name = "f32_vlrelu_test",
6436 srcs = [
6437 "test/f32-vlrelu.cc",
6438 "test/vunary-microkernel-tester.h",
6439 ] + MICROKERNEL_TEST_HDRS,
6440 deps = MICROKERNEL_TEST_DEPS,
6441)
6442
6443xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07006444 name = "f32_vneg_test",
6445 srcs = [
6446 "test/f32-vneg.cc",
6447 "test/vunary-microkernel-tester.h",
6448 ] + MICROKERNEL_TEST_HDRS,
6449 deps = MICROKERNEL_TEST_DEPS,
6450)
6451
6452xnnpack_unit_test(
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07006453 name = "f32_vrndne_test",
6454 srcs = [
6455 "test/f32-vrndne.cc",
6456 "test/vunary-microkernel-tester.h",
6457 ] + MICROKERNEL_TEST_HDRS,
6458 deps = MICROKERNEL_TEST_DEPS,
6459)
6460
6461xnnpack_unit_test(
6462 name = "f32_vrndz_test",
6463 srcs = [
6464 "test/f32-vrndz.cc",
6465 "test/vunary-microkernel-tester.h",
6466 ] + MICROKERNEL_TEST_HDRS,
6467 deps = MICROKERNEL_TEST_DEPS,
6468)
6469
6470xnnpack_unit_test(
6471 name = "f32_vrndu_test",
6472 srcs = [
6473 "test/f32-vrndu.cc",
6474 "test/vunary-microkernel-tester.h",
6475 ] + MICROKERNEL_TEST_HDRS,
6476 deps = MICROKERNEL_TEST_DEPS,
6477)
6478
6479xnnpack_unit_test(
6480 name = "f32_vrndd_test",
6481 srcs = [
6482 "test/f32-vrndd.cc",
6483 "test/vunary-microkernel-tester.h",
6484 ] + MICROKERNEL_TEST_HDRS,
6485 deps = MICROKERNEL_TEST_DEPS,
6486)
6487
6488xnnpack_unit_test(
Marat Dukhan05ac8e32019-10-21 15:39:33 -07006489 name = "f32_vscale_test",
6490 srcs = [
6491 "test/f32-vscale.cc",
6492 "test/vscale-microkernel-tester.h",
6493 ] + MICROKERNEL_TEST_HDRS,
6494 deps = MICROKERNEL_TEST_DEPS,
6495)
6496
6497xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07006498 name = "f32_vscaleexpminusmax_test",
6499 srcs = [
6500 "test/f32-vscaleexpminusmax.cc",
6501 "test/vscaleexpminusmax-microkernel-tester.h",
6502 ] + MICROKERNEL_TEST_HDRS,
6503 deps = MICROKERNEL_TEST_DEPS,
6504)
6505
6506xnnpack_unit_test(
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07006507 name = "f32_vscaleextexp_test",
6508 srcs = [
6509 "test/f32-vscaleextexp.cc",
6510 "test/vscaleextexp-microkernel-tester.h",
6511 ] + MICROKERNEL_TEST_HDRS,
6512 deps = MICROKERNEL_TEST_DEPS,
6513)
6514
6515xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07006516 name = "f32_vsqr_test",
6517 srcs = [
6518 "test/f32-vsqr.cc",
6519 "test/vunary-microkernel-tester.h",
6520 ] + MICROKERNEL_TEST_HDRS,
6521 deps = MICROKERNEL_TEST_DEPS,
6522)
6523
6524xnnpack_unit_test(
Marat Dukhan13bafb02020-06-05 00:43:11 -07006525 name = "f32_vsqrdiff_test",
6526 srcs = [
6527 "test/f32-vsqrdiff.cc",
6528 "test/vbinary-microkernel-tester.h",
6529 ] + MICROKERNEL_TEST_HDRS,
6530 deps = MICROKERNEL_TEST_DEPS,
6531)
6532
6533xnnpack_unit_test(
6534 name = "f32_vsqrdiffc_test",
6535 srcs = [
6536 "test/f32-vsqrdiffc.cc",
6537 "test/vbinaryc-microkernel-tester.h",
6538 ] + MICROKERNEL_TEST_HDRS,
6539 deps = MICROKERNEL_TEST_DEPS,
6540)
6541
6542xnnpack_unit_test(
Marat Dukhanf4db2f32020-06-30 10:55:30 -07006543 name = "f32_vsqrt_test",
6544 srcs = [
6545 "test/f32-vsqrt.cc",
6546 "test/vunary-microkernel-tester.h",
6547 ] + MICROKERNEL_TEST_HDRS,
6548 deps = MICROKERNEL_TEST_DEPS,
6549)
6550
6551xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07006552 name = "f32_vsub_test",
6553 srcs = [
6554 "test/f32-vsub.cc",
6555 "test/vbinary-microkernel-tester.h",
6556 ] + MICROKERNEL_TEST_HDRS,
6557 deps = MICROKERNEL_TEST_DEPS,
6558)
6559
6560xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006561 name = "f32_vsub_minmax_test",
Marat Dukhan97579532019-10-18 16:40:39 -07006562 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006563 "test/f32-vsub-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08006564 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08006565 ] + MICROKERNEL_TEST_HDRS,
6566 deps = MICROKERNEL_TEST_DEPS,
6567)
6568
6569xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07006570 name = "f32_vsub_relu_test",
6571 srcs = [
6572 "test/f32-vsub-relu.cc",
6573 "test/vbinary-microkernel-tester.h",
6574 ] + MICROKERNEL_TEST_HDRS,
6575 deps = MICROKERNEL_TEST_DEPS,
6576)
6577
6578xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07006579 name = "f32_vsubc_test",
6580 srcs = [
6581 "test/f32-vsubc.cc",
6582 "test/vbinaryc-microkernel-tester.h",
6583 ] + MICROKERNEL_TEST_HDRS,
6584 deps = MICROKERNEL_TEST_DEPS,
6585)
6586
6587xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006588 name = "f32_vsubc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08006589 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006590 "test/f32-vsubc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08006591 "test/vbinaryc-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08006592 ] + MICROKERNEL_TEST_HDRS,
6593 deps = MICROKERNEL_TEST_DEPS,
6594)
6595
6596xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07006597 name = "f32_vsubc_relu_test",
6598 srcs = [
6599 "test/f32-vsubc-relu.cc",
6600 "test/vbinaryc-microkernel-tester.h",
6601 ] + MICROKERNEL_TEST_HDRS,
6602 deps = MICROKERNEL_TEST_DEPS,
6603)
6604
6605xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07006606 name = "f32_vrsubc_test",
6607 srcs = [
6608 "test/f32-vrsubc.cc",
6609 "test/vbinaryc-microkernel-tester.h",
6610 ] + MICROKERNEL_TEST_HDRS,
6611 deps = MICROKERNEL_TEST_DEPS,
6612)
6613
6614xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006615 name = "f32_vrsubc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08006616 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006617 "test/f32-vrsubc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08006618 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan97579532019-10-18 16:40:39 -07006619 ] + MICROKERNEL_TEST_HDRS,
6620 deps = MICROKERNEL_TEST_DEPS,
6621)
6622
6623xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07006624 name = "f32_vrsubc_relu_test",
6625 srcs = [
6626 "test/f32-vrsubc-relu.cc",
6627 "test/vbinaryc-microkernel-tester.h",
6628 ] + MICROKERNEL_TEST_HDRS,
6629 deps = MICROKERNEL_TEST_DEPS,
6630)
6631
6632xnnpack_unit_test(
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07006633 name = "qs8_dwconv_minmax_test",
6634 srcs = [
6635 "test/qs8-dwconv-minmax.cc",
6636 "test/dwconv-microkernel-tester.h",
6637 "src/xnnpack/AlignedAllocator.h",
6638 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
6639 deps = MICROKERNEL_TEST_DEPS + [":packing"],
6640)
6641
6642xnnpack_unit_test(
Marat Dukhan4ed53f42020-08-06 01:12:55 -07006643 name = "qs8_gavgpool_minmax_test",
6644 srcs = [
6645 "test/qs8-gavgpool-minmax.cc",
6646 "test/gavgpool-microkernel-tester.h",
6647 "src/xnnpack/AlignedAllocator.h",
6648 ] + MICROKERNEL_TEST_HDRS,
6649 deps = MICROKERNEL_TEST_DEPS,
6650)
6651
6652xnnpack_unit_test(
Marat Dukhan595e1702020-07-31 10:12:52 -07006653 name = "qs8_gemm_minmax_test",
6654 srcs = [
6655 "test/qs8-gemm-minmax.cc",
6656 "test/gemm-microkernel-tester.h",
6657 "src/xnnpack/AlignedAllocator.h",
6658 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
6659 deps = MICROKERNEL_TEST_DEPS + [":packing"],
6660)
6661
6662xnnpack_unit_test(
Marat Dukhanf9480682020-07-31 14:50:24 -07006663 name = "qs8_igemm_minmax_test",
6664 srcs = [
6665 "test/qs8-igemm-minmax.cc",
6666 "test/gemm-microkernel-tester.h",
6667 "src/xnnpack/AlignedAllocator.h",
6668 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
6669 deps = MICROKERNEL_TEST_DEPS + [":packing"],
6670)
6671
6672xnnpack_unit_test(
6673 name = "qs8_requantization_test",
6674 srcs = [
6675 "src/xnnpack/requantization-stubs.h",
6676 "test/qs8-requantization.cc",
6677 "test/requantization-tester.h",
6678 ] + MICROKERNEL_TEST_HDRS,
6679 deps = MICROKERNEL_TEST_DEPS,
6680)
6681
6682xnnpack_unit_test(
Marat Dukhand9f3ad42020-08-10 12:30:58 -07006683 name = "qs8_vadd_minmax_test",
6684 srcs = [
6685 "test/qs8-vadd-minmax.cc",
6686 "test/vadd-microkernel-tester.h",
6687 ] + MICROKERNEL_TEST_HDRS,
6688 deps = MICROKERNEL_TEST_DEPS,
6689)
6690
6691xnnpack_unit_test(
Marat Dukhan0270d9f2020-08-11 00:56:46 -07006692 name = "qs8_vaddc_minmax_test",
6693 srcs = [
6694 "test/qs8-vaddc-minmax.cc",
6695 "test/vaddc-microkernel-tester.h",
6696 ] + MICROKERNEL_TEST_HDRS,
6697 deps = MICROKERNEL_TEST_DEPS,
6698)
6699
6700xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07006701 name = "qu8_avgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006702 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07006703 "test/qu8-avgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006704 "test/avgpool-microkernel-tester.h",
6705 "src/xnnpack/AlignedAllocator.h",
6706 ] + MICROKERNEL_TEST_HDRS,
6707 deps = MICROKERNEL_TEST_DEPS,
6708)
6709
6710xnnpack_unit_test(
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07006711 name = "qu8_dwconv_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006712 srcs = [
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07006713 "test/qu8-dwconv-minmax.cc",
6714 "test/dwconv-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006715 "src/xnnpack/AlignedAllocator.h",
6716 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07006717 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07006718)
6719
6720xnnpack_unit_test(
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07006721 name = "qu8_igemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006722 srcs = [
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07006723 "test/qu8-igemm-minmax.cc",
6724 "test/gemm-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006725 "src/xnnpack/AlignedAllocator.h",
6726 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07006727 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07006728)
6729
6730xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07006731 name = "qu8_gavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006732 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07006733 "test/qu8-gavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006734 "test/gavgpool-microkernel-tester.h",
6735 "src/xnnpack/AlignedAllocator.h",
6736 ] + MICROKERNEL_TEST_HDRS,
6737 deps = MICROKERNEL_TEST_DEPS,
6738)
6739
6740xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07006741 name = "qu8_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006742 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07006743 "test/qu8-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006744 "test/gemm-microkernel-tester.h",
6745 "src/xnnpack/AlignedAllocator.h",
6746 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07006747 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07006748)
6749
6750xnnpack_unit_test(
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07006751 name = "qu8_requantization_test",
6752 srcs = [
6753 "src/xnnpack/requantization-stubs.h",
6754 "test/qu8-requantization.cc",
6755 "test/requantization-tester.h",
6756 ] + MICROKERNEL_TEST_HDRS,
6757 deps = MICROKERNEL_TEST_DEPS,
6758)
6759
6760xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07006761 name = "qu8_vadd_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006762 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07006763 "test/qu8-vadd-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006764 "test/vadd-microkernel-tester.h",
6765 ] + MICROKERNEL_TEST_HDRS,
6766 deps = MICROKERNEL_TEST_DEPS,
6767)
6768
6769xnnpack_unit_test(
6770 name = "u8_clamp_test",
6771 srcs = [
6772 "test/u8-clamp.cc",
6773 "test/clamp-microkernel-tester.h",
6774 ] + MICROKERNEL_TEST_HDRS,
6775 deps = MICROKERNEL_TEST_DEPS,
6776)
6777
6778xnnpack_unit_test(
6779 name = "u8_lut32norm_test",
6780 srcs = [
6781 "test/u8-lut32norm.cc",
6782 "test/lut-norm-microkernel-tester.h",
6783 ] + MICROKERNEL_TEST_HDRS,
6784 deps = MICROKERNEL_TEST_DEPS,
6785)
6786
6787xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07006788 name = "u8_maxpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006789 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07006790 "test/u8-maxpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006791 "test/maxpool-microkernel-tester.h",
6792 ] + MICROKERNEL_TEST_HDRS,
6793 deps = MICROKERNEL_TEST_DEPS,
6794)
6795
6796xnnpack_unit_test(
6797 name = "u8_rmax_test",
6798 srcs = [
6799 "test/u8-rmax.cc",
6800 "test/rmax-microkernel-tester.h",
6801 ] + MICROKERNEL_TEST_HDRS,
6802 deps = MICROKERNEL_TEST_DEPS,
6803)
6804
6805xnnpack_unit_test(
Marat Dukhanad71b9a2020-11-20 00:01:51 -08006806 name = "x32_depthtospace2d_chw2hwc_test",
Yury Kartynnike7841862020-11-04 18:22:18 -08006807 srcs = [
Marat Dukhanad71b9a2020-11-20 00:01:51 -08006808 "test/x32-depthtospace2d-chw2hwc.cc",
6809 "test/depthtospace-microkernel-tester.h",
Yury Kartynnike7841862020-11-04 18:22:18 -08006810 ] + MICROKERNEL_TEST_HDRS,
6811 deps = MICROKERNEL_TEST_DEPS,
6812)
6813
6814xnnpack_unit_test(
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -07006815 name = "x32_fill_test",
6816 srcs = [
6817 "test/x32-fill.cc",
6818 "test/fill-microkernel-tester.h",
6819 ] + MICROKERNEL_TEST_HDRS,
6820 deps = MICROKERNEL_TEST_DEPS,
6821)
6822
6823xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07006824 name = "x32_packx_test",
6825 srcs = [
6826 "test/x32-packx.cc",
6827 "test/pack-microkernel-tester.h",
6828 "src/xnnpack/AlignedAllocator.h",
6829 ] + MICROKERNEL_TEST_HDRS,
6830 deps = MICROKERNEL_TEST_DEPS,
6831)
6832
6833xnnpack_unit_test(
6834 name = "x32_pad_test",
6835 srcs = [
6836 "test/x32-pad.cc",
6837 "test/pad-microkernel-tester.h",
6838 ] + MICROKERNEL_TEST_HDRS,
6839 deps = MICROKERNEL_TEST_DEPS,
6840)
6841
6842xnnpack_unit_test(
6843 name = "x32_unpool_test",
6844 srcs = [
6845 "test/x32-unpool.cc",
6846 "test/unpool-microkernel-tester.h",
6847 ] + MICROKERNEL_TEST_HDRS,
6848 deps = MICROKERNEL_TEST_DEPS,
6849)
6850
6851xnnpack_unit_test(
6852 name = "x32_zip_test",
6853 srcs = [
6854 "test/x32-zip.cc",
6855 "test/zip-microkernel-tester.h",
6856 ] + MICROKERNEL_TEST_HDRS,
6857 deps = MICROKERNEL_TEST_DEPS,
6858)
6859
6860xnnpack_unit_test(
6861 name = "x8_lut_test",
6862 srcs = [
6863 "test/x8-lut.cc",
6864 "test/lut-microkernel-tester.h",
6865 ] + MICROKERNEL_TEST_HDRS,
6866 deps = MICROKERNEL_TEST_DEPS,
6867)
6868
6869xnnpack_unit_test(
6870 name = "x8_zip_test",
6871 srcs = [
6872 "test/x8-zip.cc",
6873 "test/zip-microkernel-tester.h",
6874 ] + MICROKERNEL_TEST_HDRS,
6875 deps = MICROKERNEL_TEST_DEPS,
6876)
6877
Marat Dukhan20c3b922020-03-10 03:45:06 -07006878########################## Size tests for the library #########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07006879
6880xnnpack_binary(
Marat Dukhan20c3b922020-03-10 03:45:06 -07006881 name = "operator_size_test",
6882 srcs = ["test/operator-size.c"],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07006883 deps = [":xnnpack_operators_nhwc_f32"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07006884)
6885
Marat Dukhan20c3b922020-03-10 03:45:06 -07006886xnnpack_binary(
6887 name = "subgraph_size_test",
6888 srcs = ["test/subgraph-size.c"],
6889 deps = [":XNNPACK"],
6890)
6891
6892########################### Unit tests for operators ##########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07006893
6894xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07006895 name = "abs_nc_test",
6896 srcs = [
6897 "test/abs-nc.cc",
6898 "test/abs-operator-tester.h",
6899 ],
6900 deps = OPERATOR_TEST_DEPS,
6901)
6902
6903xnnpack_unit_test(
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08006904 name = "add_nd_test",
Artsiom Ablavatski2202c812021-01-22 14:16:43 -08006905 timeout = "moderate",
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08006906 srcs = [
6907 "test/add-nd.cc",
6908 "test/binary-elementwise-operator-tester.h",
6909 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006910 deps = OPERATOR_TEST_DEPS,
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08006911)
6912
6913xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006914 name = "argmax_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006915 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006916 "test/argmax-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006917 "test/argmax-pooling-operator-tester.h",
6918 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07006919 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006920)
6921
6922xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006923 name = "average_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006924 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006925 "test/average-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006926 "test/average-pooling-operator-tester.h",
6927 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07006928 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006929)
6930
6931xnnpack_unit_test(
Marat Dukhan64e52512020-06-09 13:41:16 -07006932 name = "bankers_rounding_nc_test",
6933 srcs = [
6934 "test/bankers-rounding-nc.cc",
6935 "test/bankers-rounding-operator-tester.h",
6936 ],
6937 deps = OPERATOR_TEST_DEPS,
6938)
6939
6940xnnpack_unit_test(
6941 name = "ceiling_nc_test",
6942 srcs = [
6943 "test/ceiling-nc.cc",
6944 "test/ceiling-operator-tester.h",
6945 ],
6946 deps = OPERATOR_TEST_DEPS,
6947)
6948
6949xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006950 name = "channel_shuffle_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006951 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006952 "test/channel-shuffle-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006953 "test/channel-shuffle-operator-tester.h",
6954 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006955 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006956)
6957
6958xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006959 name = "clamp_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006960 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006961 "test/clamp-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006962 "test/clamp-operator-tester.h",
6963 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006964 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006965)
6966
6967xnnpack_unit_test(
Marat Dukhan065b11e2020-05-22 09:49:41 -07006968 name = "constant_pad_nd_test",
6969 srcs = [
6970 "test/constant-pad-nd.cc",
6971 "test/constant-pad-operator-tester.h",
6972 ],
6973 deps = OPERATOR_TEST_DEPS,
6974)
6975
6976xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006977 name = "convolution_nhwc_test",
Artsiom Ablavatski2202c812021-01-22 14:16:43 -08006978 timeout = "moderate",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006979 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006980 "test/convolution-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006981 "test/convolution-operator-tester.h",
6982 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006983 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006984)
6985
6986xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006987 name = "convolution_nchw_test",
Artsiom Ablavatski2202c812021-01-22 14:16:43 -08006988 timeout = "moderate",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006989 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006990 "test/convolution-nchw.cc",
6991 "test/convolution-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006992 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006993 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006994)
6995
6996xnnpack_unit_test(
Marat Dukhan4e21b272020-06-04 18:45:01 -07006997 name = "copy_nc_test",
6998 srcs = [
6999 "test/copy-nc.cc",
7000 "test/copy-operator-tester.h",
7001 ],
7002 deps = OPERATOR_TEST_DEPS,
7003)
7004
7005xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08007006 name = "deconvolution_nhwc_test",
Artsiom Ablavatskic1aa2972020-12-08 11:23:34 -08007007 timeout = "moderate",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007008 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08007009 "test/deconvolution-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007010 "test/deconvolution-operator-tester.h",
7011 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07007012 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07007013)
7014
7015xnnpack_unit_test(
Marat Dukhan188d1042020-11-24 23:39:40 -08007016 name = "depth_to_space_nchw2nhwc_test",
Artsiom Ablavatski0f1dc182020-11-05 19:21:50 -08007017 srcs = [
Marat Dukhan188d1042020-11-24 23:39:40 -08007018 "test/depth-to-space-nchw2nhwc.cc",
Artsiom Ablavatski0f1dc182020-11-05 19:21:50 -08007019 "test/depth-to-space-operator-tester.h",
7020 ] + OPERATOR_TEST_PARAMS_HDRS,
7021 deps = OPERATOR_TEST_DEPS,
7022)
7023
7024xnnpack_unit_test(
Marat Dukhan0e521172020-11-25 13:10:04 -08007025 name = "depth_to_space_nhwc_test",
7026 srcs = [
7027 "test/depth-to-space-nhwc.cc",
7028 "test/depth-to-space-operator-tester.h",
7029 ] + OPERATOR_TEST_PARAMS_HDRS,
7030 deps = OPERATOR_TEST_DEPS,
7031)
7032
7033xnnpack_unit_test(
Marat Dukhan69180502019-12-06 15:00:31 -08007034 name = "divide_nd_test",
7035 srcs = [
7036 "test/binary-elementwise-operator-tester.h",
7037 "test/divide-nd.cc",
7038 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07007039 deps = OPERATOR_TEST_DEPS,
Marat Dukhan69180502019-12-06 15:00:31 -08007040)
7041
7042xnnpack_unit_test(
Marat Dukhanb6bd4bc2020-12-01 17:01:40 -08007043 name = "elu_nc_test",
7044 srcs = [
7045 "test/elu-nc.cc",
7046 "test/elu-operator-tester.h",
7047 ],
7048 deps = OPERATOR_TEST_DEPS,
7049)
7050
7051xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08007052 name = "fully_connected_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007053 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08007054 "test/fully-connected-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007055 "test/fully-connected-operator-tester.h",
7056 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07007057 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07007058)
7059
7060xnnpack_unit_test(
Marat Dukhan64e52512020-06-09 13:41:16 -07007061 name = "floor_nc_test",
7062 srcs = [
7063 "test/floor-nc.cc",
7064 "test/floor-operator-tester.h",
7065 ],
7066 deps = OPERATOR_TEST_DEPS,
7067)
7068
7069xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08007070 name = "global_average_pooling_nwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007071 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08007072 "test/global-average-pooling-nwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007073 "test/global-average-pooling-operator-tester.h",
Marat Dukhanef61d022020-06-19 13:54:49 -07007074 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07007075 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07007076)
7077
7078xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08007079 name = "global_average_pooling_ncw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007080 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08007081 "test/global-average-pooling-ncw.cc",
7082 "test/global-average-pooling-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007083 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07007084 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07007085)
7086
7087xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08007088 name = "hardswish_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007089 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08007090 "test/hardswish-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007091 "test/hardswish-operator-tester.h",
7092 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07007093 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07007094)
7095
7096xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08007097 name = "leaky_relu_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007098 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08007099 "test/leaky-relu-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007100 "test/leaky-relu-operator-tester.h",
7101 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07007102 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07007103)
7104
7105xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08007106 name = "max_pooling_nhwc_test",
Artsiom Ablavatski2202c812021-01-22 14:16:43 -08007107 timeout = "moderate",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007108 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08007109 "test/max-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007110 "test/max-pooling-operator-tester.h",
7111 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07007112 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07007113)
7114
7115xnnpack_unit_test(
Marat Dukhan79e7f842019-12-05 14:35:50 -08007116 name = "maximum_nd_test",
7117 srcs = [
7118 "test/binary-elementwise-operator-tester.h",
7119 "test/maximum-nd.cc",
7120 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07007121 deps = OPERATOR_TEST_DEPS,
Marat Dukhan79e7f842019-12-05 14:35:50 -08007122)
7123
7124xnnpack_unit_test(
7125 name = "minimum_nd_test",
7126 srcs = [
7127 "test/binary-elementwise-operator-tester.h",
7128 "test/minimum-nd.cc",
7129 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07007130 deps = OPERATOR_TEST_DEPS,
Marat Dukhan79e7f842019-12-05 14:35:50 -08007131)
7132
7133xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08007134 name = "multiply_nd_test",
Marat Dukhanca2733c2019-11-15 23:21:17 -08007135 srcs = [
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08007136 "test/binary-elementwise-operator-tester.h",
Marat Dukhanefc47b82019-11-18 09:25:38 -08007137 "test/multiply-nd.cc",
Marat Dukhanca2733c2019-11-15 23:21:17 -08007138 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07007139 deps = OPERATOR_TEST_DEPS,
Marat Dukhanca2733c2019-11-15 23:21:17 -08007140)
7141
7142xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07007143 name = "negate_nc_test",
7144 srcs = [
7145 "test/negate-nc.cc",
7146 "test/negate-operator-tester.h",
7147 ],
7148 deps = OPERATOR_TEST_DEPS,
7149)
7150
7151xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08007152 name = "prelu_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007153 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08007154 "test/prelu-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007155 "test/prelu-operator-tester.h",
7156 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07007157 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07007158)
7159
7160xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08007161 name = "resize_bilinear_nhwc_test",
Marat Dukhan69722492019-11-11 19:55:50 -08007162 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08007163 "test/resize-bilinear-nhwc.cc",
Marat Dukhan69722492019-11-11 19:55:50 -08007164 "test/resize-bilinear-operator-tester.h",
7165 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07007166 deps = OPERATOR_TEST_DEPS,
Marat Dukhan69722492019-11-11 19:55:50 -08007167)
7168
7169xnnpack_unit_test(
Artsiom Ablavatski97918102020-10-27 15:52:59 -07007170 name = "resize_bilinear_nchw_test",
7171 srcs = [
7172 "test/resize-bilinear-nchw.cc",
7173 "test/resize-bilinear-operator-tester.h",
7174 ] + OPERATOR_TEST_PARAMS_HDRS,
7175 deps = OPERATOR_TEST_DEPS,
7176)
7177
7178xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08007179 name = "sigmoid_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007180 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08007181 "test/sigmoid-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007182 "test/sigmoid-operator-tester.h",
7183 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07007184 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07007185)
7186
7187xnnpack_unit_test(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08007188 name = "softmax_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007189 srcs = [
Marat Dukhanfd8e6892020-01-27 15:25:25 -08007190 "test/softmax-nc.cc",
7191 "test/softmax-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007192 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07007193 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07007194)
7195
7196xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07007197 name = "square_nc_test",
7198 srcs = [
7199 "test/square-nc.cc",
7200 "test/square-operator-tester.h",
7201 ],
7202 deps = OPERATOR_TEST_DEPS,
7203)
7204
7205xnnpack_unit_test(
Marat Dukhan6804bbd2020-06-30 19:26:11 -07007206 name = "square_root_nc_test",
7207 srcs = [
7208 "test/square-root-nc.cc",
7209 "test/square-root-operator-tester.h",
7210 ],
7211 deps = OPERATOR_TEST_DEPS,
7212)
7213
7214xnnpack_unit_test(
Marat Dukhanf7399262020-06-05 10:58:44 -07007215 name = "squared_difference_nd_test",
7216 srcs = [
7217 "test/binary-elementwise-operator-tester.h",
7218 "test/squared-difference-nd.cc",
7219 ],
7220 deps = OPERATOR_TEST_DEPS,
7221)
7222
7223xnnpack_unit_test(
Marat Dukhan05f3f6d2019-12-03 15:13:53 -08007224 name = "subtract_nd_test",
7225 srcs = [
7226 "test/binary-elementwise-operator-tester.h",
7227 "test/subtract-nd.cc",
7228 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07007229 deps = OPERATOR_TEST_DEPS,
Marat Dukhan05f3f6d2019-12-03 15:13:53 -08007230)
7231
7232xnnpack_unit_test(
Marat Dukhan64e52512020-06-09 13:41:16 -07007233 name = "truncation_nc_test",
7234 srcs = [
7235 "test/truncation-nc.cc",
7236 "test/truncation-operator-tester.h",
7237 ],
7238 deps = OPERATOR_TEST_DEPS,
7239)
7240
7241xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08007242 name = "unpooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007243 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08007244 "test/unpooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007245 "test/unpooling-operator-tester.h",
7246 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07007247 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07007248)
7249
Chao Mei6ddfc602020-05-13 22:29:36 -07007250############################### Misc unit tests ###############################
7251
7252xnnpack_unit_test(
7253 name = "memory_planner_test",
7254 srcs = [
7255 "test/memory-planner-test.cc",
7256 ],
7257 deps = [
7258 ":XNNPACK",
7259 ":memory_planner",
7260 ],
7261)
7262
XNNPACK Teamab8c4c82020-10-09 08:05:51 -07007263xnnpack_unit_test(
7264 name = "subgraph_nchw_test",
7265 srcs = [
7266 "src/xnnpack/subgraph.h",
7267 "test/subgraph-nchw.cc",
7268 "test/subgraph-tester.h",
7269 ],
7270 deps = [
7271 ":XNNPACK",
7272 ],
7273)
7274
Marat Dukhan08c4a432019-10-03 09:29:21 -07007275############################# Build configurations #############################
7276
Marat Dukhanb8642352019-10-30 15:43:02 -07007277# Enables usage of assembly kernels.
Marat Dukhan08c4a432019-10-03 09:29:21 -07007278config_setting(
Marat Dukhanb8642352019-10-30 15:43:02 -07007279 name = "xnn_enable_assembly_explicit_true",
7280 define_values = {"xnn_enable_assembly": "true"},
7281)
7282
7283# Disables usage of assembly kernels.
7284config_setting(
7285 name = "xnn_enable_assembly_explicit_false",
7286 define_values = {"xnn_enable_assembly": "false"},
7287)
7288
Marat Dukhan9de90e02020-06-18 16:04:12 -07007289# Enables usage of sparse inference.
7290config_setting(
7291 name = "xnn_enable_sparse_explicit_true",
7292 define_values = {"xnn_enable_sparse": "true"},
7293)
7294
7295# Disables usage of sparse inference.
7296config_setting(
7297 name = "xnn_enable_sparse_explicit_false",
7298 define_values = {"xnn_enable_sparse": "false"},
7299)
7300
Marat Dukhan05702cf2020-03-26 15:41:33 -07007301# Disables usage of HMP-aware optimizations.
7302config_setting(
7303 name = "xnn_enable_hmp_explicit_false",
7304 define_values = {"xnn_enable_hmp": "false"},
7305)
7306
Chao Mei6ddfc602020-05-13 22:29:36 -07007307# Enable usage of optimized memory allocation
7308config_setting(
7309 name = "xnn_enable_memopt_explicit_true",
7310 define_values = {"xnnpack_enable_memopt": "true"},
7311)
7312
7313# Disable usage of optimized memory allocation
7314config_setting(
7315 name = "xnn_enable_memopt_explicit_false",
7316 define_values = {"xnnpack_enable_memopt": "false"},
7317)
7318
Marat Dukhanb8642352019-10-30 15:43:02 -07007319# Builds with -c dbg
7320config_setting(
7321 name = "debug_build",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007322 values = {
Marat Dukhanb8642352019-10-30 15:43:02 -07007323 "compilation_mode": "dbg",
7324 },
7325)
7326
7327# Builds with -c opt
7328config_setting(
7329 name = "optimized_build",
7330 values = {
7331 "compilation_mode": "opt",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007332 },
7333)
7334
7335config_setting(
Marat Dukhanb8642352019-10-30 15:43:02 -07007336 name = "linux_k8",
7337 values = {"cpu": "k8"},
7338)
7339
7340config_setting(
Marat Dukhan582094e2020-04-30 17:21:25 -07007341 name = "linux_arm",
7342 values = {"cpu": "arm"},
Marat Dukhan4e45e662019-10-03 15:40:24 -07007343)
7344
7345config_setting(
Marat Dukhanf0bd4de2020-06-15 15:53:19 -07007346 name = "linux_armeabi",
7347 values = {"cpu": "armeabi"},
7348)
7349
7350config_setting(
Terry Heo68eef3f2020-04-13 22:53:52 -07007351 name = "linux_armhf",
7352 values = {"cpu": "armhf"},
7353)
7354
7355config_setting(
Marat Dukhana720e932020-06-10 13:01:11 -07007356 name = "linux_armv7a",
7357 values = {"cpu": "armv7a"},
7358)
7359
7360config_setting(
Marat Dukhan582094e2020-04-30 17:21:25 -07007361 name = "linux_aarch64",
7362 values = {"cpu": "aarch64"},
7363)
7364
7365config_setting(
Marat Dukhan08c4a432019-10-03 09:29:21 -07007366 name = "android",
7367 values = {"crosstool_top": "//external:android/crosstool"},
7368)
7369
7370config_setting(
7371 name = "android_armv7",
7372 values = {
7373 "crosstool_top": "//external:android/crosstool",
7374 "cpu": "armeabi-v7a",
7375 },
7376)
7377
7378config_setting(
7379 name = "android_arm64",
7380 values = {
7381 "crosstool_top": "//external:android/crosstool",
7382 "cpu": "arm64-v8a",
7383 },
7384)
7385
7386config_setting(
7387 name = "android_x86",
7388 values = {
7389 "crosstool_top": "//external:android/crosstool",
7390 "cpu": "x86",
7391 },
7392)
7393
7394config_setting(
7395 name = "android_x86_64",
7396 values = {
7397 "crosstool_top": "//external:android/crosstool",
7398 "cpu": "x86_64",
7399 },
7400)
7401
7402config_setting(
Marat Dukhan10a38082020-04-17 03:58:35 -07007403 name = "windows_x86_64",
7404 values = {"cpu": "x64_windows"},
Marat Dukhan9fe932e2020-04-11 17:14:15 -07007405)
7406
7407config_setting(
Marat Dukhan10a38082020-04-17 03:58:35 -07007408 name = "windows_x86_64_clang",
7409 values = {
7410 "compiler": "clang-cl",
7411 "cpu": "x64_windows",
7412 },
7413)
7414
7415config_setting(
7416 name = "windows_x86_64_mingw",
7417 values = {
7418 "compiler": "mingw-gcc",
7419 "cpu": "x64_windows",
7420 },
7421)
7422
7423config_setting(
7424 name = "windows_x86_64_msys",
7425 values = {
7426 "compiler": "msys-gcc",
7427 "cpu": "x64_windows",
7428 },
Marat Dukhan9fe932e2020-04-11 17:14:15 -07007429)
7430
7431config_setting(
Marat Dukhan885ca242019-10-07 09:17:32 -07007432 name = "macos_x86_64",
7433 values = {
7434 "apple_platform_type": "macos",
7435 "cpu": "darwin",
7436 },
7437)
7438
7439config_setting(
Simon Maurerae33ab82021-03-03 23:38:22 +01007440 name = "macos_arm64",
7441 values = {
7442 "apple_platform_type": "macos",
7443 "cpu": "darwin_arm64",
7444 },
7445)
7446
7447config_setting(
Marat Dukhan08c4a432019-10-03 09:29:21 -07007448 name = "emscripten",
Marat Dukhan1a729ec2019-10-07 09:31:44 -07007449 values = {"crosstool_top": "//toolchain:emscripten"},
Marat Dukhan08c4a432019-10-03 09:29:21 -07007450)
7451
7452config_setting(
7453 name = "emscripten_wasm",
7454 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07007455 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007456 "cpu": "wasm",
7457 },
7458)
7459
7460config_setting(
7461 name = "emscripten_wasmsimd",
7462 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07007463 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007464 "cpu": "wasm",
Marat Dukhan81c62602020-05-29 13:22:49 -07007465 "copt": "-msimd128",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007466 },
7467)
7468
7469config_setting(
Marat Dukhan1498d1d2020-02-11 20:00:05 -08007470 name = "ios_armv7",
7471 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08007472 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08007473 "cpu": "ios_armv7",
7474 },
7475)
7476
7477config_setting(
7478 name = "ios_arm64",
7479 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08007480 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08007481 "cpu": "ios_arm64",
7482 },
7483)
7484
7485config_setting(
7486 name = "ios_arm64e",
7487 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08007488 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08007489 "cpu": "ios_arm64e",
7490 },
7491)
7492
7493config_setting(
7494 name = "ios_x86",
7495 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08007496 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08007497 "cpu": "ios_i386",
7498 },
7499)
7500
7501config_setting(
7502 name = "ios_x86_64",
7503 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08007504 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08007505 "cpu": "ios_x86_64",
7506 },
7507)
7508
7509config_setting(
7510 name = "watchos_armv7k",
7511 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08007512 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08007513 "cpu": "watchos_armv7k",
7514 },
7515)
7516
7517config_setting(
7518 name = "watchos_arm64_32",
7519 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08007520 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08007521 "cpu": "watchos_arm64_32",
7522 },
7523)
7524
7525config_setting(
7526 name = "watchos_x86",
7527 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08007528 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08007529 "cpu": "watchos_i386",
7530 },
7531)
7532
7533config_setting(
7534 name = "watchos_x86_64",
7535 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08007536 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08007537 "cpu": "watchos_x86_64",
7538 },
7539)
7540
7541config_setting(
7542 name = "tvos_arm64",
7543 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08007544 "apple_platform_type": "tvos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08007545 "cpu": "tvos_arm64",
7546 },
7547)
7548
7549config_setting(
7550 name = "tvos_x86_64",
7551 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08007552 "apple_platform_type": "tvos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08007553 "cpu": "tvos_x86_64",
7554 },
7555)