blob: 97d150c137ade180ad5b52cda91c37f337c784df [file] [log] [blame]
Frank Barchard1a953052020-11-16 18:44:58 -08001# Copyright 2020 Google LLC
Marat Dukhan08c4a432019-10-03 09:29:21 -07002#
3# This source code is licensed under the BSD-style license found in the
4# LICENSE file in the root directory of this source tree.
5#
6# Description:
7# XNNPACK - optimized floating-point neural network operators library
8
Marat Dukhan10a38082020-04-17 03:58:35 -07009load(":build_defs.bzl", "xnnpack_aggregate_library", "xnnpack_benchmark", "xnnpack_binary", "xnnpack_cc_library", "xnnpack_gcc_std_copts", "xnnpack_min_size_copts", "xnnpack_msvc_std_copts", "xnnpack_optional_armcl_copts", "xnnpack_optional_armcl_deps", "xnnpack_optional_dnnl_copts", "xnnpack_optional_dnnl_deps", "xnnpack_optional_gemmlowp_copts", "xnnpack_optional_gemmlowp_deps", "xnnpack_optional_ruy_copts", "xnnpack_optional_ruy_deps", "xnnpack_optional_tflite_copts", "xnnpack_optional_tflite_deps", "xnnpack_std_cxxopts", "xnnpack_unit_test", "xnnpack_visibility")
Marat Dukhan69c3f2c2019-11-06 12:30:01 -080010
Marat Dukhan08c4a432019-10-03 09:29:21 -070011licenses(["notice"])
12
13exports_files(["LICENSE"])
14
Marat Dukhan1b354632020-03-23 12:50:22 -070015OPERATOR_BENCHMARK_DEPS = [
16 ":XNNPACK",
17 ":bench_utils",
18 "@cpuinfo",
Frank Barchard0c849732020-06-12 13:31:32 -070019 "@FP16",
Marat Dukhan1b354632020-03-23 12:50:22 -070020 "@pthreadpool",
21]
22
Marat Dukhan08c4a432019-10-03 09:29:21 -070023MICROKERNEL_BENCHMARK_DEPS = [
24 ":ukernels",
25 ":bench_utils",
Frank Barchard7e955972019-10-11 10:34:25 -070026 ":enable_assembly",
Marat Dukhan08c4a432019-10-03 09:29:21 -070027 "@cpuinfo",
28 "@FP16",
29 "@pthreadpool",
30]
31
Marat Dukhan6adff4e2019-10-14 18:32:07 -070032ACCURACY_EVAL_DEPS = [
33 ":XNNPACK",
34 ":ukernels",
35 "@FP16",
36 "@pthreadpool",
37]
38
Marat Dukhan08c4a432019-10-03 09:29:21 -070039MICROKERNEL_TEST_DEPS = [
Marat Dukhan33fcf782020-05-24 14:27:15 -070040 ":ukernels_test_mode",
Frank Barchard7e955972019-10-11 10:34:25 -070041 ":enable_assembly",
Marat Dukhan08c4a432019-10-03 09:29:21 -070042 "@cpuinfo",
43 "@FP16",
44 "@pthreadpool",
45]
46
Marat Dukhan1b354632020-03-23 12:50:22 -070047OPERATOR_TEST_DEPS = [
Marat Dukhan33fcf782020-05-24 14:27:15 -070048 ":XNNPACK_test_mode",
Marat Dukhan1b354632020-03-23 12:50:22 -070049 "@pthreadpool",
50 "@FP16",
51]
52
Marat Dukhan08c4a432019-10-03 09:29:21 -070053OPERATOR_SRCS = [
Marat Dukhane8265432020-04-28 18:42:59 -070054 "src/operators/argmax-pooling-nhwc.c",
55 "src/operators/average-pooling-nhwc.c",
56 "src/operators/binary-elementwise-nd.c",
Marat Dukhane8265432020-04-28 18:42:59 -070057 "src/operators/channel-shuffle-nc.c",
Marat Dukhan065b11e2020-05-22 09:49:41 -070058 "src/operators/constant-pad-nd.c",
Marat Dukhane8265432020-04-28 18:42:59 -070059 "src/operators/convolution-nchw.c",
60 "src/operators/convolution-nhwc.c",
61 "src/operators/deconvolution-nhwc.c",
Marat Dukhan13b68f22020-11-12 11:55:19 -080062 "src/operators/depth-to-space-nchw2nhwc.c",
Marat Dukhan0e521172020-11-25 13:10:04 -080063 "src/operators/depth-to-space-nhwc.c",
Marat Dukhane8265432020-04-28 18:42:59 -070064 "src/operators/fully-connected-nc.c",
65 "src/operators/global-average-pooling-ncw.c",
66 "src/operators/global-average-pooling-nwc.c",
Marat Dukhane8265432020-04-28 18:42:59 -070067 "src/operators/leaky-relu-nc.c",
68 "src/operators/max-pooling-nhwc.c",
69 "src/operators/prelu-nc.c",
Artsiom Ablavatski97918102020-10-27 15:52:59 -070070 "src/operators/resize-bilinear-nchw.c",
Marat Dukhane8265432020-04-28 18:42:59 -070071 "src/operators/resize-bilinear-nhwc.c",
72 "src/operators/sigmoid-nc.c",
73 "src/operators/softmax-nc.c",
Marat Dukhanc3065f52020-06-04 13:33:32 -070074 "src/operators/unary-elementwise-nc.c",
Marat Dukhane8265432020-04-28 18:42:59 -070075 "src/operators/unpooling-nhwc.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -070076]
77
Marat Dukhan95e8b7a2020-06-03 12:46:26 -070078SUBGRAPH_SRCS = [
Marat Dukhan5fab4092020-06-10 01:28:28 -070079 "src/subgraph/abs.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -070080 "src/subgraph/add2.c",
81 "src/subgraph/argmax-pooling-2d.c",
82 "src/subgraph/average-pooling-2d.c",
Marat Dukhan5fab4092020-06-10 01:28:28 -070083 "src/subgraph/bankers-rounding.c",
84 "src/subgraph/ceiling.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -070085 "src/subgraph/clamp.c",
86 "src/subgraph/convolution-2d.c",
87 "src/subgraph/deconvolution-2d.c",
Artsiom Ablavatskibbe85062020-11-05 14:07:37 -080088 "src/subgraph/depth-to-space.c",
Frank Barchard9cef5ea2020-11-18 14:52:08 -080089 "src/subgraph/depthwise-convolution-2d.c",
Marat Dukhan9d3a4592020-06-05 16:52:42 -070090 "src/subgraph/divide.c",
Marat Dukhana1600202020-12-01 22:17:16 -080091 "src/subgraph/elu.c",
Marat Dukhan5fab4092020-06-10 01:28:28 -070092 "src/subgraph/floor.c",
Frank Barchard04336c12020-10-22 16:48:55 -070093 "src/subgraph/fully-connected.c",
Marat Dukhana059b7d2020-06-11 11:41:27 -070094 "src/subgraph/global-average-pooling-2d.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -070095 "src/subgraph/hardswish.c",
Marat Dukhan5bbebac2020-06-10 19:42:15 -070096 "src/subgraph/leaky-relu.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -070097 "src/subgraph/max-pooling-2d.c",
Marat Dukhan9d3a4592020-06-05 16:52:42 -070098 "src/subgraph/maximum2.c",
99 "src/subgraph/minimum2.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -0700100 "src/subgraph/multiply2.c",
Marat Dukhan5fab4092020-06-10 01:28:28 -0700101 "src/subgraph/negate.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -0700102 "src/subgraph/prelu.c",
103 "src/subgraph/sigmoid.c",
104 "src/subgraph/softmax.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700105 "src/subgraph/square-root.c",
106 "src/subgraph/square.c",
107 "src/subgraph/squared-difference.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -0700108 "src/subgraph/static-constant-pad.c",
Marat Dukhand27202d2020-07-09 23:43:40 -0700109 "src/subgraph/static-reshape.c",
Marat Dukhanaff24e22020-07-23 01:43:58 -0700110 "src/subgraph/static-resize-bilinear-2d.c",
Marat Dukhan9d3a4592020-06-05 16:52:42 -0700111 "src/subgraph/subtract.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -0700112 "src/subgraph/unpooling-2d.c",
113]
114
Marat Dukhan3a77ea72019-12-23 12:10:24 -0800115TABLE_SRCS = [
116 "src/tables/exp2-k-over-64.c",
117 "src/tables/exp2-k-over-2048.c",
Marat Dukhande390d42020-11-29 19:32:18 -0800118 "src/tables/exp2minus-k-over-4.c",
119 "src/tables/exp2minus-k-over-8.c",
Marat Dukhanc60742b2020-11-23 12:33:27 -0800120 "src/tables/exp2minus-k-over-16.c",
Marat Dukhan1f256fc2020-09-24 21:27:14 -0700121 "src/tables/exp2minus-k-over-64.c",
122 "src/tables/exp2minus-k-over-2048.c",
Marat Dukhan3a77ea72019-12-23 12:10:24 -0800123]
124
Marat Dukhan08c4a432019-10-03 09:29:21 -0700125SCALAR_UKERNELS = [
Marat Dukhan329da642019-11-19 21:44:39 -0800126 "src/f32-argmaxpool/4x-scalar-c1.c",
Marat Dukhan1e782c42019-11-21 17:02:40 -0800127 "src/f32-argmaxpool/9p8x-scalar-c1.c",
Marat Dukhan329da642019-11-19 21:44:39 -0800128 "src/f32-argmaxpool/9x-scalar-c1.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700129 "src/f32-avgpool/9p8x-minmax-scalar-c1.c",
130 "src/f32-avgpool/9x-minmax-scalar-c1.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -0700131 "src/f32-clamp/gen/scalar-x1.c",
132 "src/f32-clamp/gen/scalar-x2.c",
133 "src/f32-clamp/gen/scalar-x4.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700134 "src/f32-conv-hwc/3x3s2p0p1c3x4-scalar-1x1.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700135 "src/f32-conv-hwc/3x3s2p1c3x4-scalar-1x1.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -0700136 "src/f32-conv-hwc2chw/3x3s2p1c3x4-scalar-1x1.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700137 "src/f32-dwconv/gen/up1x4-minmax-scalar-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700138 "src/f32-dwconv/gen/up1x4-minmax-scalar.c",
139 "src/f32-dwconv/gen/up1x4-scalar-acc2.c",
140 "src/f32-dwconv/gen/up1x4-scalar.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700141 "src/f32-dwconv/gen/up1x9-minmax-scalar-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700142 "src/f32-dwconv/gen/up1x9-minmax-scalar.c",
143 "src/f32-dwconv/gen/up1x9-scalar-acc2.c",
144 "src/f32-dwconv/gen/up1x9-scalar.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700145 "src/f32-dwconv/gen/up1x25-minmax-scalar-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700146 "src/f32-dwconv/gen/up1x25-minmax-scalar.c",
147 "src/f32-dwconv/gen/up1x25-scalar-acc2.c",
148 "src/f32-dwconv/gen/up1x25-scalar.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700149 "src/f32-dwconv/gen/up2x4-minmax-scalar-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700150 "src/f32-dwconv/gen/up2x4-minmax-scalar.c",
151 "src/f32-dwconv/gen/up2x4-scalar-acc2.c",
152 "src/f32-dwconv/gen/up2x4-scalar.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700153 "src/f32-dwconv/gen/up2x9-minmax-scalar-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700154 "src/f32-dwconv/gen/up2x9-minmax-scalar.c",
155 "src/f32-dwconv/gen/up2x9-scalar-acc2.c",
156 "src/f32-dwconv/gen/up2x9-scalar.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700157 "src/f32-dwconv/gen/up2x25-minmax-scalar-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700158 "src/f32-dwconv/gen/up2x25-minmax-scalar.c",
159 "src/f32-dwconv/gen/up2x25-scalar-acc2.c",
160 "src/f32-dwconv/gen/up2x25-scalar.c",
Frank Barchard35db7d02020-10-26 13:37:34 -0700161 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-scalar-1x1-acc2.c",
162 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-scalar-1x1-acc3.c",
163 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-scalar-1x1-acc4.c",
Marat Dukhan91249d22020-10-24 12:02:51 -0700164 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-scalar-1x1.c",
Frank Barchard35db7d02020-10-26 13:37:34 -0700165 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-scalar-2x1-acc2.c",
Marat Dukhan91249d22020-10-24 12:02:51 -0700166 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-scalar-2x1.c",
167 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-scalar-3x1.c",
168 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-scalar-4x1.c",
169 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-scalar-5x1.c",
170 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-scalar-6x1.c",
Marat Dukhancf5b3c32020-10-25 19:21:10 -0700171 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-scalar-1x1-acc2.c",
172 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-scalar-1x1-acc3.c",
173 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-scalar-1x1-acc4.c",
Frank Barchard35db7d02020-10-26 13:37:34 -0700174 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-scalar-1x1.c",
Marat Dukhancf5b3c32020-10-25 19:21:10 -0700175 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-scalar-2x1-acc2.c",
Frank Barchard35db7d02020-10-26 13:37:34 -0700176 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-scalar-2x1.c",
177 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-scalar-3x1.c",
178 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-scalar-4x1.c",
Marat Dukhanc4efb002020-10-25 23:14:47 -0700179 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-scalar-1x1-acc2.c",
180 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-scalar-1x1-acc3.c",
181 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-scalar-1x1-acc4.c",
182 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-scalar-1x1-acc5.c",
Frank Barchard35db7d02020-10-26 13:37:34 -0700183 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-scalar-1x1.c",
Marat Dukhanc4efb002020-10-25 23:14:47 -0700184 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-scalar-2x1-acc2.c",
185 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-scalar-2x1-acc3.c",
Frank Barchard35db7d02020-10-26 13:37:34 -0700186 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-scalar-2x1.c",
Marat Dukhanc4efb002020-10-25 23:14:47 -0700187 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-scalar-3x1-acc2.c",
Frank Barchard35db7d02020-10-26 13:37:34 -0700188 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-scalar-3x1.c",
Marat Dukhan29c0c332020-10-28 22:11:00 -0700189 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-scalar-1x1-acc2.c",
190 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-scalar-1x1-acc3.c",
191 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-scalar-1x1-acc4.c",
192 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-scalar-1x1-acc5.c",
193 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-scalar-1x1.c",
194 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-scalar-2x1-acc2.c",
195 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-scalar-2x1-acc3.c",
196 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-scalar-2x1.c",
197 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-scalar-3x1-acc2.c",
198 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-scalar-3x1.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -0700199 "src/f32-gavgpool-cw/scalar-x1.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700200 "src/f32-gavgpool/7p7x-minmax-scalar-c1.c",
201 "src/f32-gavgpool/7x-minmax-scalar-c1.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700202 "src/f32-gemm/gen-inc/1x4inc-minmax-scalar.c",
203 "src/f32-gemm/gen-inc/2x4inc-minmax-scalar.c",
204 "src/f32-gemm/gen-inc/4x4inc-minmax-scalar.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700205 "src/f32-gemm/gen/1x4-minmax-scalar.c",
206 "src/f32-gemm/gen/1x4-relu-scalar.c",
207 "src/f32-gemm/gen/1x4-scalar.c",
208 "src/f32-gemm/gen/2x4-minmax-scalar.c",
209 "src/f32-gemm/gen/2x4-relu-scalar.c",
210 "src/f32-gemm/gen/2x4-scalar.c",
211 "src/f32-gemm/gen/4x2-minmax-scalar.c",
212 "src/f32-gemm/gen/4x2-relu-scalar.c",
213 "src/f32-gemm/gen/4x2-scalar.c",
214 "src/f32-gemm/gen/4x4-minmax-scalar.c",
215 "src/f32-gemm/gen/4x4-relu-scalar.c",
216 "src/f32-gemm/gen/4x4-scalar.c",
Marat Dukhanefaac272020-07-16 16:58:07 -0700217 "src/f32-hswish/gen/hswish-scalar-x1.c",
218 "src/f32-hswish/gen/hswish-scalar-x2.c",
219 "src/f32-hswish/gen/hswish-scalar-x4.c",
XNNPACK Team21432672020-10-19 19:58:48 -0700220 "src/f32-ibilinear-chw/gen/scalar-p1.c",
221 "src/f32-ibilinear-chw/gen/scalar-p2.c",
222 "src/f32-ibilinear-chw/gen/scalar-p4.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700223 "src/f32-ibilinear/gen/scalar-c1.c",
224 "src/f32-ibilinear/gen/scalar-c2.c",
225 "src/f32-ibilinear/gen/scalar-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700226 "src/f32-igemm/gen/1x4-minmax-scalar.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700227 "src/f32-igemm/gen/1x4-relu-scalar.c",
228 "src/f32-igemm/gen/1x4-scalar.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700229 "src/f32-igemm/gen/2x4-minmax-scalar.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700230 "src/f32-igemm/gen/2x4-relu-scalar.c",
231 "src/f32-igemm/gen/2x4-scalar.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700232 "src/f32-igemm/gen/4x2-minmax-scalar.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700233 "src/f32-igemm/gen/4x2-relu-scalar.c",
234 "src/f32-igemm/gen/4x2-scalar.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700235 "src/f32-igemm/gen/4x4-minmax-scalar.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700236 "src/f32-igemm/gen/4x4-relu-scalar.c",
237 "src/f32-igemm/gen/4x4-scalar.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700238 "src/f32-maxpool/9p8x-minmax-scalar-c1.c",
239 "src/f32-pavgpool/9p8x-minmax-scalar-c1.c",
240 "src/f32-pavgpool/9x-minmax-scalar-c1.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700241 "src/f32-ppmm/gen/2x4-minmax-scalar.c",
242 "src/f32-ppmm/gen/3x3-minmax-scalar.c",
243 "src/f32-ppmm/gen/4x2-minmax-scalar.c",
244 "src/f32-ppmm/gen/4x4-minmax-scalar.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800245 "src/f32-prelu/gen/scalar-2x1.c",
246 "src/f32-prelu/gen/scalar-2x4.c",
Marat Dukhanf46f6752020-01-21 11:03:49 -0800247 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x1.c",
Marat Dukhanf46f6752020-01-21 11:03:49 -0800248 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x2-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700249 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x2.c",
Marat Dukhanf46f6752020-01-21 11:03:49 -0800250 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x4-acc2.c",
251 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x4-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700252 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x4.c",
Marat Dukhanf46f6752020-01-21 11:03:49 -0800253 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x1.c",
Marat Dukhanf46f6752020-01-21 11:03:49 -0800254 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x2-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700255 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x2.c",
Marat Dukhanf46f6752020-01-21 11:03:49 -0800256 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x4-acc2.c",
257 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x4-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700258 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x4.c",
Frank Barchardfb158e22020-07-15 16:10:10 -0700259 "src/f32-relu/gen/scalar-x1.c",
260 "src/f32-relu/gen/scalar-x2.c",
261 "src/f32-relu/gen/scalar-x4.c",
Frank Barchard37297a62020-08-27 12:15:10 -0700262 "src/f32-relu/gen/scalar-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700263 "src/f32-rmax/scalar.c",
Marat Dukhan3a77ea72019-12-23 12:10:24 -0800264 "src/f32-sigmoid/gen/scalar-lut64-p2-div-x1.c",
265 "src/f32-sigmoid/gen/scalar-lut64-p2-div-x2.c",
266 "src/f32-sigmoid/gen/scalar-lut64-p2-div-x4.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700267 "src/f32-sigmoid/gen/scalar-lut2048-p1-div-x1.c",
268 "src/f32-sigmoid/gen/scalar-lut2048-p1-div-x2.c",
269 "src/f32-sigmoid/gen/scalar-lut2048-p1-div-x4.c",
Marat Dukhan3a77ea72019-12-23 12:10:24 -0800270 "src/f32-sigmoid/gen/scalar-p5-div-x1.c",
271 "src/f32-sigmoid/gen/scalar-p5-div-x2.c",
272 "src/f32-sigmoid/gen/scalar-p5-div-x4.c",
Marat Dukhan355ab432020-04-09 19:01:52 -0700273 "src/f32-spmm/gen/1x1-minmax-scalar-pipelined.c",
274 "src/f32-spmm/gen/1x1-minmax-scalar.c",
275 "src/f32-spmm/gen/2x1-minmax-scalar-pipelined.c",
276 "src/f32-spmm/gen/2x1-minmax-scalar.c",
277 "src/f32-spmm/gen/4x1-minmax-scalar-pipelined.c",
278 "src/f32-spmm/gen/4x1-minmax-scalar.c",
279 "src/f32-spmm/gen/8x1-minmax-scalar-pipelined.c",
280 "src/f32-spmm/gen/8x1-minmax-scalar.c",
281 "src/f32-spmm/gen/8x2-minmax-scalar.c",
282 "src/f32-spmm/gen/8x4-minmax-scalar.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700283 "src/f32-vbinary/gen/vadd-minmax-scalar-x1.c",
284 "src/f32-vbinary/gen/vadd-minmax-scalar-x2.c",
285 "src/f32-vbinary/gen/vadd-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700286 "src/f32-vbinary/gen/vadd-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700287 "src/f32-vbinary/gen/vadd-relu-scalar-x1.c",
288 "src/f32-vbinary/gen/vadd-relu-scalar-x2.c",
289 "src/f32-vbinary/gen/vadd-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700290 "src/f32-vbinary/gen/vadd-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700291 "src/f32-vbinary/gen/vadd-scalar-x1.c",
292 "src/f32-vbinary/gen/vadd-scalar-x2.c",
293 "src/f32-vbinary/gen/vadd-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700294 "src/f32-vbinary/gen/vadd-scalar-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700295 "src/f32-vbinary/gen/vaddc-minmax-scalar-x1.c",
296 "src/f32-vbinary/gen/vaddc-minmax-scalar-x2.c",
297 "src/f32-vbinary/gen/vaddc-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700298 "src/f32-vbinary/gen/vaddc-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700299 "src/f32-vbinary/gen/vaddc-relu-scalar-x1.c",
300 "src/f32-vbinary/gen/vaddc-relu-scalar-x2.c",
301 "src/f32-vbinary/gen/vaddc-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700302 "src/f32-vbinary/gen/vaddc-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700303 "src/f32-vbinary/gen/vaddc-scalar-x1.c",
304 "src/f32-vbinary/gen/vaddc-scalar-x2.c",
305 "src/f32-vbinary/gen/vaddc-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700306 "src/f32-vbinary/gen/vaddc-scalar-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700307 "src/f32-vbinary/gen/vdiv-minmax-scalar-x1.c",
308 "src/f32-vbinary/gen/vdiv-minmax-scalar-x2.c",
309 "src/f32-vbinary/gen/vdiv-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700310 "src/f32-vbinary/gen/vdiv-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700311 "src/f32-vbinary/gen/vdiv-relu-scalar-x1.c",
312 "src/f32-vbinary/gen/vdiv-relu-scalar-x2.c",
313 "src/f32-vbinary/gen/vdiv-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700314 "src/f32-vbinary/gen/vdiv-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700315 "src/f32-vbinary/gen/vdiv-scalar-x1.c",
316 "src/f32-vbinary/gen/vdiv-scalar-x2.c",
317 "src/f32-vbinary/gen/vdiv-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700318 "src/f32-vbinary/gen/vdiv-scalar-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700319 "src/f32-vbinary/gen/vdivc-minmax-scalar-x1.c",
320 "src/f32-vbinary/gen/vdivc-minmax-scalar-x2.c",
321 "src/f32-vbinary/gen/vdivc-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700322 "src/f32-vbinary/gen/vdivc-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700323 "src/f32-vbinary/gen/vdivc-relu-scalar-x1.c",
324 "src/f32-vbinary/gen/vdivc-relu-scalar-x2.c",
325 "src/f32-vbinary/gen/vdivc-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700326 "src/f32-vbinary/gen/vdivc-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700327 "src/f32-vbinary/gen/vdivc-scalar-x1.c",
328 "src/f32-vbinary/gen/vdivc-scalar-x2.c",
329 "src/f32-vbinary/gen/vdivc-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700330 "src/f32-vbinary/gen/vdivc-scalar-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800331 "src/f32-vbinary/gen/vmax-scalar-x1.c",
332 "src/f32-vbinary/gen/vmax-scalar-x2.c",
333 "src/f32-vbinary/gen/vmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700334 "src/f32-vbinary/gen/vmax-scalar-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800335 "src/f32-vbinary/gen/vmaxc-scalar-x1.c",
336 "src/f32-vbinary/gen/vmaxc-scalar-x2.c",
337 "src/f32-vbinary/gen/vmaxc-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700338 "src/f32-vbinary/gen/vmaxc-scalar-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800339 "src/f32-vbinary/gen/vmin-scalar-x1.c",
340 "src/f32-vbinary/gen/vmin-scalar-x2.c",
341 "src/f32-vbinary/gen/vmin-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700342 "src/f32-vbinary/gen/vmin-scalar-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800343 "src/f32-vbinary/gen/vminc-scalar-x1.c",
344 "src/f32-vbinary/gen/vminc-scalar-x2.c",
345 "src/f32-vbinary/gen/vminc-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700346 "src/f32-vbinary/gen/vminc-scalar-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700347 "src/f32-vbinary/gen/vmul-minmax-scalar-x1.c",
348 "src/f32-vbinary/gen/vmul-minmax-scalar-x2.c",
349 "src/f32-vbinary/gen/vmul-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700350 "src/f32-vbinary/gen/vmul-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700351 "src/f32-vbinary/gen/vmul-relu-scalar-x1.c",
352 "src/f32-vbinary/gen/vmul-relu-scalar-x2.c",
353 "src/f32-vbinary/gen/vmul-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700354 "src/f32-vbinary/gen/vmul-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700355 "src/f32-vbinary/gen/vmul-scalar-x1.c",
356 "src/f32-vbinary/gen/vmul-scalar-x2.c",
357 "src/f32-vbinary/gen/vmul-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700358 "src/f32-vbinary/gen/vmul-scalar-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700359 "src/f32-vbinary/gen/vmulc-minmax-scalar-x1.c",
360 "src/f32-vbinary/gen/vmulc-minmax-scalar-x2.c",
361 "src/f32-vbinary/gen/vmulc-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700362 "src/f32-vbinary/gen/vmulc-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700363 "src/f32-vbinary/gen/vmulc-relu-scalar-x1.c",
364 "src/f32-vbinary/gen/vmulc-relu-scalar-x2.c",
365 "src/f32-vbinary/gen/vmulc-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700366 "src/f32-vbinary/gen/vmulc-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700367 "src/f32-vbinary/gen/vmulc-scalar-x1.c",
368 "src/f32-vbinary/gen/vmulc-scalar-x2.c",
369 "src/f32-vbinary/gen/vmulc-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700370 "src/f32-vbinary/gen/vmulc-scalar-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700371 "src/f32-vbinary/gen/vrdivc-minmax-scalar-x1.c",
372 "src/f32-vbinary/gen/vrdivc-minmax-scalar-x2.c",
373 "src/f32-vbinary/gen/vrdivc-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700374 "src/f32-vbinary/gen/vrdivc-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700375 "src/f32-vbinary/gen/vrdivc-relu-scalar-x1.c",
376 "src/f32-vbinary/gen/vrdivc-relu-scalar-x2.c",
377 "src/f32-vbinary/gen/vrdivc-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700378 "src/f32-vbinary/gen/vrdivc-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700379 "src/f32-vbinary/gen/vrdivc-scalar-x1.c",
380 "src/f32-vbinary/gen/vrdivc-scalar-x2.c",
381 "src/f32-vbinary/gen/vrdivc-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700382 "src/f32-vbinary/gen/vrdivc-scalar-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700383 "src/f32-vbinary/gen/vrsubc-minmax-scalar-x1.c",
384 "src/f32-vbinary/gen/vrsubc-minmax-scalar-x2.c",
385 "src/f32-vbinary/gen/vrsubc-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700386 "src/f32-vbinary/gen/vrsubc-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700387 "src/f32-vbinary/gen/vrsubc-relu-scalar-x1.c",
388 "src/f32-vbinary/gen/vrsubc-relu-scalar-x2.c",
389 "src/f32-vbinary/gen/vrsubc-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700390 "src/f32-vbinary/gen/vrsubc-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700391 "src/f32-vbinary/gen/vrsubc-scalar-x1.c",
392 "src/f32-vbinary/gen/vrsubc-scalar-x2.c",
393 "src/f32-vbinary/gen/vrsubc-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700394 "src/f32-vbinary/gen/vrsubc-scalar-x8.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -0700395 "src/f32-vbinary/gen/vsqrdiff-scalar-x1.c",
396 "src/f32-vbinary/gen/vsqrdiff-scalar-x2.c",
397 "src/f32-vbinary/gen/vsqrdiff-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700398 "src/f32-vbinary/gen/vsqrdiff-scalar-x8.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -0700399 "src/f32-vbinary/gen/vsqrdiffc-scalar-x1.c",
400 "src/f32-vbinary/gen/vsqrdiffc-scalar-x2.c",
401 "src/f32-vbinary/gen/vsqrdiffc-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700402 "src/f32-vbinary/gen/vsqrdiffc-scalar-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700403 "src/f32-vbinary/gen/vsub-minmax-scalar-x1.c",
404 "src/f32-vbinary/gen/vsub-minmax-scalar-x2.c",
405 "src/f32-vbinary/gen/vsub-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700406 "src/f32-vbinary/gen/vsub-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700407 "src/f32-vbinary/gen/vsub-relu-scalar-x1.c",
408 "src/f32-vbinary/gen/vsub-relu-scalar-x2.c",
409 "src/f32-vbinary/gen/vsub-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700410 "src/f32-vbinary/gen/vsub-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700411 "src/f32-vbinary/gen/vsub-scalar-x1.c",
412 "src/f32-vbinary/gen/vsub-scalar-x2.c",
413 "src/f32-vbinary/gen/vsub-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700414 "src/f32-vbinary/gen/vsub-scalar-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700415 "src/f32-vbinary/gen/vsubc-minmax-scalar-x1.c",
416 "src/f32-vbinary/gen/vsubc-minmax-scalar-x2.c",
417 "src/f32-vbinary/gen/vsubc-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700418 "src/f32-vbinary/gen/vsubc-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700419 "src/f32-vbinary/gen/vsubc-relu-scalar-x1.c",
420 "src/f32-vbinary/gen/vsubc-relu-scalar-x2.c",
421 "src/f32-vbinary/gen/vsubc-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700422 "src/f32-vbinary/gen/vsubc-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700423 "src/f32-vbinary/gen/vsubc-scalar-x1.c",
424 "src/f32-vbinary/gen/vsubc-scalar-x2.c",
425 "src/f32-vbinary/gen/vsubc-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700426 "src/f32-vbinary/gen/vsubc-scalar-x8.c",
Marat Dukhaned6baaf2020-12-01 15:07:08 -0800427 "src/f32-velu/gen/velu-scalar-rr2-lut16-p3-x1.c",
428 "src/f32-velu/gen/velu-scalar-rr2-lut16-p3-x2.c",
429 "src/f32-velu/gen/velu-scalar-rr2-lut16-p3-x3.c",
430 "src/f32-velu/gen/velu-scalar-rr2-lut16-p3-x4.c",
431 "src/f32-velu/gen/velu-scalar-rr2-lut16-p3-x5.c",
432 "src/f32-velu/gen/velu-scalar-rr2-lut16-p3-x6.c",
433 "src/f32-velu/gen/velu-scalar-rr2-p6-x1.c",
434 "src/f32-velu/gen/velu-scalar-rr2-p6-x2.c",
435 "src/f32-velu/gen/velu-scalar-rr2-p6-x3.c",
436 "src/f32-velu/gen/velu-scalar-rr2-p6-x4.c",
437 "src/f32-velu/gen/velu-scalar-rr2-p6-x5.c",
438 "src/f32-velu/gen/velu-scalar-rr2-p6-x6.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700439 "src/f32-vlrelu/gen/vlrelu-scalar-x1.c",
440 "src/f32-vlrelu/gen/vlrelu-scalar-x2.c",
441 "src/f32-vlrelu/gen/vlrelu-scalar-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700442 "src/f32-vmulcaddc/gen/c1-minmax-scalar-2x.c",
443 "src/f32-vmulcaddc/gen/c2-minmax-scalar-2x.c",
444 "src/f32-vmulcaddc/gen/c4-minmax-scalar-2x.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -0700445 "src/f32-vrnd/gen/vrndd-scalar-libm-x1.c",
446 "src/f32-vrnd/gen/vrndd-scalar-libm-x2.c",
447 "src/f32-vrnd/gen/vrndd-scalar-libm-x4.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700448 "src/f32-vrnd/gen/vrndne-scalar-libm-x1.c",
449 "src/f32-vrnd/gen/vrndne-scalar-libm-x2.c",
450 "src/f32-vrnd/gen/vrndne-scalar-libm-x4.c",
451 "src/f32-vrnd/gen/vrndu-scalar-libm-x1.c",
452 "src/f32-vrnd/gen/vrndu-scalar-libm-x2.c",
453 "src/f32-vrnd/gen/vrndu-scalar-libm-x4.c",
454 "src/f32-vrnd/gen/vrndz-scalar-libm-x1.c",
455 "src/f32-vrnd/gen/vrndz-scalar-libm-x2.c",
456 "src/f32-vrnd/gen/vrndz-scalar-libm-x4.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -0700457 "src/f32-vsqrt/gen/scalar-sqrt-x1.c",
458 "src/f32-vsqrt/gen/scalar-sqrt-x2.c",
459 "src/f32-vsqrt/gen/scalar-sqrt-x4.c",
Marat Dukhan5020b962020-06-08 13:30:10 -0700460 "src/f32-vunary/gen/vabs-scalar-x1.c",
461 "src/f32-vunary/gen/vabs-scalar-x2.c",
462 "src/f32-vunary/gen/vabs-scalar-x4.c",
463 "src/f32-vunary/gen/vneg-scalar-x1.c",
464 "src/f32-vunary/gen/vneg-scalar-x2.c",
465 "src/f32-vunary/gen/vneg-scalar-x4.c",
466 "src/f32-vunary/gen/vsqr-scalar-x1.c",
467 "src/f32-vunary/gen/vsqr-scalar-x2.c",
468 "src/f32-vunary/gen/vsqr-scalar-x4.c",
Marat Dukhande390d42020-11-29 19:32:18 -0800469 "src/math/expm1minus-scalar-rr2-lut4-p4.c",
470 "src/math/expm1minus-scalar-rr2-lut8-p3.c",
471 "src/math/expm1minus-scalar-rr2-lut8-p4.c",
Marat Dukhanc60742b2020-11-23 12:33:27 -0800472 "src/math/expm1minus-scalar-rr2-lut16-p3.c",
473 "src/math/expm1minus-scalar-rr2-lut16-p4.c",
474 "src/math/expm1minus-scalar-rr2-p5.c",
475 "src/math/expm1minus-scalar-rr2-p6.c",
Frank Barchard22136062020-11-24 18:44:46 -0800476 "src/math/expminus-scalar-rr2-lut64-p2.c",
477 "src/math/expminus-scalar-rr2-lut2048-p1.c",
478 "src/math/expminus-scalar-rr2-p5.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -0700479 "src/math/roundd-scalar-addsub.c",
480 "src/math/roundd-scalar-cvt.c",
Marat Dukhanffbf96a2020-05-14 02:59:08 -0700481 "src/math/roundd-scalar-floor.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700482 "src/math/roundne-scalar-addsub.c",
483 "src/math/roundne-scalar-nearbyint.c",
484 "src/math/roundne-scalar-rint.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -0700485 "src/math/roundu-scalar-addsub.c",
Marat Dukhanffbf96a2020-05-14 02:59:08 -0700486 "src/math/roundu-scalar-ceil.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -0700487 "src/math/roundu-scalar-cvt.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -0700488 "src/math/roundz-scalar-addsub.c",
489 "src/math/roundz-scalar-cvt.c",
Marat Dukhanffbf96a2020-05-14 02:59:08 -0700490 "src/math/roundz-scalar-trunc.c",
Marat Dukhanf8475d62020-09-17 15:01:43 -0700491 "src/math/sigmoid-scalar-rr2-lut64-p2-div.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700492 "src/math/sigmoid-scalar-rr2-lut2048-p1-div.c",
Marat Dukhanf8475d62020-09-17 15:01:43 -0700493 "src/math/sigmoid-scalar-rr2-p5-div.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -0700494 "src/qs8-requantization/fp32-scalar-lrintf.c",
495 "src/qs8-requantization/fp32-scalar-magic.c",
496 "src/qs8-requantization/precise-scalar-signed64.c",
497 "src/qs8-requantization/precise-scalar-unsigned32.c",
498 "src/qs8-requantization/precise-scalar-unsigned64.c",
499 "src/qs8-requantization/q31-scalar.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -0700500 "src/qu8-avgpool/9p8x-minmax-scalar-c1.c",
501 "src/qu8-avgpool/9x-minmax-scalar-c1.c",
502 "src/qu8-dwconv/up1x9-minmax-scalar.c",
503 "src/qu8-gavgpool/7p7x-minmax-scalar-c1.c",
504 "src/qu8-gavgpool/7x-minmax-scalar-c1.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700505 "src/qu8-gemm/2x2-minmax-scalar.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -0700506 "src/qu8-igemm/2x2-minmax-scalar.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -0700507 "src/qu8-requantization/fp32-scalar-lrintf.c",
508 "src/qu8-requantization/fp32-scalar-magic.c",
509 "src/qu8-requantization/precise-scalar-signed64.c",
510 "src/qu8-requantization/precise-scalar-unsigned32.c",
511 "src/qu8-requantization/precise-scalar-unsigned64.c",
512 "src/qu8-requantization/q31-scalar.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -0700513 "src/qu8-vadd/minmax-scalar.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -0700514 "src/u8-clamp/scalar-x4.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700515 "src/u8-lut32norm/scalar.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700516 "src/u8-maxpool/9p8x-minmax-scalar-c1.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700517 "src/u8-rmax/scalar.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700518 "src/x8-lut/scalar.c",
519 "src/x8-zip/x2-scalar.c",
520 "src/x8-zip/x3-scalar.c",
521 "src/x8-zip/x4-scalar.c",
522 "src/x8-zip/xm-scalar.c",
Marat Dukhanad71b9a2020-11-20 00:01:51 -0800523 "src/x32-depthtospace2d-chw2hwc/scalar.c",
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -0700524 "src/x32-fill/scalar-float.c",
525 "src/x32-fill/scalar-int.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700526 "src/x32-packx/x2-scalar.c",
527 "src/x32-packx/x3-scalar.c",
528 "src/x32-packx/x4-scalar.c",
Marat Dukhan63523d42020-05-22 17:07:33 -0700529 "src/x32-pad/scalar-float.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700530 "src/x32-pad/scalar-int.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700531 "src/x32-unpool/scalar.c",
532 "src/x32-zip/x2-scalar.c",
533 "src/x32-zip/x3-scalar.c",
534 "src/x32-zip/x4-scalar.c",
535 "src/x32-zip/xm-scalar.c",
Marat Dukhan048931b2020-11-24 20:53:54 -0800536 "src/xx-copy/memcpy.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700537]
538
Marat Dukhan436ebe62019-12-04 15:10:12 -0800539WASM_UKERNELS = [
Marat Dukhan99936602020-04-11 16:47:01 -0700540 "src/f32-avgpool/9p8x-minmax-wasm-c1.c",
541 "src/f32-avgpool/9x-minmax-wasm-c1.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -0700542 "src/f32-clamp/gen/wasm-x1.c",
543 "src/f32-clamp/gen/wasm-x2.c",
544 "src/f32-clamp/gen/wasm-x4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700545 "src/f32-dwconv/gen/up1x4-minmax-wasm-acc2.c",
546 "src/f32-dwconv/gen/up1x4-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700547 "src/f32-dwconv/gen/up1x4-wasm-acc2.c",
548 "src/f32-dwconv/gen/up1x4-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700549 "src/f32-dwconv/gen/up1x9-minmax-wasm-acc2.c",
550 "src/f32-dwconv/gen/up1x9-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700551 "src/f32-dwconv/gen/up1x9-wasm-acc2.c",
552 "src/f32-dwconv/gen/up1x9-wasm.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700553 "src/f32-dwconv/gen/up1x25-minmax-wasm-acc2.c",
554 "src/f32-dwconv/gen/up1x25-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700555 "src/f32-dwconv/gen/up1x25-wasm-acc2.c",
556 "src/f32-dwconv/gen/up1x25-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700557 "src/f32-dwconv/gen/up2x4-minmax-wasm-acc2.c",
558 "src/f32-dwconv/gen/up2x4-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700559 "src/f32-dwconv/gen/up2x4-wasm-acc2.c",
560 "src/f32-dwconv/gen/up2x4-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700561 "src/f32-dwconv/gen/up2x9-minmax-wasm-acc2.c",
562 "src/f32-dwconv/gen/up2x9-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700563 "src/f32-dwconv/gen/up2x9-wasm-acc2.c",
564 "src/f32-dwconv/gen/up2x9-wasm.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700565 "src/f32-dwconv/gen/up2x25-minmax-wasm-acc2.c",
566 "src/f32-dwconv/gen/up2x25-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700567 "src/f32-dwconv/gen/up2x25-wasm-acc2.c",
568 "src/f32-dwconv/gen/up2x25-wasm.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700569 "src/f32-gavgpool/7p7x-minmax-wasm-c1.c",
570 "src/f32-gavgpool/7x-minmax-wasm-c1.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700571 "src/f32-gemm/gen-inc/1x4inc-minmax-wasm.c",
572 "src/f32-gemm/gen-inc/2x4inc-minmax-wasm.c",
573 "src/f32-gemm/gen-inc/4x4inc-minmax-wasm.c",
574 "src/f32-gemm/gen/1x4-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700575 "src/f32-gemm/gen/1x4-relu-wasm.c",
576 "src/f32-gemm/gen/1x4-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700577 "src/f32-gemm/gen/2x4-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700578 "src/f32-gemm/gen/2x4-relu-wasm.c",
579 "src/f32-gemm/gen/2x4-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700580 "src/f32-gemm/gen/4x2-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700581 "src/f32-gemm/gen/4x2-relu-wasm.c",
582 "src/f32-gemm/gen/4x2-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700583 "src/f32-gemm/gen/4x4-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700584 "src/f32-gemm/gen/4x4-relu-wasm.c",
585 "src/f32-gemm/gen/4x4-wasm.c",
Marat Dukhanefaac272020-07-16 16:58:07 -0700586 "src/f32-hswish/gen/hswish-wasm-x1.c",
587 "src/f32-hswish/gen/hswish-wasm-x2.c",
588 "src/f32-hswish/gen/hswish-wasm-x4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700589 "src/f32-igemm/gen/1x4-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700590 "src/f32-igemm/gen/1x4-relu-wasm.c",
591 "src/f32-igemm/gen/1x4-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700592 "src/f32-igemm/gen/2x4-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700593 "src/f32-igemm/gen/2x4-relu-wasm.c",
594 "src/f32-igemm/gen/2x4-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700595 "src/f32-igemm/gen/4x2-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700596 "src/f32-igemm/gen/4x2-relu-wasm.c",
597 "src/f32-igemm/gen/4x2-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700598 "src/f32-igemm/gen/4x4-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700599 "src/f32-igemm/gen/4x4-relu-wasm.c",
600 "src/f32-igemm/gen/4x4-wasm.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700601 "src/f32-maxpool/9p8x-minmax-wasm-c1.c",
602 "src/f32-pavgpool/9p8x-minmax-wasm-c1.c",
603 "src/f32-pavgpool/9x-minmax-wasm-c1.c",
Marat Dukhan7c1f8082020-06-25 13:26:20 -0700604 "src/f32-prelu/gen/wasm-2x1.c",
605 "src/f32-prelu/gen/wasm-2x4.c",
Frank Barchardfb158e22020-07-15 16:10:10 -0700606 "src/f32-relu/gen/wasm-x1.c",
607 "src/f32-relu/gen/wasm-x2.c",
608 "src/f32-relu/gen/wasm-x4.c",
Frank Barchard37297a62020-08-27 12:15:10 -0700609 "src/f32-relu/gen/wasm-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700610 "src/f32-vbinary/gen/vadd-minmax-wasm-x1.c",
611 "src/f32-vbinary/gen/vadd-minmax-wasm-x2.c",
612 "src/f32-vbinary/gen/vadd-minmax-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700613 "src/f32-vbinary/gen/vadd-minmax-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700614 "src/f32-vbinary/gen/vadd-relu-wasm-x1.c",
615 "src/f32-vbinary/gen/vadd-relu-wasm-x2.c",
616 "src/f32-vbinary/gen/vadd-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700617 "src/f32-vbinary/gen/vadd-relu-wasm-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700618 "src/f32-vbinary/gen/vaddc-minmax-wasm-x1.c",
619 "src/f32-vbinary/gen/vaddc-minmax-wasm-x2.c",
620 "src/f32-vbinary/gen/vaddc-minmax-wasm-x4.c",
621 "src/f32-vbinary/gen/vaddc-minmax-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700622 "src/f32-vbinary/gen/vaddc-relu-wasm-x1.c",
623 "src/f32-vbinary/gen/vaddc-relu-wasm-x2.c",
624 "src/f32-vbinary/gen/vaddc-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700625 "src/f32-vbinary/gen/vaddc-relu-wasm-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700626 "src/f32-vbinary/gen/vdiv-minmax-wasm-x1.c",
627 "src/f32-vbinary/gen/vdiv-minmax-wasm-x2.c",
628 "src/f32-vbinary/gen/vdiv-minmax-wasm-x4.c",
629 "src/f32-vbinary/gen/vdiv-minmax-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700630 "src/f32-vbinary/gen/vdiv-relu-wasm-x1.c",
631 "src/f32-vbinary/gen/vdiv-relu-wasm-x2.c",
632 "src/f32-vbinary/gen/vdiv-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700633 "src/f32-vbinary/gen/vdiv-relu-wasm-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700634 "src/f32-vbinary/gen/vdivc-minmax-wasm-x1.c",
635 "src/f32-vbinary/gen/vdivc-minmax-wasm-x2.c",
636 "src/f32-vbinary/gen/vdivc-minmax-wasm-x4.c",
637 "src/f32-vbinary/gen/vdivc-minmax-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700638 "src/f32-vbinary/gen/vdivc-relu-wasm-x1.c",
639 "src/f32-vbinary/gen/vdivc-relu-wasm-x2.c",
640 "src/f32-vbinary/gen/vdivc-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700641 "src/f32-vbinary/gen/vdivc-relu-wasm-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800642 "src/f32-vbinary/gen/vmax-wasm-x1.c",
643 "src/f32-vbinary/gen/vmax-wasm-x2.c",
644 "src/f32-vbinary/gen/vmax-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700645 "src/f32-vbinary/gen/vmax-wasm-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800646 "src/f32-vbinary/gen/vmaxc-wasm-x1.c",
647 "src/f32-vbinary/gen/vmaxc-wasm-x2.c",
648 "src/f32-vbinary/gen/vmaxc-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700649 "src/f32-vbinary/gen/vmaxc-wasm-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800650 "src/f32-vbinary/gen/vmin-wasm-x1.c",
651 "src/f32-vbinary/gen/vmin-wasm-x2.c",
652 "src/f32-vbinary/gen/vmin-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700653 "src/f32-vbinary/gen/vmin-wasm-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800654 "src/f32-vbinary/gen/vminc-wasm-x1.c",
655 "src/f32-vbinary/gen/vminc-wasm-x2.c",
656 "src/f32-vbinary/gen/vminc-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700657 "src/f32-vbinary/gen/vminc-wasm-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700658 "src/f32-vbinary/gen/vmul-minmax-wasm-x1.c",
659 "src/f32-vbinary/gen/vmul-minmax-wasm-x2.c",
660 "src/f32-vbinary/gen/vmul-minmax-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700661 "src/f32-vbinary/gen/vmul-minmax-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700662 "src/f32-vbinary/gen/vmul-relu-wasm-x1.c",
663 "src/f32-vbinary/gen/vmul-relu-wasm-x2.c",
664 "src/f32-vbinary/gen/vmul-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700665 "src/f32-vbinary/gen/vmul-relu-wasm-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700666 "src/f32-vbinary/gen/vmulc-minmax-wasm-x1.c",
667 "src/f32-vbinary/gen/vmulc-minmax-wasm-x2.c",
668 "src/f32-vbinary/gen/vmulc-minmax-wasm-x4.c",
669 "src/f32-vbinary/gen/vmulc-minmax-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700670 "src/f32-vbinary/gen/vmulc-relu-wasm-x1.c",
671 "src/f32-vbinary/gen/vmulc-relu-wasm-x2.c",
672 "src/f32-vbinary/gen/vmulc-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700673 "src/f32-vbinary/gen/vmulc-relu-wasm-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700674 "src/f32-vbinary/gen/vrdivc-minmax-wasm-x1.c",
675 "src/f32-vbinary/gen/vrdivc-minmax-wasm-x2.c",
676 "src/f32-vbinary/gen/vrdivc-minmax-wasm-x4.c",
677 "src/f32-vbinary/gen/vrdivc-minmax-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700678 "src/f32-vbinary/gen/vrdivc-relu-wasm-x1.c",
679 "src/f32-vbinary/gen/vrdivc-relu-wasm-x2.c",
680 "src/f32-vbinary/gen/vrdivc-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700681 "src/f32-vbinary/gen/vrdivc-relu-wasm-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700682 "src/f32-vbinary/gen/vrsubc-minmax-wasm-x1.c",
683 "src/f32-vbinary/gen/vrsubc-minmax-wasm-x2.c",
684 "src/f32-vbinary/gen/vrsubc-minmax-wasm-x4.c",
685 "src/f32-vbinary/gen/vrsubc-minmax-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700686 "src/f32-vbinary/gen/vrsubc-relu-wasm-x1.c",
687 "src/f32-vbinary/gen/vrsubc-relu-wasm-x2.c",
688 "src/f32-vbinary/gen/vrsubc-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700689 "src/f32-vbinary/gen/vrsubc-relu-wasm-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700690 "src/f32-vbinary/gen/vsub-minmax-wasm-x1.c",
691 "src/f32-vbinary/gen/vsub-minmax-wasm-x2.c",
692 "src/f32-vbinary/gen/vsub-minmax-wasm-x4.c",
693 "src/f32-vbinary/gen/vsub-minmax-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700694 "src/f32-vbinary/gen/vsub-relu-wasm-x1.c",
695 "src/f32-vbinary/gen/vsub-relu-wasm-x2.c",
696 "src/f32-vbinary/gen/vsub-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700697 "src/f32-vbinary/gen/vsub-relu-wasm-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700698 "src/f32-vbinary/gen/vsubc-minmax-wasm-x1.c",
699 "src/f32-vbinary/gen/vsubc-minmax-wasm-x2.c",
700 "src/f32-vbinary/gen/vsubc-minmax-wasm-x4.c",
701 "src/f32-vbinary/gen/vsubc-minmax-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700702 "src/f32-vbinary/gen/vsubc-relu-wasm-x1.c",
703 "src/f32-vbinary/gen/vsubc-relu-wasm-x2.c",
704 "src/f32-vbinary/gen/vsubc-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700705 "src/f32-vbinary/gen/vsubc-relu-wasm-x8.c",
Marat Dukhaned6baaf2020-12-01 15:07:08 -0800706 "src/f32-velu/gen/velu-wasm-rr2-lut16-p3-x1.c",
707 "src/f32-velu/gen/velu-wasm-rr2-lut16-p3-x2.c",
708 "src/f32-velu/gen/velu-wasm-rr2-lut16-p3-x3.c",
709 "src/f32-velu/gen/velu-wasm-rr2-lut16-p3-x4.c",
710 "src/f32-velu/gen/velu-wasm-rr2-lut16-p3-x5.c",
711 "src/f32-velu/gen/velu-wasm-rr2-lut16-p3-x6.c",
712 "src/f32-velu/gen/velu-wasm-rr2-p6-x1.c",
713 "src/f32-velu/gen/velu-wasm-rr2-p6-x2.c",
714 "src/f32-velu/gen/velu-wasm-rr2-p6-x3.c",
715 "src/f32-velu/gen/velu-wasm-rr2-p6-x4.c",
716 "src/f32-velu/gen/velu-wasm-rr2-p6-x5.c",
717 "src/f32-velu/gen/velu-wasm-rr2-p6-x6.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -0700718 "src/f32-vlrelu/gen/vlrelu-wasm-x1.c",
719 "src/f32-vlrelu/gen/vlrelu-wasm-x2.c",
720 "src/f32-vlrelu/gen/vlrelu-wasm-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700721 "src/f32-vmulcaddc/gen/c1-minmax-wasm-2x.c",
722 "src/f32-vmulcaddc/gen/c2-minmax-wasm-2x.c",
723 "src/f32-vmulcaddc/gen/c4-minmax-wasm-2x.c",
Marat Dukhan436ebe62019-12-04 15:10:12 -0800724]
725
Marat Dukhan290055c2020-06-09 12:24:29 -0700726WASMSIMD_UKERNELS = [
Marat Dukhan40f05522020-07-16 22:33:12 -0700727 "src/f32-argmaxpool/4x-wasmsimd-c4.c",
728 "src/f32-argmaxpool/9p8x-wasmsimd-c4.c",
729 "src/f32-argmaxpool/9x-wasmsimd-c4.c",
Marat Dukhan3b7432d2020-07-16 17:46:32 -0700730 "src/f32-avgpool/9p8x-minmax-wasmsimd-arm-c4.c",
731 "src/f32-avgpool/9p8x-minmax-wasmsimd-x86-c4.c",
732 "src/f32-avgpool/9x-minmax-wasmsimd-arm-c4.c",
733 "src/f32-avgpool/9x-minmax-wasmsimd-x86-c4.c",
Marat Dukhan3fa52c82020-07-08 12:54:33 -0700734 "src/f32-clamp/gen/wasmsimd-arm-x4.c",
735 "src/f32-clamp/gen/wasmsimd-arm-x8.c",
736 "src/f32-clamp/gen/wasmsimd-x86-x4.c",
737 "src/f32-clamp/gen/wasmsimd-x86-x8.c",
Frank Barchard22136062020-11-24 18:44:46 -0800738 "src/f32-conv-hwc2chw/3x3s2p1c3x4-wasmsimd-2x2.c",
Frank Barchard0725b8d2020-12-07 11:07:35 -0800739 "src/f32-dwconv/gen/up4x4-minmax-wasmsimd-arm-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700740 "src/f32-dwconv/gen/up4x4-minmax-wasmsimd-arm.c",
Frank Barchard0725b8d2020-12-07 11:07:35 -0800741 "src/f32-dwconv/gen/up4x4-minmax-wasmsimd-x86-acc2.c",
Marat Dukhanac014d72020-06-16 08:36:47 -0700742 "src/f32-dwconv/gen/up4x4-minmax-wasmsimd-x86.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700743 "src/f32-dwconv/gen/up4x4-wasmsimd.c",
Frank Barchard0725b8d2020-12-07 11:07:35 -0800744 "src/f32-dwconv/gen/up4x9-minmax-wasmsimd-arm-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700745 "src/f32-dwconv/gen/up4x9-minmax-wasmsimd-arm.c",
Frank Barchard0725b8d2020-12-07 11:07:35 -0800746 "src/f32-dwconv/gen/up4x9-minmax-wasmsimd-x86-acc2.c",
Marat Dukhanac014d72020-06-16 08:36:47 -0700747 "src/f32-dwconv/gen/up4x9-minmax-wasmsimd-x86.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700748 "src/f32-dwconv/gen/up4x9-wasmsimd.c",
Frank Barchard0725b8d2020-12-07 11:07:35 -0800749 "src/f32-dwconv/gen/up4x25-minmax-wasmsimd-arm-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700750 "src/f32-dwconv/gen/up4x25-minmax-wasmsimd-arm.c",
Frank Barchard0725b8d2020-12-07 11:07:35 -0800751 "src/f32-dwconv/gen/up4x25-minmax-wasmsimd-x86-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700752 "src/f32-dwconv/gen/up4x25-minmax-wasmsimd-x86.c",
753 "src/f32-dwconv/gen/up4x25-wasmsimd.c",
Frank Barchard0725b8d2020-12-07 11:07:35 -0800754 "src/f32-dwconv/gen/up8x4-minmax-wasmsimd-arm-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700755 "src/f32-dwconv/gen/up8x4-minmax-wasmsimd-arm.c",
Frank Barchard0725b8d2020-12-07 11:07:35 -0800756 "src/f32-dwconv/gen/up8x4-minmax-wasmsimd-x86-acc2.c",
Marat Dukhanac014d72020-06-16 08:36:47 -0700757 "src/f32-dwconv/gen/up8x4-minmax-wasmsimd-x86.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700758 "src/f32-dwconv/gen/up8x4-wasmsimd.c",
Frank Barchard0725b8d2020-12-07 11:07:35 -0800759 "src/f32-dwconv/gen/up8x9-minmax-wasmsimd-arm-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700760 "src/f32-dwconv/gen/up8x9-minmax-wasmsimd-arm.c",
Frank Barchard0725b8d2020-12-07 11:07:35 -0800761 "src/f32-dwconv/gen/up8x9-minmax-wasmsimd-x86-acc2.c",
Marat Dukhanac014d72020-06-16 08:36:47 -0700762 "src/f32-dwconv/gen/up8x9-minmax-wasmsimd-x86.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700763 "src/f32-dwconv/gen/up8x9-wasmsimd.c",
Frank Barchard0725b8d2020-12-07 11:07:35 -0800764 "src/f32-dwconv/gen/up8x25-minmax-wasmsimd-arm-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700765 "src/f32-dwconv/gen/up8x25-minmax-wasmsimd-arm.c",
Frank Barchard0725b8d2020-12-07 11:07:35 -0800766 "src/f32-dwconv/gen/up8x25-minmax-wasmsimd-x86-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700767 "src/f32-dwconv/gen/up8x25-minmax-wasmsimd-x86.c",
768 "src/f32-dwconv/gen/up8x25-wasmsimd.c",
Frank Barchard412e2f42020-12-11 11:40:50 -0800769 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-loadsplat-1x4-acc2.c",
770 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-loadsplat-1x4-acc3.c",
771 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-loadsplat-1x4-acc4.c",
772 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-loadsplat-1x4.c",
773 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-loadsplat-2x4-acc2.c",
774 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-loadsplat-2x4.c",
775 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-loadsplat-3x4.c",
776 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-loadsplat-4x4.c",
777 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-loadsplat-5x4.c",
778 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-loadsplat-6x4.c",
Frank Barchard02bb4292020-12-15 18:25:32 -0800779 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-splat-1x4-acc2.c",
780 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-splat-1x4-acc3.c",
781 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-splat-1x4-acc4.c",
782 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-splat-1x4.c",
783 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-splat-2x4-acc2.c",
784 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-splat-2x4.c",
785 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-splat-3x4.c",
786 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-splat-4x4.c",
787 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-splat-5x4.c",
788 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-splat-6x4.c",
Frank Barchard412e2f42020-12-11 11:40:50 -0800789 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-loadsplat-1x4-acc2.c",
790 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-loadsplat-1x4-acc3.c",
791 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-loadsplat-1x4-acc4.c",
792 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-loadsplat-1x4.c",
793 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-loadsplat-2x4-acc2.c",
794 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-loadsplat-2x4.c",
795 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-loadsplat-3x4.c",
796 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-loadsplat-4x4.c",
797 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-loadsplat-5x4.c",
798 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-loadsplat-6x4.c",
Frank Barchard02bb4292020-12-15 18:25:32 -0800799 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-splat-1x4-acc2.c",
800 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-splat-1x4-acc3.c",
801 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-splat-1x4-acc4.c",
802 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-splat-1x4.c",
803 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-splat-2x4-acc2.c",
804 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-splat-2x4.c",
805 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-splat-3x4.c",
806 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-splat-4x4.c",
807 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-splat-5x4.c",
808 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-splat-6x4.c",
Frank Barchardc5704bf2020-12-21 23:09:00 -0800809 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-arm-loadsplat-1x4-acc2.c",
810 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-arm-loadsplat-1x4-acc3.c",
811 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-arm-loadsplat-1x4-acc4.c",
812 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-arm-loadsplat-1x4.c",
813 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-arm-loadsplat-2x4-acc2.c",
814 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-arm-loadsplat-2x4.c",
815 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-arm-loadsplat-3x4.c",
816 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-arm-loadsplat-4x4.c",
Frank Barchard412e2f42020-12-11 11:40:50 -0800817 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-arm-splat-1x4-acc2.c",
818 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-arm-splat-1x4-acc3.c",
819 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-arm-splat-1x4-acc4.c",
820 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-arm-splat-1x4.c",
821 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-arm-splat-2x4-acc2.c",
822 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-arm-splat-2x4.c",
823 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-arm-splat-3x4.c",
824 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-arm-splat-4x4.c",
Frank Barchardcadd4222021-01-20 16:27:25 -0800825 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-x86-loadsplat-1x4-acc2.c",
826 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-x86-loadsplat-1x4-acc3.c",
827 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-x86-loadsplat-1x4-acc4.c",
828 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-x86-loadsplat-1x4.c",
829 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-x86-loadsplat-2x4-acc2.c",
830 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-x86-loadsplat-2x4.c",
831 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-x86-loadsplat-3x4.c",
832 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-x86-loadsplat-4x4.c",
Frank Barchard412e2f42020-12-11 11:40:50 -0800833 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-x86-splat-1x4-acc2.c",
834 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-x86-splat-1x4-acc3.c",
835 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-x86-splat-1x4-acc4.c",
836 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-x86-splat-1x4.c",
837 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-x86-splat-2x4-acc2.c",
838 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-x86-splat-2x4.c",
839 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-x86-splat-3x4.c",
840 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-x86-splat-4x4.c",
Frank Barchardb20dcd62020-12-15 16:46:14 -0800841 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-loadsplat-1x4-acc2.c",
842 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-loadsplat-1x4-acc3.c",
843 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-loadsplat-1x4-acc4.c",
844 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-loadsplat-1x4-acc5.c",
845 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-loadsplat-1x4.c",
846 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-loadsplat-2x4-acc2.c",
847 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-loadsplat-2x4-acc3.c",
848 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-loadsplat-2x4.c",
849 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-loadsplat-3x4-acc2.c",
850 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-loadsplat-3x4.c",
851 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-loadsplat-4x4-acc2.c",
852 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-loadsplat-4x4.c",
853 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-loadsplat-5x4.c",
Frank Barchard412e2f42020-12-11 11:40:50 -0800854 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-splat-1x4-acc2.c",
855 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-splat-1x4-acc3.c",
856 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-splat-1x4-acc4.c",
857 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-splat-1x4-acc5.c",
858 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-splat-1x4.c",
859 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-splat-2x4-acc2.c",
860 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-splat-2x4-acc3.c",
861 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-splat-2x4.c",
862 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-splat-3x4-acc2.c",
863 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-splat-3x4.c",
864 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-splat-4x4-acc2.c",
865 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-splat-4x4.c",
866 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-splat-5x4.c",
Frank Barchardb20dcd62020-12-15 16:46:14 -0800867 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-loadsplat-1x4-acc2.c",
868 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-loadsplat-1x4-acc3.c",
869 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-loadsplat-1x4-acc4.c",
870 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-loadsplat-1x4-acc5.c",
871 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-loadsplat-1x4.c",
872 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-loadsplat-2x4-acc2.c",
873 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-loadsplat-2x4-acc3.c",
874 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-loadsplat-2x4.c",
875 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-loadsplat-3x4-acc2.c",
876 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-loadsplat-3x4.c",
877 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-loadsplat-4x4-acc2.c",
878 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-loadsplat-4x4.c",
879 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-loadsplat-5x4.c",
Frank Barchard412e2f42020-12-11 11:40:50 -0800880 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-splat-1x4-acc2.c",
881 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-splat-1x4-acc3.c",
882 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-splat-1x4-acc4.c",
883 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-splat-1x4-acc5.c",
884 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-splat-1x4.c",
885 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-splat-2x4-acc2.c",
886 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-splat-2x4-acc3.c",
887 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-splat-2x4.c",
888 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-splat-3x4-acc2.c",
889 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-splat-3x4.c",
890 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-splat-4x4-acc2.c",
891 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-splat-4x4.c",
892 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-splat-5x4.c",
Frank Barchardc6889b32020-12-21 11:27:22 -0800893 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-arm-loadsplat-1x4-acc2.c",
894 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-arm-loadsplat-1x4-acc3.c",
895 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-arm-loadsplat-1x4-acc4.c",
896 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-arm-loadsplat-1x4-acc5.c",
897 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-arm-loadsplat-1x4.c",
898 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-arm-loadsplat-2x4-acc2.c",
899 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-arm-loadsplat-2x4-acc3.c",
900 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-arm-loadsplat-2x4.c",
901 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-arm-loadsplat-3x4-acc2.c",
902 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-arm-loadsplat-3x4.c",
Frank Barchard412e2f42020-12-11 11:40:50 -0800903 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-arm-splat-1x4-acc2.c",
904 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-arm-splat-1x4-acc3.c",
905 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-arm-splat-1x4-acc4.c",
906 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-arm-splat-1x4-acc5.c",
907 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-arm-splat-1x4.c",
908 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-arm-splat-2x4-acc2.c",
909 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-arm-splat-2x4-acc3.c",
910 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-arm-splat-2x4.c",
911 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-arm-splat-3x4-acc2.c",
912 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-arm-splat-3x4.c",
Frank Barchardc6889b32020-12-21 11:27:22 -0800913 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-x86-loadsplat-1x4-acc2.c",
914 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-x86-loadsplat-1x4-acc3.c",
915 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-x86-loadsplat-1x4-acc4.c",
916 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-x86-loadsplat-1x4-acc5.c",
917 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-x86-loadsplat-1x4.c",
918 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-x86-loadsplat-2x4-acc2.c",
919 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-x86-loadsplat-2x4-acc3.c",
920 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-x86-loadsplat-2x4.c",
921 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-x86-loadsplat-3x4-acc2.c",
922 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-x86-loadsplat-3x4.c",
Frank Barchard412e2f42020-12-11 11:40:50 -0800923 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-x86-splat-1x4-acc2.c",
924 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-x86-splat-1x4-acc3.c",
925 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-x86-splat-1x4-acc4.c",
926 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-x86-splat-1x4-acc5.c",
927 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-x86-splat-1x4.c",
928 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-x86-splat-2x4-acc2.c",
929 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-x86-splat-2x4-acc3.c",
930 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-x86-splat-2x4.c",
931 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-x86-splat-3x4-acc2.c",
932 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-x86-splat-3x4.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700933 "src/f32-gavgpool-cw/wasmsimd-arm-x4.c",
934 "src/f32-gavgpool-cw/wasmsimd-x86-x4.c",
Marat Dukhanc6016802020-07-16 18:51:28 -0700935 "src/f32-gavgpool/7p7x-minmax-wasmsimd-arm-c4.c",
936 "src/f32-gavgpool/7p7x-minmax-wasmsimd-x86-c4.c",
937 "src/f32-gavgpool/7x-minmax-wasmsimd-arm-c4.c",
938 "src/f32-gavgpool/7x-minmax-wasmsimd-x86-c4.c",
Frank Barchard0725b8d2020-12-07 11:07:35 -0800939 "src/f32-gemm/gen-inc/1x8inc-minmax-wasmsimd-arm-loadsplat.c",
940 "src/f32-gemm/gen-inc/1x8inc-minmax-wasmsimd-arm-splat.c",
941 "src/f32-gemm/gen-inc/1x8inc-minmax-wasmsimd-x86-loadsplat.c",
942 "src/f32-gemm/gen-inc/1x8inc-minmax-wasmsimd-x86-splat.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700943 "src/f32-gemm/gen-inc/1x8s4inc-minmax-wasmsimd-arm.c",
944 "src/f32-gemm/gen-inc/1x8s4inc-minmax-wasmsimd-x86.c",
Frank Barchard0725b8d2020-12-07 11:07:35 -0800945 "src/f32-gemm/gen-inc/3x8inc-minmax-wasmsimd-arm-loadsplat.c",
946 "src/f32-gemm/gen-inc/3x8inc-minmax-wasmsimd-arm-splat.c",
947 "src/f32-gemm/gen-inc/3x8inc-minmax-wasmsimd-x86-loadsplat.c",
948 "src/f32-gemm/gen-inc/3x8inc-minmax-wasmsimd-x86-splat.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700949 "src/f32-gemm/gen-inc/3x8s4inc-minmax-wasmsimd-arm.c",
950 "src/f32-gemm/gen-inc/3x8s4inc-minmax-wasmsimd-x86.c",
Frank Barchard0725b8d2020-12-07 11:07:35 -0800951 "src/f32-gemm/gen-inc/4x8inc-minmax-wasmsimd-arm-loadsplat.c",
952 "src/f32-gemm/gen-inc/4x8inc-minmax-wasmsimd-arm-splat.c",
953 "src/f32-gemm/gen-inc/4x8inc-minmax-wasmsimd-x86-loadsplat.c",
954 "src/f32-gemm/gen-inc/4x8inc-minmax-wasmsimd-x86-splat.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700955 "src/f32-gemm/gen-inc/4x8s4inc-minmax-wasmsimd-arm.c",
956 "src/f32-gemm/gen-inc/4x8s4inc-minmax-wasmsimd-x86.c",
Frank Barchard0725b8d2020-12-07 11:07:35 -0800957 "src/f32-gemm/gen-inc/5x8inc-minmax-wasmsimd-arm-loadsplat.c",
958 "src/f32-gemm/gen-inc/5x8inc-minmax-wasmsimd-arm-splat.c",
959 "src/f32-gemm/gen-inc/5x8inc-minmax-wasmsimd-x86-loadsplat.c",
960 "src/f32-gemm/gen-inc/5x8inc-minmax-wasmsimd-x86-splat.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700961 "src/f32-gemm/gen-inc/5x8s4inc-minmax-wasmsimd-arm.c",
962 "src/f32-gemm/gen-inc/5x8s4inc-minmax-wasmsimd-x86.c",
Frank Barchard0725b8d2020-12-07 11:07:35 -0800963 "src/f32-gemm/gen-inc/6x8inc-minmax-wasmsimd-arm-loadsplat.c",
964 "src/f32-gemm/gen-inc/6x8inc-minmax-wasmsimd-arm-splat.c",
965 "src/f32-gemm/gen-inc/6x8inc-minmax-wasmsimd-x86-loadsplat.c",
966 "src/f32-gemm/gen-inc/6x8inc-minmax-wasmsimd-x86-splat.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700967 "src/f32-gemm/gen-inc/6x8s4inc-minmax-wasmsimd-arm.c",
968 "src/f32-gemm/gen-inc/6x8s4inc-minmax-wasmsimd-x86.c",
Frank Barchard0725b8d2020-12-07 11:07:35 -0800969 "src/f32-gemm/gen/1x8-minmax-wasmsimd-arm-loadsplat.c",
970 "src/f32-gemm/gen/1x8-minmax-wasmsimd-arm-splat.c",
971 "src/f32-gemm/gen/1x8-minmax-wasmsimd-x86-loadsplat.c",
972 "src/f32-gemm/gen/1x8-minmax-wasmsimd-x86-splat.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700973 "src/f32-gemm/gen/1x8-relu-wasmsimd-splat.c",
974 "src/f32-gemm/gen/1x8-wasmsimd-splat.c",
975 "src/f32-gemm/gen/1x8s4-minmax-wasmsimd-arm.c",
976 "src/f32-gemm/gen/1x8s4-minmax-wasmsimd-x86.c",
Frank Barchard0725b8d2020-12-07 11:07:35 -0800977 "src/f32-gemm/gen/3x8-minmax-wasmsimd-arm-loadsplat.c",
978 "src/f32-gemm/gen/3x8-minmax-wasmsimd-arm-splat.c",
979 "src/f32-gemm/gen/3x8-minmax-wasmsimd-x86-loadsplat.c",
980 "src/f32-gemm/gen/3x8-minmax-wasmsimd-x86-splat.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700981 "src/f32-gemm/gen/3x8s4-minmax-wasmsimd-arm.c",
982 "src/f32-gemm/gen/3x8s4-minmax-wasmsimd-x86.c",
983 "src/f32-gemm/gen/4x2c4-minmax-wasmsimd-arm.c",
984 "src/f32-gemm/gen/4x2c4-minmax-wasmsimd-x86.c",
985 "src/f32-gemm/gen/4x2c4-relu-wasmsimd.c",
986 "src/f32-gemm/gen/4x2c4-wasmsimd.c",
Frank Barchard0725b8d2020-12-07 11:07:35 -0800987 "src/f32-gemm/gen/4x8-minmax-wasmsimd-arm-loadsplat.c",
988 "src/f32-gemm/gen/4x8-minmax-wasmsimd-arm-splat.c",
989 "src/f32-gemm/gen/4x8-minmax-wasmsimd-x86-loadsplat.c",
990 "src/f32-gemm/gen/4x8-minmax-wasmsimd-x86-splat.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700991 "src/f32-gemm/gen/4x8-relu-wasmsimd-splat.c",
992 "src/f32-gemm/gen/4x8-wasmsimd-splat.c",
993 "src/f32-gemm/gen/4x8s4-minmax-wasmsimd-arm.c",
994 "src/f32-gemm/gen/4x8s4-minmax-wasmsimd-x86.c",
Frank Barchard0725b8d2020-12-07 11:07:35 -0800995 "src/f32-gemm/gen/5x8-minmax-wasmsimd-arm-loadsplat.c",
996 "src/f32-gemm/gen/5x8-minmax-wasmsimd-arm-splat.c",
997 "src/f32-gemm/gen/5x8-minmax-wasmsimd-x86-loadsplat.c",
998 "src/f32-gemm/gen/5x8-minmax-wasmsimd-x86-splat.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700999 "src/f32-gemm/gen/5x8-relu-wasmsimd-splat.c",
1000 "src/f32-gemm/gen/5x8-wasmsimd-splat.c",
1001 "src/f32-gemm/gen/5x8s4-minmax-wasmsimd-arm.c",
1002 "src/f32-gemm/gen/5x8s4-minmax-wasmsimd-x86.c",
Frank Barchard0725b8d2020-12-07 11:07:35 -08001003 "src/f32-gemm/gen/6x8-minmax-wasmsimd-arm-loadsplat.c",
1004 "src/f32-gemm/gen/6x8-minmax-wasmsimd-arm-splat.c",
1005 "src/f32-gemm/gen/6x8-minmax-wasmsimd-x86-loadsplat.c",
1006 "src/f32-gemm/gen/6x8-minmax-wasmsimd-x86-splat.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001007 "src/f32-gemm/gen/6x8s4-minmax-wasmsimd-arm.c",
1008 "src/f32-gemm/gen/6x8s4-minmax-wasmsimd-x86.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07001009 "src/f32-hswish/gen/hswish-wasmsimd-x4.c",
1010 "src/f32-hswish/gen/hswish-wasmsimd-x8.c",
1011 "src/f32-hswish/gen/hswish-wasmsimd-x16.c",
XNNPACK Team965272b2020-10-23 21:10:15 -07001012 "src/f32-ibilinear-chw/gen/wasmsimd-p4.c",
1013 "src/f32-ibilinear-chw/gen/wasmsimd-p8.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001014 "src/f32-ibilinear/gen/wasmsimd-c4.c",
1015 "src/f32-ibilinear/gen/wasmsimd-c8.c",
Frank Barchard0725b8d2020-12-07 11:07:35 -08001016 "src/f32-igemm/gen/1x8-minmax-wasmsimd-arm-loadsplat.c",
1017 "src/f32-igemm/gen/1x8-minmax-wasmsimd-arm-splat.c",
1018 "src/f32-igemm/gen/1x8-minmax-wasmsimd-x86-loadsplat.c",
1019 "src/f32-igemm/gen/1x8-minmax-wasmsimd-x86-splat.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001020 "src/f32-igemm/gen/1x8-relu-wasmsimd-splat.c",
1021 "src/f32-igemm/gen/1x8-wasmsimd-splat.c",
1022 "src/f32-igemm/gen/1x8s4-minmax-wasmsimd-arm.c",
1023 "src/f32-igemm/gen/1x8s4-minmax-wasmsimd-x86.c",
Frank Barchard0725b8d2020-12-07 11:07:35 -08001024 "src/f32-igemm/gen/3x8-minmax-wasmsimd-arm-loadsplat.c",
1025 "src/f32-igemm/gen/3x8-minmax-wasmsimd-arm-splat.c",
1026 "src/f32-igemm/gen/3x8-minmax-wasmsimd-x86-loadsplat.c",
1027 "src/f32-igemm/gen/3x8-minmax-wasmsimd-x86-splat.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001028 "src/f32-igemm/gen/3x8s4-minmax-wasmsimd-arm.c",
1029 "src/f32-igemm/gen/3x8s4-minmax-wasmsimd-x86.c",
1030 "src/f32-igemm/gen/4x2c4-minmax-wasmsimd-arm.c",
1031 "src/f32-igemm/gen/4x2c4-minmax-wasmsimd-x86.c",
1032 "src/f32-igemm/gen/4x2c4-relu-wasmsimd.c",
1033 "src/f32-igemm/gen/4x2c4-wasmsimd.c",
Frank Barchard0725b8d2020-12-07 11:07:35 -08001034 "src/f32-igemm/gen/4x8-minmax-wasmsimd-arm-loadsplat.c",
1035 "src/f32-igemm/gen/4x8-minmax-wasmsimd-arm-splat.c",
1036 "src/f32-igemm/gen/4x8-minmax-wasmsimd-x86-loadsplat.c",
1037 "src/f32-igemm/gen/4x8-minmax-wasmsimd-x86-splat.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001038 "src/f32-igemm/gen/4x8-relu-wasmsimd-splat.c",
1039 "src/f32-igemm/gen/4x8-wasmsimd-splat.c",
1040 "src/f32-igemm/gen/4x8s4-minmax-wasmsimd-arm.c",
1041 "src/f32-igemm/gen/4x8s4-minmax-wasmsimd-x86.c",
Frank Barchard0725b8d2020-12-07 11:07:35 -08001042 "src/f32-igemm/gen/5x8-minmax-wasmsimd-arm-loadsplat.c",
1043 "src/f32-igemm/gen/5x8-minmax-wasmsimd-arm-splat.c",
1044 "src/f32-igemm/gen/5x8-minmax-wasmsimd-x86-loadsplat.c",
1045 "src/f32-igemm/gen/5x8-minmax-wasmsimd-x86-splat.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001046 "src/f32-igemm/gen/5x8-relu-wasmsimd-splat.c",
1047 "src/f32-igemm/gen/5x8-wasmsimd-splat.c",
1048 "src/f32-igemm/gen/5x8s4-minmax-wasmsimd-arm.c",
1049 "src/f32-igemm/gen/5x8s4-minmax-wasmsimd-x86.c",
Frank Barchard0725b8d2020-12-07 11:07:35 -08001050 "src/f32-igemm/gen/6x8-minmax-wasmsimd-arm-loadsplat.c",
1051 "src/f32-igemm/gen/6x8-minmax-wasmsimd-arm-splat.c",
1052 "src/f32-igemm/gen/6x8-minmax-wasmsimd-x86-loadsplat.c",
1053 "src/f32-igemm/gen/6x8-minmax-wasmsimd-x86-splat.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001054 "src/f32-igemm/gen/6x8s4-minmax-wasmsimd-arm.c",
1055 "src/f32-igemm/gen/6x8s4-minmax-wasmsimd-x86.c",
Marat Dukhanf6e24802020-07-08 22:20:40 -07001056 "src/f32-maxpool/9p8x-minmax-wasmsimd-arm-c4.c",
1057 "src/f32-maxpool/9p8x-minmax-wasmsimd-x86-c4.c",
Marat Dukhan1483c532020-07-16 18:08:19 -07001058 "src/f32-pavgpool/9p8x-minmax-wasmsimd-arm-c4.c",
1059 "src/f32-pavgpool/9p8x-minmax-wasmsimd-x86-c4.c",
1060 "src/f32-pavgpool/9x-minmax-wasmsimd-arm-c4.c",
1061 "src/f32-pavgpool/9x-minmax-wasmsimd-x86-c4.c",
Frank Barchard0725b8d2020-12-07 11:07:35 -08001062 "src/f32-ppmm/gen/4x8-minmax-wasmsimd-arm-splat.c",
1063 "src/f32-ppmm/gen/4x8-minmax-wasmsimd-x86-splat.c",
Frank Barcharda5316982020-07-23 13:19:28 -07001064 "src/f32-prelu/gen/wasmsimd-bitselect-1x4.c",
1065 "src/f32-prelu/gen/wasmsimd-bitselect-1x8.c",
1066 "src/f32-prelu/gen/wasmsimd-bitselect-1x16.c",
Marat Dukhan195f8eb2020-06-25 12:50:57 -07001067 "src/f32-prelu/gen/wasmsimd-bitselect-2x4.c",
1068 "src/f32-prelu/gen/wasmsimd-bitselect-2x8.c",
Frank Barcharda5316982020-07-23 13:19:28 -07001069 "src/f32-prelu/gen/wasmsimd-bitselect-2x16.c",
1070 "src/f32-prelu/gen/wasmsimd-bitselect-4x4.c",
1071 "src/f32-prelu/gen/wasmsimd-bitselect-4x8.c",
1072 "src/f32-prelu/gen/wasmsimd-bitselect-4x16.c",
1073 "src/f32-prelu/gen/wasmsimd-minmax-1x4.c",
1074 "src/f32-prelu/gen/wasmsimd-minmax-1x8.c",
1075 "src/f32-prelu/gen/wasmsimd-minmax-1x16.c",
Marat Dukhan195f8eb2020-06-25 12:50:57 -07001076 "src/f32-prelu/gen/wasmsimd-minmax-2x4.c",
1077 "src/f32-prelu/gen/wasmsimd-minmax-2x8.c",
Frank Barcharda5316982020-07-23 13:19:28 -07001078 "src/f32-prelu/gen/wasmsimd-minmax-2x16.c",
1079 "src/f32-prelu/gen/wasmsimd-minmax-4x4.c",
1080 "src/f32-prelu/gen/wasmsimd-minmax-4x8.c",
1081 "src/f32-prelu/gen/wasmsimd-minmax-4x16.c",
Marat Dukhan52238f02020-07-16 15:30:28 -07001082 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x4.c",
Marat Dukhan52238f02020-07-16 15:30:28 -07001083 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x8-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001084 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x8.c",
Marat Dukhan52238f02020-07-16 15:30:28 -07001085 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x12-acc2.c",
1086 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x12-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001087 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x12.c",
Marat Dukhan52238f02020-07-16 15:30:28 -07001088 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x16-acc2.c",
1089 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x16-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001090 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x16.c",
Marat Dukhan52238f02020-07-16 15:30:28 -07001091 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x20-acc2.c",
1092 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x20-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001093 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x20.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07001094 "src/f32-relu/gen/wasmsimd-x4.c",
1095 "src/f32-relu/gen/wasmsimd-x8.c",
Frank Barchard37297a62020-08-27 12:15:10 -07001096 "src/f32-relu/gen/wasmsimd-x16.c",
Marat Dukhan8c417962020-07-08 12:27:50 -07001097 "src/f32-rmax/wasmsimd-arm.c",
1098 "src/f32-rmax/wasmsimd-x86.c",
Marat Dukhand187a5b2020-07-20 01:07:17 -07001099 "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x4.c",
1100 "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x8.c",
1101 "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x12.c",
1102 "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x16.c",
1103 "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x20.c",
1104 "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x24.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001105 "src/f32-sigmoid/gen/wasmsimd-p5-div-x4.c",
1106 "src/f32-sigmoid/gen/wasmsimd-p5-div-x8.c",
1107 "src/f32-sigmoid/gen/wasmsimd-p5-div-x12.c",
1108 "src/f32-sigmoid/gen/wasmsimd-p5-div-x16.c",
1109 "src/f32-sigmoid/gen/wasmsimd-p5-div-x20.c",
1110 "src/f32-sigmoid/gen/wasmsimd-p5-div-x24.c",
Frank Barchard8ef44cd2020-11-03 12:30:23 -08001111 "src/f32-spmm/gen/4x1-minmax-wasmsimd-arm-pipelined-x2.c",
1112 "src/f32-spmm/gen/4x1-minmax-wasmsimd-arm-pipelined.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07001113 "src/f32-spmm/gen/4x1-minmax-wasmsimd-arm-x2.c",
1114 "src/f32-spmm/gen/4x1-minmax-wasmsimd-arm-x4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001115 "src/f32-spmm/gen/4x1-minmax-wasmsimd-arm.c",
Frank Barchard8ef44cd2020-11-03 12:30:23 -08001116 "src/f32-spmm/gen/4x1-minmax-wasmsimd-x86-pipelined-x2.c",
1117 "src/f32-spmm/gen/4x1-minmax-wasmsimd-x86-pipelined.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07001118 "src/f32-spmm/gen/4x1-minmax-wasmsimd-x86-x2.c",
1119 "src/f32-spmm/gen/4x1-minmax-wasmsimd-x86-x4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001120 "src/f32-spmm/gen/4x1-minmax-wasmsimd-x86.c",
Frank Barchard8ef44cd2020-11-03 12:30:23 -08001121 "src/f32-spmm/gen/8x1-minmax-wasmsimd-arm-pipelined-x2.c",
1122 "src/f32-spmm/gen/8x1-minmax-wasmsimd-arm-pipelined.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07001123 "src/f32-spmm/gen/8x1-minmax-wasmsimd-arm-x2.c",
1124 "src/f32-spmm/gen/8x1-minmax-wasmsimd-arm-x4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001125 "src/f32-spmm/gen/8x1-minmax-wasmsimd-arm.c",
Frank Barchard8ef44cd2020-11-03 12:30:23 -08001126 "src/f32-spmm/gen/8x1-minmax-wasmsimd-x86-pipelined-x2.c",
1127 "src/f32-spmm/gen/8x1-minmax-wasmsimd-x86-pipelined.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07001128 "src/f32-spmm/gen/8x1-minmax-wasmsimd-x86-x2.c",
1129 "src/f32-spmm/gen/8x1-minmax-wasmsimd-x86-x4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001130 "src/f32-spmm/gen/8x1-minmax-wasmsimd-x86.c",
Frank Barchard8ef44cd2020-11-03 12:30:23 -08001131 "src/f32-spmm/gen/16x1-minmax-wasmsimd-arm-pipelined-x2.c",
1132 "src/f32-spmm/gen/16x1-minmax-wasmsimd-arm-pipelined.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07001133 "src/f32-spmm/gen/16x1-minmax-wasmsimd-arm-x2.c",
1134 "src/f32-spmm/gen/16x1-minmax-wasmsimd-arm-x4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001135 "src/f32-spmm/gen/16x1-minmax-wasmsimd-arm.c",
Frank Barchard8ef44cd2020-11-03 12:30:23 -08001136 "src/f32-spmm/gen/16x1-minmax-wasmsimd-x86-pipelined-x2.c",
1137 "src/f32-spmm/gen/16x1-minmax-wasmsimd-x86-pipelined.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07001138 "src/f32-spmm/gen/16x1-minmax-wasmsimd-x86-x2.c",
1139 "src/f32-spmm/gen/16x1-minmax-wasmsimd-x86-x4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001140 "src/f32-spmm/gen/16x1-minmax-wasmsimd-x86.c",
Frank Barchard8ef44cd2020-11-03 12:30:23 -08001141 "src/f32-spmm/gen/32x1-minmax-wasmsimd-arm-pipelined-x2.c",
1142 "src/f32-spmm/gen/32x1-minmax-wasmsimd-arm-pipelined.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07001143 "src/f32-spmm/gen/32x1-minmax-wasmsimd-arm-x2.c",
1144 "src/f32-spmm/gen/32x1-minmax-wasmsimd-arm-x4.c",
Frank Barchard846c0c62020-10-26 15:01:39 -07001145 "src/f32-spmm/gen/32x1-minmax-wasmsimd-arm.c",
Frank Barchard8ef44cd2020-11-03 12:30:23 -08001146 "src/f32-spmm/gen/32x1-minmax-wasmsimd-x86-pipelined-x2.c",
1147 "src/f32-spmm/gen/32x1-minmax-wasmsimd-x86-pipelined.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07001148 "src/f32-spmm/gen/32x1-minmax-wasmsimd-x86-x2.c",
1149 "src/f32-spmm/gen/32x1-minmax-wasmsimd-x86-x4.c",
Frank Barchard846c0c62020-10-26 15:01:39 -07001150 "src/f32-spmm/gen/32x1-minmax-wasmsimd-x86.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -07001151 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-arm-x4.c",
1152 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001153 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-arm-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -07001154 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-x86-x4.c",
1155 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001156 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -07001157 "src/f32-vbinary/gen/vadd-relu-wasmsimd-x4.c",
1158 "src/f32-vbinary/gen/vadd-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001159 "src/f32-vbinary/gen/vadd-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07001160 "src/f32-vbinary/gen/vadd-wasmsimd-x4.c",
1161 "src/f32-vbinary/gen/vadd-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001162 "src/f32-vbinary/gen/vadd-wasmsimd-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001163 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-arm-x4.c",
1164 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001165 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-arm-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001166 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-x86-x4.c",
1167 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001168 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -07001169 "src/f32-vbinary/gen/vaddc-relu-wasmsimd-x4.c",
1170 "src/f32-vbinary/gen/vaddc-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001171 "src/f32-vbinary/gen/vaddc-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07001172 "src/f32-vbinary/gen/vaddc-wasmsimd-x4.c",
1173 "src/f32-vbinary/gen/vaddc-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001174 "src/f32-vbinary/gen/vaddc-wasmsimd-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -07001175 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-arm-x4.c",
1176 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001177 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-arm-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -07001178 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-x86-x4.c",
1179 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001180 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -07001181 "src/f32-vbinary/gen/vdiv-relu-wasmsimd-x4.c",
1182 "src/f32-vbinary/gen/vdiv-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001183 "src/f32-vbinary/gen/vdiv-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07001184 "src/f32-vbinary/gen/vdiv-wasmsimd-x4.c",
1185 "src/f32-vbinary/gen/vdiv-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001186 "src/f32-vbinary/gen/vdiv-wasmsimd-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001187 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-arm-x4.c",
1188 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001189 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-arm-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001190 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-x86-x4.c",
1191 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001192 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -07001193 "src/f32-vbinary/gen/vdivc-relu-wasmsimd-x4.c",
1194 "src/f32-vbinary/gen/vdivc-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001195 "src/f32-vbinary/gen/vdivc-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07001196 "src/f32-vbinary/gen/vdivc-wasmsimd-x4.c",
1197 "src/f32-vbinary/gen/vdivc-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001198 "src/f32-vbinary/gen/vdivc-wasmsimd-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -07001199 "src/f32-vbinary/gen/vmax-wasmsimd-arm-x4.c",
1200 "src/f32-vbinary/gen/vmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001201 "src/f32-vbinary/gen/vmax-wasmsimd-arm-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -07001202 "src/f32-vbinary/gen/vmax-wasmsimd-x86-x4.c",
1203 "src/f32-vbinary/gen/vmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001204 "src/f32-vbinary/gen/vmax-wasmsimd-x86-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001205 "src/f32-vbinary/gen/vmaxc-wasmsimd-arm-x4.c",
1206 "src/f32-vbinary/gen/vmaxc-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001207 "src/f32-vbinary/gen/vmaxc-wasmsimd-arm-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001208 "src/f32-vbinary/gen/vmaxc-wasmsimd-x86-x4.c",
1209 "src/f32-vbinary/gen/vmaxc-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001210 "src/f32-vbinary/gen/vmaxc-wasmsimd-x86-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -07001211 "src/f32-vbinary/gen/vmin-wasmsimd-arm-x4.c",
1212 "src/f32-vbinary/gen/vmin-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001213 "src/f32-vbinary/gen/vmin-wasmsimd-arm-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -07001214 "src/f32-vbinary/gen/vmin-wasmsimd-x86-x4.c",
1215 "src/f32-vbinary/gen/vmin-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001216 "src/f32-vbinary/gen/vmin-wasmsimd-x86-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001217 "src/f32-vbinary/gen/vminc-wasmsimd-arm-x4.c",
1218 "src/f32-vbinary/gen/vminc-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001219 "src/f32-vbinary/gen/vminc-wasmsimd-arm-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001220 "src/f32-vbinary/gen/vminc-wasmsimd-x86-x4.c",
1221 "src/f32-vbinary/gen/vminc-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001222 "src/f32-vbinary/gen/vminc-wasmsimd-x86-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -07001223 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-arm-x4.c",
1224 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001225 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-arm-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -07001226 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-x86-x4.c",
1227 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001228 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -07001229 "src/f32-vbinary/gen/vmul-relu-wasmsimd-x4.c",
1230 "src/f32-vbinary/gen/vmul-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001231 "src/f32-vbinary/gen/vmul-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07001232 "src/f32-vbinary/gen/vmul-wasmsimd-x4.c",
1233 "src/f32-vbinary/gen/vmul-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001234 "src/f32-vbinary/gen/vmul-wasmsimd-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001235 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-arm-x4.c",
1236 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001237 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-arm-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001238 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-x86-x4.c",
1239 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001240 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -07001241 "src/f32-vbinary/gen/vmulc-relu-wasmsimd-x4.c",
1242 "src/f32-vbinary/gen/vmulc-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001243 "src/f32-vbinary/gen/vmulc-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07001244 "src/f32-vbinary/gen/vmulc-wasmsimd-x4.c",
1245 "src/f32-vbinary/gen/vmulc-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001246 "src/f32-vbinary/gen/vmulc-wasmsimd-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001247 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-arm-x4.c",
1248 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001249 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-arm-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001250 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-x86-x4.c",
1251 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001252 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -07001253 "src/f32-vbinary/gen/vrdivc-relu-wasmsimd-x4.c",
1254 "src/f32-vbinary/gen/vrdivc-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001255 "src/f32-vbinary/gen/vrdivc-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07001256 "src/f32-vbinary/gen/vrdivc-wasmsimd-x4.c",
1257 "src/f32-vbinary/gen/vrdivc-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001258 "src/f32-vbinary/gen/vrdivc-wasmsimd-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001259 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-arm-x4.c",
1260 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001261 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-arm-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001262 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-x86-x4.c",
1263 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001264 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -07001265 "src/f32-vbinary/gen/vrsubc-relu-wasmsimd-x4.c",
1266 "src/f32-vbinary/gen/vrsubc-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001267 "src/f32-vbinary/gen/vrsubc-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07001268 "src/f32-vbinary/gen/vrsubc-wasmsimd-x4.c",
1269 "src/f32-vbinary/gen/vrsubc-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001270 "src/f32-vbinary/gen/vrsubc-wasmsimd-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -07001271 "src/f32-vbinary/gen/vsqrdiff-wasmsimd-x4.c",
1272 "src/f32-vbinary/gen/vsqrdiff-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001273 "src/f32-vbinary/gen/vsqrdiff-wasmsimd-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001274 "src/f32-vbinary/gen/vsqrdiffc-wasmsimd-x4.c",
1275 "src/f32-vbinary/gen/vsqrdiffc-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001276 "src/f32-vbinary/gen/vsqrdiffc-wasmsimd-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -07001277 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-arm-x4.c",
1278 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001279 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-arm-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -07001280 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-x86-x4.c",
1281 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001282 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -07001283 "src/f32-vbinary/gen/vsub-relu-wasmsimd-x4.c",
1284 "src/f32-vbinary/gen/vsub-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001285 "src/f32-vbinary/gen/vsub-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07001286 "src/f32-vbinary/gen/vsub-wasmsimd-x4.c",
1287 "src/f32-vbinary/gen/vsub-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001288 "src/f32-vbinary/gen/vsub-wasmsimd-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001289 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-arm-x4.c",
1290 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001291 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-arm-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001292 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-x86-x4.c",
1293 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001294 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -07001295 "src/f32-vbinary/gen/vsubc-relu-wasmsimd-x4.c",
1296 "src/f32-vbinary/gen/vsubc-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001297 "src/f32-vbinary/gen/vsubc-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07001298 "src/f32-vbinary/gen/vsubc-wasmsimd-x4.c",
1299 "src/f32-vbinary/gen/vsubc-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001300 "src/f32-vbinary/gen/vsubc-wasmsimd-x16.c",
Marat Dukhaned6baaf2020-12-01 15:07:08 -08001301 "src/f32-velu/gen/velu-wasmsimd-arm-rr2-lut16-p3-x4.c",
1302 "src/f32-velu/gen/velu-wasmsimd-arm-rr2-lut16-p3-x8.c",
1303 "src/f32-velu/gen/velu-wasmsimd-arm-rr2-lut16-p3-x12.c",
1304 "src/f32-velu/gen/velu-wasmsimd-arm-rr2-lut16-p3-x16.c",
1305 "src/f32-velu/gen/velu-wasmsimd-arm-rr2-lut16-p3-x20.c",
1306 "src/f32-velu/gen/velu-wasmsimd-arm-rr2-lut16-p3-x24.c",
Marat Dukhaned6baaf2020-12-01 15:07:08 -08001307 "src/f32-velu/gen/velu-wasmsimd-arm-rr2-p6-x4.c",
1308 "src/f32-velu/gen/velu-wasmsimd-arm-rr2-p6-x8.c",
1309 "src/f32-velu/gen/velu-wasmsimd-arm-rr2-p6-x12.c",
1310 "src/f32-velu/gen/velu-wasmsimd-arm-rr2-p6-x16.c",
1311 "src/f32-velu/gen/velu-wasmsimd-arm-rr2-p6-x20.c",
1312 "src/f32-velu/gen/velu-wasmsimd-arm-rr2-p6-x24.c",
Frank Barcharde7223ee2020-12-04 19:04:01 -08001313 "src/f32-velu/gen/velu-wasmsimd-x86-rr2-lut16-p3-x4.c",
1314 "src/f32-velu/gen/velu-wasmsimd-x86-rr2-lut16-p3-x8.c",
1315 "src/f32-velu/gen/velu-wasmsimd-x86-rr2-lut16-p3-x12.c",
1316 "src/f32-velu/gen/velu-wasmsimd-x86-rr2-lut16-p3-x16.c",
1317 "src/f32-velu/gen/velu-wasmsimd-x86-rr2-lut16-p3-x20.c",
1318 "src/f32-velu/gen/velu-wasmsimd-x86-rr2-lut16-p3-x24.c",
Marat Dukhaned6baaf2020-12-01 15:07:08 -08001319 "src/f32-velu/gen/velu-wasmsimd-x86-rr2-p6-x4.c",
1320 "src/f32-velu/gen/velu-wasmsimd-x86-rr2-p6-x8.c",
1321 "src/f32-velu/gen/velu-wasmsimd-x86-rr2-p6-x12.c",
1322 "src/f32-velu/gen/velu-wasmsimd-x86-rr2-p6-x16.c",
1323 "src/f32-velu/gen/velu-wasmsimd-x86-rr2-p6-x20.c",
1324 "src/f32-velu/gen/velu-wasmsimd-x86-rr2-p6-x24.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07001325 "src/f32-vlrelu/gen/vlrelu-wasmsimd-bitselect-x4.c",
1326 "src/f32-vlrelu/gen/vlrelu-wasmsimd-bitselect-x8.c",
1327 "src/f32-vlrelu/gen/vlrelu-wasmsimd-minmax-x4.c",
1328 "src/f32-vlrelu/gen/vlrelu-wasmsimd-minmax-x8.c",
Marat Dukhand816f622020-07-15 10:14:39 -07001329 "src/f32-vmulcaddc/gen/c4-minmax-wasmsimd-arm-2x.c",
Marat Dukhand816f622020-07-15 10:14:39 -07001330 "src/f32-vmulcaddc/gen/c4-minmax-wasmsimd-x86-2x.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001331 "src/f32-vmulcaddc/gen/c8-minmax-wasmsimd-arm-2x.c",
Marat Dukhand816f622020-07-15 10:14:39 -07001332 "src/f32-vmulcaddc/gen/c8-minmax-wasmsimd-x86-2x.c",
Marat Dukhanb82b2cd2020-07-16 02:23:42 -07001333 "src/f32-vrnd/gen/vrndd-wasmsimd-addsub-x4.c",
1334 "src/f32-vrnd/gen/vrndd-wasmsimd-addsub-x8.c",
1335 "src/f32-vrnd/gen/vrndd-wasmsimd-cvt-x4.c",
1336 "src/f32-vrnd/gen/vrndd-wasmsimd-cvt-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001337 "src/f32-vrnd/gen/vrndne-wasmsimd-addsub-x4.c",
1338 "src/f32-vrnd/gen/vrndne-wasmsimd-addsub-x8.c",
1339 "src/f32-vrnd/gen/vrndu-wasmsimd-addsub-x4.c",
1340 "src/f32-vrnd/gen/vrndu-wasmsimd-addsub-x8.c",
1341 "src/f32-vrnd/gen/vrndu-wasmsimd-cvt-x4.c",
1342 "src/f32-vrnd/gen/vrndu-wasmsimd-cvt-x8.c",
1343 "src/f32-vrnd/gen/vrndz-wasmsimd-addsub-x4.c",
1344 "src/f32-vrnd/gen/vrndz-wasmsimd-addsub-x8.c",
1345 "src/f32-vrnd/gen/vrndz-wasmsimd-cvt-x4.c",
1346 "src/f32-vrnd/gen/vrndz-wasmsimd-cvt-x8.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07001347 "src/f32-vsqrt/gen/wasmsimd-sqrt-x4.c",
1348 "src/f32-vsqrt/gen/wasmsimd-sqrt-x8.c",
Marat Dukhan37c83512020-06-29 13:25:53 -07001349 "src/f32-vunary/gen/vabs-wasmsimd-x4.c",
1350 "src/f32-vunary/gen/vabs-wasmsimd-x8.c",
1351 "src/f32-vunary/gen/vneg-wasmsimd-x4.c",
1352 "src/f32-vunary/gen/vneg-wasmsimd-x8.c",
1353 "src/f32-vunary/gen/vsqr-wasmsimd-x4.c",
1354 "src/f32-vunary/gen/vsqr-wasmsimd-x8.c",
Marat Dukhande390d42020-11-29 19:32:18 -08001355 "src/math/expm1minus-wasmsimd-rr2-lut16-p3-andnot.c",
1356 "src/math/expm1minus-wasmsimd-rr2-lut16-p3-max.c",
1357 "src/math/expm1minus-wasmsimd-rr2-p6-andnot.c",
1358 "src/math/expm1minus-wasmsimd-rr2-p6-max.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001359 "src/math/roundd-wasmsimd-addsub.c",
1360 "src/math/roundd-wasmsimd-cvt.c",
1361 "src/math/roundne-wasmsimd-addsub.c",
1362 "src/math/roundu-wasmsimd-addsub.c",
1363 "src/math/roundu-wasmsimd-cvt.c",
1364 "src/math/roundz-wasmsimd-addsub.c",
1365 "src/math/roundz-wasmsimd-cvt.c",
1366 "src/math/sigmoid-wasmsimd-rr2-lut64-p2-div.c",
1367 "src/math/sigmoid-wasmsimd-rr2-p5-div.c",
Marat Dukhancc8f34c2020-08-05 16:36:38 -07001368 "src/qs8-dwconv/gen/up8x9-minmax-wasmsimd-mul16.c",
1369 "src/qs8-dwconv/gen/up16x9-minmax-wasmsimd-mul16.c",
1370 "src/qs8-dwconv/gen/up24x9-minmax-wasmsimd-mul16.c",
Marat Dukhanb5e3d172020-08-06 13:29:53 -07001371 "src/qs8-gavgpool/gen/7p7x-minmax-wasmsimd-c8-acc2.c",
1372 "src/qs8-gavgpool/gen/7p7x-minmax-wasmsimd-c16-acc2.c",
1373 "src/qs8-gavgpool/gen/7p7x-minmax-wasmsimd-c24-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001374 "src/qs8-gavgpool/gen/7x-minmax-wasmsimd-c8-acc2.c",
1375 "src/qs8-gavgpool/gen/7x-minmax-wasmsimd-c16-acc2.c",
1376 "src/qs8-gavgpool/gen/7x-minmax-wasmsimd-c24-acc2.c",
Marat Dukhan27203da2020-08-05 15:19:03 -07001377 "src/qs8-gemm/gen/1x4c8-minmax-wasmsimd-ld64.c",
Marat Dukhan27203da2020-08-05 15:19:03 -07001378 "src/qs8-gemm/gen/1x4c8-minmax-wasmsimd-ld128.c",
Marat Dukhan27203da2020-08-05 15:19:03 -07001379 "src/qs8-gemm/gen/1x4c8-xw-minmax-wasmsimd.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001380 "src/qs8-gemm/gen/2x4c8-minmax-wasmsimd-ld64.c",
1381 "src/qs8-gemm/gen/2x4c8-minmax-wasmsimd-ld128.c",
Marat Dukhan27203da2020-08-05 15:19:03 -07001382 "src/qs8-gemm/gen/2x4c8-xw-minmax-wasmsimd.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001383 "src/qs8-gemm/gen/3x4c8-minmax-wasmsimd-ld64.c",
1384 "src/qs8-gemm/gen/3x4c8-minmax-wasmsimd-ld128.c",
Marat Dukhan27203da2020-08-05 15:19:03 -07001385 "src/qs8-gemm/gen/3x4c8-xw-minmax-wasmsimd.c",
1386 "src/qs8-igemm/gen/1x4c8-minmax-wasmsimd-ld64.c",
Marat Dukhan27203da2020-08-05 15:19:03 -07001387 "src/qs8-igemm/gen/1x4c8-minmax-wasmsimd-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001388 "src/qs8-igemm/gen/2x4c8-minmax-wasmsimd-ld64.c",
Marat Dukhan27203da2020-08-05 15:19:03 -07001389 "src/qs8-igemm/gen/2x4c8-minmax-wasmsimd-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001390 "src/qs8-igemm/gen/3x4c8-minmax-wasmsimd-ld64.c",
Marat Dukhan27203da2020-08-05 15:19:03 -07001391 "src/qs8-igemm/gen/3x4c8-minmax-wasmsimd-ld128.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -07001392 "src/qs8-requantization/fp32-wasmsimd.c",
Marat Dukhan138560c2020-08-03 18:57:34 -07001393 "src/qs8-requantization/q31-wasmsimd.c",
Marat Dukhan5df27f82020-09-02 23:59:21 -07001394 "src/qs8-vadd/gen/minmax-wasmsimd-x8.c",
1395 "src/qs8-vadd/gen/minmax-wasmsimd-x16.c",
1396 "src/qs8-vadd/gen/minmax-wasmsimd-x24.c",
1397 "src/qs8-vadd/gen/minmax-wasmsimd-x32.c",
1398 "src/qs8-vaddc/gen/minmax-wasmsimd-x8.c",
1399 "src/qs8-vaddc/gen/minmax-wasmsimd-x16.c",
1400 "src/qs8-vaddc/gen/minmax-wasmsimd-x24.c",
1401 "src/qs8-vaddc/gen/minmax-wasmsimd-x32.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07001402 "src/qu8-requantization/fp32-wasmsimd.c",
Marat Dukhan22de5e72020-08-03 21:53:18 -07001403 "src/qu8-requantization/q31-wasmsimd.c",
Marat Dukhan8ee37012020-07-16 13:17:13 -07001404 "src/x32-fill/wasmsimd.c",
Marat Dukhan66d99e92020-07-16 12:56:21 -07001405 "src/x32-packx/x4-wasmsimd.c",
Marat Dukhan9306ae02020-07-16 15:51:13 -07001406 "src/x32-pad/wasmsimd.c",
Marat Dukhan9d4bfa22020-07-16 19:07:04 -07001407 "src/x32-unpool/wasmsimd.c",
Marat Dukhane3b78762020-07-16 20:02:58 -07001408 "src/x32-zip/x2-wasmsimd.c",
1409 "src/x32-zip/x3-wasmsimd.c",
1410 "src/x32-zip/x4-wasmsimd.c",
1411 "src/x32-zip/xm-wasmsimd.c",
Marat Dukhan290055c2020-06-09 12:24:29 -07001412]
1413
Marat Dukhan08c4a432019-10-03 09:29:21 -07001414# ISA-specific micro-kernels
1415NEON_UKERNELS = [
Marat Dukhanef25c6d2020-07-24 00:59:40 -07001416 "src/f32-argmaxpool/4x-neon-c4.c",
1417 "src/f32-argmaxpool/9p8x-neon-c4.c",
1418 "src/f32-argmaxpool/9x-neon-c4.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001419 "src/f32-avgpool/9p8x-minmax-neon-c4.c",
1420 "src/f32-avgpool/9x-minmax-neon-c4.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001421 "src/f32-clamp/gen/neon-x4.c",
1422 "src/f32-clamp/gen/neon-x8.c",
Marat Dukhan56b10cd2020-05-18 09:35:49 -07001423 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neon-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001424 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neon-2x2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001425 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neon-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001426 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neon-2x2.c",
Marat Dukhan56b10cd2020-05-18 09:35:49 -07001427 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neon-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001428 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neon-2x2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001429 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neon-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001430 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neon-2x2.c",
Marat Dukhanc7634882020-12-07 15:11:12 -08001431 "src/f32-conv-hwc2chw/3x3s2p1c3x4-neon-2x2.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -07001432 "src/f32-dwconv/gen/up4x4-minmax-neon-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001433 "src/f32-dwconv/gen/up4x4-minmax-neon.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001434 "src/f32-dwconv/gen/up4x9-minmax-neon-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001435 "src/f32-dwconv/gen/up4x9-minmax-neon.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -07001436 "src/f32-dwconv/gen/up4x25-minmax-neon-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001437 "src/f32-dwconv/gen/up4x25-minmax-neon.c",
1438 "src/f32-dwconv/gen/up8x4-minmax-neon-acc2.c",
1439 "src/f32-dwconv/gen/up8x4-minmax-neon.c",
1440 "src/f32-dwconv/gen/up8x9-minmax-neon-acc2.c",
1441 "src/f32-dwconv/gen/up8x9-minmax-neon.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -07001442 "src/f32-dwconv/gen/up8x25-minmax-neon-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001443 "src/f32-dwconv/gen/up8x25-minmax-neon.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001444 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-1x4-acc2.c",
1445 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-1x4-acc3.c",
1446 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-1x4-acc4.c",
Marat Dukhanc581e482020-10-24 01:28:11 -07001447 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-1x4.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001448 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-2x4-acc2.c",
Marat Dukhanc581e482020-10-24 01:28:11 -07001449 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-2x4.c",
1450 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-3x4.c",
1451 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-4x4.c",
1452 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-5x4.c",
1453 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-6x4.c",
Marat Dukhan82f0c322020-10-25 19:17:35 -07001454 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neon-1x4-acc2.c",
1455 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neon-1x4-acc3.c",
1456 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neon-1x4-acc4.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001457 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neon-1x4.c",
Marat Dukhan82f0c322020-10-25 19:17:35 -07001458 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neon-2x4-acc2.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001459 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neon-2x4.c",
1460 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neon-3x4.c",
1461 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neon-4x4.c",
Marat Dukhan149f0ea2020-10-26 12:50:33 -07001462 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-1x4-acc2.c",
1463 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-1x4-acc3.c",
1464 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-1x4-acc4.c",
1465 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-1x4-acc5.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001466 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-1x4.c",
Marat Dukhan149f0ea2020-10-26 12:50:33 -07001467 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-2x4-acc2.c",
1468 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-2x4-acc3.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001469 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-2x4.c",
Marat Dukhan149f0ea2020-10-26 12:50:33 -07001470 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-3x4-acc2.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001471 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-3x4.c",
Marat Dukhan149f0ea2020-10-26 12:50:33 -07001472 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-4x4-acc2.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001473 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-4x4.c",
1474 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-5x4.c",
Marat Dukhan30d4b252020-10-29 16:33:22 -07001475 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neon-1x4-acc2.c",
1476 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neon-1x4-acc3.c",
1477 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neon-1x4-acc4.c",
1478 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neon-1x4-acc5.c",
1479 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neon-1x4.c",
1480 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neon-2x4-acc2.c",
1481 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neon-2x4-acc3.c",
1482 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neon-2x4.c",
Marat Dukhan30d4b252020-10-29 16:33:22 -07001483 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neon-3x4-acc2.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07001484 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neon-3x4.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -07001485 "src/f32-gavgpool-cw/neon-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001486 "src/f32-gavgpool/7p7x-minmax-neon-c4.c",
1487 "src/f32-gavgpool/7x-minmax-neon-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001488 "src/f32-gemm/gen-inc/1x8inc-minmax-neon-dup-ld64.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001489 "src/f32-gemm/gen-inc/1x8inc-minmax-neon-lane-ld64.c",
1490 "src/f32-gemm/gen-inc/1x8s4inc-minmax-neon.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001491 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-dup-ld64.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001492 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-dup-ld128.c",
1493 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-lane-ld64.c",
1494 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-lane-ld128.c",
1495 "src/f32-gemm/gen-inc/4x8s4inc-minmax-neon.c",
1496 "src/f32-gemm/gen-inc/5x8inc-minmax-neon-lane-ld64.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001497 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-dup-ld64.c",
1498 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-dup-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001499 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-lane-ld64.c",
1500 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-lane-ld128.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001501 "src/f32-gemm/gen-inc/6x8s4inc-minmax-neon.c",
1502 "src/f32-gemm/gen-inc/8x8s4inc-minmax-neon.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001503 "src/f32-gemm/gen/1x8-minmax-neon-dup-ld64.c",
1504 "src/f32-gemm/gen/1x8-minmax-neon-lane-ld64.c",
1505 "src/f32-gemm/gen/1x8s4-minmax-neon.c",
1506 "src/f32-gemm/gen/4x2-minmax-neon-lane-ld64.c",
1507 "src/f32-gemm/gen/4x8-minmax-neon-dup-ld64.c",
1508 "src/f32-gemm/gen/4x8-minmax-neon-dup-ld128.c",
1509 "src/f32-gemm/gen/4x8-minmax-neon-lane-ld64.c",
1510 "src/f32-gemm/gen/4x8-minmax-neon-lane-ld128.c",
1511 "src/f32-gemm/gen/4x8s4-minmax-neon.c",
1512 "src/f32-gemm/gen/5x8-minmax-neon-lane-ld64.c",
1513 "src/f32-gemm/gen/6x8-minmax-neon-dup-ld64.c",
1514 "src/f32-gemm/gen/6x8-minmax-neon-dup-ld128.c",
1515 "src/f32-gemm/gen/6x8-minmax-neon-lane-ld64.c",
1516 "src/f32-gemm/gen/6x8-minmax-neon-lane-ld128.c",
1517 "src/f32-gemm/gen/6x8s4-minmax-neon.c",
1518 "src/f32-gemm/gen/8x8s4-minmax-neon.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07001519 "src/f32-hswish/gen/hswish-neon-x4.c",
1520 "src/f32-hswish/gen/hswish-neon-x8.c",
1521 "src/f32-hswish/gen/hswish-neon-x16.c",
Artsiom Ablavatski2202c812021-01-22 14:16:43 -08001522 "src/f32-ibilinear-chw/gen/neon-p4.c",
1523 "src/f32-ibilinear-chw/gen/neon-p8.c",
Frank Barchard8247e212021-02-03 18:12:33 -08001524 "src/f32-ibilinear/gen/neon-c4.c",
1525 "src/f32-ibilinear/gen/neon-c8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001526 "src/f32-igemm/gen/1x8-minmax-neon-dup-ld64.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001527 "src/f32-igemm/gen/1x8-minmax-neon-lane-ld64.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001528 "src/f32-igemm/gen/1x8s4-minmax-neon.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001529 "src/f32-igemm/gen/4x2-minmax-neon-lane-ld64.c",
1530 "src/f32-igemm/gen/4x4-minmax-neon-lane-ld64.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001531 "src/f32-igemm/gen/4x8-minmax-neon-dup-ld64.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001532 "src/f32-igemm/gen/4x8-minmax-neon-dup-ld128.c",
1533 "src/f32-igemm/gen/4x8-minmax-neon-lane-ld64.c",
1534 "src/f32-igemm/gen/4x8-minmax-neon-lane-ld128.c",
1535 "src/f32-igemm/gen/4x8s4-minmax-neon.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001536 "src/f32-igemm/gen/6x8-minmax-neon-dup-ld64.c",
1537 "src/f32-igemm/gen/6x8-minmax-neon-dup-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001538 "src/f32-igemm/gen/6x8-minmax-neon-lane-ld64.c",
1539 "src/f32-igemm/gen/6x8-minmax-neon-lane-ld128.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001540 "src/f32-igemm/gen/6x8s4-minmax-neon.c",
1541 "src/f32-igemm/gen/8x8s4-minmax-neon.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001542 "src/f32-maxpool/9p8x-minmax-neon-c4.c",
1543 "src/f32-pavgpool/9p8x-minmax-neon-c4.c",
1544 "src/f32-pavgpool/9x-minmax-neon-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001545 "src/f32-ppmm/gen/4x8-minmax-neon.c",
1546 "src/f32-ppmm/gen/8x8-minmax-neon.c",
Frank Barcharda5316982020-07-23 13:19:28 -07001547 "src/f32-prelu/gen/neon-1x4.c",
1548 "src/f32-prelu/gen/neon-1x8.c",
1549 "src/f32-prelu/gen/neon-1x16.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001550 "src/f32-prelu/gen/neon-2x4.c",
1551 "src/f32-prelu/gen/neon-2x8.c",
Frank Barcharda5316982020-07-23 13:19:28 -07001552 "src/f32-prelu/gen/neon-2x16.c",
1553 "src/f32-prelu/gen/neon-4x4.c",
1554 "src/f32-prelu/gen/neon-4x8.c",
1555 "src/f32-prelu/gen/neon-4x16.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001556 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x4.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001557 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x8-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001558 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x8.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001559 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12-acc2.c",
1560 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001561 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001562 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16-acc2.c",
1563 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001564 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001565 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20-acc2.c",
1566 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001567 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20.c",
1568 "src/f32-raddstoreexpminusmax/gen/neon-p5-x4.c",
1569 "src/f32-raddstoreexpminusmax/gen/neon-p5-x8-acc2.c",
1570 "src/f32-raddstoreexpminusmax/gen/neon-p5-x8.c",
1571 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12-acc2.c",
1572 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12-acc3.c",
1573 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12.c",
1574 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16-acc2.c",
1575 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16-acc4.c",
1576 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16.c",
1577 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20-acc2.c",
1578 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20-acc5.c",
1579 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07001580 "src/f32-relu/gen/neon-x4.c",
1581 "src/f32-relu/gen/neon-x8.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001582 "src/f32-rmax/neon.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -08001583 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x4.c",
1584 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x8.c",
1585 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x12.c",
1586 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x16.c",
1587 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x20.c",
1588 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x24.c",
1589 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x4.c",
1590 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x8.c",
1591 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x12.c",
1592 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x16.c",
1593 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x20.c",
1594 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x24.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001595 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x4.c",
1596 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x8.c",
1597 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x12.c",
1598 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x16.c",
1599 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x20.c",
1600 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x24.c",
Marat Dukhan5b86c432020-12-06 19:15:03 -08001601 "src/f32-spmm/gen/4x1-minmax-neon-pipelined.c",
1602 "src/f32-spmm/gen/4x1-minmax-neon-x2.c",
1603 "src/f32-spmm/gen/4x1-minmax-neon.c",
1604 "src/f32-spmm/gen/8x1-minmax-neon-pipelined.c",
1605 "src/f32-spmm/gen/8x1-minmax-neon-x2.c",
1606 "src/f32-spmm/gen/8x1-minmax-neon.c",
1607 "src/f32-spmm/gen/12x1-minmax-neon.c",
1608 "src/f32-spmm/gen/16x1-minmax-neon-pipelined.c",
1609 "src/f32-spmm/gen/16x1-minmax-neon-x2.c",
1610 "src/f32-spmm/gen/16x1-minmax-neon.c",
1611 "src/f32-spmm/gen/32x1-minmax-neon-pipelined.c",
1612 "src/f32-spmm/gen/32x1-minmax-neon-x2.c",
1613 "src/f32-spmm/gen/32x1-minmax-neon.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001614 "src/f32-vbinary/gen/vadd-minmax-neon-x4.c",
1615 "src/f32-vbinary/gen/vadd-minmax-neon-x8.c",
1616 "src/f32-vbinary/gen/vaddc-minmax-neon-x4.c",
1617 "src/f32-vbinary/gen/vaddc-minmax-neon-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -08001618 "src/f32-vbinary/gen/vmax-neon-x4.c",
1619 "src/f32-vbinary/gen/vmax-neon-x8.c",
1620 "src/f32-vbinary/gen/vmaxc-neon-x4.c",
1621 "src/f32-vbinary/gen/vmaxc-neon-x8.c",
1622 "src/f32-vbinary/gen/vmin-neon-x4.c",
1623 "src/f32-vbinary/gen/vmin-neon-x8.c",
1624 "src/f32-vbinary/gen/vminc-neon-x4.c",
1625 "src/f32-vbinary/gen/vminc-neon-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001626 "src/f32-vbinary/gen/vmul-minmax-neon-x4.c",
1627 "src/f32-vbinary/gen/vmul-minmax-neon-x8.c",
1628 "src/f32-vbinary/gen/vmulc-minmax-neon-x4.c",
1629 "src/f32-vbinary/gen/vmulc-minmax-neon-x8.c",
1630 "src/f32-vbinary/gen/vrsubc-minmax-neon-x4.c",
1631 "src/f32-vbinary/gen/vrsubc-minmax-neon-x8.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07001632 "src/f32-vbinary/gen/vsqrdiff-neon-x4.c",
1633 "src/f32-vbinary/gen/vsqrdiff-neon-x8.c",
1634 "src/f32-vbinary/gen/vsqrdiffc-neon-x4.c",
1635 "src/f32-vbinary/gen/vsqrdiffc-neon-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001636 "src/f32-vbinary/gen/vsub-minmax-neon-x4.c",
1637 "src/f32-vbinary/gen/vsub-minmax-neon-x8.c",
1638 "src/f32-vbinary/gen/vsubc-minmax-neon-x4.c",
1639 "src/f32-vbinary/gen/vsubc-minmax-neon-x8.c",
Marat Dukhaned6baaf2020-12-01 15:07:08 -08001640 "src/f32-velu/gen/velu-neon-rr2-lut16-p3-x4.c",
1641 "src/f32-velu/gen/velu-neon-rr2-lut16-p3-x8.c",
1642 "src/f32-velu/gen/velu-neon-rr2-lut16-p3-x12.c",
1643 "src/f32-velu/gen/velu-neon-rr2-lut16-p3-x16.c",
1644 "src/f32-velu/gen/velu-neon-rr2-lut16-p3-x20.c",
1645 "src/f32-velu/gen/velu-neon-rr2-lut16-p3-x24.c",
1646 "src/f32-velu/gen/velu-neon-rr2-p6-x4.c",
1647 "src/f32-velu/gen/velu-neon-rr2-p6-x8.c",
1648 "src/f32-velu/gen/velu-neon-rr2-p6-x12.c",
1649 "src/f32-velu/gen/velu-neon-rr2-p6-x16.c",
1650 "src/f32-velu/gen/velu-neon-rr2-p6-x20.c",
1651 "src/f32-velu/gen/velu-neon-rr2-p6-x24.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07001652 "src/f32-vlrelu/gen/vlrelu-neon-x4.c",
1653 "src/f32-vlrelu/gen/vlrelu-neon-x8.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001654 "src/f32-vmulcaddc/gen/c4-minmax-neon-2x.c",
1655 "src/f32-vmulcaddc/gen/c8-minmax-neon-2x.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07001656 "src/f32-vrnd/gen/vrndd-neon-x4.c",
1657 "src/f32-vrnd/gen/vrndd-neon-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001658 "src/f32-vrnd/gen/vrndne-neon-x4.c",
1659 "src/f32-vrnd/gen/vrndne-neon-x8.c",
1660 "src/f32-vrnd/gen/vrndu-neon-x4.c",
1661 "src/f32-vrnd/gen/vrndu-neon-x8.c",
1662 "src/f32-vrnd/gen/vrndz-neon-x4.c",
1663 "src/f32-vrnd/gen/vrndz-neon-x8.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07001664 "src/f32-vunary/gen/vabs-neon-x4.c",
1665 "src/f32-vunary/gen/vabs-neon-x8.c",
1666 "src/f32-vunary/gen/vneg-neon-x4.c",
1667 "src/f32-vunary/gen/vneg-neon-x8.c",
1668 "src/f32-vunary/gen/vsqr-neon-x4.c",
1669 "src/f32-vunary/gen/vsqr-neon-x8.c",
Marat Dukhande390d42020-11-29 19:32:18 -08001670 "src/math/expm1minus-neon-rr2-lut16-p3.c",
1671 "src/math/expm1minus-neon-rr2-p6.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001672 "src/math/roundd-neon-addsub.c",
1673 "src/math/roundd-neon-cvt.c",
1674 "src/math/roundne-neon-addsub.c",
1675 "src/math/roundu-neon-addsub.c",
1676 "src/math/roundu-neon-cvt.c",
1677 "src/math/roundz-neon-addsub.c",
1678 "src/math/roundz-neon-cvt.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001679 "src/math/sigmoid-neon-rr2-lut64-p2-nr2recps.c",
1680 "src/math/sigmoid-neon-rr2-lut2048-p1-nr2recps.c",
1681 "src/math/sigmoid-neon-rr2-p5-nr2recps.c",
1682 "src/math/sqrt-neon-nr1rsqrts.c",
1683 "src/math/sqrt-neon-nr2rsqrts.c",
1684 "src/math/sqrt-neon-nr3rsqrts.c",
Marat Dukhan023bcf92020-08-10 12:40:50 -07001685 "src/qs8-dwconv/gen/up8x9-minmax-neon-mul16.c",
1686 "src/qs8-dwconv/gen/up16x9-minmax-neon-mul16.c",
1687 "src/qs8-dwconv/gen/up24x9-minmax-neon-mul16.c",
1688 "src/qs8-dwconv/gen/up32x9-minmax-neon-mul16.c",
Marat Dukhan281262d2020-08-10 13:23:21 -07001689 "src/qs8-gavgpool/gen/7p7x-minmax-neon-c8-acc2.c",
1690 "src/qs8-gavgpool/gen/7p7x-minmax-neon-c16-acc2.c",
1691 "src/qs8-gavgpool/gen/7p7x-minmax-neon-c24-acc2.c",
1692 "src/qs8-gavgpool/gen/7p7x-minmax-neon-c32-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001693 "src/qs8-gavgpool/gen/7x-minmax-neon-c8-acc2.c",
1694 "src/qs8-gavgpool/gen/7x-minmax-neon-c16-acc2.c",
1695 "src/qs8-gavgpool/gen/7x-minmax-neon-c24-acc2.c",
1696 "src/qs8-gavgpool/gen/7x-minmax-neon-c32-acc2.c",
Frank Barchardec0bf142021-01-15 10:14:05 -08001697 "src/qs8-gemm/gen/1x8-minmax-neon-mlal-lane.c",
Frank Barchardec0bf142021-01-15 10:14:05 -08001698 "src/qs8-gemm/gen/1x8-minmax-neon-mull-addw-dup.c",
Frank Barchard8247e212021-02-03 18:12:33 -08001699 "src/qs8-gemm/gen/1x8c2-minmax-neon-mlal-padal-dup.c",
Frank Barchard2302ffd2021-01-22 14:08:02 -08001700 "src/qs8-gemm/gen/1x8c2-minmax-neon-mull-padal-dup.c",
Frank Barchardda78da12021-03-02 14:28:00 -08001701 "src/qs8-gemm/gen/1x8c8-minmax-neon-mlal-padal.c",
Frank Barchard8247e212021-02-03 18:12:33 -08001702 "src/qs8-gemm/gen/1x8c8-minmax-neon-mull-padal.c",
Frank Barchard671d1b02021-03-10 15:42:08 -08001703 "src/qs8-gemm/gen/1x8c16-minmax-neon-mlal-padal.c",
Frank Barchardcadd4222021-01-20 16:27:25 -08001704 "src/qs8-gemm/gen/1x16-minmax-neon-mlal-lane.c",
1705 "src/qs8-gemm/gen/1x16-minmax-neon-mull-addw-dup.c",
Frank Barchard8247e212021-02-03 18:12:33 -08001706 "src/qs8-gemm/gen/1x16c2-minmax-neon-mlal-padal-dup.c",
Frank Barchard2302ffd2021-01-22 14:08:02 -08001707 "src/qs8-gemm/gen/1x16c2-minmax-neon-mull-padal-dup.c",
Frank Barchardda78da12021-03-02 14:28:00 -08001708 "src/qs8-gemm/gen/1x16c8-minmax-neon-mlal-padal.c",
Frank Barchard8247e212021-02-03 18:12:33 -08001709 "src/qs8-gemm/gen/1x16c8-minmax-neon-mull-padal.c",
Frank Barchard671d1b02021-03-10 15:42:08 -08001710 "src/qs8-gemm/gen/1x16c16-minmax-neon-mlal-padal.c",
Frank Barchardcadd4222021-01-20 16:27:25 -08001711 "src/qs8-gemm/gen/2x8-minmax-neon-mlal-lane.c",
Frank Barchardec0bf142021-01-15 10:14:05 -08001712 "src/qs8-gemm/gen/2x8-minmax-neon-mull-addw-dup.c",
Frank Barchard8247e212021-02-03 18:12:33 -08001713 "src/qs8-gemm/gen/2x8c2-minmax-neon-mlal-padal-dup.c",
Frank Barchard2302ffd2021-01-22 14:08:02 -08001714 "src/qs8-gemm/gen/2x8c2-minmax-neon-mull-padal-dup.c",
Frank Barchardda78da12021-03-02 14:28:00 -08001715 "src/qs8-gemm/gen/2x8c8-minmax-neon-mlal-padal.c",
Frank Barchard8247e212021-02-03 18:12:33 -08001716 "src/qs8-gemm/gen/2x8c8-minmax-neon-mull-padal.c",
Frank Barchard671d1b02021-03-10 15:42:08 -08001717 "src/qs8-gemm/gen/2x8c16-minmax-neon-mlal-padal.c",
Frank Barchardcadd4222021-01-20 16:27:25 -08001718 "src/qs8-gemm/gen/2x16-minmax-neon-mlal-lane.c",
1719 "src/qs8-gemm/gen/2x16-minmax-neon-mull-addw-dup.c",
Frank Barchard8247e212021-02-03 18:12:33 -08001720 "src/qs8-gemm/gen/2x16c2-minmax-neon-mlal-padal-dup.c",
Frank Barchard2302ffd2021-01-22 14:08:02 -08001721 "src/qs8-gemm/gen/2x16c2-minmax-neon-mull-padal-dup.c",
Frank Barchardda78da12021-03-02 14:28:00 -08001722 "src/qs8-gemm/gen/2x16c8-minmax-neon-mlal-padal.c",
Frank Barchard8247e212021-02-03 18:12:33 -08001723 "src/qs8-gemm/gen/2x16c8-minmax-neon-mull-padal.c",
Frank Barchard671d1b02021-03-10 15:42:08 -08001724 "src/qs8-gemm/gen/2x16c16-minmax-neon-mlal-padal.c",
Frank Barchard2302ffd2021-01-22 14:08:02 -08001725 "src/qs8-gemm/gen/3x8-minmax-neon-mlal-lane.c",
1726 "src/qs8-gemm/gen/3x8-minmax-neon-mull-addw-dup.c",
Frank Barchard8247e212021-02-03 18:12:33 -08001727 "src/qs8-gemm/gen/3x8c2-minmax-neon-mlal-padal-dup.c",
Frank Barchard2302ffd2021-01-22 14:08:02 -08001728 "src/qs8-gemm/gen/3x8c2-minmax-neon-mull-padal-dup.c",
Frank Barchardda78da12021-03-02 14:28:00 -08001729 "src/qs8-gemm/gen/3x8c8-minmax-neon-mlal-padal.c",
Frank Barchard8247e212021-02-03 18:12:33 -08001730 "src/qs8-gemm/gen/3x8c8-minmax-neon-mull-padal.c",
Frank Barchard671d1b02021-03-10 15:42:08 -08001731 "src/qs8-gemm/gen/3x8c16-minmax-neon-mlal-padal.c",
Frank Barchard8247e212021-02-03 18:12:33 -08001732 "src/qs8-gemm/gen/3x16-minmax-neon-mlal-lane.c",
1733 "src/qs8-gemm/gen/3x16-minmax-neon-mull-addw-dup.c",
1734 "src/qs8-gemm/gen/3x16c2-minmax-neon-mlal-padal-dup.c",
1735 "src/qs8-gemm/gen/3x16c2-minmax-neon-mull-padal-dup.c",
Frank Barchardda78da12021-03-02 14:28:00 -08001736 "src/qs8-gemm/gen/3x16c8-minmax-neon-mlal-padal.c",
Frank Barchard8247e212021-02-03 18:12:33 -08001737 "src/qs8-gemm/gen/3x16c8-minmax-neon-mull-padal.c",
Frank Barchard671d1b02021-03-10 15:42:08 -08001738 "src/qs8-gemm/gen/3x16c16-minmax-neon-mlal-padal.c",
Frank Barchardcadd4222021-01-20 16:27:25 -08001739 "src/qs8-gemm/gen/4x8-minmax-neon-mlal-lane.c",
Frank Barchardec0bf142021-01-15 10:14:05 -08001740 "src/qs8-gemm/gen/4x8-minmax-neon-mull-addw-dup.c",
Frank Barchard8247e212021-02-03 18:12:33 -08001741 "src/qs8-gemm/gen/4x8c2-minmax-neon-mlal-padal-dup.c",
Frank Barchard2302ffd2021-01-22 14:08:02 -08001742 "src/qs8-gemm/gen/4x8c2-minmax-neon-mull-padal-dup.c",
Frank Barchardda78da12021-03-02 14:28:00 -08001743 "src/qs8-gemm/gen/4x8c8-minmax-neon-mlal-padal.c",
Frank Barchard8247e212021-02-03 18:12:33 -08001744 "src/qs8-gemm/gen/4x8c8-minmax-neon-mull-padal.c",
Frank Barchard671d1b02021-03-10 15:42:08 -08001745 "src/qs8-gemm/gen/4x8c16-minmax-neon-mlal-padal.c",
Frank Barchardcadd4222021-01-20 16:27:25 -08001746 "src/qs8-gemm/gen/4x16-minmax-neon-mlal-lane.c",
1747 "src/qs8-gemm/gen/4x16-minmax-neon-mull-addw-dup.c",
Frank Barchard8247e212021-02-03 18:12:33 -08001748 "src/qs8-gemm/gen/4x16c2-minmax-neon-mlal-padal-dup.c",
Frank Barchard2302ffd2021-01-22 14:08:02 -08001749 "src/qs8-gemm/gen/4x16c2-minmax-neon-mull-padal-dup.c",
Frank Barchardda78da12021-03-02 14:28:00 -08001750 "src/qs8-gemm/gen/4x16c8-minmax-neon-mlal-padal.c",
Frank Barchard58990122021-01-29 16:09:56 -08001751 "src/qs8-gemm/gen/4x16c8-minmax-neon-mull-padal.c",
Frank Barchard71c4d1a2021-02-19 10:07:23 -08001752 "src/qs8-gemm/gen/4x16c16-minmax-neon-mlal-padal.c",
Frank Barchardec0bf142021-01-15 10:14:05 -08001753 "src/qs8-igemm/gen/1x8-minmax-neon-mlal-lane.c",
Frank Barchard671d1b02021-03-10 15:42:08 -08001754 "src/qs8-igemm/gen/1x8-minmax-neon-mull-addw-dup.c",
1755 "src/qs8-igemm/gen/1x8c2-minmax-neon-mlal-padal-dup.c",
1756 "src/qs8-igemm/gen/1x8c2-minmax-neon-mull-padal-dup.c",
Frank Barchardda78da12021-03-02 14:28:00 -08001757 "src/qs8-igemm/gen/1x8c8-minmax-neon-mlal-padal.c",
1758 "src/qs8-igemm/gen/1x8c8-minmax-neon-mull-padal.c",
Frank Barchard671d1b02021-03-10 15:42:08 -08001759 "src/qs8-igemm/gen/1x8c16-minmax-neon-mlal-padal.c",
1760 "src/qs8-igemm/gen/1x16-minmax-neon-mlal-lane.c",
1761 "src/qs8-igemm/gen/1x16-minmax-neon-mull-addw-dup.c",
1762 "src/qs8-igemm/gen/1x16c2-minmax-neon-mlal-padal-dup.c",
1763 "src/qs8-igemm/gen/1x16c2-minmax-neon-mull-padal-dup.c",
1764 "src/qs8-igemm/gen/1x16c8-minmax-neon-mlal-padal.c",
1765 "src/qs8-igemm/gen/1x16c8-minmax-neon-mull-padal.c",
1766 "src/qs8-igemm/gen/1x16c16-minmax-neon-mlal-padal.c",
1767 "src/qs8-igemm/gen/2x8-minmax-neon-mlal-lane.c",
1768 "src/qs8-igemm/gen/2x8-minmax-neon-mull-addw-dup.c",
1769 "src/qs8-igemm/gen/2x8c2-minmax-neon-mlal-padal-dup.c",
1770 "src/qs8-igemm/gen/2x8c2-minmax-neon-mull-padal-dup.c",
Frank Barchardda78da12021-03-02 14:28:00 -08001771 "src/qs8-igemm/gen/2x8c8-minmax-neon-mlal-padal.c",
1772 "src/qs8-igemm/gen/2x8c8-minmax-neon-mull-padal.c",
Frank Barchard671d1b02021-03-10 15:42:08 -08001773 "src/qs8-igemm/gen/2x8c16-minmax-neon-mlal-padal.c",
1774 "src/qs8-igemm/gen/2x16-minmax-neon-mlal-lane.c",
1775 "src/qs8-igemm/gen/2x16-minmax-neon-mull-addw-dup.c",
1776 "src/qs8-igemm/gen/2x16c2-minmax-neon-mlal-padal-dup.c",
1777 "src/qs8-igemm/gen/2x16c2-minmax-neon-mull-padal-dup.c",
1778 "src/qs8-igemm/gen/2x16c8-minmax-neon-mlal-padal.c",
1779 "src/qs8-igemm/gen/2x16c8-minmax-neon-mull-padal.c",
1780 "src/qs8-igemm/gen/2x16c16-minmax-neon-mlal-padal.c",
1781 "src/qs8-igemm/gen/3x8-minmax-neon-mlal-lane.c",
1782 "src/qs8-igemm/gen/3x8-minmax-neon-mull-addw-dup.c",
1783 "src/qs8-igemm/gen/3x8c2-minmax-neon-mlal-padal-dup.c",
1784 "src/qs8-igemm/gen/3x8c2-minmax-neon-mull-padal-dup.c",
Frank Barchardda78da12021-03-02 14:28:00 -08001785 "src/qs8-igemm/gen/3x8c8-minmax-neon-mlal-padal.c",
1786 "src/qs8-igemm/gen/3x8c8-minmax-neon-mull-padal.c",
Frank Barchard671d1b02021-03-10 15:42:08 -08001787 "src/qs8-igemm/gen/3x8c16-minmax-neon-mlal-padal.c",
1788 "src/qs8-igemm/gen/3x16-minmax-neon-mlal-lane.c",
1789 "src/qs8-igemm/gen/3x16-minmax-neon-mull-addw-dup.c",
1790 "src/qs8-igemm/gen/3x16c2-minmax-neon-mlal-padal-dup.c",
1791 "src/qs8-igemm/gen/3x16c2-minmax-neon-mull-padal-dup.c",
1792 "src/qs8-igemm/gen/3x16c8-minmax-neon-mlal-padal.c",
1793 "src/qs8-igemm/gen/3x16c8-minmax-neon-mull-padal.c",
1794 "src/qs8-igemm/gen/3x16c16-minmax-neon-mlal-padal.c",
1795 "src/qs8-igemm/gen/4x8-minmax-neon-mlal-lane.c",
1796 "src/qs8-igemm/gen/4x8-minmax-neon-mull-addw-dup.c",
1797 "src/qs8-igemm/gen/4x8c2-minmax-neon-mlal-padal-dup.c",
1798 "src/qs8-igemm/gen/4x8c2-minmax-neon-mull-padal-dup.c",
Frank Barchardda78da12021-03-02 14:28:00 -08001799 "src/qs8-igemm/gen/4x8c8-minmax-neon-mlal-padal.c",
1800 "src/qs8-igemm/gen/4x8c8-minmax-neon-mull-padal.c",
Frank Barchardda78da12021-03-02 14:28:00 -08001801 "src/qs8-igemm/gen/4x8c16-minmax-neon-mlal-padal.c",
Frank Barchard671d1b02021-03-10 15:42:08 -08001802 "src/qs8-igemm/gen/4x16-minmax-neon-mlal-lane.c",
Frank Barchardda78da12021-03-02 14:28:00 -08001803 "src/qs8-igemm/gen/4x16-minmax-neon-mull-addw-dup.c",
Frank Barchard671d1b02021-03-10 15:42:08 -08001804 "src/qs8-igemm/gen/4x16c2-minmax-neon-mlal-padal-dup.c",
1805 "src/qs8-igemm/gen/4x16c2-minmax-neon-mull-padal-dup.c",
1806 "src/qs8-igemm/gen/4x16c8-minmax-neon-mlal-padal.c",
1807 "src/qs8-igemm/gen/4x16c8-minmax-neon-mull-padal.c",
1808 "src/qs8-igemm/gen/4x16c16-minmax-neon-mlal-padal.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -07001809 "src/qs8-requantization/fp32-neon.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001810 "src/qs8-requantization/precise-neon.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -07001811 "src/qs8-requantization/q31-neon.c",
Marat Dukhanba7b2792020-09-02 14:26:45 -07001812 "src/qs8-vadd/gen/minmax-neon-ld64-x8.c",
1813 "src/qs8-vadd/gen/minmax-neon-ld64-x16.c",
1814 "src/qs8-vadd/gen/minmax-neon-ld64-x24.c",
1815 "src/qs8-vadd/gen/minmax-neon-ld64-x32.c",
1816 "src/qs8-vaddc/gen/minmax-neon-ld64-x8.c",
1817 "src/qs8-vaddc/gen/minmax-neon-ld64-x16.c",
1818 "src/qs8-vaddc/gen/minmax-neon-ld64-x24.c",
1819 "src/qs8-vaddc/gen/minmax-neon-ld64-x32.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -07001820 "src/qu8-avgpool/9p8x-minmax-neon-c8.c",
1821 "src/qu8-avgpool/9x-minmax-neon-c8.c",
1822 "src/qu8-dwconv/up8x9-minmax-neon.c",
1823 "src/qu8-gavgpool/7p7x-minmax-neon-c8.c",
1824 "src/qu8-gavgpool/7x-minmax-neon-c8.c",
1825 "src/qu8-gemm/4x8-minmax-neon.c",
1826 "src/qu8-gemm/8x8-minmax-neon.c",
1827 "src/qu8-igemm/4x8-minmax-neon.c",
1828 "src/qu8-igemm/8x8-minmax-neon.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07001829 "src/qu8-requantization/fp32-neon.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001830 "src/qu8-requantization/precise-neon.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07001831 "src/qu8-requantization/q31-neon.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -07001832 "src/qu8-vadd/minmax-neon.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001833 "src/u8-clamp/neon-x64.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001834 "src/u8-maxpool/9p8x-minmax-neon-c16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001835 "src/u8-rmax/neon.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001836 "src/x8-zip/x2-neon.c",
1837 "src/x8-zip/x3-neon.c",
1838 "src/x8-zip/x4-neon.c",
1839 "src/x8-zip/xm-neon.c",
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -07001840 "src/x32-fill/neon.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001841 "src/x32-packx/x4-neon-st4.c",
Marat Dukhan63523d42020-05-22 17:07:33 -07001842 "src/x32-pad/neon.c",
Marat Dukhan57dccd82020-04-14 00:53:10 -07001843 "src/x32-unpool/neon.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001844 "src/x32-zip/x2-neon.c",
1845 "src/x32-zip/x3-neon.c",
1846 "src/x32-zip/x4-neon.c",
1847 "src/x32-zip/xm-neon.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001848]
1849
1850NEONFMA_UKERNELS = [
Frank Barchard04336c12020-10-22 16:48:55 -07001851 "src/f32-dwconv/gen/up4x4-minmax-neonfma-acc2.c",
1852 "src/f32-dwconv/gen/up4x4-minmax-neonfma.c",
1853 "src/f32-dwconv/gen/up4x9-minmax-neonfma-acc2.c",
1854 "src/f32-dwconv/gen/up4x9-minmax-neonfma.c",
1855 "src/f32-dwconv/gen/up4x25-minmax-neonfma-acc2.c",
1856 "src/f32-dwconv/gen/up4x25-minmax-neonfma.c",
1857 "src/f32-dwconv/gen/up8x4-minmax-neonfma-acc2.c",
1858 "src/f32-dwconv/gen/up8x4-minmax-neonfma.c",
1859 "src/f32-dwconv/gen/up8x9-minmax-neonfma-acc2.c",
1860 "src/f32-dwconv/gen/up8x9-minmax-neonfma.c",
1861 "src/f32-dwconv/gen/up8x25-minmax-neonfma-acc2.c",
1862 "src/f32-dwconv/gen/up8x25-minmax-neonfma.c",
1863 "src/f32-gemm/gen-inc/1x8inc-minmax-neonfma-dup-ld64.c",
1864 "src/f32-gemm/gen-inc/1x8s4inc-minmax-neonfma.c",
1865 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-dup-ld64.c",
1866 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-dup-ld128.c",
1867 "src/f32-gemm/gen-inc/4x8s4inc-minmax-neonfma.c",
1868 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-dup-ld64.c",
1869 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-dup-ld128.c",
1870 "src/f32-gemm/gen-inc/6x8s4inc-minmax-neonfma.c",
1871 "src/f32-gemm/gen-inc/8x8s4inc-minmax-neonfma.c",
1872 "src/f32-gemm/gen/1x8-minmax-neonfma-dup-ld64.c",
1873 "src/f32-gemm/gen/1x8s4-minmax-neonfma.c",
1874 "src/f32-gemm/gen/4x8-minmax-neonfma-dup-ld64.c",
1875 "src/f32-gemm/gen/4x8-minmax-neonfma-dup-ld128.c",
1876 "src/f32-gemm/gen/4x8s4-minmax-neonfma.c",
1877 "src/f32-gemm/gen/6x8-minmax-neonfma-dup-ld64.c",
1878 "src/f32-gemm/gen/6x8-minmax-neonfma-dup-ld128.c",
1879 "src/f32-gemm/gen/6x8s4-minmax-neonfma.c",
1880 "src/f32-gemm/gen/8x8s4-minmax-neonfma.c",
Artsiom Ablavatski2202c812021-01-22 14:16:43 -08001881 "src/f32-ibilinear-chw/gen/neonfma-p4.c",
1882 "src/f32-ibilinear-chw/gen/neonfma-p8.c",
Frank Barchard8247e212021-02-03 18:12:33 -08001883 "src/f32-ibilinear/gen/neonfma-c4.c",
1884 "src/f32-ibilinear/gen/neonfma-c8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001885 "src/f32-igemm/gen/1x8-minmax-neonfma-dup-ld64.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001886 "src/f32-igemm/gen/1x8s4-minmax-neonfma.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001887 "src/f32-igemm/gen/4x8-minmax-neonfma-dup-ld64.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001888 "src/f32-igemm/gen/4x8-minmax-neonfma-dup-ld128.c",
1889 "src/f32-igemm/gen/4x8s4-minmax-neonfma.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001890 "src/f32-igemm/gen/6x8-minmax-neonfma-dup-ld64.c",
1891 "src/f32-igemm/gen/6x8-minmax-neonfma-dup-ld128.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001892 "src/f32-igemm/gen/6x8s4-minmax-neonfma.c",
1893 "src/f32-igemm/gen/8x8s4-minmax-neonfma.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001894 "src/f32-ppmm/gen/4x8-minmax-neonfma.c",
1895 "src/f32-ppmm/gen/8x8-minmax-neonfma.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001896 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x4.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001897 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x8-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001898 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x8.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001899 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12-acc2.c",
1900 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001901 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001902 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16-acc2.c",
1903 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001904 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001905 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20-acc2.c",
1906 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001907 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20.c",
1908 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x4.c",
1909 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x8-acc2.c",
1910 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x8.c",
1911 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12-acc2.c",
1912 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12-acc3.c",
1913 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12.c",
1914 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16-acc2.c",
1915 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16-acc4.c",
1916 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16.c",
1917 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20-acc2.c",
1918 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20-acc5.c",
1919 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -08001920 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x4.c",
1921 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x8.c",
1922 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x12.c",
1923 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x16.c",
1924 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x20.c",
1925 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x24.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001926 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x4.c",
1927 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x8.c",
1928 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x12.c",
1929 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x16.c",
1930 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x20.c",
1931 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x24.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -08001932 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x4.c",
1933 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x8.c",
1934 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x12.c",
1935 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x16.c",
1936 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x20.c",
1937 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x24.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -08001938 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x4.c",
1939 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x8.c",
1940 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x12.c",
1941 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x16.c",
1942 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x20.c",
1943 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x24.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001944 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x4.c",
1945 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x8.c",
1946 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x12.c",
1947 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x16.c",
1948 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x20.c",
1949 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x24.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -08001950 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x4.c",
1951 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x8.c",
1952 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x12.c",
1953 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x16.c",
1954 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x20.c",
1955 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x24.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001956 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x4.c",
1957 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x8.c",
1958 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x12.c",
1959 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x16.c",
1960 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x20.c",
1961 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x24.c",
1962 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x4.c",
1963 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x8.c",
1964 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x12.c",
1965 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x16.c",
1966 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x20.c",
1967 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x24.c",
1968 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x4.c",
1969 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x8.c",
1970 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x12.c",
1971 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x16.c",
1972 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x20.c",
1973 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x24.c",
Marat Dukhan2fa7a0c2020-12-06 19:09:02 -08001974 "src/f32-spmm/gen/4x1-minmax-neonfma-pipelined.c",
1975 "src/f32-spmm/gen/4x1-minmax-neonfma-x2.c",
1976 "src/f32-spmm/gen/4x1-minmax-neonfma.c",
1977 "src/f32-spmm/gen/8x1-minmax-neonfma-pipelined.c",
1978 "src/f32-spmm/gen/8x1-minmax-neonfma-x2.c",
1979 "src/f32-spmm/gen/8x1-minmax-neonfma.c",
1980 "src/f32-spmm/gen/12x1-minmax-neonfma.c",
1981 "src/f32-spmm/gen/16x1-minmax-neonfma-pipelined.c",
1982 "src/f32-spmm/gen/16x1-minmax-neonfma-x2.c",
1983 "src/f32-spmm/gen/16x1-minmax-neonfma.c",
1984 "src/f32-spmm/gen/32x1-minmax-neonfma-pipelined.c",
1985 "src/f32-spmm/gen/32x1-minmax-neonfma-x2.c",
1986 "src/f32-spmm/gen/32x1-minmax-neonfma.c",
Marat Dukhaned6baaf2020-12-01 15:07:08 -08001987 "src/f32-velu/gen/velu-neonfma-rr1-lut16-p3-x4.c",
1988 "src/f32-velu/gen/velu-neonfma-rr1-lut16-p3-x8.c",
1989 "src/f32-velu/gen/velu-neonfma-rr1-lut16-p3-x12.c",
1990 "src/f32-velu/gen/velu-neonfma-rr1-lut16-p3-x16.c",
1991 "src/f32-velu/gen/velu-neonfma-rr1-lut16-p3-x20.c",
1992 "src/f32-velu/gen/velu-neonfma-rr1-lut16-p3-x24.c",
1993 "src/f32-velu/gen/velu-neonfma-rr1-p6-x4.c",
1994 "src/f32-velu/gen/velu-neonfma-rr1-p6-x8.c",
1995 "src/f32-velu/gen/velu-neonfma-rr1-p6-x12.c",
1996 "src/f32-velu/gen/velu-neonfma-rr1-p6-x16.c",
1997 "src/f32-velu/gen/velu-neonfma-rr1-p6-x20.c",
1998 "src/f32-velu/gen/velu-neonfma-rr1-p6-x24.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001999 "src/f32-vmulcaddc/gen/c4-minmax-neonfma-2x.c",
2000 "src/f32-vmulcaddc/gen/c8-minmax-neonfma-2x.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07002001 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x4.c",
2002 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x8.c",
2003 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x12.c",
2004 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x16.c",
2005 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x20.c",
2006 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x24.c",
2007 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x28.c",
2008 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x32.c",
2009 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x36.c",
2010 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x40.c",
2011 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x4.c",
2012 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x8.c",
2013 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x12.c",
2014 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x16.c",
2015 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x20.c",
2016 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x24.c",
2017 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x28.c",
2018 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x32.c",
2019 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x36.c",
2020 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x40.c",
Marat Dukhanb7633f22020-11-20 16:34:56 -08002021 "src/math/exp-neonfma-rr2-lut64-p2.c",
2022 "src/math/exp-neonfma-rr2-p5.c",
Frank Barcharde7223ee2020-12-04 19:04:01 -08002023 "src/math/expm1minus-neonfma-rr1-lut16-p3.c",
2024 "src/math/expm1minus-neonfma-rr1-p6.c",
Marat Dukhan9dd119a2020-11-20 18:20:04 -08002025 "src/math/expminus-neonfma-rr2-lut64-p2.c",
2026 "src/math/expminus-neonfma-rr2-lut2048-p1.c",
2027 "src/math/expminus-neonfma-rr2-p5.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08002028 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr1recps1fma.c",
2029 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr2fma.c",
2030 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr2recps.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002031 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr1recps1fma.c",
2032 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr2fma.c",
2033 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr2recps.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08002034 "src/math/sigmoid-neonfma-rr1-p5-nr1recps1fma.c",
2035 "src/math/sigmoid-neonfma-rr1-p5-nr2fma.c",
2036 "src/math/sigmoid-neonfma-rr1-p5-nr2recps.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08002037 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr1recps1fma.c",
2038 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr2fma.c",
2039 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr2recps.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002040 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr1recps1fma.c",
2041 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr2fma.c",
2042 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr2recps.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08002043 "src/math/sigmoid-neonfma-rr2-p5-nr1recps1fma.c",
2044 "src/math/sigmoid-neonfma-rr2-p5-nr2fma.c",
2045 "src/math/sigmoid-neonfma-rr2-p5-nr2recps.c",
Marat Dukhan84000762020-06-29 18:38:43 -07002046 "src/math/sqrt-neonfma-nr1fma.c",
Marat Dukhan84000762020-06-29 18:38:43 -07002047 "src/math/sqrt-neonfma-nr1rsqrts1fma1adj.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002048 "src/math/sqrt-neonfma-nr2fma.c",
2049 "src/math/sqrt-neonfma-nr2fma1adj.c",
2050 "src/math/sqrt-neonfma-nr3fma.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002051]
2052
2053AARCH64_NEONFMA_UKERNELS = [
Marat Dukhan56b10cd2020-05-18 09:35:49 -07002054 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neonfma-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07002055 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neonfma-2x2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002056 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neonfma-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07002057 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neonfma-2x2.c",
Marat Dukhan56b10cd2020-05-18 09:35:49 -07002058 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neonfma-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07002059 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neonfma-2x2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002060 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neonfma-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07002061 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neonfma-2x2.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -07002062 "src/f32-conv-hwc2chw/3x3s2p1c3x4-neonfma-2x2.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07002063 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-1x4-acc2.c",
2064 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-1x4-acc3.c",
2065 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-1x4-acc4.c",
Marat Dukhan1268a242020-10-24 00:36:32 -07002066 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-1x4.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07002067 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-2x4-acc2.c",
Marat Dukhan1268a242020-10-24 00:36:32 -07002068 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-2x4.c",
2069 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-3x4.c",
2070 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-4x4.c",
2071 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-5x4.c",
2072 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-6x4.c",
Marat Dukhan82f0c322020-10-25 19:17:35 -07002073 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neonfma-1x4-acc2.c",
2074 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neonfma-1x4-acc3.c",
2075 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neonfma-1x4-acc4.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07002076 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neonfma-1x4.c",
Marat Dukhan82f0c322020-10-25 19:17:35 -07002077 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neonfma-2x4-acc2.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07002078 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neonfma-2x4.c",
2079 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neonfma-3x4.c",
2080 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neonfma-4x4.c",
Marat Dukhan149f0ea2020-10-26 12:50:33 -07002081 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-1x4-acc2.c",
2082 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-1x4-acc3.c",
2083 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-1x4-acc4.c",
2084 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-1x4-acc5.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07002085 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-1x4.c",
Marat Dukhan149f0ea2020-10-26 12:50:33 -07002086 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-2x4-acc2.c",
2087 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-2x4-acc3.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07002088 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-2x4.c",
Marat Dukhan149f0ea2020-10-26 12:50:33 -07002089 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-3x4-acc2.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07002090 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-3x4.c",
Marat Dukhan149f0ea2020-10-26 12:50:33 -07002091 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-4x4-acc2.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07002092 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-4x4.c",
2093 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-5x4.c",
Marat Dukhan30d4b252020-10-29 16:33:22 -07002094 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neonfma-1x4-acc2.c",
2095 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neonfma-1x4-acc3.c",
2096 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neonfma-1x4-acc4.c",
2097 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neonfma-1x4-acc5.c",
2098 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neonfma-1x4.c",
2099 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neonfma-2x4-acc2.c",
2100 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neonfma-2x4-acc3.c",
2101 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neonfma-2x4.c",
Marat Dukhan30d4b252020-10-29 16:33:22 -07002102 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neonfma-3x4-acc2.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07002103 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neonfma-3x4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002104 "src/f32-gemm/gen-inc/1x8inc-minmax-neonfma-lane-ld64.c",
2105 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-lane-ld64.c",
2106 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-lane-ld128.c",
2107 "src/f32-gemm/gen-inc/5x8inc-minmax-neonfma-lane-ld64.c",
2108 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-lane-ld64.c",
2109 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-lane-ld128.c",
2110 "src/f32-gemm/gen/1x8-minmax-neonfma-lane-ld64.c",
2111 "src/f32-gemm/gen/4x2-minmax-neonfma-lane-ld64.c",
2112 "src/f32-gemm/gen/4x8-minmax-neonfma-lane-ld64.c",
2113 "src/f32-gemm/gen/4x8-minmax-neonfma-lane-ld128.c",
2114 "src/f32-gemm/gen/5x8-minmax-neonfma-lane-ld64.c",
2115 "src/f32-gemm/gen/6x8-minmax-neonfma-lane-ld64.c",
2116 "src/f32-gemm/gen/6x8-minmax-neonfma-lane-ld128.c",
2117 "src/f32-igemm/gen/1x8-minmax-neonfma-lane-ld64.c",
2118 "src/f32-igemm/gen/4x2-minmax-neonfma-lane-ld64.c",
2119 "src/f32-igemm/gen/4x4-minmax-neonfma-lane-ld64.c",
2120 "src/f32-igemm/gen/4x8-minmax-neonfma-lane-ld64.c",
2121 "src/f32-igemm/gen/4x8-minmax-neonfma-lane-ld128.c",
2122 "src/f32-igemm/gen/6x8-minmax-neonfma-lane-ld64.c",
2123 "src/f32-igemm/gen/6x8-minmax-neonfma-lane-ld128.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -08002124 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x4.c",
2125 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x8.c",
2126 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x12.c",
2127 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x16.c",
2128 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x20.c",
2129 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x24.c",
2130 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x4.c",
2131 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x8.c",
2132 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x12.c",
2133 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x16.c",
2134 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x20.c",
2135 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x24.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002136 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x4.c",
2137 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x8.c",
2138 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x12.c",
2139 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x16.c",
2140 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x20.c",
2141 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x24.c",
Marat Dukhan355ab432020-04-09 19:01:52 -07002142 "src/f32-spmm/gen/4x2-minmax-neonfma.c",
2143 "src/f32-spmm/gen/4x4-minmax-neonfma.c",
Marat Dukhan355ab432020-04-09 19:01:52 -07002144 "src/f32-spmm/gen/8x2-minmax-neonfma.c",
2145 "src/f32-spmm/gen/8x4-minmax-neonfma.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002146 "src/f32-spmm/gen/12x2-minmax-neonfma.c",
2147 "src/f32-spmm/gen/12x4-minmax-neonfma.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002148 "src/f32-spmm/gen/16x2-minmax-neonfma.c",
2149 "src/f32-spmm/gen/16x4-minmax-neonfma.c",
Frank Barchard846c0c62020-10-26 15:01:39 -07002150 "src/f32-spmm/gen/32x2-minmax-neonfma.c",
2151 "src/f32-spmm/gen/32x4-minmax-neonfma.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002152 "src/f32-vbinary/gen/vdiv-minmax-neon-x4.c",
2153 "src/f32-vbinary/gen/vdiv-minmax-neon-x8.c",
2154 "src/f32-vbinary/gen/vdivc-minmax-neon-x4.c",
2155 "src/f32-vbinary/gen/vdivc-minmax-neon-x8.c",
2156 "src/f32-vbinary/gen/vrdivc-minmax-neon-x4.c",
2157 "src/f32-vbinary/gen/vrdivc-minmax-neon-x8.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07002158 "src/f32-vsqrt/gen/neon-sqrt-x4.c",
2159 "src/f32-vsqrt/gen/neon-sqrt-x8.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08002160 "src/math/sigmoid-neonfma-rr1-lut64-p2-div.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002161 "src/math/sigmoid-neonfma-rr1-lut2048-p1-div.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08002162 "src/math/sigmoid-neonfma-rr1-p5-div.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08002163 "src/math/sigmoid-neonfma-rr2-lut64-p2-div.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002164 "src/math/sigmoid-neonfma-rr2-lut2048-p1-div.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08002165 "src/math/sigmoid-neonfma-rr2-p5-div.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002166]
2167
Marat Dukhan8853b822020-05-07 12:19:01 -07002168NEONV8_UKERNELS = [
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07002169 "src/f32-vrnd/gen/vrndd-neonv8-x4.c",
2170 "src/f32-vrnd/gen/vrndd-neonv8-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002171 "src/f32-vrnd/gen/vrndne-neonv8-x4.c",
2172 "src/f32-vrnd/gen/vrndne-neonv8-x8.c",
2173 "src/f32-vrnd/gen/vrndu-neonv8-x4.c",
2174 "src/f32-vrnd/gen/vrndu-neonv8-x8.c",
2175 "src/f32-vrnd/gen/vrndz-neonv8-x4.c",
2176 "src/f32-vrnd/gen/vrndz-neonv8-x8.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07002177 "src/math/roundd-neonv8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002178 "src/math/roundne-neonv8.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07002179 "src/math/roundu-neonv8.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07002180 "src/math/roundz-neonv8.c",
Marat Dukhan8853b822020-05-07 12:19:01 -07002181]
2182
Marat Dukhan08c4a432019-10-03 09:29:21 -07002183AARCH64_NEONFP16ARITH_UKERNELS = [
Frank Barchardb1966592020-05-12 13:47:06 -07002184 "src/f16-clamp/gen/neonfp16arith-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002185 "src/f16-clamp/gen/neonfp16arith-x16.c",
Frank Barchard5a599a62020-06-04 20:12:44 -07002186 "src/f16-dwconv/gen/up8x4-minmax-neonfp16arith-acc2.c",
2187 "src/f16-dwconv/gen/up8x4-minmax-neonfp16arith.c",
2188 "src/f16-dwconv/gen/up8x9-minmax-neonfp16arith-acc2.c",
2189 "src/f16-dwconv/gen/up8x9-minmax-neonfp16arith.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002190 "src/f16-dwconv/gen/up8x25-minmax-neonfp16arith-acc2.c",
2191 "src/f16-dwconv/gen/up8x25-minmax-neonfp16arith.c",
2192 "src/f16-dwconv/gen/up16x4-minmax-neonfp16arith-acc2.c",
2193 "src/f16-dwconv/gen/up16x4-minmax-neonfp16arith.c",
2194 "src/f16-dwconv/gen/up16x9-minmax-neonfp16arith-acc2.c",
2195 "src/f16-dwconv/gen/up16x9-minmax-neonfp16arith.c",
2196 "src/f16-dwconv/gen/up16x25-minmax-neonfp16arith-acc2.c",
2197 "src/f16-dwconv/gen/up16x25-minmax-neonfp16arith.c",
Frank Barchard0bb49a72020-06-04 11:35:11 -07002198 "src/f16-gavgpool/7p7x-minmax-neonfp16arith-c8.c",
2199 "src/f16-gavgpool/7x-minmax-neonfp16arith-c8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002200 "src/f16-gemm/gen-inc/1x8inc-minmax-neonfp16arith-ld64.c",
2201 "src/f16-gemm/gen-inc/1x16inc-minmax-neonfp16arith-ld64.c",
2202 "src/f16-gemm/gen-inc/4x8inc-minmax-neonfp16arith-ld64.c",
2203 "src/f16-gemm/gen-inc/4x16inc-minmax-neonfp16arith-ld64.c",
2204 "src/f16-gemm/gen-inc/6x8inc-minmax-neonfp16arith-ld64.c",
2205 "src/f16-gemm/gen-inc/6x16inc-minmax-neonfp16arith-ld64.c",
2206 "src/f16-gemm/gen-inc/8x8inc-minmax-neonfp16arith-ld64.c",
2207 "src/f16-gemm/gen-inc/8x16inc-minmax-neonfp16arith-ld64.c",
2208 "src/f16-gemm/gen/1x8-minmax-neonfp16arith-ld64.c",
2209 "src/f16-gemm/gen/1x16-minmax-neonfp16arith-ld64.c",
2210 "src/f16-gemm/gen/4x8-minmax-neonfp16arith-ld64.c",
2211 "src/f16-gemm/gen/4x16-minmax-neonfp16arith-ld64.c",
2212 "src/f16-gemm/gen/6x8-minmax-neonfp16arith-ld64.c",
2213 "src/f16-gemm/gen/6x16-minmax-neonfp16arith-ld64.c",
2214 "src/f16-gemm/gen/8x8-minmax-neonfp16arith-ld64.c",
2215 "src/f16-gemm/gen/8x16-minmax-neonfp16arith-ld64.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07002216 "src/f16-hswish/gen/hswish-neonfp16arith-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002217 "src/f16-hswish/gen/hswish-neonfp16arith-x16.c",
2218 "src/f16-igemm/gen/1x8-minmax-neonfp16arith-ld64.c",
2219 "src/f16-igemm/gen/1x16-minmax-neonfp16arith-ld64.c",
2220 "src/f16-igemm/gen/4x8-minmax-neonfp16arith-ld64.c",
2221 "src/f16-igemm/gen/4x16-minmax-neonfp16arith-ld64.c",
2222 "src/f16-igemm/gen/6x8-minmax-neonfp16arith-ld64.c",
2223 "src/f16-igemm/gen/6x16-minmax-neonfp16arith-ld64.c",
2224 "src/f16-igemm/gen/8x8-minmax-neonfp16arith-ld64.c",
2225 "src/f16-igemm/gen/8x16-minmax-neonfp16arith-ld64.c",
Frank Barchardb1966592020-05-12 13:47:06 -07002226 "src/f16-prelu/gen/neonfp16arith-2x8.c",
Frank Barcharda5316982020-07-23 13:19:28 -07002227 "src/f16-prelu/gen/neonfp16arith-2x16.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07002228 "src/f16-relu/gen/neonfp16arith-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002229 "src/f16-relu/gen/neonfp16arith-x16.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07002230 "src/f16-spmm/gen/8x1-minmax-neonfp16arith-x2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002231 "src/f16-spmm/gen/8x1-minmax-neonfp16arith.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07002232 "src/f16-spmm/gen/16x1-minmax-neonfp16arith-x2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002233 "src/f16-spmm/gen/16x1-minmax-neonfp16arith.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07002234 "src/f16-spmm/gen/24x1-minmax-neonfp16arith-x2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002235 "src/f16-spmm/gen/24x1-minmax-neonfp16arith.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07002236 "src/f16-spmm/gen/32x1-minmax-neonfp16arith-x2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002237 "src/f16-spmm/gen/32x1-minmax-neonfp16arith.c",
2238 "src/f16-vbinary/gen/vadd-minmax-neonfp16arith-x8.c",
2239 "src/f16-vbinary/gen/vadd-minmax-neonfp16arith-x16.c",
2240 "src/f16-vbinary/gen/vaddc-minmax-neonfp16arith-x8.c",
2241 "src/f16-vbinary/gen/vaddc-minmax-neonfp16arith-x16.c",
2242 "src/f16-vbinary/gen/vdiv-minmax-neonfp16arith-x8.c",
2243 "src/f16-vbinary/gen/vdiv-minmax-neonfp16arith-x16.c",
2244 "src/f16-vbinary/gen/vdivc-minmax-neonfp16arith-x8.c",
2245 "src/f16-vbinary/gen/vdivc-minmax-neonfp16arith-x16.c",
2246 "src/f16-vbinary/gen/vmax-neonfp16arith-x8.c",
2247 "src/f16-vbinary/gen/vmax-neonfp16arith-x16.c",
2248 "src/f16-vbinary/gen/vmaxc-neonfp16arith-x8.c",
2249 "src/f16-vbinary/gen/vmaxc-neonfp16arith-x16.c",
2250 "src/f16-vbinary/gen/vmin-neonfp16arith-x8.c",
2251 "src/f16-vbinary/gen/vmin-neonfp16arith-x16.c",
2252 "src/f16-vbinary/gen/vminc-neonfp16arith-x8.c",
2253 "src/f16-vbinary/gen/vminc-neonfp16arith-x16.c",
2254 "src/f16-vbinary/gen/vmul-minmax-neonfp16arith-x8.c",
2255 "src/f16-vbinary/gen/vmul-minmax-neonfp16arith-x16.c",
2256 "src/f16-vbinary/gen/vmulc-minmax-neonfp16arith-x8.c",
2257 "src/f16-vbinary/gen/vmulc-minmax-neonfp16arith-x16.c",
2258 "src/f16-vbinary/gen/vrdivc-minmax-neonfp16arith-x8.c",
2259 "src/f16-vbinary/gen/vrdivc-minmax-neonfp16arith-x16.c",
2260 "src/f16-vbinary/gen/vrsubc-minmax-neonfp16arith-x8.c",
2261 "src/f16-vbinary/gen/vrsubc-minmax-neonfp16arith-x16.c",
2262 "src/f16-vbinary/gen/vsub-minmax-neonfp16arith-x8.c",
2263 "src/f16-vbinary/gen/vsub-minmax-neonfp16arith-x16.c",
2264 "src/f16-vbinary/gen/vsubc-minmax-neonfp16arith-x8.c",
2265 "src/f16-vbinary/gen/vsubc-minmax-neonfp16arith-x16.c",
2266 "src/f16-vmulcaddc/gen/c8-minmax-neonfp16arith-2x.c",
2267 "src/f16-vmulcaddc/gen/c16-minmax-neonfp16arith-2x.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002268]
2269
Benoit Jacoba9644732020-08-13 12:48:55 -07002270NEONDOT_UKERNELS = [
Marat Dukhand4c83032020-09-09 23:20:12 -07002271 "src/qs8-gemm/gen/1x8c4-minmax-neondot.c",
Marat Dukhand4c83032020-09-09 23:20:12 -07002272 "src/qs8-gemm/gen/1x16c4-minmax-neondot.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002273 "src/qs8-gemm/gen/4x8c4-minmax-neondot.c",
Benoit Jacoba9644732020-08-13 12:48:55 -07002274 "src/qs8-gemm/gen/4x16c4-minmax-neondot.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002275 "src/qs8-gemm/gen/6x8c4-minmax-neondot.c",
Frank Barchard2fa17452020-09-22 09:51:46 -07002276 "src/qs8-gemm/gen/6x16c4-minmax-neondot.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002277 "src/qs8-gemm/gen/8x8c4-minmax-neondot.c",
Frank Barcharda48848f2020-09-24 00:59:24 -07002278 "src/qs8-gemm/gen/8x16c4-minmax-neondot.c",
Marat Dukhand4c83032020-09-09 23:20:12 -07002279 "src/qs8-igemm/gen/1x8c4-minmax-neondot.c",
Marat Dukhand4c83032020-09-09 23:20:12 -07002280 "src/qs8-igemm/gen/1x16c4-minmax-neondot.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002281 "src/qs8-igemm/gen/4x8c4-minmax-neondot.c",
Benoit Jacoba05487f2020-08-14 13:37:21 -07002282 "src/qs8-igemm/gen/4x16c4-minmax-neondot.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002283 "src/qs8-igemm/gen/6x8c4-minmax-neondot.c",
Frank Barchard2fa17452020-09-22 09:51:46 -07002284 "src/qs8-igemm/gen/6x16c4-minmax-neondot.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002285 "src/qs8-igemm/gen/8x8c4-minmax-neondot.c",
Frank Barcharda48848f2020-09-24 00:59:24 -07002286 "src/qs8-igemm/gen/8x16c4-minmax-neondot.c",
Benoit Jacoba9644732020-08-13 12:48:55 -07002287]
2288
Marat Dukhan08c4a432019-10-03 09:29:21 -07002289SSE_UKERNELS = [
Marat Dukhan99936602020-04-11 16:47:01 -07002290 "src/f32-avgpool/9p8x-minmax-sse-c4.c",
2291 "src/f32-avgpool/9x-minmax-sse-c4.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -07002292 "src/f32-clamp/gen/sse-x4.c",
2293 "src/f32-clamp/gen/sse-x8.c",
Erich Elsenb1233402020-06-08 15:53:15 -07002294 "src/f32-conv-hwc2chw/3x3s2p1c3x4-sse-1x1.c",
2295 "src/f32-conv-hwc2chw/3x3s2p1c3x4-sse-2x2.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002296 "src/f32-dwconv/gen/up4x4-minmax-sse-acc2.c",
2297 "src/f32-dwconv/gen/up4x4-minmax-sse.c",
2298 "src/f32-dwconv/gen/up4x9-minmax-sse-acc2.c",
2299 "src/f32-dwconv/gen/up4x9-minmax-sse.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002300 "src/f32-dwconv/gen/up4x25-minmax-sse-acc2.c",
2301 "src/f32-dwconv/gen/up4x25-minmax-sse.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002302 "src/f32-dwconv/gen/up8x4-minmax-sse-acc2.c",
2303 "src/f32-dwconv/gen/up8x4-minmax-sse.c",
2304 "src/f32-dwconv/gen/up8x9-minmax-sse-acc2.c",
2305 "src/f32-dwconv/gen/up8x9-minmax-sse.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002306 "src/f32-dwconv/gen/up8x25-minmax-sse-acc2.c",
2307 "src/f32-dwconv/gen/up8x25-minmax-sse.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07002308 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-1x4-acc2.c",
2309 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-1x4-acc3.c",
2310 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-1x4-acc4.c",
Marat Dukhan470078a2020-10-23 22:36:52 -07002311 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-1x4.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07002312 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-2x4-acc2.c",
Marat Dukhan470078a2020-10-23 22:36:52 -07002313 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-2x4.c",
2314 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-3x4.c",
2315 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-4x4.c",
2316 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-5x4.c",
2317 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-6x4.c",
Marat Dukhan0ff97182020-10-25 19:14:03 -07002318 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-sse-1x4-acc2.c",
2319 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-sse-1x4-acc3.c",
2320 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-sse-1x4-acc4.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07002321 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-sse-1x4.c",
Marat Dukhan0ff97182020-10-25 19:14:03 -07002322 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-sse-2x4-acc2.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07002323 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-sse-2x4.c",
2324 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-sse-3x4.c",
2325 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-sse-4x4.c",
Marat Dukhand0503892020-10-30 08:22:04 -07002326 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-1x4-acc2.c",
2327 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-1x4-acc3.c",
2328 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-1x4-acc4.c",
2329 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-1x4-acc5.c",
2330 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-1x4.c",
2331 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-2x4-acc2.c",
2332 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-2x4-acc3.c",
2333 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-2x4.c",
2334 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-3x4-acc2.c",
2335 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-3x4.c",
2336 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-4x4-acc2.c",
2337 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-4x4.c",
2338 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-5x4.c",
Marat Dukhanccca2142020-10-30 17:32:45 -07002339 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-sse-1x4-acc2.c",
2340 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-sse-1x4-acc3.c",
2341 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-sse-1x4-acc4.c",
2342 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-sse-1x4-acc5.c",
2343 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-sse-1x4.c",
2344 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-sse-2x4-acc2.c",
2345 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-sse-2x4-acc3.c",
2346 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-sse-2x4.c",
Marat Dukhanccca2142020-10-30 17:32:45 -07002347 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-sse-3x4-acc2.c",
Frank Barchard8ef44cd2020-11-03 12:30:23 -08002348 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-sse-3x4.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -07002349 "src/f32-gavgpool-cw/sse-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -07002350 "src/f32-gavgpool/7p7x-minmax-sse-c4.c",
2351 "src/f32-gavgpool/7x-minmax-sse-c4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002352 "src/f32-gemm/gen-inc/1x8inc-minmax-sse-dup.c",
2353 "src/f32-gemm/gen-inc/1x8inc-minmax-sse-load1.c",
2354 "src/f32-gemm/gen-inc/1x8s4inc-minmax-sse.c",
Marat Dukhan802fcae2020-12-11 14:37:25 -08002355 "src/f32-gemm/gen-inc/3x8inc-minmax-sse-dup.c",
2356 "src/f32-gemm/gen-inc/3x8inc-minmax-sse-load1.c",
2357 "src/f32-gemm/gen-inc/3x8s4inc-minmax-sse.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002358 "src/f32-gemm/gen-inc/4x8inc-minmax-sse-dup.c",
2359 "src/f32-gemm/gen-inc/4x8inc-minmax-sse-load1.c",
2360 "src/f32-gemm/gen-inc/4x8s4inc-minmax-sse.c",
Marat Dukhan802fcae2020-12-11 14:37:25 -08002361 "src/f32-gemm/gen-inc/5x8inc-minmax-sse-dup.c",
2362 "src/f32-gemm/gen-inc/5x8inc-minmax-sse-load1.c",
2363 "src/f32-gemm/gen-inc/5x8s4inc-minmax-sse.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002364 "src/f32-gemm/gen/1x8-minmax-sse-dup.c",
2365 "src/f32-gemm/gen/1x8-minmax-sse-load1.c",
2366 "src/f32-gemm/gen/1x8s4-minmax-sse.c",
Marat Dukhan802fcae2020-12-11 14:37:25 -08002367 "src/f32-gemm/gen/3x8-minmax-sse-dup.c",
2368 "src/f32-gemm/gen/3x8-minmax-sse-load1.c",
2369 "src/f32-gemm/gen/3x8s4-minmax-sse.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002370 "src/f32-gemm/gen/4x2c4-minmax-sse.c",
2371 "src/f32-gemm/gen/4x8-minmax-sse-dup.c",
2372 "src/f32-gemm/gen/4x8-minmax-sse-load1.c",
2373 "src/f32-gemm/gen/4x8s4-minmax-sse.c",
Marat Dukhan802fcae2020-12-11 14:37:25 -08002374 "src/f32-gemm/gen/5x8-minmax-sse-dup.c",
2375 "src/f32-gemm/gen/5x8-minmax-sse-load1.c",
2376 "src/f32-gemm/gen/5x8s4-minmax-sse.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07002377 "src/f32-hswish/gen/hswish-sse-x4.c",
2378 "src/f32-hswish/gen/hswish-sse-x8.c",
Marat Dukhan660fd192020-03-10 04:55:30 -07002379 "src/f32-ibilinear/gen/sse-c4.c",
2380 "src/f32-ibilinear/gen/sse-c8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002381 "src/f32-igemm/gen/1x8-minmax-sse-dup.c",
2382 "src/f32-igemm/gen/1x8-minmax-sse-load1.c",
2383 "src/f32-igemm/gen/1x8s4-minmax-sse.c",
Marat Dukhan802fcae2020-12-11 14:37:25 -08002384 "src/f32-igemm/gen/3x8-minmax-sse-dup.c",
2385 "src/f32-igemm/gen/3x8-minmax-sse-load1.c",
2386 "src/f32-igemm/gen/3x8s4-minmax-sse.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002387 "src/f32-igemm/gen/4x2c4-minmax-sse.c",
2388 "src/f32-igemm/gen/4x8-minmax-sse-dup.c",
2389 "src/f32-igemm/gen/4x8-minmax-sse-load1.c",
2390 "src/f32-igemm/gen/4x8s4-minmax-sse.c",
Marat Dukhan802fcae2020-12-11 14:37:25 -08002391 "src/f32-igemm/gen/5x8-minmax-sse-dup.c",
2392 "src/f32-igemm/gen/5x8-minmax-sse-load1.c",
2393 "src/f32-igemm/gen/5x8s4-minmax-sse.c",
Marat Dukhan99936602020-04-11 16:47:01 -07002394 "src/f32-maxpool/9p8x-minmax-sse-c4.c",
2395 "src/f32-pavgpool/9p8x-minmax-sse-c4.c",
2396 "src/f32-pavgpool/9x-minmax-sse-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002397 "src/f32-ppmm/gen/4x8-minmax-sse.c",
Marat Dukhan39b5e942020-06-24 15:03:48 -07002398 "src/f32-prelu/gen/sse-2x4.c",
2399 "src/f32-prelu/gen/sse-2x8.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07002400 "src/f32-relu/gen/sse-x4.c",
2401 "src/f32-relu/gen/sse-x8.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002402 "src/f32-rmax/sse.c",
Marat Dukhan355ab432020-04-09 19:01:52 -07002403 "src/f32-spmm/gen/4x1-minmax-sse.c",
2404 "src/f32-spmm/gen/8x1-minmax-sse.c",
Erich Elsen6e80fdc2020-06-09 15:35:37 -07002405 "src/f32-spmm/gen/16x1-minmax-sse.c",
Frank Barchard846c0c62020-10-26 15:01:39 -07002406 "src/f32-spmm/gen/32x1-minmax-sse.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002407 "src/f32-vbinary/gen/vadd-minmax-sse-x4.c",
2408 "src/f32-vbinary/gen/vadd-minmax-sse-x8.c",
2409 "src/f32-vbinary/gen/vaddc-minmax-sse-x4.c",
2410 "src/f32-vbinary/gen/vaddc-minmax-sse-x8.c",
2411 "src/f32-vbinary/gen/vdiv-minmax-sse-x4.c",
2412 "src/f32-vbinary/gen/vdiv-minmax-sse-x8.c",
2413 "src/f32-vbinary/gen/vdivc-minmax-sse-x4.c",
2414 "src/f32-vbinary/gen/vdivc-minmax-sse-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -08002415 "src/f32-vbinary/gen/vmax-sse-x4.c",
2416 "src/f32-vbinary/gen/vmax-sse-x8.c",
2417 "src/f32-vbinary/gen/vmaxc-sse-x4.c",
2418 "src/f32-vbinary/gen/vmaxc-sse-x8.c",
2419 "src/f32-vbinary/gen/vmin-sse-x4.c",
2420 "src/f32-vbinary/gen/vmin-sse-x8.c",
2421 "src/f32-vbinary/gen/vminc-sse-x4.c",
2422 "src/f32-vbinary/gen/vminc-sse-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002423 "src/f32-vbinary/gen/vmul-minmax-sse-x4.c",
2424 "src/f32-vbinary/gen/vmul-minmax-sse-x8.c",
2425 "src/f32-vbinary/gen/vmulc-minmax-sse-x4.c",
2426 "src/f32-vbinary/gen/vmulc-minmax-sse-x8.c",
2427 "src/f32-vbinary/gen/vrdivc-minmax-sse-x4.c",
2428 "src/f32-vbinary/gen/vrdivc-minmax-sse-x8.c",
2429 "src/f32-vbinary/gen/vrsubc-minmax-sse-x4.c",
2430 "src/f32-vbinary/gen/vrsubc-minmax-sse-x8.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07002431 "src/f32-vbinary/gen/vsqrdiff-sse-x4.c",
2432 "src/f32-vbinary/gen/vsqrdiff-sse-x8.c",
2433 "src/f32-vbinary/gen/vsqrdiffc-sse-x4.c",
2434 "src/f32-vbinary/gen/vsqrdiffc-sse-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002435 "src/f32-vbinary/gen/vsub-minmax-sse-x4.c",
2436 "src/f32-vbinary/gen/vsub-minmax-sse-x8.c",
2437 "src/f32-vbinary/gen/vsubc-minmax-sse-x4.c",
2438 "src/f32-vbinary/gen/vsubc-minmax-sse-x8.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07002439 "src/f32-vlrelu/gen/vlrelu-sse-x4.c",
2440 "src/f32-vlrelu/gen/vlrelu-sse-x8.c",
Marat Dukhan99936602020-04-11 16:47:01 -07002441 "src/f32-vmulcaddc/gen/c4-minmax-sse-2x.c",
2442 "src/f32-vmulcaddc/gen/c8-minmax-sse-2x.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07002443 "src/f32-vsqrt/gen/sse-sqrt-x4.c",
2444 "src/f32-vsqrt/gen/sse-sqrt-x8.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07002445 "src/f32-vunary/gen/vabs-sse-x4.c",
2446 "src/f32-vunary/gen/vabs-sse-x8.c",
2447 "src/f32-vunary/gen/vneg-sse-x4.c",
2448 "src/f32-vunary/gen/vneg-sse-x8.c",
2449 "src/f32-vunary/gen/vsqr-sse-x4.c",
2450 "src/f32-vunary/gen/vsqr-sse-x8.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07002451 "src/math/roundd-sse-addsub.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002452 "src/math/roundne-sse-addsub.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07002453 "src/math/roundu-sse-addsub.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07002454 "src/math/roundz-sse-addsub.c",
Marat Dukhan84000762020-06-29 18:38:43 -07002455 "src/math/sqrt-sse-hh1mac.c",
2456 "src/math/sqrt-sse-nr1mac.c",
2457 "src/math/sqrt-sse-nr2mac.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002458 "src/x32-fill/sse.c",
2459 "src/x32-packx/x4-sse.c",
2460 "src/x32-pad/sse.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002461]
2462
2463SSE2_UKERNELS = [
Marat Dukhan329da642019-11-19 21:44:39 -08002464 "src/f32-argmaxpool/4x-sse2-c4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002465 "src/f32-argmaxpool/9p8x-sse2-c4.c",
Marat Dukhan329da642019-11-19 21:44:39 -08002466 "src/f32-argmaxpool/9x-sse2-c4.c",
Marat Dukhan802fcae2020-12-11 14:37:25 -08002467 "src/f32-gemm/gen-inc/1x8inc-minmax-sse2-dup.c",
2468 "src/f32-gemm/gen-inc/3x8inc-minmax-sse2-dup.c",
2469 "src/f32-gemm/gen-inc/4x8inc-minmax-sse2-dup.c",
2470 "src/f32-gemm/gen-inc/5x8inc-minmax-sse2-dup.c",
2471 "src/f32-gemm/gen/1x8-minmax-sse2-dup.c",
2472 "src/f32-gemm/gen/3x8-minmax-sse2-dup.c",
2473 "src/f32-gemm/gen/4x8-minmax-sse2-dup.c",
2474 "src/f32-gemm/gen/5x8-minmax-sse2-dup.c",
2475 "src/f32-igemm/gen/1x8-minmax-sse2-dup.c",
2476 "src/f32-igemm/gen/3x8-minmax-sse2-dup.c",
2477 "src/f32-igemm/gen/4x8-minmax-sse2-dup.c",
2478 "src/f32-igemm/gen/5x8-minmax-sse2-dup.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08002479 "src/f32-prelu/gen/sse2-2x4.c",
2480 "src/f32-prelu/gen/sse2-2x8.c",
Marat Dukhanb39689d2020-01-24 13:32:20 -08002481 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x4.c",
Marat Dukhanb39689d2020-01-24 13:32:20 -08002482 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x8-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002483 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x8.c",
Marat Dukhanb39689d2020-01-24 13:32:20 -08002484 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12-acc2.c",
2485 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002486 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12.c",
Marat Dukhanb39689d2020-01-24 13:32:20 -08002487 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16-acc2.c",
2488 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002489 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16.c",
Marat Dukhanb39689d2020-01-24 13:32:20 -08002490 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20-acc2.c",
2491 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002492 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20.c",
Marat Dukhand243c1a2020-09-17 18:48:10 -07002493 "src/f32-sigmoid/gen/sse2-lut64-p2-div-x4.c",
2494 "src/f32-sigmoid/gen/sse2-lut64-p2-div-x8.c",
2495 "src/f32-sigmoid/gen/sse2-lut64-p2-div-x12.c",
2496 "src/f32-sigmoid/gen/sse2-lut64-p2-div-x16.c",
2497 "src/f32-sigmoid/gen/sse2-lut64-p2-div-x20.c",
2498 "src/f32-sigmoid/gen/sse2-lut64-p2-div-x24.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002499 "src/f32-sigmoid/gen/sse2-p5-div-x4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08002500 "src/f32-sigmoid/gen/sse2-p5-div-x8.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002501 "src/f32-sigmoid/gen/sse2-p5-div-x12.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08002502 "src/f32-sigmoid/gen/sse2-p5-div-x16.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002503 "src/f32-sigmoid/gen/sse2-p5-div-x20.c",
2504 "src/f32-sigmoid/gen/sse2-p5-div-x24.c",
Marat Dukhaned6baaf2020-12-01 15:07:08 -08002505 "src/f32-velu/gen/velu-sse2-rr2-lut16-p3-x4.c",
2506 "src/f32-velu/gen/velu-sse2-rr2-lut16-p3-x8.c",
2507 "src/f32-velu/gen/velu-sse2-rr2-lut16-p3-x12.c",
2508 "src/f32-velu/gen/velu-sse2-rr2-lut16-p3-x16.c",
2509 "src/f32-velu/gen/velu-sse2-rr2-lut16-p3-x20.c",
2510 "src/f32-velu/gen/velu-sse2-rr2-lut16-p3-x24.c",
2511 "src/f32-velu/gen/velu-sse2-rr2-p6-x4.c",
2512 "src/f32-velu/gen/velu-sse2-rr2-p6-x8.c",
2513 "src/f32-velu/gen/velu-sse2-rr2-p6-x12.c",
2514 "src/f32-velu/gen/velu-sse2-rr2-p6-x16.c",
2515 "src/f32-velu/gen/velu-sse2-rr2-p6-x20.c",
2516 "src/f32-velu/gen/velu-sse2-rr2-p6-x24.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07002517 "src/f32-vlrelu/gen/vlrelu-sse2-x4.c",
2518 "src/f32-vlrelu/gen/vlrelu-sse2-x8.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07002519 "src/f32-vrnd/gen/vrndd-sse2-x4.c",
2520 "src/f32-vrnd/gen/vrndd-sse2-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002521 "src/f32-vrnd/gen/vrndne-sse2-x4.c",
2522 "src/f32-vrnd/gen/vrndne-sse2-x8.c",
2523 "src/f32-vrnd/gen/vrndu-sse2-x4.c",
2524 "src/f32-vrnd/gen/vrndu-sse2-x8.c",
2525 "src/f32-vrnd/gen/vrndz-sse2-x4.c",
2526 "src/f32-vrnd/gen/vrndz-sse2-x8.c",
Marat Dukhanb7633f22020-11-20 16:34:56 -08002527 "src/math/exp-sse2-rr2-lut64-p2.c",
2528 "src/math/exp-sse2-rr2-p5.c",
Marat Dukhande390d42020-11-29 19:32:18 -08002529 "src/math/expm1minus-sse2-rr2-lut16-p3.c",
Marat Dukhana438aca2020-11-20 15:45:01 -08002530 "src/math/expm1minus-sse2-rr2-p6.c",
Frank Barchard3b800452020-11-22 12:12:35 -08002531 "src/math/expminus-sse2-rr2-p5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002532 "src/math/roundd-sse2-cvt.c",
2533 "src/math/roundne-sse2-cvt.c",
2534 "src/math/roundu-sse2-cvt.c",
2535 "src/math/roundz-sse2-cvt.c",
2536 "src/math/sigmoid-sse2-rr2-lut64-p2-div.c",
2537 "src/math/sigmoid-sse2-rr2-lut64-p2-nr1.c",
2538 "src/math/sigmoid-sse2-rr2-lut64-p2-nr2.c",
2539 "src/math/sigmoid-sse2-rr2-p5-div.c",
2540 "src/math/sigmoid-sse2-rr2-p5-nr1.c",
2541 "src/math/sigmoid-sse2-rr2-p5-nr2.c",
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07002542 "src/qs8-dwconv/gen/up8x9-minmax-sse2-mul16.c",
2543 "src/qs8-dwconv/gen/up16x9-minmax-sse2-mul16.c",
2544 "src/qs8-dwconv/gen/up24x9-minmax-sse2-mul16.c",
Marat Dukhan159688f2020-08-06 10:34:29 -07002545 "src/qs8-gavgpool/gen/7p7x-minmax-sse2-c8-acc2.c",
2546 "src/qs8-gavgpool/gen/7p7x-minmax-sse2-c16-acc2.c",
2547 "src/qs8-gavgpool/gen/7p7x-minmax-sse2-c24-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002548 "src/qs8-gavgpool/gen/7x-minmax-sse2-c8-acc2.c",
2549 "src/qs8-gavgpool/gen/7x-minmax-sse2-c16-acc2.c",
2550 "src/qs8-gavgpool/gen/7x-minmax-sse2-c24-acc2.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002551 "src/qs8-gemm/gen/1x4c2-minmax-sse2-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002552 "src/qs8-gemm/gen/1x4c2-minmax-sse2-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002553 "src/qs8-gemm/gen/1x4c2-xw-minmax-sse2.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002554 "src/qs8-gemm/gen/1x4c8-minmax-sse2-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002555 "src/qs8-gemm/gen/1x4c8-minmax-sse2-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002556 "src/qs8-gemm/gen/1x4c8-xw-minmax-sse2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002557 "src/qs8-gemm/gen/2x4c8-minmax-sse2-ld64.c",
2558 "src/qs8-gemm/gen/2x4c8-minmax-sse2-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002559 "src/qs8-gemm/gen/2x4c8-xw-minmax-sse2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002560 "src/qs8-gemm/gen/3x4c8-minmax-sse2-ld64.c",
2561 "src/qs8-gemm/gen/3x4c8-minmax-sse2-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002562 "src/qs8-gemm/gen/3x4c8-xw-minmax-sse2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002563 "src/qs8-gemm/gen/4x4c2-minmax-sse2-ld64.c",
2564 "src/qs8-gemm/gen/4x4c2-minmax-sse2-ld128.c",
2565 "src/qs8-gemm/gen/4x4c2-xw-minmax-sse2.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002566 "src/qs8-igemm/gen/1x4c2-minmax-sse2-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002567 "src/qs8-igemm/gen/1x4c2-minmax-sse2-ld128.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002568 "src/qs8-igemm/gen/1x4c8-minmax-sse2-ld64.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002569 "src/qs8-igemm/gen/1x4c8-minmax-sse2-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002570 "src/qs8-igemm/gen/2x4c8-minmax-sse2-ld64.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002571 "src/qs8-igemm/gen/2x4c8-minmax-sse2-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002572 "src/qs8-igemm/gen/3x4c8-minmax-sse2-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002573 "src/qs8-igemm/gen/3x4c8-minmax-sse2-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002574 "src/qs8-igemm/gen/4x4c2-minmax-sse2-ld64.c",
2575 "src/qs8-igemm/gen/4x4c2-minmax-sse2-ld128.c",
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07002576 "src/qs8-requantization/fp32-sse2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002577 "src/qs8-requantization/precise-sse2.c",
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07002578 "src/qs8-requantization/q31-sse2.c",
Marat Dukhand9f3ad42020-08-10 12:30:58 -07002579 "src/qs8-vadd/gen/minmax-sse2-mul16-ld64-x8.c",
2580 "src/qs8-vadd/gen/minmax-sse2-mul16-ld64-x16.c",
2581 "src/qs8-vadd/gen/minmax-sse2-mul16-ld64-x24.c",
2582 "src/qs8-vadd/gen/minmax-sse2-mul16-ld64-x32.c",
Marat Dukhan0270d9f2020-08-11 00:56:46 -07002583 "src/qs8-vaddc/gen/minmax-sse2-mul16-ld64-x8.c",
2584 "src/qs8-vaddc/gen/minmax-sse2-mul16-ld64-x16.c",
2585 "src/qs8-vaddc/gen/minmax-sse2-mul16-ld64-x24.c",
2586 "src/qs8-vaddc/gen/minmax-sse2-mul16-ld64-x32.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -07002587 "src/qu8-avgpool/9p8x-minmax-sse2-c8.c",
2588 "src/qu8-avgpool/9x-minmax-sse2-c8.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -07002589 "src/qu8-dwconv/up8x9-minmax-sse2.c",
2590 "src/qu8-gavgpool/7p7x-minmax-sse2-c8.c",
2591 "src/qu8-gavgpool/7x-minmax-sse2-c8.c",
2592 "src/qu8-gemm/2x4c8-minmax-sse2.c",
2593 "src/qu8-gemm/4x4c2-minmax-sse2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002594 "src/qu8-igemm/4x4c2-minmax-sse2.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07002595 "src/qu8-requantization/fp32-sse2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002596 "src/qu8-requantization/precise-sse2.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07002597 "src/qu8-requantization/q31-sse2.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -07002598 "src/qu8-vadd/minmax-sse2.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -07002599 "src/u8-clamp/sse2-x64.c",
Marat Dukhan99936602020-04-11 16:47:01 -07002600 "src/u8-maxpool/9p8x-minmax-sse2-c16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002601 "src/u8-rmax/sse2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002602 "src/x8-zip/x2-sse2.c",
2603 "src/x8-zip/x3-sse2.c",
2604 "src/x8-zip/x4-sse2.c",
2605 "src/x8-zip/xm-sse2.c",
Marat Dukhan57dccd82020-04-14 00:53:10 -07002606 "src/x32-unpool/sse2.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002607 "src/x32-zip/x2-sse2.c",
2608 "src/x32-zip/x3-sse2.c",
2609 "src/x32-zip/x4-sse2.c",
2610 "src/x32-zip/xm-sse2.c",
Marat Dukhanfe7acb62020-03-09 19:30:05 -07002611]
2612
2613SSSE3_UKERNELS = [
Frank Barchard35db7d02020-10-26 13:37:34 -07002614 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-1x4-acc2.c",
2615 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-1x4-acc3.c",
2616 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-1x4-acc4.c",
Marat Dukhan98f2eeb2020-10-23 23:13:41 -07002617 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-1x4.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07002618 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-2x4-acc2.c",
Marat Dukhan98f2eeb2020-10-23 23:13:41 -07002619 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-2x4.c",
2620 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-3x4.c",
2621 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-4x4.c",
2622 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-5x4.c",
2623 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-6x4.c",
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07002624 "src/qs8-dwconv/gen/up8x9-minmax-ssse3-mul16.c",
2625 "src/qs8-dwconv/gen/up16x9-minmax-ssse3-mul16.c",
2626 "src/qs8-dwconv/gen/up24x9-minmax-ssse3-mul16.c",
Marat Dukhan159688f2020-08-06 10:34:29 -07002627 "src/qs8-gavgpool/gen/7p7x-minmax-ssse3-c8-acc2.c",
2628 "src/qs8-gavgpool/gen/7p7x-minmax-ssse3-c16-acc2.c",
2629 "src/qs8-gavgpool/gen/7p7x-minmax-ssse3-c24-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002630 "src/qs8-gavgpool/gen/7x-minmax-ssse3-c8-acc2.c",
2631 "src/qs8-gavgpool/gen/7x-minmax-ssse3-c16-acc2.c",
2632 "src/qs8-gavgpool/gen/7x-minmax-ssse3-c24-acc2.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002633 "src/qs8-gemm/gen/1x4c2-minmax-ssse3-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002634 "src/qs8-gemm/gen/1x4c2-minmax-ssse3-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002635 "src/qs8-gemm/gen/1x4c2-xw-minmax-ssse3.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002636 "src/qs8-gemm/gen/1x4c8-minmax-ssse3-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002637 "src/qs8-gemm/gen/1x4c8-minmax-ssse3-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002638 "src/qs8-gemm/gen/1x4c8-xw-minmax-ssse3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002639 "src/qs8-gemm/gen/2x4c8-minmax-ssse3-ld64.c",
2640 "src/qs8-gemm/gen/2x4c8-minmax-ssse3-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002641 "src/qs8-gemm/gen/2x4c8-xw-minmax-ssse3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002642 "src/qs8-gemm/gen/3x4c8-minmax-ssse3-ld64.c",
2643 "src/qs8-gemm/gen/3x4c8-minmax-ssse3-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002644 "src/qs8-gemm/gen/3x4c8-xw-minmax-ssse3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002645 "src/qs8-gemm/gen/4x4c2-minmax-ssse3-ld64.c",
2646 "src/qs8-gemm/gen/4x4c2-minmax-ssse3-ld128.c",
2647 "src/qs8-gemm/gen/4x4c2-xw-minmax-ssse3.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002648 "src/qs8-igemm/gen/1x4c2-minmax-ssse3-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002649 "src/qs8-igemm/gen/1x4c2-minmax-ssse3-ld128.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002650 "src/qs8-igemm/gen/1x4c8-minmax-ssse3-ld64.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002651 "src/qs8-igemm/gen/1x4c8-minmax-ssse3-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002652 "src/qs8-igemm/gen/2x4c8-minmax-ssse3-ld64.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002653 "src/qs8-igemm/gen/2x4c8-minmax-ssse3-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002654 "src/qs8-igemm/gen/3x4c8-minmax-ssse3-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002655 "src/qs8-igemm/gen/3x4c8-minmax-ssse3-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002656 "src/qs8-igemm/gen/4x4c2-minmax-ssse3-ld64.c",
2657 "src/qs8-igemm/gen/4x4c2-minmax-ssse3-ld128.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -07002658 "src/qs8-requantization/precise-ssse3.c",
2659 "src/qs8-requantization/q31-ssse3.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07002660 "src/qu8-requantization/precise-ssse3.c",
2661 "src/qu8-requantization/q31-ssse3.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002662]
2663
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08002664SSE41_UKERNELS = [
Marat Dukhan40a672f2019-11-25 03:08:22 -08002665 "src/f32-prelu/gen/sse41-2x4.c",
2666 "src/f32-prelu/gen/sse41-2x8.c",
Marat Dukhand243c1a2020-09-17 18:48:10 -07002667 "src/f32-sigmoid/gen/sse41-lut64-p2-div-x4.c",
2668 "src/f32-sigmoid/gen/sse41-lut64-p2-div-x8.c",
2669 "src/f32-sigmoid/gen/sse41-lut64-p2-div-x12.c",
2670 "src/f32-sigmoid/gen/sse41-lut64-p2-div-x16.c",
2671 "src/f32-sigmoid/gen/sse41-lut64-p2-div-x20.c",
2672 "src/f32-sigmoid/gen/sse41-lut64-p2-div-x24.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002673 "src/f32-sigmoid/gen/sse41-p5-div-x4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08002674 "src/f32-sigmoid/gen/sse41-p5-div-x8.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002675 "src/f32-sigmoid/gen/sse41-p5-div-x12.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08002676 "src/f32-sigmoid/gen/sse41-p5-div-x16.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002677 "src/f32-sigmoid/gen/sse41-p5-div-x20.c",
2678 "src/f32-sigmoid/gen/sse41-p5-div-x24.c",
Marat Dukhaned6baaf2020-12-01 15:07:08 -08002679 "src/f32-velu/gen/velu-sse41-rr2-lut16-p3-x4.c",
2680 "src/f32-velu/gen/velu-sse41-rr2-lut16-p3-x8.c",
2681 "src/f32-velu/gen/velu-sse41-rr2-lut16-p3-x12.c",
2682 "src/f32-velu/gen/velu-sse41-rr2-lut16-p3-x16.c",
2683 "src/f32-velu/gen/velu-sse41-rr2-lut16-p3-x20.c",
2684 "src/f32-velu/gen/velu-sse41-rr2-lut16-p3-x24.c",
2685 "src/f32-velu/gen/velu-sse41-rr2-p6-x4.c",
2686 "src/f32-velu/gen/velu-sse41-rr2-p6-x8.c",
2687 "src/f32-velu/gen/velu-sse41-rr2-p6-x12.c",
2688 "src/f32-velu/gen/velu-sse41-rr2-p6-x16.c",
2689 "src/f32-velu/gen/velu-sse41-rr2-p6-x20.c",
2690 "src/f32-velu/gen/velu-sse41-rr2-p6-x24.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07002691 "src/f32-vlrelu/gen/vlrelu-sse41-x4.c",
2692 "src/f32-vlrelu/gen/vlrelu-sse41-x8.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07002693 "src/f32-vrnd/gen/vrndd-sse41-x4.c",
2694 "src/f32-vrnd/gen/vrndd-sse41-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002695 "src/f32-vrnd/gen/vrndne-sse41-x4.c",
2696 "src/f32-vrnd/gen/vrndne-sse41-x8.c",
2697 "src/f32-vrnd/gen/vrndu-sse41-x4.c",
2698 "src/f32-vrnd/gen/vrndu-sse41-x8.c",
2699 "src/f32-vrnd/gen/vrndz-sse41-x4.c",
2700 "src/f32-vrnd/gen/vrndz-sse41-x8.c",
2701 "src/math/roundd-sse41.c",
2702 "src/math/roundne-sse41.c",
2703 "src/math/roundu-sse41.c",
2704 "src/math/roundz-sse41.c",
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07002705 "src/qs8-dwconv/gen/up8x9-minmax-sse41-mul16.c",
2706 "src/qs8-dwconv/gen/up16x9-minmax-sse41-mul16.c",
2707 "src/qs8-dwconv/gen/up24x9-minmax-sse41-mul16.c",
Marat Dukhan159688f2020-08-06 10:34:29 -07002708 "src/qs8-gavgpool/gen/7p7x-minmax-sse41-c8-acc2.c",
2709 "src/qs8-gavgpool/gen/7p7x-minmax-sse41-c16-acc2.c",
2710 "src/qs8-gavgpool/gen/7p7x-minmax-sse41-c24-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002711 "src/qs8-gavgpool/gen/7x-minmax-sse41-c8-acc2.c",
2712 "src/qs8-gavgpool/gen/7x-minmax-sse41-c16-acc2.c",
2713 "src/qs8-gavgpool/gen/7x-minmax-sse41-c24-acc2.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002714 "src/qs8-gemm/gen/1x4c2-minmax-sse41-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002715 "src/qs8-gemm/gen/1x4c2-minmax-sse41-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002716 "src/qs8-gemm/gen/1x4c2-xw-minmax-sse41.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002717 "src/qs8-gemm/gen/1x4c8-minmax-sse41-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002718 "src/qs8-gemm/gen/1x4c8-minmax-sse41-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002719 "src/qs8-gemm/gen/1x4c8-xw-minmax-sse41.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002720 "src/qs8-gemm/gen/2x4c8-minmax-sse41-ld64.c",
2721 "src/qs8-gemm/gen/2x4c8-minmax-sse41-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002722 "src/qs8-gemm/gen/2x4c8-xw-minmax-sse41.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002723 "src/qs8-gemm/gen/3x4c8-minmax-sse41-ld64.c",
2724 "src/qs8-gemm/gen/3x4c8-minmax-sse41-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002725 "src/qs8-gemm/gen/3x4c8-xw-minmax-sse41.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002726 "src/qs8-gemm/gen/4x4c2-minmax-sse41-ld64.c",
2727 "src/qs8-gemm/gen/4x4c2-minmax-sse41-ld128.c",
2728 "src/qs8-gemm/gen/4x4c2-xw-minmax-sse41.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002729 "src/qs8-igemm/gen/1x4c2-minmax-sse41-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002730 "src/qs8-igemm/gen/1x4c2-minmax-sse41-ld128.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002731 "src/qs8-igemm/gen/1x4c8-minmax-sse41-ld64.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002732 "src/qs8-igemm/gen/1x4c8-minmax-sse41-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002733 "src/qs8-igemm/gen/2x4c8-minmax-sse41-ld64.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002734 "src/qs8-igemm/gen/2x4c8-minmax-sse41-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002735 "src/qs8-igemm/gen/3x4c8-minmax-sse41-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002736 "src/qs8-igemm/gen/3x4c8-minmax-sse41-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002737 "src/qs8-igemm/gen/4x4c2-minmax-sse41-ld64.c",
2738 "src/qs8-igemm/gen/4x4c2-minmax-sse41-ld128.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -07002739 "src/qs8-requantization/fp32-sse4.c",
2740 "src/qs8-requantization/precise-sse4.c",
2741 "src/qs8-requantization/q31-sse4.c",
Marat Dukhand9f3ad42020-08-10 12:30:58 -07002742 "src/qs8-vadd/gen/minmax-sse41-mul16-ld64-x8.c",
2743 "src/qs8-vadd/gen/minmax-sse41-mul16-ld64-x16.c",
2744 "src/qs8-vadd/gen/minmax-sse41-mul16-ld64-x24.c",
2745 "src/qs8-vadd/gen/minmax-sse41-mul16-ld64-x32.c",
Marat Dukhanbb9225e2020-09-06 22:40:56 -07002746 "src/qs8-vadd/gen/minmax-sse41-mul32-ld32-x8.c",
2747 "src/qs8-vadd/gen/minmax-sse41-mul32-ld32-x16.c",
2748 "src/qs8-vadd/gen/minmax-sse41-mul32-ld32-x24.c",
2749 "src/qs8-vadd/gen/minmax-sse41-mul32-ld32-x32.c",
Marat Dukhan0270d9f2020-08-11 00:56:46 -07002750 "src/qs8-vaddc/gen/minmax-sse41-mul16-ld64-x8.c",
2751 "src/qs8-vaddc/gen/minmax-sse41-mul16-ld64-x16.c",
2752 "src/qs8-vaddc/gen/minmax-sse41-mul16-ld64-x24.c",
2753 "src/qs8-vaddc/gen/minmax-sse41-mul16-ld64-x32.c",
Marat Dukhanbb9225e2020-09-06 22:40:56 -07002754 "src/qs8-vaddc/gen/minmax-sse41-mul32-ld32-x8.c",
2755 "src/qs8-vaddc/gen/minmax-sse41-mul32-ld32-x16.c",
2756 "src/qs8-vaddc/gen/minmax-sse41-mul32-ld32-x24.c",
2757 "src/qs8-vaddc/gen/minmax-sse41-mul32-ld32-x32.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07002758 "src/qu8-requantization/precise-sse4.c",
2759 "src/qu8-requantization/q31-sse4.c",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08002760]
2761
Marat Dukhan08c4a432019-10-03 09:29:21 -07002762AVX_UKERNELS = [
Marat Dukhan5c5fa962020-03-10 18:38:33 -07002763 "src/f32-clamp/gen/avx-x8.c",
2764 "src/f32-clamp/gen/avx-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002765 "src/f32-dwconv/gen/up8x4-minmax-avx-acc2.c",
2766 "src/f32-dwconv/gen/up8x4-minmax-avx.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002767 "src/f32-dwconv/gen/up8x9-minmax-avx-acc2.c",
2768 "src/f32-dwconv/gen/up8x9-minmax-avx.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002769 "src/f32-dwconv/gen/up8x25-minmax-avx-acc2.c",
2770 "src/f32-dwconv/gen/up8x25-minmax-avx.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002771 "src/f32-dwconv/gen/up16x4-minmax-avx-acc2.c",
2772 "src/f32-dwconv/gen/up16x4-minmax-avx.c",
2773 "src/f32-dwconv/gen/up16x9-minmax-avx-acc2.c",
2774 "src/f32-dwconv/gen/up16x9-minmax-avx.c",
2775 "src/f32-dwconv/gen/up16x25-minmax-avx-acc2.c",
2776 "src/f32-dwconv/gen/up16x25-minmax-avx.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002777 "src/f32-gemm/gen-inc/1x8inc-minmax-avx-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002778 "src/f32-gemm/gen-inc/1x16inc-minmax-avx-broadcast.c",
2779 "src/f32-gemm/gen-inc/3x16inc-minmax-avx-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002780 "src/f32-gemm/gen-inc/4x8inc-minmax-avx-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002781 "src/f32-gemm/gen-inc/4x16inc-minmax-avx-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002782 "src/f32-gemm/gen-inc/5x8inc-minmax-avx-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002783 "src/f32-gemm/gen-inc/5x16inc-minmax-avx-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002784 "src/f32-gemm/gen-inc/6x8inc-minmax-avx-broadcast.c",
2785 "src/f32-gemm/gen-inc/7x8inc-minmax-avx-broadcast.c",
2786 "src/f32-gemm/gen/1x8-minmax-avx-broadcast.c",
2787 "src/f32-gemm/gen/1x16-minmax-avx-broadcast.c",
2788 "src/f32-gemm/gen/3x16-minmax-avx-broadcast.c",
2789 "src/f32-gemm/gen/4x8-minmax-avx-broadcast.c",
2790 "src/f32-gemm/gen/4x16-minmax-avx-broadcast.c",
2791 "src/f32-gemm/gen/5x8-minmax-avx-broadcast.c",
2792 "src/f32-gemm/gen/5x16-minmax-avx-broadcast.c",
2793 "src/f32-gemm/gen/6x8-minmax-avx-broadcast.c",
2794 "src/f32-gemm/gen/7x8-minmax-avx-broadcast.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07002795 "src/f32-hswish/gen/hswish-avx-x8.c",
2796 "src/f32-hswish/gen/hswish-avx-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002797 "src/f32-igemm/gen/1x8-minmax-avx-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002798 "src/f32-igemm/gen/1x16-minmax-avx-broadcast.c",
2799 "src/f32-igemm/gen/3x16-minmax-avx-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002800 "src/f32-igemm/gen/4x8-minmax-avx-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002801 "src/f32-igemm/gen/4x16-minmax-avx-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002802 "src/f32-igemm/gen/5x8-minmax-avx-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002803 "src/f32-igemm/gen/5x16-minmax-avx-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002804 "src/f32-igemm/gen/6x8-minmax-avx-broadcast.c",
2805 "src/f32-igemm/gen/7x8-minmax-avx-broadcast.c",
Marat Dukhan90eca0a2020-03-11 00:52:23 -07002806 "src/f32-prelu/gen/avx-2x8.c",
2807 "src/f32-prelu/gen/avx-2x16.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07002808 "src/f32-relu/gen/avx-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002809 "src/f32-relu/gen/avx-x16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002810 "src/f32-rmax/avx.c",
T.J. Alumbaughdc2b29c2020-10-14 13:56:08 -07002811 "src/f32-sigmoid/gen/avx-rr2-p5-div-x8.c",
2812 "src/f32-sigmoid/gen/avx-rr2-p5-div-x16.c",
2813 "src/f32-sigmoid/gen/avx-rr2-p5-div-x24.c",
2814 "src/f32-sigmoid/gen/avx-rr2-p5-div-x32.c",
2815 "src/f32-sigmoid/gen/avx-rr2-p5-div-x40.c",
2816 "src/f32-sigmoid/gen/avx-rr2-p5-div-x48.c",
2817 "src/f32-sigmoid/gen/avx-rr2-p5-div-x56.c",
2818 "src/f32-sigmoid/gen/avx-rr2-p5-div-x64.c",
2819 "src/f32-sigmoid/gen/avx-rr2-p5-div-x72.c",
2820 "src/f32-sigmoid/gen/avx-rr2-p5-div-x80.c",
2821 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x8.c",
2822 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x16.c",
2823 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x24.c",
2824 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x32.c",
2825 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x40.c",
2826 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x48.c",
2827 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x56.c",
2828 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x64.c",
2829 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x72.c",
2830 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x80.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002831 "src/f32-vbinary/gen/vadd-minmax-avx-x8.c",
2832 "src/f32-vbinary/gen/vadd-minmax-avx-x16.c",
2833 "src/f32-vbinary/gen/vaddc-minmax-avx-x8.c",
2834 "src/f32-vbinary/gen/vaddc-minmax-avx-x16.c",
2835 "src/f32-vbinary/gen/vdiv-minmax-avx-x8.c",
2836 "src/f32-vbinary/gen/vdiv-minmax-avx-x16.c",
2837 "src/f32-vbinary/gen/vdivc-minmax-avx-x8.c",
2838 "src/f32-vbinary/gen/vdivc-minmax-avx-x16.c",
Marat Dukhan9a88efe2019-12-10 15:54:24 -08002839 "src/f32-vbinary/gen/vmax-avx-x8.c",
2840 "src/f32-vbinary/gen/vmax-avx-x16.c",
2841 "src/f32-vbinary/gen/vmaxc-avx-x8.c",
2842 "src/f32-vbinary/gen/vmaxc-avx-x16.c",
2843 "src/f32-vbinary/gen/vmin-avx-x8.c",
2844 "src/f32-vbinary/gen/vmin-avx-x16.c",
2845 "src/f32-vbinary/gen/vminc-avx-x8.c",
2846 "src/f32-vbinary/gen/vminc-avx-x16.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002847 "src/f32-vbinary/gen/vmul-minmax-avx-x8.c",
2848 "src/f32-vbinary/gen/vmul-minmax-avx-x16.c",
2849 "src/f32-vbinary/gen/vmulc-minmax-avx-x8.c",
2850 "src/f32-vbinary/gen/vmulc-minmax-avx-x16.c",
2851 "src/f32-vbinary/gen/vrdivc-minmax-avx-x8.c",
2852 "src/f32-vbinary/gen/vrdivc-minmax-avx-x16.c",
2853 "src/f32-vbinary/gen/vrsubc-minmax-avx-x8.c",
2854 "src/f32-vbinary/gen/vrsubc-minmax-avx-x16.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07002855 "src/f32-vbinary/gen/vsqrdiff-avx-x8.c",
2856 "src/f32-vbinary/gen/vsqrdiff-avx-x16.c",
2857 "src/f32-vbinary/gen/vsqrdiffc-avx-x8.c",
2858 "src/f32-vbinary/gen/vsqrdiffc-avx-x16.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002859 "src/f32-vbinary/gen/vsub-minmax-avx-x8.c",
2860 "src/f32-vbinary/gen/vsub-minmax-avx-x16.c",
2861 "src/f32-vbinary/gen/vsubc-minmax-avx-x8.c",
2862 "src/f32-vbinary/gen/vsubc-minmax-avx-x16.c",
Marat Dukhaned6baaf2020-12-01 15:07:08 -08002863 "src/f32-velu/gen/velu-avx-rr2-lut4-p4-perm-x8.c",
2864 "src/f32-velu/gen/velu-avx-rr2-lut4-p4-perm-x16.c",
2865 "src/f32-velu/gen/velu-avx-rr2-lut4-p4-perm-x24.c",
2866 "src/f32-velu/gen/velu-avx-rr2-lut4-p4-perm-x32.c",
2867 "src/f32-velu/gen/velu-avx-rr2-lut4-p4-perm-x40.c",
2868 "src/f32-velu/gen/velu-avx-rr2-lut4-p4-perm-x48.c",
2869 "src/f32-velu/gen/velu-avx-rr2-lut16-p3-x8.c",
2870 "src/f32-velu/gen/velu-avx-rr2-lut16-p3-x16.c",
2871 "src/f32-velu/gen/velu-avx-rr2-lut16-p3-x24.c",
2872 "src/f32-velu/gen/velu-avx-rr2-lut16-p3-x32.c",
2873 "src/f32-velu/gen/velu-avx-rr2-lut16-p3-x40.c",
2874 "src/f32-velu/gen/velu-avx-rr2-lut16-p3-x48.c",
2875 "src/f32-velu/gen/velu-avx-rr2-p6-x8.c",
2876 "src/f32-velu/gen/velu-avx-rr2-p6-x16.c",
2877 "src/f32-velu/gen/velu-avx-rr2-p6-x24.c",
2878 "src/f32-velu/gen/velu-avx-rr2-p6-x32.c",
2879 "src/f32-velu/gen/velu-avx-rr2-p6-x40.c",
2880 "src/f32-velu/gen/velu-avx-rr2-p6-x48.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07002881 "src/f32-vlrelu/gen/vlrelu-avx-x8.c",
2882 "src/f32-vlrelu/gen/vlrelu-avx-x16.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07002883 "src/f32-vrnd/gen/vrndd-avx-x8.c",
2884 "src/f32-vrnd/gen/vrndd-avx-x16.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002885 "src/f32-vrnd/gen/vrndne-avx-x8.c",
2886 "src/f32-vrnd/gen/vrndne-avx-x16.c",
2887 "src/f32-vrnd/gen/vrndu-avx-x8.c",
2888 "src/f32-vrnd/gen/vrndu-avx-x16.c",
2889 "src/f32-vrnd/gen/vrndz-avx-x8.c",
2890 "src/f32-vrnd/gen/vrndz-avx-x16.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07002891 "src/f32-vscale/avx-x32.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07002892 "src/f32-vsqrt/gen/avx-sqrt-x8.c",
2893 "src/f32-vsqrt/gen/avx-sqrt-x16.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07002894 "src/f32-vunary/gen/vabs-avx-x8.c",
2895 "src/f32-vunary/gen/vabs-avx-x16.c",
2896 "src/f32-vunary/gen/vneg-avx-x8.c",
2897 "src/f32-vunary/gen/vneg-avx-x16.c",
2898 "src/f32-vunary/gen/vsqr-avx-x8.c",
2899 "src/f32-vunary/gen/vsqr-avx-x16.c",
Marat Dukhanb7633f22020-11-20 16:34:56 -08002900 "src/math/exp-avx-rr2-p5.c",
Marat Dukhande390d42020-11-29 19:32:18 -08002901 "src/math/expm1minus-avx-rr2-lut4-p4-perm.c",
2902 "src/math/expm1minus-avx-rr2-lut16-p3.c",
2903 "src/math/expm1minus-avx-rr2-p6.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002904 "src/math/sigmoid-avx-rr2-lut64-p2-div.c",
Marat Dukhan2e4727b2020-09-24 00:48:39 -07002905 "src/math/sigmoid-avx-rr2-p5-div.c",
Marat Dukhan39ab16f2020-10-16 14:30:35 -07002906 "src/math/sigmoid-avx-rr2-p5-nr1.c",
2907 "src/math/sigmoid-avx-rr2-p5-nr2.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002908]
2909
Marat Dukhan1566fee2020-08-02 21:55:41 -07002910XOP_UKERNELS = [
2911 "src/qs8-gemm/gen/1x4c2-minmax-xop-ld64.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002912 "src/qs8-gemm/gen/1x4c2-minmax-xop-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002913 "src/qs8-gemm/gen/1x4c2-xw-minmax-xop.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002914 "src/qs8-gemm/gen/1x4c8-minmax-xop-ld64.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002915 "src/qs8-gemm/gen/1x4c8-minmax-xop-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002916 "src/qs8-gemm/gen/1x4c8-xw-minmax-xop.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002917 "src/qs8-gemm/gen/2x4c8-minmax-xop-ld64.c",
2918 "src/qs8-gemm/gen/2x4c8-minmax-xop-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002919 "src/qs8-gemm/gen/2x4c8-xw-minmax-xop.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002920 "src/qs8-gemm/gen/3x4c8-minmax-xop-ld64.c",
2921 "src/qs8-gemm/gen/3x4c8-minmax-xop-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002922 "src/qs8-gemm/gen/3x4c8-xw-minmax-xop.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002923 "src/qs8-gemm/gen/4x4c2-minmax-xop-ld64.c",
2924 "src/qs8-gemm/gen/4x4c2-minmax-xop-ld128.c",
2925 "src/qs8-gemm/gen/4x4c2-xw-minmax-xop.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002926 "src/qs8-igemm/gen/1x4c2-minmax-xop-ld64.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002927 "src/qs8-igemm/gen/1x4c2-minmax-xop-ld128.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002928 "src/qs8-igemm/gen/1x4c8-minmax-xop-ld64.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002929 "src/qs8-igemm/gen/1x4c8-minmax-xop-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002930 "src/qs8-igemm/gen/2x4c8-minmax-xop-ld64.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002931 "src/qs8-igemm/gen/2x4c8-minmax-xop-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002932 "src/qs8-igemm/gen/3x4c8-minmax-xop-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002933 "src/qs8-igemm/gen/3x4c8-minmax-xop-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002934 "src/qs8-igemm/gen/4x4c2-minmax-xop-ld64.c",
2935 "src/qs8-igemm/gen/4x4c2-minmax-xop-ld128.c",
Marat Dukhanbb9225e2020-09-06 22:40:56 -07002936 "src/qs8-vadd/gen/minmax-xop-mul32-ld32-x8.c",
2937 "src/qs8-vadd/gen/minmax-xop-mul32-ld32-x16.c",
2938 "src/qs8-vadd/gen/minmax-xop-mul32-ld32-x24.c",
2939 "src/qs8-vadd/gen/minmax-xop-mul32-ld32-x32.c",
2940 "src/qs8-vaddc/gen/minmax-xop-mul32-ld32-x8.c",
2941 "src/qs8-vaddc/gen/minmax-xop-mul32-ld32-x16.c",
2942 "src/qs8-vaddc/gen/minmax-xop-mul32-ld32-x24.c",
2943 "src/qs8-vaddc/gen/minmax-xop-mul32-ld32-x32.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002944]
2945
Marat Dukhanfda12b82019-11-21 12:27:59 -08002946FMA3_UKERNELS = [
Marat Dukhan1c587112020-04-08 20:04:28 -07002947 "src/f32-dwconv/gen/up8x4-minmax-fma3-acc2.c",
2948 "src/f32-dwconv/gen/up8x4-minmax-fma3.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002949 "src/f32-dwconv/gen/up8x9-minmax-fma3-acc2.c",
2950 "src/f32-dwconv/gen/up8x9-minmax-fma3.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002951 "src/f32-dwconv/gen/up8x25-minmax-fma3-acc2.c",
2952 "src/f32-dwconv/gen/up8x25-minmax-fma3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002953 "src/f32-dwconv/gen/up16x4-minmax-fma3-acc2.c",
2954 "src/f32-dwconv/gen/up16x4-minmax-fma3.c",
2955 "src/f32-dwconv/gen/up16x9-minmax-fma3-acc2.c",
2956 "src/f32-dwconv/gen/up16x9-minmax-fma3.c",
2957 "src/f32-dwconv/gen/up16x25-minmax-fma3-acc2.c",
2958 "src/f32-dwconv/gen/up16x25-minmax-fma3.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002959 "src/f32-gemm/gen-inc/1x8inc-minmax-fma3-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002960 "src/f32-gemm/gen-inc/1x16inc-minmax-fma3-broadcast.c",
2961 "src/f32-gemm/gen-inc/1x16s4inc-minmax-fma3-broadcast.c",
2962 "src/f32-gemm/gen-inc/3x16inc-minmax-fma3-broadcast.c",
2963 "src/f32-gemm/gen-inc/3x16s4inc-minmax-fma3-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002964 "src/f32-gemm/gen-inc/4x8inc-minmax-fma3-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002965 "src/f32-gemm/gen-inc/4x16inc-minmax-fma3-broadcast.c",
2966 "src/f32-gemm/gen-inc/4x16s4inc-minmax-fma3-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002967 "src/f32-gemm/gen-inc/5x8inc-minmax-fma3-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002968 "src/f32-gemm/gen-inc/5x16inc-minmax-fma3-broadcast.c",
2969 "src/f32-gemm/gen-inc/5x16s4inc-minmax-fma3-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002970 "src/f32-gemm/gen-inc/6x8inc-minmax-fma3-broadcast.c",
2971 "src/f32-gemm/gen-inc/7x8inc-minmax-fma3-broadcast.c",
2972 "src/f32-gemm/gen-inc/8x8inc-minmax-fma3-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002973 "src/f32-gemm/gen/1x8-minmax-fma3-broadcast.c",
2974 "src/f32-gemm/gen/1x16-minmax-fma3-broadcast.c",
2975 "src/f32-gemm/gen/1x16s4-minmax-fma3-broadcast.c",
2976 "src/f32-gemm/gen/3x16-minmax-fma3-broadcast.c",
2977 "src/f32-gemm/gen/3x16s4-minmax-fma3-broadcast.c",
2978 "src/f32-gemm/gen/4x8-minmax-fma3-broadcast.c",
2979 "src/f32-gemm/gen/4x16-minmax-fma3-broadcast.c",
2980 "src/f32-gemm/gen/4x16s4-minmax-fma3-broadcast.c",
2981 "src/f32-gemm/gen/5x8-minmax-fma3-broadcast.c",
2982 "src/f32-gemm/gen/5x16-minmax-fma3-broadcast.c",
2983 "src/f32-gemm/gen/5x16s4-minmax-fma3-broadcast.c",
2984 "src/f32-gemm/gen/6x8-minmax-fma3-broadcast.c",
2985 "src/f32-gemm/gen/7x8-minmax-fma3-broadcast.c",
2986 "src/f32-gemm/gen/8x8-minmax-fma3-broadcast.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07002987 "src/f32-hswish/gen/hswish-fma3-x8.c",
2988 "src/f32-hswish/gen/hswish-fma3-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002989 "src/f32-igemm/gen/1x8-minmax-fma3-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002990 "src/f32-igemm/gen/1x16-minmax-fma3-broadcast.c",
2991 "src/f32-igemm/gen/1x16s4-minmax-fma3-broadcast.c",
2992 "src/f32-igemm/gen/3x16-minmax-fma3-broadcast.c",
2993 "src/f32-igemm/gen/3x16s4-minmax-fma3-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002994 "src/f32-igemm/gen/4x8-minmax-fma3-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002995 "src/f32-igemm/gen/4x16-minmax-fma3-broadcast.c",
2996 "src/f32-igemm/gen/4x16s4-minmax-fma3-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002997 "src/f32-igemm/gen/5x8-minmax-fma3-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002998 "src/f32-igemm/gen/5x16-minmax-fma3-broadcast.c",
2999 "src/f32-igemm/gen/5x16s4-minmax-fma3-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07003000 "src/f32-igemm/gen/6x8-minmax-fma3-broadcast.c",
3001 "src/f32-igemm/gen/7x8-minmax-fma3-broadcast.c",
3002 "src/f32-igemm/gen/8x8-minmax-fma3-broadcast.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07003003 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x8.c",
3004 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x16.c",
3005 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x24.c",
3006 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x32.c",
3007 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x40.c",
3008 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x48.c",
3009 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x56.c",
3010 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x64.c",
Marat Dukhan84000762020-06-29 18:38:43 -07003011 "src/math/sqrt-fma3-nr1fma.c",
Marat Dukhan84000762020-06-29 18:38:43 -07003012 "src/math/sqrt-fma3-nr1fma1adj.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003013 "src/math/sqrt-fma3-nr2fma.c",
Marat Dukhanfda12b82019-11-21 12:27:59 -08003014]
3015
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003016AVX2_UKERNELS = [
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003017 "src/f32-raddexpminusmax/gen/avx2-p5-x64-acc2.c",
3018 "src/f32-raddexpminusmax/gen/avx2-p5-x64-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003019 "src/f32-raddexpminusmax/gen/avx2-p5-x64.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003020 "src/f32-raddexpminusmax/gen/avx2-p5-x72-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003021 "src/f32-raddexpminusmax/gen/avx2-p5-x72.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003022 "src/f32-raddexpminusmax/gen/avx2-p5-x80-acc2.c",
3023 "src/f32-raddexpminusmax/gen/avx2-p5-x80-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003024 "src/f32-raddexpminusmax/gen/avx2-p5-x80.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003025 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc2.c",
3026 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc3.c",
3027 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc6.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003028 "src/f32-raddexpminusmax/gen/avx2-p5-x96.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003029 "src/f32-raddextexp/gen/avx2-p5-x64-acc2.c",
3030 "src/f32-raddextexp/gen/avx2-p5-x64-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003031 "src/f32-raddextexp/gen/avx2-p5-x64.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003032 "src/f32-raddextexp/gen/avx2-p5-x72-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003033 "src/f32-raddextexp/gen/avx2-p5-x72.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003034 "src/f32-raddextexp/gen/avx2-p5-x80-acc2.c",
3035 "src/f32-raddextexp/gen/avx2-p5-x80-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003036 "src/f32-raddextexp/gen/avx2-p5-x80.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003037 "src/f32-raddextexp/gen/avx2-p5-x96-acc2.c",
3038 "src/f32-raddextexp/gen/avx2-p5-x96-acc3.c",
3039 "src/f32-raddextexp/gen/avx2-p5-x96-acc6.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003040 "src/f32-raddextexp/gen/avx2-p5-x96.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003041 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64-acc2.c",
3042 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003043 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003044 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x72-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003045 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x72.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003046 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80-acc2.c",
3047 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003048 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003049 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc2.c",
3050 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc3.c",
3051 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc6.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003052 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96.c",
Marat Dukhanfa0a4322020-01-06 16:14:29 -08003053 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x8.c",
3054 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x16.c",
3055 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x24.c",
3056 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x32.c",
3057 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x40.c",
3058 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x48.c",
3059 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x56.c",
3060 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x64.c",
3061 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x72.c",
3062 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x80.c",
3063 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x8.c",
3064 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x16.c",
3065 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x24.c",
3066 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x32.c",
3067 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x40.c",
3068 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x48.c",
3069 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x56.c",
3070 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x64.c",
3071 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x72.c",
3072 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x80.c",
3073 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x8.c",
3074 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x16.c",
3075 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x24.c",
3076 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x32.c",
3077 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x40.c",
3078 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x48.c",
3079 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x56.c",
3080 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x64.c",
3081 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x72.c",
3082 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x80.c",
Marat Dukhaned6baaf2020-12-01 15:07:08 -08003083 "src/f32-velu/gen/velu-avx2-rr1-lut4-p4-perm-x8.c",
3084 "src/f32-velu/gen/velu-avx2-rr1-lut4-p4-perm-x16.c",
3085 "src/f32-velu/gen/velu-avx2-rr1-lut4-p4-perm-x24.c",
3086 "src/f32-velu/gen/velu-avx2-rr1-lut4-p4-perm-x32.c",
3087 "src/f32-velu/gen/velu-avx2-rr1-lut4-p4-perm-x40.c",
3088 "src/f32-velu/gen/velu-avx2-rr1-lut4-p4-perm-x48.c",
3089 "src/f32-velu/gen/velu-avx2-rr1-lut4-p4-perm-x56.c",
3090 "src/f32-velu/gen/velu-avx2-rr1-lut4-p4-perm-x64.c",
3091 "src/f32-velu/gen/velu-avx2-rr1-lut4-p4-perm-x72.c",
3092 "src/f32-velu/gen/velu-avx2-rr1-lut4-p4-perm-x80.c",
3093 "src/f32-velu/gen/velu-avx2-rr1-lut8-p4-perm-x8.c",
3094 "src/f32-velu/gen/velu-avx2-rr1-lut8-p4-perm-x16.c",
3095 "src/f32-velu/gen/velu-avx2-rr1-lut8-p4-perm-x24.c",
3096 "src/f32-velu/gen/velu-avx2-rr1-lut8-p4-perm-x32.c",
3097 "src/f32-velu/gen/velu-avx2-rr1-lut8-p4-perm-x40.c",
3098 "src/f32-velu/gen/velu-avx2-rr1-lut8-p4-perm-x48.c",
3099 "src/f32-velu/gen/velu-avx2-rr1-lut8-p4-perm-x56.c",
3100 "src/f32-velu/gen/velu-avx2-rr1-lut8-p4-perm-x64.c",
3101 "src/f32-velu/gen/velu-avx2-rr1-lut8-p4-perm-x72.c",
3102 "src/f32-velu/gen/velu-avx2-rr1-lut8-p4-perm-x80.c",
3103 "src/f32-velu/gen/velu-avx2-rr1-lut16-p3-gather-x8.c",
3104 "src/f32-velu/gen/velu-avx2-rr1-lut16-p3-gather-x16.c",
3105 "src/f32-velu/gen/velu-avx2-rr1-lut16-p3-gather-x24.c",
3106 "src/f32-velu/gen/velu-avx2-rr1-lut16-p3-gather-x32.c",
3107 "src/f32-velu/gen/velu-avx2-rr1-lut16-p3-gather-x40.c",
3108 "src/f32-velu/gen/velu-avx2-rr1-lut16-p3-gather-x48.c",
3109 "src/f32-velu/gen/velu-avx2-rr1-lut16-p3-gather-x56.c",
3110 "src/f32-velu/gen/velu-avx2-rr1-lut16-p3-gather-x64.c",
3111 "src/f32-velu/gen/velu-avx2-rr1-lut16-p3-gather-x72.c",
3112 "src/f32-velu/gen/velu-avx2-rr1-lut16-p3-gather-x80.c",
3113 "src/f32-velu/gen/velu-avx2-rr1-p6-x8.c",
3114 "src/f32-velu/gen/velu-avx2-rr1-p6-x16.c",
3115 "src/f32-velu/gen/velu-avx2-rr1-p6-x24.c",
3116 "src/f32-velu/gen/velu-avx2-rr1-p6-x32.c",
3117 "src/f32-velu/gen/velu-avx2-rr1-p6-x40.c",
3118 "src/f32-velu/gen/velu-avx2-rr1-p6-x48.c",
3119 "src/f32-velu/gen/velu-avx2-rr1-p6-x56.c",
3120 "src/f32-velu/gen/velu-avx2-rr1-p6-x64.c",
3121 "src/f32-velu/gen/velu-avx2-rr1-p6-x72.c",
3122 "src/f32-velu/gen/velu-avx2-rr1-p6-x80.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003123 "src/f32-vscaleexpminusmax/gen/avx2-p5-x8.c",
3124 "src/f32-vscaleexpminusmax/gen/avx2-p5-x16.c",
3125 "src/f32-vscaleexpminusmax/gen/avx2-p5-x24.c",
3126 "src/f32-vscaleexpminusmax/gen/avx2-p5-x32.c",
3127 "src/f32-vscaleexpminusmax/gen/avx2-p5-x40.c",
3128 "src/f32-vscaleexpminusmax/gen/avx2-p5-x48.c",
3129 "src/f32-vscaleexpminusmax/gen/avx2-p5-x56.c",
3130 "src/f32-vscaleexpminusmax/gen/avx2-p5-x64.c",
3131 "src/f32-vscaleexpminusmax/gen/avx2-p5-x72.c",
3132 "src/f32-vscaleexpminusmax/gen/avx2-p5-x80.c",
3133 "src/f32-vscaleexpminusmax/gen/avx2-p5-x88.c",
3134 "src/f32-vscaleexpminusmax/gen/avx2-p5-x96.c",
3135 "src/f32-vscaleextexp/gen/avx2-p5-x8.c",
3136 "src/f32-vscaleextexp/gen/avx2-p5-x16.c",
3137 "src/f32-vscaleextexp/gen/avx2-p5-x24.c",
3138 "src/f32-vscaleextexp/gen/avx2-p5-x32.c",
3139 "src/f32-vscaleextexp/gen/avx2-p5-x40.c",
3140 "src/f32-vscaleextexp/gen/avx2-p5-x48.c",
3141 "src/f32-vscaleextexp/gen/avx2-p5-x56.c",
3142 "src/f32-vscaleextexp/gen/avx2-p5-x64.c",
3143 "src/f32-vscaleextexp/gen/avx2-p5-x72.c",
3144 "src/f32-vscaleextexp/gen/avx2-p5-x80.c",
3145 "src/f32-vscaleextexp/gen/avx2-p5-x88.c",
3146 "src/f32-vscaleextexp/gen/avx2-p5-x96.c",
Marat Dukhanb7633f22020-11-20 16:34:56 -08003147 "src/math/exp-avx2-rr2-lut8-p3-perm.c",
3148 "src/math/exp-avx2-rr2-lut8-p4-perm.c",
3149 "src/math/exp-avx2-rr2-p5.c",
Marat Dukhande390d42020-11-29 19:32:18 -08003150 "src/math/expm1minus-avx2-rr1-lut4-p4-perm.c",
3151 "src/math/expm1minus-avx2-rr1-lut8-p4-perm.c",
3152 "src/math/expm1minus-avx2-rr1-lut16-p3-gather.c",
3153 "src/math/expm1minus-avx2-rr1-p6.c",
Frank Barcharde7223ee2020-12-04 19:04:01 -08003154 "src/math/expminus-avx2-rr2-p5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003155 "src/math/extexp-avx2-p5.c",
3156 "src/math/sigmoid-avx2-rr1-lut64-p2-gather-div.c",
3157 "src/math/sigmoid-avx2-rr1-lut64-p2-gather-nr1fma.c",
3158 "src/math/sigmoid-avx2-rr1-lut64-p2-gather-nr2fma.c",
3159 "src/math/sigmoid-avx2-rr1-lut64-p2-gather-nr2fma1adj.c",
3160 "src/math/sigmoid-avx2-rr1-p5-div.c",
3161 "src/math/sigmoid-avx2-rr1-p5-nr1fma.c",
3162 "src/math/sigmoid-avx2-rr1-p5-nr2fma.c",
3163 "src/math/sigmoid-avx2-rr2-lut64-p2-gather-div.c",
3164 "src/math/sigmoid-avx2-rr2-lut64-p2-gather-nr1fma.c",
3165 "src/math/sigmoid-avx2-rr2-lut64-p2-gather-nr2fma.c",
3166 "src/math/sigmoid-avx2-rr2-lut64-p2-gather-nr2fma1adj.c",
3167 "src/math/sigmoid-avx2-rr2-p5-div.c",
3168 "src/math/sigmoid-avx2-rr2-p5-nr1fma.c",
3169 "src/math/sigmoid-avx2-rr2-p5-nr2fma.c",
Marat Dukhand65a1522020-08-04 19:28:18 -07003170 "src/qs8-dwconv/gen/up8x9-minmax-avx2-mul32.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003171 "src/qs8-dwconv/gen/up16x9-minmax-avx2-mul16.c",
Marat Dukhand65a1522020-08-04 19:28:18 -07003172 "src/qs8-dwconv/gen/up16x9-minmax-avx2-mul32.c",
3173 "src/qs8-dwconv/gen/up24x9-minmax-avx2-mul32.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003174 "src/qs8-dwconv/gen/up32x9-minmax-avx2-mul16.c",
Marat Dukhand65a1522020-08-04 19:28:18 -07003175 "src/qs8-dwconv/gen/up32x9-minmax-avx2-mul32.c",
Marat Dukhan12809522020-08-02 22:23:51 -07003176 "src/qs8-gemm/gen/1x8c8-minmax-avx2.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07003177 "src/qs8-gemm/gen/1x8c8-xw-minmax-avx2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003178 "src/qs8-gemm/gen/2x8c8-minmax-avx2.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07003179 "src/qs8-gemm/gen/2x8c8-xw-minmax-avx2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003180 "src/qs8-gemm/gen/3x8c8-minmax-avx2.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07003181 "src/qs8-gemm/gen/3x8c8-xw-minmax-avx2.c",
Marat Dukhan12809522020-08-02 22:23:51 -07003182 "src/qs8-igemm/gen/1x8c8-minmax-avx2.c",
3183 "src/qs8-igemm/gen/2x8c8-minmax-avx2.c",
3184 "src/qs8-igemm/gen/3x8c8-minmax-avx2.c",
Marat Dukhane6dc0b62020-09-08 23:57:14 -07003185 "src/qs8-vadd/gen/minmax-avx2-mul32-ld64-x8.c",
3186 "src/qs8-vadd/gen/minmax-avx2-mul32-ld64-x16.c",
3187 "src/qs8-vadd/gen/minmax-avx2-mul32-ld64-x24.c",
3188 "src/qs8-vadd/gen/minmax-avx2-mul32-ld64-x32.c",
3189 "src/qs8-vaddc/gen/minmax-avx2-mul32-ld64-x8.c",
3190 "src/qs8-vaddc/gen/minmax-avx2-mul32-ld64-x16.c",
3191 "src/qs8-vaddc/gen/minmax-avx2-mul32-ld64-x24.c",
3192 "src/qs8-vaddc/gen/minmax-avx2-mul32-ld64-x32.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003193]
3194
Marat Dukhan08c4a432019-10-03 09:29:21 -07003195AVX512F_UKERNELS = [
Marat Dukhan5c5fa962020-03-10 18:38:33 -07003196 "src/f32-clamp/gen/avx512f-x16.c",
3197 "src/f32-clamp/gen/avx512f-x32.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07003198 "src/f32-dwconv/gen/up16x4-minmax-avx512f-acc2.c",
3199 "src/f32-dwconv/gen/up16x4-minmax-avx512f.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07003200 "src/f32-dwconv/gen/up16x9-minmax-avx512f-acc2.c",
3201 "src/f32-dwconv/gen/up16x9-minmax-avx512f.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07003202 "src/f32-dwconv/gen/up16x25-minmax-avx512f-acc2.c",
3203 "src/f32-dwconv/gen/up16x25-minmax-avx512f.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003204 "src/f32-dwconv/gen/up32x4-minmax-avx512f-acc2.c",
3205 "src/f32-dwconv/gen/up32x4-minmax-avx512f.c",
3206 "src/f32-dwconv/gen/up32x9-minmax-avx512f-acc2.c",
3207 "src/f32-dwconv/gen/up32x9-minmax-avx512f.c",
3208 "src/f32-dwconv/gen/up32x25-minmax-avx512f-acc2.c",
3209 "src/f32-dwconv/gen/up32x25-minmax-avx512f.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07003210 "src/f32-gemm/gen-inc/1x16inc-minmax-avx512f-broadcast.c",
3211 "src/f32-gemm/gen-inc/4x16inc-minmax-avx512f-broadcast.c",
3212 "src/f32-gemm/gen-inc/5x16inc-minmax-avx512f-broadcast.c",
3213 "src/f32-gemm/gen-inc/6x16inc-minmax-avx512f-broadcast.c",
3214 "src/f32-gemm/gen-inc/7x16inc-minmax-avx512f-broadcast.c",
3215 "src/f32-gemm/gen-inc/8x16inc-minmax-avx512f-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003216 "src/f32-gemm/gen/1x16-minmax-avx512f-broadcast.c",
3217 "src/f32-gemm/gen/4x16-minmax-avx512f-broadcast.c",
3218 "src/f32-gemm/gen/5x16-minmax-avx512f-broadcast.c",
3219 "src/f32-gemm/gen/6x16-minmax-avx512f-broadcast.c",
3220 "src/f32-gemm/gen/7x16-minmax-avx512f-broadcast.c",
3221 "src/f32-gemm/gen/8x16-minmax-avx512f-broadcast.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07003222 "src/f32-hswish/gen/hswish-avx512f-x16.c",
3223 "src/f32-hswish/gen/hswish-avx512f-x32.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07003224 "src/f32-igemm/gen/1x16-minmax-avx512f-broadcast.c",
3225 "src/f32-igemm/gen/4x16-minmax-avx512f-broadcast.c",
3226 "src/f32-igemm/gen/5x16-minmax-avx512f-broadcast.c",
3227 "src/f32-igemm/gen/6x16-minmax-avx512f-broadcast.c",
3228 "src/f32-igemm/gen/7x16-minmax-avx512f-broadcast.c",
3229 "src/f32-igemm/gen/8x16-minmax-avx512f-broadcast.c",
Marat Dukhan90eca0a2020-03-11 00:52:23 -07003230 "src/f32-prelu/gen/avx512f-2x16.c",
3231 "src/f32-prelu/gen/avx512f-2x32.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003232 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128-acc2.c",
3233 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003234 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003235 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x144-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003236 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x144.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003237 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160-acc2.c",
3238 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003239 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003240 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc2.c",
3241 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc3.c",
3242 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc6.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003243 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003244 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128-acc2.c",
3245 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003246 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003247 "src/f32-raddextexp/gen/avx512f-p5-scalef-x144-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003248 "src/f32-raddextexp/gen/avx512f-p5-scalef-x144.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003249 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160-acc2.c",
3250 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003251 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003252 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc2.c",
3253 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc3.c",
3254 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc6.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003255 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003256 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128-acc2.c",
3257 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003258 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003259 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x144-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003260 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x144.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003261 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160-acc2.c",
3262 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003263 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003264 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc2.c",
3265 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc3.c",
3266 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc6.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003267 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07003268 "src/f32-relu/gen/avx512f-x16.c",
3269 "src/f32-relu/gen/avx512f-x32.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003270 "src/f32-rmax/avx512f.c",
Marat Dukhand9ca7e62020-09-23 23:45:29 -07003271 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x16.c",
3272 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x32.c",
3273 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x48.c",
3274 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x64.c",
3275 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x80.c",
3276 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x96.c",
3277 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x112.c",
3278 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x128.c",
3279 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x16.c",
3280 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x32.c",
3281 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x48.c",
3282 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x64.c",
3283 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x80.c",
3284 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x96.c",
3285 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x112.c",
3286 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003287 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x16.c",
3288 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x32.c",
3289 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x48.c",
3290 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x64.c",
3291 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x80.c",
3292 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x96.c",
3293 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x112.c",
3294 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x128.c",
3295 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x16.c",
3296 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x32.c",
3297 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x48.c",
3298 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x64.c",
3299 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x80.c",
3300 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x96.c",
3301 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x112.c",
3302 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x128.c",
Marat Dukhand9ca7e62020-09-23 23:45:29 -07003303 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x16.c",
3304 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x32.c",
3305 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x48.c",
3306 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x64.c",
3307 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x80.c",
3308 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x96.c",
3309 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x112.c",
3310 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x128.c",
3311 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x16.c",
3312 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x32.c",
3313 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x48.c",
3314 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x64.c",
3315 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x80.c",
3316 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x96.c",
3317 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x112.c",
3318 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x128.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07003319 "src/f32-vbinary/gen/vadd-minmax-avx512f-x16.c",
3320 "src/f32-vbinary/gen/vadd-minmax-avx512f-x32.c",
3321 "src/f32-vbinary/gen/vaddc-minmax-avx512f-x16.c",
3322 "src/f32-vbinary/gen/vaddc-minmax-avx512f-x32.c",
3323 "src/f32-vbinary/gen/vdiv-minmax-avx512f-x16.c",
3324 "src/f32-vbinary/gen/vdiv-minmax-avx512f-x32.c",
3325 "src/f32-vbinary/gen/vdivc-minmax-avx512f-x16.c",
3326 "src/f32-vbinary/gen/vdivc-minmax-avx512f-x32.c",
Marat Dukhan9a88efe2019-12-10 15:54:24 -08003327 "src/f32-vbinary/gen/vmax-avx512f-x16.c",
3328 "src/f32-vbinary/gen/vmax-avx512f-x32.c",
3329 "src/f32-vbinary/gen/vmaxc-avx512f-x16.c",
3330 "src/f32-vbinary/gen/vmaxc-avx512f-x32.c",
3331 "src/f32-vbinary/gen/vmin-avx512f-x16.c",
3332 "src/f32-vbinary/gen/vmin-avx512f-x32.c",
3333 "src/f32-vbinary/gen/vminc-avx512f-x16.c",
3334 "src/f32-vbinary/gen/vminc-avx512f-x32.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07003335 "src/f32-vbinary/gen/vmul-minmax-avx512f-x16.c",
3336 "src/f32-vbinary/gen/vmul-minmax-avx512f-x32.c",
3337 "src/f32-vbinary/gen/vmulc-minmax-avx512f-x16.c",
3338 "src/f32-vbinary/gen/vmulc-minmax-avx512f-x32.c",
3339 "src/f32-vbinary/gen/vrdivc-minmax-avx512f-x16.c",
3340 "src/f32-vbinary/gen/vrdivc-minmax-avx512f-x32.c",
3341 "src/f32-vbinary/gen/vrsubc-minmax-avx512f-x16.c",
3342 "src/f32-vbinary/gen/vrsubc-minmax-avx512f-x32.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07003343 "src/f32-vbinary/gen/vsqrdiff-avx512f-x16.c",
3344 "src/f32-vbinary/gen/vsqrdiff-avx512f-x32.c",
3345 "src/f32-vbinary/gen/vsqrdiffc-avx512f-x16.c",
3346 "src/f32-vbinary/gen/vsqrdiffc-avx512f-x32.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07003347 "src/f32-vbinary/gen/vsub-minmax-avx512f-x16.c",
3348 "src/f32-vbinary/gen/vsub-minmax-avx512f-x32.c",
3349 "src/f32-vbinary/gen/vsubc-minmax-avx512f-x16.c",
3350 "src/f32-vbinary/gen/vsubc-minmax-avx512f-x32.c",
Marat Dukhaned6baaf2020-12-01 15:07:08 -08003351 "src/f32-velu/gen/velu-avx512f-rr1-lut16-p3-perm-x16.c",
3352 "src/f32-velu/gen/velu-avx512f-rr1-lut16-p3-perm-x32.c",
3353 "src/f32-velu/gen/velu-avx512f-rr1-lut16-p3-perm-x48.c",
3354 "src/f32-velu/gen/velu-avx512f-rr1-lut16-p3-perm-x64.c",
3355 "src/f32-velu/gen/velu-avx512f-rr1-lut16-p3-perm-x80.c",
3356 "src/f32-velu/gen/velu-avx512f-rr1-lut16-p3-perm-x96.c",
3357 "src/f32-velu/gen/velu-avx512f-rr1-lut16-p3-perm-x112.c",
3358 "src/f32-velu/gen/velu-avx512f-rr1-lut16-p3-perm-x128.c",
3359 "src/f32-velu/gen/velu-avx512f-rr1-p6-x16.c",
3360 "src/f32-velu/gen/velu-avx512f-rr1-p6-x32.c",
3361 "src/f32-velu/gen/velu-avx512f-rr1-p6-x48.c",
3362 "src/f32-velu/gen/velu-avx512f-rr1-p6-x64.c",
3363 "src/f32-velu/gen/velu-avx512f-rr1-p6-x80.c",
3364 "src/f32-velu/gen/velu-avx512f-rr1-p6-x96.c",
3365 "src/f32-velu/gen/velu-avx512f-rr1-p6-x112.c",
3366 "src/f32-velu/gen/velu-avx512f-rr1-p6-x128.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07003367 "src/f32-vlrelu/gen/vlrelu-avx512f-x16.c",
3368 "src/f32-vlrelu/gen/vlrelu-avx512f-x32.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003369 "src/f32-vrnd/gen/vrndd-avx512f-x16.c",
3370 "src/f32-vrnd/gen/vrndd-avx512f-x32.c",
3371 "src/f32-vrnd/gen/vrndne-avx512f-x16.c",
3372 "src/f32-vrnd/gen/vrndne-avx512f-x32.c",
3373 "src/f32-vrnd/gen/vrndu-avx512f-x16.c",
3374 "src/f32-vrnd/gen/vrndu-avx512f-x32.c",
3375 "src/f32-vrnd/gen/vrndz-avx512f-x16.c",
3376 "src/f32-vrnd/gen/vrndz-avx512f-x32.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07003377 "src/f32-vscale/avx512f-x64.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003378 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x16.c",
3379 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x32.c",
3380 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x48.c",
3381 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x64.c",
3382 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x80.c",
3383 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x96.c",
3384 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x112.c",
3385 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x128.c",
3386 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x144.c",
3387 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x160.c",
3388 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x176.c",
3389 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x192.c",
3390 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x16.c",
3391 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x32.c",
3392 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x48.c",
3393 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x64.c",
3394 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x80.c",
3395 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x96.c",
3396 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x112.c",
3397 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x128.c",
3398 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x144.c",
3399 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x160.c",
3400 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x176.c",
3401 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x192.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07003402 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x16.c",
3403 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x32.c",
3404 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x48.c",
3405 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x64.c",
3406 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x80.c",
3407 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x96.c",
3408 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x112.c",
3409 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x128.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07003410 "src/f32-vunary/gen/vabs-avx512f-x16.c",
3411 "src/f32-vunary/gen/vabs-avx512f-x32.c",
3412 "src/f32-vunary/gen/vneg-avx512f-x16.c",
3413 "src/f32-vunary/gen/vneg-avx512f-x32.c",
3414 "src/f32-vunary/gen/vsqr-avx512f-x16.c",
3415 "src/f32-vunary/gen/vsqr-avx512f-x32.c",
Marat Dukhanb7633f22020-11-20 16:34:56 -08003416 "src/math/exp-avx512f-rr2-lut16-p3-perm-scalef.c",
3417 "src/math/exp-avx512f-rr2-lut16-p3-perm.c",
3418 "src/math/exp-avx512f-rr2-lut32-p2-perm2-scalef.c",
3419 "src/math/exp-avx512f-rr2-lut32-p2-perm2.c",
3420 "src/math/exp-avx512f-rr2-p5-scalef.c",
3421 "src/math/exp-avx512f-rr2-p5.c",
Marat Dukhande390d42020-11-29 19:32:18 -08003422 "src/math/expm1minus-avx512f-rr1-lut16-p3-perm.c",
3423 "src/math/expm1minus-avx512f-rr1-p6.c",
Marat Dukhan98ba4412019-10-23 02:14:28 -07003424 "src/math/extexp-avx512f-p5.c",
Marat Dukhan6a34c5f2020-09-22 21:44:15 -07003425 "src/math/sigmoid-avx512f-rr1-lut16-p3-perm-scalef-div.c",
Marat Dukhan6a34c5f2020-09-22 21:44:15 -07003426 "src/math/sigmoid-avx512f-rr1-lut16-p3-perm-scalef-nr1fma.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003427 "src/math/sigmoid-avx512f-rr1-lut16-p3-perm-scalef-nr1fma1adj.c",
Marat Dukhan6a34c5f2020-09-22 21:44:15 -07003428 "src/math/sigmoid-avx512f-rr1-lut32-p2-perm2-scalef-div.c",
Marat Dukhan6a34c5f2020-09-22 21:44:15 -07003429 "src/math/sigmoid-avx512f-rr1-lut32-p2-perm2-scalef-nr1fma.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003430 "src/math/sigmoid-avx512f-rr1-lut32-p2-perm2-scalef-nr1fma1adj.c",
Marat Dukhan36173d22020-10-15 17:14:26 -07003431 "src/math/sigmoid-avx512f-rr1-lut64-p2-gather-scalef-div.c",
Marat Dukhan36173d22020-10-15 17:14:26 -07003432 "src/math/sigmoid-avx512f-rr1-lut64-p2-gather-scalef-nr1fma.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003433 "src/math/sigmoid-avx512f-rr1-lut64-p2-gather-scalef-nr1fma1adj.c",
3434 "src/math/sigmoid-avx512f-rr1-p5-scalef-div.c",
3435 "src/math/sigmoid-avx512f-rr1-p5-scalef-nr1fma.c",
3436 "src/math/sigmoid-avx512f-rr1-p5-scalef-nr1fma1adj.c",
3437 "src/math/sigmoid-avx512f-rr2-lut16-p3-perm-scalef-div.c",
3438 "src/math/sigmoid-avx512f-rr2-lut16-p3-perm-scalef-nr1fma.c",
3439 "src/math/sigmoid-avx512f-rr2-lut16-p3-perm-scalef-nr1fma1adj.c",
3440 "src/math/sigmoid-avx512f-rr2-lut32-p2-perm2-scalef-div.c",
3441 "src/math/sigmoid-avx512f-rr2-lut32-p2-perm2-scalef-nr1fma.c",
3442 "src/math/sigmoid-avx512f-rr2-lut32-p2-perm2-scalef-nr1fma1adj.c",
Marat Dukhan36173d22020-10-15 17:14:26 -07003443 "src/math/sigmoid-avx512f-rr2-lut64-p2-gather-scalef-div.c",
Marat Dukhan36173d22020-10-15 17:14:26 -07003444 "src/math/sigmoid-avx512f-rr2-lut64-p2-gather-scalef-nr1fma.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003445 "src/math/sigmoid-avx512f-rr2-lut64-p2-gather-scalef-nr1fma1adj.c",
3446 "src/math/sigmoid-avx512f-rr2-p5-scalef-div.c",
3447 "src/math/sigmoid-avx512f-rr2-p5-scalef-nr1fma.c",
3448 "src/math/sigmoid-avx512f-rr2-p5-scalef-nr1fma1adj.c",
Marat Dukhan84000762020-06-29 18:38:43 -07003449 "src/math/sqrt-avx512f-nr1fma.c",
Marat Dukhan84000762020-06-29 18:38:43 -07003450 "src/math/sqrt-avx512f-nr1fma1adj.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003451 "src/math/sqrt-avx512f-nr2fma.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003452]
3453
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07003454AVX512SKX_UKERNELS = [
Marat Dukhan2ffc5e62020-09-06 22:33:38 -07003455 "src/qs8-dwconv/gen/up16x9-minmax-avx512skx-mul32.c",
3456 "src/qs8-dwconv/gen/up32x9-minmax-avx512skx-mul32.c",
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07003457 "src/qs8-gemm/gen/1x16c8-minmax-avx512skx.c",
3458 "src/qs8-gemm/gen/2x16c8-minmax-avx512skx.c",
3459 "src/qs8-gemm/gen/3x16c8-minmax-avx512skx.c",
3460 "src/qs8-gemm/gen/4x16c8-minmax-avx512skx.c",
3461 "src/qs8-igemm/gen/1x16c8-minmax-avx512skx.c",
3462 "src/qs8-igemm/gen/2x16c8-minmax-avx512skx.c",
3463 "src/qs8-igemm/gen/3x16c8-minmax-avx512skx.c",
3464 "src/qs8-igemm/gen/4x16c8-minmax-avx512skx.c",
3465]
3466
Frank Barchardbcedc082020-08-17 18:00:51 -07003467WASM32_ASM_UKERNELS = [
3468 "src/f32-relu/wasm_shr_x1.S",
3469 "src/f32-relu/wasm_shr_x2.S",
3470 "src/f32-relu/wasm_shr_x4.S",
3471]
3472
Marat Dukhan08c4a432019-10-03 09:29:21 -07003473AARCH32_ASM_UKERNELS = [
Marat Dukhan32f93812020-05-17 20:31:21 -07003474 "src/f32-gemm/4x4-aarch32-vfp-ld64.S",
Marat Dukhan3b98f6b2020-05-17 10:09:22 -07003475 "src/f32-gemm/4x4-minmax-aarch32-vfp-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003476 "src/f32-gemm/4x8-minmax-aarch32-neon-cortex-a53.S",
3477 "src/f32-gemm/4x8-minmax-aarch32-neon-cortex-a55.S",
Frank Barchard490febe2020-07-16 18:42:17 -07003478 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-cortex-a7.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003479 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-cortex-a75.S",
Frank Barchard569561d2020-06-17 13:11:12 -07003480 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-ld64.S",
Frank Barchard490febe2020-07-16 18:42:17 -07003481 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-pld-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003482 "src/f32-igemm/4x8-minmax-aarch32-neon-cortex-a53.S",
3483 "src/f32-igemm/4x8-minmax-aarch32-neon-cortex-a55.S",
Frank Barchard490febe2020-07-16 18:42:17 -07003484 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-cortex-a7.S",
3485 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-cortex-a75.S",
3486 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-ld64.S",
3487 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-pld-cortex-a75.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003488]
3489
3490AARCH64_ASM_UKERNELS = [
Frank Barchardbddfbcd2020-04-15 12:32:41 -07003491 "src/f16-gemm/gen-inc/1x8inc-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchard04336c12020-10-22 16:48:55 -07003492 "src/f16-gemm/gen-inc/1x16inc-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchardbddfbcd2020-04-15 12:32:41 -07003493 "src/f16-gemm/gen-inc/4x8inc-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchard04336c12020-10-22 16:48:55 -07003494 "src/f16-gemm/gen-inc/4x16inc-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchardbddfbcd2020-04-15 12:32:41 -07003495 "src/f16-gemm/gen-inc/6x8inc-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchard04336c12020-10-22 16:48:55 -07003496 "src/f16-gemm/gen-inc/6x16inc-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchard3b8e5662020-04-20 12:12:53 -07003497 "src/f16-gemm/gen-inc/8x8inc-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchard04336c12020-10-22 16:48:55 -07003498 "src/f16-gemm/gen/1x8-minmax-aarch64-neonfp16arith-ld64.S",
3499 "src/f16-gemm/gen/1x16-minmax-aarch64-neonfp16arith-ld32.S",
3500 "src/f16-gemm/gen/4x8-minmax-aarch64-neonfp16arith-ld64.S",
3501 "src/f16-gemm/gen/4x16-minmax-aarch64-neonfp16arith-ld32.S",
3502 "src/f16-gemm/gen/6x8-minmax-aarch64-neonfp16arith-ld64.S",
3503 "src/f16-gemm/gen/6x16-minmax-aarch64-neonfp16arith-ld32.S",
3504 "src/f16-gemm/gen/8x8-minmax-aarch64-neonfp16arith-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003505 "src/f32-dwconv/up4x9-minmax-aarch64-neonfma-cortex-a55.S",
3506 "src/f32-dwconv/up4x9-minmax-aarch64-neonfma.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003507 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a53.S",
3508 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a57.S",
3509 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a75.S",
Frank Barchard04336c12020-10-22 16:48:55 -07003510 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-ld64.S",
3511 "src/f32-gemm/gen-inc/1x12inc-minmax-aarch64-neonfma-cortex-a53.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003512 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a53.S",
3513 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a55.S",
3514 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a57.S",
3515 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003516 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-ld64.S",
Frank Barchard04336c12020-10-22 16:48:55 -07003517 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-ld128.S",
3518 "src/f32-gemm/gen-inc/4x12inc-minmax-aarch64-neonfma-cortex-a53.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003519 "src/f32-gemm/gen-inc/5x8inc-minmax-aarch64-neonfma-cortex-a57.S",
3520 "src/f32-gemm/gen-inc/5x8inc-minmax-aarch64-neonfma-cortex-a75.S",
3521 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a53.S",
3522 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a55.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003523 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a57.S",
Frank Barchard04336c12020-10-22 16:48:55 -07003524 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a73.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003525 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003526 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-ld64.S",
Frank Barchard04336c12020-10-22 16:48:55 -07003527 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-ld128.S",
3528 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a53.S",
3529 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a57.S",
3530 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a75.S",
3531 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-ld64.S",
3532 "src/f32-gemm/gen/1x12-minmax-aarch64-neonfma-cortex-a53.S",
3533 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a53.S",
3534 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a55.S",
3535 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a57.S",
3536 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a75.S",
3537 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-ld64.S",
3538 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-ld128.S",
3539 "src/f32-gemm/gen/4x12-minmax-aarch64-neonfma-cortex-a53.S",
3540 "src/f32-gemm/gen/5x8-minmax-aarch64-neonfma-cortex-a57.S",
3541 "src/f32-gemm/gen/5x8-minmax-aarch64-neonfma-cortex-a75.S",
3542 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a53.S",
3543 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a55.S",
3544 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a57.S",
3545 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a73.S",
3546 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a75.S",
3547 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-ld64.S",
3548 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-ld128.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003549 "src/f32-igemm/1x8-minmax-aarch64-neonfma-cortex-a53.S",
Frank Barchard04336c12020-10-22 16:48:55 -07003550 "src/f32-igemm/1x12-minmax-aarch64-neonfma-cortex-a53.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003551 "src/f32-igemm/4x8-minmax-aarch64-neonfma-cortex-a53.S",
3552 "src/f32-igemm/4x8-minmax-aarch64-neonfma-cortex-a55.S",
Frank Barchard04336c12020-10-22 16:48:55 -07003553 "src/f32-igemm/4x12-minmax-aarch64-neonfma-cortex-a53.S",
3554 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a53.S",
3555 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a55.S",
3556 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a73.S",
3557 "src/f32-igemm/gen/1x8-minmax-aarch64-neonfma-cortex-a57.S",
3558 "src/f32-igemm/gen/1x8-minmax-aarch64-neonfma-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003559 "src/f32-igemm/gen/4x8-minmax-aarch64-neonfma-cortex-a57.S",
3560 "src/f32-igemm/gen/4x8-minmax-aarch64-neonfma-cortex-a75.S",
3561 "src/f32-igemm/gen/5x8-minmax-aarch64-neonfma-cortex-a57.S",
3562 "src/f32-igemm/gen/5x8-minmax-aarch64-neonfma-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003563 "src/f32-igemm/gen/6x8-minmax-aarch64-neonfma-cortex-a57.S",
3564 "src/f32-igemm/gen/6x8-minmax-aarch64-neonfma-cortex-a75.S",
Frank Barchard4a4be4e2021-03-01 15:00:28 -08003565 "src/qs8-gemm/1x16c4-aarch64-neondot-ld32.S",
Frank Barchardf1fd89e2020-10-06 23:27:11 -07003566 "src/qs8-gemm/1x16c4-aarch64-neondot-ld64.S",
Frank Barchard1dc9fef2021-03-10 23:50:19 -08003567 "src/qs8-gemm/2x8c8-aarch64-neon-mull-padal.S",
Frank Barchardcbb8e702021-03-12 01:22:55 -08003568 "src/qs8-gemm/2x8c8-aarch64-neon-mlal-padal.S",
Frank Barchard5655cb72021-03-09 20:57:14 -08003569 "src/qs8-gemm/2x8c16-aarch64-neon-mlal-padal.S",
Frank Barchard671d1b02021-03-10 15:42:08 -08003570 "src/qs8-gemm/4x16c4-aarch64-neondot-cortex-a55.S",
Frank Barchardda78da12021-03-02 14:28:00 -08003571 "src/qs8-gemm/4x16c4-aarch64-neondot-ld32.S",
Frank Barchard671d1b02021-03-10 15:42:08 -08003572 "src/qs8-gemm/4x16c4-aarch64-neondot-ld64.S",
Frank Barchardcbb8e702021-03-12 01:22:55 -08003573 "src/qs8-igemm/2x8c8-aarch64-neon-mlal-padal.S",
Frank Barchard7ca54df2021-03-12 00:09:02 -08003574 "src/qs8-igemm/2x8c16-aarch64-neon-mlal-padal.S",
Frank Barchard89e12f82021-03-08 11:42:44 -08003575 "src/qs8-igemm/4x16c4-aarch64-neondot-cortex-a55.S",
Frank Barchard671d1b02021-03-10 15:42:08 -08003576 "src/qs8-igemm/4x16c4-aarch64-neondot-ld64.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003577]
3578
Marat Dukhan1b354632020-03-23 12:50:22 -07003579INTERNAL_MICROKERNEL_HDRS = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003580 "src/xnnpack/argmaxpool.h",
3581 "src/xnnpack/avgpool.h",
3582 "src/xnnpack/clamp.h",
3583 "src/xnnpack/common.h",
3584 "src/xnnpack/conv.h",
Yury Kartynnike7841862020-11-04 18:22:18 -08003585 "src/xnnpack/depthtospace.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003586 "src/xnnpack/dwconv.h",
Frank Barchard04336c12020-10-22 16:48:55 -07003587 "src/xnnpack/fill.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003588 "src/xnnpack/gavgpool.h",
3589 "src/xnnpack/gemm.h",
3590 "src/xnnpack/hswish.h",
Marat Dukhan660fd192020-03-10 04:55:30 -07003591 "src/xnnpack/ibilinear.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003592 "src/xnnpack/igemm.h",
Marat Dukhancfb31342019-12-05 10:42:57 -08003593 "src/xnnpack/intrinsics-polyfill.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003594 "src/xnnpack/lut.h",
3595 "src/xnnpack/math.h",
3596 "src/xnnpack/maxpool.h",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003597 "src/xnnpack/memory.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003598 "src/xnnpack/packx.h",
3599 "src/xnnpack/pad.h",
3600 "src/xnnpack/params.h",
3601 "src/xnnpack/pavgpool.h",
3602 "src/xnnpack/ppmm.h",
3603 "src/xnnpack/prelu.h",
Marat Dukhan97579532019-10-18 16:40:39 -07003604 "src/xnnpack/raddexpminusmax.h",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07003605 "src/xnnpack/raddextexp.h",
Marat Dukhan97579532019-10-18 16:40:39 -07003606 "src/xnnpack/raddstoreexpminusmax.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003607 "src/xnnpack/rmax.h",
3608 "src/xnnpack/scalar-utils.h",
3609 "src/xnnpack/spmm.h",
3610 "src/xnnpack/unpool.h",
3611 "src/xnnpack/vadd.h",
Marat Dukhan1e782c42019-11-21 17:02:40 -08003612 "src/xnnpack/vbinary.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003613 "src/xnnpack/vmulcaddc.h",
Marat Dukhan05ac8e32019-10-21 15:39:33 -07003614 "src/xnnpack/vscale.h",
Marat Dukhan97579532019-10-18 16:40:39 -07003615 "src/xnnpack/vscaleexpminusmax.h",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07003616 "src/xnnpack/vscaleextexp.h",
Marat Dukhan1e782c42019-11-21 17:02:40 -08003617 "src/xnnpack/vunary.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003618 "src/xnnpack/zip.h",
Marat Dukhan1b354632020-03-23 12:50:22 -07003619]
3620
3621INTERNAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003622 "include/xnnpack.h",
3623 "src/xnnpack/allocator.h",
3624 "src/xnnpack/compute.h",
3625 "src/xnnpack/im2col.h",
3626 "src/xnnpack/indirection.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003627 "src/xnnpack/math-stubs.h",
Chao Mei6ddfc602020-05-13 22:29:36 -07003628 "src/xnnpack/memory-planner.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003629 "src/xnnpack/operator.h",
3630 "src/xnnpack/pack.h",
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07003631 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003632 "src/xnnpack/requantization-stubs.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003633 "src/xnnpack/requantization.h",
Marat Dukhan1d75a542020-02-03 12:23:01 -08003634 "src/xnnpack/subgraph.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003635]
3636
Marat Dukhan1b354632020-03-23 12:50:22 -07003637ACCURACY_EVAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003638 "src/xnnpack/math-stubs.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003639]
3640
Marat Dukhan1b354632020-03-23 12:50:22 -07003641MICROKERNEL_BENCHMARK_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003642 "include/xnnpack.h",
Frank Barchard35db7d02020-10-26 13:37:34 -07003643 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003644]
3645
Marat Dukhan1b354632020-03-23 12:50:22 -07003646MICROKERNEL_TEST_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Frank Barchard35db7d02020-10-26 13:37:34 -07003647 "include/xnnpack.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003648 "src/xnnpack/isa-checks.h",
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07003649 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003650 "src/xnnpack/requantization.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003651]
3652
3653OPERATOR_TEST_PARAMS_HDRS = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003654 "src/xnnpack/common.h",
Frank Barchard04336c12020-10-22 16:48:55 -07003655 "src/xnnpack/params.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003656]
3657
3658WEIGHTS_PACK_HDRS = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003659 "src/xnnpack/compute.h",
Frank Barchard04336c12020-10-22 16:48:55 -07003660 "src/xnnpack/operator.h",
3661 "src/xnnpack/pack.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003662]
3663
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07003664LOGGING_COPTS = select({
3665 # No logging in optimized mode
3666 ":optimized_build": ["-DXNN_LOG_LEVEL=0"],
3667 # Full logging in debug mode
3668 ":debug_build": ["-DXNN_LOG_LEVEL=5"],
3669 # Error-only logging in default (fastbuild) mode
3670 "//conditions:default": ["-DXNN_LOG_LEVEL=2"],
3671})
3672
Marat Dukhan3b59de22020-06-03 20:15:19 -07003673LOGGING_SRCS = select({
3674 # No logging in optimized mode
3675 ":optimized_build": [],
3676 "//conditions:default": [
3677 "src/operator-strings.c",
3678 "src/subgraph-strings.c",
3679 ],
3680})
3681
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07003682LOGGING_HDRS = [
3683 "src/xnnpack/log.h",
3684]
3685
Marat Dukhan08c4a432019-10-03 09:29:21 -07003686xnnpack_cc_library(
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003687 name = "tables",
3688 srcs = TABLE_SRCS,
3689 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003690 gcc_copts = xnnpack_gcc_std_copts(),
3691 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003692)
3693
3694xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003695 name = "scalar_ukernels",
3696 srcs = SCALAR_UKERNELS,
3697 hdrs = INTERNAL_HDRS,
3698 aarch32_copts = ["-marm"],
Marat Dukhan10a38082020-04-17 03:58:35 -07003699 gcc_copts = xnnpack_gcc_std_copts(),
3700 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003701 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003702 ":tables",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003703 "@FP16",
3704 "@FXdiv",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003705 "@pthreadpool",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003706 ],
3707)
3708
3709xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003710 name = "scalar_ukernels_test_mode",
3711 srcs = SCALAR_UKERNELS,
3712 hdrs = INTERNAL_HDRS,
3713 aarch32_copts = ["-marm"],
3714 copts = [
3715 "-UNDEBUG",
3716 "-DXNN_TEST_MODE=1",
3717 ],
3718 gcc_copts = xnnpack_gcc_std_copts(),
3719 msvc_copts = xnnpack_msvc_std_copts(),
3720 deps = [
3721 ":tables",
3722 "@FP16",
3723 "@FXdiv",
3724 "@pthreadpool",
3725 ],
3726)
3727
3728xnnpack_cc_library(
Marat Dukhan436ebe62019-12-04 15:10:12 -08003729 name = "wasm_ukernels",
3730 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003731 gcc_copts = xnnpack_gcc_std_copts(),
3732 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan436ebe62019-12-04 15:10:12 -08003733 wasm_srcs = WASM_UKERNELS,
Marat Dukhan1c6e3892020-06-25 23:56:10 -07003734 wasmsimd_srcs = WASM_UKERNELS + WASMSIMD_UKERNELS,
Marat Dukhan436ebe62019-12-04 15:10:12 -08003735 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003736 ":tables",
Marat Dukhan436ebe62019-12-04 15:10:12 -08003737 "@FP16",
3738 "@FXdiv",
3739 "@pthreadpool",
3740 ],
3741)
3742
3743xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003744 name = "wasm_ukernels_test_mode",
3745 hdrs = INTERNAL_HDRS,
3746 copts = [
3747 "-UNDEBUG",
3748 "-DXNN_TEST_MODE=1",
3749 ],
3750 gcc_copts = xnnpack_gcc_std_copts(),
3751 msvc_copts = xnnpack_msvc_std_copts(),
3752 wasm_srcs = WASM_UKERNELS,
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07003753 wasmsimd_srcs = WASM_UKERNELS + WASMSIMD_UKERNELS,
Marat Dukhan33fcf782020-05-24 14:27:15 -07003754 deps = [
3755 ":tables",
3756 "@FP16",
3757 "@FXdiv",
3758 "@pthreadpool",
3759 ],
3760)
3761
3762xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003763 name = "neon_ukernels",
3764 hdrs = INTERNAL_HDRS,
3765 aarch32_copts = [
3766 "-marm",
Marat Dukhan8853b822020-05-07 12:19:01 -07003767 "-march=armv7-a",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003768 "-mfpu=neon",
3769 ],
3770 aarch32_srcs = NEON_UKERNELS,
3771 aarch64_srcs = NEON_UKERNELS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003772 gcc_copts = xnnpack_gcc_std_copts(),
3773 msvc_copts = xnnpack_msvc_std_copts(),
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 Dukhan08c4a432019-10-03 09:29:21 -07003779)
3780
3781xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003782 name = "neon_ukernels_test_mode",
3783 hdrs = INTERNAL_HDRS,
3784 aarch32_copts = [
3785 "-marm",
3786 "-march=armv7-a",
3787 "-mfpu=neon",
3788 ],
3789 aarch32_srcs = NEON_UKERNELS,
3790 aarch64_srcs = NEON_UKERNELS,
3791 copts = [
3792 "-UNDEBUG",
3793 "-DXNN_TEST_MODE=1",
3794 ],
3795 gcc_copts = xnnpack_gcc_std_copts(),
3796 msvc_copts = xnnpack_msvc_std_copts(),
3797 deps = [
3798 ":tables",
3799 "@FP16",
3800 "@pthreadpool",
3801 ],
3802)
3803
3804xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003805 name = "neonfma_ukernels",
3806 hdrs = INTERNAL_HDRS,
3807 aarch32_copts = [
3808 "-marm",
Marat Dukhan8853b822020-05-07 12:19:01 -07003809 "-march=armv7-a",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003810 "-mfpu=neon-vfpv4",
3811 ],
3812 aarch32_srcs = NEONFMA_UKERNELS,
3813 aarch64_srcs = NEONFMA_UKERNELS + AARCH64_NEONFMA_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07003814 apple_aarch32_copts = [
3815 "-mcpu=swift",
3816 "-mtune=generic",
3817 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07003818 gcc_copts = xnnpack_gcc_std_copts(),
3819 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08003820 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003821 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003822 "@FP16",
3823 "@pthreadpool",
3824 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003825)
3826
3827xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003828 name = "neonfma_ukernels_test_mode",
3829 hdrs = INTERNAL_HDRS,
3830 aarch32_copts = [
3831 "-marm",
3832 "-march=armv7-a",
3833 "-mfpu=neon-vfpv4",
3834 ],
3835 aarch32_srcs = NEONFMA_UKERNELS,
3836 aarch64_srcs = NEONFMA_UKERNELS + AARCH64_NEONFMA_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07003837 apple_aarch32_copts = [
3838 "-mcpu=swift",
3839 "-mtune=generic",
3840 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07003841 copts = [
3842 "-UNDEBUG",
3843 "-DXNN_TEST_MODE=1",
3844 ],
3845 gcc_copts = xnnpack_gcc_std_copts(),
3846 msvc_copts = xnnpack_msvc_std_copts(),
3847 deps = [
3848 ":tables",
3849 "@FP16",
3850 "@pthreadpool",
3851 ],
3852)
3853
3854xnnpack_cc_library(
Marat Dukhan8853b822020-05-07 12:19:01 -07003855 name = "neonv8_ukernels",
3856 hdrs = INTERNAL_HDRS,
3857 aarch32_copts = [
3858 "-marm",
3859 "-march=armv8-a",
3860 "-mfpu=neon-fp-armv8",
3861 ],
3862 aarch32_srcs = NEONV8_UKERNELS,
3863 aarch64_srcs = NEONV8_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07003864 apple_aarch32_copts = [
3865 "-mcpu=cyclone",
3866 "-mtune=generic",
3867 ],
Marat Dukhan8853b822020-05-07 12:19:01 -07003868 gcc_copts = xnnpack_gcc_std_copts(),
3869 msvc_copts = xnnpack_msvc_std_copts(),
3870 deps = [
3871 ":tables",
3872 "@FP16",
3873 "@pthreadpool",
3874 ],
3875)
3876
3877xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003878 name = "neonv8_ukernels_test_mode",
3879 hdrs = INTERNAL_HDRS,
3880 aarch32_copts = [
3881 "-marm",
3882 "-march=armv8-a",
3883 "-mfpu=neon-fp-armv8",
3884 ],
3885 aarch32_srcs = NEONV8_UKERNELS,
3886 aarch64_srcs = NEONV8_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07003887 apple_aarch32_copts = [
3888 "-mcpu=cyclone",
3889 "-mtune=generic",
3890 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07003891 copts = [
3892 "-UNDEBUG",
3893 "-DXNN_TEST_MODE=1",
3894 ],
3895 gcc_copts = xnnpack_gcc_std_copts(),
3896 msvc_copts = xnnpack_msvc_std_copts(),
3897 deps = [
3898 ":tables",
3899 "@FP16",
3900 "@pthreadpool",
3901 ],
3902)
3903
3904xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003905 name = "neonfp16arith_ukernels",
3906 hdrs = INTERNAL_HDRS,
3907 aarch64_copts = ["-march=armv8.2-a+fp16"],
3908 aarch64_srcs = AARCH64_NEONFP16ARITH_UKERNELS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003909 gcc_copts = xnnpack_gcc_std_copts(),
3910 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08003911 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003912 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003913 "@FP16",
3914 "@pthreadpool",
3915 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003916)
3917
3918xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003919 name = "neonfp16arith_ukernels_test_mode",
3920 hdrs = INTERNAL_HDRS,
3921 aarch64_copts = ["-march=armv8.2-a+fp16"],
3922 aarch64_srcs = AARCH64_NEONFP16ARITH_UKERNELS,
3923 copts = [
3924 "-UNDEBUG",
3925 "-DXNN_TEST_MODE=1",
3926 ],
3927 gcc_copts = xnnpack_gcc_std_copts(),
3928 msvc_copts = xnnpack_msvc_std_copts(),
3929 deps = [
3930 ":tables",
3931 "@FP16",
3932 "@pthreadpool",
3933 ],
3934)
3935
3936xnnpack_cc_library(
Benoit Jacoba9644732020-08-13 12:48:55 -07003937 name = "neondot_ukernels",
3938 hdrs = INTERNAL_HDRS,
Marat Dukhan799ac752020-08-13 16:08:03 -07003939 aarch32_copts = [
3940 "-marm",
3941 "-march=armv8.2-a+dotprod",
3942 "-mfpu=neon-fp-armv8",
3943 ],
Benoit Jacoba9644732020-08-13 12:48:55 -07003944 aarch32_srcs = NEONDOT_UKERNELS,
3945 aarch64_copts = ["-march=armv8.2-a+dotprod"],
3946 aarch64_srcs = NEONDOT_UKERNELS,
3947 gcc_copts = xnnpack_gcc_std_copts(),
3948 msvc_copts = xnnpack_msvc_std_copts(),
3949 deps = [
3950 ":tables",
3951 "@FP16",
3952 "@pthreadpool",
3953 ],
3954)
3955
3956xnnpack_cc_library(
3957 name = "neondot_ukernels_test_mode",
3958 hdrs = INTERNAL_HDRS,
Marat Dukhan799ac752020-08-13 16:08:03 -07003959 aarch32_copts = [
3960 "-marm",
3961 "-march=armv8.2-a+dotprod",
3962 "-mfpu=neon-fp-armv8",
3963 ],
Benoit Jacoba9644732020-08-13 12:48:55 -07003964 aarch32_srcs = NEONDOT_UKERNELS,
3965 aarch64_copts = ["-march=armv8.2-a+dotprod"],
3966 aarch64_srcs = NEONDOT_UKERNELS,
3967 copts = [
3968 "-UNDEBUG",
3969 "-DXNN_TEST_MODE=1",
3970 ],
3971 gcc_copts = xnnpack_gcc_std_copts(),
3972 msvc_copts = xnnpack_msvc_std_copts(),
3973 deps = [
3974 ":tables",
3975 "@FP16",
3976 "@pthreadpool",
3977 ],
3978)
3979
3980xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003981 name = "sse2_ukernels",
3982 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003983 gcc_copts = xnnpack_gcc_std_copts(),
3984 gcc_x86_copts = ["-msse2"],
3985 msvc_copts = xnnpack_msvc_std_copts(),
3986 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003987 x86_srcs = SSE_UKERNELS + SSE2_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08003988 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003989 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003990 "@FP16",
3991 "@pthreadpool",
3992 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003993)
3994
3995xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003996 name = "sse2_ukernels_test_mode",
3997 hdrs = INTERNAL_HDRS,
3998 copts = [
3999 "-UNDEBUG",
4000 "-DXNN_TEST_MODE=1",
4001 ],
4002 gcc_copts = xnnpack_gcc_std_copts(),
4003 gcc_x86_copts = ["-msse2"],
4004 msvc_copts = xnnpack_msvc_std_copts(),
4005 msvc_x86_32_copts = ["/arch:SSE2"],
4006 x86_srcs = SSE_UKERNELS + SSE2_UKERNELS,
4007 deps = [
4008 ":tables",
4009 "@FP16",
4010 "@pthreadpool",
4011 ],
4012)
4013
4014xnnpack_cc_library(
Marat Dukhanfe7acb62020-03-09 19:30:05 -07004015 name = "ssse3_ukernels",
4016 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004017 gcc_copts = xnnpack_gcc_std_copts(),
4018 gcc_x86_copts = ["-mssse3"],
4019 msvc_copts = xnnpack_msvc_std_copts(),
4020 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhanfe7acb62020-03-09 19:30:05 -07004021 x86_srcs = SSSE3_UKERNELS,
4022 deps = [
4023 ":tables",
4024 "@FP16",
4025 "@pthreadpool",
4026 ],
4027)
4028
4029xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07004030 name = "ssse3_ukernels_test_mode",
4031 hdrs = INTERNAL_HDRS,
4032 copts = [
4033 "-UNDEBUG",
4034 "-DXNN_TEST_MODE=1",
4035 ],
4036 gcc_copts = xnnpack_gcc_std_copts(),
4037 gcc_x86_copts = ["-mssse3"],
4038 msvc_copts = xnnpack_msvc_std_copts(),
4039 msvc_x86_32_copts = ["/arch:SSE2"],
4040 x86_srcs = SSSE3_UKERNELS,
4041 deps = [
4042 ":tables",
4043 "@FP16",
4044 "@pthreadpool",
4045 ],
4046)
4047
4048xnnpack_cc_library(
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08004049 name = "sse41_ukernels",
4050 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004051 gcc_copts = xnnpack_gcc_std_copts(),
4052 gcc_x86_copts = ["-msse4.1"],
4053 msvc_copts = xnnpack_msvc_std_copts(),
4054 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08004055 x86_srcs = SSE41_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08004056 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08004057 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08004058 "@FP16",
4059 "@pthreadpool",
4060 ],
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08004061)
4062
4063xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07004064 name = "sse41_ukernels_test_mode",
4065 hdrs = INTERNAL_HDRS,
4066 copts = [
4067 "-UNDEBUG",
4068 "-DXNN_TEST_MODE=1",
4069 ],
4070 gcc_copts = xnnpack_gcc_std_copts(),
4071 gcc_x86_copts = ["-msse4.1"],
4072 msvc_copts = xnnpack_msvc_std_copts(),
4073 msvc_x86_32_copts = ["/arch:SSE2"],
4074 x86_srcs = SSE41_UKERNELS,
4075 deps = [
4076 ":tables",
4077 "@FP16",
4078 "@pthreadpool",
4079 ],
4080)
4081
4082xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004083 name = "avx_ukernels",
4084 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004085 gcc_copts = xnnpack_gcc_std_copts(),
4086 gcc_x86_copts = ["-mavx"],
4087 msvc_copts = xnnpack_msvc_std_copts(),
4088 msvc_x86_32_copts = ["/arch:AVX"],
4089 msvc_x86_64_copts = ["/arch:AVX"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004090 x86_srcs = AVX_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08004091 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08004092 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08004093 "@FP16",
4094 "@pthreadpool",
4095 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004096)
4097
4098xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07004099 name = "avx_ukernels_test_mode",
4100 hdrs = INTERNAL_HDRS,
4101 copts = [
4102 "-UNDEBUG",
4103 "-DXNN_TEST_MODE=1",
4104 ],
4105 gcc_copts = xnnpack_gcc_std_copts(),
4106 gcc_x86_copts = ["-mavx"],
4107 msvc_copts = xnnpack_msvc_std_copts(),
4108 msvc_x86_32_copts = ["/arch:AVX"],
4109 msvc_x86_64_copts = ["/arch:AVX"],
4110 x86_srcs = AVX_UKERNELS,
4111 deps = [
4112 ":tables",
4113 "@FP16",
4114 "@pthreadpool",
4115 ],
4116)
4117
4118xnnpack_cc_library(
Marat Dukhan1566fee2020-08-02 21:55:41 -07004119 name = "xop_ukernels",
4120 hdrs = INTERNAL_HDRS,
4121 gcc_copts = xnnpack_gcc_std_copts(),
4122 gcc_x86_copts = ["-mxop"],
4123 msvc_copts = xnnpack_msvc_std_copts(),
4124 msvc_x86_32_copts = ["/arch:AVX"],
4125 msvc_x86_64_copts = ["/arch:AVX"],
4126 x86_srcs = XOP_UKERNELS,
4127 deps = [
4128 ":tables",
4129 "@FP16",
4130 "@pthreadpool",
4131 ],
4132)
4133
4134xnnpack_cc_library(
4135 name = "xop_ukernels_test_mode",
4136 hdrs = INTERNAL_HDRS,
4137 copts = [
4138 "-UNDEBUG",
4139 "-DXNN_TEST_MODE=1",
4140 ],
4141 gcc_copts = xnnpack_gcc_std_copts(),
4142 gcc_x86_copts = ["-mxop"],
4143 msvc_copts = xnnpack_msvc_std_copts(),
4144 msvc_x86_32_copts = ["/arch:AVX"],
4145 msvc_x86_64_copts = ["/arch:AVX"],
4146 x86_srcs = XOP_UKERNELS,
4147 deps = [
4148 ":tables",
4149 "@FP16",
4150 "@pthreadpool",
4151 ],
4152)
4153
4154xnnpack_cc_library(
Marat Dukhanfda12b82019-11-21 12:27:59 -08004155 name = "fma3_ukernels",
4156 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004157 gcc_copts = xnnpack_gcc_std_copts(),
4158 gcc_x86_copts = ["-mfma"],
4159 msvc_copts = xnnpack_msvc_std_copts(),
4160 msvc_x86_32_copts = ["/arch:AVX"],
4161 msvc_x86_64_copts = ["/arch:AVX"],
Marat Dukhanfda12b82019-11-21 12:27:59 -08004162 x86_srcs = FMA3_UKERNELS,
4163 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08004164 ":tables",
Marat Dukhanfda12b82019-11-21 12:27:59 -08004165 "@FP16",
4166 "@pthreadpool",
4167 ],
4168)
4169
4170xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07004171 name = "fma3_ukernels_test_mode",
4172 hdrs = INTERNAL_HDRS,
4173 copts = [
4174 "-UNDEBUG",
4175 "-DXNN_TEST_MODE=1",
4176 ],
4177 gcc_copts = xnnpack_gcc_std_copts(),
4178 gcc_x86_copts = ["-mfma"],
4179 msvc_copts = xnnpack_msvc_std_copts(),
4180 msvc_x86_32_copts = ["/arch:AVX"],
4181 msvc_x86_64_copts = ["/arch:AVX"],
4182 x86_srcs = FMA3_UKERNELS,
4183 deps = [
4184 ":tables",
4185 "@FP16",
4186 "@pthreadpool",
4187 ],
4188)
4189
4190xnnpack_cc_library(
Marat Dukhan6adff4e2019-10-14 18:32:07 -07004191 name = "avx2_ukernels",
4192 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004193 gcc_copts = xnnpack_gcc_std_copts(),
4194 gcc_x86_copts = [
Marat Dukhan6adff4e2019-10-14 18:32:07 -07004195 "-mfma",
4196 "-mavx2",
4197 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07004198 msvc_copts = xnnpack_msvc_std_copts(),
4199 msvc_x86_32_copts = ["/arch:AVX2"],
4200 msvc_x86_64_copts = ["/arch:AVX2"],
Marat Dukhan6adff4e2019-10-14 18:32:07 -07004201 x86_srcs = AVX2_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08004202 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08004203 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08004204 "@FP16",
4205 "@pthreadpool",
4206 ],
Marat Dukhan6adff4e2019-10-14 18:32:07 -07004207)
4208
4209xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07004210 name = "avx2_ukernels_test_mode",
4211 hdrs = INTERNAL_HDRS,
4212 copts = [
4213 "-UNDEBUG",
4214 "-DXNN_TEST_MODE=1",
4215 ],
4216 gcc_copts = xnnpack_gcc_std_copts(),
4217 gcc_x86_copts = [
4218 "-mfma",
4219 "-mavx2",
4220 ],
4221 msvc_copts = xnnpack_msvc_std_copts(),
4222 msvc_x86_32_copts = ["/arch:AVX2"],
4223 msvc_x86_64_copts = ["/arch:AVX2"],
4224 x86_srcs = AVX2_UKERNELS,
4225 deps = [
4226 ":tables",
4227 "@FP16",
4228 "@pthreadpool",
4229 ],
4230)
4231
4232xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004233 name = "avx512f_ukernels",
4234 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004235 gcc_copts = xnnpack_gcc_std_copts(),
4236 gcc_x86_copts = ["-mavx512f"],
4237 mingw_copts = ["-fno-asynchronous-unwind-tables"],
4238 msvc_copts = xnnpack_msvc_std_copts(),
4239 msvc_x86_32_copts = ["/arch:AVX512"],
4240 msvc_x86_64_copts = ["/arch:AVX512"],
4241 msys_copts = ["-fno-asynchronous-unwind-tables"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004242 x86_srcs = AVX512F_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08004243 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08004244 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08004245 "@FP16",
4246 "@pthreadpool",
4247 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004248)
4249
4250xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07004251 name = "avx512f_ukernels_test_mode",
4252 hdrs = INTERNAL_HDRS,
4253 copts = [
4254 "-UNDEBUG",
4255 "-DXNN_TEST_MODE=1",
4256 ],
4257 gcc_copts = xnnpack_gcc_std_copts(),
4258 gcc_x86_copts = ["-mavx512f"],
4259 mingw_copts = ["-fno-asynchronous-unwind-tables"],
4260 msvc_copts = xnnpack_msvc_std_copts(),
4261 msvc_x86_32_copts = ["/arch:AVX512"],
4262 msvc_x86_64_copts = ["/arch:AVX512"],
4263 msys_copts = ["-fno-asynchronous-unwind-tables"],
4264 x86_srcs = AVX512F_UKERNELS,
4265 deps = [
4266 ":tables",
4267 "@FP16",
4268 "@pthreadpool",
4269 ],
4270)
4271
4272xnnpack_cc_library(
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07004273 name = "avx512skx_ukernels",
4274 hdrs = INTERNAL_HDRS,
4275 gcc_copts = xnnpack_gcc_std_copts(),
4276 gcc_x86_copts = [
4277 "-mavx512f",
4278 "-mavx512cd",
4279 "-mavx512bw",
4280 "-mavx512dq",
4281 "-mavx512vl",
4282 ],
4283 mingw_copts = ["-fno-asynchronous-unwind-tables"],
4284 msvc_copts = xnnpack_msvc_std_copts(),
4285 msvc_x86_32_copts = ["/arch:AVX512"],
4286 msvc_x86_64_copts = ["/arch:AVX512"],
4287 msys_copts = ["-fno-asynchronous-unwind-tables"],
4288 x86_srcs = AVX512SKX_UKERNELS,
4289 deps = [
4290 ":tables",
4291 "@FP16",
4292 "@pthreadpool",
4293 ],
4294)
4295
4296xnnpack_cc_library(
4297 name = "avx512skx_ukernels_test_mode",
4298 hdrs = INTERNAL_HDRS,
4299 copts = [
4300 "-UNDEBUG",
4301 "-DXNN_TEST_MODE=1",
4302 ],
4303 gcc_copts = xnnpack_gcc_std_copts(),
4304 gcc_x86_copts = [
4305 "-mavx512f",
4306 "-mavx512cd",
4307 "-mavx512bw",
4308 "-mavx512dq",
4309 "-mavx512vl",
4310 ],
4311 mingw_copts = ["-fno-asynchronous-unwind-tables"],
4312 msvc_copts = xnnpack_msvc_std_copts(),
4313 msvc_x86_32_copts = ["/arch:AVX512"],
4314 msvc_x86_64_copts = ["/arch:AVX512"],
4315 msys_copts = ["-fno-asynchronous-unwind-tables"],
4316 x86_srcs = AVX512SKX_UKERNELS,
4317 deps = [
4318 ":tables",
4319 "@FP16",
4320 "@pthreadpool",
4321 ],
4322)
4323
4324xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004325 name = "asm_ukernels",
4326 hdrs = ["src/xnnpack/assembly.h"],
4327 aarch32_srcs = AARCH32_ASM_UKERNELS,
Frank Barchard31bb45b2020-10-06 00:26:33 -07004328 aarch64_copts = ["-march=armv8.2-a+fp16+dotprod"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004329 aarch64_srcs = AARCH64_ASM_UKERNELS,
Frank Barchardbcedc082020-08-17 18:00:51 -07004330 wasm_srcs = WASM32_ASM_UKERNELS,
4331 wasmsimd_srcs = WASM32_ASM_UKERNELS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07004332)
4333
Marat Dukhan3b59de22020-06-03 20:15:19 -07004334xnnpack_cc_library(
4335 name = "logging_utils",
4336 srcs = LOGGING_SRCS,
4337 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
4338 copts = LOGGING_COPTS + [
4339 "-Isrc",
4340 "-Iinclude",
4341 ] + select({
4342 ":debug_build": [],
4343 "//conditions:default": xnnpack_min_size_copts(),
4344 }),
4345 gcc_copts = xnnpack_gcc_std_copts(),
4346 msvc_copts = xnnpack_msvc_std_copts(),
4347 visibility = xnnpack_visibility(),
4348 deps = [
4349 "@FP16",
4350 "@clog",
4351 "@pthreadpool",
4352 ],
4353)
4354
Marat Dukhan08c4a432019-10-03 09:29:21 -07004355xnnpack_aggregate_library(
4356 name = "ukernels",
4357 aarch32_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07004358 ":neon_ukernels",
4359 ":neonfma_ukernels",
Marat Dukhan8853b822020-05-07 12:19:01 -07004360 ":neonv8_ukernels",
Benoit Jacoba9644732020-08-13 12:48:55 -07004361 ":neondot_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004362 ":asm_ukernels",
4363 ],
4364 aarch64_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07004365 ":neon_ukernels",
4366 ":neonfma_ukernels",
Marat Dukhan8853b822020-05-07 12:19:01 -07004367 ":neonv8_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004368 ":neonfp16arith_ukernels",
Benoit Jacoba9644732020-08-13 12:48:55 -07004369 ":neondot_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004370 ":asm_ukernels",
4371 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07004372 generic_deps = [
4373 ":scalar_ukernels",
4374 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07004375 wasm_deps = [
4376 ":wasm_ukernels",
Frank Barchardbcedc082020-08-17 18:00:51 -07004377 ":asm_ukernels",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004378 ],
4379 wasmsimd_deps = [
4380 ":wasm_ukernels",
Frank Barchardbcedc082020-08-17 18:00:51 -07004381 ":asm_ukernels",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004382 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004383 x86_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07004384 ":sse2_ukernels",
Marat Dukhanfe7acb62020-03-09 19:30:05 -07004385 ":ssse3_ukernels",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08004386 ":sse41_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004387 ":avx_ukernels",
Marat Dukhan1566fee2020-08-02 21:55:41 -07004388 ":xop_ukernels",
Marat Dukhanfda12b82019-11-21 12:27:59 -08004389 ":fma3_ukernels",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07004390 ":avx2_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004391 ":avx512f_ukernels",
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07004392 ":avx512skx_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004393 ],
4394)
4395
Marat Dukhan33fcf782020-05-24 14:27:15 -07004396xnnpack_aggregate_library(
4397 name = "ukernels_test_mode",
4398 aarch32_deps = [
4399 ":neon_ukernels_test_mode",
4400 ":neonfma_ukernels_test_mode",
4401 ":neonv8_ukernels_test_mode",
Benoit Jacoba9644732020-08-13 12:48:55 -07004402 ":neondot_ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004403 ":asm_ukernels",
4404 ],
4405 aarch64_deps = [
4406 ":neon_ukernels_test_mode",
4407 ":neonfma_ukernels_test_mode",
4408 ":neonv8_ukernels_test_mode",
4409 ":neonfp16arith_ukernels_test_mode",
Benoit Jacoba9644732020-08-13 12:48:55 -07004410 ":neondot_ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004411 ":asm_ukernels",
4412 ],
4413 generic_deps = [
4414 ":scalar_ukernels_test_mode",
4415 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07004416 wasm_deps = [
4417 ":wasm_ukernels_test_mode",
Frank Barchardbcedc082020-08-17 18:00:51 -07004418 ":asm_ukernels",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004419 ],
4420 wasmsimd_deps = [
4421 ":wasm_ukernels_test_mode",
Frank Barchardbcedc082020-08-17 18:00:51 -07004422 ":asm_ukernels",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004423 ],
4424 x86_deps = [
4425 ":sse2_ukernels_test_mode",
4426 ":ssse3_ukernels_test_mode",
4427 ":sse41_ukernels_test_mode",
4428 ":avx_ukernels_test_mode",
Marat Dukhan1566fee2020-08-02 21:55:41 -07004429 ":xop_ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004430 ":fma3_ukernels_test_mode",
4431 ":avx2_ukernels_test_mode",
4432 ":avx512f_ukernels_test_mode",
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07004433 ":avx512skx_ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004434 ],
4435)
4436
Marat Dukhan86df3842020-06-17 14:03:31 -07004437# This target must be compatible with Android NDK r17 and compile with Clang 6.
4438# It is similar to :ukernels target, but excludes microkernels which require
4439# NEON FP16 arithmetics, NEON Dot Product, and newer intrinsics.
Marat Dukhanae046f52020-06-15 13:16:14 -07004440xnnpack_aggregate_library(
Marat Dukhan86df3842020-06-17 14:03:31 -07004441 name = "ukernels_max_compatibility",
Marat Dukhanae046f52020-06-15 13:16:14 -07004442 aarch32_deps = [
4443 ":neon_ukernels",
4444 ":neonfma_ukernels",
4445 ":neonv8_ukernels",
4446 ":asm_ukernels",
4447 ],
4448 aarch64_deps = [
4449 ":neon_ukernels",
4450 ":neonfma_ukernels",
4451 ":neonv8_ukernels",
4452 ":asm_ukernels",
4453 ],
4454 generic_deps = [
4455 ":scalar_ukernels",
4456 ],
Marat Dukhanae046f52020-06-15 13:16:14 -07004457 wasm_deps = [
4458 ":wasm_ukernels",
4459 ],
4460 wasmsimd_deps = [
4461 ":wasm_ukernels",
4462 ],
4463 x86_deps = [
4464 ":sse2_ukernels",
4465 ":ssse3_ukernels",
4466 ":sse41_ukernels",
4467 ":avx_ukernels",
Marat Dukhan1566fee2020-08-02 21:55:41 -07004468 ":xop_ukernels",
Marat Dukhanae046f52020-06-15 13:16:14 -07004469 ":fma3_ukernels",
4470 ":avx2_ukernels",
4471 ":avx512f_ukernels",
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07004472 ":avx512skx_ukernels",
Marat Dukhanae046f52020-06-15 13:16:14 -07004473 ],
4474)
4475
Marat Dukhan08c4a432019-10-03 09:29:21 -07004476xnnpack_cc_library(
4477 name = "im2col",
4478 srcs = ["src/im2col.c"],
4479 hdrs = [
4480 "src/xnnpack/common.h",
4481 "src/xnnpack/im2col.h",
4482 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07004483 gcc_copts = xnnpack_gcc_std_copts(),
4484 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004485)
4486
4487xnnpack_cc_library(
4488 name = "indirection",
4489 srcs = ["src/indirection.c"],
4490 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004491 gcc_copts = xnnpack_gcc_std_copts(),
4492 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004493 deps = [
4494 "@FP16",
4495 "@FXdiv",
4496 "@pthreadpool",
4497 ],
4498)
4499
4500xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07004501 name = "indirection_test_mode",
4502 srcs = ["src/indirection.c"],
4503 hdrs = INTERNAL_HDRS,
4504 copts = [
4505 "-UNDEBUG",
4506 "-DXNN_TEST_MODE=1",
4507 ],
4508 gcc_copts = xnnpack_gcc_std_copts(),
4509 msvc_copts = xnnpack_msvc_std_copts(),
4510 deps = [
4511 "@FP16",
4512 "@FXdiv",
4513 "@pthreadpool",
4514 ],
4515)
4516
4517xnnpack_cc_library(
Marat Dukhanab582382020-07-06 13:32:08 -07004518 name = "packing",
4519 srcs = ["src/packing.c"],
4520 hdrs = INTERNAL_HDRS,
4521 gcc_copts = xnnpack_gcc_std_copts(),
4522 msvc_copts = xnnpack_msvc_std_copts(),
4523 deps = [
4524 "@FP16",
4525 "@FXdiv",
4526 "@pthreadpool",
4527 ],
4528)
4529
4530xnnpack_cc_library(
4531 name = "packing_test_mode",
4532 srcs = ["src/packing.c"],
4533 hdrs = INTERNAL_HDRS,
4534 copts = [
4535 "-UNDEBUG",
4536 "-DXNN_TEST_MODE=1",
4537 ],
4538 gcc_copts = xnnpack_gcc_std_copts(),
4539 msvc_copts = xnnpack_msvc_std_copts(),
4540 deps = [
4541 "@FP16",
4542 "@FXdiv",
4543 "@pthreadpool",
4544 ],
4545)
4546
4547xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004548 name = "operator_run",
4549 srcs = ["src/operator-run.c"],
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07004550 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004551 copts = LOGGING_COPTS + select({
Marat Dukhan05702cf2020-03-26 15:41:33 -07004552 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4553 "//conditions:default": [],
4554 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07004555 gcc_copts = xnnpack_gcc_std_copts(),
4556 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004557 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07004558 ":logging_utils",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004559 "@FP16",
4560 "@FXdiv",
4561 "@clog",
4562 "@pthreadpool",
4563 ],
4564)
4565
Chao Mei6ddfc602020-05-13 22:29:36 -07004566xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07004567 name = "operator_run_test_mode",
4568 srcs = ["src/operator-run.c"],
4569 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
4570 copts = LOGGING_COPTS + [
4571 "-UNDEBUG",
4572 "-DXNN_TEST_MODE=1",
4573 ] + select({
4574 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4575 "//conditions:default": [],
4576 }),
4577 gcc_copts = xnnpack_gcc_std_copts(),
4578 msvc_copts = xnnpack_msvc_std_copts(),
4579 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07004580 ":logging_utils",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004581 "@FP16",
4582 "@FXdiv",
4583 "@clog",
4584 "@pthreadpool",
4585 ],
4586)
4587
4588xnnpack_cc_library(
Chao Mei6ddfc602020-05-13 22:29:36 -07004589 name = "memory_planner",
4590 srcs = ["src/memory-planner.c"],
4591 hdrs = INTERNAL_HDRS,
4592 defines = select({
4593 ":xnn_enable_memopt_explicit_true": ["XNN_ENABLE_MEMOPT=1"],
4594 ":xnn_enable_memopt_explicit_false": ["XNN_ENABLE_MEMOPT=0"],
4595 "//conditions:default": ["XNN_ENABLE_MEMOPT=1"],
4596 }),
4597 gcc_copts = xnnpack_gcc_std_copts(),
4598 msvc_copts = xnnpack_msvc_std_copts(),
4599 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07004600 ":logging_utils",
Chao Mei6ddfc602020-05-13 22:29:36 -07004601 "@pthreadpool",
4602 ],
4603)
4604
Marat Dukhan33fcf782020-05-24 14:27:15 -07004605xnnpack_cc_library(
4606 name = "memory_planner_test_mode",
4607 srcs = ["src/memory-planner.c"],
4608 hdrs = INTERNAL_HDRS,
4609 copts = [
4610 "-UNDEBUG",
4611 "-DXNN_TEST_MODE=1",
4612 ],
4613 defines = select({
4614 ":xnn_enable_memopt_explicit_true": ["XNN_ENABLE_MEMOPT=1"],
4615 ":xnn_enable_memopt_explicit_false": ["XNN_ENABLE_MEMOPT=0"],
4616 "//conditions:default": ["XNN_ENABLE_MEMOPT=1"],
4617 }),
4618 gcc_copts = xnnpack_gcc_std_copts(),
4619 msvc_copts = xnnpack_msvc_std_copts(),
4620 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07004621 ":logging_utils",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004622 "@pthreadpool",
4623 ],
4624)
4625
Marat Dukhan08c4a432019-10-03 09:29:21 -07004626cc_library(
4627 name = "enable_assembly",
4628 defines = select({
4629 ":xnn_enable_assembly_explicit_true": ["XNN_ENABLE_ASSEMBLY=1"],
4630 ":xnn_enable_assembly_explicit_false": ["XNN_ENABLE_ASSEMBLY=0"],
Frank Barchard810171d2019-10-10 10:34:51 -07004631 "//conditions:default": ["XNN_ENABLE_ASSEMBLY=1"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004632 }),
4633)
4634
Marat Dukhan9de90e02020-06-18 16:04:12 -07004635cc_library(
4636 name = "enable_sparse",
4637 defines = select({
4638 ":xnn_enable_sparse_explicit_true": ["XNN_ENABLE_SPARSE=1"],
4639 ":xnn_enable_sparse_explicit_false": ["XNN_ENABLE_SPARSE=0"],
Marat Dukhanb36582b2020-12-08 11:16:28 -08004640 "//conditions:default": ["XNN_ENABLE_SPARSE=1"],
Marat Dukhan9de90e02020-06-18 16:04:12 -07004641 }),
4642)
4643
Marat Dukhancf056b22019-10-07 10:26:29 -07004644xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004645 name = "operators",
4646 srcs = OPERATOR_SRCS + [
Marat Dukhan04f03be2019-11-19 12:36:47 -08004647 "src/memory.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004648 "src/operator-delete.c",
Marat Dukhancf056b22019-10-07 10:26:29 -07004649 ],
4650 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004651 copts = LOGGING_COPTS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07004652 "-Isrc",
4653 "-Iinclude",
4654 ] + select({
4655 ":debug_build": [],
4656 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07004657 }) + select({
4658 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4659 "//conditions:default": [],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004660 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07004661 gcc_copts = xnnpack_gcc_std_copts(),
4662 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004663 deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07004664 ":indirection",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004665 ":logging_utils",
Marat Dukhanab582382020-07-06 13:32:08 -07004666 ":packing",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004667 "@FP16",
4668 "@FXdiv",
4669 "@clog",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004670 "@pthreadpool",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004671 ],
4672)
4673
Marat Dukhan10a38082020-04-17 03:58:35 -07004674xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07004675 name = "operators_test_mode",
4676 srcs = OPERATOR_SRCS + [
4677 "src/memory.c",
4678 "src/operator-delete.c",
4679 ],
4680 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
4681 copts = LOGGING_COPTS + [
4682 "-Isrc",
4683 "-Iinclude",
4684 "-UNDEBUG",
4685 "-DXNN_TEST_MODE=1",
4686 ] + select({
4687 ":debug_build": [],
4688 "//conditions:default": xnnpack_min_size_copts(),
4689 }) + select({
4690 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4691 "//conditions:default": [],
4692 }),
4693 gcc_copts = xnnpack_gcc_std_copts(),
4694 msvc_copts = xnnpack_msvc_std_copts(),
4695 deps = [
4696 ":indirection_test_mode",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004697 ":logging_utils",
Marat Dukhanab582382020-07-06 13:32:08 -07004698 ":packing_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004699 "@FP16",
4700 "@FXdiv",
4701 "@clog",
4702 "@pthreadpool",
4703 ],
4704)
4705
4706xnnpack_cc_library(
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004707 name = "XNNPACK",
4708 srcs = [
4709 "src/init.c",
Marat Dukhanccfdbd12020-02-03 14:27:45 -08004710 "src/runtime.c",
4711 "src/subgraph.c",
4712 "src/tensor.c",
Marat Dukhanf03da0d2020-06-10 16:00:20 -07004713 ] + SUBGRAPH_SRCS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004714 hdrs = ["include/xnnpack.h"],
4715 copts = LOGGING_COPTS + [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004716 "-Isrc",
4717 "-Iinclude",
4718 ] + select({
4719 ":debug_build": [],
4720 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07004721 }) + select({
4722 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4723 "//conditions:default": [],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004724 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07004725 gcc_copts = xnnpack_gcc_std_copts(),
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004726 includes = ["include"],
Marat Dukhan10a38082020-04-17 03:58:35 -07004727 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004728 visibility = xnnpack_visibility(),
4729 deps = [
4730 ":enable_assembly",
Marat Dukhan9de90e02020-06-18 16:04:12 -07004731 ":enable_sparse",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004732 ":logging_utils",
Chao Mei6ddfc602020-05-13 22:29:36 -07004733 ":memory_planner",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004734 ":operator_run",
4735 ":operators",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004736 ":ukernels",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004737 "@clog",
Marat Dukhanab2946c2020-05-21 20:04:13 -07004738 "@FP16",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004739 "@pthreadpool",
Marat Dukhand343c222019-10-07 09:22:14 -07004740 ] + select({
4741 ":emscripten": [],
4742 "//conditions:default": ["@cpuinfo"],
4743 }),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004744)
4745
Marat Dukhan10a38082020-04-17 03:58:35 -07004746xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07004747 name = "XNNPACK_test_mode",
4748 srcs = [
4749 "src/init.c",
4750 "src/runtime.c",
4751 "src/subgraph.c",
4752 "src/tensor.c",
Marat Dukhanf03da0d2020-06-10 16:00:20 -07004753 ] + SUBGRAPH_SRCS,
Marat Dukhan33fcf782020-05-24 14:27:15 -07004754 hdrs = ["include/xnnpack.h"],
4755 copts = LOGGING_COPTS + [
4756 "-Isrc",
4757 "-Iinclude",
4758 "-UNDEBUG",
4759 "-DXNN_TEST_MODE=1",
4760 ] + select({
4761 ":debug_build": [],
4762 "//conditions:default": xnnpack_min_size_copts(),
4763 }) + select({
4764 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4765 "//conditions:default": [],
4766 }),
4767 gcc_copts = xnnpack_gcc_std_copts(),
4768 includes = ["include"],
4769 msvc_copts = xnnpack_msvc_std_copts(),
4770 visibility = xnnpack_visibility(),
4771 deps = [
4772 ":enable_assembly",
Marat Dukhan9de90e02020-06-18 16:04:12 -07004773 ":enable_sparse",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004774 ":logging_utils",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004775 ":memory_planner_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004776 ":operator_run_test_mode",
4777 ":operators_test_mode",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004778 ":ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004779 "@clog",
4780 "@FP16",
4781 "@pthreadpool",
4782 ] + select({
4783 ":emscripten": [],
4784 "//conditions:default": ["@cpuinfo"],
4785 }),
4786)
4787
4788xnnpack_cc_library(
Marat Dukhanae046f52020-06-15 13:16:14 -07004789 name = "xnnpack_f32",
4790 srcs = [
4791 "src/init.c",
4792 "src/runtime.c",
4793 "src/subgraph.c",
4794 "src/tensor.c",
4795 ] + SUBGRAPH_SRCS,
4796 hdrs = ["include/xnnpack.h"],
4797 copts = LOGGING_COPTS + [
4798 "-Isrc",
4799 "-Iinclude",
4800 ] + select({
4801 ":debug_build": [],
4802 "//conditions:default": xnnpack_min_size_copts(),
4803 }) + select({
4804 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4805 "//conditions:default": [],
4806 }),
4807 defines = [
Marat Dukhan16f1e1a2020-08-04 16:38:22 -07004808 "XNN_NO_QS8_OPERATORS",
Marat Dukhan08b7a972020-07-14 18:17:29 -07004809 "XNN_NO_QU8_OPERATORS",
Marat Dukhanae046f52020-06-15 13:16:14 -07004810 "XNN_NO_U8_OPERATORS",
4811 "XNN_NO_X8_OPERATORS",
4812 "XNN_NO_F16_OPERATORS",
4813 "XNN_NO_X16_OPERATORS",
4814 ],
4815 gcc_copts = xnnpack_gcc_std_copts(),
4816 includes = ["include"],
4817 msvc_copts = xnnpack_msvc_std_copts(),
4818 visibility = xnnpack_visibility(),
4819 deps = [
4820 ":enable_assembly",
Marat Dukhan9de90e02020-06-18 16:04:12 -07004821 ":enable_sparse",
Marat Dukhanae046f52020-06-15 13:16:14 -07004822 ":logging_utils",
4823 ":memory_planner",
4824 ":operator_run",
4825 ":operators",
Marat Dukhan86df3842020-06-17 14:03:31 -07004826 ":ukernels_max_compatibility",
Marat Dukhanae046f52020-06-15 13:16:14 -07004827 "@clog",
4828 "@FP16",
4829 "@pthreadpool",
4830 ] + select({
4831 ":emscripten": [],
4832 "//conditions:default": ["@cpuinfo"],
4833 }),
4834)
4835
4836xnnpack_cc_library(
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004837 name = "xnnpack_operators_nhwc_f32",
4838 srcs = [
4839 "src/init.c",
4840 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07004841 hdrs = ["include/xnnpack.h"],
4842 copts = LOGGING_COPTS + [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004843 "-Isrc",
4844 "-Iinclude",
4845 ] + select({
4846 ":debug_build": [],
4847 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07004848 }) + select({
4849 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4850 "//conditions:default": [],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004851 }),
4852 defines = [
Marat Dukhan16f1e1a2020-08-04 16:38:22 -07004853 "XNN_NO_QS8_OPERATORS",
Marat Dukhan08b7a972020-07-14 18:17:29 -07004854 "XNN_NO_QU8_OPERATORS",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004855 "XNN_NO_U8_OPERATORS",
4856 "XNN_NO_X8_OPERATORS",
Marat Dukhanefc47b82019-11-18 09:25:38 -08004857 "XNN_NO_NCHW_OPERATORS",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004858 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07004859 gcc_copts = xnnpack_gcc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004860 includes = ["include"],
Marat Dukhan10a38082020-04-17 03:58:35 -07004861 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004862 visibility = xnnpack_visibility(),
4863 deps = [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004864 ":enable_assembly",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004865 ":logging_utils",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004866 ":operator_run",
4867 ":operators",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004868 ":ukernels",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004869 "@clog",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004870 "@pthreadpool",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004871 ] + select({
4872 ":emscripten": [],
4873 "//conditions:default": ["@cpuinfo"],
4874 }),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004875)
4876
Marat Dukhancf056b22019-10-07 10:26:29 -07004877xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004878 name = "bench_utils",
4879 srcs = ["bench/utils.cc"],
4880 hdrs = ["bench/utils.h"],
Marat Dukhanbad48fe2019-11-04 10:35:22 -08004881 deps = [
4882 "@com_google_benchmark//:benchmark",
4883 "@cpuinfo",
4884 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004885)
4886
Frank Barchard7e955972019-10-11 10:34:25 -07004887######################### Benchmarks for micro-kernels #########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07004888
4889xnnpack_benchmark(
Marat Dukhan595e1702020-07-31 10:12:52 -07004890 name = "qs8_gemm_bench",
4891 srcs = [
4892 "bench/gemm.h",
4893 "bench/qs8-gemm.cc",
4894 "src/xnnpack/AlignedAllocator.h",
4895 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Frank Barchard31328cb2020-10-12 11:55:18 -07004896 copts = xnnpack_optional_ruy_copts() + xnnpack_optional_gemmlowp_copts(),
4897 deps = MICROKERNEL_BENCHMARK_DEPS + [":packing"] + xnnpack_optional_ruy_deps() + xnnpack_optional_gemmlowp_deps(),
Marat Dukhan595e1702020-07-31 10:12:52 -07004898)
4899
4900xnnpack_benchmark(
Marat Dukhan56bdd4a2020-08-03 19:47:04 -07004901 name = "qs8_requantization_bench",
4902 srcs = [
4903 "bench/qs8-requantization.cc",
Marat Dukhan56bdd4a2020-08-03 19:47:04 -07004904 "src/xnnpack/AlignedAllocator.h",
Frank Barchard04336c12020-10-22 16:48:55 -07004905 "src/xnnpack/requantization-stubs.h",
Marat Dukhan56bdd4a2020-08-03 19:47:04 -07004906 ] + MICROKERNEL_BENCHMARK_HDRS,
4907 deps = MICROKERNEL_BENCHMARK_DEPS,
4908)
4909
4910xnnpack_benchmark(
Marat Dukhan08b7a972020-07-14 18:17:29 -07004911 name = "qu8_gemm_bench",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004912 srcs = [
4913 "bench/gemm.h",
Marat Dukhan08b7a972020-07-14 18:17:29 -07004914 "bench/qu8-gemm.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004915 "src/xnnpack/AlignedAllocator.h",
4916 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004917 copts = xnnpack_optional_ruy_copts() + xnnpack_optional_gemmlowp_copts(),
Marat Dukhanab582382020-07-06 13:32:08 -07004918 deps = MICROKERNEL_BENCHMARK_DEPS + [":packing"] + xnnpack_optional_ruy_deps() + xnnpack_optional_gemmlowp_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004919)
4920
4921xnnpack_benchmark(
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07004922 name = "qu8_requantization_bench",
4923 srcs = [
4924 "bench/qu8-requantization.cc",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07004925 "src/xnnpack/AlignedAllocator.h",
Frank Barchard04336c12020-10-22 16:48:55 -07004926 "src/xnnpack/requantization-stubs.h",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07004927 ] + MICROKERNEL_BENCHMARK_HDRS,
4928 deps = MICROKERNEL_BENCHMARK_DEPS,
4929)
4930
4931xnnpack_benchmark(
Frank Barchard40d20fe2020-05-05 00:37:45 -07004932 name = "f16_igemm_bench",
4933 srcs = [
4934 "bench/f16-igemm.cc",
4935 "bench/conv.h",
4936 "bench/google/conv.h",
4937 "src/xnnpack/AlignedAllocator.h",
4938 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004939 deps = MICROKERNEL_BENCHMARK_DEPS + [
4940 ":indirection",
4941 ":packing",
4942 ],
Frank Barchard40d20fe2020-05-05 00:37:45 -07004943)
4944
4945xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004946 name = "f16_gemm_bench",
4947 srcs = [
4948 "bench/f16-gemm.cc",
4949 "bench/gemm.h",
4950 "src/xnnpack/AlignedAllocator.h",
4951 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004952 deps = MICROKERNEL_BENCHMARK_DEPS + [
4953 ":packing",
4954 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004955)
4956
4957xnnpack_benchmark(
Marat Dukhanbdb56f52020-02-05 21:42:49 -08004958 name = "f16_spmm_bench",
4959 srcs = [
4960 "bench/f16-spmm.cc",
Marat Dukhan1631e3e2020-12-06 19:29:31 -08004961 "bench/spmm.h",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08004962 "src/xnnpack/AlignedAllocator.h",
4963 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanbdb56f52020-02-05 21:42:49 -08004964 deps = MICROKERNEL_BENCHMARK_DEPS,
4965)
4966
4967xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004968 name = "f32_igemm_bench",
4969 srcs = [
4970 "bench/f32-igemm.cc",
4971 "bench/conv.h",
4972 "src/xnnpack/AlignedAllocator.h",
4973 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004974 deps = MICROKERNEL_BENCHMARK_DEPS + [
4975 ":indirection",
4976 ":packing",
4977 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004978)
4979
4980xnnpack_benchmark(
Frank Barchardc2010062020-07-19 10:04:12 -07004981 name = "f16_relu_bench",
4982 srcs = [
4983 "bench/f16-relu.cc",
4984 "src/xnnpack/AlignedAllocator.h",
4985 ] + MICROKERNEL_BENCHMARK_HDRS,
4986 deps = MICROKERNEL_BENCHMARK_DEPS,
4987)
4988
4989xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004990 name = "f32_conv_hwc_bench",
4991 srcs = [
4992 "bench/f32-conv-hwc.cc",
4993 "bench/dconv.h",
4994 "src/xnnpack/AlignedAllocator.h",
4995 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004996 deps = MICROKERNEL_BENCHMARK_DEPS + [
4997 ":packing",
4998 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004999)
5000
5001xnnpack_benchmark(
Marat Dukhan1f29b802020-05-15 23:46:39 -07005002 name = "f32_conv_hwc2chw_bench",
Erich Elsen563df5f2019-10-23 08:02:21 -07005003 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07005004 "bench/f32-conv-hwc2chw.cc",
Erich Elsen563df5f2019-10-23 08:02:21 -07005005 "bench/dconv.h",
5006 "src/xnnpack/AlignedAllocator.h",
5007 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005008 deps = MICROKERNEL_BENCHMARK_DEPS + [
5009 ":packing",
5010 ],
Erich Elsen563df5f2019-10-23 08:02:21 -07005011)
5012
5013xnnpack_benchmark(
Frank Barchard5a599a62020-06-04 20:12:44 -07005014 name = "f16_dwconv_bench",
5015 srcs = [
5016 "bench/f16-dwconv.cc",
5017 "bench/dwconv.h",
5018 "bench/google/dwconv.h",
5019 "src/xnnpack/AlignedAllocator.h",
5020 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005021 deps = MICROKERNEL_BENCHMARK_DEPS + [
5022 ":indirection",
5023 ":packing",
5024 ],
Frank Barchard5a599a62020-06-04 20:12:44 -07005025)
5026
5027xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005028 name = "f32_dwconv_bench",
5029 srcs = [
5030 "bench/f32-dwconv.cc",
5031 "bench/dwconv.h",
5032 "src/xnnpack/AlignedAllocator.h",
5033 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005034 deps = MICROKERNEL_BENCHMARK_DEPS + [
5035 ":indirection",
5036 ":packing",
5037 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005038)
5039
5040xnnpack_benchmark(
Marat Dukhanbf715f92020-10-23 20:17:00 -07005041 name = "f32_dwconv2d_chw_bench",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005042 srcs = [
Marat Dukhanbf715f92020-10-23 20:17:00 -07005043 "bench/f32-dwconv2d-chw.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005044 "bench/dwconv.h",
5045 "src/xnnpack/AlignedAllocator.h",
5046 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005047 deps = MICROKERNEL_BENCHMARK_DEPS + [
5048 ":indirection",
5049 ":packing",
5050 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005051)
5052
5053xnnpack_benchmark(
5054 name = "f32_gemm_bench",
5055 srcs = [
5056 "bench/f32-gemm.cc",
5057 "bench/gemm.h",
5058 "src/xnnpack/AlignedAllocator.h",
5059 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07005060 copts = xnnpack_optional_ruy_copts(),
Marat Dukhanab582382020-07-06 13:32:08 -07005061 deps = MICROKERNEL_BENCHMARK_DEPS + [":packing"] + xnnpack_optional_ruy_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07005062)
5063
5064xnnpack_benchmark(
Marat Dukhana11ca342020-06-25 23:45:07 -07005065 name = "f32_hswish_bench",
5066 srcs = [
5067 "bench/f32-hswish.cc",
5068 "src/xnnpack/AlignedAllocator.h",
5069 ] + MICROKERNEL_BENCHMARK_HDRS,
5070 deps = MICROKERNEL_BENCHMARK_DEPS,
5071)
5072
5073xnnpack_benchmark(
Marat Dukhan4c4eb002019-12-08 21:27:49 -08005074 name = "f32_raddexpminusmax_bench",
5075 srcs = [
5076 "bench/f32-raddexpminusmax.cc",
5077 "src/xnnpack/AlignedAllocator.h",
5078 ] + MICROKERNEL_BENCHMARK_HDRS,
5079 deps = MICROKERNEL_BENCHMARK_DEPS,
5080)
5081
5082xnnpack_benchmark(
5083 name = "f32_raddextexp_bench",
5084 srcs = [
5085 "bench/f32-raddextexp.cc",
5086 "src/xnnpack/AlignedAllocator.h",
5087 ] + MICROKERNEL_BENCHMARK_HDRS,
5088 deps = MICROKERNEL_BENCHMARK_DEPS,
5089)
5090
5091xnnpack_benchmark(
5092 name = "f32_raddstoreexpminusmax_bench",
5093 srcs = [
5094 "bench/f32-raddstoreexpminusmax.cc",
5095 "src/xnnpack/AlignedAllocator.h",
5096 ] + MICROKERNEL_BENCHMARK_HDRS,
5097 deps = MICROKERNEL_BENCHMARK_DEPS,
5098)
5099
5100xnnpack_benchmark(
Frank Barchard3c4a9522020-07-18 16:11:22 -07005101 name = "f32_relu_bench",
5102 srcs = [
5103 "bench/f32-relu.cc",
5104 "src/xnnpack/AlignedAllocator.h",
5105 ] + MICROKERNEL_BENCHMARK_HDRS,
5106 deps = MICROKERNEL_BENCHMARK_DEPS,
5107)
5108
5109xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005110 name = "f32_rmax_bench",
5111 srcs = [
5112 "bench/f32-rmax.cc",
5113 "src/xnnpack/AlignedAllocator.h",
5114 ] + MICROKERNEL_BENCHMARK_HDRS,
5115 deps = MICROKERNEL_BENCHMARK_DEPS,
5116)
5117
5118xnnpack_benchmark(
Marat Dukhan14bec502019-11-18 11:35:31 -08005119 name = "f32_sigmoid_bench",
5120 srcs = [
5121 "bench/f32-sigmoid.cc",
5122 "src/xnnpack/AlignedAllocator.h",
5123 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan14bec502019-11-18 11:35:31 -08005124 deps = MICROKERNEL_BENCHMARK_DEPS,
5125)
5126
5127xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005128 name = "f32_spmm_bench",
5129 srcs = [
5130 "bench/f32-spmm.cc",
Marat Dukhan1631e3e2020-12-06 19:29:31 -08005131 "bench/spmm.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005132 "src/xnnpack/AlignedAllocator.h",
5133 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005134 deps = MICROKERNEL_BENCHMARK_DEPS,
5135)
5136
5137xnnpack_benchmark(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08005138 name = "f32_softmax_bench",
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07005139 srcs = [
Marat Dukhanfd8e6892020-01-27 15:25:25 -08005140 "bench/f32-softmax.cc",
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07005141 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07005142 copts = xnnpack_optional_dnnl_copts(),
Marat Dukhan8d3c6932020-03-06 20:27:27 -08005143 deps = MICROKERNEL_BENCHMARK_DEPS + xnnpack_optional_dnnl_deps(),
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07005144)
5145
5146xnnpack_benchmark(
Marat Dukhaned6baaf2020-12-01 15:07:08 -08005147 name = "f32_velu_bench",
5148 srcs = [
5149 "bench/f32-velu.cc",
5150 "src/xnnpack/AlignedAllocator.h",
5151 ] + MICROKERNEL_BENCHMARK_HDRS,
5152 deps = MICROKERNEL_BENCHMARK_DEPS,
5153)
5154
5155xnnpack_benchmark(
Marat Dukhan4c4eb002019-12-08 21:27:49 -08005156 name = "f32_vscaleexpminusmax_bench",
5157 srcs = [
5158 "bench/f32-vscaleexpminusmax.cc",
5159 "src/xnnpack/AlignedAllocator.h",
5160 ] + MICROKERNEL_BENCHMARK_HDRS,
5161 deps = MICROKERNEL_BENCHMARK_DEPS,
5162)
5163
5164xnnpack_benchmark(
5165 name = "f32_vscaleextexp_bench",
5166 srcs = [
5167 "bench/f32-vscaleextexp.cc",
5168 "src/xnnpack/AlignedAllocator.h",
5169 ] + MICROKERNEL_BENCHMARK_HDRS,
5170 deps = MICROKERNEL_BENCHMARK_DEPS,
5171)
5172
5173xnnpack_benchmark(
Marat Dukhanf4db2f32020-06-30 10:55:30 -07005174 name = "f32_vsqrt_bench",
5175 srcs = [
5176 "bench/f32-vsqrt.cc",
5177 "src/xnnpack/AlignedAllocator.h",
5178 ] + MICROKERNEL_BENCHMARK_HDRS,
5179 deps = MICROKERNEL_BENCHMARK_DEPS,
5180)
5181
5182xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005183 name = "f32_im2col_gemm_bench",
5184 srcs = [
5185 "bench/f32-im2col-gemm.cc",
5186 "bench/conv.h",
5187 "src/xnnpack/AlignedAllocator.h",
5188 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005189 deps = MICROKERNEL_BENCHMARK_DEPS + [
5190 ":im2col",
5191 ":packing",
5192 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005193)
5194
Marat Dukhanfe7acb62020-03-09 19:30:05 -07005195xnnpack_benchmark(
Marat Dukhanffbf96a2020-05-14 02:59:08 -07005196 name = "rounding_bench",
5197 srcs = [
5198 "bench/rounding.cc",
Marat Dukhanffbf96a2020-05-14 02:59:08 -07005199 "src/xnnpack/AlignedAllocator.h",
Frank Barchard04336c12020-10-22 16:48:55 -07005200 "src/xnnpack/math-stubs.h",
Marat Dukhanffbf96a2020-05-14 02:59:08 -07005201 ] + MICROKERNEL_BENCHMARK_HDRS,
5202 deps = MICROKERNEL_BENCHMARK_DEPS,
5203)
5204
Marat Dukhan08c4a432019-10-03 09:29:21 -07005205########################### Benchmarks for operators ###########################
5206
5207xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005208 name = "average_pooling_bench",
5209 srcs = ["bench/average-pooling.cc"],
Marat Dukhan7a16d8b2020-03-11 04:22:44 -07005210 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07005211 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07005212 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07005213)
5214
5215xnnpack_benchmark(
Marat Dukhan87727142020-06-24 15:24:10 -07005216 name = "bankers_rounding_bench",
5217 srcs = ["bench/bankers-rounding.cc"],
5218 copts = xnnpack_optional_tflite_copts(),
5219 tags = ["nowin32"],
5220 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
5221)
5222
5223xnnpack_benchmark(
5224 name = "ceiling_bench",
5225 srcs = ["bench/ceiling.cc"],
5226 copts = xnnpack_optional_tflite_copts(),
5227 tags = ["nowin32"],
5228 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
5229)
5230
5231xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005232 name = "channel_shuffle_bench",
5233 srcs = ["bench/channel-shuffle.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07005234 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005235)
5236
5237xnnpack_benchmark(
5238 name = "convolution_bench",
5239 srcs = ["bench/convolution.cc"],
5240 copts = xnnpack_optional_tflite_copts() + xnnpack_optional_armcl_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07005241 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07005242 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps() + xnnpack_optional_armcl_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07005243)
5244
5245xnnpack_benchmark(
5246 name = "deconvolution_bench",
5247 srcs = ["bench/deconvolution.cc"],
5248 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07005249 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07005250 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07005251)
5252
5253xnnpack_benchmark(
Marat Dukhanb6bd4bc2020-12-01 17:01:40 -08005254 name = "elu_bench",
5255 srcs = ["bench/elu.cc"],
5256 copts = xnnpack_optional_tflite_copts(),
5257 tags = ["nowin32"],
5258 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
5259)
5260
5261xnnpack_benchmark(
Marat Dukhan87727142020-06-24 15:24:10 -07005262 name = "floor_bench",
5263 srcs = ["bench/floor.cc"],
5264 copts = xnnpack_optional_tflite_copts(),
5265 tags = ["nowin32"],
5266 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
5267)
5268
5269xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005270 name = "global_average_pooling_bench",
5271 srcs = ["bench/global-average-pooling.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07005272 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005273)
5274
5275xnnpack_benchmark(
Marat Dukhanad352602020-06-25 21:50:54 -07005276 name = "hardswish_bench",
5277 srcs = ["bench/hardswish.cc"],
5278 copts = xnnpack_optional_tflite_copts(),
5279 tags = ["nowin32"],
5280 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
5281)
5282
5283xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005284 name = "max_pooling_bench",
5285 srcs = ["bench/max-pooling.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07005286 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005287)
5288
5289xnnpack_benchmark(
5290 name = "sigmoid_bench",
5291 srcs = ["bench/sigmoid.cc"],
Marat Dukhanc3b9e862019-11-17 13:18:54 -08005292 copts = xnnpack_optional_tflite_copts(),
Marat Dukhanca2ba702020-04-24 01:31:47 -07005293 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07005294 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07005295)
5296
5297xnnpack_benchmark(
Marat Dukhan95b22432019-10-30 16:30:14 -07005298 name = "prelu_bench",
5299 srcs = ["bench/prelu.cc"],
5300 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07005301 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07005302 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan95b22432019-10-30 16:30:14 -07005303)
5304
5305xnnpack_benchmark(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08005306 name = "softmax_bench",
5307 srcs = ["bench/softmax.cc"],
Marat Dukhan9c0db962020-01-28 12:30:14 -08005308 copts = xnnpack_optional_tflite_copts(),
Marat Dukhanca2ba702020-04-24 01:31:47 -07005309 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07005310 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07005311)
5312
Marat Dukhan87727142020-06-24 15:24:10 -07005313xnnpack_benchmark(
Marat Dukhan6804bbd2020-06-30 19:26:11 -07005314 name = "square_root_bench",
5315 srcs = ["bench/square-root.cc"],
5316 copts = xnnpack_optional_tflite_copts(),
5317 tags = ["nowin32"],
5318 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
5319)
5320
5321xnnpack_benchmark(
Marat Dukhan87727142020-06-24 15:24:10 -07005322 name = "truncation_bench",
5323 srcs = ["bench/truncation.cc"],
5324 deps = OPERATOR_BENCHMARK_DEPS,
5325)
5326
Marat Dukhanc068bb62019-10-04 13:24:39 -07005327############################# End-to-end benchmarks ############################
5328
5329cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07005330 name = "fp32_mobilenet_v1",
5331 srcs = ["models/fp32-mobilenet-v1.cc"],
Marat Dukhanc068bb62019-10-04 13:24:39 -07005332 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08005333 copts = xnnpack_std_cxxopts(),
Marat Dukhanc068bb62019-10-04 13:24:39 -07005334 linkstatic = True,
5335 deps = [
5336 ":XNNPACK",
5337 "@pthreadpool",
5338 ],
5339)
5340
5341cc_library(
Marat Dukhan4cea2322021-03-09 09:35:36 -08005342 name = "fp32_sparse_mobilenet_v1",
5343 srcs = ["models/fp32-sparse-mobilenet-v1.cc"],
5344 hdrs = ["models/models.h"],
5345 copts = xnnpack_std_cxxopts(),
5346 linkstatic = True,
5347 deps = [
5348 ":XNNPACK",
5349 "@pthreadpool",
5350 ],
5351)
5352
5353cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07005354 name = "fp16_mobilenet_v1",
5355 srcs = ["models/fp16-mobilenet-v1.cc"],
5356 hdrs = ["models/models.h"],
5357 copts = xnnpack_std_cxxopts(),
5358 linkstatic = True,
5359 deps = [
5360 ":XNNPACK",
5361 "@FP16",
5362 "@pthreadpool",
5363 ],
5364)
5365
5366cc_library(
Marat Dukhan0743cdf2020-08-04 18:52:07 -07005367 name = "qs8_mobilenet_v1",
5368 srcs = ["models/qs8-mobilenet-v1.cc"],
5369 hdrs = ["models/models.h"],
5370 copts = xnnpack_std_cxxopts(),
5371 linkstatic = True,
5372 deps = [
5373 ":XNNPACK",
5374 "@pthreadpool",
5375 ],
5376)
5377
5378cc_library(
Marat Dukhan70a96182020-09-03 17:13:58 -07005379 name = "qs8_mobilenet_v2",
5380 srcs = ["models/qs8-mobilenet-v2.cc"],
5381 hdrs = ["models/models.h"],
5382 copts = xnnpack_std_cxxopts(),
5383 linkstatic = True,
5384 deps = [
5385 ":XNNPACK",
5386 "@pthreadpool",
5387 ],
5388)
5389
5390cc_library(
Marat Dukhan12a23bb2021-03-08 08:13:21 -08005391 name = "qu8_mobilenet_v1",
5392 srcs = ["models/qu8-mobilenet-v1.cc"],
5393 hdrs = ["models/models.h"],
5394 copts = xnnpack_std_cxxopts(),
5395 linkstatic = True,
5396 deps = [
5397 ":XNNPACK",
5398 "@pthreadpool",
5399 ],
5400)
5401
5402cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07005403 name = "fp32_mobilenet_v2",
5404 srcs = ["models/fp32-mobilenet-v2.cc"],
Marat Dukhanc068bb62019-10-04 13:24:39 -07005405 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08005406 copts = xnnpack_std_cxxopts(),
Marat Dukhanc068bb62019-10-04 13:24:39 -07005407 linkstatic = True,
5408 deps = [
5409 ":XNNPACK",
5410 "@pthreadpool",
5411 ],
5412)
5413
Marat Dukhanc08cdf52019-12-09 09:17:51 -08005414cc_library(
Marat Dukhan4cea2322021-03-09 09:35:36 -08005415 name = "fp32_sparse_mobilenet_v2",
5416 srcs = ["models/fp32-sparse-mobilenet-v2.cc"],
5417 hdrs = ["models/models.h"],
5418 copts = xnnpack_std_cxxopts(),
5419 linkstatic = True,
5420 deps = [
5421 ":XNNPACK",
5422 "@pthreadpool",
5423 ],
5424)
5425
5426cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07005427 name = "fp16_mobilenet_v2",
5428 srcs = ["models/fp16-mobilenet-v2.cc"],
5429 hdrs = ["models/models.h"],
5430 copts = xnnpack_std_cxxopts(),
5431 linkstatic = True,
5432 deps = [
5433 ":XNNPACK",
5434 "@FP16",
5435 "@pthreadpool",
5436 ],
5437)
5438
5439cc_library(
5440 name = "fp32_mobilenet_v3_large",
5441 srcs = ["models/fp32-mobilenet-v3-large.cc"],
Marat Dukhanc08cdf52019-12-09 09:17:51 -08005442 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08005443 copts = xnnpack_std_cxxopts(),
Marat Dukhanc08cdf52019-12-09 09:17:51 -08005444 linkstatic = True,
5445 deps = [
5446 ":XNNPACK",
5447 "@pthreadpool",
5448 ],
5449)
5450
5451cc_library(
Marat Dukhan4cea2322021-03-09 09:35:36 -08005452 name = "fp32_sparse_mobilenet_v3_large",
5453 srcs = ["models/fp32-sparse-mobilenet-v3-large.cc"],
5454 hdrs = ["models/models.h"],
5455 copts = xnnpack_std_cxxopts(),
5456 linkstatic = True,
5457 deps = [
5458 ":XNNPACK",
5459 "@pthreadpool",
5460 ],
5461)
5462
5463cc_library(
Marat Dukhan66f3ccd2020-09-14 16:09:38 -07005464 name = "fp16_mobilenet_v3_large",
5465 srcs = ["models/fp16-mobilenet-v3-large.cc"],
5466 hdrs = ["models/models.h"],
5467 copts = xnnpack_std_cxxopts(),
5468 linkstatic = True,
5469 deps = [
5470 ":XNNPACK",
5471 "@FP16",
5472 "@pthreadpool",
5473 ],
5474)
5475
5476cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07005477 name = "fp32_mobilenet_v3_small",
5478 srcs = ["models/fp32-mobilenet-v3-small.cc"],
Marat Dukhanc08cdf52019-12-09 09:17:51 -08005479 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08005480 copts = xnnpack_std_cxxopts(),
Marat Dukhanc08cdf52019-12-09 09:17:51 -08005481 linkstatic = True,
5482 deps = [
5483 ":XNNPACK",
5484 "@pthreadpool",
5485 ],
5486)
5487
Marat Dukhan66f3ccd2020-09-14 16:09:38 -07005488cc_library(
Marat Dukhan4cea2322021-03-09 09:35:36 -08005489 name = "fp32_sparse_mobilenet_v3_small",
5490 srcs = ["models/fp32-sparse-mobilenet-v3-small.cc"],
5491 hdrs = ["models/models.h"],
5492 copts = xnnpack_std_cxxopts(),
5493 linkstatic = True,
5494 deps = [
5495 ":XNNPACK",
5496 "@pthreadpool",
5497 ],
5498)
5499
5500cc_library(
Marat Dukhan66f3ccd2020-09-14 16:09:38 -07005501 name = "fp16_mobilenet_v3_small",
5502 srcs = ["models/fp16-mobilenet-v3-small.cc"],
5503 hdrs = ["models/models.h"],
5504 copts = xnnpack_std_cxxopts(),
5505 linkstatic = True,
5506 deps = [
5507 ":XNNPACK",
5508 "@FP16",
5509 "@pthreadpool",
5510 ],
5511)
5512
Marat Dukhanc068bb62019-10-04 13:24:39 -07005513xnnpack_benchmark(
Marat Dukhanef4416e2019-10-31 13:44:40 -07005514 name = "f32_dwconv_e2e_bench",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08005515 srcs = [
5516 "bench/f32-dwconv-e2e.cc",
5517 "bench/end2end.h",
5518 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanef4416e2019-10-31 13:44:40 -07005519 deps = MICROKERNEL_BENCHMARK_DEPS + [
5520 ":XNNPACK",
Marat Dukhan270a2c42020-06-26 16:45:52 -07005521 ":fp32_mobilenet_v1",
5522 ":fp32_mobilenet_v2",
5523 ":fp32_mobilenet_v3_large",
5524 ":fp32_mobilenet_v3_small",
Marat Dukhanef4416e2019-10-31 13:44:40 -07005525 ],
5526)
5527
5528xnnpack_benchmark(
Marat Dukhan5f18d262019-10-31 10:24:14 -07005529 name = "f32_gemm_e2e_bench",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08005530 srcs = [
5531 "bench/f32-gemm-e2e.cc",
5532 "bench/end2end.h",
5533 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan5f18d262019-10-31 10:24:14 -07005534 deps = MICROKERNEL_BENCHMARK_DEPS + [
5535 ":XNNPACK",
Marat Dukhan270a2c42020-06-26 16:45:52 -07005536 ":fp32_mobilenet_v1",
5537 ":fp32_mobilenet_v2",
5538 ":fp32_mobilenet_v3_large",
5539 ":fp32_mobilenet_v3_small",
Marat Dukhan5f18d262019-10-31 10:24:14 -07005540 ],
5541)
5542
5543xnnpack_benchmark(
Frank Barcharddc909cb2021-02-08 13:59:31 -08005544 name = "qs8_gemm_e2e_bench",
5545 srcs = [
5546 "bench/qs8-gemm-e2e.cc",
5547 "bench/end2end.h",
5548 ] + MICROKERNEL_BENCHMARK_HDRS,
5549 deps = MICROKERNEL_BENCHMARK_DEPS + [
5550 ":XNNPACK",
5551 ":qs8_mobilenet_v1",
5552 ":qs8_mobilenet_v2",
5553 ],
5554)
5555
5556xnnpack_benchmark(
Marat Dukhanc068bb62019-10-04 13:24:39 -07005557 name = "end2end_bench",
5558 srcs = ["bench/end2end.cc"],
5559 deps = [
5560 ":XNNPACK",
Frank Barchardc712fa42019-10-31 14:00:21 -07005561 ":bench_utils",
Marat Dukhan270a2c42020-06-26 16:45:52 -07005562 ":fp16_mobilenet_v1",
5563 ":fp16_mobilenet_v2",
Marat Dukhan66f3ccd2020-09-14 16:09:38 -07005564 ":fp16_mobilenet_v3_large",
5565 ":fp16_mobilenet_v3_small",
Marat Dukhan270a2c42020-06-26 16:45:52 -07005566 ":fp32_mobilenet_v1",
5567 ":fp32_mobilenet_v2",
5568 ":fp32_mobilenet_v3_large",
5569 ":fp32_mobilenet_v3_small",
Marat Dukhan4cea2322021-03-09 09:35:36 -08005570 ":fp32_sparse_mobilenet_v1",
5571 ":fp32_sparse_mobilenet_v2",
5572 ":fp32_sparse_mobilenet_v3_large",
5573 ":fp32_sparse_mobilenet_v3_small",
Marat Dukhan0743cdf2020-08-04 18:52:07 -07005574 ":qs8_mobilenet_v1",
Marat Dukhan70a96182020-09-03 17:13:58 -07005575 ":qs8_mobilenet_v2",
Marat Dukhan12a23bb2021-03-08 08:13:21 -08005576 ":qu8_mobilenet_v1",
Marat Dukhanc068bb62019-10-04 13:24:39 -07005577 "@pthreadpool",
5578 ],
5579)
5580
Marat Dukhan6adff4e2019-10-14 18:32:07 -07005581#################### Accuracy evaluation for math functions ####################
5582
5583xnnpack_benchmark(
Marat Dukhancf67ec62020-12-14 09:56:37 -08005584 name = "f32_exp_ulp_eval",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07005585 srcs = [
Marat Dukhancf67ec62020-12-14 09:56:37 -08005586 "eval/f32-exp-ulp.cc",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07005587 "src/xnnpack/AlignedAllocator.h",
5588 ] + ACCURACY_EVAL_HDRS,
Marat Dukhan3a305212020-12-06 19:24:27 -08005589 deps = ACCURACY_EVAL_DEPS + [
5590 ":bench_utils",
5591 "@cpuinfo",
5592 ],
Marat Dukhan6adff4e2019-10-14 18:32:07 -07005593)
5594
Marat Dukhan515c9772019-10-17 18:07:57 -07005595xnnpack_benchmark(
Marat Dukhancf67ec62020-12-14 09:56:37 -08005596 name = "f32_expminus_ulp_eval",
Marat Dukhan515c9772019-10-17 18:07:57 -07005597 srcs = [
Marat Dukhancf67ec62020-12-14 09:56:37 -08005598 "eval/f32-expminus-ulp.cc",
Marat Dukhan515c9772019-10-17 18:07:57 -07005599 "src/xnnpack/AlignedAllocator.h",
5600 ] + ACCURACY_EVAL_HDRS,
Marat Dukhan3a305212020-12-06 19:24:27 -08005601 deps = ACCURACY_EVAL_DEPS + [
5602 ":bench_utils",
5603 "@cpuinfo",
5604 ],
Marat Dukhan515c9772019-10-17 18:07:57 -07005605)
5606
Marat Dukhan98ba4412019-10-23 02:14:28 -07005607xnnpack_benchmark(
Marat Dukhancf67ec62020-12-14 09:56:37 -08005608 name = "f32_expm1minus_ulp_eval",
Marat Dukhana438aca2020-11-20 15:45:01 -08005609 srcs = [
Marat Dukhancf67ec62020-12-14 09:56:37 -08005610 "eval/f32-expm1minus-ulp.cc",
Marat Dukhana438aca2020-11-20 15:45:01 -08005611 "src/xnnpack/AlignedAllocator.h",
5612 ] + ACCURACY_EVAL_HDRS,
Marat Dukhande390d42020-11-29 19:32:18 -08005613 deps = ACCURACY_EVAL_DEPS + [
5614 ":bench_utils",
Marat Dukhan3a305212020-12-06 19:24:27 -08005615 "@cpuinfo",
Marat Dukhande390d42020-11-29 19:32:18 -08005616 ],
Marat Dukhana438aca2020-11-20 15:45:01 -08005617)
5618
5619xnnpack_benchmark(
Marat Dukhancf67ec62020-12-14 09:56:37 -08005620 name = "f32_extexp_ulp_eval",
Marat Dukhan98ba4412019-10-23 02:14:28 -07005621 srcs = [
Marat Dukhancf67ec62020-12-14 09:56:37 -08005622 "eval/f32-extexp-ulp.cc",
Marat Dukhan98ba4412019-10-23 02:14:28 -07005623 "src/xnnpack/AlignedAllocator.h",
5624 ] + ACCURACY_EVAL_HDRS,
Marat Dukhan3a305212020-12-06 19:24:27 -08005625 deps = ACCURACY_EVAL_DEPS + [
5626 ":bench_utils",
5627 "@cpuinfo",
5628 ],
Marat Dukhan98ba4412019-10-23 02:14:28 -07005629)
5630
Marat Dukhanf44f0222020-12-14 11:53:27 -08005631xnnpack_benchmark(
5632 name = "f32_sigmoid_ulp_eval",
5633 srcs = [
5634 "eval/f32-sigmoid-ulp.cc",
5635 "src/xnnpack/AlignedAllocator.h",
5636 ] + ACCURACY_EVAL_HDRS,
5637 deps = ACCURACY_EVAL_DEPS + [
5638 ":bench_utils",
5639 "@cpuinfo",
5640 ],
5641)
5642
5643xnnpack_benchmark(
5644 name = "f32_sqrt_ulp_eval",
5645 srcs = [
5646 "eval/f32-sqrt-ulp.cc",
5647 "src/xnnpack/AlignedAllocator.h",
5648 ] + ACCURACY_EVAL_HDRS,
5649 deps = ACCURACY_EVAL_DEPS + [
5650 ":bench_utils",
5651 "@cpuinfo",
5652 ],
5653)
5654
5655################### Accuracy verification for math functions ##################
5656
5657xnnpack_unit_test(
Marat Dukhanf7291fc2020-12-15 11:02:50 -08005658 name = "f32_exp_eval",
5659 srcs = [
5660 "eval/f32-exp.cc",
5661 "src/xnnpack/AlignedAllocator.h",
5662 "src/xnnpack/math-stubs.h",
5663 ] + MICROKERNEL_TEST_HDRS,
5664 automatic = False,
5665 deps = MICROKERNEL_TEST_DEPS,
5666)
5667
5668xnnpack_unit_test(
Marat Dukhanf44f0222020-12-14 11:53:27 -08005669 name = "f32_expm1minus_eval",
5670 srcs = [
5671 "eval/f32-expm1minus.cc",
5672 "src/xnnpack/AlignedAllocator.h",
5673 "src/xnnpack/math-stubs.h",
5674 ] + MICROKERNEL_TEST_HDRS,
5675 automatic = False,
5676 deps = MICROKERNEL_TEST_DEPS,
5677)
5678
Marat Dukhan8853b822020-05-07 12:19:01 -07005679xnnpack_unit_test(
Marat Dukhand28a5a22020-12-14 15:27:22 -08005680 name = "f32_expminus_eval",
5681 srcs = [
5682 "eval/f32-expminus.cc",
5683 "src/xnnpack/AlignedAllocator.h",
5684 "src/xnnpack/math-stubs.h",
5685 ] + MICROKERNEL_TEST_HDRS,
5686 automatic = False,
5687 deps = MICROKERNEL_TEST_DEPS,
5688)
5689
5690xnnpack_unit_test(
Marat Dukhan8853b822020-05-07 12:19:01 -07005691 name = "f32_roundne_eval",
5692 srcs = [
5693 "eval/f32-roundne.cc",
5694 "src/xnnpack/AlignedAllocator.h",
5695 "src/xnnpack/math-stubs.h",
5696 ] + MICROKERNEL_TEST_HDRS,
Marat Dukhan22eed3d2020-05-11 20:13:37 -07005697 automatic = False,
Marat Dukhan8853b822020-05-07 12:19:01 -07005698 deps = MICROKERNEL_TEST_DEPS,
5699)
5700
Marat Dukhan2dbb9442020-05-12 20:43:43 -07005701xnnpack_unit_test(
Marat Dukhanc9852ba2020-05-13 17:21:29 -07005702 name = "f32_roundd_eval",
5703 srcs = [
5704 "eval/f32-roundd.cc",
5705 "src/xnnpack/AlignedAllocator.h",
5706 "src/xnnpack/math-stubs.h",
5707 ] + MICROKERNEL_TEST_HDRS,
5708 automatic = False,
5709 deps = MICROKERNEL_TEST_DEPS,
5710)
5711
5712xnnpack_unit_test(
5713 name = "f32_roundu_eval",
5714 srcs = [
5715 "eval/f32-roundu.cc",
5716 "src/xnnpack/AlignedAllocator.h",
5717 "src/xnnpack/math-stubs.h",
5718 ] + MICROKERNEL_TEST_HDRS,
5719 automatic = False,
5720 deps = MICROKERNEL_TEST_DEPS,
5721)
5722
5723xnnpack_unit_test(
Marat Dukhan2dbb9442020-05-12 20:43:43 -07005724 name = "f32_roundz_eval",
5725 srcs = [
5726 "eval/f32-roundz.cc",
5727 "src/xnnpack/AlignedAllocator.h",
5728 "src/xnnpack/math-stubs.h",
5729 ] + MICROKERNEL_TEST_HDRS,
Marat Dukhanc9852ba2020-05-13 17:21:29 -07005730 automatic = False,
Marat Dukhan2dbb9442020-05-12 20:43:43 -07005731 deps = MICROKERNEL_TEST_DEPS,
5732)
5733
Marat Dukhan08c4a432019-10-03 09:29:21 -07005734######################### Unit tests for micro-kernels #########################
5735
5736xnnpack_unit_test(
Marat Dukhande06f492020-04-09 00:19:31 -07005737 name = "f16_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005738 srcs = [
Marat Dukhande06f492020-04-09 00:19:31 -07005739 "test/f16-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005740 "test/gemm-microkernel-tester.h",
5741 "src/xnnpack/AlignedAllocator.h",
5742 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005743 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005744)
5745
5746xnnpack_unit_test(
Marat Dukhan355ab432020-04-09 19:01:52 -07005747 name = "f16_spmm_minmax_test",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08005748 srcs = [
Marat Dukhan355ab432020-04-09 19:01:52 -07005749 "test/f16-spmm-minmax.cc",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08005750 "test/spmm-microkernel-tester.h",
5751 "src/xnnpack/AlignedAllocator.h",
5752 ] + MICROKERNEL_TEST_HDRS,
5753 deps = MICROKERNEL_TEST_DEPS,
5754)
5755
5756xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005757 name = "f32_argmaxpool_test",
5758 srcs = [
5759 "test/f32-argmaxpool.cc",
5760 "test/argmaxpool-microkernel-tester.h",
5761 "src/xnnpack/AlignedAllocator.h",
5762 ] + MICROKERNEL_TEST_HDRS,
5763 deps = MICROKERNEL_TEST_DEPS,
5764)
5765
5766xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07005767 name = "f32_avgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005768 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07005769 "test/f32-avgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005770 "test/avgpool-microkernel-tester.h",
5771 "src/xnnpack/AlignedAllocator.h",
5772 ] + MICROKERNEL_TEST_HDRS,
5773 deps = MICROKERNEL_TEST_DEPS,
5774)
5775
5776xnnpack_unit_test(
Marat Dukhan660fd192020-03-10 04:55:30 -07005777 name = "f32_ibilinear_test",
Marat Dukhan35dacfb2019-11-07 19:18:16 -08005778 srcs = [
Marat Dukhan660fd192020-03-10 04:55:30 -07005779 "test/f32-ibilinear.cc",
5780 "test/ibilinear-microkernel-tester.h",
Marat Dukhan35dacfb2019-11-07 19:18:16 -08005781 "src/xnnpack/AlignedAllocator.h",
5782 ] + MICROKERNEL_TEST_HDRS,
5783 deps = MICROKERNEL_TEST_DEPS,
5784)
5785
5786xnnpack_unit_test(
XNNPACK Team21432672020-10-19 19:58:48 -07005787 name = "f32_ibilinear_chw_test",
5788 srcs = [
5789 "test/f32-ibilinear-chw.cc",
XNNPACK Team6be46b22020-10-22 23:34:54 -07005790 "test/ibilinear-microkernel-tester.h",
XNNPACK Team21432672020-10-19 19:58:48 -07005791 "src/xnnpack/AlignedAllocator.h",
5792 ] + MICROKERNEL_TEST_HDRS,
5793 deps = MICROKERNEL_TEST_DEPS,
5794)
5795
5796xnnpack_unit_test(
Frank Barchardb1966592020-05-12 13:47:06 -07005797 name = "f16_clamp_test",
5798 srcs = [
5799 "test/f16-clamp.cc",
5800 "test/clamp-microkernel-tester.h",
5801 ] + MICROKERNEL_TEST_HDRS,
5802 deps = MICROKERNEL_TEST_DEPS,
5803)
5804
5805xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005806 name = "f32_clamp_test",
5807 srcs = [
5808 "test/f32-clamp.cc",
5809 "test/clamp-microkernel-tester.h",
5810 ] + MICROKERNEL_TEST_HDRS,
5811 deps = MICROKERNEL_TEST_DEPS,
5812)
5813
5814xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07005815 name = "f32_igemm_test",
5816 srcs = [
5817 "test/f32-igemm.cc",
5818 "test/gemm-microkernel-tester.h",
5819 "src/xnnpack/AlignedAllocator.h",
5820 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005821 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan163a7e62020-04-09 04:19:26 -07005822)
5823
5824xnnpack_unit_test(
Marat Dukhan467f6362020-05-22 23:21:55 -07005825 name = "f32_igemm_relu_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005826 srcs = [
Marat Dukhan467f6362020-05-22 23:21:55 -07005827 "test/f32-igemm-relu.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005828 "test/gemm-microkernel-tester.h",
5829 "src/xnnpack/AlignedAllocator.h",
5830 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005831 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005832)
5833
5834xnnpack_unit_test(
Marat Dukhane207b7b2020-05-28 16:27:42 -07005835 name = "f32_igemm_minmax_test",
5836 srcs = [
5837 "test/f32-igemm-minmax.cc",
5838 "test/gemm-microkernel-tester.h",
5839 "src/xnnpack/AlignedAllocator.h",
5840 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005841 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhane207b7b2020-05-28 16:27:42 -07005842)
5843
5844xnnpack_unit_test(
Frank Barchardb0e4fae2020-05-04 15:27:51 -07005845 name = "f16_igemm_minmax_test",
5846 srcs = [
5847 "test/f16-igemm-minmax.cc",
5848 "test/gemm-microkernel-tester.h",
5849 "src/xnnpack/AlignedAllocator.h",
5850 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005851 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Frank Barchardb0e4fae2020-05-04 15:27:51 -07005852)
5853
5854xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005855 name = "f32_conv_hwc_test",
5856 srcs = [
5857 "test/f32-conv-hwc.cc",
5858 "test/conv-hwc-microkernel-tester.h",
5859 "src/xnnpack/AlignedAllocator.h",
5860 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005861 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005862)
5863
5864xnnpack_unit_test(
Marat Dukhan1f29b802020-05-15 23:46:39 -07005865 name = "f32_conv_hwc2chw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005866 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07005867 "test/f32-conv-hwc2chw.cc",
5868 "test/conv-hwc2chw-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005869 "src/xnnpack/AlignedAllocator.h",
5870 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005871 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005872)
5873
5874xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07005875 name = "f32_dwconv_test",
5876 srcs = [
5877 "test/f32-dwconv.cc",
5878 "test/dwconv-microkernel-tester.h",
5879 "src/xnnpack/AlignedAllocator.h",
5880 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005881 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan163a7e62020-04-09 04:19:26 -07005882)
5883
5884xnnpack_unit_test(
Frank Barchard5a599a62020-06-04 20:12:44 -07005885 name = "f16_dwconv_minmax_test",
5886 srcs = [
5887 "test/f16-dwconv-minmax.cc",
5888 "test/dwconv-microkernel-tester.h",
5889 "src/xnnpack/AlignedAllocator.h",
5890 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005891 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Frank Barchard5a599a62020-06-04 20:12:44 -07005892)
5893
5894xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07005895 name = "f32_dwconv_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005896 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07005897 "test/f32-dwconv-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005898 "test/dwconv-microkernel-tester.h",
5899 "src/xnnpack/AlignedAllocator.h",
5900 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005901 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005902)
5903
5904xnnpack_unit_test(
Marat Dukhanbf715f92020-10-23 20:17:00 -07005905 name = "f32_dwconv2d_chw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005906 srcs = [
Marat Dukhanbf715f92020-10-23 20:17:00 -07005907 "test/f32-dwconv2d-chw.cc",
5908 "test/dwconv2d-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005909 "src/xnnpack/AlignedAllocator.h",
5910 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005911 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005912)
5913
5914xnnpack_unit_test(
Frank Barchard0bb49a72020-06-04 11:35:11 -07005915 name = "f16_gavgpool_minmax_test",
5916 srcs = [
5917 "test/f16-gavgpool-minmax.cc",
5918 "test/gavgpool-microkernel-tester.h",
5919 "src/xnnpack/AlignedAllocator.h",
5920 ] + MICROKERNEL_TEST_HDRS,
5921 deps = MICROKERNEL_TEST_DEPS,
5922)
5923
5924xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07005925 name = "f32_gavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005926 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07005927 "test/f32-gavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005928 "test/gavgpool-microkernel-tester.h",
5929 "src/xnnpack/AlignedAllocator.h",
5930 ] + MICROKERNEL_TEST_HDRS,
5931 deps = MICROKERNEL_TEST_DEPS,
5932)
5933
5934xnnpack_unit_test(
Marat Dukhan1f29b802020-05-15 23:46:39 -07005935 name = "f32_gavgpool_cw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005936 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07005937 "test/f32-gavgpool-cw.cc",
5938 "test/gavgpool-cw-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005939 "src/xnnpack/AlignedAllocator.h",
5940 ] + MICROKERNEL_TEST_HDRS,
5941 deps = MICROKERNEL_TEST_DEPS,
5942)
5943
5944xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07005945 name = "f32_gemm_test",
5946 srcs = [
5947 "test/f32-gemm.cc",
5948 "test/gemm-microkernel-tester.h",
5949 "src/xnnpack/AlignedAllocator.h",
5950 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005951 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan163a7e62020-04-09 04:19:26 -07005952)
5953
5954xnnpack_unit_test(
Marat Dukhan467f6362020-05-22 23:21:55 -07005955 name = "f32_gemm_relu_test",
5956 srcs = [
5957 "test/f32-gemm-relu.cc",
5958 "test/gemm-microkernel-tester.h",
5959 "src/xnnpack/AlignedAllocator.h",
5960 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005961 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan467f6362020-05-22 23:21:55 -07005962)
5963
5964xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07005965 name = "f32_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005966 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07005967 "test/f32-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005968 "test/gemm-microkernel-tester.h",
5969 "src/xnnpack/AlignedAllocator.h",
5970 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005971 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005972)
5973
5974xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07005975 name = "f32_gemminc_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005976 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07005977 "test/f32-gemminc-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005978 "test/gemm-microkernel-tester.h",
5979 "src/xnnpack/AlignedAllocator.h",
5980 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005981 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005982)
5983
5984xnnpack_unit_test(
Frank Barchardb1966592020-05-12 13:47:06 -07005985 name = "f16_hswish_test",
5986 srcs = [
5987 "test/f16-hswish.cc",
5988 "test/hswish-microkernel-tester.h",
5989 ] + MICROKERNEL_TEST_HDRS,
5990 deps = MICROKERNEL_TEST_DEPS,
5991)
5992
5993xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005994 name = "f32_hswish_test",
5995 srcs = [
5996 "test/f32-hswish.cc",
5997 "test/hswish-microkernel-tester.h",
5998 ] + MICROKERNEL_TEST_HDRS,
5999 deps = MICROKERNEL_TEST_DEPS,
6000)
6001
6002xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07006003 name = "f32_maxpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006004 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07006005 "test/f32-maxpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006006 "test/maxpool-microkernel-tester.h",
6007 ] + MICROKERNEL_TEST_HDRS,
6008 deps = MICROKERNEL_TEST_DEPS,
6009)
6010
6011xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07006012 name = "f32_pavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006013 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07006014 "test/f32-pavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006015 "test/avgpool-microkernel-tester.h",
6016 "src/xnnpack/AlignedAllocator.h",
6017 ] + MICROKERNEL_TEST_HDRS,
6018 deps = MICROKERNEL_TEST_DEPS,
6019)
6020
6021xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07006022 name = "f32_ppmm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006023 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07006024 "test/f32-ppmm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006025 "test/gemm-microkernel-tester.h",
6026 "src/xnnpack/AlignedAllocator.h",
6027 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07006028 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07006029)
6030
6031xnnpack_unit_test(
Frank Barchardb1966592020-05-12 13:47:06 -07006032 name = "f16_prelu_test",
6033 srcs = [
6034 "test/f16-prelu.cc",
6035 "test/prelu-microkernel-tester.h",
6036 "src/xnnpack/AlignedAllocator.h",
6037 ] + MICROKERNEL_TEST_HDRS,
6038 deps = MICROKERNEL_TEST_DEPS,
6039)
6040
6041xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07006042 name = "f32_prelu_test",
6043 srcs = [
6044 "test/f32-prelu.cc",
6045 "test/prelu-microkernel-tester.h",
6046 "src/xnnpack/AlignedAllocator.h",
6047 ] + MICROKERNEL_TEST_HDRS,
6048 deps = MICROKERNEL_TEST_DEPS,
6049)
6050
6051xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07006052 name = "f32_raddexpminusmax_test",
6053 srcs = [
6054 "test/f32-raddexpminusmax.cc",
6055 "test/raddexpminusmax-microkernel-tester.h",
6056 ] + MICROKERNEL_TEST_HDRS,
6057 deps = MICROKERNEL_TEST_DEPS,
6058)
6059
6060xnnpack_unit_test(
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07006061 name = "f32_raddextexp_test",
6062 srcs = [
6063 "test/f32-raddextexp.cc",
6064 "test/raddextexp-microkernel-tester.h",
6065 ] + MICROKERNEL_TEST_HDRS,
6066 deps = MICROKERNEL_TEST_DEPS,
6067)
6068
6069xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07006070 name = "f32_raddstoreexpminusmax_test",
6071 srcs = [
6072 "test/f32-raddstoreexpminusmax.cc",
6073 "test/raddstoreexpminusmax-microkernel-tester.h",
6074 ] + MICROKERNEL_TEST_HDRS,
6075 deps = MICROKERNEL_TEST_DEPS,
6076)
6077
6078xnnpack_unit_test(
Frank Barchardfb158e22020-07-15 16:10:10 -07006079 name = "f32_relu_test",
6080 srcs = [
6081 "test/f32-relu.cc",
6082 "test/vunary-microkernel-tester.h",
6083 ] + MICROKERNEL_TEST_HDRS,
6084 deps = MICROKERNEL_TEST_DEPS,
6085)
6086
6087xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07006088 name = "f32_rmax_test",
6089 srcs = [
6090 "test/f32-rmax.cc",
6091 "test/rmax-microkernel-tester.h",
6092 ] + MICROKERNEL_TEST_HDRS,
6093 deps = MICROKERNEL_TEST_DEPS,
6094)
6095
6096xnnpack_unit_test(
Marat Dukhan346a9e52019-11-15 09:06:30 -08006097 name = "f32_sigmoid_test",
6098 srcs = [
6099 "test/f32-sigmoid.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08006100 "test/vunary-microkernel-tester.h",
Marat Dukhan346a9e52019-11-15 09:06:30 -08006101 ] + MICROKERNEL_TEST_HDRS,
6102 deps = MICROKERNEL_TEST_DEPS,
6103)
6104
6105xnnpack_unit_test(
Marat Dukhan355ab432020-04-09 19:01:52 -07006106 name = "f32_spmm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006107 srcs = [
Marat Dukhan355ab432020-04-09 19:01:52 -07006108 "test/f32-spmm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006109 "test/spmm-microkernel-tester.h",
6110 "src/xnnpack/AlignedAllocator.h",
6111 ] + MICROKERNEL_TEST_HDRS,
6112 deps = MICROKERNEL_TEST_DEPS,
6113)
6114
6115xnnpack_unit_test(
Frank Barchardd793f6c2020-05-08 13:37:43 -07006116 name = "f16_vadd_minmax_test",
6117 srcs = [
6118 "test/f16-vadd-minmax.cc",
6119 "test/vbinary-microkernel-tester.h",
6120 ] + MICROKERNEL_TEST_HDRS,
6121 deps = MICROKERNEL_TEST_DEPS,
6122)
6123
6124xnnpack_unit_test(
6125 name = "f16_vaddc_minmax_test",
6126 srcs = [
6127 "test/f16-vaddc-minmax.cc",
6128 "test/vbinaryc-microkernel-tester.h",
6129 ] + MICROKERNEL_TEST_HDRS,
6130 deps = MICROKERNEL_TEST_DEPS,
6131)
6132
6133xnnpack_unit_test(
6134 name = "f16_vdiv_minmax_test",
6135 srcs = [
6136 "test/f16-vdiv-minmax.cc",
6137 "test/vbinary-microkernel-tester.h",
6138 ] + MICROKERNEL_TEST_HDRS,
6139 deps = MICROKERNEL_TEST_DEPS,
6140)
6141
6142xnnpack_unit_test(
6143 name = "f16_vdivc_minmax_test",
6144 srcs = [
6145 "test/f16-vdivc-minmax.cc",
6146 "test/vbinaryc-microkernel-tester.h",
6147 ] + MICROKERNEL_TEST_HDRS,
6148 deps = MICROKERNEL_TEST_DEPS,
6149)
6150
6151xnnpack_unit_test(
6152 name = "f16_vrdivc_minmax_test",
6153 srcs = [
6154 "test/f16-vrdivc-minmax.cc",
6155 "test/vbinaryc-microkernel-tester.h",
6156 ] + MICROKERNEL_TEST_HDRS,
6157 deps = MICROKERNEL_TEST_DEPS,
6158)
6159
6160xnnpack_unit_test(
6161 name = "f16_vmax_test",
6162 srcs = [
6163 "test/f16-vmax.cc",
6164 "test/vbinary-microkernel-tester.h",
6165 ] + MICROKERNEL_TEST_HDRS,
6166 deps = MICROKERNEL_TEST_DEPS,
6167)
6168
6169xnnpack_unit_test(
6170 name = "f16_vmaxc_test",
6171 srcs = [
6172 "test/f16-vmaxc.cc",
6173 "test/vbinaryc-microkernel-tester.h",
6174 ] + MICROKERNEL_TEST_HDRS,
6175 deps = MICROKERNEL_TEST_DEPS,
6176)
6177
6178xnnpack_unit_test(
6179 name = "f16_vmin_test",
6180 srcs = [
6181 "test/f16-vmin.cc",
6182 "test/vbinary-microkernel-tester.h",
6183 ] + MICROKERNEL_TEST_HDRS,
6184 deps = MICROKERNEL_TEST_DEPS,
6185)
6186
6187xnnpack_unit_test(
6188 name = "f16_vminc_test",
6189 srcs = [
6190 "test/f16-vminc.cc",
6191 "test/vbinaryc-microkernel-tester.h",
6192 ] + MICROKERNEL_TEST_HDRS,
6193 deps = MICROKERNEL_TEST_DEPS,
6194)
6195
6196xnnpack_unit_test(
6197 name = "f16_vmul_minmax_test",
6198 srcs = [
6199 "test/f16-vmul-minmax.cc",
6200 "test/vbinary-microkernel-tester.h",
6201 ] + MICROKERNEL_TEST_HDRS,
6202 deps = MICROKERNEL_TEST_DEPS,
6203)
6204
6205xnnpack_unit_test(
6206 name = "f16_vmulc_minmax_test",
6207 srcs = [
6208 "test/f16-vmulc-minmax.cc",
6209 "test/vbinaryc-microkernel-tester.h",
6210 ] + MICROKERNEL_TEST_HDRS,
6211 deps = MICROKERNEL_TEST_DEPS,
6212)
6213
6214xnnpack_unit_test(
6215 name = "f16_vsub_minmax_test",
6216 srcs = [
6217 "test/f16-vsub-minmax.cc",
6218 "test/vbinary-microkernel-tester.h",
6219 ] + MICROKERNEL_TEST_HDRS,
6220 deps = MICROKERNEL_TEST_DEPS,
6221)
6222
6223xnnpack_unit_test(
6224 name = "f16_vsubc_minmax_test",
6225 srcs = [
6226 "test/f16-vsubc-minmax.cc",
6227 "test/vbinaryc-microkernel-tester.h",
6228 ] + MICROKERNEL_TEST_HDRS,
6229 deps = MICROKERNEL_TEST_DEPS,
6230)
6231
6232xnnpack_unit_test(
6233 name = "f16_vrsubc_minmax_test",
6234 srcs = [
6235 "test/f16-vrsubc-minmax.cc",
6236 "test/vbinaryc-microkernel-tester.h",
6237 ] + MICROKERNEL_TEST_HDRS,
6238 deps = MICROKERNEL_TEST_DEPS,
6239)
6240
6241xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07006242 name = "f32_vabs_test",
6243 srcs = [
6244 "test/f32-vabs.cc",
6245 "test/vunary-microkernel-tester.h",
6246 ] + MICROKERNEL_TEST_HDRS,
6247 deps = MICROKERNEL_TEST_DEPS,
6248)
6249
6250xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07006251 name = "f32_vadd_test",
6252 srcs = [
6253 "test/f32-vadd.cc",
6254 "test/vbinary-microkernel-tester.h",
6255 ] + MICROKERNEL_TEST_HDRS,
6256 deps = MICROKERNEL_TEST_DEPS,
6257)
6258
6259xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006260 name = "f32_vadd_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006261 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006262 "test/f32-vadd-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08006263 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08006264 ] + MICROKERNEL_TEST_HDRS,
6265 deps = MICROKERNEL_TEST_DEPS,
6266)
6267
6268xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07006269 name = "f32_vadd_relu_test",
6270 srcs = [
6271 "test/f32-vadd-relu.cc",
6272 "test/vbinary-microkernel-tester.h",
6273 ] + MICROKERNEL_TEST_HDRS,
6274 deps = MICROKERNEL_TEST_DEPS,
6275)
6276
6277xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07006278 name = "f32_vaddc_test",
6279 srcs = [
6280 "test/f32-vaddc.cc",
6281 "test/vbinaryc-microkernel-tester.h",
6282 ] + MICROKERNEL_TEST_HDRS,
6283 deps = MICROKERNEL_TEST_DEPS,
6284)
6285
6286xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006287 name = "f32_vaddc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08006288 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006289 "test/f32-vaddc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08006290 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006291 ] + MICROKERNEL_TEST_HDRS,
6292 deps = MICROKERNEL_TEST_DEPS,
6293)
6294
6295xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07006296 name = "f32_vaddc_relu_test",
6297 srcs = [
6298 "test/f32-vaddc-relu.cc",
6299 "test/vbinaryc-microkernel-tester.h",
6300 ] + MICROKERNEL_TEST_HDRS,
6301 deps = MICROKERNEL_TEST_DEPS,
6302)
6303
6304xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07006305 name = "f32_vdiv_test",
6306 srcs = [
6307 "test/f32-vdiv.cc",
6308 "test/vbinary-microkernel-tester.h",
6309 ] + MICROKERNEL_TEST_HDRS,
6310 deps = MICROKERNEL_TEST_DEPS,
6311)
6312
6313xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006314 name = "f32_vdiv_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08006315 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006316 "test/f32-vdiv-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08006317 "test/vbinary-microkernel-tester.h",
6318 ] + MICROKERNEL_TEST_HDRS,
6319 deps = MICROKERNEL_TEST_DEPS,
6320)
6321
6322xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07006323 name = "f32_vdiv_relu_test",
6324 srcs = [
6325 "test/f32-vdiv-relu.cc",
6326 "test/vbinary-microkernel-tester.h",
6327 ] + MICROKERNEL_TEST_HDRS,
6328 deps = MICROKERNEL_TEST_DEPS,
6329)
6330
6331xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07006332 name = "f32_vdivc_test",
6333 srcs = [
6334 "test/f32-vdivc.cc",
6335 "test/vbinaryc-microkernel-tester.h",
6336 ] + MICROKERNEL_TEST_HDRS,
6337 deps = MICROKERNEL_TEST_DEPS,
6338)
6339
6340xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006341 name = "f32_vdivc_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08006342 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006343 "test/f32-vdivc-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08006344 "test/vbinaryc-microkernel-tester.h",
6345 ] + MICROKERNEL_TEST_HDRS,
6346 deps = MICROKERNEL_TEST_DEPS,
6347)
6348
6349xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07006350 name = "f32_vdivc_relu_test",
6351 srcs = [
6352 "test/f32-vdivc-relu.cc",
6353 "test/vbinaryc-microkernel-tester.h",
6354 ] + MICROKERNEL_TEST_HDRS,
6355 deps = MICROKERNEL_TEST_DEPS,
6356)
6357
6358xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07006359 name = "f32_vrdivc_test",
6360 srcs = [
6361 "test/f32-vrdivc.cc",
6362 "test/vbinaryc-microkernel-tester.h",
6363 ] + MICROKERNEL_TEST_HDRS,
6364 deps = MICROKERNEL_TEST_DEPS,
6365)
6366
6367xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006368 name = "f32_vrdivc_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08006369 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006370 "test/f32-vrdivc-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08006371 "test/vbinaryc-microkernel-tester.h",
6372 ] + MICROKERNEL_TEST_HDRS,
6373 deps = MICROKERNEL_TEST_DEPS,
6374)
6375
6376xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07006377 name = "f32_vrdivc_relu_test",
6378 srcs = [
6379 "test/f32-vrdivc-relu.cc",
6380 "test/vbinaryc-microkernel-tester.h",
6381 ] + MICROKERNEL_TEST_HDRS,
6382 deps = MICROKERNEL_TEST_DEPS,
6383)
6384
6385xnnpack_unit_test(
Marat Dukhaned6baaf2020-12-01 15:07:08 -08006386 name = "f32_velu_test",
6387 srcs = [
6388 "test/f32-velu.cc",
6389 "test/vunary-microkernel-tester.h",
6390 ] + MICROKERNEL_TEST_HDRS,
6391 deps = MICROKERNEL_TEST_DEPS,
6392)
6393
6394xnnpack_unit_test(
Marat Dukhan403b7d42019-12-05 12:49:11 -08006395 name = "f32_vmax_test",
6396 srcs = [
6397 "test/f32-vmax.cc",
6398 "test/vbinary-microkernel-tester.h",
6399 ] + MICROKERNEL_TEST_HDRS,
6400 deps = MICROKERNEL_TEST_DEPS,
6401)
6402
6403xnnpack_unit_test(
6404 name = "f32_vmaxc_test",
6405 srcs = [
6406 "test/f32-vmaxc.cc",
6407 "test/vbinaryc-microkernel-tester.h",
6408 ] + MICROKERNEL_TEST_HDRS,
6409 deps = MICROKERNEL_TEST_DEPS,
6410)
6411
6412xnnpack_unit_test(
6413 name = "f32_vmin_test",
6414 srcs = [
6415 "test/f32-vmin.cc",
6416 "test/vbinary-microkernel-tester.h",
6417 ] + MICROKERNEL_TEST_HDRS,
6418 deps = MICROKERNEL_TEST_DEPS,
6419)
6420
6421xnnpack_unit_test(
6422 name = "f32_vminc_test",
6423 srcs = [
6424 "test/f32-vminc.cc",
6425 "test/vbinaryc-microkernel-tester.h",
6426 ] + MICROKERNEL_TEST_HDRS,
6427 deps = MICROKERNEL_TEST_DEPS,
6428)
6429
6430xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07006431 name = "f32_vmul_test",
6432 srcs = [
6433 "test/f32-vmul.cc",
6434 "test/vbinary-microkernel-tester.h",
6435 ] + MICROKERNEL_TEST_HDRS,
6436 deps = MICROKERNEL_TEST_DEPS,
6437)
6438
6439xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006440 name = "f32_vmul_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006441 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006442 "test/f32-vmul-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08006443 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08006444 ] + MICROKERNEL_TEST_HDRS,
6445 deps = MICROKERNEL_TEST_DEPS,
6446)
6447
6448xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07006449 name = "f32_vmul_relu_test",
6450 srcs = [
6451 "test/f32-vmul-relu.cc",
6452 "test/vbinary-microkernel-tester.h",
6453 ] + MICROKERNEL_TEST_HDRS,
6454 deps = MICROKERNEL_TEST_DEPS,
6455)
6456
6457xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07006458 name = "f32_vmulc_test",
6459 srcs = [
6460 "test/f32-vmulc.cc",
6461 "test/vbinaryc-microkernel-tester.h",
6462 ] + MICROKERNEL_TEST_HDRS,
6463 deps = MICROKERNEL_TEST_DEPS,
6464)
6465
6466xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006467 name = "f32_vmulc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08006468 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006469 "test/f32-vmulc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08006470 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006471 ] + MICROKERNEL_TEST_HDRS,
6472 deps = MICROKERNEL_TEST_DEPS,
6473)
6474
6475xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07006476 name = "f32_vmulc_relu_test",
6477 srcs = [
6478 "test/f32-vmulc-relu.cc",
6479 "test/vbinaryc-microkernel-tester.h",
6480 ] + MICROKERNEL_TEST_HDRS,
6481 deps = MICROKERNEL_TEST_DEPS,
6482)
6483
6484xnnpack_unit_test(
Frank Barchard2a1049c2020-06-03 02:31:27 -07006485 name = "f16_vmulcaddc_minmax_test",
6486 srcs = [
6487 "test/f16-vmulcaddc-minmax.cc",
6488 "test/vmulcaddc-microkernel-tester.h",
6489 "src/xnnpack/AlignedAllocator.h",
6490 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07006491 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Frank Barchard2a1049c2020-06-03 02:31:27 -07006492)
6493
6494xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07006495 name = "f32_vmulcaddc_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006496 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07006497 "test/f32-vmulcaddc-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006498 "test/vmulcaddc-microkernel-tester.h",
6499 "src/xnnpack/AlignedAllocator.h",
6500 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07006501 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07006502)
6503
6504xnnpack_unit_test(
Marat Dukhan8cc7efe2020-06-10 16:24:27 -07006505 name = "f32_vlrelu_test",
6506 srcs = [
6507 "test/f32-vlrelu.cc",
6508 "test/vunary-microkernel-tester.h",
6509 ] + MICROKERNEL_TEST_HDRS,
6510 deps = MICROKERNEL_TEST_DEPS,
6511)
6512
6513xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07006514 name = "f32_vneg_test",
6515 srcs = [
6516 "test/f32-vneg.cc",
6517 "test/vunary-microkernel-tester.h",
6518 ] + MICROKERNEL_TEST_HDRS,
6519 deps = MICROKERNEL_TEST_DEPS,
6520)
6521
6522xnnpack_unit_test(
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07006523 name = "f32_vrndne_test",
6524 srcs = [
6525 "test/f32-vrndne.cc",
6526 "test/vunary-microkernel-tester.h",
6527 ] + MICROKERNEL_TEST_HDRS,
6528 deps = MICROKERNEL_TEST_DEPS,
6529)
6530
6531xnnpack_unit_test(
6532 name = "f32_vrndz_test",
6533 srcs = [
6534 "test/f32-vrndz.cc",
6535 "test/vunary-microkernel-tester.h",
6536 ] + MICROKERNEL_TEST_HDRS,
6537 deps = MICROKERNEL_TEST_DEPS,
6538)
6539
6540xnnpack_unit_test(
6541 name = "f32_vrndu_test",
6542 srcs = [
6543 "test/f32-vrndu.cc",
6544 "test/vunary-microkernel-tester.h",
6545 ] + MICROKERNEL_TEST_HDRS,
6546 deps = MICROKERNEL_TEST_DEPS,
6547)
6548
6549xnnpack_unit_test(
6550 name = "f32_vrndd_test",
6551 srcs = [
6552 "test/f32-vrndd.cc",
6553 "test/vunary-microkernel-tester.h",
6554 ] + MICROKERNEL_TEST_HDRS,
6555 deps = MICROKERNEL_TEST_DEPS,
6556)
6557
6558xnnpack_unit_test(
Marat Dukhan05ac8e32019-10-21 15:39:33 -07006559 name = "f32_vscale_test",
6560 srcs = [
6561 "test/f32-vscale.cc",
6562 "test/vscale-microkernel-tester.h",
6563 ] + MICROKERNEL_TEST_HDRS,
6564 deps = MICROKERNEL_TEST_DEPS,
6565)
6566
6567xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07006568 name = "f32_vscaleexpminusmax_test",
6569 srcs = [
6570 "test/f32-vscaleexpminusmax.cc",
6571 "test/vscaleexpminusmax-microkernel-tester.h",
6572 ] + MICROKERNEL_TEST_HDRS,
6573 deps = MICROKERNEL_TEST_DEPS,
6574)
6575
6576xnnpack_unit_test(
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07006577 name = "f32_vscaleextexp_test",
6578 srcs = [
6579 "test/f32-vscaleextexp.cc",
6580 "test/vscaleextexp-microkernel-tester.h",
6581 ] + MICROKERNEL_TEST_HDRS,
6582 deps = MICROKERNEL_TEST_DEPS,
6583)
6584
6585xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07006586 name = "f32_vsqr_test",
6587 srcs = [
6588 "test/f32-vsqr.cc",
6589 "test/vunary-microkernel-tester.h",
6590 ] + MICROKERNEL_TEST_HDRS,
6591 deps = MICROKERNEL_TEST_DEPS,
6592)
6593
6594xnnpack_unit_test(
Marat Dukhan13bafb02020-06-05 00:43:11 -07006595 name = "f32_vsqrdiff_test",
6596 srcs = [
6597 "test/f32-vsqrdiff.cc",
6598 "test/vbinary-microkernel-tester.h",
6599 ] + MICROKERNEL_TEST_HDRS,
6600 deps = MICROKERNEL_TEST_DEPS,
6601)
6602
6603xnnpack_unit_test(
6604 name = "f32_vsqrdiffc_test",
6605 srcs = [
6606 "test/f32-vsqrdiffc.cc",
6607 "test/vbinaryc-microkernel-tester.h",
6608 ] + MICROKERNEL_TEST_HDRS,
6609 deps = MICROKERNEL_TEST_DEPS,
6610)
6611
6612xnnpack_unit_test(
Marat Dukhanf4db2f32020-06-30 10:55:30 -07006613 name = "f32_vsqrt_test",
6614 srcs = [
6615 "test/f32-vsqrt.cc",
6616 "test/vunary-microkernel-tester.h",
6617 ] + MICROKERNEL_TEST_HDRS,
6618 deps = MICROKERNEL_TEST_DEPS,
6619)
6620
6621xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07006622 name = "f32_vsub_test",
6623 srcs = [
6624 "test/f32-vsub.cc",
6625 "test/vbinary-microkernel-tester.h",
6626 ] + MICROKERNEL_TEST_HDRS,
6627 deps = MICROKERNEL_TEST_DEPS,
6628)
6629
6630xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006631 name = "f32_vsub_minmax_test",
Marat Dukhan97579532019-10-18 16:40:39 -07006632 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006633 "test/f32-vsub-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08006634 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08006635 ] + MICROKERNEL_TEST_HDRS,
6636 deps = MICROKERNEL_TEST_DEPS,
6637)
6638
6639xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07006640 name = "f32_vsub_relu_test",
6641 srcs = [
6642 "test/f32-vsub-relu.cc",
6643 "test/vbinary-microkernel-tester.h",
6644 ] + MICROKERNEL_TEST_HDRS,
6645 deps = MICROKERNEL_TEST_DEPS,
6646)
6647
6648xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07006649 name = "f32_vsubc_test",
6650 srcs = [
6651 "test/f32-vsubc.cc",
6652 "test/vbinaryc-microkernel-tester.h",
6653 ] + MICROKERNEL_TEST_HDRS,
6654 deps = MICROKERNEL_TEST_DEPS,
6655)
6656
6657xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006658 name = "f32_vsubc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08006659 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006660 "test/f32-vsubc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08006661 "test/vbinaryc-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08006662 ] + MICROKERNEL_TEST_HDRS,
6663 deps = MICROKERNEL_TEST_DEPS,
6664)
6665
6666xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07006667 name = "f32_vsubc_relu_test",
6668 srcs = [
6669 "test/f32-vsubc-relu.cc",
6670 "test/vbinaryc-microkernel-tester.h",
6671 ] + MICROKERNEL_TEST_HDRS,
6672 deps = MICROKERNEL_TEST_DEPS,
6673)
6674
6675xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07006676 name = "f32_vrsubc_test",
6677 srcs = [
6678 "test/f32-vrsubc.cc",
6679 "test/vbinaryc-microkernel-tester.h",
6680 ] + MICROKERNEL_TEST_HDRS,
6681 deps = MICROKERNEL_TEST_DEPS,
6682)
6683
6684xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006685 name = "f32_vrsubc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08006686 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006687 "test/f32-vrsubc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08006688 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan97579532019-10-18 16:40:39 -07006689 ] + MICROKERNEL_TEST_HDRS,
6690 deps = MICROKERNEL_TEST_DEPS,
6691)
6692
6693xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07006694 name = "f32_vrsubc_relu_test",
6695 srcs = [
6696 "test/f32-vrsubc-relu.cc",
6697 "test/vbinaryc-microkernel-tester.h",
6698 ] + MICROKERNEL_TEST_HDRS,
6699 deps = MICROKERNEL_TEST_DEPS,
6700)
6701
6702xnnpack_unit_test(
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07006703 name = "qs8_dwconv_minmax_test",
6704 srcs = [
6705 "test/qs8-dwconv-minmax.cc",
6706 "test/dwconv-microkernel-tester.h",
6707 "src/xnnpack/AlignedAllocator.h",
6708 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
6709 deps = MICROKERNEL_TEST_DEPS + [":packing"],
6710)
6711
6712xnnpack_unit_test(
Marat Dukhan4ed53f42020-08-06 01:12:55 -07006713 name = "qs8_gavgpool_minmax_test",
6714 srcs = [
6715 "test/qs8-gavgpool-minmax.cc",
6716 "test/gavgpool-microkernel-tester.h",
6717 "src/xnnpack/AlignedAllocator.h",
6718 ] + MICROKERNEL_TEST_HDRS,
6719 deps = MICROKERNEL_TEST_DEPS,
6720)
6721
6722xnnpack_unit_test(
Marat Dukhan595e1702020-07-31 10:12:52 -07006723 name = "qs8_gemm_minmax_test",
6724 srcs = [
6725 "test/qs8-gemm-minmax.cc",
6726 "test/gemm-microkernel-tester.h",
6727 "src/xnnpack/AlignedAllocator.h",
6728 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
6729 deps = MICROKERNEL_TEST_DEPS + [":packing"],
6730)
6731
6732xnnpack_unit_test(
Marat Dukhanf9480682020-07-31 14:50:24 -07006733 name = "qs8_igemm_minmax_test",
6734 srcs = [
6735 "test/qs8-igemm-minmax.cc",
6736 "test/gemm-microkernel-tester.h",
6737 "src/xnnpack/AlignedAllocator.h",
6738 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
6739 deps = MICROKERNEL_TEST_DEPS + [":packing"],
6740)
6741
6742xnnpack_unit_test(
6743 name = "qs8_requantization_test",
6744 srcs = [
6745 "src/xnnpack/requantization-stubs.h",
6746 "test/qs8-requantization.cc",
6747 "test/requantization-tester.h",
6748 ] + MICROKERNEL_TEST_HDRS,
6749 deps = MICROKERNEL_TEST_DEPS,
6750)
6751
6752xnnpack_unit_test(
Marat Dukhand9f3ad42020-08-10 12:30:58 -07006753 name = "qs8_vadd_minmax_test",
6754 srcs = [
6755 "test/qs8-vadd-minmax.cc",
6756 "test/vadd-microkernel-tester.h",
6757 ] + MICROKERNEL_TEST_HDRS,
6758 deps = MICROKERNEL_TEST_DEPS,
6759)
6760
6761xnnpack_unit_test(
Marat Dukhan0270d9f2020-08-11 00:56:46 -07006762 name = "qs8_vaddc_minmax_test",
6763 srcs = [
6764 "test/qs8-vaddc-minmax.cc",
6765 "test/vaddc-microkernel-tester.h",
6766 ] + MICROKERNEL_TEST_HDRS,
6767 deps = MICROKERNEL_TEST_DEPS,
6768)
6769
6770xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07006771 name = "qu8_avgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006772 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07006773 "test/qu8-avgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006774 "test/avgpool-microkernel-tester.h",
6775 "src/xnnpack/AlignedAllocator.h",
6776 ] + MICROKERNEL_TEST_HDRS,
6777 deps = MICROKERNEL_TEST_DEPS,
6778)
6779
6780xnnpack_unit_test(
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07006781 name = "qu8_dwconv_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006782 srcs = [
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07006783 "test/qu8-dwconv-minmax.cc",
6784 "test/dwconv-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006785 "src/xnnpack/AlignedAllocator.h",
6786 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07006787 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07006788)
6789
6790xnnpack_unit_test(
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07006791 name = "qu8_igemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006792 srcs = [
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07006793 "test/qu8-igemm-minmax.cc",
6794 "test/gemm-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006795 "src/xnnpack/AlignedAllocator.h",
6796 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07006797 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07006798)
6799
6800xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07006801 name = "qu8_gavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006802 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07006803 "test/qu8-gavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006804 "test/gavgpool-microkernel-tester.h",
6805 "src/xnnpack/AlignedAllocator.h",
6806 ] + MICROKERNEL_TEST_HDRS,
6807 deps = MICROKERNEL_TEST_DEPS,
6808)
6809
6810xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07006811 name = "qu8_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006812 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07006813 "test/qu8-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006814 "test/gemm-microkernel-tester.h",
6815 "src/xnnpack/AlignedAllocator.h",
6816 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07006817 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07006818)
6819
6820xnnpack_unit_test(
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07006821 name = "qu8_requantization_test",
6822 srcs = [
6823 "src/xnnpack/requantization-stubs.h",
6824 "test/qu8-requantization.cc",
6825 "test/requantization-tester.h",
6826 ] + MICROKERNEL_TEST_HDRS,
6827 deps = MICROKERNEL_TEST_DEPS,
6828)
6829
6830xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07006831 name = "qu8_vadd_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006832 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07006833 "test/qu8-vadd-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006834 "test/vadd-microkernel-tester.h",
6835 ] + MICROKERNEL_TEST_HDRS,
6836 deps = MICROKERNEL_TEST_DEPS,
6837)
6838
6839xnnpack_unit_test(
6840 name = "u8_clamp_test",
6841 srcs = [
6842 "test/u8-clamp.cc",
6843 "test/clamp-microkernel-tester.h",
6844 ] + MICROKERNEL_TEST_HDRS,
6845 deps = MICROKERNEL_TEST_DEPS,
6846)
6847
6848xnnpack_unit_test(
6849 name = "u8_lut32norm_test",
6850 srcs = [
6851 "test/u8-lut32norm.cc",
6852 "test/lut-norm-microkernel-tester.h",
6853 ] + MICROKERNEL_TEST_HDRS,
6854 deps = MICROKERNEL_TEST_DEPS,
6855)
6856
6857xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07006858 name = "u8_maxpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006859 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07006860 "test/u8-maxpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006861 "test/maxpool-microkernel-tester.h",
6862 ] + MICROKERNEL_TEST_HDRS,
6863 deps = MICROKERNEL_TEST_DEPS,
6864)
6865
6866xnnpack_unit_test(
6867 name = "u8_rmax_test",
6868 srcs = [
6869 "test/u8-rmax.cc",
6870 "test/rmax-microkernel-tester.h",
6871 ] + MICROKERNEL_TEST_HDRS,
6872 deps = MICROKERNEL_TEST_DEPS,
6873)
6874
6875xnnpack_unit_test(
Marat Dukhanad71b9a2020-11-20 00:01:51 -08006876 name = "x32_depthtospace2d_chw2hwc_test",
Yury Kartynnike7841862020-11-04 18:22:18 -08006877 srcs = [
Marat Dukhanad71b9a2020-11-20 00:01:51 -08006878 "test/x32-depthtospace2d-chw2hwc.cc",
6879 "test/depthtospace-microkernel-tester.h",
Yury Kartynnike7841862020-11-04 18:22:18 -08006880 ] + MICROKERNEL_TEST_HDRS,
6881 deps = MICROKERNEL_TEST_DEPS,
6882)
6883
6884xnnpack_unit_test(
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -07006885 name = "x32_fill_test",
6886 srcs = [
6887 "test/x32-fill.cc",
6888 "test/fill-microkernel-tester.h",
6889 ] + MICROKERNEL_TEST_HDRS,
6890 deps = MICROKERNEL_TEST_DEPS,
6891)
6892
6893xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07006894 name = "x32_packx_test",
6895 srcs = [
6896 "test/x32-packx.cc",
6897 "test/pack-microkernel-tester.h",
6898 "src/xnnpack/AlignedAllocator.h",
6899 ] + MICROKERNEL_TEST_HDRS,
6900 deps = MICROKERNEL_TEST_DEPS,
6901)
6902
6903xnnpack_unit_test(
6904 name = "x32_pad_test",
6905 srcs = [
6906 "test/x32-pad.cc",
6907 "test/pad-microkernel-tester.h",
6908 ] + MICROKERNEL_TEST_HDRS,
6909 deps = MICROKERNEL_TEST_DEPS,
6910)
6911
6912xnnpack_unit_test(
6913 name = "x32_unpool_test",
6914 srcs = [
6915 "test/x32-unpool.cc",
6916 "test/unpool-microkernel-tester.h",
6917 ] + MICROKERNEL_TEST_HDRS,
6918 deps = MICROKERNEL_TEST_DEPS,
6919)
6920
6921xnnpack_unit_test(
6922 name = "x32_zip_test",
6923 srcs = [
6924 "test/x32-zip.cc",
6925 "test/zip-microkernel-tester.h",
6926 ] + MICROKERNEL_TEST_HDRS,
6927 deps = MICROKERNEL_TEST_DEPS,
6928)
6929
6930xnnpack_unit_test(
6931 name = "x8_lut_test",
6932 srcs = [
6933 "test/x8-lut.cc",
6934 "test/lut-microkernel-tester.h",
6935 ] + MICROKERNEL_TEST_HDRS,
6936 deps = MICROKERNEL_TEST_DEPS,
6937)
6938
6939xnnpack_unit_test(
6940 name = "x8_zip_test",
6941 srcs = [
6942 "test/x8-zip.cc",
6943 "test/zip-microkernel-tester.h",
6944 ] + MICROKERNEL_TEST_HDRS,
6945 deps = MICROKERNEL_TEST_DEPS,
6946)
6947
Marat Dukhan20c3b922020-03-10 03:45:06 -07006948########################## Size tests for the library #########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07006949
6950xnnpack_binary(
Marat Dukhan20c3b922020-03-10 03:45:06 -07006951 name = "operator_size_test",
6952 srcs = ["test/operator-size.c"],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07006953 deps = [":xnnpack_operators_nhwc_f32"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07006954)
6955
Marat Dukhan20c3b922020-03-10 03:45:06 -07006956xnnpack_binary(
6957 name = "subgraph_size_test",
6958 srcs = ["test/subgraph-size.c"],
6959 deps = [":XNNPACK"],
6960)
6961
6962########################### Unit tests for operators ##########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07006963
6964xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07006965 name = "abs_nc_test",
6966 srcs = [
6967 "test/abs-nc.cc",
6968 "test/abs-operator-tester.h",
6969 ],
6970 deps = OPERATOR_TEST_DEPS,
6971)
6972
6973xnnpack_unit_test(
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08006974 name = "add_nd_test",
Artsiom Ablavatski2202c812021-01-22 14:16:43 -08006975 timeout = "moderate",
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08006976 srcs = [
6977 "test/add-nd.cc",
6978 "test/binary-elementwise-operator-tester.h",
6979 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006980 deps = OPERATOR_TEST_DEPS,
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08006981)
6982
6983xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006984 name = "argmax_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006985 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006986 "test/argmax-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006987 "test/argmax-pooling-operator-tester.h",
6988 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07006989 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006990)
6991
6992xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006993 name = "average_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006994 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006995 "test/average-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006996 "test/average-pooling-operator-tester.h",
6997 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07006998 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006999)
7000
7001xnnpack_unit_test(
Marat Dukhan64e52512020-06-09 13:41:16 -07007002 name = "bankers_rounding_nc_test",
7003 srcs = [
7004 "test/bankers-rounding-nc.cc",
7005 "test/bankers-rounding-operator-tester.h",
7006 ],
7007 deps = OPERATOR_TEST_DEPS,
7008)
7009
7010xnnpack_unit_test(
7011 name = "ceiling_nc_test",
7012 srcs = [
7013 "test/ceiling-nc.cc",
7014 "test/ceiling-operator-tester.h",
7015 ],
7016 deps = OPERATOR_TEST_DEPS,
7017)
7018
7019xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08007020 name = "channel_shuffle_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007021 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08007022 "test/channel-shuffle-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007023 "test/channel-shuffle-operator-tester.h",
7024 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07007025 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07007026)
7027
7028xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08007029 name = "clamp_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007030 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08007031 "test/clamp-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007032 "test/clamp-operator-tester.h",
7033 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07007034 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07007035)
7036
7037xnnpack_unit_test(
Marat Dukhan065b11e2020-05-22 09:49:41 -07007038 name = "constant_pad_nd_test",
7039 srcs = [
7040 "test/constant-pad-nd.cc",
7041 "test/constant-pad-operator-tester.h",
7042 ],
7043 deps = OPERATOR_TEST_DEPS,
7044)
7045
7046xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08007047 name = "convolution_nhwc_test",
Artsiom Ablavatski2202c812021-01-22 14:16:43 -08007048 timeout = "moderate",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007049 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08007050 "test/convolution-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007051 "test/convolution-operator-tester.h",
7052 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07007053 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07007054)
7055
7056xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08007057 name = "convolution_nchw_test",
Artsiom Ablavatski2202c812021-01-22 14:16:43 -08007058 timeout = "moderate",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007059 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08007060 "test/convolution-nchw.cc",
7061 "test/convolution-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007062 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07007063 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07007064)
7065
7066xnnpack_unit_test(
Marat Dukhan4e21b272020-06-04 18:45:01 -07007067 name = "copy_nc_test",
7068 srcs = [
7069 "test/copy-nc.cc",
7070 "test/copy-operator-tester.h",
7071 ],
7072 deps = OPERATOR_TEST_DEPS,
7073)
7074
7075xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08007076 name = "deconvolution_nhwc_test",
Artsiom Ablavatskic1aa2972020-12-08 11:23:34 -08007077 timeout = "moderate",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007078 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08007079 "test/deconvolution-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007080 "test/deconvolution-operator-tester.h",
7081 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07007082 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07007083)
7084
7085xnnpack_unit_test(
Marat Dukhan188d1042020-11-24 23:39:40 -08007086 name = "depth_to_space_nchw2nhwc_test",
Artsiom Ablavatski0f1dc182020-11-05 19:21:50 -08007087 srcs = [
Marat Dukhan188d1042020-11-24 23:39:40 -08007088 "test/depth-to-space-nchw2nhwc.cc",
Artsiom Ablavatski0f1dc182020-11-05 19:21:50 -08007089 "test/depth-to-space-operator-tester.h",
7090 ] + OPERATOR_TEST_PARAMS_HDRS,
7091 deps = OPERATOR_TEST_DEPS,
7092)
7093
7094xnnpack_unit_test(
Marat Dukhan0e521172020-11-25 13:10:04 -08007095 name = "depth_to_space_nhwc_test",
7096 srcs = [
7097 "test/depth-to-space-nhwc.cc",
7098 "test/depth-to-space-operator-tester.h",
7099 ] + OPERATOR_TEST_PARAMS_HDRS,
7100 deps = OPERATOR_TEST_DEPS,
7101)
7102
7103xnnpack_unit_test(
Marat Dukhan69180502019-12-06 15:00:31 -08007104 name = "divide_nd_test",
7105 srcs = [
7106 "test/binary-elementwise-operator-tester.h",
7107 "test/divide-nd.cc",
7108 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07007109 deps = OPERATOR_TEST_DEPS,
Marat Dukhan69180502019-12-06 15:00:31 -08007110)
7111
7112xnnpack_unit_test(
Marat Dukhanb6bd4bc2020-12-01 17:01:40 -08007113 name = "elu_nc_test",
7114 srcs = [
7115 "test/elu-nc.cc",
7116 "test/elu-operator-tester.h",
7117 ],
7118 deps = OPERATOR_TEST_DEPS,
7119)
7120
7121xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08007122 name = "fully_connected_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007123 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08007124 "test/fully-connected-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007125 "test/fully-connected-operator-tester.h",
7126 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07007127 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07007128)
7129
7130xnnpack_unit_test(
Marat Dukhan64e52512020-06-09 13:41:16 -07007131 name = "floor_nc_test",
7132 srcs = [
7133 "test/floor-nc.cc",
7134 "test/floor-operator-tester.h",
7135 ],
7136 deps = OPERATOR_TEST_DEPS,
7137)
7138
7139xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08007140 name = "global_average_pooling_nwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007141 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08007142 "test/global-average-pooling-nwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007143 "test/global-average-pooling-operator-tester.h",
Marat Dukhanef61d022020-06-19 13:54:49 -07007144 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07007145 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07007146)
7147
7148xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08007149 name = "global_average_pooling_ncw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007150 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08007151 "test/global-average-pooling-ncw.cc",
7152 "test/global-average-pooling-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007153 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07007154 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07007155)
7156
7157xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08007158 name = "hardswish_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007159 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08007160 "test/hardswish-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007161 "test/hardswish-operator-tester.h",
7162 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07007163 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07007164)
7165
7166xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08007167 name = "leaky_relu_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007168 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08007169 "test/leaky-relu-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007170 "test/leaky-relu-operator-tester.h",
7171 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07007172 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07007173)
7174
7175xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08007176 name = "max_pooling_nhwc_test",
Artsiom Ablavatski2202c812021-01-22 14:16:43 -08007177 timeout = "moderate",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007178 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08007179 "test/max-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007180 "test/max-pooling-operator-tester.h",
7181 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07007182 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07007183)
7184
7185xnnpack_unit_test(
Marat Dukhan79e7f842019-12-05 14:35:50 -08007186 name = "maximum_nd_test",
7187 srcs = [
7188 "test/binary-elementwise-operator-tester.h",
7189 "test/maximum-nd.cc",
7190 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07007191 deps = OPERATOR_TEST_DEPS,
Marat Dukhan79e7f842019-12-05 14:35:50 -08007192)
7193
7194xnnpack_unit_test(
7195 name = "minimum_nd_test",
7196 srcs = [
7197 "test/binary-elementwise-operator-tester.h",
7198 "test/minimum-nd.cc",
7199 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07007200 deps = OPERATOR_TEST_DEPS,
Marat Dukhan79e7f842019-12-05 14:35:50 -08007201)
7202
7203xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08007204 name = "multiply_nd_test",
Marat Dukhanca2733c2019-11-15 23:21:17 -08007205 srcs = [
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08007206 "test/binary-elementwise-operator-tester.h",
Marat Dukhanefc47b82019-11-18 09:25:38 -08007207 "test/multiply-nd.cc",
Marat Dukhanca2733c2019-11-15 23:21:17 -08007208 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07007209 deps = OPERATOR_TEST_DEPS,
Marat Dukhanca2733c2019-11-15 23:21:17 -08007210)
7211
7212xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07007213 name = "negate_nc_test",
7214 srcs = [
7215 "test/negate-nc.cc",
7216 "test/negate-operator-tester.h",
7217 ],
7218 deps = OPERATOR_TEST_DEPS,
7219)
7220
7221xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08007222 name = "prelu_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007223 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08007224 "test/prelu-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007225 "test/prelu-operator-tester.h",
7226 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07007227 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07007228)
7229
7230xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08007231 name = "resize_bilinear_nhwc_test",
Marat Dukhan69722492019-11-11 19:55:50 -08007232 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08007233 "test/resize-bilinear-nhwc.cc",
Marat Dukhan69722492019-11-11 19:55:50 -08007234 "test/resize-bilinear-operator-tester.h",
7235 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07007236 deps = OPERATOR_TEST_DEPS,
Marat Dukhan69722492019-11-11 19:55:50 -08007237)
7238
7239xnnpack_unit_test(
Artsiom Ablavatski97918102020-10-27 15:52:59 -07007240 name = "resize_bilinear_nchw_test",
7241 srcs = [
7242 "test/resize-bilinear-nchw.cc",
7243 "test/resize-bilinear-operator-tester.h",
7244 ] + OPERATOR_TEST_PARAMS_HDRS,
7245 deps = OPERATOR_TEST_DEPS,
7246)
7247
7248xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08007249 name = "sigmoid_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007250 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08007251 "test/sigmoid-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007252 "test/sigmoid-operator-tester.h",
7253 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07007254 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07007255)
7256
7257xnnpack_unit_test(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08007258 name = "softmax_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007259 srcs = [
Marat Dukhanfd8e6892020-01-27 15:25:25 -08007260 "test/softmax-nc.cc",
7261 "test/softmax-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007262 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07007263 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07007264)
7265
7266xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07007267 name = "square_nc_test",
7268 srcs = [
7269 "test/square-nc.cc",
7270 "test/square-operator-tester.h",
7271 ],
7272 deps = OPERATOR_TEST_DEPS,
7273)
7274
7275xnnpack_unit_test(
Marat Dukhan6804bbd2020-06-30 19:26:11 -07007276 name = "square_root_nc_test",
7277 srcs = [
7278 "test/square-root-nc.cc",
7279 "test/square-root-operator-tester.h",
7280 ],
7281 deps = OPERATOR_TEST_DEPS,
7282)
7283
7284xnnpack_unit_test(
Marat Dukhanf7399262020-06-05 10:58:44 -07007285 name = "squared_difference_nd_test",
7286 srcs = [
7287 "test/binary-elementwise-operator-tester.h",
7288 "test/squared-difference-nd.cc",
7289 ],
7290 deps = OPERATOR_TEST_DEPS,
7291)
7292
7293xnnpack_unit_test(
Marat Dukhan05f3f6d2019-12-03 15:13:53 -08007294 name = "subtract_nd_test",
7295 srcs = [
7296 "test/binary-elementwise-operator-tester.h",
7297 "test/subtract-nd.cc",
7298 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07007299 deps = OPERATOR_TEST_DEPS,
Marat Dukhan05f3f6d2019-12-03 15:13:53 -08007300)
7301
7302xnnpack_unit_test(
Marat Dukhan64e52512020-06-09 13:41:16 -07007303 name = "truncation_nc_test",
7304 srcs = [
7305 "test/truncation-nc.cc",
7306 "test/truncation-operator-tester.h",
7307 ],
7308 deps = OPERATOR_TEST_DEPS,
7309)
7310
7311xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08007312 name = "unpooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007313 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08007314 "test/unpooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007315 "test/unpooling-operator-tester.h",
7316 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07007317 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07007318)
7319
Chao Mei6ddfc602020-05-13 22:29:36 -07007320############################### Misc unit tests ###############################
7321
7322xnnpack_unit_test(
7323 name = "memory_planner_test",
7324 srcs = [
7325 "test/memory-planner-test.cc",
7326 ],
7327 deps = [
7328 ":XNNPACK",
7329 ":memory_planner",
7330 ],
7331)
7332
XNNPACK Teamab8c4c82020-10-09 08:05:51 -07007333xnnpack_unit_test(
7334 name = "subgraph_nchw_test",
7335 srcs = [
7336 "src/xnnpack/subgraph.h",
7337 "test/subgraph-nchw.cc",
7338 "test/subgraph-tester.h",
7339 ],
7340 deps = [
7341 ":XNNPACK",
7342 ],
7343)
7344
Marat Dukhan08c4a432019-10-03 09:29:21 -07007345############################# Build configurations #############################
7346
Marat Dukhanb8642352019-10-30 15:43:02 -07007347# Enables usage of assembly kernels.
Marat Dukhan08c4a432019-10-03 09:29:21 -07007348config_setting(
Marat Dukhanb8642352019-10-30 15:43:02 -07007349 name = "xnn_enable_assembly_explicit_true",
7350 define_values = {"xnn_enable_assembly": "true"},
7351)
7352
7353# Disables usage of assembly kernels.
7354config_setting(
7355 name = "xnn_enable_assembly_explicit_false",
7356 define_values = {"xnn_enable_assembly": "false"},
7357)
7358
Marat Dukhan9de90e02020-06-18 16:04:12 -07007359# Enables usage of sparse inference.
7360config_setting(
7361 name = "xnn_enable_sparse_explicit_true",
7362 define_values = {"xnn_enable_sparse": "true"},
7363)
7364
7365# Disables usage of sparse inference.
7366config_setting(
7367 name = "xnn_enable_sparse_explicit_false",
7368 define_values = {"xnn_enable_sparse": "false"},
7369)
7370
Marat Dukhan05702cf2020-03-26 15:41:33 -07007371# Disables usage of HMP-aware optimizations.
7372config_setting(
7373 name = "xnn_enable_hmp_explicit_false",
7374 define_values = {"xnn_enable_hmp": "false"},
7375)
7376
Chao Mei6ddfc602020-05-13 22:29:36 -07007377# Enable usage of optimized memory allocation
7378config_setting(
7379 name = "xnn_enable_memopt_explicit_true",
7380 define_values = {"xnnpack_enable_memopt": "true"},
7381)
7382
7383# Disable usage of optimized memory allocation
7384config_setting(
7385 name = "xnn_enable_memopt_explicit_false",
7386 define_values = {"xnnpack_enable_memopt": "false"},
7387)
7388
Marat Dukhanb8642352019-10-30 15:43:02 -07007389# Builds with -c dbg
7390config_setting(
7391 name = "debug_build",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007392 values = {
Marat Dukhanb8642352019-10-30 15:43:02 -07007393 "compilation_mode": "dbg",
7394 },
7395)
7396
7397# Builds with -c opt
7398config_setting(
7399 name = "optimized_build",
7400 values = {
7401 "compilation_mode": "opt",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007402 },
7403)
7404
7405config_setting(
Marat Dukhanb8642352019-10-30 15:43:02 -07007406 name = "linux_k8",
7407 values = {"cpu": "k8"},
7408)
7409
7410config_setting(
Marat Dukhan582094e2020-04-30 17:21:25 -07007411 name = "linux_arm",
7412 values = {"cpu": "arm"},
Marat Dukhan4e45e662019-10-03 15:40:24 -07007413)
7414
7415config_setting(
Marat Dukhanf0bd4de2020-06-15 15:53:19 -07007416 name = "linux_armeabi",
7417 values = {"cpu": "armeabi"},
7418)
7419
7420config_setting(
Terry Heo68eef3f2020-04-13 22:53:52 -07007421 name = "linux_armhf",
7422 values = {"cpu": "armhf"},
7423)
7424
7425config_setting(
Marat Dukhana720e932020-06-10 13:01:11 -07007426 name = "linux_armv7a",
7427 values = {"cpu": "armv7a"},
7428)
7429
7430config_setting(
Marat Dukhan582094e2020-04-30 17:21:25 -07007431 name = "linux_aarch64",
7432 values = {"cpu": "aarch64"},
7433)
7434
7435config_setting(
Marat Dukhan08c4a432019-10-03 09:29:21 -07007436 name = "android",
7437 values = {"crosstool_top": "//external:android/crosstool"},
7438)
7439
7440config_setting(
7441 name = "android_armv7",
7442 values = {
7443 "crosstool_top": "//external:android/crosstool",
7444 "cpu": "armeabi-v7a",
7445 },
7446)
7447
7448config_setting(
7449 name = "android_arm64",
7450 values = {
7451 "crosstool_top": "//external:android/crosstool",
7452 "cpu": "arm64-v8a",
7453 },
7454)
7455
7456config_setting(
7457 name = "android_x86",
7458 values = {
7459 "crosstool_top": "//external:android/crosstool",
7460 "cpu": "x86",
7461 },
7462)
7463
7464config_setting(
7465 name = "android_x86_64",
7466 values = {
7467 "crosstool_top": "//external:android/crosstool",
7468 "cpu": "x86_64",
7469 },
7470)
7471
7472config_setting(
Marat Dukhan10a38082020-04-17 03:58:35 -07007473 name = "windows_x86_64",
7474 values = {"cpu": "x64_windows"},
Marat Dukhan9fe932e2020-04-11 17:14:15 -07007475)
7476
7477config_setting(
Marat Dukhan10a38082020-04-17 03:58:35 -07007478 name = "windows_x86_64_clang",
7479 values = {
7480 "compiler": "clang-cl",
7481 "cpu": "x64_windows",
7482 },
7483)
7484
7485config_setting(
7486 name = "windows_x86_64_mingw",
7487 values = {
7488 "compiler": "mingw-gcc",
7489 "cpu": "x64_windows",
7490 },
7491)
7492
7493config_setting(
7494 name = "windows_x86_64_msys",
7495 values = {
7496 "compiler": "msys-gcc",
7497 "cpu": "x64_windows",
7498 },
Marat Dukhan9fe932e2020-04-11 17:14:15 -07007499)
7500
7501config_setting(
Marat Dukhan885ca242019-10-07 09:17:32 -07007502 name = "macos_x86_64",
7503 values = {
7504 "apple_platform_type": "macos",
7505 "cpu": "darwin",
7506 },
7507)
7508
7509config_setting(
Simon Maurerae33ab82021-03-03 23:38:22 +01007510 name = "macos_arm64",
7511 values = {
7512 "apple_platform_type": "macos",
7513 "cpu": "darwin_arm64",
7514 },
7515)
7516
7517config_setting(
Marat Dukhan08c4a432019-10-03 09:29:21 -07007518 name = "emscripten",
XNNPACK Teamf69ac5d2021-03-03 16:47:33 -08007519 values = {"crosstool_top": "//emscripten_toolchain:everything"},
Marat Dukhan08c4a432019-10-03 09:29:21 -07007520)
7521
7522config_setting(
7523 name = "emscripten_wasm",
7524 values = {
XNNPACK Teamf69ac5d2021-03-03 16:47:33 -08007525 "crosstool_top": "//emscripten_toolchain:everything",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007526 "cpu": "wasm",
7527 },
7528)
7529
7530config_setting(
7531 name = "emscripten_wasmsimd",
7532 values = {
XNNPACK Teamf69ac5d2021-03-03 16:47:33 -08007533 "crosstool_top": "//emscripten_toolchain:everything",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007534 "cpu": "wasm",
Marat Dukhan81c62602020-05-29 13:22:49 -07007535 "copt": "-msimd128",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007536 },
7537)
7538
7539config_setting(
Marat Dukhan1498d1d2020-02-11 20:00:05 -08007540 name = "ios_armv7",
7541 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08007542 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08007543 "cpu": "ios_armv7",
7544 },
7545)
7546
7547config_setting(
7548 name = "ios_arm64",
7549 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08007550 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08007551 "cpu": "ios_arm64",
7552 },
7553)
7554
7555config_setting(
7556 name = "ios_arm64e",
7557 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08007558 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08007559 "cpu": "ios_arm64e",
7560 },
7561)
7562
7563config_setting(
7564 name = "ios_x86",
7565 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08007566 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08007567 "cpu": "ios_i386",
7568 },
7569)
7570
7571config_setting(
7572 name = "ios_x86_64",
7573 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08007574 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08007575 "cpu": "ios_x86_64",
7576 },
7577)
7578
7579config_setting(
7580 name = "watchos_armv7k",
7581 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08007582 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08007583 "cpu": "watchos_armv7k",
7584 },
7585)
7586
7587config_setting(
7588 name = "watchos_arm64_32",
7589 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08007590 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08007591 "cpu": "watchos_arm64_32",
7592 },
7593)
7594
7595config_setting(
7596 name = "watchos_x86",
7597 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08007598 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08007599 "cpu": "watchos_i386",
7600 },
7601)
7602
7603config_setting(
7604 name = "watchos_x86_64",
7605 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08007606 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08007607 "cpu": "watchos_x86_64",
7608 },
7609)
7610
7611config_setting(
7612 name = "tvos_arm64",
7613 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08007614 "apple_platform_type": "tvos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08007615 "cpu": "tvos_arm64",
7616 },
7617)
7618
7619config_setting(
7620 name = "tvos_x86_64",
7621 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08007622 "apple_platform_type": "tvos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08007623 "cpu": "tvos_x86_64",
7624 },
7625)