blob: 845637be1a040fcc992945f731b92ee6ffd6c1dd [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 Dukhane8265432020-04-28 18:42:59 -070063 "src/operators/fully-connected-nc.c",
64 "src/operators/global-average-pooling-ncw.c",
65 "src/operators/global-average-pooling-nwc.c",
Marat Dukhane8265432020-04-28 18:42:59 -070066 "src/operators/leaky-relu-nc.c",
67 "src/operators/max-pooling-nhwc.c",
68 "src/operators/prelu-nc.c",
Artsiom Ablavatski97918102020-10-27 15:52:59 -070069 "src/operators/resize-bilinear-nchw.c",
Marat Dukhane8265432020-04-28 18:42:59 -070070 "src/operators/resize-bilinear-nhwc.c",
71 "src/operators/sigmoid-nc.c",
72 "src/operators/softmax-nc.c",
Marat Dukhanc3065f52020-06-04 13:33:32 -070073 "src/operators/unary-elementwise-nc.c",
Marat Dukhane8265432020-04-28 18:42:59 -070074 "src/operators/unpooling-nhwc.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -070075]
76
Marat Dukhan95e8b7a2020-06-03 12:46:26 -070077SUBGRAPH_SRCS = [
Marat Dukhan5fab4092020-06-10 01:28:28 -070078 "src/subgraph/abs.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -070079 "src/subgraph/add2.c",
80 "src/subgraph/argmax-pooling-2d.c",
81 "src/subgraph/average-pooling-2d.c",
Marat Dukhan5fab4092020-06-10 01:28:28 -070082 "src/subgraph/bankers-rounding.c",
83 "src/subgraph/ceiling.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -070084 "src/subgraph/clamp.c",
85 "src/subgraph/convolution-2d.c",
86 "src/subgraph/deconvolution-2d.c",
Artsiom Ablavatskibbe85062020-11-05 14:07:37 -080087 "src/subgraph/depth-to-space.c",
Frank Barchard9cef5ea2020-11-18 14:52:08 -080088 "src/subgraph/depthwise-convolution-2d.c",
Marat Dukhan9d3a4592020-06-05 16:52:42 -070089 "src/subgraph/divide.c",
Marat Dukhan5fab4092020-06-10 01:28:28 -070090 "src/subgraph/floor.c",
Frank Barchard04336c12020-10-22 16:48:55 -070091 "src/subgraph/fully-connected.c",
Marat Dukhana059b7d2020-06-11 11:41:27 -070092 "src/subgraph/global-average-pooling-2d.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -070093 "src/subgraph/hardswish.c",
Marat Dukhan5bbebac2020-06-10 19:42:15 -070094 "src/subgraph/leaky-relu.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -070095 "src/subgraph/max-pooling-2d.c",
Marat Dukhan9d3a4592020-06-05 16:52:42 -070096 "src/subgraph/maximum2.c",
97 "src/subgraph/minimum2.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -070098 "src/subgraph/multiply2.c",
Marat Dukhan5fab4092020-06-10 01:28:28 -070099 "src/subgraph/negate.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -0700100 "src/subgraph/prelu.c",
101 "src/subgraph/sigmoid.c",
102 "src/subgraph/softmax.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700103 "src/subgraph/square-root.c",
104 "src/subgraph/square.c",
105 "src/subgraph/squared-difference.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -0700106 "src/subgraph/static-constant-pad.c",
Marat Dukhand27202d2020-07-09 23:43:40 -0700107 "src/subgraph/static-reshape.c",
Marat Dukhanaff24e22020-07-23 01:43:58 -0700108 "src/subgraph/static-resize-bilinear-2d.c",
Marat Dukhan9d3a4592020-06-05 16:52:42 -0700109 "src/subgraph/subtract.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -0700110 "src/subgraph/unpooling-2d.c",
111]
112
Marat Dukhan3a77ea72019-12-23 12:10:24 -0800113TABLE_SRCS = [
114 "src/tables/exp2-k-over-64.c",
115 "src/tables/exp2-k-over-2048.c",
Marat Dukhanc60742b2020-11-23 12:33:27 -0800116 "src/tables/exp2minus-k-over-16.c",
Marat Dukhan1f256fc2020-09-24 21:27:14 -0700117 "src/tables/exp2minus-k-over-64.c",
118 "src/tables/exp2minus-k-over-2048.c",
Marat Dukhan3a77ea72019-12-23 12:10:24 -0800119]
120
Marat Dukhan08c4a432019-10-03 09:29:21 -0700121SCALAR_UKERNELS = [
Marat Dukhan329da642019-11-19 21:44:39 -0800122 "src/f32-argmaxpool/4x-scalar-c1.c",
Marat Dukhan1e782c42019-11-21 17:02:40 -0800123 "src/f32-argmaxpool/9p8x-scalar-c1.c",
Marat Dukhan329da642019-11-19 21:44:39 -0800124 "src/f32-argmaxpool/9x-scalar-c1.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700125 "src/f32-avgpool/9p8x-minmax-scalar-c1.c",
126 "src/f32-avgpool/9x-minmax-scalar-c1.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -0700127 "src/f32-clamp/gen/scalar-x1.c",
128 "src/f32-clamp/gen/scalar-x2.c",
129 "src/f32-clamp/gen/scalar-x4.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700130 "src/f32-conv-hwc/3x3s2p0p1c3x4-scalar-1x1.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700131 "src/f32-conv-hwc/3x3s2p1c3x4-scalar-1x1.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -0700132 "src/f32-conv-hwc2chw/3x3s2p1c3x4-scalar-1x1.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700133 "src/f32-dwconv/gen/up1x4-minmax-scalar-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700134 "src/f32-dwconv/gen/up1x4-minmax-scalar.c",
135 "src/f32-dwconv/gen/up1x4-scalar-acc2.c",
136 "src/f32-dwconv/gen/up1x4-scalar.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700137 "src/f32-dwconv/gen/up1x9-minmax-scalar-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700138 "src/f32-dwconv/gen/up1x9-minmax-scalar.c",
139 "src/f32-dwconv/gen/up1x9-scalar-acc2.c",
140 "src/f32-dwconv/gen/up1x9-scalar.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700141 "src/f32-dwconv/gen/up1x25-minmax-scalar-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700142 "src/f32-dwconv/gen/up1x25-minmax-scalar.c",
143 "src/f32-dwconv/gen/up1x25-scalar-acc2.c",
144 "src/f32-dwconv/gen/up1x25-scalar.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700145 "src/f32-dwconv/gen/up2x4-minmax-scalar-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700146 "src/f32-dwconv/gen/up2x4-minmax-scalar.c",
147 "src/f32-dwconv/gen/up2x4-scalar-acc2.c",
148 "src/f32-dwconv/gen/up2x4-scalar.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700149 "src/f32-dwconv/gen/up2x9-minmax-scalar-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700150 "src/f32-dwconv/gen/up2x9-minmax-scalar.c",
151 "src/f32-dwconv/gen/up2x9-scalar-acc2.c",
152 "src/f32-dwconv/gen/up2x9-scalar.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700153 "src/f32-dwconv/gen/up2x25-minmax-scalar-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700154 "src/f32-dwconv/gen/up2x25-minmax-scalar.c",
155 "src/f32-dwconv/gen/up2x25-scalar-acc2.c",
156 "src/f32-dwconv/gen/up2x25-scalar.c",
Frank Barchard35db7d02020-10-26 13:37:34 -0700157 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-scalar-1x1-acc2.c",
158 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-scalar-1x1-acc3.c",
159 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-scalar-1x1-acc4.c",
Marat Dukhan91249d22020-10-24 12:02:51 -0700160 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-scalar-1x1.c",
Frank Barchard35db7d02020-10-26 13:37:34 -0700161 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-scalar-2x1-acc2.c",
Marat Dukhan91249d22020-10-24 12:02:51 -0700162 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-scalar-2x1.c",
163 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-scalar-3x1.c",
164 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-scalar-4x1.c",
165 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-scalar-5x1.c",
166 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-scalar-6x1.c",
Marat Dukhancf5b3c32020-10-25 19:21:10 -0700167 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-scalar-1x1-acc2.c",
168 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-scalar-1x1-acc3.c",
169 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-scalar-1x1-acc4.c",
Frank Barchard35db7d02020-10-26 13:37:34 -0700170 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-scalar-1x1.c",
Marat Dukhancf5b3c32020-10-25 19:21:10 -0700171 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-scalar-2x1-acc2.c",
Frank Barchard35db7d02020-10-26 13:37:34 -0700172 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-scalar-2x1.c",
173 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-scalar-3x1.c",
174 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-scalar-4x1.c",
Marat Dukhanc4efb002020-10-25 23:14:47 -0700175 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-scalar-1x1-acc2.c",
176 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-scalar-1x1-acc3.c",
177 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-scalar-1x1-acc4.c",
178 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-scalar-1x1-acc5.c",
Frank Barchard35db7d02020-10-26 13:37:34 -0700179 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-scalar-1x1.c",
Marat Dukhanc4efb002020-10-25 23:14:47 -0700180 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-scalar-2x1-acc2.c",
181 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-scalar-2x1-acc3.c",
Frank Barchard35db7d02020-10-26 13:37:34 -0700182 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-scalar-2x1.c",
Marat Dukhanc4efb002020-10-25 23:14:47 -0700183 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-scalar-3x1-acc2.c",
Frank Barchard35db7d02020-10-26 13:37:34 -0700184 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-scalar-3x1.c",
Marat Dukhan29c0c332020-10-28 22:11:00 -0700185 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-scalar-1x1-acc2.c",
186 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-scalar-1x1-acc3.c",
187 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-scalar-1x1-acc4.c",
188 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-scalar-1x1-acc5.c",
189 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-scalar-1x1.c",
190 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-scalar-2x1-acc2.c",
191 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-scalar-2x1-acc3.c",
192 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-scalar-2x1.c",
193 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-scalar-3x1-acc2.c",
194 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-scalar-3x1.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -0700195 "src/f32-gavgpool-cw/scalar-x1.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700196 "src/f32-gavgpool/7p7x-minmax-scalar-c1.c",
197 "src/f32-gavgpool/7x-minmax-scalar-c1.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700198 "src/f32-gemm/gen-inc/1x4inc-minmax-scalar.c",
199 "src/f32-gemm/gen-inc/2x4inc-minmax-scalar.c",
200 "src/f32-gemm/gen-inc/4x4inc-minmax-scalar.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700201 "src/f32-gemm/gen/1x4-minmax-scalar.c",
202 "src/f32-gemm/gen/1x4-relu-scalar.c",
203 "src/f32-gemm/gen/1x4-scalar.c",
204 "src/f32-gemm/gen/2x4-minmax-scalar.c",
205 "src/f32-gemm/gen/2x4-relu-scalar.c",
206 "src/f32-gemm/gen/2x4-scalar.c",
207 "src/f32-gemm/gen/4x2-minmax-scalar.c",
208 "src/f32-gemm/gen/4x2-relu-scalar.c",
209 "src/f32-gemm/gen/4x2-scalar.c",
210 "src/f32-gemm/gen/4x4-minmax-scalar.c",
211 "src/f32-gemm/gen/4x4-relu-scalar.c",
212 "src/f32-gemm/gen/4x4-scalar.c",
Marat Dukhanefaac272020-07-16 16:58:07 -0700213 "src/f32-hswish/gen/hswish-scalar-x1.c",
214 "src/f32-hswish/gen/hswish-scalar-x2.c",
215 "src/f32-hswish/gen/hswish-scalar-x4.c",
XNNPACK Team21432672020-10-19 19:58:48 -0700216 "src/f32-ibilinear-chw/gen/scalar-p1.c",
217 "src/f32-ibilinear-chw/gen/scalar-p2.c",
218 "src/f32-ibilinear-chw/gen/scalar-p4.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700219 "src/f32-ibilinear/gen/scalar-c1.c",
220 "src/f32-ibilinear/gen/scalar-c2.c",
221 "src/f32-ibilinear/gen/scalar-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700222 "src/f32-igemm/gen/1x4-minmax-scalar.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700223 "src/f32-igemm/gen/1x4-relu-scalar.c",
224 "src/f32-igemm/gen/1x4-scalar.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700225 "src/f32-igemm/gen/2x4-minmax-scalar.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700226 "src/f32-igemm/gen/2x4-relu-scalar.c",
227 "src/f32-igemm/gen/2x4-scalar.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700228 "src/f32-igemm/gen/4x2-minmax-scalar.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700229 "src/f32-igemm/gen/4x2-relu-scalar.c",
230 "src/f32-igemm/gen/4x2-scalar.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700231 "src/f32-igemm/gen/4x4-minmax-scalar.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700232 "src/f32-igemm/gen/4x4-relu-scalar.c",
233 "src/f32-igemm/gen/4x4-scalar.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700234 "src/f32-maxpool/9p8x-minmax-scalar-c1.c",
235 "src/f32-pavgpool/9p8x-minmax-scalar-c1.c",
236 "src/f32-pavgpool/9x-minmax-scalar-c1.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700237 "src/f32-ppmm/gen/2x4-minmax-scalar.c",
238 "src/f32-ppmm/gen/3x3-minmax-scalar.c",
239 "src/f32-ppmm/gen/4x2-minmax-scalar.c",
240 "src/f32-ppmm/gen/4x4-minmax-scalar.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800241 "src/f32-prelu/gen/scalar-2x1.c",
242 "src/f32-prelu/gen/scalar-2x4.c",
Marat Dukhanf46f6752020-01-21 11:03:49 -0800243 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x1.c",
Marat Dukhanf46f6752020-01-21 11:03:49 -0800244 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x2-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700245 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x2.c",
Marat Dukhanf46f6752020-01-21 11:03:49 -0800246 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x4-acc2.c",
247 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x4-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700248 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x4.c",
Marat Dukhanf46f6752020-01-21 11:03:49 -0800249 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x1.c",
Marat Dukhanf46f6752020-01-21 11:03:49 -0800250 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x2-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700251 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x2.c",
Marat Dukhanf46f6752020-01-21 11:03:49 -0800252 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x4-acc2.c",
253 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x4-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700254 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x4.c",
Frank Barchardfb158e22020-07-15 16:10:10 -0700255 "src/f32-relu/gen/scalar-x1.c",
256 "src/f32-relu/gen/scalar-x2.c",
257 "src/f32-relu/gen/scalar-x4.c",
Frank Barchard37297a62020-08-27 12:15:10 -0700258 "src/f32-relu/gen/scalar-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700259 "src/f32-rmax/scalar.c",
Marat Dukhan3a77ea72019-12-23 12:10:24 -0800260 "src/f32-sigmoid/gen/scalar-lut64-p2-div-x1.c",
261 "src/f32-sigmoid/gen/scalar-lut64-p2-div-x2.c",
262 "src/f32-sigmoid/gen/scalar-lut64-p2-div-x4.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700263 "src/f32-sigmoid/gen/scalar-lut2048-p1-div-x1.c",
264 "src/f32-sigmoid/gen/scalar-lut2048-p1-div-x2.c",
265 "src/f32-sigmoid/gen/scalar-lut2048-p1-div-x4.c",
Marat Dukhan3a77ea72019-12-23 12:10:24 -0800266 "src/f32-sigmoid/gen/scalar-p5-div-x1.c",
267 "src/f32-sigmoid/gen/scalar-p5-div-x2.c",
268 "src/f32-sigmoid/gen/scalar-p5-div-x4.c",
Marat Dukhan355ab432020-04-09 19:01:52 -0700269 "src/f32-spmm/gen/1x1-minmax-scalar-pipelined.c",
270 "src/f32-spmm/gen/1x1-minmax-scalar.c",
271 "src/f32-spmm/gen/2x1-minmax-scalar-pipelined.c",
272 "src/f32-spmm/gen/2x1-minmax-scalar.c",
273 "src/f32-spmm/gen/4x1-minmax-scalar-pipelined.c",
274 "src/f32-spmm/gen/4x1-minmax-scalar.c",
275 "src/f32-spmm/gen/8x1-minmax-scalar-pipelined.c",
276 "src/f32-spmm/gen/8x1-minmax-scalar.c",
277 "src/f32-spmm/gen/8x2-minmax-scalar.c",
278 "src/f32-spmm/gen/8x4-minmax-scalar.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700279 "src/f32-vbinary/gen/vadd-minmax-scalar-x1.c",
280 "src/f32-vbinary/gen/vadd-minmax-scalar-x2.c",
281 "src/f32-vbinary/gen/vadd-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700282 "src/f32-vbinary/gen/vadd-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700283 "src/f32-vbinary/gen/vadd-relu-scalar-x1.c",
284 "src/f32-vbinary/gen/vadd-relu-scalar-x2.c",
285 "src/f32-vbinary/gen/vadd-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700286 "src/f32-vbinary/gen/vadd-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700287 "src/f32-vbinary/gen/vadd-scalar-x1.c",
288 "src/f32-vbinary/gen/vadd-scalar-x2.c",
289 "src/f32-vbinary/gen/vadd-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700290 "src/f32-vbinary/gen/vadd-scalar-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700291 "src/f32-vbinary/gen/vaddc-minmax-scalar-x1.c",
292 "src/f32-vbinary/gen/vaddc-minmax-scalar-x2.c",
293 "src/f32-vbinary/gen/vaddc-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700294 "src/f32-vbinary/gen/vaddc-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700295 "src/f32-vbinary/gen/vaddc-relu-scalar-x1.c",
296 "src/f32-vbinary/gen/vaddc-relu-scalar-x2.c",
297 "src/f32-vbinary/gen/vaddc-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700298 "src/f32-vbinary/gen/vaddc-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700299 "src/f32-vbinary/gen/vaddc-scalar-x1.c",
300 "src/f32-vbinary/gen/vaddc-scalar-x2.c",
301 "src/f32-vbinary/gen/vaddc-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700302 "src/f32-vbinary/gen/vaddc-scalar-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700303 "src/f32-vbinary/gen/vdiv-minmax-scalar-x1.c",
304 "src/f32-vbinary/gen/vdiv-minmax-scalar-x2.c",
305 "src/f32-vbinary/gen/vdiv-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700306 "src/f32-vbinary/gen/vdiv-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700307 "src/f32-vbinary/gen/vdiv-relu-scalar-x1.c",
308 "src/f32-vbinary/gen/vdiv-relu-scalar-x2.c",
309 "src/f32-vbinary/gen/vdiv-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700310 "src/f32-vbinary/gen/vdiv-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700311 "src/f32-vbinary/gen/vdiv-scalar-x1.c",
312 "src/f32-vbinary/gen/vdiv-scalar-x2.c",
313 "src/f32-vbinary/gen/vdiv-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700314 "src/f32-vbinary/gen/vdiv-scalar-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700315 "src/f32-vbinary/gen/vdivc-minmax-scalar-x1.c",
316 "src/f32-vbinary/gen/vdivc-minmax-scalar-x2.c",
317 "src/f32-vbinary/gen/vdivc-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700318 "src/f32-vbinary/gen/vdivc-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700319 "src/f32-vbinary/gen/vdivc-relu-scalar-x1.c",
320 "src/f32-vbinary/gen/vdivc-relu-scalar-x2.c",
321 "src/f32-vbinary/gen/vdivc-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700322 "src/f32-vbinary/gen/vdivc-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700323 "src/f32-vbinary/gen/vdivc-scalar-x1.c",
324 "src/f32-vbinary/gen/vdivc-scalar-x2.c",
325 "src/f32-vbinary/gen/vdivc-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700326 "src/f32-vbinary/gen/vdivc-scalar-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800327 "src/f32-vbinary/gen/vmax-scalar-x1.c",
328 "src/f32-vbinary/gen/vmax-scalar-x2.c",
329 "src/f32-vbinary/gen/vmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700330 "src/f32-vbinary/gen/vmax-scalar-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800331 "src/f32-vbinary/gen/vmaxc-scalar-x1.c",
332 "src/f32-vbinary/gen/vmaxc-scalar-x2.c",
333 "src/f32-vbinary/gen/vmaxc-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700334 "src/f32-vbinary/gen/vmaxc-scalar-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800335 "src/f32-vbinary/gen/vmin-scalar-x1.c",
336 "src/f32-vbinary/gen/vmin-scalar-x2.c",
337 "src/f32-vbinary/gen/vmin-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700338 "src/f32-vbinary/gen/vmin-scalar-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800339 "src/f32-vbinary/gen/vminc-scalar-x1.c",
340 "src/f32-vbinary/gen/vminc-scalar-x2.c",
341 "src/f32-vbinary/gen/vminc-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700342 "src/f32-vbinary/gen/vminc-scalar-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700343 "src/f32-vbinary/gen/vmul-minmax-scalar-x1.c",
344 "src/f32-vbinary/gen/vmul-minmax-scalar-x2.c",
345 "src/f32-vbinary/gen/vmul-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700346 "src/f32-vbinary/gen/vmul-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700347 "src/f32-vbinary/gen/vmul-relu-scalar-x1.c",
348 "src/f32-vbinary/gen/vmul-relu-scalar-x2.c",
349 "src/f32-vbinary/gen/vmul-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700350 "src/f32-vbinary/gen/vmul-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700351 "src/f32-vbinary/gen/vmul-scalar-x1.c",
352 "src/f32-vbinary/gen/vmul-scalar-x2.c",
353 "src/f32-vbinary/gen/vmul-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700354 "src/f32-vbinary/gen/vmul-scalar-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700355 "src/f32-vbinary/gen/vmulc-minmax-scalar-x1.c",
356 "src/f32-vbinary/gen/vmulc-minmax-scalar-x2.c",
357 "src/f32-vbinary/gen/vmulc-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700358 "src/f32-vbinary/gen/vmulc-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700359 "src/f32-vbinary/gen/vmulc-relu-scalar-x1.c",
360 "src/f32-vbinary/gen/vmulc-relu-scalar-x2.c",
361 "src/f32-vbinary/gen/vmulc-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700362 "src/f32-vbinary/gen/vmulc-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700363 "src/f32-vbinary/gen/vmulc-scalar-x1.c",
364 "src/f32-vbinary/gen/vmulc-scalar-x2.c",
365 "src/f32-vbinary/gen/vmulc-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700366 "src/f32-vbinary/gen/vmulc-scalar-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700367 "src/f32-vbinary/gen/vrdivc-minmax-scalar-x1.c",
368 "src/f32-vbinary/gen/vrdivc-minmax-scalar-x2.c",
369 "src/f32-vbinary/gen/vrdivc-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700370 "src/f32-vbinary/gen/vrdivc-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700371 "src/f32-vbinary/gen/vrdivc-relu-scalar-x1.c",
372 "src/f32-vbinary/gen/vrdivc-relu-scalar-x2.c",
373 "src/f32-vbinary/gen/vrdivc-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700374 "src/f32-vbinary/gen/vrdivc-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700375 "src/f32-vbinary/gen/vrdivc-scalar-x1.c",
376 "src/f32-vbinary/gen/vrdivc-scalar-x2.c",
377 "src/f32-vbinary/gen/vrdivc-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700378 "src/f32-vbinary/gen/vrdivc-scalar-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700379 "src/f32-vbinary/gen/vrsubc-minmax-scalar-x1.c",
380 "src/f32-vbinary/gen/vrsubc-minmax-scalar-x2.c",
381 "src/f32-vbinary/gen/vrsubc-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700382 "src/f32-vbinary/gen/vrsubc-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700383 "src/f32-vbinary/gen/vrsubc-relu-scalar-x1.c",
384 "src/f32-vbinary/gen/vrsubc-relu-scalar-x2.c",
385 "src/f32-vbinary/gen/vrsubc-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700386 "src/f32-vbinary/gen/vrsubc-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700387 "src/f32-vbinary/gen/vrsubc-scalar-x1.c",
388 "src/f32-vbinary/gen/vrsubc-scalar-x2.c",
389 "src/f32-vbinary/gen/vrsubc-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700390 "src/f32-vbinary/gen/vrsubc-scalar-x8.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -0700391 "src/f32-vbinary/gen/vsqrdiff-scalar-x1.c",
392 "src/f32-vbinary/gen/vsqrdiff-scalar-x2.c",
393 "src/f32-vbinary/gen/vsqrdiff-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700394 "src/f32-vbinary/gen/vsqrdiff-scalar-x8.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -0700395 "src/f32-vbinary/gen/vsqrdiffc-scalar-x1.c",
396 "src/f32-vbinary/gen/vsqrdiffc-scalar-x2.c",
397 "src/f32-vbinary/gen/vsqrdiffc-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700398 "src/f32-vbinary/gen/vsqrdiffc-scalar-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700399 "src/f32-vbinary/gen/vsub-minmax-scalar-x1.c",
400 "src/f32-vbinary/gen/vsub-minmax-scalar-x2.c",
401 "src/f32-vbinary/gen/vsub-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700402 "src/f32-vbinary/gen/vsub-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700403 "src/f32-vbinary/gen/vsub-relu-scalar-x1.c",
404 "src/f32-vbinary/gen/vsub-relu-scalar-x2.c",
405 "src/f32-vbinary/gen/vsub-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700406 "src/f32-vbinary/gen/vsub-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700407 "src/f32-vbinary/gen/vsub-scalar-x1.c",
408 "src/f32-vbinary/gen/vsub-scalar-x2.c",
409 "src/f32-vbinary/gen/vsub-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700410 "src/f32-vbinary/gen/vsub-scalar-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700411 "src/f32-vbinary/gen/vsubc-minmax-scalar-x1.c",
412 "src/f32-vbinary/gen/vsubc-minmax-scalar-x2.c",
413 "src/f32-vbinary/gen/vsubc-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700414 "src/f32-vbinary/gen/vsubc-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700415 "src/f32-vbinary/gen/vsubc-relu-scalar-x1.c",
416 "src/f32-vbinary/gen/vsubc-relu-scalar-x2.c",
417 "src/f32-vbinary/gen/vsubc-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700418 "src/f32-vbinary/gen/vsubc-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700419 "src/f32-vbinary/gen/vsubc-scalar-x1.c",
420 "src/f32-vbinary/gen/vsubc-scalar-x2.c",
421 "src/f32-vbinary/gen/vsubc-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700422 "src/f32-vbinary/gen/vsubc-scalar-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700423 "src/f32-vlrelu/gen/vlrelu-scalar-x1.c",
424 "src/f32-vlrelu/gen/vlrelu-scalar-x2.c",
425 "src/f32-vlrelu/gen/vlrelu-scalar-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700426 "src/f32-vmulcaddc/gen/c1-minmax-scalar-2x.c",
427 "src/f32-vmulcaddc/gen/c2-minmax-scalar-2x.c",
428 "src/f32-vmulcaddc/gen/c4-minmax-scalar-2x.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -0700429 "src/f32-vrnd/gen/vrndd-scalar-libm-x1.c",
430 "src/f32-vrnd/gen/vrndd-scalar-libm-x2.c",
431 "src/f32-vrnd/gen/vrndd-scalar-libm-x4.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700432 "src/f32-vrnd/gen/vrndne-scalar-libm-x1.c",
433 "src/f32-vrnd/gen/vrndne-scalar-libm-x2.c",
434 "src/f32-vrnd/gen/vrndne-scalar-libm-x4.c",
435 "src/f32-vrnd/gen/vrndu-scalar-libm-x1.c",
436 "src/f32-vrnd/gen/vrndu-scalar-libm-x2.c",
437 "src/f32-vrnd/gen/vrndu-scalar-libm-x4.c",
438 "src/f32-vrnd/gen/vrndz-scalar-libm-x1.c",
439 "src/f32-vrnd/gen/vrndz-scalar-libm-x2.c",
440 "src/f32-vrnd/gen/vrndz-scalar-libm-x4.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -0700441 "src/f32-vsqrt/gen/scalar-sqrt-x1.c",
442 "src/f32-vsqrt/gen/scalar-sqrt-x2.c",
443 "src/f32-vsqrt/gen/scalar-sqrt-x4.c",
Marat Dukhan5020b962020-06-08 13:30:10 -0700444 "src/f32-vunary/gen/vabs-scalar-x1.c",
445 "src/f32-vunary/gen/vabs-scalar-x2.c",
446 "src/f32-vunary/gen/vabs-scalar-x4.c",
447 "src/f32-vunary/gen/vneg-scalar-x1.c",
448 "src/f32-vunary/gen/vneg-scalar-x2.c",
449 "src/f32-vunary/gen/vneg-scalar-x4.c",
450 "src/f32-vunary/gen/vsqr-scalar-x1.c",
451 "src/f32-vunary/gen/vsqr-scalar-x2.c",
452 "src/f32-vunary/gen/vsqr-scalar-x4.c",
Marat Dukhanc60742b2020-11-23 12:33:27 -0800453 "src/math/expm1minus-scalar-rr2-lut16-p3.c",
454 "src/math/expm1minus-scalar-rr2-lut16-p4.c",
455 "src/math/expm1minus-scalar-rr2-p5.c",
456 "src/math/expm1minus-scalar-rr2-p6.c",
Frank Barchard22136062020-11-24 18:44:46 -0800457 "src/math/expminus-scalar-rr2-lut64-p2.c",
458 "src/math/expminus-scalar-rr2-lut2048-p1.c",
459 "src/math/expminus-scalar-rr2-p5.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -0700460 "src/math/roundd-scalar-addsub.c",
461 "src/math/roundd-scalar-cvt.c",
Marat Dukhanffbf96a2020-05-14 02:59:08 -0700462 "src/math/roundd-scalar-floor.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700463 "src/math/roundne-scalar-addsub.c",
464 "src/math/roundne-scalar-nearbyint.c",
465 "src/math/roundne-scalar-rint.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -0700466 "src/math/roundu-scalar-addsub.c",
Marat Dukhanffbf96a2020-05-14 02:59:08 -0700467 "src/math/roundu-scalar-ceil.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -0700468 "src/math/roundu-scalar-cvt.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -0700469 "src/math/roundz-scalar-addsub.c",
470 "src/math/roundz-scalar-cvt.c",
Marat Dukhanffbf96a2020-05-14 02:59:08 -0700471 "src/math/roundz-scalar-trunc.c",
Marat Dukhanf8475d62020-09-17 15:01:43 -0700472 "src/math/sigmoid-scalar-rr2-lut64-p2-div.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700473 "src/math/sigmoid-scalar-rr2-lut2048-p1-div.c",
Marat Dukhanf8475d62020-09-17 15:01:43 -0700474 "src/math/sigmoid-scalar-rr2-p5-div.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700475 "src/qs8-gemm/gen/8x8c4-minmax-scalar.c",
476 "src/qs8-gemm/gen/12x4c4-minmax-scalar.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -0700477 "src/qs8-requantization/fp32-scalar-lrintf.c",
478 "src/qs8-requantization/fp32-scalar-magic.c",
479 "src/qs8-requantization/precise-scalar-signed64.c",
480 "src/qs8-requantization/precise-scalar-unsigned32.c",
481 "src/qs8-requantization/precise-scalar-unsigned64.c",
482 "src/qs8-requantization/q31-scalar.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -0700483 "src/qu8-avgpool/9p8x-minmax-scalar-c1.c",
484 "src/qu8-avgpool/9x-minmax-scalar-c1.c",
485 "src/qu8-dwconv/up1x9-minmax-scalar.c",
486 "src/qu8-gavgpool/7p7x-minmax-scalar-c1.c",
487 "src/qu8-gavgpool/7x-minmax-scalar-c1.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700488 "src/qu8-gemm/2x2-minmax-scalar.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700489 "src/qu8-gemm/gen/8x8c4-minmax-scalar.c",
490 "src/qu8-gemm/gen/12x4c4-minmax-scalar.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -0700491 "src/qu8-igemm/2x2-minmax-scalar.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -0700492 "src/qu8-requantization/fp32-scalar-lrintf.c",
493 "src/qu8-requantization/fp32-scalar-magic.c",
494 "src/qu8-requantization/precise-scalar-signed64.c",
495 "src/qu8-requantization/precise-scalar-unsigned32.c",
496 "src/qu8-requantization/precise-scalar-unsigned64.c",
497 "src/qu8-requantization/q31-scalar.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -0700498 "src/qu8-vadd/minmax-scalar.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -0700499 "src/u8-clamp/scalar-x4.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700500 "src/u8-lut32norm/scalar.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700501 "src/u8-maxpool/9p8x-minmax-scalar-c1.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700502 "src/u8-rmax/scalar.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700503 "src/x8-lut/scalar.c",
504 "src/x8-zip/x2-scalar.c",
505 "src/x8-zip/x3-scalar.c",
506 "src/x8-zip/x4-scalar.c",
507 "src/x8-zip/xm-scalar.c",
Marat Dukhanad71b9a2020-11-20 00:01:51 -0800508 "src/x32-depthtospace2d-chw2hwc/scalar.c",
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -0700509 "src/x32-fill/scalar-float.c",
510 "src/x32-fill/scalar-int.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700511 "src/x32-packx/x2-scalar.c",
512 "src/x32-packx/x3-scalar.c",
513 "src/x32-packx/x4-scalar.c",
Marat Dukhan63523d42020-05-22 17:07:33 -0700514 "src/x32-pad/scalar-float.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700515 "src/x32-pad/scalar-int.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700516 "src/x32-unpool/scalar.c",
517 "src/x32-zip/x2-scalar.c",
518 "src/x32-zip/x3-scalar.c",
519 "src/x32-zip/x4-scalar.c",
520 "src/x32-zip/xm-scalar.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700521]
522
Marat Dukhan436ebe62019-12-04 15:10:12 -0800523WASM_UKERNELS = [
Marat Dukhan99936602020-04-11 16:47:01 -0700524 "src/f32-avgpool/9p8x-minmax-wasm-c1.c",
525 "src/f32-avgpool/9x-minmax-wasm-c1.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -0700526 "src/f32-clamp/gen/wasm-x1.c",
527 "src/f32-clamp/gen/wasm-x2.c",
528 "src/f32-clamp/gen/wasm-x4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700529 "src/f32-dwconv/gen/up1x4-minmax-wasm-acc2.c",
530 "src/f32-dwconv/gen/up1x4-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700531 "src/f32-dwconv/gen/up1x4-wasm-acc2.c",
532 "src/f32-dwconv/gen/up1x4-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700533 "src/f32-dwconv/gen/up1x9-minmax-wasm-acc2.c",
534 "src/f32-dwconv/gen/up1x9-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700535 "src/f32-dwconv/gen/up1x9-wasm-acc2.c",
536 "src/f32-dwconv/gen/up1x9-wasm.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700537 "src/f32-dwconv/gen/up1x25-minmax-wasm-acc2.c",
538 "src/f32-dwconv/gen/up1x25-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700539 "src/f32-dwconv/gen/up1x25-wasm-acc2.c",
540 "src/f32-dwconv/gen/up1x25-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700541 "src/f32-dwconv/gen/up2x4-minmax-wasm-acc2.c",
542 "src/f32-dwconv/gen/up2x4-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700543 "src/f32-dwconv/gen/up2x4-wasm-acc2.c",
544 "src/f32-dwconv/gen/up2x4-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700545 "src/f32-dwconv/gen/up2x9-minmax-wasm-acc2.c",
546 "src/f32-dwconv/gen/up2x9-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700547 "src/f32-dwconv/gen/up2x9-wasm-acc2.c",
548 "src/f32-dwconv/gen/up2x9-wasm.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700549 "src/f32-dwconv/gen/up2x25-minmax-wasm-acc2.c",
550 "src/f32-dwconv/gen/up2x25-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700551 "src/f32-dwconv/gen/up2x25-wasm-acc2.c",
552 "src/f32-dwconv/gen/up2x25-wasm.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700553 "src/f32-gavgpool/7p7x-minmax-wasm-c1.c",
554 "src/f32-gavgpool/7x-minmax-wasm-c1.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700555 "src/f32-gemm/gen-inc/1x4inc-minmax-wasm.c",
556 "src/f32-gemm/gen-inc/2x4inc-minmax-wasm.c",
557 "src/f32-gemm/gen-inc/4x4inc-minmax-wasm.c",
558 "src/f32-gemm/gen/1x4-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700559 "src/f32-gemm/gen/1x4-relu-wasm.c",
560 "src/f32-gemm/gen/1x4-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700561 "src/f32-gemm/gen/2x4-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700562 "src/f32-gemm/gen/2x4-relu-wasm.c",
563 "src/f32-gemm/gen/2x4-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700564 "src/f32-gemm/gen/4x2-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700565 "src/f32-gemm/gen/4x2-relu-wasm.c",
566 "src/f32-gemm/gen/4x2-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700567 "src/f32-gemm/gen/4x4-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700568 "src/f32-gemm/gen/4x4-relu-wasm.c",
569 "src/f32-gemm/gen/4x4-wasm.c",
Marat Dukhanefaac272020-07-16 16:58:07 -0700570 "src/f32-hswish/gen/hswish-wasm-x1.c",
571 "src/f32-hswish/gen/hswish-wasm-x2.c",
572 "src/f32-hswish/gen/hswish-wasm-x4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700573 "src/f32-igemm/gen/1x4-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700574 "src/f32-igemm/gen/1x4-relu-wasm.c",
575 "src/f32-igemm/gen/1x4-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700576 "src/f32-igemm/gen/2x4-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700577 "src/f32-igemm/gen/2x4-relu-wasm.c",
578 "src/f32-igemm/gen/2x4-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700579 "src/f32-igemm/gen/4x2-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700580 "src/f32-igemm/gen/4x2-relu-wasm.c",
581 "src/f32-igemm/gen/4x2-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700582 "src/f32-igemm/gen/4x4-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700583 "src/f32-igemm/gen/4x4-relu-wasm.c",
584 "src/f32-igemm/gen/4x4-wasm.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700585 "src/f32-maxpool/9p8x-minmax-wasm-c1.c",
586 "src/f32-pavgpool/9p8x-minmax-wasm-c1.c",
587 "src/f32-pavgpool/9x-minmax-wasm-c1.c",
Marat Dukhan7c1f8082020-06-25 13:26:20 -0700588 "src/f32-prelu/gen/wasm-2x1.c",
589 "src/f32-prelu/gen/wasm-2x4.c",
Frank Barchardfb158e22020-07-15 16:10:10 -0700590 "src/f32-relu/gen/wasm-x1.c",
591 "src/f32-relu/gen/wasm-x2.c",
592 "src/f32-relu/gen/wasm-x4.c",
Frank Barchard37297a62020-08-27 12:15:10 -0700593 "src/f32-relu/gen/wasm-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700594 "src/f32-vbinary/gen/vadd-minmax-wasm-x1.c",
595 "src/f32-vbinary/gen/vadd-minmax-wasm-x2.c",
596 "src/f32-vbinary/gen/vadd-minmax-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700597 "src/f32-vbinary/gen/vadd-minmax-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700598 "src/f32-vbinary/gen/vadd-relu-wasm-x1.c",
599 "src/f32-vbinary/gen/vadd-relu-wasm-x2.c",
600 "src/f32-vbinary/gen/vadd-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700601 "src/f32-vbinary/gen/vadd-relu-wasm-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700602 "src/f32-vbinary/gen/vaddc-minmax-wasm-x1.c",
603 "src/f32-vbinary/gen/vaddc-minmax-wasm-x2.c",
604 "src/f32-vbinary/gen/vaddc-minmax-wasm-x4.c",
605 "src/f32-vbinary/gen/vaddc-minmax-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700606 "src/f32-vbinary/gen/vaddc-relu-wasm-x1.c",
607 "src/f32-vbinary/gen/vaddc-relu-wasm-x2.c",
608 "src/f32-vbinary/gen/vaddc-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700609 "src/f32-vbinary/gen/vaddc-relu-wasm-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700610 "src/f32-vbinary/gen/vdiv-minmax-wasm-x1.c",
611 "src/f32-vbinary/gen/vdiv-minmax-wasm-x2.c",
612 "src/f32-vbinary/gen/vdiv-minmax-wasm-x4.c",
613 "src/f32-vbinary/gen/vdiv-minmax-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700614 "src/f32-vbinary/gen/vdiv-relu-wasm-x1.c",
615 "src/f32-vbinary/gen/vdiv-relu-wasm-x2.c",
616 "src/f32-vbinary/gen/vdiv-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700617 "src/f32-vbinary/gen/vdiv-relu-wasm-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700618 "src/f32-vbinary/gen/vdivc-minmax-wasm-x1.c",
619 "src/f32-vbinary/gen/vdivc-minmax-wasm-x2.c",
620 "src/f32-vbinary/gen/vdivc-minmax-wasm-x4.c",
621 "src/f32-vbinary/gen/vdivc-minmax-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700622 "src/f32-vbinary/gen/vdivc-relu-wasm-x1.c",
623 "src/f32-vbinary/gen/vdivc-relu-wasm-x2.c",
624 "src/f32-vbinary/gen/vdivc-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700625 "src/f32-vbinary/gen/vdivc-relu-wasm-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800626 "src/f32-vbinary/gen/vmax-wasm-x1.c",
627 "src/f32-vbinary/gen/vmax-wasm-x2.c",
628 "src/f32-vbinary/gen/vmax-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700629 "src/f32-vbinary/gen/vmax-wasm-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800630 "src/f32-vbinary/gen/vmaxc-wasm-x1.c",
631 "src/f32-vbinary/gen/vmaxc-wasm-x2.c",
632 "src/f32-vbinary/gen/vmaxc-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700633 "src/f32-vbinary/gen/vmaxc-wasm-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800634 "src/f32-vbinary/gen/vmin-wasm-x1.c",
635 "src/f32-vbinary/gen/vmin-wasm-x2.c",
636 "src/f32-vbinary/gen/vmin-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700637 "src/f32-vbinary/gen/vmin-wasm-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800638 "src/f32-vbinary/gen/vminc-wasm-x1.c",
639 "src/f32-vbinary/gen/vminc-wasm-x2.c",
640 "src/f32-vbinary/gen/vminc-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700641 "src/f32-vbinary/gen/vminc-wasm-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700642 "src/f32-vbinary/gen/vmul-minmax-wasm-x1.c",
643 "src/f32-vbinary/gen/vmul-minmax-wasm-x2.c",
644 "src/f32-vbinary/gen/vmul-minmax-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700645 "src/f32-vbinary/gen/vmul-minmax-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700646 "src/f32-vbinary/gen/vmul-relu-wasm-x1.c",
647 "src/f32-vbinary/gen/vmul-relu-wasm-x2.c",
648 "src/f32-vbinary/gen/vmul-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700649 "src/f32-vbinary/gen/vmul-relu-wasm-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700650 "src/f32-vbinary/gen/vmulc-minmax-wasm-x1.c",
651 "src/f32-vbinary/gen/vmulc-minmax-wasm-x2.c",
652 "src/f32-vbinary/gen/vmulc-minmax-wasm-x4.c",
653 "src/f32-vbinary/gen/vmulc-minmax-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700654 "src/f32-vbinary/gen/vmulc-relu-wasm-x1.c",
655 "src/f32-vbinary/gen/vmulc-relu-wasm-x2.c",
656 "src/f32-vbinary/gen/vmulc-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700657 "src/f32-vbinary/gen/vmulc-relu-wasm-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700658 "src/f32-vbinary/gen/vrdivc-minmax-wasm-x1.c",
659 "src/f32-vbinary/gen/vrdivc-minmax-wasm-x2.c",
660 "src/f32-vbinary/gen/vrdivc-minmax-wasm-x4.c",
661 "src/f32-vbinary/gen/vrdivc-minmax-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700662 "src/f32-vbinary/gen/vrdivc-relu-wasm-x1.c",
663 "src/f32-vbinary/gen/vrdivc-relu-wasm-x2.c",
664 "src/f32-vbinary/gen/vrdivc-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700665 "src/f32-vbinary/gen/vrdivc-relu-wasm-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700666 "src/f32-vbinary/gen/vrsubc-minmax-wasm-x1.c",
667 "src/f32-vbinary/gen/vrsubc-minmax-wasm-x2.c",
668 "src/f32-vbinary/gen/vrsubc-minmax-wasm-x4.c",
669 "src/f32-vbinary/gen/vrsubc-minmax-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700670 "src/f32-vbinary/gen/vrsubc-relu-wasm-x1.c",
671 "src/f32-vbinary/gen/vrsubc-relu-wasm-x2.c",
672 "src/f32-vbinary/gen/vrsubc-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700673 "src/f32-vbinary/gen/vrsubc-relu-wasm-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700674 "src/f32-vbinary/gen/vsub-minmax-wasm-x1.c",
675 "src/f32-vbinary/gen/vsub-minmax-wasm-x2.c",
676 "src/f32-vbinary/gen/vsub-minmax-wasm-x4.c",
677 "src/f32-vbinary/gen/vsub-minmax-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700678 "src/f32-vbinary/gen/vsub-relu-wasm-x1.c",
679 "src/f32-vbinary/gen/vsub-relu-wasm-x2.c",
680 "src/f32-vbinary/gen/vsub-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700681 "src/f32-vbinary/gen/vsub-relu-wasm-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700682 "src/f32-vbinary/gen/vsubc-minmax-wasm-x1.c",
683 "src/f32-vbinary/gen/vsubc-minmax-wasm-x2.c",
684 "src/f32-vbinary/gen/vsubc-minmax-wasm-x4.c",
685 "src/f32-vbinary/gen/vsubc-minmax-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700686 "src/f32-vbinary/gen/vsubc-relu-wasm-x1.c",
687 "src/f32-vbinary/gen/vsubc-relu-wasm-x2.c",
688 "src/f32-vbinary/gen/vsubc-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700689 "src/f32-vbinary/gen/vsubc-relu-wasm-x8.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -0700690 "src/f32-vlrelu/gen/vlrelu-wasm-x1.c",
691 "src/f32-vlrelu/gen/vlrelu-wasm-x2.c",
692 "src/f32-vlrelu/gen/vlrelu-wasm-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700693 "src/f32-vmulcaddc/gen/c1-minmax-wasm-2x.c",
694 "src/f32-vmulcaddc/gen/c2-minmax-wasm-2x.c",
695 "src/f32-vmulcaddc/gen/c4-minmax-wasm-2x.c",
Marat Dukhan436ebe62019-12-04 15:10:12 -0800696]
697
Marat Dukhan290055c2020-06-09 12:24:29 -0700698WASMSIMD_UKERNELS = [
Marat Dukhan40f05522020-07-16 22:33:12 -0700699 "src/f32-argmaxpool/4x-wasmsimd-c4.c",
700 "src/f32-argmaxpool/9p8x-wasmsimd-c4.c",
701 "src/f32-argmaxpool/9x-wasmsimd-c4.c",
Marat Dukhan3b7432d2020-07-16 17:46:32 -0700702 "src/f32-avgpool/9p8x-minmax-wasmsimd-arm-c4.c",
703 "src/f32-avgpool/9p8x-minmax-wasmsimd-x86-c4.c",
704 "src/f32-avgpool/9x-minmax-wasmsimd-arm-c4.c",
705 "src/f32-avgpool/9x-minmax-wasmsimd-x86-c4.c",
Marat Dukhan3fa52c82020-07-08 12:54:33 -0700706 "src/f32-clamp/gen/wasmsimd-arm-x4.c",
707 "src/f32-clamp/gen/wasmsimd-arm-x8.c",
708 "src/f32-clamp/gen/wasmsimd-x86-x4.c",
709 "src/f32-clamp/gen/wasmsimd-x86-x8.c",
Frank Barchard22136062020-11-24 18:44:46 -0800710 "src/f32-conv-hwc2chw/3x3s2p1c3x4-wasmsimd-2x2.c",
Marat Dukhanac014d72020-06-16 08:36:47 -0700711 "src/f32-dwconv/gen/up4x4-minmax-wasmsimd-acc2-arm.c",
Marat Dukhanac014d72020-06-16 08:36:47 -0700712 "src/f32-dwconv/gen/up4x4-minmax-wasmsimd-acc2-x86.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700713 "src/f32-dwconv/gen/up4x4-minmax-wasmsimd-arm.c",
Marat Dukhanac014d72020-06-16 08:36:47 -0700714 "src/f32-dwconv/gen/up4x4-minmax-wasmsimd-x86.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700715 "src/f32-dwconv/gen/up4x4-wasmsimd.c",
716 "src/f32-dwconv/gen/up4x9-minmax-wasmsimd-acc2-arm.c",
Marat Dukhanac014d72020-06-16 08:36:47 -0700717 "src/f32-dwconv/gen/up4x9-minmax-wasmsimd-acc2-x86.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700718 "src/f32-dwconv/gen/up4x9-minmax-wasmsimd-arm.c",
Marat Dukhanac014d72020-06-16 08:36:47 -0700719 "src/f32-dwconv/gen/up4x9-minmax-wasmsimd-x86.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700720 "src/f32-dwconv/gen/up4x9-wasmsimd.c",
721 "src/f32-dwconv/gen/up4x25-minmax-wasmsimd-acc2-arm.c",
722 "src/f32-dwconv/gen/up4x25-minmax-wasmsimd-acc2-x86.c",
723 "src/f32-dwconv/gen/up4x25-minmax-wasmsimd-arm.c",
724 "src/f32-dwconv/gen/up4x25-minmax-wasmsimd-x86.c",
725 "src/f32-dwconv/gen/up4x25-wasmsimd.c",
726 "src/f32-dwconv/gen/up8x4-minmax-wasmsimd-acc2-arm.c",
Marat Dukhanac014d72020-06-16 08:36:47 -0700727 "src/f32-dwconv/gen/up8x4-minmax-wasmsimd-acc2-x86.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700728 "src/f32-dwconv/gen/up8x4-minmax-wasmsimd-arm.c",
Marat Dukhanac014d72020-06-16 08:36:47 -0700729 "src/f32-dwconv/gen/up8x4-minmax-wasmsimd-x86.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700730 "src/f32-dwconv/gen/up8x4-wasmsimd.c",
731 "src/f32-dwconv/gen/up8x9-minmax-wasmsimd-acc2-arm.c",
Marat Dukhanac014d72020-06-16 08:36:47 -0700732 "src/f32-dwconv/gen/up8x9-minmax-wasmsimd-acc2-x86.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700733 "src/f32-dwconv/gen/up8x9-minmax-wasmsimd-arm.c",
Marat Dukhanac014d72020-06-16 08:36:47 -0700734 "src/f32-dwconv/gen/up8x9-minmax-wasmsimd-x86.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700735 "src/f32-dwconv/gen/up8x9-wasmsimd.c",
736 "src/f32-dwconv/gen/up8x25-minmax-wasmsimd-acc2-arm.c",
737 "src/f32-dwconv/gen/up8x25-minmax-wasmsimd-acc2-x86.c",
738 "src/f32-dwconv/gen/up8x25-minmax-wasmsimd-arm.c",
739 "src/f32-dwconv/gen/up8x25-minmax-wasmsimd-x86.c",
740 "src/f32-dwconv/gen/up8x25-wasmsimd.c",
Frank Barchard3b800452020-11-22 12:12:35 -0800741 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-1x4-acc2.c",
742 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-1x4-acc3.c",
743 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-1x4-acc4.c",
744 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-1x4.c",
745 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-2x4-acc2.c",
746 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-2x4.c",
747 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-3x4.c",
748 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-4x4.c",
749 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-5x4.c",
750 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-6x4.c",
751 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-1x4-acc2.c",
752 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-1x4-acc3.c",
753 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-1x4-acc4.c",
754 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-1x4.c",
755 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-2x4-acc2.c",
756 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-2x4.c",
757 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-3x4.c",
758 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-4x4.c",
759 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-5x4.c",
760 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-6x4.c",
Frank Barchard9cef5ea2020-11-18 14:52:08 -0800761 "src/f32-dwconv2d-chw/gen/3x3s2p1-wasmsimd-1x4-acc3-arm.c",
Frank Barcharddb5c32d2020-11-16 23:58:42 -0800762 "src/f32-dwconv2d-chw/gen/3x3s2p1-wasmsimd-1x4-acc3-x86.c",
Frank Barchard9cef5ea2020-11-18 14:52:08 -0800763 "src/f32-dwconv2d-chw/gen/5x5p2-wasmsimd-3x4-arm.c",
Frank Barcharddb5c32d2020-11-16 23:58:42 -0800764 "src/f32-dwconv2d-chw/gen/5x5p2-wasmsimd-3x4-x86.c",
Frank Barchard9cef5ea2020-11-18 14:52:08 -0800765 "src/f32-dwconv2d-chw/gen/5x5s2p2-wasmsimd-1x4-acc2-arm.c",
Frank Barcharddb5c32d2020-11-16 23:58:42 -0800766 "src/f32-dwconv2d-chw/gen/5x5s2p2-wasmsimd-1x4-acc2-x86.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700767 "src/f32-gavgpool-cw/wasmsimd-arm-x4.c",
768 "src/f32-gavgpool-cw/wasmsimd-x86-x4.c",
Marat Dukhanc6016802020-07-16 18:51:28 -0700769 "src/f32-gavgpool/7p7x-minmax-wasmsimd-arm-c4.c",
770 "src/f32-gavgpool/7p7x-minmax-wasmsimd-x86-c4.c",
771 "src/f32-gavgpool/7x-minmax-wasmsimd-arm-c4.c",
772 "src/f32-gavgpool/7x-minmax-wasmsimd-x86-c4.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700773 "src/f32-gemm/gen-inc/1x8inc-minmax-wasmsimd-loadsplat-arm.c",
774 "src/f32-gemm/gen-inc/1x8inc-minmax-wasmsimd-loadsplat-x86.c",
775 "src/f32-gemm/gen-inc/1x8inc-minmax-wasmsimd-splat-arm.c",
776 "src/f32-gemm/gen-inc/1x8inc-minmax-wasmsimd-splat-x86.c",
777 "src/f32-gemm/gen-inc/1x8s4inc-minmax-wasmsimd-arm.c",
778 "src/f32-gemm/gen-inc/1x8s4inc-minmax-wasmsimd-x86.c",
779 "src/f32-gemm/gen-inc/3x8inc-minmax-wasmsimd-loadsplat-arm.c",
780 "src/f32-gemm/gen-inc/3x8inc-minmax-wasmsimd-loadsplat-x86.c",
781 "src/f32-gemm/gen-inc/3x8inc-minmax-wasmsimd-splat-arm.c",
782 "src/f32-gemm/gen-inc/3x8inc-minmax-wasmsimd-splat-x86.c",
783 "src/f32-gemm/gen-inc/3x8s4inc-minmax-wasmsimd-arm.c",
784 "src/f32-gemm/gen-inc/3x8s4inc-minmax-wasmsimd-x86.c",
785 "src/f32-gemm/gen-inc/4x8inc-minmax-wasmsimd-loadsplat-arm.c",
786 "src/f32-gemm/gen-inc/4x8inc-minmax-wasmsimd-loadsplat-x86.c",
787 "src/f32-gemm/gen-inc/4x8inc-minmax-wasmsimd-splat-arm.c",
788 "src/f32-gemm/gen-inc/4x8inc-minmax-wasmsimd-splat-x86.c",
789 "src/f32-gemm/gen-inc/4x8s4inc-minmax-wasmsimd-arm.c",
790 "src/f32-gemm/gen-inc/4x8s4inc-minmax-wasmsimd-x86.c",
791 "src/f32-gemm/gen-inc/5x8inc-minmax-wasmsimd-loadsplat-arm.c",
792 "src/f32-gemm/gen-inc/5x8inc-minmax-wasmsimd-loadsplat-x86.c",
793 "src/f32-gemm/gen-inc/5x8inc-minmax-wasmsimd-splat-arm.c",
794 "src/f32-gemm/gen-inc/5x8inc-minmax-wasmsimd-splat-x86.c",
795 "src/f32-gemm/gen-inc/5x8s4inc-minmax-wasmsimd-arm.c",
796 "src/f32-gemm/gen-inc/5x8s4inc-minmax-wasmsimd-x86.c",
797 "src/f32-gemm/gen-inc/6x8inc-minmax-wasmsimd-loadsplat-arm.c",
798 "src/f32-gemm/gen-inc/6x8inc-minmax-wasmsimd-loadsplat-x86.c",
799 "src/f32-gemm/gen-inc/6x8inc-minmax-wasmsimd-splat-arm.c",
800 "src/f32-gemm/gen-inc/6x8inc-minmax-wasmsimd-splat-x86.c",
801 "src/f32-gemm/gen-inc/6x8s4inc-minmax-wasmsimd-arm.c",
802 "src/f32-gemm/gen-inc/6x8s4inc-minmax-wasmsimd-x86.c",
803 "src/f32-gemm/gen/1x8-minmax-wasmsimd-loadsplat-arm.c",
804 "src/f32-gemm/gen/1x8-minmax-wasmsimd-loadsplat-x86.c",
805 "src/f32-gemm/gen/1x8-minmax-wasmsimd-splat-arm.c",
806 "src/f32-gemm/gen/1x8-minmax-wasmsimd-splat-x86.c",
807 "src/f32-gemm/gen/1x8-relu-wasmsimd-splat.c",
808 "src/f32-gemm/gen/1x8-wasmsimd-splat.c",
809 "src/f32-gemm/gen/1x8s4-minmax-wasmsimd-arm.c",
810 "src/f32-gemm/gen/1x8s4-minmax-wasmsimd-x86.c",
811 "src/f32-gemm/gen/3x8-minmax-wasmsimd-loadsplat-arm.c",
812 "src/f32-gemm/gen/3x8-minmax-wasmsimd-loadsplat-x86.c",
813 "src/f32-gemm/gen/3x8-minmax-wasmsimd-splat-arm.c",
814 "src/f32-gemm/gen/3x8-minmax-wasmsimd-splat-x86.c",
815 "src/f32-gemm/gen/3x8s4-minmax-wasmsimd-arm.c",
816 "src/f32-gemm/gen/3x8s4-minmax-wasmsimd-x86.c",
817 "src/f32-gemm/gen/4x2c4-minmax-wasmsimd-arm.c",
818 "src/f32-gemm/gen/4x2c4-minmax-wasmsimd-x86.c",
819 "src/f32-gemm/gen/4x2c4-relu-wasmsimd.c",
820 "src/f32-gemm/gen/4x2c4-wasmsimd.c",
821 "src/f32-gemm/gen/4x8-minmax-wasmsimd-loadsplat-arm.c",
822 "src/f32-gemm/gen/4x8-minmax-wasmsimd-loadsplat-x86.c",
823 "src/f32-gemm/gen/4x8-minmax-wasmsimd-splat-arm.c",
824 "src/f32-gemm/gen/4x8-minmax-wasmsimd-splat-x86.c",
825 "src/f32-gemm/gen/4x8-relu-wasmsimd-splat.c",
826 "src/f32-gemm/gen/4x8-wasmsimd-splat.c",
827 "src/f32-gemm/gen/4x8s4-minmax-wasmsimd-arm.c",
828 "src/f32-gemm/gen/4x8s4-minmax-wasmsimd-x86.c",
829 "src/f32-gemm/gen/5x8-minmax-wasmsimd-loadsplat-arm.c",
830 "src/f32-gemm/gen/5x8-minmax-wasmsimd-loadsplat-x86.c",
831 "src/f32-gemm/gen/5x8-minmax-wasmsimd-splat-arm.c",
832 "src/f32-gemm/gen/5x8-minmax-wasmsimd-splat-x86.c",
833 "src/f32-gemm/gen/5x8-relu-wasmsimd-splat.c",
834 "src/f32-gemm/gen/5x8-wasmsimd-splat.c",
835 "src/f32-gemm/gen/5x8s4-minmax-wasmsimd-arm.c",
836 "src/f32-gemm/gen/5x8s4-minmax-wasmsimd-x86.c",
837 "src/f32-gemm/gen/6x8-minmax-wasmsimd-loadsplat-arm.c",
838 "src/f32-gemm/gen/6x8-minmax-wasmsimd-loadsplat-x86.c",
839 "src/f32-gemm/gen/6x8-minmax-wasmsimd-splat-arm.c",
840 "src/f32-gemm/gen/6x8-minmax-wasmsimd-splat-x86.c",
841 "src/f32-gemm/gen/6x8s4-minmax-wasmsimd-arm.c",
842 "src/f32-gemm/gen/6x8s4-minmax-wasmsimd-x86.c",
Marat Dukhanefaac272020-07-16 16:58:07 -0700843 "src/f32-hswish/gen/hswish-wasmsimd-x4.c",
844 "src/f32-hswish/gen/hswish-wasmsimd-x8.c",
845 "src/f32-hswish/gen/hswish-wasmsimd-x16.c",
XNNPACK Team965272b2020-10-23 21:10:15 -0700846 "src/f32-ibilinear-chw/gen/wasmsimd-p4.c",
847 "src/f32-ibilinear-chw/gen/wasmsimd-p8.c",
Frank Barchard35db7d02020-10-26 13:37:34 -0700848 "src/f32-ibilinear/gen/wasmsimd-c4.c",
849 "src/f32-ibilinear/gen/wasmsimd-c8.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700850 "src/f32-igemm/gen/1x8-minmax-wasmsimd-loadsplat-arm.c",
851 "src/f32-igemm/gen/1x8-minmax-wasmsimd-loadsplat-x86.c",
852 "src/f32-igemm/gen/1x8-minmax-wasmsimd-splat-arm.c",
853 "src/f32-igemm/gen/1x8-minmax-wasmsimd-splat-x86.c",
854 "src/f32-igemm/gen/1x8-relu-wasmsimd-splat.c",
855 "src/f32-igemm/gen/1x8-wasmsimd-splat.c",
856 "src/f32-igemm/gen/1x8s4-minmax-wasmsimd-arm.c",
857 "src/f32-igemm/gen/1x8s4-minmax-wasmsimd-x86.c",
858 "src/f32-igemm/gen/3x8-minmax-wasmsimd-loadsplat-arm.c",
859 "src/f32-igemm/gen/3x8-minmax-wasmsimd-loadsplat-x86.c",
860 "src/f32-igemm/gen/3x8-minmax-wasmsimd-splat-arm.c",
861 "src/f32-igemm/gen/3x8-minmax-wasmsimd-splat-x86.c",
862 "src/f32-igemm/gen/3x8s4-minmax-wasmsimd-arm.c",
863 "src/f32-igemm/gen/3x8s4-minmax-wasmsimd-x86.c",
864 "src/f32-igemm/gen/4x2c4-minmax-wasmsimd-arm.c",
865 "src/f32-igemm/gen/4x2c4-minmax-wasmsimd-x86.c",
866 "src/f32-igemm/gen/4x2c4-relu-wasmsimd.c",
867 "src/f32-igemm/gen/4x2c4-wasmsimd.c",
868 "src/f32-igemm/gen/4x8-minmax-wasmsimd-loadsplat-arm.c",
869 "src/f32-igemm/gen/4x8-minmax-wasmsimd-loadsplat-x86.c",
870 "src/f32-igemm/gen/4x8-minmax-wasmsimd-splat-arm.c",
871 "src/f32-igemm/gen/4x8-minmax-wasmsimd-splat-x86.c",
872 "src/f32-igemm/gen/4x8-relu-wasmsimd-splat.c",
873 "src/f32-igemm/gen/4x8-wasmsimd-splat.c",
874 "src/f32-igemm/gen/4x8s4-minmax-wasmsimd-arm.c",
875 "src/f32-igemm/gen/4x8s4-minmax-wasmsimd-x86.c",
876 "src/f32-igemm/gen/5x8-minmax-wasmsimd-loadsplat-arm.c",
877 "src/f32-igemm/gen/5x8-minmax-wasmsimd-loadsplat-x86.c",
878 "src/f32-igemm/gen/5x8-minmax-wasmsimd-splat-arm.c",
879 "src/f32-igemm/gen/5x8-minmax-wasmsimd-splat-x86.c",
880 "src/f32-igemm/gen/5x8-relu-wasmsimd-splat.c",
881 "src/f32-igemm/gen/5x8-wasmsimd-splat.c",
882 "src/f32-igemm/gen/5x8s4-minmax-wasmsimd-arm.c",
883 "src/f32-igemm/gen/5x8s4-minmax-wasmsimd-x86.c",
884 "src/f32-igemm/gen/6x8-minmax-wasmsimd-loadsplat-arm.c",
885 "src/f32-igemm/gen/6x8-minmax-wasmsimd-loadsplat-x86.c",
886 "src/f32-igemm/gen/6x8-minmax-wasmsimd-splat-arm.c",
887 "src/f32-igemm/gen/6x8-minmax-wasmsimd-splat-x86.c",
888 "src/f32-igemm/gen/6x8s4-minmax-wasmsimd-arm.c",
889 "src/f32-igemm/gen/6x8s4-minmax-wasmsimd-x86.c",
Marat Dukhanf6e24802020-07-08 22:20:40 -0700890 "src/f32-maxpool/9p8x-minmax-wasmsimd-arm-c4.c",
891 "src/f32-maxpool/9p8x-minmax-wasmsimd-x86-c4.c",
Marat Dukhan1483c532020-07-16 18:08:19 -0700892 "src/f32-pavgpool/9p8x-minmax-wasmsimd-arm-c4.c",
893 "src/f32-pavgpool/9p8x-minmax-wasmsimd-x86-c4.c",
894 "src/f32-pavgpool/9x-minmax-wasmsimd-arm-c4.c",
895 "src/f32-pavgpool/9x-minmax-wasmsimd-x86-c4.c",
Marat Dukhan0d0d8822020-07-23 23:37:56 -0700896 "src/f32-ppmm/gen/4x8-minmax-wasmsimd-splat-arm.c",
897 "src/f32-ppmm/gen/4x8-minmax-wasmsimd-splat-x86.c",
Frank Barcharda5316982020-07-23 13:19:28 -0700898 "src/f32-prelu/gen/wasmsimd-bitselect-1x4.c",
899 "src/f32-prelu/gen/wasmsimd-bitselect-1x8.c",
900 "src/f32-prelu/gen/wasmsimd-bitselect-1x16.c",
Marat Dukhan195f8eb2020-06-25 12:50:57 -0700901 "src/f32-prelu/gen/wasmsimd-bitselect-2x4.c",
902 "src/f32-prelu/gen/wasmsimd-bitselect-2x8.c",
Frank Barcharda5316982020-07-23 13:19:28 -0700903 "src/f32-prelu/gen/wasmsimd-bitselect-2x16.c",
904 "src/f32-prelu/gen/wasmsimd-bitselect-4x4.c",
905 "src/f32-prelu/gen/wasmsimd-bitselect-4x8.c",
906 "src/f32-prelu/gen/wasmsimd-bitselect-4x16.c",
907 "src/f32-prelu/gen/wasmsimd-minmax-1x4.c",
908 "src/f32-prelu/gen/wasmsimd-minmax-1x8.c",
909 "src/f32-prelu/gen/wasmsimd-minmax-1x16.c",
Marat Dukhan195f8eb2020-06-25 12:50:57 -0700910 "src/f32-prelu/gen/wasmsimd-minmax-2x4.c",
911 "src/f32-prelu/gen/wasmsimd-minmax-2x8.c",
Frank Barcharda5316982020-07-23 13:19:28 -0700912 "src/f32-prelu/gen/wasmsimd-minmax-2x16.c",
913 "src/f32-prelu/gen/wasmsimd-minmax-4x4.c",
914 "src/f32-prelu/gen/wasmsimd-minmax-4x8.c",
915 "src/f32-prelu/gen/wasmsimd-minmax-4x16.c",
Marat Dukhan52238f02020-07-16 15:30:28 -0700916 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x4.c",
Marat Dukhan52238f02020-07-16 15:30:28 -0700917 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x8-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700918 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x8.c",
Marat Dukhan52238f02020-07-16 15:30:28 -0700919 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x12-acc2.c",
920 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x12-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700921 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x12.c",
Marat Dukhan52238f02020-07-16 15:30:28 -0700922 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x16-acc2.c",
923 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x16-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700924 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x16.c",
Marat Dukhan52238f02020-07-16 15:30:28 -0700925 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x20-acc2.c",
926 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x20-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700927 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x20.c",
Frank Barchardfb158e22020-07-15 16:10:10 -0700928 "src/f32-relu/gen/wasmsimd-x4.c",
929 "src/f32-relu/gen/wasmsimd-x8.c",
Frank Barchard37297a62020-08-27 12:15:10 -0700930 "src/f32-relu/gen/wasmsimd-x16.c",
Marat Dukhan8c417962020-07-08 12:27:50 -0700931 "src/f32-rmax/wasmsimd-arm.c",
932 "src/f32-rmax/wasmsimd-x86.c",
Marat Dukhand187a5b2020-07-20 01:07:17 -0700933 "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x4.c",
934 "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x8.c",
935 "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x12.c",
936 "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x16.c",
937 "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x20.c",
938 "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x24.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700939 "src/f32-sigmoid/gen/wasmsimd-p5-div-x4.c",
940 "src/f32-sigmoid/gen/wasmsimd-p5-div-x8.c",
941 "src/f32-sigmoid/gen/wasmsimd-p5-div-x12.c",
942 "src/f32-sigmoid/gen/wasmsimd-p5-div-x16.c",
943 "src/f32-sigmoid/gen/wasmsimd-p5-div-x20.c",
944 "src/f32-sigmoid/gen/wasmsimd-p5-div-x24.c",
Frank Barchard8ef44cd2020-11-03 12:30:23 -0800945 "src/f32-spmm/gen/4x1-minmax-wasmsimd-arm-pipelined-x2.c",
946 "src/f32-spmm/gen/4x1-minmax-wasmsimd-arm-pipelined.c",
Frank Barchardbeca6522020-10-30 22:34:35 -0700947 "src/f32-spmm/gen/4x1-minmax-wasmsimd-arm-x2.c",
948 "src/f32-spmm/gen/4x1-minmax-wasmsimd-arm-x4.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700949 "src/f32-spmm/gen/4x1-minmax-wasmsimd-arm.c",
Frank Barchard8ef44cd2020-11-03 12:30:23 -0800950 "src/f32-spmm/gen/4x1-minmax-wasmsimd-x86-pipelined-x2.c",
951 "src/f32-spmm/gen/4x1-minmax-wasmsimd-x86-pipelined.c",
Frank Barchardbeca6522020-10-30 22:34:35 -0700952 "src/f32-spmm/gen/4x1-minmax-wasmsimd-x86-x2.c",
953 "src/f32-spmm/gen/4x1-minmax-wasmsimd-x86-x4.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700954 "src/f32-spmm/gen/4x1-minmax-wasmsimd-x86.c",
Frank Barchard8ef44cd2020-11-03 12:30:23 -0800955 "src/f32-spmm/gen/8x1-minmax-wasmsimd-arm-pipelined-x2.c",
956 "src/f32-spmm/gen/8x1-minmax-wasmsimd-arm-pipelined.c",
Frank Barchardbeca6522020-10-30 22:34:35 -0700957 "src/f32-spmm/gen/8x1-minmax-wasmsimd-arm-x2.c",
958 "src/f32-spmm/gen/8x1-minmax-wasmsimd-arm-x4.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700959 "src/f32-spmm/gen/8x1-minmax-wasmsimd-arm.c",
Frank Barchard8ef44cd2020-11-03 12:30:23 -0800960 "src/f32-spmm/gen/8x1-minmax-wasmsimd-x86-pipelined-x2.c",
961 "src/f32-spmm/gen/8x1-minmax-wasmsimd-x86-pipelined.c",
Frank Barchardbeca6522020-10-30 22:34:35 -0700962 "src/f32-spmm/gen/8x1-minmax-wasmsimd-x86-x2.c",
963 "src/f32-spmm/gen/8x1-minmax-wasmsimd-x86-x4.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700964 "src/f32-spmm/gen/8x1-minmax-wasmsimd-x86.c",
Frank Barchard8ef44cd2020-11-03 12:30:23 -0800965 "src/f32-spmm/gen/16x1-minmax-wasmsimd-arm-pipelined-x2.c",
966 "src/f32-spmm/gen/16x1-minmax-wasmsimd-arm-pipelined.c",
Frank Barchardbeca6522020-10-30 22:34:35 -0700967 "src/f32-spmm/gen/16x1-minmax-wasmsimd-arm-x2.c",
968 "src/f32-spmm/gen/16x1-minmax-wasmsimd-arm-x4.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700969 "src/f32-spmm/gen/16x1-minmax-wasmsimd-arm.c",
Frank Barchard8ef44cd2020-11-03 12:30:23 -0800970 "src/f32-spmm/gen/16x1-minmax-wasmsimd-x86-pipelined-x2.c",
971 "src/f32-spmm/gen/16x1-minmax-wasmsimd-x86-pipelined.c",
Frank Barchardbeca6522020-10-30 22:34:35 -0700972 "src/f32-spmm/gen/16x1-minmax-wasmsimd-x86-x2.c",
973 "src/f32-spmm/gen/16x1-minmax-wasmsimd-x86-x4.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700974 "src/f32-spmm/gen/16x1-minmax-wasmsimd-x86.c",
Frank Barchard8ef44cd2020-11-03 12:30:23 -0800975 "src/f32-spmm/gen/32x1-minmax-wasmsimd-arm-pipelined-x2.c",
976 "src/f32-spmm/gen/32x1-minmax-wasmsimd-arm-pipelined.c",
Frank Barchardbeca6522020-10-30 22:34:35 -0700977 "src/f32-spmm/gen/32x1-minmax-wasmsimd-arm-x2.c",
978 "src/f32-spmm/gen/32x1-minmax-wasmsimd-arm-x4.c",
Frank Barchard846c0c62020-10-26 15:01:39 -0700979 "src/f32-spmm/gen/32x1-minmax-wasmsimd-arm.c",
Frank Barchard8ef44cd2020-11-03 12:30:23 -0800980 "src/f32-spmm/gen/32x1-minmax-wasmsimd-x86-pipelined-x2.c",
981 "src/f32-spmm/gen/32x1-minmax-wasmsimd-x86-pipelined.c",
Frank Barchardbeca6522020-10-30 22:34:35 -0700982 "src/f32-spmm/gen/32x1-minmax-wasmsimd-x86-x2.c",
983 "src/f32-spmm/gen/32x1-minmax-wasmsimd-x86-x4.c",
Frank Barchard846c0c62020-10-26 15:01:39 -0700984 "src/f32-spmm/gen/32x1-minmax-wasmsimd-x86.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700985 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-arm-x4.c",
986 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700987 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-arm-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700988 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-x86-x4.c",
989 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700990 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700991 "src/f32-vbinary/gen/vadd-relu-wasmsimd-x4.c",
992 "src/f32-vbinary/gen/vadd-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700993 "src/f32-vbinary/gen/vadd-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700994 "src/f32-vbinary/gen/vadd-wasmsimd-x4.c",
995 "src/f32-vbinary/gen/vadd-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700996 "src/f32-vbinary/gen/vadd-wasmsimd-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700997 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-arm-x4.c",
998 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700999 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-arm-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001000 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-x86-x4.c",
1001 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001002 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -07001003 "src/f32-vbinary/gen/vaddc-relu-wasmsimd-x4.c",
1004 "src/f32-vbinary/gen/vaddc-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001005 "src/f32-vbinary/gen/vaddc-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07001006 "src/f32-vbinary/gen/vaddc-wasmsimd-x4.c",
1007 "src/f32-vbinary/gen/vaddc-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001008 "src/f32-vbinary/gen/vaddc-wasmsimd-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -07001009 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-arm-x4.c",
1010 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001011 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-arm-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -07001012 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-x86-x4.c",
1013 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001014 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -07001015 "src/f32-vbinary/gen/vdiv-relu-wasmsimd-x4.c",
1016 "src/f32-vbinary/gen/vdiv-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001017 "src/f32-vbinary/gen/vdiv-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07001018 "src/f32-vbinary/gen/vdiv-wasmsimd-x4.c",
1019 "src/f32-vbinary/gen/vdiv-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001020 "src/f32-vbinary/gen/vdiv-wasmsimd-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001021 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-arm-x4.c",
1022 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001023 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-arm-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001024 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-x86-x4.c",
1025 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001026 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -07001027 "src/f32-vbinary/gen/vdivc-relu-wasmsimd-x4.c",
1028 "src/f32-vbinary/gen/vdivc-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001029 "src/f32-vbinary/gen/vdivc-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07001030 "src/f32-vbinary/gen/vdivc-wasmsimd-x4.c",
1031 "src/f32-vbinary/gen/vdivc-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001032 "src/f32-vbinary/gen/vdivc-wasmsimd-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -07001033 "src/f32-vbinary/gen/vmax-wasmsimd-arm-x4.c",
1034 "src/f32-vbinary/gen/vmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001035 "src/f32-vbinary/gen/vmax-wasmsimd-arm-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -07001036 "src/f32-vbinary/gen/vmax-wasmsimd-x86-x4.c",
1037 "src/f32-vbinary/gen/vmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001038 "src/f32-vbinary/gen/vmax-wasmsimd-x86-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001039 "src/f32-vbinary/gen/vmaxc-wasmsimd-arm-x4.c",
1040 "src/f32-vbinary/gen/vmaxc-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001041 "src/f32-vbinary/gen/vmaxc-wasmsimd-arm-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001042 "src/f32-vbinary/gen/vmaxc-wasmsimd-x86-x4.c",
1043 "src/f32-vbinary/gen/vmaxc-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001044 "src/f32-vbinary/gen/vmaxc-wasmsimd-x86-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -07001045 "src/f32-vbinary/gen/vmin-wasmsimd-arm-x4.c",
1046 "src/f32-vbinary/gen/vmin-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001047 "src/f32-vbinary/gen/vmin-wasmsimd-arm-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -07001048 "src/f32-vbinary/gen/vmin-wasmsimd-x86-x4.c",
1049 "src/f32-vbinary/gen/vmin-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001050 "src/f32-vbinary/gen/vmin-wasmsimd-x86-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001051 "src/f32-vbinary/gen/vminc-wasmsimd-arm-x4.c",
1052 "src/f32-vbinary/gen/vminc-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001053 "src/f32-vbinary/gen/vminc-wasmsimd-arm-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001054 "src/f32-vbinary/gen/vminc-wasmsimd-x86-x4.c",
1055 "src/f32-vbinary/gen/vminc-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001056 "src/f32-vbinary/gen/vminc-wasmsimd-x86-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -07001057 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-arm-x4.c",
1058 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001059 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-arm-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -07001060 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-x86-x4.c",
1061 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001062 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -07001063 "src/f32-vbinary/gen/vmul-relu-wasmsimd-x4.c",
1064 "src/f32-vbinary/gen/vmul-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001065 "src/f32-vbinary/gen/vmul-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07001066 "src/f32-vbinary/gen/vmul-wasmsimd-x4.c",
1067 "src/f32-vbinary/gen/vmul-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001068 "src/f32-vbinary/gen/vmul-wasmsimd-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001069 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-arm-x4.c",
1070 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001071 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-arm-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001072 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-x86-x4.c",
1073 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001074 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -07001075 "src/f32-vbinary/gen/vmulc-relu-wasmsimd-x4.c",
1076 "src/f32-vbinary/gen/vmulc-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001077 "src/f32-vbinary/gen/vmulc-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07001078 "src/f32-vbinary/gen/vmulc-wasmsimd-x4.c",
1079 "src/f32-vbinary/gen/vmulc-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001080 "src/f32-vbinary/gen/vmulc-wasmsimd-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001081 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-arm-x4.c",
1082 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001083 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-arm-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001084 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-x86-x4.c",
1085 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001086 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -07001087 "src/f32-vbinary/gen/vrdivc-relu-wasmsimd-x4.c",
1088 "src/f32-vbinary/gen/vrdivc-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001089 "src/f32-vbinary/gen/vrdivc-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07001090 "src/f32-vbinary/gen/vrdivc-wasmsimd-x4.c",
1091 "src/f32-vbinary/gen/vrdivc-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001092 "src/f32-vbinary/gen/vrdivc-wasmsimd-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001093 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-arm-x4.c",
1094 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001095 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-arm-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001096 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-x86-x4.c",
1097 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001098 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -07001099 "src/f32-vbinary/gen/vrsubc-relu-wasmsimd-x4.c",
1100 "src/f32-vbinary/gen/vrsubc-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001101 "src/f32-vbinary/gen/vrsubc-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07001102 "src/f32-vbinary/gen/vrsubc-wasmsimd-x4.c",
1103 "src/f32-vbinary/gen/vrsubc-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001104 "src/f32-vbinary/gen/vrsubc-wasmsimd-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -07001105 "src/f32-vbinary/gen/vsqrdiff-wasmsimd-x4.c",
1106 "src/f32-vbinary/gen/vsqrdiff-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001107 "src/f32-vbinary/gen/vsqrdiff-wasmsimd-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001108 "src/f32-vbinary/gen/vsqrdiffc-wasmsimd-x4.c",
1109 "src/f32-vbinary/gen/vsqrdiffc-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001110 "src/f32-vbinary/gen/vsqrdiffc-wasmsimd-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -07001111 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-arm-x4.c",
1112 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001113 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-arm-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -07001114 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-x86-x4.c",
1115 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001116 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -07001117 "src/f32-vbinary/gen/vsub-relu-wasmsimd-x4.c",
1118 "src/f32-vbinary/gen/vsub-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001119 "src/f32-vbinary/gen/vsub-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07001120 "src/f32-vbinary/gen/vsub-wasmsimd-x4.c",
1121 "src/f32-vbinary/gen/vsub-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001122 "src/f32-vbinary/gen/vsub-wasmsimd-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001123 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-arm-x4.c",
1124 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001125 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-arm-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001126 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-x86-x4.c",
1127 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001128 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -07001129 "src/f32-vbinary/gen/vsubc-relu-wasmsimd-x4.c",
1130 "src/f32-vbinary/gen/vsubc-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001131 "src/f32-vbinary/gen/vsubc-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07001132 "src/f32-vbinary/gen/vsubc-wasmsimd-x4.c",
1133 "src/f32-vbinary/gen/vsubc-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001134 "src/f32-vbinary/gen/vsubc-wasmsimd-x16.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07001135 "src/f32-vlrelu/gen/vlrelu-wasmsimd-bitselect-x4.c",
1136 "src/f32-vlrelu/gen/vlrelu-wasmsimd-bitselect-x8.c",
1137 "src/f32-vlrelu/gen/vlrelu-wasmsimd-minmax-x4.c",
1138 "src/f32-vlrelu/gen/vlrelu-wasmsimd-minmax-x8.c",
Marat Dukhand816f622020-07-15 10:14:39 -07001139 "src/f32-vmulcaddc/gen/c4-minmax-wasmsimd-arm-2x.c",
Marat Dukhand816f622020-07-15 10:14:39 -07001140 "src/f32-vmulcaddc/gen/c4-minmax-wasmsimd-x86-2x.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001141 "src/f32-vmulcaddc/gen/c8-minmax-wasmsimd-arm-2x.c",
Marat Dukhand816f622020-07-15 10:14:39 -07001142 "src/f32-vmulcaddc/gen/c8-minmax-wasmsimd-x86-2x.c",
Marat Dukhanb82b2cd2020-07-16 02:23:42 -07001143 "src/f32-vrnd/gen/vrndd-wasmsimd-addsub-x4.c",
1144 "src/f32-vrnd/gen/vrndd-wasmsimd-addsub-x8.c",
1145 "src/f32-vrnd/gen/vrndd-wasmsimd-cvt-x4.c",
1146 "src/f32-vrnd/gen/vrndd-wasmsimd-cvt-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001147 "src/f32-vrnd/gen/vrndne-wasmsimd-addsub-x4.c",
1148 "src/f32-vrnd/gen/vrndne-wasmsimd-addsub-x8.c",
1149 "src/f32-vrnd/gen/vrndu-wasmsimd-addsub-x4.c",
1150 "src/f32-vrnd/gen/vrndu-wasmsimd-addsub-x8.c",
1151 "src/f32-vrnd/gen/vrndu-wasmsimd-cvt-x4.c",
1152 "src/f32-vrnd/gen/vrndu-wasmsimd-cvt-x8.c",
1153 "src/f32-vrnd/gen/vrndz-wasmsimd-addsub-x4.c",
1154 "src/f32-vrnd/gen/vrndz-wasmsimd-addsub-x8.c",
1155 "src/f32-vrnd/gen/vrndz-wasmsimd-cvt-x4.c",
1156 "src/f32-vrnd/gen/vrndz-wasmsimd-cvt-x8.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07001157 "src/f32-vsqrt/gen/wasmsimd-sqrt-x4.c",
1158 "src/f32-vsqrt/gen/wasmsimd-sqrt-x8.c",
Marat Dukhan37c83512020-06-29 13:25:53 -07001159 "src/f32-vunary/gen/vabs-wasmsimd-x4.c",
1160 "src/f32-vunary/gen/vabs-wasmsimd-x8.c",
1161 "src/f32-vunary/gen/vneg-wasmsimd-x4.c",
1162 "src/f32-vunary/gen/vneg-wasmsimd-x8.c",
1163 "src/f32-vunary/gen/vsqr-wasmsimd-x4.c",
1164 "src/f32-vunary/gen/vsqr-wasmsimd-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001165 "src/math/roundd-wasmsimd-addsub.c",
1166 "src/math/roundd-wasmsimd-cvt.c",
1167 "src/math/roundne-wasmsimd-addsub.c",
1168 "src/math/roundu-wasmsimd-addsub.c",
1169 "src/math/roundu-wasmsimd-cvt.c",
1170 "src/math/roundz-wasmsimd-addsub.c",
1171 "src/math/roundz-wasmsimd-cvt.c",
1172 "src/math/sigmoid-wasmsimd-rr2-lut64-p2-div.c",
1173 "src/math/sigmoid-wasmsimd-rr2-p5-div.c",
Marat Dukhancc8f34c2020-08-05 16:36:38 -07001174 "src/qs8-dwconv/gen/up8x9-minmax-wasmsimd-mul16.c",
1175 "src/qs8-dwconv/gen/up16x9-minmax-wasmsimd-mul16.c",
1176 "src/qs8-dwconv/gen/up24x9-minmax-wasmsimd-mul16.c",
Marat Dukhanb5e3d172020-08-06 13:29:53 -07001177 "src/qs8-gavgpool/gen/7p7x-minmax-wasmsimd-c8-acc2.c",
1178 "src/qs8-gavgpool/gen/7p7x-minmax-wasmsimd-c16-acc2.c",
1179 "src/qs8-gavgpool/gen/7p7x-minmax-wasmsimd-c24-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001180 "src/qs8-gavgpool/gen/7x-minmax-wasmsimd-c8-acc2.c",
1181 "src/qs8-gavgpool/gen/7x-minmax-wasmsimd-c16-acc2.c",
1182 "src/qs8-gavgpool/gen/7x-minmax-wasmsimd-c24-acc2.c",
Marat Dukhan27203da2020-08-05 15:19:03 -07001183 "src/qs8-gemm/gen/1x4c8-minmax-wasmsimd-ld64.c",
Marat Dukhan27203da2020-08-05 15:19:03 -07001184 "src/qs8-gemm/gen/1x4c8-minmax-wasmsimd-ld128.c",
Marat Dukhan27203da2020-08-05 15:19:03 -07001185 "src/qs8-gemm/gen/1x4c8-xw-minmax-wasmsimd.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001186 "src/qs8-gemm/gen/2x4c8-minmax-wasmsimd-ld64.c",
1187 "src/qs8-gemm/gen/2x4c8-minmax-wasmsimd-ld128.c",
Marat Dukhan27203da2020-08-05 15:19:03 -07001188 "src/qs8-gemm/gen/2x4c8-xw-minmax-wasmsimd.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001189 "src/qs8-gemm/gen/3x4c8-minmax-wasmsimd-ld64.c",
1190 "src/qs8-gemm/gen/3x4c8-minmax-wasmsimd-ld128.c",
Marat Dukhan27203da2020-08-05 15:19:03 -07001191 "src/qs8-gemm/gen/3x4c8-xw-minmax-wasmsimd.c",
1192 "src/qs8-igemm/gen/1x4c8-minmax-wasmsimd-ld64.c",
Marat Dukhan27203da2020-08-05 15:19:03 -07001193 "src/qs8-igemm/gen/1x4c8-minmax-wasmsimd-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001194 "src/qs8-igemm/gen/2x4c8-minmax-wasmsimd-ld64.c",
Marat Dukhan27203da2020-08-05 15:19:03 -07001195 "src/qs8-igemm/gen/2x4c8-minmax-wasmsimd-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001196 "src/qs8-igemm/gen/3x4c8-minmax-wasmsimd-ld64.c",
Marat Dukhan27203da2020-08-05 15:19:03 -07001197 "src/qs8-igemm/gen/3x4c8-minmax-wasmsimd-ld128.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -07001198 "src/qs8-requantization/fp32-wasmsimd.c",
Marat Dukhan138560c2020-08-03 18:57:34 -07001199 "src/qs8-requantization/q31-wasmsimd.c",
Marat Dukhan5df27f82020-09-02 23:59:21 -07001200 "src/qs8-vadd/gen/minmax-wasmsimd-x8.c",
1201 "src/qs8-vadd/gen/minmax-wasmsimd-x16.c",
1202 "src/qs8-vadd/gen/minmax-wasmsimd-x24.c",
1203 "src/qs8-vadd/gen/minmax-wasmsimd-x32.c",
1204 "src/qs8-vaddc/gen/minmax-wasmsimd-x8.c",
1205 "src/qs8-vaddc/gen/minmax-wasmsimd-x16.c",
1206 "src/qs8-vaddc/gen/minmax-wasmsimd-x24.c",
1207 "src/qs8-vaddc/gen/minmax-wasmsimd-x32.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07001208 "src/qu8-requantization/fp32-wasmsimd.c",
Marat Dukhan22de5e72020-08-03 21:53:18 -07001209 "src/qu8-requantization/q31-wasmsimd.c",
Marat Dukhan8ee37012020-07-16 13:17:13 -07001210 "src/x32-fill/wasmsimd.c",
Marat Dukhan66d99e92020-07-16 12:56:21 -07001211 "src/x32-packx/x4-wasmsimd.c",
Marat Dukhan9306ae02020-07-16 15:51:13 -07001212 "src/x32-pad/wasmsimd.c",
Marat Dukhan9d4bfa22020-07-16 19:07:04 -07001213 "src/x32-unpool/wasmsimd.c",
Marat Dukhane3b78762020-07-16 20:02:58 -07001214 "src/x32-zip/x2-wasmsimd.c",
1215 "src/x32-zip/x3-wasmsimd.c",
1216 "src/x32-zip/x4-wasmsimd.c",
1217 "src/x32-zip/xm-wasmsimd.c",
Marat Dukhan290055c2020-06-09 12:24:29 -07001218]
1219
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001220PSIMD_ACCMATH_UKERNELS = [
Marat Dukhan2e23d2b2020-07-29 16:01:37 -07001221 "src/qs8-requantization/fp32-psimd.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001222 "src/qs8-requantization/precise-psimd.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07001223 "src/qu8-requantization/fp32-psimd.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001224 "src/qu8-requantization/precise-psimd.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001225]
1226
Marat Dukhan08c4a432019-10-03 09:29:21 -07001227# ISA-specific micro-kernels
1228NEON_UKERNELS = [
Marat Dukhanef25c6d2020-07-24 00:59:40 -07001229 "src/f32-argmaxpool/4x-neon-c4.c",
1230 "src/f32-argmaxpool/9p8x-neon-c4.c",
1231 "src/f32-argmaxpool/9x-neon-c4.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001232 "src/f32-avgpool/9p8x-minmax-neon-c4.c",
1233 "src/f32-avgpool/9x-minmax-neon-c4.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001234 "src/f32-clamp/gen/neon-x4.c",
1235 "src/f32-clamp/gen/neon-x8.c",
Marat Dukhan56b10cd2020-05-18 09:35:49 -07001236 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neon-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001237 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neon-2x2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001238 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neon-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001239 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neon-2x2.c",
Marat Dukhan56b10cd2020-05-18 09:35:49 -07001240 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neon-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001241 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neon-2x2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001242 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neon-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001243 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neon-2x2.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -07001244 "src/f32-dwconv/gen/up4x4-minmax-neon-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001245 "src/f32-dwconv/gen/up4x4-minmax-neon.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001246 "src/f32-dwconv/gen/up4x9-minmax-neon-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001247 "src/f32-dwconv/gen/up4x9-minmax-neon.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -07001248 "src/f32-dwconv/gen/up4x25-minmax-neon-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001249 "src/f32-dwconv/gen/up4x25-minmax-neon.c",
1250 "src/f32-dwconv/gen/up8x4-minmax-neon-acc2.c",
1251 "src/f32-dwconv/gen/up8x4-minmax-neon.c",
1252 "src/f32-dwconv/gen/up8x9-minmax-neon-acc2.c",
1253 "src/f32-dwconv/gen/up8x9-minmax-neon.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -07001254 "src/f32-dwconv/gen/up8x25-minmax-neon-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001255 "src/f32-dwconv/gen/up8x25-minmax-neon.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001256 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-1x4-acc2.c",
1257 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-1x4-acc3.c",
1258 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-1x4-acc4.c",
Marat Dukhanc581e482020-10-24 01:28:11 -07001259 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-1x4.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001260 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-2x4-acc2.c",
Marat Dukhanc581e482020-10-24 01:28:11 -07001261 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-2x4.c",
1262 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-3x4.c",
1263 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-4x4.c",
1264 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-5x4.c",
1265 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-6x4.c",
Marat Dukhan82f0c322020-10-25 19:17:35 -07001266 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neon-1x4-acc2.c",
1267 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neon-1x4-acc3.c",
1268 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neon-1x4-acc4.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001269 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neon-1x4.c",
Marat Dukhan82f0c322020-10-25 19:17:35 -07001270 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neon-2x4-acc2.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001271 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neon-2x4.c",
1272 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neon-3x4.c",
1273 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neon-4x4.c",
Marat Dukhan149f0ea2020-10-26 12:50:33 -07001274 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-1x4-acc2.c",
1275 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-1x4-acc3.c",
1276 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-1x4-acc4.c",
1277 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-1x4-acc5.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001278 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-1x4.c",
Marat Dukhan149f0ea2020-10-26 12:50:33 -07001279 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-2x4-acc2.c",
1280 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-2x4-acc3.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001281 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-2x4.c",
Marat Dukhan149f0ea2020-10-26 12:50:33 -07001282 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-3x4-acc2.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001283 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-3x4.c",
Marat Dukhan149f0ea2020-10-26 12:50:33 -07001284 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-4x4-acc2.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001285 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-4x4.c",
1286 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-5x4.c",
Marat Dukhan30d4b252020-10-29 16:33:22 -07001287 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neon-1x4-acc2.c",
1288 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neon-1x4-acc3.c",
1289 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neon-1x4-acc4.c",
1290 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neon-1x4-acc5.c",
1291 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neon-1x4.c",
1292 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neon-2x4-acc2.c",
1293 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neon-2x4-acc3.c",
1294 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neon-2x4.c",
Marat Dukhan30d4b252020-10-29 16:33:22 -07001295 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neon-3x4-acc2.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07001296 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neon-3x4.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -07001297 "src/f32-gavgpool-cw/neon-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001298 "src/f32-gavgpool/7p7x-minmax-neon-c4.c",
1299 "src/f32-gavgpool/7x-minmax-neon-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001300 "src/f32-gemm/gen-inc/1x8inc-minmax-neon-dup-ld64.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001301 "src/f32-gemm/gen-inc/1x8inc-minmax-neon-lane-ld64.c",
1302 "src/f32-gemm/gen-inc/1x8s4inc-minmax-neon.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001303 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-dup-ld64.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001304 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-dup-ld128.c",
1305 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-lane-ld64.c",
1306 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-lane-ld128.c",
1307 "src/f32-gemm/gen-inc/4x8s4inc-minmax-neon.c",
1308 "src/f32-gemm/gen-inc/5x8inc-minmax-neon-lane-ld64.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001309 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-dup-ld64.c",
1310 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-dup-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001311 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-lane-ld64.c",
1312 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-lane-ld128.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001313 "src/f32-gemm/gen-inc/6x8s4inc-minmax-neon.c",
1314 "src/f32-gemm/gen-inc/8x8s4inc-minmax-neon.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001315 "src/f32-gemm/gen/1x8-minmax-neon-dup-ld64.c",
1316 "src/f32-gemm/gen/1x8-minmax-neon-lane-ld64.c",
1317 "src/f32-gemm/gen/1x8s4-minmax-neon.c",
1318 "src/f32-gemm/gen/4x2-minmax-neon-lane-ld64.c",
1319 "src/f32-gemm/gen/4x8-minmax-neon-dup-ld64.c",
1320 "src/f32-gemm/gen/4x8-minmax-neon-dup-ld128.c",
1321 "src/f32-gemm/gen/4x8-minmax-neon-lane-ld64.c",
1322 "src/f32-gemm/gen/4x8-minmax-neon-lane-ld128.c",
1323 "src/f32-gemm/gen/4x8s4-minmax-neon.c",
1324 "src/f32-gemm/gen/5x8-minmax-neon-lane-ld64.c",
1325 "src/f32-gemm/gen/6x8-minmax-neon-dup-ld64.c",
1326 "src/f32-gemm/gen/6x8-minmax-neon-dup-ld128.c",
1327 "src/f32-gemm/gen/6x8-minmax-neon-lane-ld64.c",
1328 "src/f32-gemm/gen/6x8-minmax-neon-lane-ld128.c",
1329 "src/f32-gemm/gen/6x8s4-minmax-neon.c",
1330 "src/f32-gemm/gen/8x8s4-minmax-neon.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07001331 "src/f32-hswish/gen/hswish-neon-x4.c",
1332 "src/f32-hswish/gen/hswish-neon-x8.c",
1333 "src/f32-hswish/gen/hswish-neon-x16.c",
Marat Dukhan660fd192020-03-10 04:55:30 -07001334 "src/f32-ibilinear/gen/neon-c4.c",
1335 "src/f32-ibilinear/gen/neon-c8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001336 "src/f32-igemm/gen/1x8-minmax-neon-dup-ld64.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001337 "src/f32-igemm/gen/1x8-minmax-neon-lane-ld64.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001338 "src/f32-igemm/gen/1x8s4-minmax-neon.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001339 "src/f32-igemm/gen/4x2-minmax-neon-lane-ld64.c",
1340 "src/f32-igemm/gen/4x4-minmax-neon-lane-ld64.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001341 "src/f32-igemm/gen/4x8-minmax-neon-dup-ld64.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001342 "src/f32-igemm/gen/4x8-minmax-neon-dup-ld128.c",
1343 "src/f32-igemm/gen/4x8-minmax-neon-lane-ld64.c",
1344 "src/f32-igemm/gen/4x8-minmax-neon-lane-ld128.c",
1345 "src/f32-igemm/gen/4x8s4-minmax-neon.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001346 "src/f32-igemm/gen/6x8-minmax-neon-dup-ld64.c",
1347 "src/f32-igemm/gen/6x8-minmax-neon-dup-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001348 "src/f32-igemm/gen/6x8-minmax-neon-lane-ld64.c",
1349 "src/f32-igemm/gen/6x8-minmax-neon-lane-ld128.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001350 "src/f32-igemm/gen/6x8s4-minmax-neon.c",
1351 "src/f32-igemm/gen/8x8s4-minmax-neon.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001352 "src/f32-maxpool/9p8x-minmax-neon-c4.c",
1353 "src/f32-pavgpool/9p8x-minmax-neon-c4.c",
1354 "src/f32-pavgpool/9x-minmax-neon-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001355 "src/f32-ppmm/gen/4x8-minmax-neon.c",
1356 "src/f32-ppmm/gen/8x8-minmax-neon.c",
Frank Barcharda5316982020-07-23 13:19:28 -07001357 "src/f32-prelu/gen/neon-1x4.c",
1358 "src/f32-prelu/gen/neon-1x8.c",
1359 "src/f32-prelu/gen/neon-1x16.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001360 "src/f32-prelu/gen/neon-2x4.c",
1361 "src/f32-prelu/gen/neon-2x8.c",
Frank Barcharda5316982020-07-23 13:19:28 -07001362 "src/f32-prelu/gen/neon-2x16.c",
1363 "src/f32-prelu/gen/neon-4x4.c",
1364 "src/f32-prelu/gen/neon-4x8.c",
1365 "src/f32-prelu/gen/neon-4x16.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001366 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x4.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001367 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x8-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001368 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x8.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001369 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12-acc2.c",
1370 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001371 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001372 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16-acc2.c",
1373 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001374 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001375 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20-acc2.c",
1376 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001377 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20.c",
1378 "src/f32-raddstoreexpminusmax/gen/neon-p5-x4.c",
1379 "src/f32-raddstoreexpminusmax/gen/neon-p5-x8-acc2.c",
1380 "src/f32-raddstoreexpminusmax/gen/neon-p5-x8.c",
1381 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12-acc2.c",
1382 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12-acc3.c",
1383 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12.c",
1384 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16-acc2.c",
1385 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16-acc4.c",
1386 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16.c",
1387 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20-acc2.c",
1388 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20-acc5.c",
1389 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07001390 "src/f32-relu/gen/neon-x4.c",
1391 "src/f32-relu/gen/neon-x8.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001392 "src/f32-rmax/neon.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001393 "src/f32-sigmoid/gen/neon-frac-p9-p10-nr1recps-x16.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -08001394 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x4.c",
1395 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x8.c",
1396 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x12.c",
1397 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x16.c",
1398 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x20.c",
1399 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x24.c",
1400 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x4.c",
1401 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x8.c",
1402 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x12.c",
1403 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x16.c",
1404 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x20.c",
1405 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x24.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001406 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x4.c",
1407 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x8.c",
1408 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x12.c",
1409 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x16.c",
1410 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x20.c",
1411 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x24.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001412 "src/f32-vbinary/gen/vadd-minmax-neon-x4.c",
1413 "src/f32-vbinary/gen/vadd-minmax-neon-x8.c",
1414 "src/f32-vbinary/gen/vaddc-minmax-neon-x4.c",
1415 "src/f32-vbinary/gen/vaddc-minmax-neon-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -08001416 "src/f32-vbinary/gen/vmax-neon-x4.c",
1417 "src/f32-vbinary/gen/vmax-neon-x8.c",
1418 "src/f32-vbinary/gen/vmaxc-neon-x4.c",
1419 "src/f32-vbinary/gen/vmaxc-neon-x8.c",
1420 "src/f32-vbinary/gen/vmin-neon-x4.c",
1421 "src/f32-vbinary/gen/vmin-neon-x8.c",
1422 "src/f32-vbinary/gen/vminc-neon-x4.c",
1423 "src/f32-vbinary/gen/vminc-neon-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001424 "src/f32-vbinary/gen/vmul-minmax-neon-x4.c",
1425 "src/f32-vbinary/gen/vmul-minmax-neon-x8.c",
1426 "src/f32-vbinary/gen/vmulc-minmax-neon-x4.c",
1427 "src/f32-vbinary/gen/vmulc-minmax-neon-x8.c",
1428 "src/f32-vbinary/gen/vrsubc-minmax-neon-x4.c",
1429 "src/f32-vbinary/gen/vrsubc-minmax-neon-x8.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07001430 "src/f32-vbinary/gen/vsqrdiff-neon-x4.c",
1431 "src/f32-vbinary/gen/vsqrdiff-neon-x8.c",
1432 "src/f32-vbinary/gen/vsqrdiffc-neon-x4.c",
1433 "src/f32-vbinary/gen/vsqrdiffc-neon-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001434 "src/f32-vbinary/gen/vsub-minmax-neon-x4.c",
1435 "src/f32-vbinary/gen/vsub-minmax-neon-x8.c",
1436 "src/f32-vbinary/gen/vsubc-minmax-neon-x4.c",
1437 "src/f32-vbinary/gen/vsubc-minmax-neon-x8.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07001438 "src/f32-vlrelu/gen/vlrelu-neon-x4.c",
1439 "src/f32-vlrelu/gen/vlrelu-neon-x8.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001440 "src/f32-vmulcaddc/gen/c4-minmax-neon-2x.c",
1441 "src/f32-vmulcaddc/gen/c8-minmax-neon-2x.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07001442 "src/f32-vrnd/gen/vrndd-neon-x4.c",
1443 "src/f32-vrnd/gen/vrndd-neon-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001444 "src/f32-vrnd/gen/vrndne-neon-x4.c",
1445 "src/f32-vrnd/gen/vrndne-neon-x8.c",
1446 "src/f32-vrnd/gen/vrndu-neon-x4.c",
1447 "src/f32-vrnd/gen/vrndu-neon-x8.c",
1448 "src/f32-vrnd/gen/vrndz-neon-x4.c",
1449 "src/f32-vrnd/gen/vrndz-neon-x8.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07001450 "src/f32-vunary/gen/vabs-neon-x4.c",
1451 "src/f32-vunary/gen/vabs-neon-x8.c",
1452 "src/f32-vunary/gen/vneg-neon-x4.c",
1453 "src/f32-vunary/gen/vneg-neon-x8.c",
1454 "src/f32-vunary/gen/vsqr-neon-x4.c",
1455 "src/f32-vunary/gen/vsqr-neon-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001456 "src/math/roundd-neon-addsub.c",
1457 "src/math/roundd-neon-cvt.c",
1458 "src/math/roundne-neon-addsub.c",
1459 "src/math/roundu-neon-addsub.c",
1460 "src/math/roundu-neon-cvt.c",
1461 "src/math/roundz-neon-addsub.c",
1462 "src/math/roundz-neon-cvt.c",
1463 "src/math/sigmoid-neon-frac-p9-p10-nr1recps.c",
1464 "src/math/sigmoid-neon-rr1-lut64-p2-nr2recps.c",
1465 "src/math/sigmoid-neon-rr1-lut2048-p1-nr2recps.c",
1466 "src/math/sigmoid-neon-rr1-p5-nr2recps.c",
1467 "src/math/sigmoid-neon-rr2-lut64-p2-nr2recps.c",
1468 "src/math/sigmoid-neon-rr2-lut2048-p1-nr2recps.c",
1469 "src/math/sigmoid-neon-rr2-p5-nr2recps.c",
1470 "src/math/sqrt-neon-nr1rsqrts.c",
1471 "src/math/sqrt-neon-nr2rsqrts.c",
1472 "src/math/sqrt-neon-nr3rsqrts.c",
Marat Dukhan023bcf92020-08-10 12:40:50 -07001473 "src/qs8-dwconv/gen/up8x9-minmax-neon-mul16.c",
1474 "src/qs8-dwconv/gen/up16x9-minmax-neon-mul16.c",
1475 "src/qs8-dwconv/gen/up24x9-minmax-neon-mul16.c",
1476 "src/qs8-dwconv/gen/up32x9-minmax-neon-mul16.c",
Marat Dukhan281262d2020-08-10 13:23:21 -07001477 "src/qs8-gavgpool/gen/7p7x-minmax-neon-c8-acc2.c",
1478 "src/qs8-gavgpool/gen/7p7x-minmax-neon-c16-acc2.c",
1479 "src/qs8-gavgpool/gen/7p7x-minmax-neon-c24-acc2.c",
1480 "src/qs8-gavgpool/gen/7p7x-minmax-neon-c32-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001481 "src/qs8-gavgpool/gen/7x-minmax-neon-c8-acc2.c",
1482 "src/qs8-gavgpool/gen/7x-minmax-neon-c16-acc2.c",
1483 "src/qs8-gavgpool/gen/7x-minmax-neon-c24-acc2.c",
1484 "src/qs8-gavgpool/gen/7x-minmax-neon-c32-acc2.c",
Marat Dukhan40bbafe2020-08-04 02:04:22 -07001485 "src/qs8-gemm/gen/1x8-minmax-neon-mlal-lane.c",
Marat Dukhan40bbafe2020-08-04 02:04:22 -07001486 "src/qs8-gemm/gen/1x16-minmax-neon-mlal-lane.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001487 "src/qs8-gemm/gen/2x8-minmax-neon-mlal-lane.c",
Marat Dukhan40bbafe2020-08-04 02:04:22 -07001488 "src/qs8-gemm/gen/2x16-minmax-neon-mlal-lane.c",
1489 "src/qs8-igemm/gen/1x8-minmax-neon-mlal-lane.c",
Marat Dukhan40bbafe2020-08-04 02:04:22 -07001490 "src/qs8-igemm/gen/1x16-minmax-neon-mlal-lane.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001491 "src/qs8-igemm/gen/2x8-minmax-neon-mlal-lane.c",
Marat Dukhan40bbafe2020-08-04 02:04:22 -07001492 "src/qs8-igemm/gen/2x16-minmax-neon-mlal-lane.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -07001493 "src/qs8-requantization/fp32-neon.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001494 "src/qs8-requantization/precise-neon.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -07001495 "src/qs8-requantization/q31-neon.c",
Marat Dukhanba7b2792020-09-02 14:26:45 -07001496 "src/qs8-vadd/gen/minmax-neon-ld64-x8.c",
1497 "src/qs8-vadd/gen/minmax-neon-ld64-x16.c",
1498 "src/qs8-vadd/gen/minmax-neon-ld64-x24.c",
1499 "src/qs8-vadd/gen/minmax-neon-ld64-x32.c",
1500 "src/qs8-vaddc/gen/minmax-neon-ld64-x8.c",
1501 "src/qs8-vaddc/gen/minmax-neon-ld64-x16.c",
1502 "src/qs8-vaddc/gen/minmax-neon-ld64-x24.c",
1503 "src/qs8-vaddc/gen/minmax-neon-ld64-x32.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -07001504 "src/qu8-avgpool/9p8x-minmax-neon-c8.c",
1505 "src/qu8-avgpool/9x-minmax-neon-c8.c",
1506 "src/qu8-dwconv/up8x9-minmax-neon.c",
1507 "src/qu8-gavgpool/7p7x-minmax-neon-c8.c",
1508 "src/qu8-gavgpool/7x-minmax-neon-c8.c",
1509 "src/qu8-gemm/4x8-minmax-neon.c",
1510 "src/qu8-gemm/8x8-minmax-neon.c",
1511 "src/qu8-igemm/4x8-minmax-neon.c",
1512 "src/qu8-igemm/8x8-minmax-neon.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07001513 "src/qu8-requantization/fp32-neon.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001514 "src/qu8-requantization/precise-neon.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07001515 "src/qu8-requantization/q31-neon.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -07001516 "src/qu8-vadd/minmax-neon.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001517 "src/u8-clamp/neon-x64.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001518 "src/u8-maxpool/9p8x-minmax-neon-c16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001519 "src/u8-rmax/neon.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001520 "src/x8-zip/x2-neon.c",
1521 "src/x8-zip/x3-neon.c",
1522 "src/x8-zip/x4-neon.c",
1523 "src/x8-zip/xm-neon.c",
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -07001524 "src/x32-fill/neon.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001525 "src/x32-packx/x4-neon-st4.c",
Marat Dukhan63523d42020-05-22 17:07:33 -07001526 "src/x32-pad/neon.c",
Marat Dukhan57dccd82020-04-14 00:53:10 -07001527 "src/x32-unpool/neon.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001528 "src/x32-zip/x2-neon.c",
1529 "src/x32-zip/x3-neon.c",
1530 "src/x32-zip/x4-neon.c",
1531 "src/x32-zip/xm-neon.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001532]
1533
1534NEONFMA_UKERNELS = [
Frank Barchard04336c12020-10-22 16:48:55 -07001535 "src/f32-dwconv/gen/up4x4-minmax-neonfma-acc2.c",
1536 "src/f32-dwconv/gen/up4x4-minmax-neonfma.c",
1537 "src/f32-dwconv/gen/up4x9-minmax-neonfma-acc2.c",
1538 "src/f32-dwconv/gen/up4x9-minmax-neonfma.c",
1539 "src/f32-dwconv/gen/up4x25-minmax-neonfma-acc2.c",
1540 "src/f32-dwconv/gen/up4x25-minmax-neonfma.c",
1541 "src/f32-dwconv/gen/up8x4-minmax-neonfma-acc2.c",
1542 "src/f32-dwconv/gen/up8x4-minmax-neonfma.c",
1543 "src/f32-dwconv/gen/up8x9-minmax-neonfma-acc2.c",
1544 "src/f32-dwconv/gen/up8x9-minmax-neonfma.c",
1545 "src/f32-dwconv/gen/up8x25-minmax-neonfma-acc2.c",
1546 "src/f32-dwconv/gen/up8x25-minmax-neonfma.c",
1547 "src/f32-gemm/gen-inc/1x8inc-minmax-neonfma-dup-ld64.c",
1548 "src/f32-gemm/gen-inc/1x8s4inc-minmax-neonfma.c",
1549 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-dup-ld64.c",
1550 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-dup-ld128.c",
1551 "src/f32-gemm/gen-inc/4x8s4inc-minmax-neonfma.c",
1552 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-dup-ld64.c",
1553 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-dup-ld128.c",
1554 "src/f32-gemm/gen-inc/6x8s4inc-minmax-neonfma.c",
1555 "src/f32-gemm/gen-inc/8x8s4inc-minmax-neonfma.c",
1556 "src/f32-gemm/gen/1x8-minmax-neonfma-dup-ld64.c",
1557 "src/f32-gemm/gen/1x8s4-minmax-neonfma.c",
1558 "src/f32-gemm/gen/4x8-minmax-neonfma-dup-ld64.c",
1559 "src/f32-gemm/gen/4x8-minmax-neonfma-dup-ld128.c",
1560 "src/f32-gemm/gen/4x8s4-minmax-neonfma.c",
1561 "src/f32-gemm/gen/6x8-minmax-neonfma-dup-ld64.c",
1562 "src/f32-gemm/gen/6x8-minmax-neonfma-dup-ld128.c",
1563 "src/f32-gemm/gen/6x8s4-minmax-neonfma.c",
1564 "src/f32-gemm/gen/8x8s4-minmax-neonfma.c",
Marat Dukhan660fd192020-03-10 04:55:30 -07001565 "src/f32-ibilinear/gen/neonfma-c4.c",
1566 "src/f32-ibilinear/gen/neonfma-c8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001567 "src/f32-igemm/gen/1x8-minmax-neonfma-dup-ld64.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001568 "src/f32-igemm/gen/1x8s4-minmax-neonfma.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001569 "src/f32-igemm/gen/4x8-minmax-neonfma-dup-ld64.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001570 "src/f32-igemm/gen/4x8-minmax-neonfma-dup-ld128.c",
1571 "src/f32-igemm/gen/4x8s4-minmax-neonfma.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001572 "src/f32-igemm/gen/6x8-minmax-neonfma-dup-ld64.c",
1573 "src/f32-igemm/gen/6x8-minmax-neonfma-dup-ld128.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001574 "src/f32-igemm/gen/6x8s4-minmax-neonfma.c",
1575 "src/f32-igemm/gen/8x8s4-minmax-neonfma.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001576 "src/f32-ppmm/gen/4x8-minmax-neonfma.c",
1577 "src/f32-ppmm/gen/8x8-minmax-neonfma.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001578 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x4.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001579 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x8-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001580 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x8.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001581 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12-acc2.c",
1582 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001583 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001584 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16-acc2.c",
1585 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001586 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001587 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20-acc2.c",
1588 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001589 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20.c",
1590 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x4.c",
1591 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x8-acc2.c",
1592 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x8.c",
1593 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12-acc2.c",
1594 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12-acc3.c",
1595 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12.c",
1596 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16-acc2.c",
1597 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16-acc4.c",
1598 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16.c",
1599 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20-acc2.c",
1600 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20-acc5.c",
1601 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -08001602 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x4.c",
1603 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x8.c",
1604 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x12.c",
1605 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x16.c",
1606 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x20.c",
1607 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x24.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001608 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x4.c",
1609 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x8.c",
1610 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x12.c",
1611 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x16.c",
1612 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x20.c",
1613 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x24.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -08001614 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x4.c",
1615 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x8.c",
1616 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x12.c",
1617 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x16.c",
1618 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x20.c",
1619 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x24.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -08001620 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x4.c",
1621 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x8.c",
1622 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x12.c",
1623 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x16.c",
1624 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x20.c",
1625 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x24.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001626 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x4.c",
1627 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x8.c",
1628 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x12.c",
1629 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x16.c",
1630 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x20.c",
1631 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x24.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -08001632 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x4.c",
1633 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x8.c",
1634 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x12.c",
1635 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x16.c",
1636 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x20.c",
1637 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x24.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001638 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x4.c",
1639 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x8.c",
1640 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x12.c",
1641 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x16.c",
1642 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x20.c",
1643 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x24.c",
1644 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x4.c",
1645 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x8.c",
1646 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x12.c",
1647 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x16.c",
1648 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x20.c",
1649 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x24.c",
1650 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x4.c",
1651 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x8.c",
1652 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x12.c",
1653 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x16.c",
1654 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x20.c",
1655 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x24.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001656 "src/f32-vmulcaddc/gen/c4-minmax-neonfma-2x.c",
1657 "src/f32-vmulcaddc/gen/c8-minmax-neonfma-2x.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07001658 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x4.c",
1659 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x8.c",
1660 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x12.c",
1661 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x16.c",
1662 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x20.c",
1663 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x24.c",
1664 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x28.c",
1665 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x32.c",
1666 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x36.c",
1667 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x40.c",
1668 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x4.c",
1669 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x8.c",
1670 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x12.c",
1671 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x16.c",
1672 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x20.c",
1673 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x24.c",
1674 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x28.c",
1675 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x32.c",
1676 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x36.c",
1677 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x40.c",
Marat Dukhanb7633f22020-11-20 16:34:56 -08001678 "src/math/exp-neonfma-rr2-lut64-p2.c",
1679 "src/math/exp-neonfma-rr2-p5.c",
Marat Dukhan9dd119a2020-11-20 18:20:04 -08001680 "src/math/expminus-neonfma-rr2-lut64-p2.c",
1681 "src/math/expminus-neonfma-rr2-lut2048-p1.c",
1682 "src/math/expminus-neonfma-rr2-p5.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001683 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr1recps1fma.c",
1684 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr2fma.c",
1685 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr2recps.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001686 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr1recps1fma.c",
1687 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr2fma.c",
1688 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr2recps.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001689 "src/math/sigmoid-neonfma-rr1-p5-nr1recps1fma.c",
1690 "src/math/sigmoid-neonfma-rr1-p5-nr2fma.c",
1691 "src/math/sigmoid-neonfma-rr1-p5-nr2recps.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001692 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr1recps1fma.c",
1693 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr2fma.c",
1694 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr2recps.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001695 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr1recps1fma.c",
1696 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr2fma.c",
1697 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr2recps.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001698 "src/math/sigmoid-neonfma-rr2-p5-nr1recps1fma.c",
1699 "src/math/sigmoid-neonfma-rr2-p5-nr2fma.c",
1700 "src/math/sigmoid-neonfma-rr2-p5-nr2recps.c",
Marat Dukhan84000762020-06-29 18:38:43 -07001701 "src/math/sqrt-neonfma-nr1fma.c",
Marat Dukhan84000762020-06-29 18:38:43 -07001702 "src/math/sqrt-neonfma-nr1rsqrts1fma1adj.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001703 "src/math/sqrt-neonfma-nr2fma.c",
1704 "src/math/sqrt-neonfma-nr2fma1adj.c",
1705 "src/math/sqrt-neonfma-nr3fma.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001706]
1707
1708AARCH64_NEONFMA_UKERNELS = [
Marat Dukhan56b10cd2020-05-18 09:35:49 -07001709 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neonfma-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001710 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neonfma-2x2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001711 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neonfma-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001712 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neonfma-2x2.c",
Marat Dukhan56b10cd2020-05-18 09:35:49 -07001713 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neonfma-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001714 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neonfma-2x2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001715 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neonfma-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001716 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neonfma-2x2.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -07001717 "src/f32-conv-hwc2chw/3x3s2p1c3x4-neonfma-2x2.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001718 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-1x4-acc2.c",
1719 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-1x4-acc3.c",
1720 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-1x4-acc4.c",
Marat Dukhan1268a242020-10-24 00:36:32 -07001721 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-1x4.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001722 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-2x4-acc2.c",
Marat Dukhan1268a242020-10-24 00:36:32 -07001723 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-2x4.c",
1724 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-3x4.c",
1725 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-4x4.c",
1726 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-5x4.c",
1727 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-6x4.c",
Marat Dukhan82f0c322020-10-25 19:17:35 -07001728 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neonfma-1x4-acc2.c",
1729 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neonfma-1x4-acc3.c",
1730 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neonfma-1x4-acc4.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001731 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neonfma-1x4.c",
Marat Dukhan82f0c322020-10-25 19:17:35 -07001732 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neonfma-2x4-acc2.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001733 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neonfma-2x4.c",
1734 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neonfma-3x4.c",
1735 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neonfma-4x4.c",
Marat Dukhan149f0ea2020-10-26 12:50:33 -07001736 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-1x4-acc2.c",
1737 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-1x4-acc3.c",
1738 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-1x4-acc4.c",
1739 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-1x4-acc5.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001740 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-1x4.c",
Marat Dukhan149f0ea2020-10-26 12:50:33 -07001741 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-2x4-acc2.c",
1742 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-2x4-acc3.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001743 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-2x4.c",
Marat Dukhan149f0ea2020-10-26 12:50:33 -07001744 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-3x4-acc2.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001745 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-3x4.c",
Marat Dukhan149f0ea2020-10-26 12:50:33 -07001746 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-4x4-acc2.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001747 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-4x4.c",
1748 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-5x4.c",
Marat Dukhan30d4b252020-10-29 16:33:22 -07001749 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neonfma-1x4-acc2.c",
1750 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neonfma-1x4-acc3.c",
1751 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neonfma-1x4-acc4.c",
1752 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neonfma-1x4-acc5.c",
1753 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neonfma-1x4.c",
1754 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neonfma-2x4-acc2.c",
1755 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neonfma-2x4-acc3.c",
1756 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neonfma-2x4.c",
Marat Dukhan30d4b252020-10-29 16:33:22 -07001757 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neonfma-3x4-acc2.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07001758 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neonfma-3x4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001759 "src/f32-gemm/gen-inc/1x8inc-minmax-neonfma-lane-ld64.c",
1760 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-lane-ld64.c",
1761 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-lane-ld128.c",
1762 "src/f32-gemm/gen-inc/5x8inc-minmax-neonfma-lane-ld64.c",
1763 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-lane-ld64.c",
1764 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-lane-ld128.c",
1765 "src/f32-gemm/gen/1x8-minmax-neonfma-lane-ld64.c",
1766 "src/f32-gemm/gen/4x2-minmax-neonfma-lane-ld64.c",
1767 "src/f32-gemm/gen/4x8-minmax-neonfma-lane-ld64.c",
1768 "src/f32-gemm/gen/4x8-minmax-neonfma-lane-ld128.c",
1769 "src/f32-gemm/gen/5x8-minmax-neonfma-lane-ld64.c",
1770 "src/f32-gemm/gen/6x8-minmax-neonfma-lane-ld64.c",
1771 "src/f32-gemm/gen/6x8-minmax-neonfma-lane-ld128.c",
1772 "src/f32-igemm/gen/1x8-minmax-neonfma-lane-ld64.c",
1773 "src/f32-igemm/gen/4x2-minmax-neonfma-lane-ld64.c",
1774 "src/f32-igemm/gen/4x4-minmax-neonfma-lane-ld64.c",
1775 "src/f32-igemm/gen/4x8-minmax-neonfma-lane-ld64.c",
1776 "src/f32-igemm/gen/4x8-minmax-neonfma-lane-ld128.c",
1777 "src/f32-igemm/gen/6x8-minmax-neonfma-lane-ld64.c",
1778 "src/f32-igemm/gen/6x8-minmax-neonfma-lane-ld128.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -08001779 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x4.c",
1780 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x8.c",
1781 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x12.c",
1782 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x16.c",
1783 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x20.c",
1784 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x24.c",
1785 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x4.c",
1786 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x8.c",
1787 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x12.c",
1788 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x16.c",
1789 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x20.c",
1790 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x24.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001791 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x4.c",
1792 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x8.c",
1793 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x12.c",
1794 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x16.c",
1795 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x20.c",
1796 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x24.c",
Marat Dukhan355ab432020-04-09 19:01:52 -07001797 "src/f32-spmm/gen/4x1-minmax-neonfma-pipelined.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07001798 "src/f32-spmm/gen/4x1-minmax-neonfma-x2.c",
Marat Dukhan355ab432020-04-09 19:01:52 -07001799 "src/f32-spmm/gen/4x1-minmax-neonfma.c",
1800 "src/f32-spmm/gen/4x2-minmax-neonfma.c",
1801 "src/f32-spmm/gen/4x4-minmax-neonfma.c",
1802 "src/f32-spmm/gen/8x1-minmax-neonfma-pipelined.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07001803 "src/f32-spmm/gen/8x1-minmax-neonfma-x2.c",
Marat Dukhan355ab432020-04-09 19:01:52 -07001804 "src/f32-spmm/gen/8x1-minmax-neonfma.c",
1805 "src/f32-spmm/gen/8x2-minmax-neonfma.c",
1806 "src/f32-spmm/gen/8x4-minmax-neonfma.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001807 "src/f32-spmm/gen/12x1-minmax-neonfma.c",
1808 "src/f32-spmm/gen/12x2-minmax-neonfma.c",
1809 "src/f32-spmm/gen/12x4-minmax-neonfma.c",
1810 "src/f32-spmm/gen/16x1-minmax-neonfma-pipelined.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07001811 "src/f32-spmm/gen/16x1-minmax-neonfma-x2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001812 "src/f32-spmm/gen/16x1-minmax-neonfma.c",
1813 "src/f32-spmm/gen/16x2-minmax-neonfma.c",
1814 "src/f32-spmm/gen/16x4-minmax-neonfma.c",
Frank Barchard846c0c62020-10-26 15:01:39 -07001815 "src/f32-spmm/gen/32x1-minmax-neonfma-pipelined.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07001816 "src/f32-spmm/gen/32x1-minmax-neonfma-x2.c",
Frank Barchard846c0c62020-10-26 15:01:39 -07001817 "src/f32-spmm/gen/32x1-minmax-neonfma.c",
1818 "src/f32-spmm/gen/32x2-minmax-neonfma.c",
1819 "src/f32-spmm/gen/32x4-minmax-neonfma.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001820 "src/f32-vbinary/gen/vdiv-minmax-neon-x4.c",
1821 "src/f32-vbinary/gen/vdiv-minmax-neon-x8.c",
1822 "src/f32-vbinary/gen/vdivc-minmax-neon-x4.c",
1823 "src/f32-vbinary/gen/vdivc-minmax-neon-x8.c",
1824 "src/f32-vbinary/gen/vrdivc-minmax-neon-x4.c",
1825 "src/f32-vbinary/gen/vrdivc-minmax-neon-x8.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07001826 "src/f32-vsqrt/gen/neon-sqrt-x4.c",
1827 "src/f32-vsqrt/gen/neon-sqrt-x8.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001828 "src/math/sigmoid-neonfma-rr1-lut64-p2-div.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001829 "src/math/sigmoid-neonfma-rr1-lut2048-p1-div.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001830 "src/math/sigmoid-neonfma-rr1-p5-div.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001831 "src/math/sigmoid-neonfma-rr2-lut64-p2-div.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001832 "src/math/sigmoid-neonfma-rr2-lut2048-p1-div.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001833 "src/math/sigmoid-neonfma-rr2-p5-div.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001834]
1835
Marat Dukhan8853b822020-05-07 12:19:01 -07001836NEONV8_UKERNELS = [
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07001837 "src/f32-vrnd/gen/vrndd-neonv8-x4.c",
1838 "src/f32-vrnd/gen/vrndd-neonv8-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001839 "src/f32-vrnd/gen/vrndne-neonv8-x4.c",
1840 "src/f32-vrnd/gen/vrndne-neonv8-x8.c",
1841 "src/f32-vrnd/gen/vrndu-neonv8-x4.c",
1842 "src/f32-vrnd/gen/vrndu-neonv8-x8.c",
1843 "src/f32-vrnd/gen/vrndz-neonv8-x4.c",
1844 "src/f32-vrnd/gen/vrndz-neonv8-x8.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001845 "src/math/roundd-neonv8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001846 "src/math/roundne-neonv8.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001847 "src/math/roundu-neonv8.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07001848 "src/math/roundz-neonv8.c",
Marat Dukhan8853b822020-05-07 12:19:01 -07001849]
1850
Marat Dukhan08c4a432019-10-03 09:29:21 -07001851AARCH64_NEONFP16ARITH_UKERNELS = [
Frank Barchardb1966592020-05-12 13:47:06 -07001852 "src/f16-clamp/gen/neonfp16arith-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001853 "src/f16-clamp/gen/neonfp16arith-x16.c",
Frank Barchard5a599a62020-06-04 20:12:44 -07001854 "src/f16-dwconv/gen/up8x4-minmax-neonfp16arith-acc2.c",
1855 "src/f16-dwconv/gen/up8x4-minmax-neonfp16arith.c",
1856 "src/f16-dwconv/gen/up8x9-minmax-neonfp16arith-acc2.c",
1857 "src/f16-dwconv/gen/up8x9-minmax-neonfp16arith.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001858 "src/f16-dwconv/gen/up8x25-minmax-neonfp16arith-acc2.c",
1859 "src/f16-dwconv/gen/up8x25-minmax-neonfp16arith.c",
1860 "src/f16-dwconv/gen/up16x4-minmax-neonfp16arith-acc2.c",
1861 "src/f16-dwconv/gen/up16x4-minmax-neonfp16arith.c",
1862 "src/f16-dwconv/gen/up16x9-minmax-neonfp16arith-acc2.c",
1863 "src/f16-dwconv/gen/up16x9-minmax-neonfp16arith.c",
1864 "src/f16-dwconv/gen/up16x25-minmax-neonfp16arith-acc2.c",
1865 "src/f16-dwconv/gen/up16x25-minmax-neonfp16arith.c",
Frank Barchard0bb49a72020-06-04 11:35:11 -07001866 "src/f16-gavgpool/7p7x-minmax-neonfp16arith-c8.c",
1867 "src/f16-gavgpool/7x-minmax-neonfp16arith-c8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001868 "src/f16-gemm/gen-inc/1x8inc-minmax-neonfp16arith-ld64.c",
1869 "src/f16-gemm/gen-inc/1x16inc-minmax-neonfp16arith-ld64.c",
1870 "src/f16-gemm/gen-inc/4x8inc-minmax-neonfp16arith-ld64.c",
1871 "src/f16-gemm/gen-inc/4x16inc-minmax-neonfp16arith-ld64.c",
1872 "src/f16-gemm/gen-inc/6x8inc-minmax-neonfp16arith-ld64.c",
1873 "src/f16-gemm/gen-inc/6x16inc-minmax-neonfp16arith-ld64.c",
1874 "src/f16-gemm/gen-inc/8x8inc-minmax-neonfp16arith-ld64.c",
1875 "src/f16-gemm/gen-inc/8x16inc-minmax-neonfp16arith-ld64.c",
1876 "src/f16-gemm/gen/1x8-minmax-neonfp16arith-ld64.c",
1877 "src/f16-gemm/gen/1x16-minmax-neonfp16arith-ld64.c",
1878 "src/f16-gemm/gen/4x8-minmax-neonfp16arith-ld64.c",
1879 "src/f16-gemm/gen/4x16-minmax-neonfp16arith-ld64.c",
1880 "src/f16-gemm/gen/6x8-minmax-neonfp16arith-ld64.c",
1881 "src/f16-gemm/gen/6x16-minmax-neonfp16arith-ld64.c",
1882 "src/f16-gemm/gen/8x8-minmax-neonfp16arith-ld64.c",
1883 "src/f16-gemm/gen/8x16-minmax-neonfp16arith-ld64.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07001884 "src/f16-hswish/gen/hswish-neonfp16arith-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001885 "src/f16-hswish/gen/hswish-neonfp16arith-x16.c",
1886 "src/f16-igemm/gen/1x8-minmax-neonfp16arith-ld64.c",
1887 "src/f16-igemm/gen/1x16-minmax-neonfp16arith-ld64.c",
1888 "src/f16-igemm/gen/4x8-minmax-neonfp16arith-ld64.c",
1889 "src/f16-igemm/gen/4x16-minmax-neonfp16arith-ld64.c",
1890 "src/f16-igemm/gen/6x8-minmax-neonfp16arith-ld64.c",
1891 "src/f16-igemm/gen/6x16-minmax-neonfp16arith-ld64.c",
1892 "src/f16-igemm/gen/8x8-minmax-neonfp16arith-ld64.c",
1893 "src/f16-igemm/gen/8x16-minmax-neonfp16arith-ld64.c",
Frank Barchardb1966592020-05-12 13:47:06 -07001894 "src/f16-prelu/gen/neonfp16arith-2x8.c",
Frank Barcharda5316982020-07-23 13:19:28 -07001895 "src/f16-prelu/gen/neonfp16arith-2x16.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07001896 "src/f16-relu/gen/neonfp16arith-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001897 "src/f16-relu/gen/neonfp16arith-x16.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07001898 "src/f16-spmm/gen/8x1-minmax-neonfp16arith-x2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001899 "src/f16-spmm/gen/8x1-minmax-neonfp16arith.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07001900 "src/f16-spmm/gen/16x1-minmax-neonfp16arith-x2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001901 "src/f16-spmm/gen/16x1-minmax-neonfp16arith.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07001902 "src/f16-spmm/gen/24x1-minmax-neonfp16arith-x2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001903 "src/f16-spmm/gen/24x1-minmax-neonfp16arith.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07001904 "src/f16-spmm/gen/32x1-minmax-neonfp16arith-x2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001905 "src/f16-spmm/gen/32x1-minmax-neonfp16arith.c",
1906 "src/f16-vbinary/gen/vadd-minmax-neonfp16arith-x8.c",
1907 "src/f16-vbinary/gen/vadd-minmax-neonfp16arith-x16.c",
1908 "src/f16-vbinary/gen/vaddc-minmax-neonfp16arith-x8.c",
1909 "src/f16-vbinary/gen/vaddc-minmax-neonfp16arith-x16.c",
1910 "src/f16-vbinary/gen/vdiv-minmax-neonfp16arith-x8.c",
1911 "src/f16-vbinary/gen/vdiv-minmax-neonfp16arith-x16.c",
1912 "src/f16-vbinary/gen/vdivc-minmax-neonfp16arith-x8.c",
1913 "src/f16-vbinary/gen/vdivc-minmax-neonfp16arith-x16.c",
1914 "src/f16-vbinary/gen/vmax-neonfp16arith-x8.c",
1915 "src/f16-vbinary/gen/vmax-neonfp16arith-x16.c",
1916 "src/f16-vbinary/gen/vmaxc-neonfp16arith-x8.c",
1917 "src/f16-vbinary/gen/vmaxc-neonfp16arith-x16.c",
1918 "src/f16-vbinary/gen/vmin-neonfp16arith-x8.c",
1919 "src/f16-vbinary/gen/vmin-neonfp16arith-x16.c",
1920 "src/f16-vbinary/gen/vminc-neonfp16arith-x8.c",
1921 "src/f16-vbinary/gen/vminc-neonfp16arith-x16.c",
1922 "src/f16-vbinary/gen/vmul-minmax-neonfp16arith-x8.c",
1923 "src/f16-vbinary/gen/vmul-minmax-neonfp16arith-x16.c",
1924 "src/f16-vbinary/gen/vmulc-minmax-neonfp16arith-x8.c",
1925 "src/f16-vbinary/gen/vmulc-minmax-neonfp16arith-x16.c",
1926 "src/f16-vbinary/gen/vrdivc-minmax-neonfp16arith-x8.c",
1927 "src/f16-vbinary/gen/vrdivc-minmax-neonfp16arith-x16.c",
1928 "src/f16-vbinary/gen/vrsubc-minmax-neonfp16arith-x8.c",
1929 "src/f16-vbinary/gen/vrsubc-minmax-neonfp16arith-x16.c",
1930 "src/f16-vbinary/gen/vsub-minmax-neonfp16arith-x8.c",
1931 "src/f16-vbinary/gen/vsub-minmax-neonfp16arith-x16.c",
1932 "src/f16-vbinary/gen/vsubc-minmax-neonfp16arith-x8.c",
1933 "src/f16-vbinary/gen/vsubc-minmax-neonfp16arith-x16.c",
1934 "src/f16-vmulcaddc/gen/c8-minmax-neonfp16arith-2x.c",
1935 "src/f16-vmulcaddc/gen/c16-minmax-neonfp16arith-2x.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001936]
1937
Benoit Jacoba9644732020-08-13 12:48:55 -07001938NEONDOT_UKERNELS = [
Marat Dukhand4c83032020-09-09 23:20:12 -07001939 "src/qs8-gemm/gen/1x8c4-minmax-neondot.c",
Marat Dukhand4c83032020-09-09 23:20:12 -07001940 "src/qs8-gemm/gen/1x16c4-minmax-neondot.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001941 "src/qs8-gemm/gen/4x8c4-minmax-neondot.c",
Benoit Jacoba9644732020-08-13 12:48:55 -07001942 "src/qs8-gemm/gen/4x16c4-minmax-neondot.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001943 "src/qs8-gemm/gen/6x8c4-minmax-neondot.c",
Frank Barchard2fa17452020-09-22 09:51:46 -07001944 "src/qs8-gemm/gen/6x16c4-minmax-neondot.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001945 "src/qs8-gemm/gen/8x8c4-minmax-neondot.c",
Frank Barcharda48848f2020-09-24 00:59:24 -07001946 "src/qs8-gemm/gen/8x16c4-minmax-neondot.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001947 "src/qs8-gemm/gen/12x8c4-minmax-neondot.c",
Marat Dukhand4c83032020-09-09 23:20:12 -07001948 "src/qs8-igemm/gen/1x8c4-minmax-neondot.c",
Marat Dukhand4c83032020-09-09 23:20:12 -07001949 "src/qs8-igemm/gen/1x16c4-minmax-neondot.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001950 "src/qs8-igemm/gen/4x8c4-minmax-neondot.c",
Benoit Jacoba05487f2020-08-14 13:37:21 -07001951 "src/qs8-igemm/gen/4x16c4-minmax-neondot.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001952 "src/qs8-igemm/gen/6x8c4-minmax-neondot.c",
Frank Barchard2fa17452020-09-22 09:51:46 -07001953 "src/qs8-igemm/gen/6x16c4-minmax-neondot.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001954 "src/qs8-igemm/gen/8x8c4-minmax-neondot.c",
Frank Barcharda48848f2020-09-24 00:59:24 -07001955 "src/qs8-igemm/gen/8x16c4-minmax-neondot.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001956 "src/qs8-igemm/gen/12x8c4-minmax-neondot.c",
Benoit Jacoba9644732020-08-13 12:48:55 -07001957]
1958
Marat Dukhan08c4a432019-10-03 09:29:21 -07001959SSE_UKERNELS = [
Marat Dukhan99936602020-04-11 16:47:01 -07001960 "src/f32-avgpool/9p8x-minmax-sse-c4.c",
1961 "src/f32-avgpool/9x-minmax-sse-c4.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001962 "src/f32-clamp/gen/sse-x4.c",
1963 "src/f32-clamp/gen/sse-x8.c",
Erich Elsenb1233402020-06-08 15:53:15 -07001964 "src/f32-conv-hwc2chw/3x3s2p1c3x4-sse-1x1.c",
1965 "src/f32-conv-hwc2chw/3x3s2p1c3x4-sse-2x2.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001966 "src/f32-dwconv/gen/up4x4-minmax-sse-acc2.c",
1967 "src/f32-dwconv/gen/up4x4-minmax-sse.c",
1968 "src/f32-dwconv/gen/up4x9-minmax-sse-acc2.c",
1969 "src/f32-dwconv/gen/up4x9-minmax-sse.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001970 "src/f32-dwconv/gen/up4x25-minmax-sse-acc2.c",
1971 "src/f32-dwconv/gen/up4x25-minmax-sse.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001972 "src/f32-dwconv/gen/up8x4-minmax-sse-acc2.c",
1973 "src/f32-dwconv/gen/up8x4-minmax-sse.c",
1974 "src/f32-dwconv/gen/up8x9-minmax-sse-acc2.c",
1975 "src/f32-dwconv/gen/up8x9-minmax-sse.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001976 "src/f32-dwconv/gen/up8x25-minmax-sse-acc2.c",
1977 "src/f32-dwconv/gen/up8x25-minmax-sse.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001978 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-1x4-acc2.c",
1979 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-1x4-acc3.c",
1980 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-1x4-acc4.c",
Marat Dukhan470078a2020-10-23 22:36:52 -07001981 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-1x4.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001982 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-2x4-acc2.c",
Marat Dukhan470078a2020-10-23 22:36:52 -07001983 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-2x4.c",
1984 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-3x4.c",
1985 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-4x4.c",
1986 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-5x4.c",
1987 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-6x4.c",
Marat Dukhan0ff97182020-10-25 19:14:03 -07001988 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-sse-1x4-acc2.c",
1989 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-sse-1x4-acc3.c",
1990 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-sse-1x4-acc4.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001991 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-sse-1x4.c",
Marat Dukhan0ff97182020-10-25 19:14:03 -07001992 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-sse-2x4-acc2.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001993 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-sse-2x4.c",
1994 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-sse-3x4.c",
1995 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-sse-4x4.c",
Marat Dukhand0503892020-10-30 08:22:04 -07001996 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-1x4-acc2.c",
1997 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-1x4-acc3.c",
1998 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-1x4-acc4.c",
1999 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-1x4-acc5.c",
2000 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-1x4.c",
2001 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-2x4-acc2.c",
2002 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-2x4-acc3.c",
2003 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-2x4.c",
2004 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-3x4-acc2.c",
2005 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-3x4.c",
2006 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-4x4-acc2.c",
2007 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-4x4.c",
2008 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-5x4.c",
Marat Dukhanccca2142020-10-30 17:32:45 -07002009 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-sse-1x4-acc2.c",
2010 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-sse-1x4-acc3.c",
2011 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-sse-1x4-acc4.c",
2012 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-sse-1x4-acc5.c",
2013 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-sse-1x4.c",
2014 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-sse-2x4-acc2.c",
2015 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-sse-2x4-acc3.c",
2016 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-sse-2x4.c",
Marat Dukhanccca2142020-10-30 17:32:45 -07002017 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-sse-3x4-acc2.c",
Frank Barchard8ef44cd2020-11-03 12:30:23 -08002018 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-sse-3x4.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -07002019 "src/f32-gavgpool-cw/sse-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -07002020 "src/f32-gavgpool/7p7x-minmax-sse-c4.c",
2021 "src/f32-gavgpool/7x-minmax-sse-c4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002022 "src/f32-gemm/gen-inc/1x8inc-minmax-sse-dup.c",
2023 "src/f32-gemm/gen-inc/1x8inc-minmax-sse-load1.c",
2024 "src/f32-gemm/gen-inc/1x8s4inc-minmax-sse.c",
2025 "src/f32-gemm/gen-inc/4x8inc-minmax-sse-dup.c",
2026 "src/f32-gemm/gen-inc/4x8inc-minmax-sse-load1.c",
2027 "src/f32-gemm/gen-inc/4x8s4inc-minmax-sse.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002028 "src/f32-gemm/gen/1x8-minmax-sse-dup.c",
2029 "src/f32-gemm/gen/1x8-minmax-sse-load1.c",
2030 "src/f32-gemm/gen/1x8s4-minmax-sse.c",
2031 "src/f32-gemm/gen/4x2c4-minmax-sse.c",
2032 "src/f32-gemm/gen/4x8-minmax-sse-dup.c",
2033 "src/f32-gemm/gen/4x8-minmax-sse-load1.c",
2034 "src/f32-gemm/gen/4x8s4-minmax-sse.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07002035 "src/f32-hswish/gen/hswish-sse-x4.c",
2036 "src/f32-hswish/gen/hswish-sse-x8.c",
Marat Dukhan660fd192020-03-10 04:55:30 -07002037 "src/f32-ibilinear/gen/sse-c4.c",
2038 "src/f32-ibilinear/gen/sse-c8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002039 "src/f32-igemm/gen/1x8-minmax-sse-dup.c",
2040 "src/f32-igemm/gen/1x8-minmax-sse-load1.c",
2041 "src/f32-igemm/gen/1x8s4-minmax-sse.c",
2042 "src/f32-igemm/gen/4x2c4-minmax-sse.c",
2043 "src/f32-igemm/gen/4x8-minmax-sse-dup.c",
2044 "src/f32-igemm/gen/4x8-minmax-sse-load1.c",
2045 "src/f32-igemm/gen/4x8s4-minmax-sse.c",
Marat Dukhan99936602020-04-11 16:47:01 -07002046 "src/f32-maxpool/9p8x-minmax-sse-c4.c",
2047 "src/f32-pavgpool/9p8x-minmax-sse-c4.c",
2048 "src/f32-pavgpool/9x-minmax-sse-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002049 "src/f32-ppmm/gen/4x8-minmax-sse.c",
Marat Dukhan39b5e942020-06-24 15:03:48 -07002050 "src/f32-prelu/gen/sse-2x4.c",
2051 "src/f32-prelu/gen/sse-2x8.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07002052 "src/f32-relu/gen/sse-x4.c",
2053 "src/f32-relu/gen/sse-x8.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002054 "src/f32-rmax/sse.c",
Marat Dukhan355ab432020-04-09 19:01:52 -07002055 "src/f32-spmm/gen/4x1-minmax-sse.c",
2056 "src/f32-spmm/gen/8x1-minmax-sse.c",
Erich Elsen6e80fdc2020-06-09 15:35:37 -07002057 "src/f32-spmm/gen/16x1-minmax-sse.c",
Frank Barchard846c0c62020-10-26 15:01:39 -07002058 "src/f32-spmm/gen/32x1-minmax-sse.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002059 "src/f32-vbinary/gen/vadd-minmax-sse-x4.c",
2060 "src/f32-vbinary/gen/vadd-minmax-sse-x8.c",
2061 "src/f32-vbinary/gen/vaddc-minmax-sse-x4.c",
2062 "src/f32-vbinary/gen/vaddc-minmax-sse-x8.c",
2063 "src/f32-vbinary/gen/vdiv-minmax-sse-x4.c",
2064 "src/f32-vbinary/gen/vdiv-minmax-sse-x8.c",
2065 "src/f32-vbinary/gen/vdivc-minmax-sse-x4.c",
2066 "src/f32-vbinary/gen/vdivc-minmax-sse-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -08002067 "src/f32-vbinary/gen/vmax-sse-x4.c",
2068 "src/f32-vbinary/gen/vmax-sse-x8.c",
2069 "src/f32-vbinary/gen/vmaxc-sse-x4.c",
2070 "src/f32-vbinary/gen/vmaxc-sse-x8.c",
2071 "src/f32-vbinary/gen/vmin-sse-x4.c",
2072 "src/f32-vbinary/gen/vmin-sse-x8.c",
2073 "src/f32-vbinary/gen/vminc-sse-x4.c",
2074 "src/f32-vbinary/gen/vminc-sse-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002075 "src/f32-vbinary/gen/vmul-minmax-sse-x4.c",
2076 "src/f32-vbinary/gen/vmul-minmax-sse-x8.c",
2077 "src/f32-vbinary/gen/vmulc-minmax-sse-x4.c",
2078 "src/f32-vbinary/gen/vmulc-minmax-sse-x8.c",
2079 "src/f32-vbinary/gen/vrdivc-minmax-sse-x4.c",
2080 "src/f32-vbinary/gen/vrdivc-minmax-sse-x8.c",
2081 "src/f32-vbinary/gen/vrsubc-minmax-sse-x4.c",
2082 "src/f32-vbinary/gen/vrsubc-minmax-sse-x8.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07002083 "src/f32-vbinary/gen/vsqrdiff-sse-x4.c",
2084 "src/f32-vbinary/gen/vsqrdiff-sse-x8.c",
2085 "src/f32-vbinary/gen/vsqrdiffc-sse-x4.c",
2086 "src/f32-vbinary/gen/vsqrdiffc-sse-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002087 "src/f32-vbinary/gen/vsub-minmax-sse-x4.c",
2088 "src/f32-vbinary/gen/vsub-minmax-sse-x8.c",
2089 "src/f32-vbinary/gen/vsubc-minmax-sse-x4.c",
2090 "src/f32-vbinary/gen/vsubc-minmax-sse-x8.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07002091 "src/f32-vlrelu/gen/vlrelu-sse-x4.c",
2092 "src/f32-vlrelu/gen/vlrelu-sse-x8.c",
Marat Dukhan99936602020-04-11 16:47:01 -07002093 "src/f32-vmulcaddc/gen/c4-minmax-sse-2x.c",
2094 "src/f32-vmulcaddc/gen/c8-minmax-sse-2x.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07002095 "src/f32-vsqrt/gen/sse-sqrt-x4.c",
2096 "src/f32-vsqrt/gen/sse-sqrt-x8.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07002097 "src/f32-vunary/gen/vabs-sse-x4.c",
2098 "src/f32-vunary/gen/vabs-sse-x8.c",
2099 "src/f32-vunary/gen/vneg-sse-x4.c",
2100 "src/f32-vunary/gen/vneg-sse-x8.c",
2101 "src/f32-vunary/gen/vsqr-sse-x4.c",
2102 "src/f32-vunary/gen/vsqr-sse-x8.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07002103 "src/math/roundd-sse-addsub.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002104 "src/math/roundne-sse-addsub.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07002105 "src/math/roundu-sse-addsub.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07002106 "src/math/roundz-sse-addsub.c",
Marat Dukhan84000762020-06-29 18:38:43 -07002107 "src/math/sqrt-sse-hh1mac.c",
2108 "src/math/sqrt-sse-nr1mac.c",
2109 "src/math/sqrt-sse-nr2mac.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002110 "src/x32-fill/sse.c",
2111 "src/x32-packx/x4-sse.c",
2112 "src/x32-pad/sse.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002113]
2114
2115SSE2_UKERNELS = [
Marat Dukhan329da642019-11-19 21:44:39 -08002116 "src/f32-argmaxpool/4x-sse2-c4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002117 "src/f32-argmaxpool/9p8x-sse2-c4.c",
Marat Dukhan329da642019-11-19 21:44:39 -08002118 "src/f32-argmaxpool/9x-sse2-c4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08002119 "src/f32-prelu/gen/sse2-2x4.c",
2120 "src/f32-prelu/gen/sse2-2x8.c",
Marat Dukhanb39689d2020-01-24 13:32:20 -08002121 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x4.c",
Marat Dukhanb39689d2020-01-24 13:32:20 -08002122 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x8-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002123 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x8.c",
Marat Dukhanb39689d2020-01-24 13:32:20 -08002124 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12-acc2.c",
2125 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002126 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12.c",
Marat Dukhanb39689d2020-01-24 13:32:20 -08002127 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16-acc2.c",
2128 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002129 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16.c",
Marat Dukhanb39689d2020-01-24 13:32:20 -08002130 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20-acc2.c",
2131 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002132 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20.c",
Marat Dukhand243c1a2020-09-17 18:48:10 -07002133 "src/f32-sigmoid/gen/sse2-lut64-p2-div-x4.c",
2134 "src/f32-sigmoid/gen/sse2-lut64-p2-div-x8.c",
2135 "src/f32-sigmoid/gen/sse2-lut64-p2-div-x12.c",
2136 "src/f32-sigmoid/gen/sse2-lut64-p2-div-x16.c",
2137 "src/f32-sigmoid/gen/sse2-lut64-p2-div-x20.c",
2138 "src/f32-sigmoid/gen/sse2-lut64-p2-div-x24.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002139 "src/f32-sigmoid/gen/sse2-p5-div-x4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08002140 "src/f32-sigmoid/gen/sse2-p5-div-x8.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002141 "src/f32-sigmoid/gen/sse2-p5-div-x12.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08002142 "src/f32-sigmoid/gen/sse2-p5-div-x16.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002143 "src/f32-sigmoid/gen/sse2-p5-div-x20.c",
2144 "src/f32-sigmoid/gen/sse2-p5-div-x24.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07002145 "src/f32-vlrelu/gen/vlrelu-sse2-x4.c",
2146 "src/f32-vlrelu/gen/vlrelu-sse2-x8.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07002147 "src/f32-vrnd/gen/vrndd-sse2-x4.c",
2148 "src/f32-vrnd/gen/vrndd-sse2-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002149 "src/f32-vrnd/gen/vrndne-sse2-x4.c",
2150 "src/f32-vrnd/gen/vrndne-sse2-x8.c",
2151 "src/f32-vrnd/gen/vrndu-sse2-x4.c",
2152 "src/f32-vrnd/gen/vrndu-sse2-x8.c",
2153 "src/f32-vrnd/gen/vrndz-sse2-x4.c",
2154 "src/f32-vrnd/gen/vrndz-sse2-x8.c",
Marat Dukhanb7633f22020-11-20 16:34:56 -08002155 "src/math/exp-sse2-rr2-lut64-p2.c",
2156 "src/math/exp-sse2-rr2-p5.c",
Marat Dukhand6d321e2020-11-20 18:35:55 -08002157 "src/math/expm1minus-sse2-rr2-p5.c",
Marat Dukhana438aca2020-11-20 15:45:01 -08002158 "src/math/expm1minus-sse2-rr2-p6.c",
Frank Barchard3b800452020-11-22 12:12:35 -08002159 "src/math/expminus-sse2-rr2-p5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002160 "src/math/roundd-sse2-cvt.c",
2161 "src/math/roundne-sse2-cvt.c",
2162 "src/math/roundu-sse2-cvt.c",
2163 "src/math/roundz-sse2-cvt.c",
2164 "src/math/sigmoid-sse2-rr2-lut64-p2-div.c",
2165 "src/math/sigmoid-sse2-rr2-lut64-p2-nr1.c",
2166 "src/math/sigmoid-sse2-rr2-lut64-p2-nr2.c",
2167 "src/math/sigmoid-sse2-rr2-p5-div.c",
2168 "src/math/sigmoid-sse2-rr2-p5-nr1.c",
2169 "src/math/sigmoid-sse2-rr2-p5-nr2.c",
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07002170 "src/qs8-dwconv/gen/up8x9-minmax-sse2-mul16.c",
2171 "src/qs8-dwconv/gen/up16x9-minmax-sse2-mul16.c",
2172 "src/qs8-dwconv/gen/up24x9-minmax-sse2-mul16.c",
Marat Dukhan159688f2020-08-06 10:34:29 -07002173 "src/qs8-gavgpool/gen/7p7x-minmax-sse2-c8-acc2.c",
2174 "src/qs8-gavgpool/gen/7p7x-minmax-sse2-c16-acc2.c",
2175 "src/qs8-gavgpool/gen/7p7x-minmax-sse2-c24-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002176 "src/qs8-gavgpool/gen/7x-minmax-sse2-c8-acc2.c",
2177 "src/qs8-gavgpool/gen/7x-minmax-sse2-c16-acc2.c",
2178 "src/qs8-gavgpool/gen/7x-minmax-sse2-c24-acc2.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002179 "src/qs8-gemm/gen/1x4c2-minmax-sse2-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002180 "src/qs8-gemm/gen/1x4c2-minmax-sse2-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002181 "src/qs8-gemm/gen/1x4c2-xw-minmax-sse2.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002182 "src/qs8-gemm/gen/1x4c8-minmax-sse2-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002183 "src/qs8-gemm/gen/1x4c8-minmax-sse2-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002184 "src/qs8-gemm/gen/1x4c8-xw-minmax-sse2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002185 "src/qs8-gemm/gen/2x4c8-minmax-sse2-ld64.c",
2186 "src/qs8-gemm/gen/2x4c8-minmax-sse2-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002187 "src/qs8-gemm/gen/2x4c8-xw-minmax-sse2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002188 "src/qs8-gemm/gen/3x4c8-minmax-sse2-ld64.c",
2189 "src/qs8-gemm/gen/3x4c8-minmax-sse2-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002190 "src/qs8-gemm/gen/3x4c8-xw-minmax-sse2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002191 "src/qs8-gemm/gen/4x4c2-minmax-sse2-ld64.c",
2192 "src/qs8-gemm/gen/4x4c2-minmax-sse2-ld128.c",
2193 "src/qs8-gemm/gen/4x4c2-xw-minmax-sse2.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002194 "src/qs8-igemm/gen/1x4c2-minmax-sse2-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002195 "src/qs8-igemm/gen/1x4c2-minmax-sse2-ld128.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002196 "src/qs8-igemm/gen/1x4c8-minmax-sse2-ld64.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002197 "src/qs8-igemm/gen/1x4c8-minmax-sse2-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002198 "src/qs8-igemm/gen/2x4c8-minmax-sse2-ld64.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002199 "src/qs8-igemm/gen/2x4c8-minmax-sse2-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002200 "src/qs8-igemm/gen/3x4c8-minmax-sse2-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002201 "src/qs8-igemm/gen/3x4c8-minmax-sse2-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002202 "src/qs8-igemm/gen/4x4c2-minmax-sse2-ld64.c",
2203 "src/qs8-igemm/gen/4x4c2-minmax-sse2-ld128.c",
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07002204 "src/qs8-requantization/fp32-sse2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002205 "src/qs8-requantization/precise-sse2.c",
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07002206 "src/qs8-requantization/q31-sse2.c",
Marat Dukhand9f3ad42020-08-10 12:30:58 -07002207 "src/qs8-vadd/gen/minmax-sse2-mul16-ld64-x8.c",
2208 "src/qs8-vadd/gen/minmax-sse2-mul16-ld64-x16.c",
2209 "src/qs8-vadd/gen/minmax-sse2-mul16-ld64-x24.c",
2210 "src/qs8-vadd/gen/minmax-sse2-mul16-ld64-x32.c",
Marat Dukhan0270d9f2020-08-11 00:56:46 -07002211 "src/qs8-vaddc/gen/minmax-sse2-mul16-ld64-x8.c",
2212 "src/qs8-vaddc/gen/minmax-sse2-mul16-ld64-x16.c",
2213 "src/qs8-vaddc/gen/minmax-sse2-mul16-ld64-x24.c",
2214 "src/qs8-vaddc/gen/minmax-sse2-mul16-ld64-x32.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -07002215 "src/qu8-avgpool/9p8x-minmax-sse2-c8.c",
2216 "src/qu8-avgpool/9x-minmax-sse2-c8.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -07002217 "src/qu8-dwconv/up8x9-minmax-sse2.c",
2218 "src/qu8-gavgpool/7p7x-minmax-sse2-c8.c",
2219 "src/qu8-gavgpool/7x-minmax-sse2-c8.c",
2220 "src/qu8-gemm/2x4c8-minmax-sse2.c",
2221 "src/qu8-gemm/4x4c2-minmax-sse2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002222 "src/qu8-igemm/4x4c2-minmax-sse2.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07002223 "src/qu8-requantization/fp32-sse2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002224 "src/qu8-requantization/precise-sse2.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07002225 "src/qu8-requantization/q31-sse2.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -07002226 "src/qu8-vadd/minmax-sse2.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -07002227 "src/u8-clamp/sse2-x64.c",
Marat Dukhan99936602020-04-11 16:47:01 -07002228 "src/u8-maxpool/9p8x-minmax-sse2-c16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002229 "src/u8-rmax/sse2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002230 "src/x8-zip/x2-sse2.c",
2231 "src/x8-zip/x3-sse2.c",
2232 "src/x8-zip/x4-sse2.c",
2233 "src/x8-zip/xm-sse2.c",
Marat Dukhan57dccd82020-04-14 00:53:10 -07002234 "src/x32-unpool/sse2.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002235 "src/x32-zip/x2-sse2.c",
2236 "src/x32-zip/x3-sse2.c",
2237 "src/x32-zip/x4-sse2.c",
2238 "src/x32-zip/xm-sse2.c",
Marat Dukhanfe7acb62020-03-09 19:30:05 -07002239]
2240
2241SSSE3_UKERNELS = [
Frank Barchard35db7d02020-10-26 13:37:34 -07002242 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-1x4-acc2.c",
2243 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-1x4-acc3.c",
2244 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-1x4-acc4.c",
Marat Dukhan98f2eeb2020-10-23 23:13:41 -07002245 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-1x4.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07002246 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-2x4-acc2.c",
Marat Dukhan98f2eeb2020-10-23 23:13:41 -07002247 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-2x4.c",
2248 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-3x4.c",
2249 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-4x4.c",
2250 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-5x4.c",
2251 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-6x4.c",
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07002252 "src/qs8-dwconv/gen/up8x9-minmax-ssse3-mul16.c",
2253 "src/qs8-dwconv/gen/up16x9-minmax-ssse3-mul16.c",
2254 "src/qs8-dwconv/gen/up24x9-minmax-ssse3-mul16.c",
Marat Dukhan159688f2020-08-06 10:34:29 -07002255 "src/qs8-gavgpool/gen/7p7x-minmax-ssse3-c8-acc2.c",
2256 "src/qs8-gavgpool/gen/7p7x-minmax-ssse3-c16-acc2.c",
2257 "src/qs8-gavgpool/gen/7p7x-minmax-ssse3-c24-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002258 "src/qs8-gavgpool/gen/7x-minmax-ssse3-c8-acc2.c",
2259 "src/qs8-gavgpool/gen/7x-minmax-ssse3-c16-acc2.c",
2260 "src/qs8-gavgpool/gen/7x-minmax-ssse3-c24-acc2.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002261 "src/qs8-gemm/gen/1x4c2-minmax-ssse3-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002262 "src/qs8-gemm/gen/1x4c2-minmax-ssse3-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002263 "src/qs8-gemm/gen/1x4c2-xw-minmax-ssse3.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002264 "src/qs8-gemm/gen/1x4c8-minmax-ssse3-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002265 "src/qs8-gemm/gen/1x4c8-minmax-ssse3-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002266 "src/qs8-gemm/gen/1x4c8-xw-minmax-ssse3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002267 "src/qs8-gemm/gen/2x4c8-minmax-ssse3-ld64.c",
2268 "src/qs8-gemm/gen/2x4c8-minmax-ssse3-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002269 "src/qs8-gemm/gen/2x4c8-xw-minmax-ssse3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002270 "src/qs8-gemm/gen/3x4c8-minmax-ssse3-ld64.c",
2271 "src/qs8-gemm/gen/3x4c8-minmax-ssse3-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002272 "src/qs8-gemm/gen/3x4c8-xw-minmax-ssse3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002273 "src/qs8-gemm/gen/4x4c2-minmax-ssse3-ld64.c",
2274 "src/qs8-gemm/gen/4x4c2-minmax-ssse3-ld128.c",
2275 "src/qs8-gemm/gen/4x4c2-xw-minmax-ssse3.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002276 "src/qs8-igemm/gen/1x4c2-minmax-ssse3-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002277 "src/qs8-igemm/gen/1x4c2-minmax-ssse3-ld128.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002278 "src/qs8-igemm/gen/1x4c8-minmax-ssse3-ld64.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002279 "src/qs8-igemm/gen/1x4c8-minmax-ssse3-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002280 "src/qs8-igemm/gen/2x4c8-minmax-ssse3-ld64.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002281 "src/qs8-igemm/gen/2x4c8-minmax-ssse3-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002282 "src/qs8-igemm/gen/3x4c8-minmax-ssse3-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002283 "src/qs8-igemm/gen/3x4c8-minmax-ssse3-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002284 "src/qs8-igemm/gen/4x4c2-minmax-ssse3-ld64.c",
2285 "src/qs8-igemm/gen/4x4c2-minmax-ssse3-ld128.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -07002286 "src/qs8-requantization/precise-ssse3.c",
2287 "src/qs8-requantization/q31-ssse3.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07002288 "src/qu8-requantization/precise-ssse3.c",
2289 "src/qu8-requantization/q31-ssse3.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002290]
2291
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08002292SSE41_UKERNELS = [
Marat Dukhan40a672f2019-11-25 03:08:22 -08002293 "src/f32-prelu/gen/sse41-2x4.c",
2294 "src/f32-prelu/gen/sse41-2x8.c",
Marat Dukhand243c1a2020-09-17 18:48:10 -07002295 "src/f32-sigmoid/gen/sse41-lut64-p2-div-x4.c",
2296 "src/f32-sigmoid/gen/sse41-lut64-p2-div-x8.c",
2297 "src/f32-sigmoid/gen/sse41-lut64-p2-div-x12.c",
2298 "src/f32-sigmoid/gen/sse41-lut64-p2-div-x16.c",
2299 "src/f32-sigmoid/gen/sse41-lut64-p2-div-x20.c",
2300 "src/f32-sigmoid/gen/sse41-lut64-p2-div-x24.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002301 "src/f32-sigmoid/gen/sse41-p5-div-x4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08002302 "src/f32-sigmoid/gen/sse41-p5-div-x8.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002303 "src/f32-sigmoid/gen/sse41-p5-div-x12.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08002304 "src/f32-sigmoid/gen/sse41-p5-div-x16.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002305 "src/f32-sigmoid/gen/sse41-p5-div-x20.c",
2306 "src/f32-sigmoid/gen/sse41-p5-div-x24.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07002307 "src/f32-vlrelu/gen/vlrelu-sse41-x4.c",
2308 "src/f32-vlrelu/gen/vlrelu-sse41-x8.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07002309 "src/f32-vrnd/gen/vrndd-sse41-x4.c",
2310 "src/f32-vrnd/gen/vrndd-sse41-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002311 "src/f32-vrnd/gen/vrndne-sse41-x4.c",
2312 "src/f32-vrnd/gen/vrndne-sse41-x8.c",
2313 "src/f32-vrnd/gen/vrndu-sse41-x4.c",
2314 "src/f32-vrnd/gen/vrndu-sse41-x8.c",
2315 "src/f32-vrnd/gen/vrndz-sse41-x4.c",
2316 "src/f32-vrnd/gen/vrndz-sse41-x8.c",
2317 "src/math/roundd-sse41.c",
2318 "src/math/roundne-sse41.c",
2319 "src/math/roundu-sse41.c",
2320 "src/math/roundz-sse41.c",
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07002321 "src/qs8-dwconv/gen/up8x9-minmax-sse41-mul16.c",
2322 "src/qs8-dwconv/gen/up16x9-minmax-sse41-mul16.c",
2323 "src/qs8-dwconv/gen/up24x9-minmax-sse41-mul16.c",
Marat Dukhan159688f2020-08-06 10:34:29 -07002324 "src/qs8-gavgpool/gen/7p7x-minmax-sse41-c8-acc2.c",
2325 "src/qs8-gavgpool/gen/7p7x-minmax-sse41-c16-acc2.c",
2326 "src/qs8-gavgpool/gen/7p7x-minmax-sse41-c24-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002327 "src/qs8-gavgpool/gen/7x-minmax-sse41-c8-acc2.c",
2328 "src/qs8-gavgpool/gen/7x-minmax-sse41-c16-acc2.c",
2329 "src/qs8-gavgpool/gen/7x-minmax-sse41-c24-acc2.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002330 "src/qs8-gemm/gen/1x4c2-minmax-sse41-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002331 "src/qs8-gemm/gen/1x4c2-minmax-sse41-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002332 "src/qs8-gemm/gen/1x4c2-xw-minmax-sse41.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002333 "src/qs8-gemm/gen/1x4c8-minmax-sse41-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002334 "src/qs8-gemm/gen/1x4c8-minmax-sse41-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002335 "src/qs8-gemm/gen/1x4c8-xw-minmax-sse41.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002336 "src/qs8-gemm/gen/2x4c8-minmax-sse41-ld64.c",
2337 "src/qs8-gemm/gen/2x4c8-minmax-sse41-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002338 "src/qs8-gemm/gen/2x4c8-xw-minmax-sse41.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002339 "src/qs8-gemm/gen/3x4c8-minmax-sse41-ld64.c",
2340 "src/qs8-gemm/gen/3x4c8-minmax-sse41-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002341 "src/qs8-gemm/gen/3x4c8-xw-minmax-sse41.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002342 "src/qs8-gemm/gen/4x4c2-minmax-sse41-ld64.c",
2343 "src/qs8-gemm/gen/4x4c2-minmax-sse41-ld128.c",
2344 "src/qs8-gemm/gen/4x4c2-xw-minmax-sse41.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002345 "src/qs8-igemm/gen/1x4c2-minmax-sse41-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002346 "src/qs8-igemm/gen/1x4c2-minmax-sse41-ld128.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002347 "src/qs8-igemm/gen/1x4c8-minmax-sse41-ld64.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002348 "src/qs8-igemm/gen/1x4c8-minmax-sse41-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002349 "src/qs8-igemm/gen/2x4c8-minmax-sse41-ld64.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002350 "src/qs8-igemm/gen/2x4c8-minmax-sse41-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002351 "src/qs8-igemm/gen/3x4c8-minmax-sse41-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002352 "src/qs8-igemm/gen/3x4c8-minmax-sse41-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002353 "src/qs8-igemm/gen/4x4c2-minmax-sse41-ld64.c",
2354 "src/qs8-igemm/gen/4x4c2-minmax-sse41-ld128.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -07002355 "src/qs8-requantization/fp32-sse4.c",
2356 "src/qs8-requantization/precise-sse4.c",
2357 "src/qs8-requantization/q31-sse4.c",
Marat Dukhand9f3ad42020-08-10 12:30:58 -07002358 "src/qs8-vadd/gen/minmax-sse41-mul16-ld64-x8.c",
2359 "src/qs8-vadd/gen/minmax-sse41-mul16-ld64-x16.c",
2360 "src/qs8-vadd/gen/minmax-sse41-mul16-ld64-x24.c",
2361 "src/qs8-vadd/gen/minmax-sse41-mul16-ld64-x32.c",
Marat Dukhanbb9225e2020-09-06 22:40:56 -07002362 "src/qs8-vadd/gen/minmax-sse41-mul32-ld32-x8.c",
2363 "src/qs8-vadd/gen/minmax-sse41-mul32-ld32-x16.c",
2364 "src/qs8-vadd/gen/minmax-sse41-mul32-ld32-x24.c",
2365 "src/qs8-vadd/gen/minmax-sse41-mul32-ld32-x32.c",
Marat Dukhan0270d9f2020-08-11 00:56:46 -07002366 "src/qs8-vaddc/gen/minmax-sse41-mul16-ld64-x8.c",
2367 "src/qs8-vaddc/gen/minmax-sse41-mul16-ld64-x16.c",
2368 "src/qs8-vaddc/gen/minmax-sse41-mul16-ld64-x24.c",
2369 "src/qs8-vaddc/gen/minmax-sse41-mul16-ld64-x32.c",
Marat Dukhanbb9225e2020-09-06 22:40:56 -07002370 "src/qs8-vaddc/gen/minmax-sse41-mul32-ld32-x8.c",
2371 "src/qs8-vaddc/gen/minmax-sse41-mul32-ld32-x16.c",
2372 "src/qs8-vaddc/gen/minmax-sse41-mul32-ld32-x24.c",
2373 "src/qs8-vaddc/gen/minmax-sse41-mul32-ld32-x32.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07002374 "src/qu8-requantization/precise-sse4.c",
2375 "src/qu8-requantization/q31-sse4.c",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08002376]
2377
Marat Dukhan08c4a432019-10-03 09:29:21 -07002378AVX_UKERNELS = [
Marat Dukhan5c5fa962020-03-10 18:38:33 -07002379 "src/f32-clamp/gen/avx-x8.c",
2380 "src/f32-clamp/gen/avx-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002381 "src/f32-dwconv/gen/up8x4-minmax-avx-acc2.c",
2382 "src/f32-dwconv/gen/up8x4-minmax-avx.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002383 "src/f32-dwconv/gen/up8x9-minmax-avx-acc2.c",
2384 "src/f32-dwconv/gen/up8x9-minmax-avx.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002385 "src/f32-dwconv/gen/up8x25-minmax-avx-acc2.c",
2386 "src/f32-dwconv/gen/up8x25-minmax-avx.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002387 "src/f32-dwconv/gen/up16x4-minmax-avx-acc2.c",
2388 "src/f32-dwconv/gen/up16x4-minmax-avx.c",
2389 "src/f32-dwconv/gen/up16x9-minmax-avx-acc2.c",
2390 "src/f32-dwconv/gen/up16x9-minmax-avx.c",
2391 "src/f32-dwconv/gen/up16x25-minmax-avx-acc2.c",
2392 "src/f32-dwconv/gen/up16x25-minmax-avx.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002393 "src/f32-gemm/gen-inc/1x8inc-minmax-avx-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002394 "src/f32-gemm/gen-inc/1x16inc-minmax-avx-broadcast.c",
2395 "src/f32-gemm/gen-inc/3x16inc-minmax-avx-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002396 "src/f32-gemm/gen-inc/4x8inc-minmax-avx-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002397 "src/f32-gemm/gen-inc/4x16inc-minmax-avx-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002398 "src/f32-gemm/gen-inc/5x8inc-minmax-avx-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002399 "src/f32-gemm/gen-inc/5x16inc-minmax-avx-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002400 "src/f32-gemm/gen-inc/6x8inc-minmax-avx-broadcast.c",
2401 "src/f32-gemm/gen-inc/7x8inc-minmax-avx-broadcast.c",
2402 "src/f32-gemm/gen/1x8-minmax-avx-broadcast.c",
2403 "src/f32-gemm/gen/1x16-minmax-avx-broadcast.c",
2404 "src/f32-gemm/gen/3x16-minmax-avx-broadcast.c",
2405 "src/f32-gemm/gen/4x8-minmax-avx-broadcast.c",
2406 "src/f32-gemm/gen/4x16-minmax-avx-broadcast.c",
2407 "src/f32-gemm/gen/5x8-minmax-avx-broadcast.c",
2408 "src/f32-gemm/gen/5x16-minmax-avx-broadcast.c",
2409 "src/f32-gemm/gen/6x8-minmax-avx-broadcast.c",
2410 "src/f32-gemm/gen/7x8-minmax-avx-broadcast.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07002411 "src/f32-hswish/gen/hswish-avx-x8.c",
2412 "src/f32-hswish/gen/hswish-avx-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002413 "src/f32-igemm/gen/1x8-minmax-avx-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002414 "src/f32-igemm/gen/1x16-minmax-avx-broadcast.c",
2415 "src/f32-igemm/gen/3x16-minmax-avx-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002416 "src/f32-igemm/gen/4x8-minmax-avx-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002417 "src/f32-igemm/gen/4x16-minmax-avx-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002418 "src/f32-igemm/gen/5x8-minmax-avx-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002419 "src/f32-igemm/gen/5x16-minmax-avx-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002420 "src/f32-igemm/gen/6x8-minmax-avx-broadcast.c",
2421 "src/f32-igemm/gen/7x8-minmax-avx-broadcast.c",
Marat Dukhan90eca0a2020-03-11 00:52:23 -07002422 "src/f32-prelu/gen/avx-2x8.c",
2423 "src/f32-prelu/gen/avx-2x16.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07002424 "src/f32-relu/gen/avx-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002425 "src/f32-relu/gen/avx-x16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002426 "src/f32-rmax/avx.c",
T.J. Alumbaughdc2b29c2020-10-14 13:56:08 -07002427 "src/f32-sigmoid/gen/avx-rr2-p5-div-x8.c",
2428 "src/f32-sigmoid/gen/avx-rr2-p5-div-x16.c",
2429 "src/f32-sigmoid/gen/avx-rr2-p5-div-x24.c",
2430 "src/f32-sigmoid/gen/avx-rr2-p5-div-x32.c",
2431 "src/f32-sigmoid/gen/avx-rr2-p5-div-x40.c",
2432 "src/f32-sigmoid/gen/avx-rr2-p5-div-x48.c",
2433 "src/f32-sigmoid/gen/avx-rr2-p5-div-x56.c",
2434 "src/f32-sigmoid/gen/avx-rr2-p5-div-x64.c",
2435 "src/f32-sigmoid/gen/avx-rr2-p5-div-x72.c",
2436 "src/f32-sigmoid/gen/avx-rr2-p5-div-x80.c",
2437 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x8.c",
2438 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x16.c",
2439 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x24.c",
2440 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x32.c",
2441 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x40.c",
2442 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x48.c",
2443 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x56.c",
2444 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x64.c",
2445 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x72.c",
2446 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x80.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002447 "src/f32-vbinary/gen/vadd-minmax-avx-x8.c",
2448 "src/f32-vbinary/gen/vadd-minmax-avx-x16.c",
2449 "src/f32-vbinary/gen/vaddc-minmax-avx-x8.c",
2450 "src/f32-vbinary/gen/vaddc-minmax-avx-x16.c",
2451 "src/f32-vbinary/gen/vdiv-minmax-avx-x8.c",
2452 "src/f32-vbinary/gen/vdiv-minmax-avx-x16.c",
2453 "src/f32-vbinary/gen/vdivc-minmax-avx-x8.c",
2454 "src/f32-vbinary/gen/vdivc-minmax-avx-x16.c",
Marat Dukhan9a88efe2019-12-10 15:54:24 -08002455 "src/f32-vbinary/gen/vmax-avx-x8.c",
2456 "src/f32-vbinary/gen/vmax-avx-x16.c",
2457 "src/f32-vbinary/gen/vmaxc-avx-x8.c",
2458 "src/f32-vbinary/gen/vmaxc-avx-x16.c",
2459 "src/f32-vbinary/gen/vmin-avx-x8.c",
2460 "src/f32-vbinary/gen/vmin-avx-x16.c",
2461 "src/f32-vbinary/gen/vminc-avx-x8.c",
2462 "src/f32-vbinary/gen/vminc-avx-x16.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002463 "src/f32-vbinary/gen/vmul-minmax-avx-x8.c",
2464 "src/f32-vbinary/gen/vmul-minmax-avx-x16.c",
2465 "src/f32-vbinary/gen/vmulc-minmax-avx-x8.c",
2466 "src/f32-vbinary/gen/vmulc-minmax-avx-x16.c",
2467 "src/f32-vbinary/gen/vrdivc-minmax-avx-x8.c",
2468 "src/f32-vbinary/gen/vrdivc-minmax-avx-x16.c",
2469 "src/f32-vbinary/gen/vrsubc-minmax-avx-x8.c",
2470 "src/f32-vbinary/gen/vrsubc-minmax-avx-x16.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07002471 "src/f32-vbinary/gen/vsqrdiff-avx-x8.c",
2472 "src/f32-vbinary/gen/vsqrdiff-avx-x16.c",
2473 "src/f32-vbinary/gen/vsqrdiffc-avx-x8.c",
2474 "src/f32-vbinary/gen/vsqrdiffc-avx-x16.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002475 "src/f32-vbinary/gen/vsub-minmax-avx-x8.c",
2476 "src/f32-vbinary/gen/vsub-minmax-avx-x16.c",
2477 "src/f32-vbinary/gen/vsubc-minmax-avx-x8.c",
2478 "src/f32-vbinary/gen/vsubc-minmax-avx-x16.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07002479 "src/f32-vlrelu/gen/vlrelu-avx-x8.c",
2480 "src/f32-vlrelu/gen/vlrelu-avx-x16.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07002481 "src/f32-vrnd/gen/vrndd-avx-x8.c",
2482 "src/f32-vrnd/gen/vrndd-avx-x16.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002483 "src/f32-vrnd/gen/vrndne-avx-x8.c",
2484 "src/f32-vrnd/gen/vrndne-avx-x16.c",
2485 "src/f32-vrnd/gen/vrndu-avx-x8.c",
2486 "src/f32-vrnd/gen/vrndu-avx-x16.c",
2487 "src/f32-vrnd/gen/vrndz-avx-x8.c",
2488 "src/f32-vrnd/gen/vrndz-avx-x16.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07002489 "src/f32-vscale/avx-x32.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07002490 "src/f32-vsqrt/gen/avx-sqrt-x8.c",
2491 "src/f32-vsqrt/gen/avx-sqrt-x16.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07002492 "src/f32-vunary/gen/vabs-avx-x8.c",
2493 "src/f32-vunary/gen/vabs-avx-x16.c",
2494 "src/f32-vunary/gen/vneg-avx-x8.c",
2495 "src/f32-vunary/gen/vneg-avx-x16.c",
2496 "src/f32-vunary/gen/vsqr-avx-x8.c",
2497 "src/f32-vunary/gen/vsqr-avx-x16.c",
Marat Dukhanb7633f22020-11-20 16:34:56 -08002498 "src/math/exp-avx-rr2-p5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002499 "src/math/sigmoid-avx-rr2-lut64-p2-div.c",
Marat Dukhan2e4727b2020-09-24 00:48:39 -07002500 "src/math/sigmoid-avx-rr2-p5-div.c",
Marat Dukhan39ab16f2020-10-16 14:30:35 -07002501 "src/math/sigmoid-avx-rr2-p5-nr1.c",
2502 "src/math/sigmoid-avx-rr2-p5-nr2.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002503]
2504
Marat Dukhan1566fee2020-08-02 21:55:41 -07002505XOP_UKERNELS = [
2506 "src/qs8-gemm/gen/1x4c2-minmax-xop-ld64.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002507 "src/qs8-gemm/gen/1x4c2-minmax-xop-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002508 "src/qs8-gemm/gen/1x4c2-xw-minmax-xop.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002509 "src/qs8-gemm/gen/1x4c8-minmax-xop-ld64.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002510 "src/qs8-gemm/gen/1x4c8-minmax-xop-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002511 "src/qs8-gemm/gen/1x4c8-xw-minmax-xop.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002512 "src/qs8-gemm/gen/2x4c8-minmax-xop-ld64.c",
2513 "src/qs8-gemm/gen/2x4c8-minmax-xop-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002514 "src/qs8-gemm/gen/2x4c8-xw-minmax-xop.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002515 "src/qs8-gemm/gen/3x4c8-minmax-xop-ld64.c",
2516 "src/qs8-gemm/gen/3x4c8-minmax-xop-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002517 "src/qs8-gemm/gen/3x4c8-xw-minmax-xop.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002518 "src/qs8-gemm/gen/4x4c2-minmax-xop-ld64.c",
2519 "src/qs8-gemm/gen/4x4c2-minmax-xop-ld128.c",
2520 "src/qs8-gemm/gen/4x4c2-xw-minmax-xop.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002521 "src/qs8-igemm/gen/1x4c2-minmax-xop-ld64.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002522 "src/qs8-igemm/gen/1x4c2-minmax-xop-ld128.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002523 "src/qs8-igemm/gen/1x4c8-minmax-xop-ld64.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002524 "src/qs8-igemm/gen/1x4c8-minmax-xop-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002525 "src/qs8-igemm/gen/2x4c8-minmax-xop-ld64.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002526 "src/qs8-igemm/gen/2x4c8-minmax-xop-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002527 "src/qs8-igemm/gen/3x4c8-minmax-xop-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002528 "src/qs8-igemm/gen/3x4c8-minmax-xop-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002529 "src/qs8-igemm/gen/4x4c2-minmax-xop-ld64.c",
2530 "src/qs8-igemm/gen/4x4c2-minmax-xop-ld128.c",
Marat Dukhanbb9225e2020-09-06 22:40:56 -07002531 "src/qs8-vadd/gen/minmax-xop-mul32-ld32-x8.c",
2532 "src/qs8-vadd/gen/minmax-xop-mul32-ld32-x16.c",
2533 "src/qs8-vadd/gen/minmax-xop-mul32-ld32-x24.c",
2534 "src/qs8-vadd/gen/minmax-xop-mul32-ld32-x32.c",
2535 "src/qs8-vaddc/gen/minmax-xop-mul32-ld32-x8.c",
2536 "src/qs8-vaddc/gen/minmax-xop-mul32-ld32-x16.c",
2537 "src/qs8-vaddc/gen/minmax-xop-mul32-ld32-x24.c",
2538 "src/qs8-vaddc/gen/minmax-xop-mul32-ld32-x32.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002539]
2540
Marat Dukhanfda12b82019-11-21 12:27:59 -08002541FMA3_UKERNELS = [
Marat Dukhan1c587112020-04-08 20:04:28 -07002542 "src/f32-dwconv/gen/up8x4-minmax-fma3-acc2.c",
2543 "src/f32-dwconv/gen/up8x4-minmax-fma3.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002544 "src/f32-dwconv/gen/up8x9-minmax-fma3-acc2.c",
2545 "src/f32-dwconv/gen/up8x9-minmax-fma3.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002546 "src/f32-dwconv/gen/up8x25-minmax-fma3-acc2.c",
2547 "src/f32-dwconv/gen/up8x25-minmax-fma3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002548 "src/f32-dwconv/gen/up16x4-minmax-fma3-acc2.c",
2549 "src/f32-dwconv/gen/up16x4-minmax-fma3.c",
2550 "src/f32-dwconv/gen/up16x9-minmax-fma3-acc2.c",
2551 "src/f32-dwconv/gen/up16x9-minmax-fma3.c",
2552 "src/f32-dwconv/gen/up16x25-minmax-fma3-acc2.c",
2553 "src/f32-dwconv/gen/up16x25-minmax-fma3.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002554 "src/f32-gemm/gen-inc/1x8inc-minmax-fma3-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002555 "src/f32-gemm/gen-inc/1x16inc-minmax-fma3-broadcast.c",
2556 "src/f32-gemm/gen-inc/1x16s4inc-minmax-fma3-broadcast.c",
2557 "src/f32-gemm/gen-inc/3x16inc-minmax-fma3-broadcast.c",
2558 "src/f32-gemm/gen-inc/3x16s4inc-minmax-fma3-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002559 "src/f32-gemm/gen-inc/4x8inc-minmax-fma3-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002560 "src/f32-gemm/gen-inc/4x16inc-minmax-fma3-broadcast.c",
2561 "src/f32-gemm/gen-inc/4x16s4inc-minmax-fma3-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002562 "src/f32-gemm/gen-inc/5x8inc-minmax-fma3-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002563 "src/f32-gemm/gen-inc/5x16inc-minmax-fma3-broadcast.c",
2564 "src/f32-gemm/gen-inc/5x16s4inc-minmax-fma3-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002565 "src/f32-gemm/gen-inc/6x8inc-minmax-fma3-broadcast.c",
2566 "src/f32-gemm/gen-inc/7x8inc-minmax-fma3-broadcast.c",
2567 "src/f32-gemm/gen-inc/8x8inc-minmax-fma3-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002568 "src/f32-gemm/gen/1x8-minmax-fma3-broadcast.c",
2569 "src/f32-gemm/gen/1x16-minmax-fma3-broadcast.c",
2570 "src/f32-gemm/gen/1x16s4-minmax-fma3-broadcast.c",
2571 "src/f32-gemm/gen/3x16-minmax-fma3-broadcast.c",
2572 "src/f32-gemm/gen/3x16s4-minmax-fma3-broadcast.c",
2573 "src/f32-gemm/gen/4x8-minmax-fma3-broadcast.c",
2574 "src/f32-gemm/gen/4x16-minmax-fma3-broadcast.c",
2575 "src/f32-gemm/gen/4x16s4-minmax-fma3-broadcast.c",
2576 "src/f32-gemm/gen/5x8-minmax-fma3-broadcast.c",
2577 "src/f32-gemm/gen/5x16-minmax-fma3-broadcast.c",
2578 "src/f32-gemm/gen/5x16s4-minmax-fma3-broadcast.c",
2579 "src/f32-gemm/gen/6x8-minmax-fma3-broadcast.c",
2580 "src/f32-gemm/gen/7x8-minmax-fma3-broadcast.c",
2581 "src/f32-gemm/gen/8x8-minmax-fma3-broadcast.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07002582 "src/f32-hswish/gen/hswish-fma3-x8.c",
2583 "src/f32-hswish/gen/hswish-fma3-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002584 "src/f32-igemm/gen/1x8-minmax-fma3-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002585 "src/f32-igemm/gen/1x16-minmax-fma3-broadcast.c",
2586 "src/f32-igemm/gen/1x16s4-minmax-fma3-broadcast.c",
2587 "src/f32-igemm/gen/3x16-minmax-fma3-broadcast.c",
2588 "src/f32-igemm/gen/3x16s4-minmax-fma3-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002589 "src/f32-igemm/gen/4x8-minmax-fma3-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002590 "src/f32-igemm/gen/4x16-minmax-fma3-broadcast.c",
2591 "src/f32-igemm/gen/4x16s4-minmax-fma3-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002592 "src/f32-igemm/gen/5x8-minmax-fma3-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002593 "src/f32-igemm/gen/5x16-minmax-fma3-broadcast.c",
2594 "src/f32-igemm/gen/5x16s4-minmax-fma3-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002595 "src/f32-igemm/gen/6x8-minmax-fma3-broadcast.c",
2596 "src/f32-igemm/gen/7x8-minmax-fma3-broadcast.c",
2597 "src/f32-igemm/gen/8x8-minmax-fma3-broadcast.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07002598 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x8.c",
2599 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x16.c",
2600 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x24.c",
2601 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x32.c",
2602 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x40.c",
2603 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x48.c",
2604 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x56.c",
2605 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x64.c",
Marat Dukhan84000762020-06-29 18:38:43 -07002606 "src/math/sqrt-fma3-nr1fma.c",
Marat Dukhan84000762020-06-29 18:38:43 -07002607 "src/math/sqrt-fma3-nr1fma1adj.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002608 "src/math/sqrt-fma3-nr2fma.c",
Marat Dukhanfda12b82019-11-21 12:27:59 -08002609]
2610
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002611AVX2_UKERNELS = [
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002612 "src/f32-raddexpminusmax/gen/avx2-p5-x64-acc2.c",
2613 "src/f32-raddexpminusmax/gen/avx2-p5-x64-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002614 "src/f32-raddexpminusmax/gen/avx2-p5-x64.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002615 "src/f32-raddexpminusmax/gen/avx2-p5-x72-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002616 "src/f32-raddexpminusmax/gen/avx2-p5-x72.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002617 "src/f32-raddexpminusmax/gen/avx2-p5-x80-acc2.c",
2618 "src/f32-raddexpminusmax/gen/avx2-p5-x80-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002619 "src/f32-raddexpminusmax/gen/avx2-p5-x80.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002620 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc2.c",
2621 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc3.c",
2622 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc6.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002623 "src/f32-raddexpminusmax/gen/avx2-p5-x96.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002624 "src/f32-raddextexp/gen/avx2-p5-x64-acc2.c",
2625 "src/f32-raddextexp/gen/avx2-p5-x64-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002626 "src/f32-raddextexp/gen/avx2-p5-x64.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002627 "src/f32-raddextexp/gen/avx2-p5-x72-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002628 "src/f32-raddextexp/gen/avx2-p5-x72.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002629 "src/f32-raddextexp/gen/avx2-p5-x80-acc2.c",
2630 "src/f32-raddextexp/gen/avx2-p5-x80-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002631 "src/f32-raddextexp/gen/avx2-p5-x80.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002632 "src/f32-raddextexp/gen/avx2-p5-x96-acc2.c",
2633 "src/f32-raddextexp/gen/avx2-p5-x96-acc3.c",
2634 "src/f32-raddextexp/gen/avx2-p5-x96-acc6.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002635 "src/f32-raddextexp/gen/avx2-p5-x96.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002636 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64-acc2.c",
2637 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002638 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002639 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x72-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002640 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x72.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002641 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80-acc2.c",
2642 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002643 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002644 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc2.c",
2645 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc3.c",
2646 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc6.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002647 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96.c",
Marat Dukhanfa0a4322020-01-06 16:14:29 -08002648 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x8.c",
2649 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x16.c",
2650 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x24.c",
2651 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x32.c",
2652 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x40.c",
2653 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x48.c",
2654 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x56.c",
2655 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x64.c",
2656 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x72.c",
2657 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x80.c",
2658 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x8.c",
2659 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x16.c",
2660 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x24.c",
2661 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x32.c",
2662 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x40.c",
2663 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x48.c",
2664 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x56.c",
2665 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x64.c",
2666 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x72.c",
2667 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x80.c",
2668 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x8.c",
2669 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x16.c",
2670 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x24.c",
2671 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x32.c",
2672 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x40.c",
2673 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x48.c",
2674 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x56.c",
2675 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x64.c",
2676 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x72.c",
2677 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x80.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002678 "src/f32-vscaleexpminusmax/gen/avx2-p5-x8.c",
2679 "src/f32-vscaleexpminusmax/gen/avx2-p5-x16.c",
2680 "src/f32-vscaleexpminusmax/gen/avx2-p5-x24.c",
2681 "src/f32-vscaleexpminusmax/gen/avx2-p5-x32.c",
2682 "src/f32-vscaleexpminusmax/gen/avx2-p5-x40.c",
2683 "src/f32-vscaleexpminusmax/gen/avx2-p5-x48.c",
2684 "src/f32-vscaleexpminusmax/gen/avx2-p5-x56.c",
2685 "src/f32-vscaleexpminusmax/gen/avx2-p5-x64.c",
2686 "src/f32-vscaleexpminusmax/gen/avx2-p5-x72.c",
2687 "src/f32-vscaleexpminusmax/gen/avx2-p5-x80.c",
2688 "src/f32-vscaleexpminusmax/gen/avx2-p5-x88.c",
2689 "src/f32-vscaleexpminusmax/gen/avx2-p5-x96.c",
2690 "src/f32-vscaleextexp/gen/avx2-p5-x8.c",
2691 "src/f32-vscaleextexp/gen/avx2-p5-x16.c",
2692 "src/f32-vscaleextexp/gen/avx2-p5-x24.c",
2693 "src/f32-vscaleextexp/gen/avx2-p5-x32.c",
2694 "src/f32-vscaleextexp/gen/avx2-p5-x40.c",
2695 "src/f32-vscaleextexp/gen/avx2-p5-x48.c",
2696 "src/f32-vscaleextexp/gen/avx2-p5-x56.c",
2697 "src/f32-vscaleextexp/gen/avx2-p5-x64.c",
2698 "src/f32-vscaleextexp/gen/avx2-p5-x72.c",
2699 "src/f32-vscaleextexp/gen/avx2-p5-x80.c",
2700 "src/f32-vscaleextexp/gen/avx2-p5-x88.c",
2701 "src/f32-vscaleextexp/gen/avx2-p5-x96.c",
Marat Dukhanb7633f22020-11-20 16:34:56 -08002702 "src/math/exp-avx2-rr2-lut8-p3-perm.c",
2703 "src/math/exp-avx2-rr2-lut8-p4-perm.c",
2704 "src/math/exp-avx2-rr2-p5.c",
Marat Dukhan9dd119a2020-11-20 18:20:04 -08002705 "src/math/expminus-avx2-rr2-p5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002706 "src/math/extexp-avx2-p5.c",
2707 "src/math/sigmoid-avx2-rr1-lut64-p2-gather-div.c",
2708 "src/math/sigmoid-avx2-rr1-lut64-p2-gather-nr1fma.c",
2709 "src/math/sigmoid-avx2-rr1-lut64-p2-gather-nr2fma.c",
2710 "src/math/sigmoid-avx2-rr1-lut64-p2-gather-nr2fma1adj.c",
2711 "src/math/sigmoid-avx2-rr1-p5-div.c",
2712 "src/math/sigmoid-avx2-rr1-p5-nr1fma.c",
2713 "src/math/sigmoid-avx2-rr1-p5-nr2fma.c",
2714 "src/math/sigmoid-avx2-rr2-lut64-p2-gather-div.c",
2715 "src/math/sigmoid-avx2-rr2-lut64-p2-gather-nr1fma.c",
2716 "src/math/sigmoid-avx2-rr2-lut64-p2-gather-nr2fma.c",
2717 "src/math/sigmoid-avx2-rr2-lut64-p2-gather-nr2fma1adj.c",
2718 "src/math/sigmoid-avx2-rr2-p5-div.c",
2719 "src/math/sigmoid-avx2-rr2-p5-nr1fma.c",
2720 "src/math/sigmoid-avx2-rr2-p5-nr2fma.c",
Marat Dukhand65a1522020-08-04 19:28:18 -07002721 "src/qs8-dwconv/gen/up8x9-minmax-avx2-mul32.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002722 "src/qs8-dwconv/gen/up16x9-minmax-avx2-mul16.c",
Marat Dukhand65a1522020-08-04 19:28:18 -07002723 "src/qs8-dwconv/gen/up16x9-minmax-avx2-mul32.c",
2724 "src/qs8-dwconv/gen/up24x9-minmax-avx2-mul32.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002725 "src/qs8-dwconv/gen/up32x9-minmax-avx2-mul16.c",
Marat Dukhand65a1522020-08-04 19:28:18 -07002726 "src/qs8-dwconv/gen/up32x9-minmax-avx2-mul32.c",
Marat Dukhan12809522020-08-02 22:23:51 -07002727 "src/qs8-gemm/gen/1x8c8-minmax-avx2.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002728 "src/qs8-gemm/gen/1x8c8-xw-minmax-avx2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002729 "src/qs8-gemm/gen/2x8c8-minmax-avx2.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002730 "src/qs8-gemm/gen/2x8c8-xw-minmax-avx2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002731 "src/qs8-gemm/gen/3x8c8-minmax-avx2.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002732 "src/qs8-gemm/gen/3x8c8-xw-minmax-avx2.c",
Marat Dukhan12809522020-08-02 22:23:51 -07002733 "src/qs8-igemm/gen/1x8c8-minmax-avx2.c",
2734 "src/qs8-igemm/gen/2x8c8-minmax-avx2.c",
2735 "src/qs8-igemm/gen/3x8c8-minmax-avx2.c",
Marat Dukhane6dc0b62020-09-08 23:57:14 -07002736 "src/qs8-vadd/gen/minmax-avx2-mul32-ld64-x8.c",
2737 "src/qs8-vadd/gen/minmax-avx2-mul32-ld64-x16.c",
2738 "src/qs8-vadd/gen/minmax-avx2-mul32-ld64-x24.c",
2739 "src/qs8-vadd/gen/minmax-avx2-mul32-ld64-x32.c",
2740 "src/qs8-vaddc/gen/minmax-avx2-mul32-ld64-x8.c",
2741 "src/qs8-vaddc/gen/minmax-avx2-mul32-ld64-x16.c",
2742 "src/qs8-vaddc/gen/minmax-avx2-mul32-ld64-x24.c",
2743 "src/qs8-vaddc/gen/minmax-avx2-mul32-ld64-x32.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002744]
2745
Marat Dukhan08c4a432019-10-03 09:29:21 -07002746AVX512F_UKERNELS = [
Marat Dukhan5c5fa962020-03-10 18:38:33 -07002747 "src/f32-clamp/gen/avx512f-x16.c",
2748 "src/f32-clamp/gen/avx512f-x32.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002749 "src/f32-dwconv/gen/up16x4-minmax-avx512f-acc2.c",
2750 "src/f32-dwconv/gen/up16x4-minmax-avx512f.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002751 "src/f32-dwconv/gen/up16x9-minmax-avx512f-acc2.c",
2752 "src/f32-dwconv/gen/up16x9-minmax-avx512f.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002753 "src/f32-dwconv/gen/up16x25-minmax-avx512f-acc2.c",
2754 "src/f32-dwconv/gen/up16x25-minmax-avx512f.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002755 "src/f32-dwconv/gen/up32x4-minmax-avx512f-acc2.c",
2756 "src/f32-dwconv/gen/up32x4-minmax-avx512f.c",
2757 "src/f32-dwconv/gen/up32x9-minmax-avx512f-acc2.c",
2758 "src/f32-dwconv/gen/up32x9-minmax-avx512f.c",
2759 "src/f32-dwconv/gen/up32x25-minmax-avx512f-acc2.c",
2760 "src/f32-dwconv/gen/up32x25-minmax-avx512f.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002761 "src/f32-gemm/gen-inc/1x16inc-minmax-avx512f-broadcast.c",
2762 "src/f32-gemm/gen-inc/4x16inc-minmax-avx512f-broadcast.c",
2763 "src/f32-gemm/gen-inc/5x16inc-minmax-avx512f-broadcast.c",
2764 "src/f32-gemm/gen-inc/6x16inc-minmax-avx512f-broadcast.c",
2765 "src/f32-gemm/gen-inc/7x16inc-minmax-avx512f-broadcast.c",
2766 "src/f32-gemm/gen-inc/8x16inc-minmax-avx512f-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002767 "src/f32-gemm/gen/1x16-minmax-avx512f-broadcast.c",
2768 "src/f32-gemm/gen/4x16-minmax-avx512f-broadcast.c",
2769 "src/f32-gemm/gen/5x16-minmax-avx512f-broadcast.c",
2770 "src/f32-gemm/gen/6x16-minmax-avx512f-broadcast.c",
2771 "src/f32-gemm/gen/7x16-minmax-avx512f-broadcast.c",
2772 "src/f32-gemm/gen/8x16-minmax-avx512f-broadcast.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07002773 "src/f32-hswish/gen/hswish-avx512f-x16.c",
2774 "src/f32-hswish/gen/hswish-avx512f-x32.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002775 "src/f32-igemm/gen/1x16-minmax-avx512f-broadcast.c",
2776 "src/f32-igemm/gen/4x16-minmax-avx512f-broadcast.c",
2777 "src/f32-igemm/gen/5x16-minmax-avx512f-broadcast.c",
2778 "src/f32-igemm/gen/6x16-minmax-avx512f-broadcast.c",
2779 "src/f32-igemm/gen/7x16-minmax-avx512f-broadcast.c",
2780 "src/f32-igemm/gen/8x16-minmax-avx512f-broadcast.c",
Marat Dukhan90eca0a2020-03-11 00:52:23 -07002781 "src/f32-prelu/gen/avx512f-2x16.c",
2782 "src/f32-prelu/gen/avx512f-2x32.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002783 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128-acc2.c",
2784 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002785 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002786 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x144-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002787 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x144.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002788 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160-acc2.c",
2789 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002790 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002791 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc2.c",
2792 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc3.c",
2793 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc6.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002794 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002795 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128-acc2.c",
2796 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002797 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002798 "src/f32-raddextexp/gen/avx512f-p5-scalef-x144-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002799 "src/f32-raddextexp/gen/avx512f-p5-scalef-x144.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002800 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160-acc2.c",
2801 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002802 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002803 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc2.c",
2804 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc3.c",
2805 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc6.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002806 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002807 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128-acc2.c",
2808 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002809 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002810 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x144-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002811 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x144.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002812 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160-acc2.c",
2813 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002814 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002815 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc2.c",
2816 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc3.c",
2817 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc6.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002818 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07002819 "src/f32-relu/gen/avx512f-x16.c",
2820 "src/f32-relu/gen/avx512f-x32.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002821 "src/f32-rmax/avx512f.c",
Marat Dukhand9ca7e62020-09-23 23:45:29 -07002822 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x16.c",
2823 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x32.c",
2824 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x48.c",
2825 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x64.c",
2826 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x80.c",
2827 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x96.c",
2828 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x112.c",
2829 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x128.c",
2830 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x16.c",
2831 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x32.c",
2832 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x48.c",
2833 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x64.c",
2834 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x80.c",
2835 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x96.c",
2836 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x112.c",
2837 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002838 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x16.c",
2839 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x32.c",
2840 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x48.c",
2841 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x64.c",
2842 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x80.c",
2843 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x96.c",
2844 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x112.c",
2845 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x128.c",
2846 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x16.c",
2847 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x32.c",
2848 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x48.c",
2849 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x64.c",
2850 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x80.c",
2851 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x96.c",
2852 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x112.c",
2853 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x128.c",
Marat Dukhand9ca7e62020-09-23 23:45:29 -07002854 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x16.c",
2855 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x32.c",
2856 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x48.c",
2857 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x64.c",
2858 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x80.c",
2859 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x96.c",
2860 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x112.c",
2861 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x128.c",
2862 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x16.c",
2863 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x32.c",
2864 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x48.c",
2865 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x64.c",
2866 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x80.c",
2867 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x96.c",
2868 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x112.c",
2869 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x128.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002870 "src/f32-vbinary/gen/vadd-minmax-avx512f-x16.c",
2871 "src/f32-vbinary/gen/vadd-minmax-avx512f-x32.c",
2872 "src/f32-vbinary/gen/vaddc-minmax-avx512f-x16.c",
2873 "src/f32-vbinary/gen/vaddc-minmax-avx512f-x32.c",
2874 "src/f32-vbinary/gen/vdiv-minmax-avx512f-x16.c",
2875 "src/f32-vbinary/gen/vdiv-minmax-avx512f-x32.c",
2876 "src/f32-vbinary/gen/vdivc-minmax-avx512f-x16.c",
2877 "src/f32-vbinary/gen/vdivc-minmax-avx512f-x32.c",
Marat Dukhan9a88efe2019-12-10 15:54:24 -08002878 "src/f32-vbinary/gen/vmax-avx512f-x16.c",
2879 "src/f32-vbinary/gen/vmax-avx512f-x32.c",
2880 "src/f32-vbinary/gen/vmaxc-avx512f-x16.c",
2881 "src/f32-vbinary/gen/vmaxc-avx512f-x32.c",
2882 "src/f32-vbinary/gen/vmin-avx512f-x16.c",
2883 "src/f32-vbinary/gen/vmin-avx512f-x32.c",
2884 "src/f32-vbinary/gen/vminc-avx512f-x16.c",
2885 "src/f32-vbinary/gen/vminc-avx512f-x32.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002886 "src/f32-vbinary/gen/vmul-minmax-avx512f-x16.c",
2887 "src/f32-vbinary/gen/vmul-minmax-avx512f-x32.c",
2888 "src/f32-vbinary/gen/vmulc-minmax-avx512f-x16.c",
2889 "src/f32-vbinary/gen/vmulc-minmax-avx512f-x32.c",
2890 "src/f32-vbinary/gen/vrdivc-minmax-avx512f-x16.c",
2891 "src/f32-vbinary/gen/vrdivc-minmax-avx512f-x32.c",
2892 "src/f32-vbinary/gen/vrsubc-minmax-avx512f-x16.c",
2893 "src/f32-vbinary/gen/vrsubc-minmax-avx512f-x32.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07002894 "src/f32-vbinary/gen/vsqrdiff-avx512f-x16.c",
2895 "src/f32-vbinary/gen/vsqrdiff-avx512f-x32.c",
2896 "src/f32-vbinary/gen/vsqrdiffc-avx512f-x16.c",
2897 "src/f32-vbinary/gen/vsqrdiffc-avx512f-x32.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002898 "src/f32-vbinary/gen/vsub-minmax-avx512f-x16.c",
2899 "src/f32-vbinary/gen/vsub-minmax-avx512f-x32.c",
2900 "src/f32-vbinary/gen/vsubc-minmax-avx512f-x16.c",
2901 "src/f32-vbinary/gen/vsubc-minmax-avx512f-x32.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07002902 "src/f32-vlrelu/gen/vlrelu-avx512f-x16.c",
2903 "src/f32-vlrelu/gen/vlrelu-avx512f-x32.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002904 "src/f32-vrnd/gen/vrndd-avx512f-x16.c",
2905 "src/f32-vrnd/gen/vrndd-avx512f-x32.c",
2906 "src/f32-vrnd/gen/vrndne-avx512f-x16.c",
2907 "src/f32-vrnd/gen/vrndne-avx512f-x32.c",
2908 "src/f32-vrnd/gen/vrndu-avx512f-x16.c",
2909 "src/f32-vrnd/gen/vrndu-avx512f-x32.c",
2910 "src/f32-vrnd/gen/vrndz-avx512f-x16.c",
2911 "src/f32-vrnd/gen/vrndz-avx512f-x32.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07002912 "src/f32-vscale/avx512f-x64.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002913 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x16.c",
2914 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x32.c",
2915 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x48.c",
2916 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x64.c",
2917 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x80.c",
2918 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x96.c",
2919 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x112.c",
2920 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x128.c",
2921 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x144.c",
2922 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x160.c",
2923 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x176.c",
2924 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x192.c",
2925 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x16.c",
2926 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x32.c",
2927 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x48.c",
2928 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x64.c",
2929 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x80.c",
2930 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x96.c",
2931 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x112.c",
2932 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x128.c",
2933 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x144.c",
2934 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x160.c",
2935 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x176.c",
2936 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x192.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07002937 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x16.c",
2938 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x32.c",
2939 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x48.c",
2940 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x64.c",
2941 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x80.c",
2942 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x96.c",
2943 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x112.c",
2944 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x128.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07002945 "src/f32-vunary/gen/vabs-avx512f-x16.c",
2946 "src/f32-vunary/gen/vabs-avx512f-x32.c",
2947 "src/f32-vunary/gen/vneg-avx512f-x16.c",
2948 "src/f32-vunary/gen/vneg-avx512f-x32.c",
2949 "src/f32-vunary/gen/vsqr-avx512f-x16.c",
2950 "src/f32-vunary/gen/vsqr-avx512f-x32.c",
Marat Dukhanb7633f22020-11-20 16:34:56 -08002951 "src/math/exp-avx512f-rr2-lut16-p3-perm-scalef.c",
2952 "src/math/exp-avx512f-rr2-lut16-p3-perm.c",
2953 "src/math/exp-avx512f-rr2-lut32-p2-perm2-scalef.c",
2954 "src/math/exp-avx512f-rr2-lut32-p2-perm2.c",
2955 "src/math/exp-avx512f-rr2-p5-scalef.c",
2956 "src/math/exp-avx512f-rr2-p5.c",
Marat Dukhan98ba4412019-10-23 02:14:28 -07002957 "src/math/extexp-avx512f-p5.c",
Marat Dukhan6a34c5f2020-09-22 21:44:15 -07002958 "src/math/sigmoid-avx512f-rr1-lut16-p3-perm-scalef-div.c",
Marat Dukhan6a34c5f2020-09-22 21:44:15 -07002959 "src/math/sigmoid-avx512f-rr1-lut16-p3-perm-scalef-nr1fma.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002960 "src/math/sigmoid-avx512f-rr1-lut16-p3-perm-scalef-nr1fma1adj.c",
Marat Dukhan6a34c5f2020-09-22 21:44:15 -07002961 "src/math/sigmoid-avx512f-rr1-lut32-p2-perm2-scalef-div.c",
Marat Dukhan6a34c5f2020-09-22 21:44:15 -07002962 "src/math/sigmoid-avx512f-rr1-lut32-p2-perm2-scalef-nr1fma.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002963 "src/math/sigmoid-avx512f-rr1-lut32-p2-perm2-scalef-nr1fma1adj.c",
Marat Dukhan36173d22020-10-15 17:14:26 -07002964 "src/math/sigmoid-avx512f-rr1-lut64-p2-gather-scalef-div.c",
Marat Dukhan36173d22020-10-15 17:14:26 -07002965 "src/math/sigmoid-avx512f-rr1-lut64-p2-gather-scalef-nr1fma.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002966 "src/math/sigmoid-avx512f-rr1-lut64-p2-gather-scalef-nr1fma1adj.c",
2967 "src/math/sigmoid-avx512f-rr1-p5-scalef-div.c",
2968 "src/math/sigmoid-avx512f-rr1-p5-scalef-nr1fma.c",
2969 "src/math/sigmoid-avx512f-rr1-p5-scalef-nr1fma1adj.c",
2970 "src/math/sigmoid-avx512f-rr2-lut16-p3-perm-scalef-div.c",
2971 "src/math/sigmoid-avx512f-rr2-lut16-p3-perm-scalef-nr1fma.c",
2972 "src/math/sigmoid-avx512f-rr2-lut16-p3-perm-scalef-nr1fma1adj.c",
2973 "src/math/sigmoid-avx512f-rr2-lut32-p2-perm2-scalef-div.c",
2974 "src/math/sigmoid-avx512f-rr2-lut32-p2-perm2-scalef-nr1fma.c",
2975 "src/math/sigmoid-avx512f-rr2-lut32-p2-perm2-scalef-nr1fma1adj.c",
Marat Dukhan36173d22020-10-15 17:14:26 -07002976 "src/math/sigmoid-avx512f-rr2-lut64-p2-gather-scalef-div.c",
Marat Dukhan36173d22020-10-15 17:14:26 -07002977 "src/math/sigmoid-avx512f-rr2-lut64-p2-gather-scalef-nr1fma.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002978 "src/math/sigmoid-avx512f-rr2-lut64-p2-gather-scalef-nr1fma1adj.c",
2979 "src/math/sigmoid-avx512f-rr2-p5-scalef-div.c",
2980 "src/math/sigmoid-avx512f-rr2-p5-scalef-nr1fma.c",
2981 "src/math/sigmoid-avx512f-rr2-p5-scalef-nr1fma1adj.c",
Marat Dukhan84000762020-06-29 18:38:43 -07002982 "src/math/sqrt-avx512f-nr1fma.c",
Marat Dukhan84000762020-06-29 18:38:43 -07002983 "src/math/sqrt-avx512f-nr1fma1adj.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002984 "src/math/sqrt-avx512f-nr2fma.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002985]
2986
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07002987AVX512SKX_UKERNELS = [
Marat Dukhan2ffc5e62020-09-06 22:33:38 -07002988 "src/qs8-dwconv/gen/up16x9-minmax-avx512skx-mul32.c",
2989 "src/qs8-dwconv/gen/up32x9-minmax-avx512skx-mul32.c",
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07002990 "src/qs8-gemm/gen/1x16c8-minmax-avx512skx.c",
2991 "src/qs8-gemm/gen/2x16c8-minmax-avx512skx.c",
2992 "src/qs8-gemm/gen/3x16c8-minmax-avx512skx.c",
2993 "src/qs8-gemm/gen/4x16c8-minmax-avx512skx.c",
2994 "src/qs8-igemm/gen/1x16c8-minmax-avx512skx.c",
2995 "src/qs8-igemm/gen/2x16c8-minmax-avx512skx.c",
2996 "src/qs8-igemm/gen/3x16c8-minmax-avx512skx.c",
2997 "src/qs8-igemm/gen/4x16c8-minmax-avx512skx.c",
2998]
2999
Frank Barchardbcedc082020-08-17 18:00:51 -07003000WASM32_ASM_UKERNELS = [
3001 "src/f32-relu/wasm_shr_x1.S",
3002 "src/f32-relu/wasm_shr_x2.S",
3003 "src/f32-relu/wasm_shr_x4.S",
3004]
3005
Marat Dukhan08c4a432019-10-03 09:29:21 -07003006AARCH32_ASM_UKERNELS = [
Marat Dukhan32f93812020-05-17 20:31:21 -07003007 "src/f32-gemm/4x4-aarch32-vfp-ld64.S",
Marat Dukhan3b98f6b2020-05-17 10:09:22 -07003008 "src/f32-gemm/4x4-minmax-aarch32-vfp-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003009 "src/f32-gemm/4x8-minmax-aarch32-neon-cortex-a53.S",
3010 "src/f32-gemm/4x8-minmax-aarch32-neon-cortex-a55.S",
Frank Barchard490febe2020-07-16 18:42:17 -07003011 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-cortex-a7.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003012 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-cortex-a75.S",
Frank Barchard569561d2020-06-17 13:11:12 -07003013 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-ld64.S",
Frank Barchard490febe2020-07-16 18:42:17 -07003014 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-pld-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003015 "src/f32-igemm/4x8-minmax-aarch32-neon-cortex-a53.S",
3016 "src/f32-igemm/4x8-minmax-aarch32-neon-cortex-a55.S",
Frank Barchard490febe2020-07-16 18:42:17 -07003017 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-cortex-a7.S",
3018 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-cortex-a75.S",
3019 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-ld64.S",
3020 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-pld-cortex-a75.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003021]
3022
3023AARCH64_ASM_UKERNELS = [
Frank Barchardbddfbcd2020-04-15 12:32:41 -07003024 "src/f16-gemm/gen-inc/1x8inc-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchard04336c12020-10-22 16:48:55 -07003025 "src/f16-gemm/gen-inc/1x16inc-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchardbddfbcd2020-04-15 12:32:41 -07003026 "src/f16-gemm/gen-inc/4x8inc-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchard04336c12020-10-22 16:48:55 -07003027 "src/f16-gemm/gen-inc/4x16inc-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchardbddfbcd2020-04-15 12:32:41 -07003028 "src/f16-gemm/gen-inc/6x8inc-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchard04336c12020-10-22 16:48:55 -07003029 "src/f16-gemm/gen-inc/6x16inc-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchard3b8e5662020-04-20 12:12:53 -07003030 "src/f16-gemm/gen-inc/8x8inc-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchard04336c12020-10-22 16:48:55 -07003031 "src/f16-gemm/gen/1x8-minmax-aarch64-neonfp16arith-ld64.S",
3032 "src/f16-gemm/gen/1x16-minmax-aarch64-neonfp16arith-ld32.S",
3033 "src/f16-gemm/gen/4x8-minmax-aarch64-neonfp16arith-ld64.S",
3034 "src/f16-gemm/gen/4x16-minmax-aarch64-neonfp16arith-ld32.S",
3035 "src/f16-gemm/gen/6x8-minmax-aarch64-neonfp16arith-ld64.S",
3036 "src/f16-gemm/gen/6x16-minmax-aarch64-neonfp16arith-ld32.S",
3037 "src/f16-gemm/gen/8x8-minmax-aarch64-neonfp16arith-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003038 "src/f32-dwconv/up4x9-minmax-aarch64-neonfma-cortex-a55.S",
3039 "src/f32-dwconv/up4x9-minmax-aarch64-neonfma.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003040 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a53.S",
3041 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a57.S",
3042 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a75.S",
Frank Barchard04336c12020-10-22 16:48:55 -07003043 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-ld64.S",
3044 "src/f32-gemm/gen-inc/1x12inc-minmax-aarch64-neonfma-cortex-a53.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003045 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a53.S",
3046 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a55.S",
3047 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a57.S",
3048 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003049 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-ld64.S",
Frank Barchard04336c12020-10-22 16:48:55 -07003050 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-ld128.S",
3051 "src/f32-gemm/gen-inc/4x12inc-minmax-aarch64-neonfma-cortex-a53.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003052 "src/f32-gemm/gen-inc/5x8inc-minmax-aarch64-neonfma-cortex-a57.S",
3053 "src/f32-gemm/gen-inc/5x8inc-minmax-aarch64-neonfma-cortex-a75.S",
3054 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a53.S",
3055 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a55.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003056 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a57.S",
Frank Barchard04336c12020-10-22 16:48:55 -07003057 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a73.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003058 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003059 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-ld64.S",
Frank Barchard04336c12020-10-22 16:48:55 -07003060 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-ld128.S",
3061 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a53.S",
3062 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a57.S",
3063 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a75.S",
3064 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-ld64.S",
3065 "src/f32-gemm/gen/1x12-minmax-aarch64-neonfma-cortex-a53.S",
3066 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a53.S",
3067 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a55.S",
3068 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a57.S",
3069 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a75.S",
3070 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-ld64.S",
3071 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-ld128.S",
3072 "src/f32-gemm/gen/4x12-minmax-aarch64-neonfma-cortex-a53.S",
3073 "src/f32-gemm/gen/5x8-minmax-aarch64-neonfma-cortex-a57.S",
3074 "src/f32-gemm/gen/5x8-minmax-aarch64-neonfma-cortex-a75.S",
3075 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a53.S",
3076 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a55.S",
3077 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a57.S",
3078 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a73.S",
3079 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a75.S",
3080 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-ld64.S",
3081 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-ld128.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003082 "src/f32-igemm/1x8-minmax-aarch64-neonfma-cortex-a53.S",
Frank Barchard04336c12020-10-22 16:48:55 -07003083 "src/f32-igemm/1x12-minmax-aarch64-neonfma-cortex-a53.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003084 "src/f32-igemm/4x8-minmax-aarch64-neonfma-cortex-a53.S",
3085 "src/f32-igemm/4x8-minmax-aarch64-neonfma-cortex-a55.S",
Frank Barchard04336c12020-10-22 16:48:55 -07003086 "src/f32-igemm/4x12-minmax-aarch64-neonfma-cortex-a53.S",
3087 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a53.S",
3088 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a55.S",
3089 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a73.S",
3090 "src/f32-igemm/gen/1x8-minmax-aarch64-neonfma-cortex-a57.S",
3091 "src/f32-igemm/gen/1x8-minmax-aarch64-neonfma-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003092 "src/f32-igemm/gen/4x8-minmax-aarch64-neonfma-cortex-a57.S",
3093 "src/f32-igemm/gen/4x8-minmax-aarch64-neonfma-cortex-a75.S",
3094 "src/f32-igemm/gen/5x8-minmax-aarch64-neonfma-cortex-a57.S",
3095 "src/f32-igemm/gen/5x8-minmax-aarch64-neonfma-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003096 "src/f32-igemm/gen/6x8-minmax-aarch64-neonfma-cortex-a57.S",
3097 "src/f32-igemm/gen/6x8-minmax-aarch64-neonfma-cortex-a75.S",
Frank Barchardf1fd89e2020-10-06 23:27:11 -07003098 "src/qs8-gemm/1x16c4-aarch64-neondot-ld64.S",
Frank Barchardf2742c42020-10-12 21:16:46 -07003099 "src/qs8-gemm/4x16c4-aarch64-neondot-cortex-a55.S",
Frank Barcharda5237a52020-10-06 16:18:46 -07003100 "src/qs8-gemm/4x16c4-aarch64-neondot-ld64.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003101]
3102
Marat Dukhan1b354632020-03-23 12:50:22 -07003103INTERNAL_MICROKERNEL_HDRS = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003104 "src/xnnpack/argmaxpool.h",
3105 "src/xnnpack/avgpool.h",
3106 "src/xnnpack/clamp.h",
3107 "src/xnnpack/common.h",
3108 "src/xnnpack/conv.h",
Yury Kartynnike7841862020-11-04 18:22:18 -08003109 "src/xnnpack/depthtospace.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003110 "src/xnnpack/dwconv.h",
Frank Barchard04336c12020-10-22 16:48:55 -07003111 "src/xnnpack/fill.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003112 "src/xnnpack/gavgpool.h",
3113 "src/xnnpack/gemm.h",
3114 "src/xnnpack/hswish.h",
Marat Dukhan660fd192020-03-10 04:55:30 -07003115 "src/xnnpack/ibilinear.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003116 "src/xnnpack/igemm.h",
Marat Dukhancfb31342019-12-05 10:42:57 -08003117 "src/xnnpack/intrinsics-polyfill.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003118 "src/xnnpack/lut.h",
3119 "src/xnnpack/math.h",
3120 "src/xnnpack/maxpool.h",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003121 "src/xnnpack/memory.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003122 "src/xnnpack/packx.h",
3123 "src/xnnpack/pad.h",
3124 "src/xnnpack/params.h",
3125 "src/xnnpack/pavgpool.h",
3126 "src/xnnpack/ppmm.h",
3127 "src/xnnpack/prelu.h",
Marat Dukhan97579532019-10-18 16:40:39 -07003128 "src/xnnpack/raddexpminusmax.h",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07003129 "src/xnnpack/raddextexp.h",
Marat Dukhan97579532019-10-18 16:40:39 -07003130 "src/xnnpack/raddstoreexpminusmax.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003131 "src/xnnpack/rmax.h",
3132 "src/xnnpack/scalar-utils.h",
3133 "src/xnnpack/spmm.h",
3134 "src/xnnpack/unpool.h",
3135 "src/xnnpack/vadd.h",
Marat Dukhan1e782c42019-11-21 17:02:40 -08003136 "src/xnnpack/vbinary.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003137 "src/xnnpack/vmulcaddc.h",
Marat Dukhan05ac8e32019-10-21 15:39:33 -07003138 "src/xnnpack/vscale.h",
Marat Dukhan97579532019-10-18 16:40:39 -07003139 "src/xnnpack/vscaleexpminusmax.h",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07003140 "src/xnnpack/vscaleextexp.h",
Marat Dukhan1e782c42019-11-21 17:02:40 -08003141 "src/xnnpack/vunary.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003142 "src/xnnpack/zip.h",
Marat Dukhan1b354632020-03-23 12:50:22 -07003143]
3144
3145INTERNAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003146 "include/xnnpack.h",
3147 "src/xnnpack/allocator.h",
3148 "src/xnnpack/compute.h",
3149 "src/xnnpack/im2col.h",
3150 "src/xnnpack/indirection.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003151 "src/xnnpack/math-stubs.h",
Chao Mei6ddfc602020-05-13 22:29:36 -07003152 "src/xnnpack/memory-planner.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003153 "src/xnnpack/operator.h",
3154 "src/xnnpack/pack.h",
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07003155 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003156 "src/xnnpack/requantization-stubs.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003157 "src/xnnpack/requantization.h",
Marat Dukhan1d75a542020-02-03 12:23:01 -08003158 "src/xnnpack/subgraph.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003159]
3160
Marat Dukhan1b354632020-03-23 12:50:22 -07003161ACCURACY_EVAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003162 "src/xnnpack/math-stubs.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003163]
3164
Marat Dukhan1b354632020-03-23 12:50:22 -07003165MICROKERNEL_BENCHMARK_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003166 "include/xnnpack.h",
Frank Barchard35db7d02020-10-26 13:37:34 -07003167 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003168]
3169
Marat Dukhan1b354632020-03-23 12:50:22 -07003170MICROKERNEL_TEST_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Frank Barchard35db7d02020-10-26 13:37:34 -07003171 "include/xnnpack.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003172 "src/xnnpack/isa-checks.h",
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07003173 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003174 "src/xnnpack/requantization.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003175]
3176
3177OPERATOR_TEST_PARAMS_HDRS = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003178 "src/xnnpack/common.h",
Frank Barchard04336c12020-10-22 16:48:55 -07003179 "src/xnnpack/params.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003180]
3181
3182WEIGHTS_PACK_HDRS = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003183 "src/xnnpack/compute.h",
Frank Barchard04336c12020-10-22 16:48:55 -07003184 "src/xnnpack/operator.h",
3185 "src/xnnpack/pack.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003186]
3187
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07003188LOGGING_COPTS = select({
3189 # No logging in optimized mode
3190 ":optimized_build": ["-DXNN_LOG_LEVEL=0"],
3191 # Full logging in debug mode
3192 ":debug_build": ["-DXNN_LOG_LEVEL=5"],
3193 # Error-only logging in default (fastbuild) mode
3194 "//conditions:default": ["-DXNN_LOG_LEVEL=2"],
3195})
3196
Marat Dukhan3b59de22020-06-03 20:15:19 -07003197LOGGING_SRCS = select({
3198 # No logging in optimized mode
3199 ":optimized_build": [],
3200 "//conditions:default": [
3201 "src/operator-strings.c",
3202 "src/subgraph-strings.c",
3203 ],
3204})
3205
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07003206LOGGING_HDRS = [
3207 "src/xnnpack/log.h",
3208]
3209
Marat Dukhan08c4a432019-10-03 09:29:21 -07003210xnnpack_cc_library(
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003211 name = "tables",
3212 srcs = TABLE_SRCS,
3213 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003214 gcc_copts = xnnpack_gcc_std_copts(),
3215 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003216)
3217
3218xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003219 name = "scalar_ukernels",
3220 srcs = SCALAR_UKERNELS,
3221 hdrs = INTERNAL_HDRS,
3222 aarch32_copts = ["-marm"],
Marat Dukhan10a38082020-04-17 03:58:35 -07003223 gcc_copts = xnnpack_gcc_std_copts(),
3224 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003225 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003226 ":tables",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003227 "@FP16",
3228 "@FXdiv",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003229 "@pthreadpool",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003230 ],
3231)
3232
3233xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003234 name = "scalar_ukernels_test_mode",
3235 srcs = SCALAR_UKERNELS,
3236 hdrs = INTERNAL_HDRS,
3237 aarch32_copts = ["-marm"],
3238 copts = [
3239 "-UNDEBUG",
3240 "-DXNN_TEST_MODE=1",
3241 ],
3242 gcc_copts = xnnpack_gcc_std_copts(),
3243 msvc_copts = xnnpack_msvc_std_copts(),
3244 deps = [
3245 ":tables",
3246 "@FP16",
3247 "@FXdiv",
3248 "@pthreadpool",
3249 ],
3250)
3251
3252xnnpack_cc_library(
Marat Dukhan436ebe62019-12-04 15:10:12 -08003253 name = "wasm_ukernels",
3254 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003255 gcc_copts = xnnpack_gcc_std_copts(),
3256 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan436ebe62019-12-04 15:10:12 -08003257 wasm_srcs = WASM_UKERNELS,
Marat Dukhan1c6e3892020-06-25 23:56:10 -07003258 wasmsimd_srcs = WASM_UKERNELS + WASMSIMD_UKERNELS,
Marat Dukhan436ebe62019-12-04 15:10:12 -08003259 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003260 ":tables",
Marat Dukhan436ebe62019-12-04 15:10:12 -08003261 "@FP16",
3262 "@FXdiv",
3263 "@pthreadpool",
3264 ],
3265)
3266
3267xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003268 name = "wasm_ukernels_test_mode",
3269 hdrs = INTERNAL_HDRS,
3270 copts = [
3271 "-UNDEBUG",
3272 "-DXNN_TEST_MODE=1",
3273 ],
3274 gcc_copts = xnnpack_gcc_std_copts(),
3275 msvc_copts = xnnpack_msvc_std_copts(),
3276 wasm_srcs = WASM_UKERNELS,
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07003277 wasmsimd_srcs = WASM_UKERNELS + WASMSIMD_UKERNELS,
Marat Dukhan33fcf782020-05-24 14:27:15 -07003278 deps = [
3279 ":tables",
3280 "@FP16",
3281 "@FXdiv",
3282 "@pthreadpool",
3283 ],
3284)
3285
3286xnnpack_cc_library(
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08003287 name = "psimd_accmath_ukernels",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08003288 hdrs = INTERNAL_HDRS,
3289 aarch32_copts = [
3290 "-marm",
3291 "-mfpu=neon",
3292 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07003293 gcc_copts = xnnpack_gcc_std_copts(),
3294 gcc_x86_copts = ["-msse2"],
3295 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08003296 optimized_copts = [
3297 "-O3",
3298 ],
Marat Dukhan500b8892020-04-15 17:09:50 -07003299 psimd_srcs = PSIMD_ACCMATH_UKERNELS,
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08003300 deps = [
3301 ":tables",
3302 "@FP16",
3303 "@psimd",
3304 "@pthreadpool",
3305 ],
3306)
3307
3308xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003309 name = "psimd_accmath_ukernels_test_mode",
3310 hdrs = INTERNAL_HDRS,
3311 aarch32_copts = [
3312 "-marm",
3313 "-mfpu=neon",
3314 ],
3315 copts = [
3316 "-UNDEBUG",
3317 "-DXNN_TEST_MODE=1",
3318 ],
3319 gcc_copts = xnnpack_gcc_std_copts(),
3320 gcc_x86_copts = ["-msse2"],
3321 msvc_copts = xnnpack_msvc_std_copts(),
3322 optimized_copts = [
3323 "-O3",
3324 ],
3325 psimd_srcs = PSIMD_ACCMATH_UKERNELS,
3326 deps = [
3327 ":tables",
3328 "@FP16",
3329 "@psimd",
3330 "@pthreadpool",
3331 ],
3332)
3333
3334xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003335 name = "neon_ukernels",
3336 hdrs = INTERNAL_HDRS,
3337 aarch32_copts = [
3338 "-marm",
Marat Dukhan8853b822020-05-07 12:19:01 -07003339 "-march=armv7-a",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003340 "-mfpu=neon",
3341 ],
3342 aarch32_srcs = NEON_UKERNELS,
3343 aarch64_srcs = NEON_UKERNELS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003344 gcc_copts = xnnpack_gcc_std_copts(),
3345 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08003346 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003347 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003348 "@FP16",
3349 "@pthreadpool",
3350 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003351)
3352
3353xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003354 name = "neon_ukernels_test_mode",
3355 hdrs = INTERNAL_HDRS,
3356 aarch32_copts = [
3357 "-marm",
3358 "-march=armv7-a",
3359 "-mfpu=neon",
3360 ],
3361 aarch32_srcs = NEON_UKERNELS,
3362 aarch64_srcs = NEON_UKERNELS,
3363 copts = [
3364 "-UNDEBUG",
3365 "-DXNN_TEST_MODE=1",
3366 ],
3367 gcc_copts = xnnpack_gcc_std_copts(),
3368 msvc_copts = xnnpack_msvc_std_copts(),
3369 deps = [
3370 ":tables",
3371 "@FP16",
3372 "@pthreadpool",
3373 ],
3374)
3375
3376xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003377 name = "neonfma_ukernels",
3378 hdrs = INTERNAL_HDRS,
3379 aarch32_copts = [
3380 "-marm",
Marat Dukhan8853b822020-05-07 12:19:01 -07003381 "-march=armv7-a",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003382 "-mfpu=neon-vfpv4",
3383 ],
3384 aarch32_srcs = NEONFMA_UKERNELS,
3385 aarch64_srcs = NEONFMA_UKERNELS + AARCH64_NEONFMA_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07003386 apple_aarch32_copts = [
3387 "-mcpu=swift",
3388 "-mtune=generic",
3389 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07003390 gcc_copts = xnnpack_gcc_std_copts(),
3391 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08003392 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003393 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003394 "@FP16",
3395 "@pthreadpool",
3396 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003397)
3398
3399xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003400 name = "neonfma_ukernels_test_mode",
3401 hdrs = INTERNAL_HDRS,
3402 aarch32_copts = [
3403 "-marm",
3404 "-march=armv7-a",
3405 "-mfpu=neon-vfpv4",
3406 ],
3407 aarch32_srcs = NEONFMA_UKERNELS,
3408 aarch64_srcs = NEONFMA_UKERNELS + AARCH64_NEONFMA_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07003409 apple_aarch32_copts = [
3410 "-mcpu=swift",
3411 "-mtune=generic",
3412 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07003413 copts = [
3414 "-UNDEBUG",
3415 "-DXNN_TEST_MODE=1",
3416 ],
3417 gcc_copts = xnnpack_gcc_std_copts(),
3418 msvc_copts = xnnpack_msvc_std_copts(),
3419 deps = [
3420 ":tables",
3421 "@FP16",
3422 "@pthreadpool",
3423 ],
3424)
3425
3426xnnpack_cc_library(
Marat Dukhan8853b822020-05-07 12:19:01 -07003427 name = "neonv8_ukernels",
3428 hdrs = INTERNAL_HDRS,
3429 aarch32_copts = [
3430 "-marm",
3431 "-march=armv8-a",
3432 "-mfpu=neon-fp-armv8",
3433 ],
3434 aarch32_srcs = NEONV8_UKERNELS,
3435 aarch64_srcs = NEONV8_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07003436 apple_aarch32_copts = [
3437 "-mcpu=cyclone",
3438 "-mtune=generic",
3439 ],
Marat Dukhan8853b822020-05-07 12:19:01 -07003440 gcc_copts = xnnpack_gcc_std_copts(),
3441 msvc_copts = xnnpack_msvc_std_copts(),
3442 deps = [
3443 ":tables",
3444 "@FP16",
3445 "@pthreadpool",
3446 ],
3447)
3448
3449xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003450 name = "neonv8_ukernels_test_mode",
3451 hdrs = INTERNAL_HDRS,
3452 aarch32_copts = [
3453 "-marm",
3454 "-march=armv8-a",
3455 "-mfpu=neon-fp-armv8",
3456 ],
3457 aarch32_srcs = NEONV8_UKERNELS,
3458 aarch64_srcs = NEONV8_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07003459 apple_aarch32_copts = [
3460 "-mcpu=cyclone",
3461 "-mtune=generic",
3462 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07003463 copts = [
3464 "-UNDEBUG",
3465 "-DXNN_TEST_MODE=1",
3466 ],
3467 gcc_copts = xnnpack_gcc_std_copts(),
3468 msvc_copts = xnnpack_msvc_std_copts(),
3469 deps = [
3470 ":tables",
3471 "@FP16",
3472 "@pthreadpool",
3473 ],
3474)
3475
3476xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003477 name = "neonfp16arith_ukernels",
3478 hdrs = INTERNAL_HDRS,
3479 aarch64_copts = ["-march=armv8.2-a+fp16"],
3480 aarch64_srcs = AARCH64_NEONFP16ARITH_UKERNELS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003481 gcc_copts = xnnpack_gcc_std_copts(),
3482 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08003483 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003484 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003485 "@FP16",
3486 "@pthreadpool",
3487 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003488)
3489
3490xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003491 name = "neonfp16arith_ukernels_test_mode",
3492 hdrs = INTERNAL_HDRS,
3493 aarch64_copts = ["-march=armv8.2-a+fp16"],
3494 aarch64_srcs = AARCH64_NEONFP16ARITH_UKERNELS,
3495 copts = [
3496 "-UNDEBUG",
3497 "-DXNN_TEST_MODE=1",
3498 ],
3499 gcc_copts = xnnpack_gcc_std_copts(),
3500 msvc_copts = xnnpack_msvc_std_copts(),
3501 deps = [
3502 ":tables",
3503 "@FP16",
3504 "@pthreadpool",
3505 ],
3506)
3507
3508xnnpack_cc_library(
Benoit Jacoba9644732020-08-13 12:48:55 -07003509 name = "neondot_ukernels",
3510 hdrs = INTERNAL_HDRS,
Marat Dukhan799ac752020-08-13 16:08:03 -07003511 aarch32_copts = [
3512 "-marm",
3513 "-march=armv8.2-a+dotprod",
3514 "-mfpu=neon-fp-armv8",
3515 ],
Benoit Jacoba9644732020-08-13 12:48:55 -07003516 aarch32_srcs = NEONDOT_UKERNELS,
3517 aarch64_copts = ["-march=armv8.2-a+dotprod"],
3518 aarch64_srcs = NEONDOT_UKERNELS,
3519 gcc_copts = xnnpack_gcc_std_copts(),
3520 msvc_copts = xnnpack_msvc_std_copts(),
3521 deps = [
3522 ":tables",
3523 "@FP16",
3524 "@pthreadpool",
3525 ],
3526)
3527
3528xnnpack_cc_library(
3529 name = "neondot_ukernels_test_mode",
3530 hdrs = INTERNAL_HDRS,
Marat Dukhan799ac752020-08-13 16:08:03 -07003531 aarch32_copts = [
3532 "-marm",
3533 "-march=armv8.2-a+dotprod",
3534 "-mfpu=neon-fp-armv8",
3535 ],
Benoit Jacoba9644732020-08-13 12:48:55 -07003536 aarch32_srcs = NEONDOT_UKERNELS,
3537 aarch64_copts = ["-march=armv8.2-a+dotprod"],
3538 aarch64_srcs = NEONDOT_UKERNELS,
3539 copts = [
3540 "-UNDEBUG",
3541 "-DXNN_TEST_MODE=1",
3542 ],
3543 gcc_copts = xnnpack_gcc_std_copts(),
3544 msvc_copts = xnnpack_msvc_std_copts(),
3545 deps = [
3546 ":tables",
3547 "@FP16",
3548 "@pthreadpool",
3549 ],
3550)
3551
3552xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003553 name = "sse2_ukernels",
3554 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003555 gcc_copts = xnnpack_gcc_std_copts(),
3556 gcc_x86_copts = ["-msse2"],
3557 msvc_copts = xnnpack_msvc_std_copts(),
3558 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003559 x86_srcs = SSE_UKERNELS + SSE2_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08003560 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003561 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003562 "@FP16",
3563 "@pthreadpool",
3564 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003565)
3566
3567xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003568 name = "sse2_ukernels_test_mode",
3569 hdrs = INTERNAL_HDRS,
3570 copts = [
3571 "-UNDEBUG",
3572 "-DXNN_TEST_MODE=1",
3573 ],
3574 gcc_copts = xnnpack_gcc_std_copts(),
3575 gcc_x86_copts = ["-msse2"],
3576 msvc_copts = xnnpack_msvc_std_copts(),
3577 msvc_x86_32_copts = ["/arch:SSE2"],
3578 x86_srcs = SSE_UKERNELS + SSE2_UKERNELS,
3579 deps = [
3580 ":tables",
3581 "@FP16",
3582 "@pthreadpool",
3583 ],
3584)
3585
3586xnnpack_cc_library(
Marat Dukhanfe7acb62020-03-09 19:30:05 -07003587 name = "ssse3_ukernels",
3588 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003589 gcc_copts = xnnpack_gcc_std_copts(),
3590 gcc_x86_copts = ["-mssse3"],
3591 msvc_copts = xnnpack_msvc_std_copts(),
3592 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhanfe7acb62020-03-09 19:30:05 -07003593 x86_srcs = SSSE3_UKERNELS,
3594 deps = [
3595 ":tables",
3596 "@FP16",
3597 "@pthreadpool",
3598 ],
3599)
3600
3601xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003602 name = "ssse3_ukernels_test_mode",
3603 hdrs = INTERNAL_HDRS,
3604 copts = [
3605 "-UNDEBUG",
3606 "-DXNN_TEST_MODE=1",
3607 ],
3608 gcc_copts = xnnpack_gcc_std_copts(),
3609 gcc_x86_copts = ["-mssse3"],
3610 msvc_copts = xnnpack_msvc_std_copts(),
3611 msvc_x86_32_copts = ["/arch:SSE2"],
3612 x86_srcs = SSSE3_UKERNELS,
3613 deps = [
3614 ":tables",
3615 "@FP16",
3616 "@pthreadpool",
3617 ],
3618)
3619
3620xnnpack_cc_library(
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08003621 name = "sse41_ukernels",
3622 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003623 gcc_copts = xnnpack_gcc_std_copts(),
3624 gcc_x86_copts = ["-msse4.1"],
3625 msvc_copts = xnnpack_msvc_std_copts(),
3626 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08003627 x86_srcs = SSE41_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08003628 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003629 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003630 "@FP16",
3631 "@pthreadpool",
3632 ],
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08003633)
3634
3635xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003636 name = "sse41_ukernels_test_mode",
3637 hdrs = INTERNAL_HDRS,
3638 copts = [
3639 "-UNDEBUG",
3640 "-DXNN_TEST_MODE=1",
3641 ],
3642 gcc_copts = xnnpack_gcc_std_copts(),
3643 gcc_x86_copts = ["-msse4.1"],
3644 msvc_copts = xnnpack_msvc_std_copts(),
3645 msvc_x86_32_copts = ["/arch:SSE2"],
3646 x86_srcs = SSE41_UKERNELS,
3647 deps = [
3648 ":tables",
3649 "@FP16",
3650 "@pthreadpool",
3651 ],
3652)
3653
3654xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003655 name = "avx_ukernels",
3656 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003657 gcc_copts = xnnpack_gcc_std_copts(),
3658 gcc_x86_copts = ["-mavx"],
3659 msvc_copts = xnnpack_msvc_std_copts(),
3660 msvc_x86_32_copts = ["/arch:AVX"],
3661 msvc_x86_64_copts = ["/arch:AVX"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003662 x86_srcs = AVX_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08003663 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003664 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003665 "@FP16",
3666 "@pthreadpool",
3667 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003668)
3669
3670xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003671 name = "avx_ukernels_test_mode",
3672 hdrs = INTERNAL_HDRS,
3673 copts = [
3674 "-UNDEBUG",
3675 "-DXNN_TEST_MODE=1",
3676 ],
3677 gcc_copts = xnnpack_gcc_std_copts(),
3678 gcc_x86_copts = ["-mavx"],
3679 msvc_copts = xnnpack_msvc_std_copts(),
3680 msvc_x86_32_copts = ["/arch:AVX"],
3681 msvc_x86_64_copts = ["/arch:AVX"],
3682 x86_srcs = AVX_UKERNELS,
3683 deps = [
3684 ":tables",
3685 "@FP16",
3686 "@pthreadpool",
3687 ],
3688)
3689
3690xnnpack_cc_library(
Marat Dukhan1566fee2020-08-02 21:55:41 -07003691 name = "xop_ukernels",
3692 hdrs = INTERNAL_HDRS,
3693 gcc_copts = xnnpack_gcc_std_copts(),
3694 gcc_x86_copts = ["-mxop"],
3695 msvc_copts = xnnpack_msvc_std_copts(),
3696 msvc_x86_32_copts = ["/arch:AVX"],
3697 msvc_x86_64_copts = ["/arch:AVX"],
3698 x86_srcs = XOP_UKERNELS,
3699 deps = [
3700 ":tables",
3701 "@FP16",
3702 "@pthreadpool",
3703 ],
3704)
3705
3706xnnpack_cc_library(
3707 name = "xop_ukernels_test_mode",
3708 hdrs = INTERNAL_HDRS,
3709 copts = [
3710 "-UNDEBUG",
3711 "-DXNN_TEST_MODE=1",
3712 ],
3713 gcc_copts = xnnpack_gcc_std_copts(),
3714 gcc_x86_copts = ["-mxop"],
3715 msvc_copts = xnnpack_msvc_std_copts(),
3716 msvc_x86_32_copts = ["/arch:AVX"],
3717 msvc_x86_64_copts = ["/arch:AVX"],
3718 x86_srcs = XOP_UKERNELS,
3719 deps = [
3720 ":tables",
3721 "@FP16",
3722 "@pthreadpool",
3723 ],
3724)
3725
3726xnnpack_cc_library(
Marat Dukhanfda12b82019-11-21 12:27:59 -08003727 name = "fma3_ukernels",
3728 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003729 gcc_copts = xnnpack_gcc_std_copts(),
3730 gcc_x86_copts = ["-mfma"],
3731 msvc_copts = xnnpack_msvc_std_copts(),
3732 msvc_x86_32_copts = ["/arch:AVX"],
3733 msvc_x86_64_copts = ["/arch:AVX"],
Marat Dukhanfda12b82019-11-21 12:27:59 -08003734 x86_srcs = FMA3_UKERNELS,
3735 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003736 ":tables",
Marat Dukhanfda12b82019-11-21 12:27:59 -08003737 "@FP16",
3738 "@pthreadpool",
3739 ],
3740)
3741
3742xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003743 name = "fma3_ukernels_test_mode",
3744 hdrs = INTERNAL_HDRS,
3745 copts = [
3746 "-UNDEBUG",
3747 "-DXNN_TEST_MODE=1",
3748 ],
3749 gcc_copts = xnnpack_gcc_std_copts(),
3750 gcc_x86_copts = ["-mfma"],
3751 msvc_copts = xnnpack_msvc_std_copts(),
3752 msvc_x86_32_copts = ["/arch:AVX"],
3753 msvc_x86_64_copts = ["/arch:AVX"],
3754 x86_srcs = FMA3_UKERNELS,
3755 deps = [
3756 ":tables",
3757 "@FP16",
3758 "@pthreadpool",
3759 ],
3760)
3761
3762xnnpack_cc_library(
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003763 name = "avx2_ukernels",
3764 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003765 gcc_copts = xnnpack_gcc_std_copts(),
3766 gcc_x86_copts = [
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003767 "-mfma",
3768 "-mavx2",
3769 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07003770 msvc_copts = xnnpack_msvc_std_copts(),
3771 msvc_x86_32_copts = ["/arch:AVX2"],
3772 msvc_x86_64_copts = ["/arch:AVX2"],
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003773 x86_srcs = AVX2_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08003774 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003775 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003776 "@FP16",
3777 "@pthreadpool",
3778 ],
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003779)
3780
3781xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003782 name = "avx2_ukernels_test_mode",
3783 hdrs = INTERNAL_HDRS,
3784 copts = [
3785 "-UNDEBUG",
3786 "-DXNN_TEST_MODE=1",
3787 ],
3788 gcc_copts = xnnpack_gcc_std_copts(),
3789 gcc_x86_copts = [
3790 "-mfma",
3791 "-mavx2",
3792 ],
3793 msvc_copts = xnnpack_msvc_std_copts(),
3794 msvc_x86_32_copts = ["/arch:AVX2"],
3795 msvc_x86_64_copts = ["/arch:AVX2"],
3796 x86_srcs = AVX2_UKERNELS,
3797 deps = [
3798 ":tables",
3799 "@FP16",
3800 "@pthreadpool",
3801 ],
3802)
3803
3804xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003805 name = "avx512f_ukernels",
3806 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003807 gcc_copts = xnnpack_gcc_std_copts(),
3808 gcc_x86_copts = ["-mavx512f"],
3809 mingw_copts = ["-fno-asynchronous-unwind-tables"],
3810 msvc_copts = xnnpack_msvc_std_copts(),
3811 msvc_x86_32_copts = ["/arch:AVX512"],
3812 msvc_x86_64_copts = ["/arch:AVX512"],
3813 msys_copts = ["-fno-asynchronous-unwind-tables"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003814 x86_srcs = AVX512F_UKERNELS,
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 = "avx512f_ukernels_test_mode",
3824 hdrs = INTERNAL_HDRS,
3825 copts = [
3826 "-UNDEBUG",
3827 "-DXNN_TEST_MODE=1",
3828 ],
3829 gcc_copts = xnnpack_gcc_std_copts(),
3830 gcc_x86_copts = ["-mavx512f"],
3831 mingw_copts = ["-fno-asynchronous-unwind-tables"],
3832 msvc_copts = xnnpack_msvc_std_copts(),
3833 msvc_x86_32_copts = ["/arch:AVX512"],
3834 msvc_x86_64_copts = ["/arch:AVX512"],
3835 msys_copts = ["-fno-asynchronous-unwind-tables"],
3836 x86_srcs = AVX512F_UKERNELS,
3837 deps = [
3838 ":tables",
3839 "@FP16",
3840 "@pthreadpool",
3841 ],
3842)
3843
3844xnnpack_cc_library(
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07003845 name = "avx512skx_ukernels",
3846 hdrs = INTERNAL_HDRS,
3847 gcc_copts = xnnpack_gcc_std_copts(),
3848 gcc_x86_copts = [
3849 "-mavx512f",
3850 "-mavx512cd",
3851 "-mavx512bw",
3852 "-mavx512dq",
3853 "-mavx512vl",
3854 ],
3855 mingw_copts = ["-fno-asynchronous-unwind-tables"],
3856 msvc_copts = xnnpack_msvc_std_copts(),
3857 msvc_x86_32_copts = ["/arch:AVX512"],
3858 msvc_x86_64_copts = ["/arch:AVX512"],
3859 msys_copts = ["-fno-asynchronous-unwind-tables"],
3860 x86_srcs = AVX512SKX_UKERNELS,
3861 deps = [
3862 ":tables",
3863 "@FP16",
3864 "@pthreadpool",
3865 ],
3866)
3867
3868xnnpack_cc_library(
3869 name = "avx512skx_ukernels_test_mode",
3870 hdrs = INTERNAL_HDRS,
3871 copts = [
3872 "-UNDEBUG",
3873 "-DXNN_TEST_MODE=1",
3874 ],
3875 gcc_copts = xnnpack_gcc_std_copts(),
3876 gcc_x86_copts = [
3877 "-mavx512f",
3878 "-mavx512cd",
3879 "-mavx512bw",
3880 "-mavx512dq",
3881 "-mavx512vl",
3882 ],
3883 mingw_copts = ["-fno-asynchronous-unwind-tables"],
3884 msvc_copts = xnnpack_msvc_std_copts(),
3885 msvc_x86_32_copts = ["/arch:AVX512"],
3886 msvc_x86_64_copts = ["/arch:AVX512"],
3887 msys_copts = ["-fno-asynchronous-unwind-tables"],
3888 x86_srcs = AVX512SKX_UKERNELS,
3889 deps = [
3890 ":tables",
3891 "@FP16",
3892 "@pthreadpool",
3893 ],
3894)
3895
3896xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003897 name = "asm_ukernels",
3898 hdrs = ["src/xnnpack/assembly.h"],
3899 aarch32_srcs = AARCH32_ASM_UKERNELS,
Frank Barchard31bb45b2020-10-06 00:26:33 -07003900 aarch64_copts = ["-march=armv8.2-a+fp16+dotprod"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003901 aarch64_srcs = AARCH64_ASM_UKERNELS,
Frank Barchardbcedc082020-08-17 18:00:51 -07003902 wasm_srcs = WASM32_ASM_UKERNELS,
3903 wasmsimd_srcs = WASM32_ASM_UKERNELS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003904)
3905
Marat Dukhan3b59de22020-06-03 20:15:19 -07003906xnnpack_cc_library(
3907 name = "logging_utils",
3908 srcs = LOGGING_SRCS,
3909 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
3910 copts = LOGGING_COPTS + [
3911 "-Isrc",
3912 "-Iinclude",
3913 ] + select({
3914 ":debug_build": [],
3915 "//conditions:default": xnnpack_min_size_copts(),
3916 }),
3917 gcc_copts = xnnpack_gcc_std_copts(),
3918 msvc_copts = xnnpack_msvc_std_copts(),
3919 visibility = xnnpack_visibility(),
3920 deps = [
3921 "@FP16",
3922 "@clog",
3923 "@pthreadpool",
3924 ],
3925)
3926
Marat Dukhan08c4a432019-10-03 09:29:21 -07003927xnnpack_aggregate_library(
3928 name = "ukernels",
3929 aarch32_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003930 ":neon_ukernels",
3931 ":neonfma_ukernels",
Marat Dukhan8853b822020-05-07 12:19:01 -07003932 ":neonv8_ukernels",
Benoit Jacoba9644732020-08-13 12:48:55 -07003933 ":neondot_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003934 ":asm_ukernels",
3935 ],
3936 aarch64_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003937 ":neon_ukernels",
3938 ":neonfma_ukernels",
Marat Dukhan8853b822020-05-07 12:19:01 -07003939 ":neonv8_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003940 ":neonfp16arith_ukernels",
Benoit Jacoba9644732020-08-13 12:48:55 -07003941 ":neondot_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003942 ":asm_ukernels",
3943 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07003944 generic_deps = [
3945 ":scalar_ukernels",
3946 ],
Marat Dukhan500b8892020-04-15 17:09:50 -07003947 psimd_deps = [
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08003948 ":psimd_accmath_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003949 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07003950 wasm_deps = [
3951 ":wasm_ukernels",
Frank Barchardbcedc082020-08-17 18:00:51 -07003952 ":asm_ukernels",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003953 ],
3954 wasmsimd_deps = [
3955 ":wasm_ukernels",
Frank Barchardbcedc082020-08-17 18:00:51 -07003956 ":asm_ukernels",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003957 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003958 x86_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003959 ":sse2_ukernels",
Marat Dukhanfe7acb62020-03-09 19:30:05 -07003960 ":ssse3_ukernels",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08003961 ":sse41_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003962 ":avx_ukernels",
Marat Dukhan1566fee2020-08-02 21:55:41 -07003963 ":xop_ukernels",
Marat Dukhanfda12b82019-11-21 12:27:59 -08003964 ":fma3_ukernels",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003965 ":avx2_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003966 ":avx512f_ukernels",
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07003967 ":avx512skx_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003968 ],
3969)
3970
Marat Dukhan33fcf782020-05-24 14:27:15 -07003971xnnpack_aggregate_library(
3972 name = "ukernels_test_mode",
3973 aarch32_deps = [
3974 ":neon_ukernels_test_mode",
3975 ":neonfma_ukernels_test_mode",
3976 ":neonv8_ukernels_test_mode",
Benoit Jacoba9644732020-08-13 12:48:55 -07003977 ":neondot_ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003978 ":asm_ukernels",
3979 ],
3980 aarch64_deps = [
3981 ":neon_ukernels_test_mode",
3982 ":neonfma_ukernels_test_mode",
3983 ":neonv8_ukernels_test_mode",
3984 ":neonfp16arith_ukernels_test_mode",
Benoit Jacoba9644732020-08-13 12:48:55 -07003985 ":neondot_ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003986 ":asm_ukernels",
3987 ],
3988 generic_deps = [
3989 ":scalar_ukernels_test_mode",
3990 ],
3991 psimd_deps = [
Marat Dukhan33fcf782020-05-24 14:27:15 -07003992 ":psimd_accmath_ukernels_test_mode",
3993 ],
3994 wasm_deps = [
3995 ":wasm_ukernels_test_mode",
Frank Barchardbcedc082020-08-17 18:00:51 -07003996 ":asm_ukernels",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003997 ],
3998 wasmsimd_deps = [
3999 ":wasm_ukernels_test_mode",
Frank Barchardbcedc082020-08-17 18:00:51 -07004000 ":asm_ukernels",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004001 ],
4002 x86_deps = [
4003 ":sse2_ukernels_test_mode",
4004 ":ssse3_ukernels_test_mode",
4005 ":sse41_ukernels_test_mode",
4006 ":avx_ukernels_test_mode",
Marat Dukhan1566fee2020-08-02 21:55:41 -07004007 ":xop_ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004008 ":fma3_ukernels_test_mode",
4009 ":avx2_ukernels_test_mode",
4010 ":avx512f_ukernels_test_mode",
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07004011 ":avx512skx_ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004012 ],
4013)
4014
Marat Dukhan86df3842020-06-17 14:03:31 -07004015# This target must be compatible with Android NDK r17 and compile with Clang 6.
4016# It is similar to :ukernels target, but excludes microkernels which require
4017# NEON FP16 arithmetics, NEON Dot Product, and newer intrinsics.
Marat Dukhanae046f52020-06-15 13:16:14 -07004018xnnpack_aggregate_library(
Marat Dukhan86df3842020-06-17 14:03:31 -07004019 name = "ukernels_max_compatibility",
Marat Dukhanae046f52020-06-15 13:16:14 -07004020 aarch32_deps = [
4021 ":neon_ukernels",
4022 ":neonfma_ukernels",
4023 ":neonv8_ukernels",
4024 ":asm_ukernels",
4025 ],
4026 aarch64_deps = [
4027 ":neon_ukernels",
4028 ":neonfma_ukernels",
4029 ":neonv8_ukernels",
4030 ":asm_ukernels",
4031 ],
4032 generic_deps = [
4033 ":scalar_ukernels",
4034 ],
4035 psimd_deps = [
Marat Dukhanae046f52020-06-15 13:16:14 -07004036 ":psimd_accmath_ukernels",
4037 ],
4038 wasm_deps = [
4039 ":wasm_ukernels",
4040 ],
4041 wasmsimd_deps = [
4042 ":wasm_ukernels",
4043 ],
4044 x86_deps = [
4045 ":sse2_ukernels",
4046 ":ssse3_ukernels",
4047 ":sse41_ukernels",
4048 ":avx_ukernels",
Marat Dukhan1566fee2020-08-02 21:55:41 -07004049 ":xop_ukernels",
Marat Dukhanae046f52020-06-15 13:16:14 -07004050 ":fma3_ukernels",
4051 ":avx2_ukernels",
4052 ":avx512f_ukernels",
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07004053 ":avx512skx_ukernels",
Marat Dukhanae046f52020-06-15 13:16:14 -07004054 ],
4055)
4056
Marat Dukhan08c4a432019-10-03 09:29:21 -07004057xnnpack_cc_library(
4058 name = "im2col",
4059 srcs = ["src/im2col.c"],
4060 hdrs = [
4061 "src/xnnpack/common.h",
4062 "src/xnnpack/im2col.h",
4063 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07004064 gcc_copts = xnnpack_gcc_std_copts(),
4065 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004066)
4067
4068xnnpack_cc_library(
4069 name = "indirection",
4070 srcs = ["src/indirection.c"],
4071 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004072 gcc_copts = xnnpack_gcc_std_copts(),
4073 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004074 deps = [
4075 "@FP16",
4076 "@FXdiv",
4077 "@pthreadpool",
4078 ],
4079)
4080
4081xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07004082 name = "indirection_test_mode",
4083 srcs = ["src/indirection.c"],
4084 hdrs = INTERNAL_HDRS,
4085 copts = [
4086 "-UNDEBUG",
4087 "-DXNN_TEST_MODE=1",
4088 ],
4089 gcc_copts = xnnpack_gcc_std_copts(),
4090 msvc_copts = xnnpack_msvc_std_copts(),
4091 deps = [
4092 "@FP16",
4093 "@FXdiv",
4094 "@pthreadpool",
4095 ],
4096)
4097
4098xnnpack_cc_library(
Marat Dukhanab582382020-07-06 13:32:08 -07004099 name = "packing",
4100 srcs = ["src/packing.c"],
4101 hdrs = INTERNAL_HDRS,
4102 gcc_copts = xnnpack_gcc_std_copts(),
4103 msvc_copts = xnnpack_msvc_std_copts(),
4104 deps = [
4105 "@FP16",
4106 "@FXdiv",
4107 "@pthreadpool",
4108 ],
4109)
4110
4111xnnpack_cc_library(
4112 name = "packing_test_mode",
4113 srcs = ["src/packing.c"],
4114 hdrs = INTERNAL_HDRS,
4115 copts = [
4116 "-UNDEBUG",
4117 "-DXNN_TEST_MODE=1",
4118 ],
4119 gcc_copts = xnnpack_gcc_std_copts(),
4120 msvc_copts = xnnpack_msvc_std_copts(),
4121 deps = [
4122 "@FP16",
4123 "@FXdiv",
4124 "@pthreadpool",
4125 ],
4126)
4127
4128xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004129 name = "operator_run",
4130 srcs = ["src/operator-run.c"],
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07004131 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004132 copts = LOGGING_COPTS + select({
Marat Dukhan05702cf2020-03-26 15:41:33 -07004133 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4134 "//conditions:default": [],
4135 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07004136 gcc_copts = xnnpack_gcc_std_copts(),
4137 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004138 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07004139 ":logging_utils",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004140 "@FP16",
4141 "@FXdiv",
4142 "@clog",
4143 "@pthreadpool",
4144 ],
4145)
4146
Chao Mei6ddfc602020-05-13 22:29:36 -07004147xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07004148 name = "operator_run_test_mode",
4149 srcs = ["src/operator-run.c"],
4150 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
4151 copts = LOGGING_COPTS + [
4152 "-UNDEBUG",
4153 "-DXNN_TEST_MODE=1",
4154 ] + select({
4155 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4156 "//conditions:default": [],
4157 }),
4158 gcc_copts = xnnpack_gcc_std_copts(),
4159 msvc_copts = xnnpack_msvc_std_copts(),
4160 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07004161 ":logging_utils",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004162 "@FP16",
4163 "@FXdiv",
4164 "@clog",
4165 "@pthreadpool",
4166 ],
4167)
4168
4169xnnpack_cc_library(
Chao Mei6ddfc602020-05-13 22:29:36 -07004170 name = "memory_planner",
4171 srcs = ["src/memory-planner.c"],
4172 hdrs = INTERNAL_HDRS,
4173 defines = select({
4174 ":xnn_enable_memopt_explicit_true": ["XNN_ENABLE_MEMOPT=1"],
4175 ":xnn_enable_memopt_explicit_false": ["XNN_ENABLE_MEMOPT=0"],
4176 "//conditions:default": ["XNN_ENABLE_MEMOPT=1"],
4177 }),
4178 gcc_copts = xnnpack_gcc_std_copts(),
4179 msvc_copts = xnnpack_msvc_std_copts(),
4180 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07004181 ":logging_utils",
Chao Mei6ddfc602020-05-13 22:29:36 -07004182 "@pthreadpool",
4183 ],
4184)
4185
Marat Dukhan33fcf782020-05-24 14:27:15 -07004186xnnpack_cc_library(
4187 name = "memory_planner_test_mode",
4188 srcs = ["src/memory-planner.c"],
4189 hdrs = INTERNAL_HDRS,
4190 copts = [
4191 "-UNDEBUG",
4192 "-DXNN_TEST_MODE=1",
4193 ],
4194 defines = select({
4195 ":xnn_enable_memopt_explicit_true": ["XNN_ENABLE_MEMOPT=1"],
4196 ":xnn_enable_memopt_explicit_false": ["XNN_ENABLE_MEMOPT=0"],
4197 "//conditions:default": ["XNN_ENABLE_MEMOPT=1"],
4198 }),
4199 gcc_copts = xnnpack_gcc_std_copts(),
4200 msvc_copts = xnnpack_msvc_std_copts(),
4201 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07004202 ":logging_utils",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004203 "@pthreadpool",
4204 ],
4205)
4206
Marat Dukhan08c4a432019-10-03 09:29:21 -07004207cc_library(
4208 name = "enable_assembly",
4209 defines = select({
4210 ":xnn_enable_assembly_explicit_true": ["XNN_ENABLE_ASSEMBLY=1"],
4211 ":xnn_enable_assembly_explicit_false": ["XNN_ENABLE_ASSEMBLY=0"],
Frank Barchard810171d2019-10-10 10:34:51 -07004212 "//conditions:default": ["XNN_ENABLE_ASSEMBLY=1"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004213 }),
4214)
4215
Marat Dukhan9de90e02020-06-18 16:04:12 -07004216cc_library(
4217 name = "enable_sparse",
4218 defines = select({
4219 ":xnn_enable_sparse_explicit_true": ["XNN_ENABLE_SPARSE=1"],
4220 ":xnn_enable_sparse_explicit_false": ["XNN_ENABLE_SPARSE=0"],
Marat Dukhan8b283aa2020-06-18 16:48:57 -07004221 "//conditions:default": ["XNN_ENABLE_SPARSE=0"],
Marat Dukhan9de90e02020-06-18 16:04:12 -07004222 }),
4223)
4224
Marat Dukhancf056b22019-10-07 10:26:29 -07004225xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004226 name = "operators",
4227 srcs = OPERATOR_SRCS + [
Marat Dukhan04f03be2019-11-19 12:36:47 -08004228 "src/memory.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004229 "src/operator-delete.c",
Marat Dukhancf056b22019-10-07 10:26:29 -07004230 ],
4231 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004232 copts = LOGGING_COPTS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07004233 "-Isrc",
4234 "-Iinclude",
4235 ] + select({
4236 ":debug_build": [],
4237 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07004238 }) + select({
4239 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4240 "//conditions:default": [],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004241 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07004242 gcc_copts = xnnpack_gcc_std_copts(),
4243 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004244 deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07004245 ":indirection",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004246 ":logging_utils",
Marat Dukhanab582382020-07-06 13:32:08 -07004247 ":packing",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004248 "@FP16",
4249 "@FXdiv",
4250 "@clog",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004251 "@pthreadpool",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004252 ],
4253)
4254
Marat Dukhan10a38082020-04-17 03:58:35 -07004255xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07004256 name = "operators_test_mode",
4257 srcs = OPERATOR_SRCS + [
4258 "src/memory.c",
4259 "src/operator-delete.c",
4260 ],
4261 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
4262 copts = LOGGING_COPTS + [
4263 "-Isrc",
4264 "-Iinclude",
4265 "-UNDEBUG",
4266 "-DXNN_TEST_MODE=1",
4267 ] + select({
4268 ":debug_build": [],
4269 "//conditions:default": xnnpack_min_size_copts(),
4270 }) + select({
4271 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4272 "//conditions:default": [],
4273 }),
4274 gcc_copts = xnnpack_gcc_std_copts(),
4275 msvc_copts = xnnpack_msvc_std_copts(),
4276 deps = [
4277 ":indirection_test_mode",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004278 ":logging_utils",
Marat Dukhanab582382020-07-06 13:32:08 -07004279 ":packing_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004280 "@FP16",
4281 "@FXdiv",
4282 "@clog",
4283 "@pthreadpool",
4284 ],
4285)
4286
4287xnnpack_cc_library(
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004288 name = "XNNPACK",
4289 srcs = [
4290 "src/init.c",
Marat Dukhanccfdbd12020-02-03 14:27:45 -08004291 "src/runtime.c",
4292 "src/subgraph.c",
4293 "src/tensor.c",
Marat Dukhanf03da0d2020-06-10 16:00:20 -07004294 ] + SUBGRAPH_SRCS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004295 hdrs = ["include/xnnpack.h"],
4296 copts = LOGGING_COPTS + [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004297 "-Isrc",
4298 "-Iinclude",
4299 ] + select({
4300 ":debug_build": [],
4301 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07004302 }) + select({
4303 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4304 "//conditions:default": [],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004305 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07004306 gcc_copts = xnnpack_gcc_std_copts(),
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004307 includes = ["include"],
Marat Dukhan10a38082020-04-17 03:58:35 -07004308 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004309 visibility = xnnpack_visibility(),
4310 deps = [
4311 ":enable_assembly",
Marat Dukhan9de90e02020-06-18 16:04:12 -07004312 ":enable_sparse",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004313 ":logging_utils",
Chao Mei6ddfc602020-05-13 22:29:36 -07004314 ":memory_planner",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004315 ":operator_run",
4316 ":operators",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004317 ":ukernels",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004318 "@clog",
Marat Dukhanab2946c2020-05-21 20:04:13 -07004319 "@FP16",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004320 "@pthreadpool",
Marat Dukhand343c222019-10-07 09:22:14 -07004321 ] + select({
4322 ":emscripten": [],
4323 "//conditions:default": ["@cpuinfo"],
4324 }),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004325)
4326
Marat Dukhan10a38082020-04-17 03:58:35 -07004327xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07004328 name = "XNNPACK_test_mode",
4329 srcs = [
4330 "src/init.c",
4331 "src/runtime.c",
4332 "src/subgraph.c",
4333 "src/tensor.c",
Marat Dukhanf03da0d2020-06-10 16:00:20 -07004334 ] + SUBGRAPH_SRCS,
Marat Dukhan33fcf782020-05-24 14:27:15 -07004335 hdrs = ["include/xnnpack.h"],
4336 copts = LOGGING_COPTS + [
4337 "-Isrc",
4338 "-Iinclude",
4339 "-UNDEBUG",
4340 "-DXNN_TEST_MODE=1",
4341 ] + select({
4342 ":debug_build": [],
4343 "//conditions:default": xnnpack_min_size_copts(),
4344 }) + select({
4345 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4346 "//conditions:default": [],
4347 }),
4348 gcc_copts = xnnpack_gcc_std_copts(),
4349 includes = ["include"],
4350 msvc_copts = xnnpack_msvc_std_copts(),
4351 visibility = xnnpack_visibility(),
4352 deps = [
4353 ":enable_assembly",
Marat Dukhan9de90e02020-06-18 16:04:12 -07004354 ":enable_sparse",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004355 ":logging_utils",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004356 ":memory_planner_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004357 ":operator_run_test_mode",
4358 ":operators_test_mode",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004359 ":ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004360 "@clog",
4361 "@FP16",
4362 "@pthreadpool",
4363 ] + select({
4364 ":emscripten": [],
4365 "//conditions:default": ["@cpuinfo"],
4366 }),
4367)
4368
4369xnnpack_cc_library(
Marat Dukhanae046f52020-06-15 13:16:14 -07004370 name = "xnnpack_f32",
4371 srcs = [
4372 "src/init.c",
4373 "src/runtime.c",
4374 "src/subgraph.c",
4375 "src/tensor.c",
4376 ] + SUBGRAPH_SRCS,
4377 hdrs = ["include/xnnpack.h"],
4378 copts = LOGGING_COPTS + [
4379 "-Isrc",
4380 "-Iinclude",
4381 ] + select({
4382 ":debug_build": [],
4383 "//conditions:default": xnnpack_min_size_copts(),
4384 }) + select({
4385 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4386 "//conditions:default": [],
4387 }),
4388 defines = [
Marat Dukhan16f1e1a2020-08-04 16:38:22 -07004389 "XNN_NO_QS8_OPERATORS",
Marat Dukhan08b7a972020-07-14 18:17:29 -07004390 "XNN_NO_QU8_OPERATORS",
Marat Dukhanae046f52020-06-15 13:16:14 -07004391 "XNN_NO_U8_OPERATORS",
4392 "XNN_NO_X8_OPERATORS",
4393 "XNN_NO_F16_OPERATORS",
4394 "XNN_NO_X16_OPERATORS",
4395 ],
4396 gcc_copts = xnnpack_gcc_std_copts(),
4397 includes = ["include"],
4398 msvc_copts = xnnpack_msvc_std_copts(),
4399 visibility = xnnpack_visibility(),
4400 deps = [
4401 ":enable_assembly",
Marat Dukhan9de90e02020-06-18 16:04:12 -07004402 ":enable_sparse",
Marat Dukhanae046f52020-06-15 13:16:14 -07004403 ":logging_utils",
4404 ":memory_planner",
4405 ":operator_run",
4406 ":operators",
Marat Dukhan86df3842020-06-17 14:03:31 -07004407 ":ukernels_max_compatibility",
Marat Dukhanae046f52020-06-15 13:16:14 -07004408 "@clog",
4409 "@FP16",
4410 "@pthreadpool",
4411 ] + select({
4412 ":emscripten": [],
4413 "//conditions:default": ["@cpuinfo"],
4414 }),
4415)
4416
4417xnnpack_cc_library(
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004418 name = "xnnpack_operators_nhwc_f32",
4419 srcs = [
4420 "src/init.c",
4421 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07004422 hdrs = ["include/xnnpack.h"],
4423 copts = LOGGING_COPTS + [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004424 "-Isrc",
4425 "-Iinclude",
4426 ] + select({
4427 ":debug_build": [],
4428 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07004429 }) + select({
4430 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4431 "//conditions:default": [],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004432 }),
4433 defines = [
Marat Dukhan16f1e1a2020-08-04 16:38:22 -07004434 "XNN_NO_QS8_OPERATORS",
Marat Dukhan08b7a972020-07-14 18:17:29 -07004435 "XNN_NO_QU8_OPERATORS",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004436 "XNN_NO_U8_OPERATORS",
4437 "XNN_NO_X8_OPERATORS",
Marat Dukhanefc47b82019-11-18 09:25:38 -08004438 "XNN_NO_NCHW_OPERATORS",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004439 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07004440 gcc_copts = xnnpack_gcc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004441 includes = ["include"],
Marat Dukhan10a38082020-04-17 03:58:35 -07004442 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004443 visibility = xnnpack_visibility(),
4444 deps = [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004445 ":enable_assembly",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004446 ":logging_utils",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004447 ":operator_run",
4448 ":operators",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004449 ":ukernels",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004450 "@clog",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004451 "@pthreadpool",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004452 ] + select({
4453 ":emscripten": [],
4454 "//conditions:default": ["@cpuinfo"],
4455 }),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004456)
4457
Marat Dukhancf056b22019-10-07 10:26:29 -07004458xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004459 name = "bench_utils",
4460 srcs = ["bench/utils.cc"],
4461 hdrs = ["bench/utils.h"],
Marat Dukhanbad48fe2019-11-04 10:35:22 -08004462 deps = [
4463 "@com_google_benchmark//:benchmark",
4464 "@cpuinfo",
4465 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004466)
4467
Frank Barchard7e955972019-10-11 10:34:25 -07004468######################### Benchmarks for micro-kernels #########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07004469
4470xnnpack_benchmark(
Marat Dukhan595e1702020-07-31 10:12:52 -07004471 name = "qs8_gemm_bench",
4472 srcs = [
4473 "bench/gemm.h",
4474 "bench/qs8-gemm.cc",
4475 "src/xnnpack/AlignedAllocator.h",
4476 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Frank Barchard31328cb2020-10-12 11:55:18 -07004477 copts = xnnpack_optional_ruy_copts() + xnnpack_optional_gemmlowp_copts(),
4478 deps = MICROKERNEL_BENCHMARK_DEPS + [":packing"] + xnnpack_optional_ruy_deps() + xnnpack_optional_gemmlowp_deps(),
Marat Dukhan595e1702020-07-31 10:12:52 -07004479)
4480
4481xnnpack_benchmark(
Marat Dukhan56bdd4a2020-08-03 19:47:04 -07004482 name = "qs8_requantization_bench",
4483 srcs = [
4484 "bench/qs8-requantization.cc",
Marat Dukhan56bdd4a2020-08-03 19:47:04 -07004485 "src/xnnpack/AlignedAllocator.h",
Frank Barchard04336c12020-10-22 16:48:55 -07004486 "src/xnnpack/requantization-stubs.h",
Marat Dukhan56bdd4a2020-08-03 19:47:04 -07004487 ] + MICROKERNEL_BENCHMARK_HDRS,
4488 deps = MICROKERNEL_BENCHMARK_DEPS,
4489)
4490
4491xnnpack_benchmark(
Marat Dukhan08b7a972020-07-14 18:17:29 -07004492 name = "qu8_gemm_bench",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004493 srcs = [
4494 "bench/gemm.h",
Marat Dukhan08b7a972020-07-14 18:17:29 -07004495 "bench/qu8-gemm.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004496 "src/xnnpack/AlignedAllocator.h",
4497 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004498 copts = xnnpack_optional_ruy_copts() + xnnpack_optional_gemmlowp_copts(),
Marat Dukhanab582382020-07-06 13:32:08 -07004499 deps = MICROKERNEL_BENCHMARK_DEPS + [":packing"] + xnnpack_optional_ruy_deps() + xnnpack_optional_gemmlowp_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004500)
4501
4502xnnpack_benchmark(
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07004503 name = "qu8_requantization_bench",
4504 srcs = [
4505 "bench/qu8-requantization.cc",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07004506 "src/xnnpack/AlignedAllocator.h",
Frank Barchard04336c12020-10-22 16:48:55 -07004507 "src/xnnpack/requantization-stubs.h",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07004508 ] + MICROKERNEL_BENCHMARK_HDRS,
4509 deps = MICROKERNEL_BENCHMARK_DEPS,
4510)
4511
4512xnnpack_benchmark(
Frank Barchard40d20fe2020-05-05 00:37:45 -07004513 name = "f16_igemm_bench",
4514 srcs = [
4515 "bench/f16-igemm.cc",
4516 "bench/conv.h",
4517 "bench/google/conv.h",
4518 "src/xnnpack/AlignedAllocator.h",
4519 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004520 deps = MICROKERNEL_BENCHMARK_DEPS + [
4521 ":indirection",
4522 ":packing",
4523 ],
Frank Barchard40d20fe2020-05-05 00:37:45 -07004524)
4525
4526xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004527 name = "f16_gemm_bench",
4528 srcs = [
4529 "bench/f16-gemm.cc",
4530 "bench/gemm.h",
4531 "src/xnnpack/AlignedAllocator.h",
4532 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004533 deps = MICROKERNEL_BENCHMARK_DEPS + [
4534 ":packing",
4535 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004536)
4537
4538xnnpack_benchmark(
Marat Dukhanbdb56f52020-02-05 21:42:49 -08004539 name = "f16_spmm_bench",
4540 srcs = [
4541 "bench/f16-spmm.cc",
4542 "bench/gemm.h",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08004543 "src/xnnpack/AlignedAllocator.h",
4544 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanbdb56f52020-02-05 21:42:49 -08004545 deps = MICROKERNEL_BENCHMARK_DEPS,
4546)
4547
4548xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004549 name = "f32_igemm_bench",
4550 srcs = [
4551 "bench/f32-igemm.cc",
4552 "bench/conv.h",
4553 "src/xnnpack/AlignedAllocator.h",
4554 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004555 deps = MICROKERNEL_BENCHMARK_DEPS + [
4556 ":indirection",
4557 ":packing",
4558 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004559)
4560
4561xnnpack_benchmark(
Frank Barchardc2010062020-07-19 10:04:12 -07004562 name = "f16_relu_bench",
4563 srcs = [
4564 "bench/f16-relu.cc",
4565 "src/xnnpack/AlignedAllocator.h",
4566 ] + MICROKERNEL_BENCHMARK_HDRS,
4567 deps = MICROKERNEL_BENCHMARK_DEPS,
4568)
4569
4570xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004571 name = "f32_conv_hwc_bench",
4572 srcs = [
4573 "bench/f32-conv-hwc.cc",
4574 "bench/dconv.h",
4575 "src/xnnpack/AlignedAllocator.h",
4576 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004577 deps = MICROKERNEL_BENCHMARK_DEPS + [
4578 ":packing",
4579 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004580)
4581
4582xnnpack_benchmark(
Marat Dukhan1f29b802020-05-15 23:46:39 -07004583 name = "f32_conv_hwc2chw_bench",
Erich Elsen563df5f2019-10-23 08:02:21 -07004584 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07004585 "bench/f32-conv-hwc2chw.cc",
Erich Elsen563df5f2019-10-23 08:02:21 -07004586 "bench/dconv.h",
4587 "src/xnnpack/AlignedAllocator.h",
4588 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004589 deps = MICROKERNEL_BENCHMARK_DEPS + [
4590 ":packing",
4591 ],
Erich Elsen563df5f2019-10-23 08:02:21 -07004592)
4593
4594xnnpack_benchmark(
Frank Barchard5a599a62020-06-04 20:12:44 -07004595 name = "f16_dwconv_bench",
4596 srcs = [
4597 "bench/f16-dwconv.cc",
4598 "bench/dwconv.h",
4599 "bench/google/dwconv.h",
4600 "src/xnnpack/AlignedAllocator.h",
4601 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004602 deps = MICROKERNEL_BENCHMARK_DEPS + [
4603 ":indirection",
4604 ":packing",
4605 ],
Frank Barchard5a599a62020-06-04 20:12:44 -07004606)
4607
4608xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004609 name = "f32_dwconv_bench",
4610 srcs = [
4611 "bench/f32-dwconv.cc",
4612 "bench/dwconv.h",
4613 "src/xnnpack/AlignedAllocator.h",
4614 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004615 deps = MICROKERNEL_BENCHMARK_DEPS + [
4616 ":indirection",
4617 ":packing",
4618 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004619)
4620
4621xnnpack_benchmark(
Marat Dukhanbf715f92020-10-23 20:17:00 -07004622 name = "f32_dwconv2d_chw_bench",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004623 srcs = [
Marat Dukhanbf715f92020-10-23 20:17:00 -07004624 "bench/f32-dwconv2d-chw.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004625 "bench/dwconv.h",
4626 "src/xnnpack/AlignedAllocator.h",
4627 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004628 deps = MICROKERNEL_BENCHMARK_DEPS + [
4629 ":indirection",
4630 ":packing",
4631 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004632)
4633
4634xnnpack_benchmark(
4635 name = "f32_gemm_bench",
4636 srcs = [
4637 "bench/f32-gemm.cc",
4638 "bench/gemm.h",
4639 "src/xnnpack/AlignedAllocator.h",
4640 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004641 copts = xnnpack_optional_ruy_copts(),
Marat Dukhanab582382020-07-06 13:32:08 -07004642 deps = MICROKERNEL_BENCHMARK_DEPS + [":packing"] + xnnpack_optional_ruy_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004643)
4644
4645xnnpack_benchmark(
Marat Dukhana11ca342020-06-25 23:45:07 -07004646 name = "f32_hswish_bench",
4647 srcs = [
4648 "bench/f32-hswish.cc",
4649 "src/xnnpack/AlignedAllocator.h",
4650 ] + MICROKERNEL_BENCHMARK_HDRS,
4651 deps = MICROKERNEL_BENCHMARK_DEPS,
4652)
4653
4654xnnpack_benchmark(
Marat Dukhan4c4eb002019-12-08 21:27:49 -08004655 name = "f32_raddexpminusmax_bench",
4656 srcs = [
4657 "bench/f32-raddexpminusmax.cc",
4658 "src/xnnpack/AlignedAllocator.h",
4659 ] + MICROKERNEL_BENCHMARK_HDRS,
4660 deps = MICROKERNEL_BENCHMARK_DEPS,
4661)
4662
4663xnnpack_benchmark(
4664 name = "f32_raddextexp_bench",
4665 srcs = [
4666 "bench/f32-raddextexp.cc",
4667 "src/xnnpack/AlignedAllocator.h",
4668 ] + MICROKERNEL_BENCHMARK_HDRS,
4669 deps = MICROKERNEL_BENCHMARK_DEPS,
4670)
4671
4672xnnpack_benchmark(
4673 name = "f32_raddstoreexpminusmax_bench",
4674 srcs = [
4675 "bench/f32-raddstoreexpminusmax.cc",
4676 "src/xnnpack/AlignedAllocator.h",
4677 ] + MICROKERNEL_BENCHMARK_HDRS,
4678 deps = MICROKERNEL_BENCHMARK_DEPS,
4679)
4680
4681xnnpack_benchmark(
Frank Barchard3c4a9522020-07-18 16:11:22 -07004682 name = "f32_relu_bench",
4683 srcs = [
4684 "bench/f32-relu.cc",
4685 "src/xnnpack/AlignedAllocator.h",
4686 ] + MICROKERNEL_BENCHMARK_HDRS,
4687 deps = MICROKERNEL_BENCHMARK_DEPS,
4688)
4689
4690xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004691 name = "f32_rmax_bench",
4692 srcs = [
4693 "bench/f32-rmax.cc",
4694 "src/xnnpack/AlignedAllocator.h",
4695 ] + MICROKERNEL_BENCHMARK_HDRS,
4696 deps = MICROKERNEL_BENCHMARK_DEPS,
4697)
4698
4699xnnpack_benchmark(
Marat Dukhan14bec502019-11-18 11:35:31 -08004700 name = "f32_sigmoid_bench",
4701 srcs = [
4702 "bench/f32-sigmoid.cc",
4703 "src/xnnpack/AlignedAllocator.h",
4704 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan14bec502019-11-18 11:35:31 -08004705 deps = MICROKERNEL_BENCHMARK_DEPS,
4706)
4707
4708xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004709 name = "f32_spmm_bench",
4710 srcs = [
4711 "bench/f32-spmm.cc",
4712 "bench/gemm.h",
4713 "src/xnnpack/AlignedAllocator.h",
4714 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07004715 deps = MICROKERNEL_BENCHMARK_DEPS,
4716)
4717
4718xnnpack_benchmark(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08004719 name = "f32_softmax_bench",
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07004720 srcs = [
Marat Dukhanfd8e6892020-01-27 15:25:25 -08004721 "bench/f32-softmax.cc",
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07004722 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004723 copts = xnnpack_optional_dnnl_copts(),
Marat Dukhan8d3c6932020-03-06 20:27:27 -08004724 deps = MICROKERNEL_BENCHMARK_DEPS + xnnpack_optional_dnnl_deps(),
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07004725)
4726
4727xnnpack_benchmark(
Marat Dukhan4c4eb002019-12-08 21:27:49 -08004728 name = "f32_vscaleexpminusmax_bench",
4729 srcs = [
4730 "bench/f32-vscaleexpminusmax.cc",
4731 "src/xnnpack/AlignedAllocator.h",
4732 ] + MICROKERNEL_BENCHMARK_HDRS,
4733 deps = MICROKERNEL_BENCHMARK_DEPS,
4734)
4735
4736xnnpack_benchmark(
4737 name = "f32_vscaleextexp_bench",
4738 srcs = [
4739 "bench/f32-vscaleextexp.cc",
4740 "src/xnnpack/AlignedAllocator.h",
4741 ] + MICROKERNEL_BENCHMARK_HDRS,
4742 deps = MICROKERNEL_BENCHMARK_DEPS,
4743)
4744
4745xnnpack_benchmark(
Marat Dukhanf4db2f32020-06-30 10:55:30 -07004746 name = "f32_vsqrt_bench",
4747 srcs = [
4748 "bench/f32-vsqrt.cc",
4749 "src/xnnpack/AlignedAllocator.h",
4750 ] + MICROKERNEL_BENCHMARK_HDRS,
4751 deps = MICROKERNEL_BENCHMARK_DEPS,
4752)
4753
4754xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004755 name = "f32_im2col_gemm_bench",
4756 srcs = [
4757 "bench/f32-im2col-gemm.cc",
4758 "bench/conv.h",
4759 "src/xnnpack/AlignedAllocator.h",
4760 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004761 deps = MICROKERNEL_BENCHMARK_DEPS + [
4762 ":im2col",
4763 ":packing",
4764 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004765)
4766
Marat Dukhanfe7acb62020-03-09 19:30:05 -07004767xnnpack_benchmark(
Marat Dukhanffbf96a2020-05-14 02:59:08 -07004768 name = "rounding_bench",
4769 srcs = [
4770 "bench/rounding.cc",
Marat Dukhanffbf96a2020-05-14 02:59:08 -07004771 "src/xnnpack/AlignedAllocator.h",
Frank Barchard04336c12020-10-22 16:48:55 -07004772 "src/xnnpack/math-stubs.h",
Marat Dukhanffbf96a2020-05-14 02:59:08 -07004773 ] + MICROKERNEL_BENCHMARK_HDRS,
4774 deps = MICROKERNEL_BENCHMARK_DEPS,
4775)
4776
Marat Dukhan08c4a432019-10-03 09:29:21 -07004777########################### Benchmarks for operators ###########################
4778
4779xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004780 name = "average_pooling_bench",
4781 srcs = ["bench/average-pooling.cc"],
Marat Dukhan7a16d8b2020-03-11 04:22:44 -07004782 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07004783 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004784 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004785)
4786
4787xnnpack_benchmark(
Marat Dukhan87727142020-06-24 15:24:10 -07004788 name = "bankers_rounding_bench",
4789 srcs = ["bench/bankers-rounding.cc"],
4790 copts = xnnpack_optional_tflite_copts(),
4791 tags = ["nowin32"],
4792 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
4793)
4794
4795xnnpack_benchmark(
4796 name = "ceiling_bench",
4797 srcs = ["bench/ceiling.cc"],
4798 copts = xnnpack_optional_tflite_copts(),
4799 tags = ["nowin32"],
4800 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
4801)
4802
4803xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004804 name = "channel_shuffle_bench",
4805 srcs = ["bench/channel-shuffle.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004806 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07004807)
4808
4809xnnpack_benchmark(
4810 name = "convolution_bench",
4811 srcs = ["bench/convolution.cc"],
4812 copts = xnnpack_optional_tflite_copts() + xnnpack_optional_armcl_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07004813 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004814 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps() + xnnpack_optional_armcl_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004815)
4816
4817xnnpack_benchmark(
4818 name = "deconvolution_bench",
4819 srcs = ["bench/deconvolution.cc"],
4820 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07004821 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004822 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004823)
4824
4825xnnpack_benchmark(
Marat Dukhan87727142020-06-24 15:24:10 -07004826 name = "floor_bench",
4827 srcs = ["bench/floor.cc"],
4828 copts = xnnpack_optional_tflite_copts(),
4829 tags = ["nowin32"],
4830 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
4831)
4832
4833xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004834 name = "global_average_pooling_bench",
4835 srcs = ["bench/global-average-pooling.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004836 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07004837)
4838
4839xnnpack_benchmark(
Marat Dukhanad352602020-06-25 21:50:54 -07004840 name = "hardswish_bench",
4841 srcs = ["bench/hardswish.cc"],
4842 copts = xnnpack_optional_tflite_copts(),
4843 tags = ["nowin32"],
4844 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
4845)
4846
4847xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004848 name = "max_pooling_bench",
4849 srcs = ["bench/max-pooling.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004850 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07004851)
4852
4853xnnpack_benchmark(
4854 name = "sigmoid_bench",
4855 srcs = ["bench/sigmoid.cc"],
Marat Dukhanc3b9e862019-11-17 13:18:54 -08004856 copts = xnnpack_optional_tflite_copts(),
Marat Dukhanca2ba702020-04-24 01:31:47 -07004857 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004858 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004859)
4860
4861xnnpack_benchmark(
Marat Dukhan95b22432019-10-30 16:30:14 -07004862 name = "prelu_bench",
4863 srcs = ["bench/prelu.cc"],
4864 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07004865 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004866 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan95b22432019-10-30 16:30:14 -07004867)
4868
4869xnnpack_benchmark(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08004870 name = "softmax_bench",
4871 srcs = ["bench/softmax.cc"],
Marat Dukhan9c0db962020-01-28 12:30:14 -08004872 copts = xnnpack_optional_tflite_copts(),
Marat Dukhanca2ba702020-04-24 01:31:47 -07004873 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004874 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004875)
4876
Marat Dukhan87727142020-06-24 15:24:10 -07004877xnnpack_benchmark(
Marat Dukhan6804bbd2020-06-30 19:26:11 -07004878 name = "square_root_bench",
4879 srcs = ["bench/square-root.cc"],
4880 copts = xnnpack_optional_tflite_copts(),
4881 tags = ["nowin32"],
4882 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
4883)
4884
4885xnnpack_benchmark(
Marat Dukhan87727142020-06-24 15:24:10 -07004886 name = "truncation_bench",
4887 srcs = ["bench/truncation.cc"],
4888 deps = OPERATOR_BENCHMARK_DEPS,
4889)
4890
Marat Dukhanc068bb62019-10-04 13:24:39 -07004891############################# End-to-end benchmarks ############################
4892
4893cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07004894 name = "fp32_mobilenet_v1",
4895 srcs = ["models/fp32-mobilenet-v1.cc"],
Marat Dukhanc068bb62019-10-04 13:24:39 -07004896 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08004897 copts = xnnpack_std_cxxopts(),
Marat Dukhanc068bb62019-10-04 13:24:39 -07004898 linkstatic = True,
4899 deps = [
4900 ":XNNPACK",
4901 "@pthreadpool",
4902 ],
4903)
4904
4905cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07004906 name = "fp16_mobilenet_v1",
4907 srcs = ["models/fp16-mobilenet-v1.cc"],
4908 hdrs = ["models/models.h"],
4909 copts = xnnpack_std_cxxopts(),
4910 linkstatic = True,
4911 deps = [
4912 ":XNNPACK",
4913 "@FP16",
4914 "@pthreadpool",
4915 ],
4916)
4917
4918cc_library(
Marat Dukhan0743cdf2020-08-04 18:52:07 -07004919 name = "qs8_mobilenet_v1",
4920 srcs = ["models/qs8-mobilenet-v1.cc"],
4921 hdrs = ["models/models.h"],
4922 copts = xnnpack_std_cxxopts(),
4923 linkstatic = True,
4924 deps = [
4925 ":XNNPACK",
4926 "@pthreadpool",
4927 ],
4928)
4929
4930cc_library(
Marat Dukhan70a96182020-09-03 17:13:58 -07004931 name = "qs8_mobilenet_v2",
4932 srcs = ["models/qs8-mobilenet-v2.cc"],
4933 hdrs = ["models/models.h"],
4934 copts = xnnpack_std_cxxopts(),
4935 linkstatic = True,
4936 deps = [
4937 ":XNNPACK",
4938 "@pthreadpool",
4939 ],
4940)
4941
4942cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07004943 name = "fp32_mobilenet_v2",
4944 srcs = ["models/fp32-mobilenet-v2.cc"],
Marat Dukhanc068bb62019-10-04 13:24:39 -07004945 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08004946 copts = xnnpack_std_cxxopts(),
Marat Dukhanc068bb62019-10-04 13:24:39 -07004947 linkstatic = True,
4948 deps = [
4949 ":XNNPACK",
4950 "@pthreadpool",
4951 ],
4952)
4953
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004954cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07004955 name = "fp16_mobilenet_v2",
4956 srcs = ["models/fp16-mobilenet-v2.cc"],
4957 hdrs = ["models/models.h"],
4958 copts = xnnpack_std_cxxopts(),
4959 linkstatic = True,
4960 deps = [
4961 ":XNNPACK",
4962 "@FP16",
4963 "@pthreadpool",
4964 ],
4965)
4966
4967cc_library(
4968 name = "fp32_mobilenet_v3_large",
4969 srcs = ["models/fp32-mobilenet-v3-large.cc"],
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004970 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08004971 copts = xnnpack_std_cxxopts(),
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004972 linkstatic = True,
4973 deps = [
4974 ":XNNPACK",
4975 "@pthreadpool",
4976 ],
4977)
4978
4979cc_library(
Marat Dukhan66f3ccd2020-09-14 16:09:38 -07004980 name = "fp16_mobilenet_v3_large",
4981 srcs = ["models/fp16-mobilenet-v3-large.cc"],
4982 hdrs = ["models/models.h"],
4983 copts = xnnpack_std_cxxopts(),
4984 linkstatic = True,
4985 deps = [
4986 ":XNNPACK",
4987 "@FP16",
4988 "@pthreadpool",
4989 ],
4990)
4991
4992cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07004993 name = "fp32_mobilenet_v3_small",
4994 srcs = ["models/fp32-mobilenet-v3-small.cc"],
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004995 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08004996 copts = xnnpack_std_cxxopts(),
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004997 linkstatic = True,
4998 deps = [
4999 ":XNNPACK",
5000 "@pthreadpool",
5001 ],
5002)
5003
Marat Dukhan66f3ccd2020-09-14 16:09:38 -07005004cc_library(
5005 name = "fp16_mobilenet_v3_small",
5006 srcs = ["models/fp16-mobilenet-v3-small.cc"],
5007 hdrs = ["models/models.h"],
5008 copts = xnnpack_std_cxxopts(),
5009 linkstatic = True,
5010 deps = [
5011 ":XNNPACK",
5012 "@FP16",
5013 "@pthreadpool",
5014 ],
5015)
5016
Marat Dukhanc068bb62019-10-04 13:24:39 -07005017xnnpack_benchmark(
Marat Dukhanef4416e2019-10-31 13:44:40 -07005018 name = "f32_dwconv_e2e_bench",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08005019 srcs = [
5020 "bench/f32-dwconv-e2e.cc",
5021 "bench/end2end.h",
5022 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanef4416e2019-10-31 13:44:40 -07005023 deps = MICROKERNEL_BENCHMARK_DEPS + [
5024 ":XNNPACK",
Marat Dukhan270a2c42020-06-26 16:45:52 -07005025 ":fp32_mobilenet_v1",
5026 ":fp32_mobilenet_v2",
5027 ":fp32_mobilenet_v3_large",
5028 ":fp32_mobilenet_v3_small",
Marat Dukhanef4416e2019-10-31 13:44:40 -07005029 ],
5030)
5031
5032xnnpack_benchmark(
Marat Dukhan5f18d262019-10-31 10:24:14 -07005033 name = "f32_gemm_e2e_bench",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08005034 srcs = [
5035 "bench/f32-gemm-e2e.cc",
5036 "bench/end2end.h",
5037 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan5f18d262019-10-31 10:24:14 -07005038 deps = MICROKERNEL_BENCHMARK_DEPS + [
5039 ":XNNPACK",
Marat Dukhan270a2c42020-06-26 16:45:52 -07005040 ":fp32_mobilenet_v1",
5041 ":fp32_mobilenet_v2",
5042 ":fp32_mobilenet_v3_large",
5043 ":fp32_mobilenet_v3_small",
Marat Dukhan5f18d262019-10-31 10:24:14 -07005044 ],
5045)
5046
5047xnnpack_benchmark(
Marat Dukhanc068bb62019-10-04 13:24:39 -07005048 name = "end2end_bench",
5049 srcs = ["bench/end2end.cc"],
5050 deps = [
5051 ":XNNPACK",
Frank Barchardc712fa42019-10-31 14:00:21 -07005052 ":bench_utils",
Marat Dukhan270a2c42020-06-26 16:45:52 -07005053 ":fp16_mobilenet_v1",
5054 ":fp16_mobilenet_v2",
Marat Dukhan66f3ccd2020-09-14 16:09:38 -07005055 ":fp16_mobilenet_v3_large",
5056 ":fp16_mobilenet_v3_small",
Marat Dukhan270a2c42020-06-26 16:45:52 -07005057 ":fp32_mobilenet_v1",
5058 ":fp32_mobilenet_v2",
5059 ":fp32_mobilenet_v3_large",
5060 ":fp32_mobilenet_v3_small",
Marat Dukhan0743cdf2020-08-04 18:52:07 -07005061 ":qs8_mobilenet_v1",
Marat Dukhan70a96182020-09-03 17:13:58 -07005062 ":qs8_mobilenet_v2",
Marat Dukhanc068bb62019-10-04 13:24:39 -07005063 "@pthreadpool",
5064 ],
5065)
5066
Marat Dukhan6adff4e2019-10-14 18:32:07 -07005067#################### Accuracy evaluation for math functions ####################
5068
5069xnnpack_benchmark(
5070 name = "f32_exp_eval",
5071 srcs = [
5072 "eval/f32-exp.cc",
5073 "src/xnnpack/AlignedAllocator.h",
5074 ] + ACCURACY_EVAL_HDRS,
5075 deps = ACCURACY_EVAL_DEPS,
5076)
5077
Marat Dukhan515c9772019-10-17 18:07:57 -07005078xnnpack_benchmark(
5079 name = "f32_expminus_eval",
5080 srcs = [
5081 "eval/f32-expminus.cc",
5082 "src/xnnpack/AlignedAllocator.h",
5083 ] + ACCURACY_EVAL_HDRS,
5084 deps = ACCURACY_EVAL_DEPS,
5085)
5086
Marat Dukhan98ba4412019-10-23 02:14:28 -07005087xnnpack_benchmark(
Marat Dukhana438aca2020-11-20 15:45:01 -08005088 name = "f32_expm1minus_eval",
5089 srcs = [
5090 "eval/f32-expm1minus.cc",
5091 "src/xnnpack/AlignedAllocator.h",
5092 ] + ACCURACY_EVAL_HDRS,
5093 deps = ACCURACY_EVAL_DEPS,
5094)
5095
5096xnnpack_benchmark(
Marat Dukhan98ba4412019-10-23 02:14:28 -07005097 name = "f32_extexp_eval",
5098 srcs = [
5099 "eval/f32-extexp.cc",
5100 "src/xnnpack/AlignedAllocator.h",
5101 ] + ACCURACY_EVAL_HDRS,
5102 deps = ACCURACY_EVAL_DEPS,
5103)
5104
Marat Dukhan8853b822020-05-07 12:19:01 -07005105xnnpack_unit_test(
5106 name = "f32_roundne_eval",
5107 srcs = [
5108 "eval/f32-roundne.cc",
5109 "src/xnnpack/AlignedAllocator.h",
5110 "src/xnnpack/math-stubs.h",
5111 ] + MICROKERNEL_TEST_HDRS,
Marat Dukhan22eed3d2020-05-11 20:13:37 -07005112 automatic = False,
Marat Dukhan8853b822020-05-07 12:19:01 -07005113 deps = MICROKERNEL_TEST_DEPS,
5114)
5115
Marat Dukhan2dbb9442020-05-12 20:43:43 -07005116xnnpack_unit_test(
Marat Dukhanc9852ba2020-05-13 17:21:29 -07005117 name = "f32_roundd_eval",
5118 srcs = [
5119 "eval/f32-roundd.cc",
5120 "src/xnnpack/AlignedAllocator.h",
5121 "src/xnnpack/math-stubs.h",
5122 ] + MICROKERNEL_TEST_HDRS,
5123 automatic = False,
5124 deps = MICROKERNEL_TEST_DEPS,
5125)
5126
5127xnnpack_unit_test(
5128 name = "f32_roundu_eval",
5129 srcs = [
5130 "eval/f32-roundu.cc",
5131 "src/xnnpack/AlignedAllocator.h",
5132 "src/xnnpack/math-stubs.h",
5133 ] + MICROKERNEL_TEST_HDRS,
5134 automatic = False,
5135 deps = MICROKERNEL_TEST_DEPS,
5136)
5137
5138xnnpack_unit_test(
Marat Dukhan2dbb9442020-05-12 20:43:43 -07005139 name = "f32_roundz_eval",
5140 srcs = [
5141 "eval/f32-roundz.cc",
5142 "src/xnnpack/AlignedAllocator.h",
5143 "src/xnnpack/math-stubs.h",
5144 ] + MICROKERNEL_TEST_HDRS,
Marat Dukhanc9852ba2020-05-13 17:21:29 -07005145 automatic = False,
Marat Dukhan2dbb9442020-05-12 20:43:43 -07005146 deps = MICROKERNEL_TEST_DEPS,
5147)
5148
Marat Dukhan346a9e52019-11-15 09:06:30 -08005149xnnpack_benchmark(
5150 name = "f32_sigmoid_eval",
5151 srcs = [
5152 "eval/f32-sigmoid.cc",
5153 "src/xnnpack/AlignedAllocator.h",
5154 ] + ACCURACY_EVAL_HDRS,
5155 deps = ACCURACY_EVAL_DEPS,
5156)
5157
Marat Dukhan84000762020-06-29 18:38:43 -07005158xnnpack_benchmark(
5159 name = "f32_sqrt_eval",
5160 srcs = [
5161 "eval/f32-sqrt.cc",
5162 "src/xnnpack/AlignedAllocator.h",
5163 ] + ACCURACY_EVAL_HDRS,
5164 deps = ACCURACY_EVAL_DEPS + [
5165 ":bench_utils",
5166 ],
5167)
5168
Marat Dukhan08c4a432019-10-03 09:29:21 -07005169######################### Unit tests for micro-kernels #########################
5170
5171xnnpack_unit_test(
Marat Dukhande06f492020-04-09 00:19:31 -07005172 name = "f16_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005173 srcs = [
Marat Dukhande06f492020-04-09 00:19:31 -07005174 "test/f16-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005175 "test/gemm-microkernel-tester.h",
5176 "src/xnnpack/AlignedAllocator.h",
5177 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005178 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005179)
5180
5181xnnpack_unit_test(
Marat Dukhan355ab432020-04-09 19:01:52 -07005182 name = "f16_spmm_minmax_test",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08005183 srcs = [
Marat Dukhan355ab432020-04-09 19:01:52 -07005184 "test/f16-spmm-minmax.cc",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08005185 "test/spmm-microkernel-tester.h",
5186 "src/xnnpack/AlignedAllocator.h",
5187 ] + MICROKERNEL_TEST_HDRS,
5188 deps = MICROKERNEL_TEST_DEPS,
5189)
5190
5191xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005192 name = "f32_argmaxpool_test",
5193 srcs = [
5194 "test/f32-argmaxpool.cc",
5195 "test/argmaxpool-microkernel-tester.h",
5196 "src/xnnpack/AlignedAllocator.h",
5197 ] + MICROKERNEL_TEST_HDRS,
5198 deps = MICROKERNEL_TEST_DEPS,
5199)
5200
5201xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07005202 name = "f32_avgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005203 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07005204 "test/f32-avgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005205 "test/avgpool-microkernel-tester.h",
5206 "src/xnnpack/AlignedAllocator.h",
5207 ] + MICROKERNEL_TEST_HDRS,
5208 deps = MICROKERNEL_TEST_DEPS,
5209)
5210
5211xnnpack_unit_test(
Marat Dukhan660fd192020-03-10 04:55:30 -07005212 name = "f32_ibilinear_test",
Marat Dukhan35dacfb2019-11-07 19:18:16 -08005213 srcs = [
Marat Dukhan660fd192020-03-10 04:55:30 -07005214 "test/f32-ibilinear.cc",
5215 "test/ibilinear-microkernel-tester.h",
Marat Dukhan35dacfb2019-11-07 19:18:16 -08005216 "src/xnnpack/AlignedAllocator.h",
5217 ] + MICROKERNEL_TEST_HDRS,
5218 deps = MICROKERNEL_TEST_DEPS,
5219)
5220
5221xnnpack_unit_test(
XNNPACK Team21432672020-10-19 19:58:48 -07005222 name = "f32_ibilinear_chw_test",
5223 srcs = [
5224 "test/f32-ibilinear-chw.cc",
XNNPACK Team6be46b22020-10-22 23:34:54 -07005225 "test/ibilinear-microkernel-tester.h",
XNNPACK Team21432672020-10-19 19:58:48 -07005226 "src/xnnpack/AlignedAllocator.h",
5227 ] + MICROKERNEL_TEST_HDRS,
5228 deps = MICROKERNEL_TEST_DEPS,
5229)
5230
5231xnnpack_unit_test(
Frank Barchardb1966592020-05-12 13:47:06 -07005232 name = "f16_clamp_test",
5233 srcs = [
5234 "test/f16-clamp.cc",
5235 "test/clamp-microkernel-tester.h",
5236 ] + MICROKERNEL_TEST_HDRS,
5237 deps = MICROKERNEL_TEST_DEPS,
5238)
5239
5240xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005241 name = "f32_clamp_test",
5242 srcs = [
5243 "test/f32-clamp.cc",
5244 "test/clamp-microkernel-tester.h",
5245 ] + MICROKERNEL_TEST_HDRS,
5246 deps = MICROKERNEL_TEST_DEPS,
5247)
5248
5249xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07005250 name = "f32_igemm_test",
5251 srcs = [
5252 "test/f32-igemm.cc",
5253 "test/gemm-microkernel-tester.h",
5254 "src/xnnpack/AlignedAllocator.h",
5255 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005256 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan163a7e62020-04-09 04:19:26 -07005257)
5258
5259xnnpack_unit_test(
Marat Dukhan467f6362020-05-22 23:21:55 -07005260 name = "f32_igemm_relu_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005261 srcs = [
Marat Dukhan467f6362020-05-22 23:21:55 -07005262 "test/f32-igemm-relu.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005263 "test/gemm-microkernel-tester.h",
5264 "src/xnnpack/AlignedAllocator.h",
5265 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005266 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005267)
5268
5269xnnpack_unit_test(
Marat Dukhane207b7b2020-05-28 16:27:42 -07005270 name = "f32_igemm_minmax_test",
5271 srcs = [
5272 "test/f32-igemm-minmax.cc",
5273 "test/gemm-microkernel-tester.h",
5274 "src/xnnpack/AlignedAllocator.h",
5275 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005276 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhane207b7b2020-05-28 16:27:42 -07005277)
5278
5279xnnpack_unit_test(
Frank Barchardb0e4fae2020-05-04 15:27:51 -07005280 name = "f16_igemm_minmax_test",
5281 srcs = [
5282 "test/f16-igemm-minmax.cc",
5283 "test/gemm-microkernel-tester.h",
5284 "src/xnnpack/AlignedAllocator.h",
5285 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005286 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Frank Barchardb0e4fae2020-05-04 15:27:51 -07005287)
5288
5289xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005290 name = "f32_conv_hwc_test",
5291 srcs = [
5292 "test/f32-conv-hwc.cc",
5293 "test/conv-hwc-microkernel-tester.h",
5294 "src/xnnpack/AlignedAllocator.h",
5295 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005296 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005297)
5298
5299xnnpack_unit_test(
Marat Dukhan1f29b802020-05-15 23:46:39 -07005300 name = "f32_conv_hwc2chw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005301 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07005302 "test/f32-conv-hwc2chw.cc",
5303 "test/conv-hwc2chw-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005304 "src/xnnpack/AlignedAllocator.h",
5305 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005306 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005307)
5308
5309xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07005310 name = "f32_dwconv_test",
5311 srcs = [
5312 "test/f32-dwconv.cc",
5313 "test/dwconv-microkernel-tester.h",
5314 "src/xnnpack/AlignedAllocator.h",
5315 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005316 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan163a7e62020-04-09 04:19:26 -07005317)
5318
5319xnnpack_unit_test(
Frank Barchard5a599a62020-06-04 20:12:44 -07005320 name = "f16_dwconv_minmax_test",
5321 srcs = [
5322 "test/f16-dwconv-minmax.cc",
5323 "test/dwconv-microkernel-tester.h",
5324 "src/xnnpack/AlignedAllocator.h",
5325 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005326 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Frank Barchard5a599a62020-06-04 20:12:44 -07005327)
5328
5329xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07005330 name = "f32_dwconv_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005331 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07005332 "test/f32-dwconv-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005333 "test/dwconv-microkernel-tester.h",
5334 "src/xnnpack/AlignedAllocator.h",
5335 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005336 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005337)
5338
5339xnnpack_unit_test(
Marat Dukhanbf715f92020-10-23 20:17:00 -07005340 name = "f32_dwconv2d_chw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005341 srcs = [
Marat Dukhanbf715f92020-10-23 20:17:00 -07005342 "test/f32-dwconv2d-chw.cc",
5343 "test/dwconv2d-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005344 "src/xnnpack/AlignedAllocator.h",
5345 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005346 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005347)
5348
5349xnnpack_unit_test(
Frank Barchard0bb49a72020-06-04 11:35:11 -07005350 name = "f16_gavgpool_minmax_test",
5351 srcs = [
5352 "test/f16-gavgpool-minmax.cc",
5353 "test/gavgpool-microkernel-tester.h",
5354 "src/xnnpack/AlignedAllocator.h",
5355 ] + MICROKERNEL_TEST_HDRS,
5356 deps = MICROKERNEL_TEST_DEPS,
5357)
5358
5359xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07005360 name = "f32_gavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005361 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07005362 "test/f32-gavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005363 "test/gavgpool-microkernel-tester.h",
5364 "src/xnnpack/AlignedAllocator.h",
5365 ] + MICROKERNEL_TEST_HDRS,
5366 deps = MICROKERNEL_TEST_DEPS,
5367)
5368
5369xnnpack_unit_test(
Marat Dukhan1f29b802020-05-15 23:46:39 -07005370 name = "f32_gavgpool_cw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005371 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07005372 "test/f32-gavgpool-cw.cc",
5373 "test/gavgpool-cw-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005374 "src/xnnpack/AlignedAllocator.h",
5375 ] + MICROKERNEL_TEST_HDRS,
5376 deps = MICROKERNEL_TEST_DEPS,
5377)
5378
5379xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07005380 name = "f32_gemm_test",
5381 srcs = [
5382 "test/f32-gemm.cc",
5383 "test/gemm-microkernel-tester.h",
5384 "src/xnnpack/AlignedAllocator.h",
5385 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005386 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan163a7e62020-04-09 04:19:26 -07005387)
5388
5389xnnpack_unit_test(
Marat Dukhan467f6362020-05-22 23:21:55 -07005390 name = "f32_gemm_relu_test",
5391 srcs = [
5392 "test/f32-gemm-relu.cc",
5393 "test/gemm-microkernel-tester.h",
5394 "src/xnnpack/AlignedAllocator.h",
5395 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005396 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan467f6362020-05-22 23:21:55 -07005397)
5398
5399xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07005400 name = "f32_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005401 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07005402 "test/f32-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005403 "test/gemm-microkernel-tester.h",
5404 "src/xnnpack/AlignedAllocator.h",
5405 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005406 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005407)
5408
5409xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07005410 name = "f32_gemminc_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005411 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07005412 "test/f32-gemminc-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005413 "test/gemm-microkernel-tester.h",
5414 "src/xnnpack/AlignedAllocator.h",
5415 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005416 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005417)
5418
5419xnnpack_unit_test(
Frank Barchardb1966592020-05-12 13:47:06 -07005420 name = "f16_hswish_test",
5421 srcs = [
5422 "test/f16-hswish.cc",
5423 "test/hswish-microkernel-tester.h",
5424 ] + MICROKERNEL_TEST_HDRS,
5425 deps = MICROKERNEL_TEST_DEPS,
5426)
5427
5428xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005429 name = "f32_hswish_test",
5430 srcs = [
5431 "test/f32-hswish.cc",
5432 "test/hswish-microkernel-tester.h",
5433 ] + MICROKERNEL_TEST_HDRS,
5434 deps = MICROKERNEL_TEST_DEPS,
5435)
5436
5437xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07005438 name = "f32_maxpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005439 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07005440 "test/f32-maxpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005441 "test/maxpool-microkernel-tester.h",
5442 ] + MICROKERNEL_TEST_HDRS,
5443 deps = MICROKERNEL_TEST_DEPS,
5444)
5445
5446xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07005447 name = "f32_pavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005448 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07005449 "test/f32-pavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005450 "test/avgpool-microkernel-tester.h",
5451 "src/xnnpack/AlignedAllocator.h",
5452 ] + MICROKERNEL_TEST_HDRS,
5453 deps = MICROKERNEL_TEST_DEPS,
5454)
5455
5456xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07005457 name = "f32_ppmm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005458 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07005459 "test/f32-ppmm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005460 "test/gemm-microkernel-tester.h",
5461 "src/xnnpack/AlignedAllocator.h",
5462 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005463 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005464)
5465
5466xnnpack_unit_test(
Frank Barchardb1966592020-05-12 13:47:06 -07005467 name = "f16_prelu_test",
5468 srcs = [
5469 "test/f16-prelu.cc",
5470 "test/prelu-microkernel-tester.h",
5471 "src/xnnpack/AlignedAllocator.h",
5472 ] + MICROKERNEL_TEST_HDRS,
5473 deps = MICROKERNEL_TEST_DEPS,
5474)
5475
5476xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005477 name = "f32_prelu_test",
5478 srcs = [
5479 "test/f32-prelu.cc",
5480 "test/prelu-microkernel-tester.h",
5481 "src/xnnpack/AlignedAllocator.h",
5482 ] + MICROKERNEL_TEST_HDRS,
5483 deps = MICROKERNEL_TEST_DEPS,
5484)
5485
5486xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07005487 name = "f32_raddexpminusmax_test",
5488 srcs = [
5489 "test/f32-raddexpminusmax.cc",
5490 "test/raddexpminusmax-microkernel-tester.h",
5491 ] + MICROKERNEL_TEST_HDRS,
5492 deps = MICROKERNEL_TEST_DEPS,
5493)
5494
5495xnnpack_unit_test(
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07005496 name = "f32_raddextexp_test",
5497 srcs = [
5498 "test/f32-raddextexp.cc",
5499 "test/raddextexp-microkernel-tester.h",
5500 ] + MICROKERNEL_TEST_HDRS,
5501 deps = MICROKERNEL_TEST_DEPS,
5502)
5503
5504xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07005505 name = "f32_raddstoreexpminusmax_test",
5506 srcs = [
5507 "test/f32-raddstoreexpminusmax.cc",
5508 "test/raddstoreexpminusmax-microkernel-tester.h",
5509 ] + MICROKERNEL_TEST_HDRS,
5510 deps = MICROKERNEL_TEST_DEPS,
5511)
5512
5513xnnpack_unit_test(
Frank Barchardfb158e22020-07-15 16:10:10 -07005514 name = "f32_relu_test",
5515 srcs = [
5516 "test/f32-relu.cc",
5517 "test/vunary-microkernel-tester.h",
5518 ] + MICROKERNEL_TEST_HDRS,
5519 deps = MICROKERNEL_TEST_DEPS,
5520)
5521
5522xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005523 name = "f32_rmax_test",
5524 srcs = [
5525 "test/f32-rmax.cc",
5526 "test/rmax-microkernel-tester.h",
5527 ] + MICROKERNEL_TEST_HDRS,
5528 deps = MICROKERNEL_TEST_DEPS,
5529)
5530
5531xnnpack_unit_test(
Marat Dukhan346a9e52019-11-15 09:06:30 -08005532 name = "f32_sigmoid_test",
5533 srcs = [
5534 "test/f32-sigmoid.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005535 "test/vunary-microkernel-tester.h",
Marat Dukhan346a9e52019-11-15 09:06:30 -08005536 ] + MICROKERNEL_TEST_HDRS,
5537 deps = MICROKERNEL_TEST_DEPS,
5538)
5539
5540xnnpack_unit_test(
Marat Dukhan355ab432020-04-09 19:01:52 -07005541 name = "f32_spmm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005542 srcs = [
Marat Dukhan355ab432020-04-09 19:01:52 -07005543 "test/f32-spmm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005544 "test/spmm-microkernel-tester.h",
5545 "src/xnnpack/AlignedAllocator.h",
5546 ] + MICROKERNEL_TEST_HDRS,
5547 deps = MICROKERNEL_TEST_DEPS,
5548)
5549
5550xnnpack_unit_test(
Frank Barchardd793f6c2020-05-08 13:37:43 -07005551 name = "f16_vadd_minmax_test",
5552 srcs = [
5553 "test/f16-vadd-minmax.cc",
5554 "test/vbinary-microkernel-tester.h",
5555 ] + MICROKERNEL_TEST_HDRS,
5556 deps = MICROKERNEL_TEST_DEPS,
5557)
5558
5559xnnpack_unit_test(
5560 name = "f16_vaddc_minmax_test",
5561 srcs = [
5562 "test/f16-vaddc-minmax.cc",
5563 "test/vbinaryc-microkernel-tester.h",
5564 ] + MICROKERNEL_TEST_HDRS,
5565 deps = MICROKERNEL_TEST_DEPS,
5566)
5567
5568xnnpack_unit_test(
5569 name = "f16_vdiv_minmax_test",
5570 srcs = [
5571 "test/f16-vdiv-minmax.cc",
5572 "test/vbinary-microkernel-tester.h",
5573 ] + MICROKERNEL_TEST_HDRS,
5574 deps = MICROKERNEL_TEST_DEPS,
5575)
5576
5577xnnpack_unit_test(
5578 name = "f16_vdivc_minmax_test",
5579 srcs = [
5580 "test/f16-vdivc-minmax.cc",
5581 "test/vbinaryc-microkernel-tester.h",
5582 ] + MICROKERNEL_TEST_HDRS,
5583 deps = MICROKERNEL_TEST_DEPS,
5584)
5585
5586xnnpack_unit_test(
5587 name = "f16_vrdivc_minmax_test",
5588 srcs = [
5589 "test/f16-vrdivc-minmax.cc",
5590 "test/vbinaryc-microkernel-tester.h",
5591 ] + MICROKERNEL_TEST_HDRS,
5592 deps = MICROKERNEL_TEST_DEPS,
5593)
5594
5595xnnpack_unit_test(
5596 name = "f16_vmax_test",
5597 srcs = [
5598 "test/f16-vmax.cc",
5599 "test/vbinary-microkernel-tester.h",
5600 ] + MICROKERNEL_TEST_HDRS,
5601 deps = MICROKERNEL_TEST_DEPS,
5602)
5603
5604xnnpack_unit_test(
5605 name = "f16_vmaxc_test",
5606 srcs = [
5607 "test/f16-vmaxc.cc",
5608 "test/vbinaryc-microkernel-tester.h",
5609 ] + MICROKERNEL_TEST_HDRS,
5610 deps = MICROKERNEL_TEST_DEPS,
5611)
5612
5613xnnpack_unit_test(
5614 name = "f16_vmin_test",
5615 srcs = [
5616 "test/f16-vmin.cc",
5617 "test/vbinary-microkernel-tester.h",
5618 ] + MICROKERNEL_TEST_HDRS,
5619 deps = MICROKERNEL_TEST_DEPS,
5620)
5621
5622xnnpack_unit_test(
5623 name = "f16_vminc_test",
5624 srcs = [
5625 "test/f16-vminc.cc",
5626 "test/vbinaryc-microkernel-tester.h",
5627 ] + MICROKERNEL_TEST_HDRS,
5628 deps = MICROKERNEL_TEST_DEPS,
5629)
5630
5631xnnpack_unit_test(
5632 name = "f16_vmul_minmax_test",
5633 srcs = [
5634 "test/f16-vmul-minmax.cc",
5635 "test/vbinary-microkernel-tester.h",
5636 ] + MICROKERNEL_TEST_HDRS,
5637 deps = MICROKERNEL_TEST_DEPS,
5638)
5639
5640xnnpack_unit_test(
5641 name = "f16_vmulc_minmax_test",
5642 srcs = [
5643 "test/f16-vmulc-minmax.cc",
5644 "test/vbinaryc-microkernel-tester.h",
5645 ] + MICROKERNEL_TEST_HDRS,
5646 deps = MICROKERNEL_TEST_DEPS,
5647)
5648
5649xnnpack_unit_test(
5650 name = "f16_vsub_minmax_test",
5651 srcs = [
5652 "test/f16-vsub-minmax.cc",
5653 "test/vbinary-microkernel-tester.h",
5654 ] + MICROKERNEL_TEST_HDRS,
5655 deps = MICROKERNEL_TEST_DEPS,
5656)
5657
5658xnnpack_unit_test(
5659 name = "f16_vsubc_minmax_test",
5660 srcs = [
5661 "test/f16-vsubc-minmax.cc",
5662 "test/vbinaryc-microkernel-tester.h",
5663 ] + MICROKERNEL_TEST_HDRS,
5664 deps = MICROKERNEL_TEST_DEPS,
5665)
5666
5667xnnpack_unit_test(
5668 name = "f16_vrsubc_minmax_test",
5669 srcs = [
5670 "test/f16-vrsubc-minmax.cc",
5671 "test/vbinaryc-microkernel-tester.h",
5672 ] + MICROKERNEL_TEST_HDRS,
5673 deps = MICROKERNEL_TEST_DEPS,
5674)
5675
5676xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07005677 name = "f32_vabs_test",
5678 srcs = [
5679 "test/f32-vabs.cc",
5680 "test/vunary-microkernel-tester.h",
5681 ] + MICROKERNEL_TEST_HDRS,
5682 deps = MICROKERNEL_TEST_DEPS,
5683)
5684
5685xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005686 name = "f32_vadd_test",
5687 srcs = [
5688 "test/f32-vadd.cc",
5689 "test/vbinary-microkernel-tester.h",
5690 ] + MICROKERNEL_TEST_HDRS,
5691 deps = MICROKERNEL_TEST_DEPS,
5692)
5693
5694xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005695 name = "f32_vadd_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005696 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005697 "test/f32-vadd-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005698 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005699 ] + MICROKERNEL_TEST_HDRS,
5700 deps = MICROKERNEL_TEST_DEPS,
5701)
5702
5703xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005704 name = "f32_vadd_relu_test",
5705 srcs = [
5706 "test/f32-vadd-relu.cc",
5707 "test/vbinary-microkernel-tester.h",
5708 ] + MICROKERNEL_TEST_HDRS,
5709 deps = MICROKERNEL_TEST_DEPS,
5710)
5711
5712xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005713 name = "f32_vaddc_test",
5714 srcs = [
5715 "test/f32-vaddc.cc",
5716 "test/vbinaryc-microkernel-tester.h",
5717 ] + MICROKERNEL_TEST_HDRS,
5718 deps = MICROKERNEL_TEST_DEPS,
5719)
5720
5721xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005722 name = "f32_vaddc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005723 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005724 "test/f32-vaddc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005725 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005726 ] + MICROKERNEL_TEST_HDRS,
5727 deps = MICROKERNEL_TEST_DEPS,
5728)
5729
5730xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005731 name = "f32_vaddc_relu_test",
5732 srcs = [
5733 "test/f32-vaddc-relu.cc",
5734 "test/vbinaryc-microkernel-tester.h",
5735 ] + MICROKERNEL_TEST_HDRS,
5736 deps = MICROKERNEL_TEST_DEPS,
5737)
5738
5739xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005740 name = "f32_vdiv_test",
5741 srcs = [
5742 "test/f32-vdiv.cc",
5743 "test/vbinary-microkernel-tester.h",
5744 ] + MICROKERNEL_TEST_HDRS,
5745 deps = MICROKERNEL_TEST_DEPS,
5746)
5747
5748xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005749 name = "f32_vdiv_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08005750 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005751 "test/f32-vdiv-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08005752 "test/vbinary-microkernel-tester.h",
5753 ] + MICROKERNEL_TEST_HDRS,
5754 deps = MICROKERNEL_TEST_DEPS,
5755)
5756
5757xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005758 name = "f32_vdiv_relu_test",
5759 srcs = [
5760 "test/f32-vdiv-relu.cc",
5761 "test/vbinary-microkernel-tester.h",
5762 ] + MICROKERNEL_TEST_HDRS,
5763 deps = MICROKERNEL_TEST_DEPS,
5764)
5765
5766xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005767 name = "f32_vdivc_test",
5768 srcs = [
5769 "test/f32-vdivc.cc",
5770 "test/vbinaryc-microkernel-tester.h",
5771 ] + MICROKERNEL_TEST_HDRS,
5772 deps = MICROKERNEL_TEST_DEPS,
5773)
5774
5775xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005776 name = "f32_vdivc_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08005777 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005778 "test/f32-vdivc-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08005779 "test/vbinaryc-microkernel-tester.h",
5780 ] + MICROKERNEL_TEST_HDRS,
5781 deps = MICROKERNEL_TEST_DEPS,
5782)
5783
5784xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005785 name = "f32_vdivc_relu_test",
5786 srcs = [
5787 "test/f32-vdivc-relu.cc",
5788 "test/vbinaryc-microkernel-tester.h",
5789 ] + MICROKERNEL_TEST_HDRS,
5790 deps = MICROKERNEL_TEST_DEPS,
5791)
5792
5793xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005794 name = "f32_vrdivc_test",
5795 srcs = [
5796 "test/f32-vrdivc.cc",
5797 "test/vbinaryc-microkernel-tester.h",
5798 ] + MICROKERNEL_TEST_HDRS,
5799 deps = MICROKERNEL_TEST_DEPS,
5800)
5801
5802xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005803 name = "f32_vrdivc_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08005804 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005805 "test/f32-vrdivc-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08005806 "test/vbinaryc-microkernel-tester.h",
5807 ] + MICROKERNEL_TEST_HDRS,
5808 deps = MICROKERNEL_TEST_DEPS,
5809)
5810
5811xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005812 name = "f32_vrdivc_relu_test",
5813 srcs = [
5814 "test/f32-vrdivc-relu.cc",
5815 "test/vbinaryc-microkernel-tester.h",
5816 ] + MICROKERNEL_TEST_HDRS,
5817 deps = MICROKERNEL_TEST_DEPS,
5818)
5819
5820xnnpack_unit_test(
Marat Dukhan403b7d42019-12-05 12:49:11 -08005821 name = "f32_vmax_test",
5822 srcs = [
5823 "test/f32-vmax.cc",
5824 "test/vbinary-microkernel-tester.h",
5825 ] + MICROKERNEL_TEST_HDRS,
5826 deps = MICROKERNEL_TEST_DEPS,
5827)
5828
5829xnnpack_unit_test(
5830 name = "f32_vmaxc_test",
5831 srcs = [
5832 "test/f32-vmaxc.cc",
5833 "test/vbinaryc-microkernel-tester.h",
5834 ] + MICROKERNEL_TEST_HDRS,
5835 deps = MICROKERNEL_TEST_DEPS,
5836)
5837
5838xnnpack_unit_test(
5839 name = "f32_vmin_test",
5840 srcs = [
5841 "test/f32-vmin.cc",
5842 "test/vbinary-microkernel-tester.h",
5843 ] + MICROKERNEL_TEST_HDRS,
5844 deps = MICROKERNEL_TEST_DEPS,
5845)
5846
5847xnnpack_unit_test(
5848 name = "f32_vminc_test",
5849 srcs = [
5850 "test/f32-vminc.cc",
5851 "test/vbinaryc-microkernel-tester.h",
5852 ] + MICROKERNEL_TEST_HDRS,
5853 deps = MICROKERNEL_TEST_DEPS,
5854)
5855
5856xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005857 name = "f32_vmul_test",
5858 srcs = [
5859 "test/f32-vmul.cc",
5860 "test/vbinary-microkernel-tester.h",
5861 ] + MICROKERNEL_TEST_HDRS,
5862 deps = MICROKERNEL_TEST_DEPS,
5863)
5864
5865xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005866 name = "f32_vmul_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005867 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005868 "test/f32-vmul-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005869 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005870 ] + MICROKERNEL_TEST_HDRS,
5871 deps = MICROKERNEL_TEST_DEPS,
5872)
5873
5874xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005875 name = "f32_vmul_relu_test",
5876 srcs = [
5877 "test/f32-vmul-relu.cc",
5878 "test/vbinary-microkernel-tester.h",
5879 ] + MICROKERNEL_TEST_HDRS,
5880 deps = MICROKERNEL_TEST_DEPS,
5881)
5882
5883xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005884 name = "f32_vmulc_test",
5885 srcs = [
5886 "test/f32-vmulc.cc",
5887 "test/vbinaryc-microkernel-tester.h",
5888 ] + MICROKERNEL_TEST_HDRS,
5889 deps = MICROKERNEL_TEST_DEPS,
5890)
5891
5892xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005893 name = "f32_vmulc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005894 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005895 "test/f32-vmulc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005896 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005897 ] + MICROKERNEL_TEST_HDRS,
5898 deps = MICROKERNEL_TEST_DEPS,
5899)
5900
5901xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005902 name = "f32_vmulc_relu_test",
5903 srcs = [
5904 "test/f32-vmulc-relu.cc",
5905 "test/vbinaryc-microkernel-tester.h",
5906 ] + MICROKERNEL_TEST_HDRS,
5907 deps = MICROKERNEL_TEST_DEPS,
5908)
5909
5910xnnpack_unit_test(
Frank Barchard2a1049c2020-06-03 02:31:27 -07005911 name = "f16_vmulcaddc_minmax_test",
5912 srcs = [
5913 "test/f16-vmulcaddc-minmax.cc",
5914 "test/vmulcaddc-microkernel-tester.h",
5915 "src/xnnpack/AlignedAllocator.h",
5916 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005917 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Frank Barchard2a1049c2020-06-03 02:31:27 -07005918)
5919
5920xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07005921 name = "f32_vmulcaddc_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005922 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07005923 "test/f32-vmulcaddc-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005924 "test/vmulcaddc-microkernel-tester.h",
5925 "src/xnnpack/AlignedAllocator.h",
5926 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005927 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005928)
5929
5930xnnpack_unit_test(
Marat Dukhan8cc7efe2020-06-10 16:24:27 -07005931 name = "f32_vlrelu_test",
5932 srcs = [
5933 "test/f32-vlrelu.cc",
5934 "test/vunary-microkernel-tester.h",
5935 ] + MICROKERNEL_TEST_HDRS,
5936 deps = MICROKERNEL_TEST_DEPS,
5937)
5938
5939xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07005940 name = "f32_vneg_test",
5941 srcs = [
5942 "test/f32-vneg.cc",
5943 "test/vunary-microkernel-tester.h",
5944 ] + MICROKERNEL_TEST_HDRS,
5945 deps = MICROKERNEL_TEST_DEPS,
5946)
5947
5948xnnpack_unit_test(
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07005949 name = "f32_vrndne_test",
5950 srcs = [
5951 "test/f32-vrndne.cc",
5952 "test/vunary-microkernel-tester.h",
5953 ] + MICROKERNEL_TEST_HDRS,
5954 deps = MICROKERNEL_TEST_DEPS,
5955)
5956
5957xnnpack_unit_test(
5958 name = "f32_vrndz_test",
5959 srcs = [
5960 "test/f32-vrndz.cc",
5961 "test/vunary-microkernel-tester.h",
5962 ] + MICROKERNEL_TEST_HDRS,
5963 deps = MICROKERNEL_TEST_DEPS,
5964)
5965
5966xnnpack_unit_test(
5967 name = "f32_vrndu_test",
5968 srcs = [
5969 "test/f32-vrndu.cc",
5970 "test/vunary-microkernel-tester.h",
5971 ] + MICROKERNEL_TEST_HDRS,
5972 deps = MICROKERNEL_TEST_DEPS,
5973)
5974
5975xnnpack_unit_test(
5976 name = "f32_vrndd_test",
5977 srcs = [
5978 "test/f32-vrndd.cc",
5979 "test/vunary-microkernel-tester.h",
5980 ] + MICROKERNEL_TEST_HDRS,
5981 deps = MICROKERNEL_TEST_DEPS,
5982)
5983
5984xnnpack_unit_test(
Marat Dukhan05ac8e32019-10-21 15:39:33 -07005985 name = "f32_vscale_test",
5986 srcs = [
5987 "test/f32-vscale.cc",
5988 "test/vscale-microkernel-tester.h",
5989 ] + MICROKERNEL_TEST_HDRS,
5990 deps = MICROKERNEL_TEST_DEPS,
5991)
5992
5993xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07005994 name = "f32_vscaleexpminusmax_test",
5995 srcs = [
5996 "test/f32-vscaleexpminusmax.cc",
5997 "test/vscaleexpminusmax-microkernel-tester.h",
5998 ] + MICROKERNEL_TEST_HDRS,
5999 deps = MICROKERNEL_TEST_DEPS,
6000)
6001
6002xnnpack_unit_test(
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07006003 name = "f32_vscaleextexp_test",
6004 srcs = [
6005 "test/f32-vscaleextexp.cc",
6006 "test/vscaleextexp-microkernel-tester.h",
6007 ] + MICROKERNEL_TEST_HDRS,
6008 deps = MICROKERNEL_TEST_DEPS,
6009)
6010
6011xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07006012 name = "f32_vsqr_test",
6013 srcs = [
6014 "test/f32-vsqr.cc",
6015 "test/vunary-microkernel-tester.h",
6016 ] + MICROKERNEL_TEST_HDRS,
6017 deps = MICROKERNEL_TEST_DEPS,
6018)
6019
6020xnnpack_unit_test(
Marat Dukhan13bafb02020-06-05 00:43:11 -07006021 name = "f32_vsqrdiff_test",
6022 srcs = [
6023 "test/f32-vsqrdiff.cc",
6024 "test/vbinary-microkernel-tester.h",
6025 ] + MICROKERNEL_TEST_HDRS,
6026 deps = MICROKERNEL_TEST_DEPS,
6027)
6028
6029xnnpack_unit_test(
6030 name = "f32_vsqrdiffc_test",
6031 srcs = [
6032 "test/f32-vsqrdiffc.cc",
6033 "test/vbinaryc-microkernel-tester.h",
6034 ] + MICROKERNEL_TEST_HDRS,
6035 deps = MICROKERNEL_TEST_DEPS,
6036)
6037
6038xnnpack_unit_test(
Marat Dukhanf4db2f32020-06-30 10:55:30 -07006039 name = "f32_vsqrt_test",
6040 srcs = [
6041 "test/f32-vsqrt.cc",
6042 "test/vunary-microkernel-tester.h",
6043 ] + MICROKERNEL_TEST_HDRS,
6044 deps = MICROKERNEL_TEST_DEPS,
6045)
6046
6047xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07006048 name = "f32_vsub_test",
6049 srcs = [
6050 "test/f32-vsub.cc",
6051 "test/vbinary-microkernel-tester.h",
6052 ] + MICROKERNEL_TEST_HDRS,
6053 deps = MICROKERNEL_TEST_DEPS,
6054)
6055
6056xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006057 name = "f32_vsub_minmax_test",
Marat Dukhan97579532019-10-18 16:40:39 -07006058 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006059 "test/f32-vsub-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08006060 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08006061 ] + MICROKERNEL_TEST_HDRS,
6062 deps = MICROKERNEL_TEST_DEPS,
6063)
6064
6065xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07006066 name = "f32_vsub_relu_test",
6067 srcs = [
6068 "test/f32-vsub-relu.cc",
6069 "test/vbinary-microkernel-tester.h",
6070 ] + MICROKERNEL_TEST_HDRS,
6071 deps = MICROKERNEL_TEST_DEPS,
6072)
6073
6074xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07006075 name = "f32_vsubc_test",
6076 srcs = [
6077 "test/f32-vsubc.cc",
6078 "test/vbinaryc-microkernel-tester.h",
6079 ] + MICROKERNEL_TEST_HDRS,
6080 deps = MICROKERNEL_TEST_DEPS,
6081)
6082
6083xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006084 name = "f32_vsubc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08006085 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006086 "test/f32-vsubc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08006087 "test/vbinaryc-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08006088 ] + MICROKERNEL_TEST_HDRS,
6089 deps = MICROKERNEL_TEST_DEPS,
6090)
6091
6092xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07006093 name = "f32_vsubc_relu_test",
6094 srcs = [
6095 "test/f32-vsubc-relu.cc",
6096 "test/vbinaryc-microkernel-tester.h",
6097 ] + MICROKERNEL_TEST_HDRS,
6098 deps = MICROKERNEL_TEST_DEPS,
6099)
6100
6101xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07006102 name = "f32_vrsubc_test",
6103 srcs = [
6104 "test/f32-vrsubc.cc",
6105 "test/vbinaryc-microkernel-tester.h",
6106 ] + MICROKERNEL_TEST_HDRS,
6107 deps = MICROKERNEL_TEST_DEPS,
6108)
6109
6110xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006111 name = "f32_vrsubc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08006112 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006113 "test/f32-vrsubc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08006114 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan97579532019-10-18 16:40:39 -07006115 ] + MICROKERNEL_TEST_HDRS,
6116 deps = MICROKERNEL_TEST_DEPS,
6117)
6118
6119xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07006120 name = "f32_vrsubc_relu_test",
6121 srcs = [
6122 "test/f32-vrsubc-relu.cc",
6123 "test/vbinaryc-microkernel-tester.h",
6124 ] + MICROKERNEL_TEST_HDRS,
6125 deps = MICROKERNEL_TEST_DEPS,
6126)
6127
6128xnnpack_unit_test(
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07006129 name = "qs8_dwconv_minmax_test",
6130 srcs = [
6131 "test/qs8-dwconv-minmax.cc",
6132 "test/dwconv-microkernel-tester.h",
6133 "src/xnnpack/AlignedAllocator.h",
6134 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
6135 deps = MICROKERNEL_TEST_DEPS + [":packing"],
6136)
6137
6138xnnpack_unit_test(
Marat Dukhan4ed53f42020-08-06 01:12:55 -07006139 name = "qs8_gavgpool_minmax_test",
6140 srcs = [
6141 "test/qs8-gavgpool-minmax.cc",
6142 "test/gavgpool-microkernel-tester.h",
6143 "src/xnnpack/AlignedAllocator.h",
6144 ] + MICROKERNEL_TEST_HDRS,
6145 deps = MICROKERNEL_TEST_DEPS,
6146)
6147
6148xnnpack_unit_test(
Marat Dukhan595e1702020-07-31 10:12:52 -07006149 name = "qs8_gemm_minmax_test",
6150 srcs = [
6151 "test/qs8-gemm-minmax.cc",
6152 "test/gemm-microkernel-tester.h",
6153 "src/xnnpack/AlignedAllocator.h",
6154 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
6155 deps = MICROKERNEL_TEST_DEPS + [":packing"],
6156)
6157
6158xnnpack_unit_test(
Marat Dukhanf9480682020-07-31 14:50:24 -07006159 name = "qs8_igemm_minmax_test",
6160 srcs = [
6161 "test/qs8-igemm-minmax.cc",
6162 "test/gemm-microkernel-tester.h",
6163 "src/xnnpack/AlignedAllocator.h",
6164 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
6165 deps = MICROKERNEL_TEST_DEPS + [":packing"],
6166)
6167
6168xnnpack_unit_test(
6169 name = "qs8_requantization_test",
6170 srcs = [
6171 "src/xnnpack/requantization-stubs.h",
6172 "test/qs8-requantization.cc",
6173 "test/requantization-tester.h",
6174 ] + MICROKERNEL_TEST_HDRS,
6175 deps = MICROKERNEL_TEST_DEPS,
6176)
6177
6178xnnpack_unit_test(
Marat Dukhand9f3ad42020-08-10 12:30:58 -07006179 name = "qs8_vadd_minmax_test",
6180 srcs = [
6181 "test/qs8-vadd-minmax.cc",
6182 "test/vadd-microkernel-tester.h",
6183 ] + MICROKERNEL_TEST_HDRS,
6184 deps = MICROKERNEL_TEST_DEPS,
6185)
6186
6187xnnpack_unit_test(
Marat Dukhan0270d9f2020-08-11 00:56:46 -07006188 name = "qs8_vaddc_minmax_test",
6189 srcs = [
6190 "test/qs8-vaddc-minmax.cc",
6191 "test/vaddc-microkernel-tester.h",
6192 ] + MICROKERNEL_TEST_HDRS,
6193 deps = MICROKERNEL_TEST_DEPS,
6194)
6195
6196xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07006197 name = "qu8_avgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006198 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07006199 "test/qu8-avgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006200 "test/avgpool-microkernel-tester.h",
6201 "src/xnnpack/AlignedAllocator.h",
6202 ] + MICROKERNEL_TEST_HDRS,
6203 deps = MICROKERNEL_TEST_DEPS,
6204)
6205
6206xnnpack_unit_test(
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07006207 name = "qu8_dwconv_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006208 srcs = [
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07006209 "test/qu8-dwconv-minmax.cc",
6210 "test/dwconv-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006211 "src/xnnpack/AlignedAllocator.h",
6212 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07006213 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07006214)
6215
6216xnnpack_unit_test(
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07006217 name = "qu8_igemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006218 srcs = [
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07006219 "test/qu8-igemm-minmax.cc",
6220 "test/gemm-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006221 "src/xnnpack/AlignedAllocator.h",
6222 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07006223 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07006224)
6225
6226xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07006227 name = "qu8_gavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006228 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07006229 "test/qu8-gavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006230 "test/gavgpool-microkernel-tester.h",
6231 "src/xnnpack/AlignedAllocator.h",
6232 ] + MICROKERNEL_TEST_HDRS,
6233 deps = MICROKERNEL_TEST_DEPS,
6234)
6235
6236xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07006237 name = "qu8_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006238 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07006239 "test/qu8-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006240 "test/gemm-microkernel-tester.h",
6241 "src/xnnpack/AlignedAllocator.h",
6242 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07006243 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07006244)
6245
6246xnnpack_unit_test(
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07006247 name = "qu8_requantization_test",
6248 srcs = [
6249 "src/xnnpack/requantization-stubs.h",
6250 "test/qu8-requantization.cc",
6251 "test/requantization-tester.h",
6252 ] + MICROKERNEL_TEST_HDRS,
6253 deps = MICROKERNEL_TEST_DEPS,
6254)
6255
6256xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07006257 name = "qu8_vadd_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006258 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07006259 "test/qu8-vadd-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006260 "test/vadd-microkernel-tester.h",
6261 ] + MICROKERNEL_TEST_HDRS,
6262 deps = MICROKERNEL_TEST_DEPS,
6263)
6264
6265xnnpack_unit_test(
6266 name = "u8_clamp_test",
6267 srcs = [
6268 "test/u8-clamp.cc",
6269 "test/clamp-microkernel-tester.h",
6270 ] + MICROKERNEL_TEST_HDRS,
6271 deps = MICROKERNEL_TEST_DEPS,
6272)
6273
6274xnnpack_unit_test(
6275 name = "u8_lut32norm_test",
6276 srcs = [
6277 "test/u8-lut32norm.cc",
6278 "test/lut-norm-microkernel-tester.h",
6279 ] + MICROKERNEL_TEST_HDRS,
6280 deps = MICROKERNEL_TEST_DEPS,
6281)
6282
6283xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07006284 name = "u8_maxpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006285 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07006286 "test/u8-maxpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006287 "test/maxpool-microkernel-tester.h",
6288 ] + MICROKERNEL_TEST_HDRS,
6289 deps = MICROKERNEL_TEST_DEPS,
6290)
6291
6292xnnpack_unit_test(
6293 name = "u8_rmax_test",
6294 srcs = [
6295 "test/u8-rmax.cc",
6296 "test/rmax-microkernel-tester.h",
6297 ] + MICROKERNEL_TEST_HDRS,
6298 deps = MICROKERNEL_TEST_DEPS,
6299)
6300
6301xnnpack_unit_test(
Marat Dukhanad71b9a2020-11-20 00:01:51 -08006302 name = "x32_depthtospace2d_chw2hwc_test",
Yury Kartynnike7841862020-11-04 18:22:18 -08006303 srcs = [
Marat Dukhanad71b9a2020-11-20 00:01:51 -08006304 "test/x32-depthtospace2d-chw2hwc.cc",
6305 "test/depthtospace-microkernel-tester.h",
Yury Kartynnike7841862020-11-04 18:22:18 -08006306 ] + MICROKERNEL_TEST_HDRS,
6307 deps = MICROKERNEL_TEST_DEPS,
6308)
6309
6310xnnpack_unit_test(
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -07006311 name = "x32_fill_test",
6312 srcs = [
6313 "test/x32-fill.cc",
6314 "test/fill-microkernel-tester.h",
6315 ] + MICROKERNEL_TEST_HDRS,
6316 deps = MICROKERNEL_TEST_DEPS,
6317)
6318
6319xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07006320 name = "x32_packx_test",
6321 srcs = [
6322 "test/x32-packx.cc",
6323 "test/pack-microkernel-tester.h",
6324 "src/xnnpack/AlignedAllocator.h",
6325 ] + MICROKERNEL_TEST_HDRS,
6326 deps = MICROKERNEL_TEST_DEPS,
6327)
6328
6329xnnpack_unit_test(
6330 name = "x32_pad_test",
6331 srcs = [
6332 "test/x32-pad.cc",
6333 "test/pad-microkernel-tester.h",
6334 ] + MICROKERNEL_TEST_HDRS,
6335 deps = MICROKERNEL_TEST_DEPS,
6336)
6337
6338xnnpack_unit_test(
6339 name = "x32_unpool_test",
6340 srcs = [
6341 "test/x32-unpool.cc",
6342 "test/unpool-microkernel-tester.h",
6343 ] + MICROKERNEL_TEST_HDRS,
6344 deps = MICROKERNEL_TEST_DEPS,
6345)
6346
6347xnnpack_unit_test(
6348 name = "x32_zip_test",
6349 srcs = [
6350 "test/x32-zip.cc",
6351 "test/zip-microkernel-tester.h",
6352 ] + MICROKERNEL_TEST_HDRS,
6353 deps = MICROKERNEL_TEST_DEPS,
6354)
6355
6356xnnpack_unit_test(
6357 name = "x8_lut_test",
6358 srcs = [
6359 "test/x8-lut.cc",
6360 "test/lut-microkernel-tester.h",
6361 ] + MICROKERNEL_TEST_HDRS,
6362 deps = MICROKERNEL_TEST_DEPS,
6363)
6364
6365xnnpack_unit_test(
6366 name = "x8_zip_test",
6367 srcs = [
6368 "test/x8-zip.cc",
6369 "test/zip-microkernel-tester.h",
6370 ] + MICROKERNEL_TEST_HDRS,
6371 deps = MICROKERNEL_TEST_DEPS,
6372)
6373
Marat Dukhan20c3b922020-03-10 03:45:06 -07006374########################## Size tests for the library #########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07006375
6376xnnpack_binary(
Marat Dukhan20c3b922020-03-10 03:45:06 -07006377 name = "operator_size_test",
6378 srcs = ["test/operator-size.c"],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07006379 deps = [":xnnpack_operators_nhwc_f32"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07006380)
6381
Marat Dukhan20c3b922020-03-10 03:45:06 -07006382xnnpack_binary(
6383 name = "subgraph_size_test",
6384 srcs = ["test/subgraph-size.c"],
6385 deps = [":XNNPACK"],
6386)
6387
6388########################### Unit tests for operators ##########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07006389
6390xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07006391 name = "abs_nc_test",
6392 srcs = [
6393 "test/abs-nc.cc",
6394 "test/abs-operator-tester.h",
6395 ],
6396 deps = OPERATOR_TEST_DEPS,
6397)
6398
6399xnnpack_unit_test(
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08006400 name = "add_nd_test",
6401 srcs = [
6402 "test/add-nd.cc",
6403 "test/binary-elementwise-operator-tester.h",
6404 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006405 deps = OPERATOR_TEST_DEPS,
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08006406)
6407
6408xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006409 name = "argmax_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006410 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006411 "test/argmax-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006412 "test/argmax-pooling-operator-tester.h",
6413 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07006414 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006415)
6416
6417xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006418 name = "average_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006419 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006420 "test/average-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006421 "test/average-pooling-operator-tester.h",
6422 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07006423 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006424)
6425
6426xnnpack_unit_test(
Marat Dukhan64e52512020-06-09 13:41:16 -07006427 name = "bankers_rounding_nc_test",
6428 srcs = [
6429 "test/bankers-rounding-nc.cc",
6430 "test/bankers-rounding-operator-tester.h",
6431 ],
6432 deps = OPERATOR_TEST_DEPS,
6433)
6434
6435xnnpack_unit_test(
6436 name = "ceiling_nc_test",
6437 srcs = [
6438 "test/ceiling-nc.cc",
6439 "test/ceiling-operator-tester.h",
6440 ],
6441 deps = OPERATOR_TEST_DEPS,
6442)
6443
6444xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006445 name = "channel_shuffle_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006446 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006447 "test/channel-shuffle-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006448 "test/channel-shuffle-operator-tester.h",
6449 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006450 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006451)
6452
6453xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006454 name = "clamp_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006455 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006456 "test/clamp-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006457 "test/clamp-operator-tester.h",
6458 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006459 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006460)
6461
6462xnnpack_unit_test(
Marat Dukhan065b11e2020-05-22 09:49:41 -07006463 name = "constant_pad_nd_test",
6464 srcs = [
6465 "test/constant-pad-nd.cc",
6466 "test/constant-pad-operator-tester.h",
6467 ],
6468 deps = OPERATOR_TEST_DEPS,
6469)
6470
6471xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006472 name = "convolution_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006473 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006474 "test/convolution-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006475 "test/convolution-operator-tester.h",
6476 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006477 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006478)
6479
6480xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006481 name = "convolution_nchw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006482 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006483 "test/convolution-nchw.cc",
6484 "test/convolution-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006485 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006486 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006487)
6488
6489xnnpack_unit_test(
Marat Dukhan4e21b272020-06-04 18:45:01 -07006490 name = "copy_nc_test",
6491 srcs = [
6492 "test/copy-nc.cc",
6493 "test/copy-operator-tester.h",
6494 ],
6495 deps = OPERATOR_TEST_DEPS,
6496)
6497
6498xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006499 name = "deconvolution_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006500 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006501 "test/deconvolution-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006502 "test/deconvolution-operator-tester.h",
6503 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07006504 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006505)
6506
6507xnnpack_unit_test(
Artsiom Ablavatski0f1dc182020-11-05 19:21:50 -08006508 name = "depth_to_space_test",
6509 srcs = [
6510 "test/depth-to-space.cc",
6511 "test/depth-to-space-operator-tester.h",
6512 ] + OPERATOR_TEST_PARAMS_HDRS,
6513 deps = OPERATOR_TEST_DEPS,
6514)
6515
6516xnnpack_unit_test(
Marat Dukhan69180502019-12-06 15:00:31 -08006517 name = "divide_nd_test",
6518 srcs = [
6519 "test/binary-elementwise-operator-tester.h",
6520 "test/divide-nd.cc",
6521 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006522 deps = OPERATOR_TEST_DEPS,
Marat Dukhan69180502019-12-06 15:00:31 -08006523)
6524
6525xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006526 name = "fully_connected_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006527 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006528 "test/fully-connected-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006529 "test/fully-connected-operator-tester.h",
6530 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006531 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006532)
6533
6534xnnpack_unit_test(
Marat Dukhan64e52512020-06-09 13:41:16 -07006535 name = "floor_nc_test",
6536 srcs = [
6537 "test/floor-nc.cc",
6538 "test/floor-operator-tester.h",
6539 ],
6540 deps = OPERATOR_TEST_DEPS,
6541)
6542
6543xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006544 name = "global_average_pooling_nwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006545 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006546 "test/global-average-pooling-nwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006547 "test/global-average-pooling-operator-tester.h",
Marat Dukhanef61d022020-06-19 13:54:49 -07006548 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07006549 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006550)
6551
6552xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006553 name = "global_average_pooling_ncw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006554 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006555 "test/global-average-pooling-ncw.cc",
6556 "test/global-average-pooling-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006557 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006558 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006559)
6560
6561xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006562 name = "hardswish_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006563 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006564 "test/hardswish-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006565 "test/hardswish-operator-tester.h",
6566 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006567 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006568)
6569
6570xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006571 name = "leaky_relu_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006572 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006573 "test/leaky-relu-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006574 "test/leaky-relu-operator-tester.h",
6575 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006576 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006577)
6578
6579xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006580 name = "max_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006581 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006582 "test/max-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006583 "test/max-pooling-operator-tester.h",
6584 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07006585 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006586)
6587
6588xnnpack_unit_test(
Marat Dukhan79e7f842019-12-05 14:35:50 -08006589 name = "maximum_nd_test",
6590 srcs = [
6591 "test/binary-elementwise-operator-tester.h",
6592 "test/maximum-nd.cc",
6593 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006594 deps = OPERATOR_TEST_DEPS,
Marat Dukhan79e7f842019-12-05 14:35:50 -08006595)
6596
6597xnnpack_unit_test(
6598 name = "minimum_nd_test",
6599 srcs = [
6600 "test/binary-elementwise-operator-tester.h",
6601 "test/minimum-nd.cc",
6602 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006603 deps = OPERATOR_TEST_DEPS,
Marat Dukhan79e7f842019-12-05 14:35:50 -08006604)
6605
6606xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006607 name = "multiply_nd_test",
Marat Dukhanca2733c2019-11-15 23:21:17 -08006608 srcs = [
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08006609 "test/binary-elementwise-operator-tester.h",
Marat Dukhanefc47b82019-11-18 09:25:38 -08006610 "test/multiply-nd.cc",
Marat Dukhanca2733c2019-11-15 23:21:17 -08006611 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006612 deps = OPERATOR_TEST_DEPS,
Marat Dukhanca2733c2019-11-15 23:21:17 -08006613)
6614
6615xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07006616 name = "negate_nc_test",
6617 srcs = [
6618 "test/negate-nc.cc",
6619 "test/negate-operator-tester.h",
6620 ],
6621 deps = OPERATOR_TEST_DEPS,
6622)
6623
6624xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006625 name = "prelu_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006626 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006627 "test/prelu-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006628 "test/prelu-operator-tester.h",
6629 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07006630 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006631)
6632
6633xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006634 name = "resize_bilinear_nhwc_test",
Marat Dukhan69722492019-11-11 19:55:50 -08006635 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006636 "test/resize-bilinear-nhwc.cc",
Marat Dukhan69722492019-11-11 19:55:50 -08006637 "test/resize-bilinear-operator-tester.h",
6638 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07006639 deps = OPERATOR_TEST_DEPS,
Marat Dukhan69722492019-11-11 19:55:50 -08006640)
6641
6642xnnpack_unit_test(
Artsiom Ablavatski97918102020-10-27 15:52:59 -07006643 name = "resize_bilinear_nchw_test",
6644 srcs = [
6645 "test/resize-bilinear-nchw.cc",
6646 "test/resize-bilinear-operator-tester.h",
6647 ] + OPERATOR_TEST_PARAMS_HDRS,
6648 deps = OPERATOR_TEST_DEPS,
6649)
6650
6651xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006652 name = "sigmoid_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006653 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006654 "test/sigmoid-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006655 "test/sigmoid-operator-tester.h",
6656 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006657 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006658)
6659
6660xnnpack_unit_test(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08006661 name = "softmax_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006662 srcs = [
Marat Dukhanfd8e6892020-01-27 15:25:25 -08006663 "test/softmax-nc.cc",
6664 "test/softmax-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006665 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006666 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006667)
6668
6669xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07006670 name = "square_nc_test",
6671 srcs = [
6672 "test/square-nc.cc",
6673 "test/square-operator-tester.h",
6674 ],
6675 deps = OPERATOR_TEST_DEPS,
6676)
6677
6678xnnpack_unit_test(
Marat Dukhan6804bbd2020-06-30 19:26:11 -07006679 name = "square_root_nc_test",
6680 srcs = [
6681 "test/square-root-nc.cc",
6682 "test/square-root-operator-tester.h",
6683 ],
6684 deps = OPERATOR_TEST_DEPS,
6685)
6686
6687xnnpack_unit_test(
Marat Dukhanf7399262020-06-05 10:58:44 -07006688 name = "squared_difference_nd_test",
6689 srcs = [
6690 "test/binary-elementwise-operator-tester.h",
6691 "test/squared-difference-nd.cc",
6692 ],
6693 deps = OPERATOR_TEST_DEPS,
6694)
6695
6696xnnpack_unit_test(
Marat Dukhan05f3f6d2019-12-03 15:13:53 -08006697 name = "subtract_nd_test",
6698 srcs = [
6699 "test/binary-elementwise-operator-tester.h",
6700 "test/subtract-nd.cc",
6701 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006702 deps = OPERATOR_TEST_DEPS,
Marat Dukhan05f3f6d2019-12-03 15:13:53 -08006703)
6704
6705xnnpack_unit_test(
Marat Dukhan64e52512020-06-09 13:41:16 -07006706 name = "truncation_nc_test",
6707 srcs = [
6708 "test/truncation-nc.cc",
6709 "test/truncation-operator-tester.h",
6710 ],
6711 deps = OPERATOR_TEST_DEPS,
6712)
6713
6714xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006715 name = "unpooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006716 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006717 "test/unpooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006718 "test/unpooling-operator-tester.h",
6719 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006720 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006721)
6722
Chao Mei6ddfc602020-05-13 22:29:36 -07006723############################### Misc unit tests ###############################
6724
6725xnnpack_unit_test(
6726 name = "memory_planner_test",
6727 srcs = [
6728 "test/memory-planner-test.cc",
6729 ],
6730 deps = [
6731 ":XNNPACK",
6732 ":memory_planner",
6733 ],
6734)
6735
XNNPACK Teamab8c4c82020-10-09 08:05:51 -07006736xnnpack_unit_test(
6737 name = "subgraph_nchw_test",
6738 srcs = [
6739 "src/xnnpack/subgraph.h",
6740 "test/subgraph-nchw.cc",
6741 "test/subgraph-tester.h",
6742 ],
6743 deps = [
6744 ":XNNPACK",
6745 ],
6746)
6747
Marat Dukhan08c4a432019-10-03 09:29:21 -07006748############################# Build configurations #############################
6749
Marat Dukhanb8642352019-10-30 15:43:02 -07006750# Enables usage of assembly kernels.
Marat Dukhan08c4a432019-10-03 09:29:21 -07006751config_setting(
Marat Dukhanb8642352019-10-30 15:43:02 -07006752 name = "xnn_enable_assembly_explicit_true",
6753 define_values = {"xnn_enable_assembly": "true"},
6754)
6755
6756# Disables usage of assembly kernels.
6757config_setting(
6758 name = "xnn_enable_assembly_explicit_false",
6759 define_values = {"xnn_enable_assembly": "false"},
6760)
6761
Marat Dukhan9de90e02020-06-18 16:04:12 -07006762# Enables usage of sparse inference.
6763config_setting(
6764 name = "xnn_enable_sparse_explicit_true",
6765 define_values = {"xnn_enable_sparse": "true"},
6766)
6767
6768# Disables usage of sparse inference.
6769config_setting(
6770 name = "xnn_enable_sparse_explicit_false",
6771 define_values = {"xnn_enable_sparse": "false"},
6772)
6773
Marat Dukhan05702cf2020-03-26 15:41:33 -07006774# Disables usage of HMP-aware optimizations.
6775config_setting(
6776 name = "xnn_enable_hmp_explicit_false",
6777 define_values = {"xnn_enable_hmp": "false"},
6778)
6779
Chao Mei6ddfc602020-05-13 22:29:36 -07006780# Enable usage of optimized memory allocation
6781config_setting(
6782 name = "xnn_enable_memopt_explicit_true",
6783 define_values = {"xnnpack_enable_memopt": "true"},
6784)
6785
6786# Disable usage of optimized memory allocation
6787config_setting(
6788 name = "xnn_enable_memopt_explicit_false",
6789 define_values = {"xnnpack_enable_memopt": "false"},
6790)
6791
Marat Dukhanb8642352019-10-30 15:43:02 -07006792# Builds with -c dbg
6793config_setting(
6794 name = "debug_build",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006795 values = {
Marat Dukhanb8642352019-10-30 15:43:02 -07006796 "compilation_mode": "dbg",
6797 },
6798)
6799
6800# Builds with -c opt
6801config_setting(
6802 name = "optimized_build",
6803 values = {
6804 "compilation_mode": "opt",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006805 },
6806)
6807
6808config_setting(
Marat Dukhanb8642352019-10-30 15:43:02 -07006809 name = "linux_k8",
6810 values = {"cpu": "k8"},
6811)
6812
6813config_setting(
Marat Dukhan582094e2020-04-30 17:21:25 -07006814 name = "linux_arm",
6815 values = {"cpu": "arm"},
Marat Dukhan4e45e662019-10-03 15:40:24 -07006816)
6817
6818config_setting(
Marat Dukhanf0bd4de2020-06-15 15:53:19 -07006819 name = "linux_armeabi",
6820 values = {"cpu": "armeabi"},
6821)
6822
6823config_setting(
Terry Heo68eef3f2020-04-13 22:53:52 -07006824 name = "linux_armhf",
6825 values = {"cpu": "armhf"},
6826)
6827
6828config_setting(
Marat Dukhana720e932020-06-10 13:01:11 -07006829 name = "linux_armv7a",
6830 values = {"cpu": "armv7a"},
6831)
6832
6833config_setting(
Marat Dukhan582094e2020-04-30 17:21:25 -07006834 name = "linux_aarch64",
6835 values = {"cpu": "aarch64"},
6836)
6837
6838config_setting(
Marat Dukhan08c4a432019-10-03 09:29:21 -07006839 name = "android",
6840 values = {"crosstool_top": "//external:android/crosstool"},
6841)
6842
6843config_setting(
6844 name = "android_armv7",
6845 values = {
6846 "crosstool_top": "//external:android/crosstool",
6847 "cpu": "armeabi-v7a",
6848 },
6849)
6850
6851config_setting(
6852 name = "android_arm64",
6853 values = {
6854 "crosstool_top": "//external:android/crosstool",
6855 "cpu": "arm64-v8a",
6856 },
6857)
6858
6859config_setting(
6860 name = "android_x86",
6861 values = {
6862 "crosstool_top": "//external:android/crosstool",
6863 "cpu": "x86",
6864 },
6865)
6866
6867config_setting(
6868 name = "android_x86_64",
6869 values = {
6870 "crosstool_top": "//external:android/crosstool",
6871 "cpu": "x86_64",
6872 },
6873)
6874
6875config_setting(
Marat Dukhan10a38082020-04-17 03:58:35 -07006876 name = "windows_x86_64",
6877 values = {"cpu": "x64_windows"},
Marat Dukhan9fe932e2020-04-11 17:14:15 -07006878)
6879
6880config_setting(
Marat Dukhan10a38082020-04-17 03:58:35 -07006881 name = "windows_x86_64_clang",
6882 values = {
6883 "compiler": "clang-cl",
6884 "cpu": "x64_windows",
6885 },
6886)
6887
6888config_setting(
6889 name = "windows_x86_64_mingw",
6890 values = {
6891 "compiler": "mingw-gcc",
6892 "cpu": "x64_windows",
6893 },
6894)
6895
6896config_setting(
6897 name = "windows_x86_64_msys",
6898 values = {
6899 "compiler": "msys-gcc",
6900 "cpu": "x64_windows",
6901 },
Marat Dukhan9fe932e2020-04-11 17:14:15 -07006902)
6903
6904config_setting(
Marat Dukhan885ca242019-10-07 09:17:32 -07006905 name = "macos_x86_64",
6906 values = {
6907 "apple_platform_type": "macos",
6908 "cpu": "darwin",
6909 },
6910)
6911
6912config_setting(
Marat Dukhan08c4a432019-10-03 09:29:21 -07006913 name = "emscripten",
Marat Dukhan1a729ec2019-10-07 09:31:44 -07006914 values = {"crosstool_top": "//toolchain:emscripten"},
Marat Dukhan08c4a432019-10-03 09:29:21 -07006915)
6916
6917config_setting(
6918 name = "emscripten_wasm",
6919 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07006920 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006921 "cpu": "wasm",
6922 },
6923)
6924
6925config_setting(
6926 name = "emscripten_wasmsimd",
6927 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07006928 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006929 "cpu": "wasm",
Marat Dukhan81c62602020-05-29 13:22:49 -07006930 "copt": "-msimd128",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006931 },
6932)
6933
6934config_setting(
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006935 name = "ios_armv7",
6936 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006937 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006938 "cpu": "ios_armv7",
6939 },
6940)
6941
6942config_setting(
6943 name = "ios_arm64",
6944 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006945 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006946 "cpu": "ios_arm64",
6947 },
6948)
6949
6950config_setting(
6951 name = "ios_arm64e",
6952 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006953 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006954 "cpu": "ios_arm64e",
6955 },
6956)
6957
6958config_setting(
6959 name = "ios_x86",
6960 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006961 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006962 "cpu": "ios_i386",
6963 },
6964)
6965
6966config_setting(
6967 name = "ios_x86_64",
6968 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006969 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006970 "cpu": "ios_x86_64",
6971 },
6972)
6973
6974config_setting(
6975 name = "watchos_armv7k",
6976 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006977 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006978 "cpu": "watchos_armv7k",
6979 },
6980)
6981
6982config_setting(
6983 name = "watchos_arm64_32",
6984 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006985 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006986 "cpu": "watchos_arm64_32",
6987 },
6988)
6989
6990config_setting(
6991 name = "watchos_x86",
6992 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006993 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006994 "cpu": "watchos_i386",
6995 },
6996)
6997
6998config_setting(
6999 name = "watchos_x86_64",
7000 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08007001 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08007002 "cpu": "watchos_x86_64",
7003 },
7004)
7005
7006config_setting(
7007 name = "tvos_arm64",
7008 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08007009 "apple_platform_type": "tvos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08007010 "cpu": "tvos_arm64",
7011 },
7012)
7013
7014config_setting(
7015 name = "tvos_x86_64",
7016 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08007017 "apple_platform_type": "tvos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08007018 "cpu": "tvos_x86_64",
7019 },
7020)